diff options
author | kobalicek <kobalicek.petr@gmail.com> | 2021-03-14 00:25:43 +0300 |
---|---|---|
committer | kobalicek <kobalicek.petr@gmail.com> | 2021-03-14 01:05:48 +0300 |
commit | 7836449c3099dd564c8cc7666f640a1532151af2 (patch) | |
tree | 06c6c0f7a0e3dd6709aa25bf6337e09b279dff0e | |
parent | c9cebc67bc195060f56f0a4f7fa9f066a0901e25 (diff) |
Added asmjit_test_perf, which replaces asmjit_bench and provides much better performance overview
Removed asmjit_test_opcode (not needed anymore as we have asmjit_test_assembler and asmjit_test_perf)
-rw-r--r-- | .github/workflows/build-config.json | 8 | ||||
-rw-r--r-- | CMakeLists.txt | 38 | ||||
-rw-r--r-- | test/asmjit_bench_x86.cpp | 177 | ||||
-rw-r--r-- | test/asmjit_test_compiler.cpp | 23 | ||||
-rw-r--r-- | test/asmjit_test_compiler_x86.cpp | 2 | ||||
-rw-r--r-- | test/asmjit_test_misc.h | 179 | ||||
-rw-r--r-- | test/asmjit_test_opcode.cpp | 115 | ||||
-rw-r--r-- | test/asmjit_test_opcode.h | 6055 | ||||
-rw-r--r-- | test/asmjit_test_perf.cpp | 69 | ||||
-rw-r--r-- | test/asmjit_test_perf.h | 81 | ||||
-rw-r--r-- | test/asmjit_test_perf_x86.cpp | 5049 | ||||
-rw-r--r-- | test/cmdline.h | 6 | ||||
-rw-r--r-- | test/performancetimer.h | 59 |
13 files changed, 5421 insertions, 6440 deletions
diff --git a/.github/workflows/build-config.json b/.github/workflows/build-config.json index 857ddeb..3c76bd0 100644 --- a/.github/workflows/build-config.json +++ b/.github/workflows/build-config.json @@ -16,15 +16,15 @@ "optional": true }, { - "cmd": ["asmjit_test_opcode", "--quiet"], + "cmd": ["asmjit_test_assembler", "--quiet"], "optional": true }, { - "cmd": ["asmjit_test_assembler", "--quiet"], + "cmd": ["asmjit_test_emitters"], "optional": true }, { - "cmd": ["asmjit_test_emitters"], + "cmd": ["asmjit_test_compiler"], "optional": true }, { @@ -36,7 +36,7 @@ "optional": true }, { - "cmd": ["asmjit_test_compiler"], + "cmd": ["asmjit_test_perf", "--quick"], "optional": true } ] diff --git a/CMakeLists.txt b/CMakeLists.txt index 2153f3f..5ba57ea 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -497,17 +497,6 @@ if (NOT ASMJIT_EMBED) CFLAGS_REL ${ASMJIT_PRIVATE_CFLAGS_REL}) target_include_directories(asmjit_test_unit BEFORE PRIVATE ${ASMJIT_INCLUDE_DIRS}) - foreach(_target asmjit_test_opcode - asmjit_test_emitters - asmjit_test_x86_sections) - asmjit_add_target(${_target} TEST - SOURCES test/${_target}.cpp - LIBRARIES asmjit::asmjit - CFLAGS ${ASMJIT_PRIVATE_CFLAGS} - CFLAGS_DBG ${ASMJIT_PRIVATE_CFLAGS_DBG} - CFLAGS_REL ${ASMJIT_PRIVATE_CFLAGS_REL}) - endforeach() - asmjit_add_target(asmjit_test_assembler TEST SOURCES test/asmjit_test_assembler.cpp test/asmjit_test_assembler_x64.cpp @@ -518,6 +507,25 @@ if (NOT ASMJIT_EMBED) CFLAGS_DBG ${ASMJIT_PRIVATE_CFLAGS_DBG} CFLAGS_REL ${ASMJIT_PRIVATE_CFLAGS_REL}) + asmjit_add_target(asmjit_test_perf EXECUTABLE + SOURCES test/asmjit_test_perf.cpp + test/asmjit_test_perf_x86.cpp + SOURCES test/asmjit_test_perf.h + LIBRARIES asmjit::asmjit + CFLAGS ${ASMJIT_PRIVATE_CFLAGS} + CFLAGS_DBG ${ASMJIT_PRIVATE_CFLAGS_DBG} + CFLAGS_REL ${ASMJIT_PRIVATE_CFLAGS_REL}) + + foreach(_target asmjit_test_emitters + asmjit_test_x86_sections) + asmjit_add_target(${_target} TEST + SOURCES test/${_target}.cpp + LIBRARIES asmjit::asmjit + CFLAGS ${ASMJIT_PRIVATE_CFLAGS} + CFLAGS_DBG ${ASMJIT_PRIVATE_CFLAGS_DBG} + CFLAGS_REL ${ASMJIT_PRIVATE_CFLAGS_REL}) + endforeach() + if (NOT ASMJIT_NO_INTROSPECTION) asmjit_add_target(asmjit_test_instinfo TEST SOURCES test/asmjit_test_instinfo.cpp @@ -545,14 +553,6 @@ if (NOT ASMJIT_EMBED) CFLAGS_REL ${ASMJIT_PRIVATE_CFLAGS_REL}) endif() - foreach(_target asmjit_bench_x86) - asmjit_add_target(${_target} EXECUTABLE - SOURCES test/${_target}.cpp - LIBRARIES asmjit::asmjit - CFLAGS ${ASMJIT_PRIVATE_CFLAGS} - CFLAGS_DBG ${ASMJIT_PRIVATE_CFLAGS_DBG} - CFLAGS_REL ${ASMJIT_PRIVATE_CFLAGS_REL}) - endforeach() endif() endif() diff --git a/test/asmjit_bench_x86.cpp b/test/asmjit_bench_x86.cpp deleted file mode 100644 index 6f75916..0000000 --- a/test/asmjit_bench_x86.cpp +++ /dev/null @@ -1,177 +0,0 @@ -// AsmJit - Machine code generation for C++ -// -// * Official AsmJit Home Page: https://asmjit.com -// * Official Github Repository: https://github.com/asmjit/asmjit -// -// Copyright (c) 2008-2020 The AsmJit Authors -// -// This software is provided 'as-is', without any express or implied -// warranty. In no event will the authors be held liable for any damages -// arising from the use of this software. -// -// Permission is granted to anyone to use this software for any purpose, -// including commercial applications, and to alter it and redistribute it -// freely, subject to the following restrictions: -// -// 1. The origin of this software must not be misrepresented; you must not -// claim that you wrote the original software. If you use this software -// in a product, an acknowledgment in the product documentation would be -// appreciated but is not required. -// 2. Altered source versions must be plainly marked as such, and must not be -// misrepresented as being the original software. -// 3. This notice may not be removed or altered from any source distribution. - -#include <asmjit/core.h> - -#ifdef ASMJIT_BUILD_X86 -#include <asmjit/x86.h> -#endif - -#include <stdio.h> -#include <string.h> - -#include "asmjit_test_opcode.h" - -#ifndef ASMJIT_NO_COMPILER - #include "asmjit_test_misc.h" -#endif - -using namespace asmjit; - -// ============================================================================ -// [Configuration] -// ============================================================================ - -static constexpr uint32_t kNumRepeats = 20; -static constexpr uint32_t kNumIterations = 1000; - -// ============================================================================ -// [BenchUtils] -// ============================================================================ - -namespace BenchUtils { - class Performance { - public: - inline Performance() noexcept { reset(); } - - inline void reset() noexcept { - tick = 0u; - best = 0xFFFFFFFFu; - } - - inline uint32_t start() noexcept { return (tick = now()); } - inline uint32_t diff() const noexcept { return now() - tick; } - - inline uint32_t end() noexcept { - tick = diff(); - if (best > tick) - best = tick; - return tick; - } - - static inline uint32_t now() noexcept { - return OSUtils::getTickCount(); - } - - uint32_t tick; - uint32_t best; - }; - - static double mbps(uint32_t time, uint64_t outputSize) noexcept { - if (!time) return 0.0; - - double bytesTotal = double(outputSize); - return (bytesTotal * 1000) / (double(time) * 1024 * 1024); - } - - template<typename EmitterT, typename FuncT> - static void bench(CodeHolder& code, uint32_t arch, const char* testName, const FuncT& func) noexcept { - EmitterT emitter; - - const char* archName = - arch == Environment::kArchX86 ? "X86" : - arch == Environment::kArchX64 ? "X64" : "???"; - - const char* emitterName = - emitter.isAssembler() ? "Assembler" : - emitter.isCompiler() ? "Compiler" : - emitter.isBuilder() ? "Builder" : "Unknown"; - - Performance perf; - uint64_t codeSize = 0; - - Environment env(arch); - - for (uint32_t r = 0; r < kNumRepeats; r++) { - perf.start(); - codeSize = 0; - for (uint32_t i = 0; i < kNumIterations; i++) { - code.init(env); - code.attach(&emitter); - - func(emitter); - codeSize += code.codeSize(); - - code.reset(); - } - perf.end(); - } - - printf("[%s] %-9s %-10s | Time:%6u [ms] | ", archName, emitterName, testName, perf.best); - if (codeSize) - printf("Speed: %7.3f [MB/s]", mbps(perf.best, codeSize)); - else - printf("Speed: N/A"); - printf("\n"); - } -} - -// ============================================================================ -// [Main] -// ============================================================================ - -#ifdef ASMJIT_BUILD_X86 -static void benchX86(uint32_t arch) noexcept { - CodeHolder code; - - BenchUtils::bench<x86::Assembler>(code, arch, "[fast]", [](x86::Assembler& a) { - asmtest::generateOpcodes(a.as<x86::Emitter>()); - }); - - BenchUtils::bench<x86::Assembler>(code, arch, "[validate]", [](x86::Assembler& a) { - a.addValidationOptions(BaseEmitter::kValidationOptionAssembler); - asmtest::generateOpcodes(a.as<x86::Emitter>()); - }); - -#ifndef ASMJIT_NO_BUILDER - BenchUtils::bench<x86::Builder>(code, arch, "[no-asm]", [](x86::Builder& cb) { - asmtest::generateOpcodes(cb.as<x86::Emitter>()); - }); - - BenchUtils::bench<x86::Builder>(code, arch, "[asm]", [](x86::Builder& cb) { - asmtest::generateOpcodes(cb.as<x86::Emitter>()); - cb.finalize(); - }); -#endif - -#ifndef ASMJIT_NO_COMPILER - BenchUtils::bench<x86::Compiler>(code, arch, "[no-asm]", [](x86::Compiler& cc) { - asmtest::generateAlphaBlend(cc); - }); - - BenchUtils::bench<x86::Compiler>(code, arch, "[asm]", [](x86::Compiler& cc) { - asmtest::generateAlphaBlend(cc); - cc.finalize(); - }); -#endif -} -#endif - -int main() { -#ifdef ASMJIT_BUILD_X86 - benchX86(Environment::kArchX86); - benchX86(Environment::kArchX64); -#endif - - return 0; -} diff --git a/test/asmjit_test_compiler.cpp b/test/asmjit_test_compiler.cpp index c13ba18..e7a5eeb 100644 --- a/test/asmjit_test_compiler.cpp +++ b/test/asmjit_test_compiler.cpp @@ -32,6 +32,8 @@ #include <chrono> #include "cmdline.h" +#include "performancetimer.h" + #include "asmjit_test_compiler.h" #if defined(ASMJIT_BUILD_X86) && ASMJIT_ARCH_X86 @@ -54,27 +56,6 @@ void compiler_add_a64_tests(TestApp& app); using namespace asmjit; -class PerformanceTimer { -public: - typedef std::chrono::high_resolution_clock::time_point TimePoint; - - TimePoint _startTime {}; - TimePoint _endTime {}; - - inline void start() { - _startTime = std::chrono::high_resolution_clock::now(); - } - - inline void stop() { - _endTime = std::chrono::high_resolution_clock::now(); - } - - inline double duration() const { - std::chrono::duration<double> elapsed = _endTime - _startTime; - return elapsed.count() * 1000; - } -}; - // ============================================================================ // [TestApp] // ============================================================================ diff --git a/test/asmjit_test_compiler_x86.cpp b/test/asmjit_test_compiler_x86.cpp index 75b465e..1c82e72 100644 --- a/test/asmjit_test_compiler_x86.cpp +++ b/test/asmjit_test_compiler_x86.cpp @@ -2291,7 +2291,7 @@ public: } virtual void compile(x86::Compiler& cc) { - asmtest::generateAlphaBlend(cc); + asmtest::generateSseAlphaBlend(cc, true); } virtual bool run(void* _func, String& result, String& expect) { diff --git a/test/asmjit_test_misc.h b/test/asmjit_test_misc.h index 0327327..f0d156f 100644 --- a/test/asmjit_test_misc.h +++ b/test/asmjit_test_misc.h @@ -28,42 +28,42 @@ namespace asmtest { -// Generate a typical alpha blend function using SSE2 instruction set. Used -// for benchmarking and also in test86. The generated code should be stable -// and fully functional. -static void generateAlphaBlend(asmjit::x86::Compiler& cc) { - using namespace asmjit; - using namespace asmjit::x86; - - Gp dst = cc.newIntPtr("dst"); - Gp src = cc.newIntPtr("src"); - - Gp i = cc.newIntPtr("i"); - Gp j = cc.newIntPtr("j"); - Gp t = cc.newIntPtr("t"); - - Xmm vzero = cc.newXmm("vzero"); - Xmm v0080 = cc.newXmm("v0080"); - Xmm v0101 = cc.newXmm("v0101"); +using namespace asmjit; + +// Generates a typical alpha blend function that uses SSE2 instruction set. +// This function combines emitting instructions with control flow constructs +// like binding Labels and jumping to them. This should be pretty representative. +template<typename Emitter> +static void generateSseAlphaBlendInternal( + Emitter& cc, + const x86::Gp& dst, const x86::Gp& src, const x86::Gp& n, + const x86::Gp& gp0, + const x86::Xmm& simd0, const x86::Xmm& simd1, const x86::Xmm& simd2, const x86::Xmm& simd3, + const x86::Xmm& simd4, const x86::Xmm& simd5, const x86::Xmm& simd6, const x86::Xmm& simd7) { + + x86::Gp i = n; + x86::Gp j = gp0; + + x86::Xmm vzero = simd0; + x86::Xmm v0080 = simd1; + x86::Xmm v0101 = simd2; Label L_SmallLoop = cc.newLabel(); Label L_SmallEnd = cc.newLabel(); Label L_LargeLoop = cc.newLabel(); Label L_LargeEnd = cc.newLabel(); - Label L_DataPool = cc.newLabel(); + Label L_Done = cc.newLabel(); - cc.addFunc(FuncSignatureT<void, void*, const void*, size_t>(CallConv::kIdHost)); - - cc.setArg(0, dst); - cc.setArg(1, src); - cc.setArg(2, i); - - // How many pixels have to be processed to make the loop aligned. - cc.lea(t, x86::ptr(L_DataPool)); + // Load SIMD Constants. cc.xorps(vzero, vzero); - cc.movaps(v0080, x86::ptr(t, 0)); - cc.movaps(v0101, x86::ptr(t, 16)); + cc.mov(gp0.r32(), 0x00800080); + cc.movd(v0080, gp0.r32()); + cc.mov(gp0.r32(), 0x01010101); + cc.movd(v0101, gp0.r32()); + cc.pshufd(v0080, v0080, x86::Predicate::shuf(0, 0, 0, 0)); + cc.pshufd(v0101, v0101, x86::Predicate::shuf(0, 0, 0, 0)); + // How many pixels have to be processed to make the loop aligned. cc.xor_(j, j); cc.sub(j, dst); cc.and_(j, 15); @@ -71,15 +71,15 @@ static void generateAlphaBlend(asmjit::x86::Compiler& cc) { cc.jz(L_SmallEnd); cc.cmp(j, i); - cc.cmovg(j, i); // j = min(i, j). - cc.sub(i, j); // i -= j. + cc.cmovg(j, i); // j = min(i, j) + cc.sub(i, j); // i -= j // Small loop. cc.bind(L_SmallLoop); { - Xmm x0 = cc.newXmm("x0"); - Xmm y0 = cc.newXmm("y0"); - Xmm a0 = cc.newXmm("a0"); + x86::Xmm x0 = simd3; + x86::Xmm y0 = simd4; + x86::Xmm a0 = simd5; cc.movd(y0, x86::ptr(src)); cc.movd(x0, x86::ptr(dst)); @@ -113,7 +113,7 @@ static void generateAlphaBlend(asmjit::x86::Compiler& cc) { cc.test(i, i); cc.mov(j, i); - cc.jz(cc.func()->exitLabel()); + cc.jz(L_Done); cc.and_(j, 3); cc.shr(i, 2); @@ -122,11 +122,11 @@ static void generateAlphaBlend(asmjit::x86::Compiler& cc) { // Aligned loop. cc.bind(L_LargeLoop); { - Xmm x0 = cc.newXmm("x0"); - Xmm x1 = cc.newXmm("x1"); - Xmm y0 = cc.newXmm("y0"); - Xmm a0 = cc.newXmm("a0"); - Xmm a1 = cc.newXmm("a1"); + x86::Xmm x0 = simd3; + x86::Xmm x1 = simd4; + x86::Xmm y0 = simd5; + x86::Xmm a0 = simd6; + x86::Xmm a1 = simd7; cc.movups(y0, x86::ptr(src)); cc.movaps(x0, x86::ptr(dst)); @@ -172,13 +172,102 @@ static void generateAlphaBlend(asmjit::x86::Compiler& cc) { cc.test(j, j); cc.jnz(L_SmallLoop); - cc.endFunc(); + cc.bind(L_Done); +} + +static void generateSseAlphaBlend(asmjit::BaseEmitter& emitter, bool emitPrologEpilog) { + using namespace asmjit::x86; - // Data. - cc.align(kAlignData, 16); - cc.bind(L_DataPool); - cc.embedUInt16(uint16_t(0x0080u), 8); - cc.embedUInt16(uint16_t(0x0101u), 8); + if (emitter.isAssembler()) { + Assembler& cc = *emitter.as<Assembler>(); + + x86::Gp dst = cc.zax(); + x86::Gp src = cc.zcx(); + x86::Gp i = cc.zdx(); + x86::Gp j = cc.zdi(); + + if (emitPrologEpilog) { + FuncDetail func; + func.init(FuncSignatureT<void, void*, const void*, size_t>(CallConv::kIdHost), cc.environment()); + + FuncFrame frame; + frame.init(func); + frame.addDirtyRegs(dst, src, i, j); + frame.addDirtyRegs(xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7); + + FuncArgsAssignment args(&func); + args.assignAll(dst, src, i); + args.updateFuncFrame(frame); + frame.finalize(); + + cc.emitProlog(frame); + cc.emitArgsAssignment(frame, args); + generateSseAlphaBlendInternal(cc, dst, src, i, j, xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7); + cc.emitEpilog(frame); + } + else { + generateSseAlphaBlendInternal(cc, dst, src, i, j, xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7); + } + } +#ifndef ASMJIT_NO_BUILDER + else if (emitter.isBuilder()) { + Builder& cc = *emitter.as<Builder>(); + + x86::Gp dst = cc.zax(); + x86::Gp src = cc.zcx(); + x86::Gp i = cc.zdx(); + x86::Gp j = cc.zdi(); + + if (emitPrologEpilog) { + FuncDetail func; + func.init(FuncSignatureT<void, void*, const void*, size_t>(CallConv::kIdHost), cc.environment()); + + FuncFrame frame; + frame.init(func); + frame.addDirtyRegs(dst, src, i, j); + frame.addDirtyRegs(xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7); + + FuncArgsAssignment args(&func); + args.assignAll(dst, src, i); + args.updateFuncFrame(frame); + frame.finalize(); + + cc.emitProlog(frame); + cc.emitArgsAssignment(frame, args); + generateSseAlphaBlendInternal(cc, dst, src, i, j, xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7); + cc.emitEpilog(frame); + } + else { + generateSseAlphaBlendInternal(cc, dst, src, i, j, xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7); + } + } +#endif +#ifndef ASMJIT_NO_COMPILER + else if (emitter.isCompiler()) { + Compiler& cc = *emitter.as<Compiler>(); + + Gp dst = cc.newIntPtr("dst"); + Gp src = cc.newIntPtr("src"); + Gp i = cc.newIntPtr("i"); + Gp j = cc.newIntPtr("j"); + + Xmm v0 = cc.newXmm("v0"); + Xmm v1 = cc.newXmm("v1"); + Xmm v2 = cc.newXmm("v2"); + Xmm v3 = cc.newXmm("v3"); + Xmm v4 = cc.newXmm("v4"); + Xmm v5 = cc.newXmm("v5"); + Xmm v6 = cc.newXmm("v6"); + Xmm v7 = cc.newXmm("v7"); + + cc.addFunc(FuncSignatureT<void, void*, const void*, size_t>(CallConv::kIdHost)); + cc.setArg(0, dst); + cc.setArg(1, src); + cc.setArg(2, i); + generateSseAlphaBlendInternal(cc, dst, src, i, j, v0, v1, v2, v3, v4, v5, v6, v7); + cc.endFunc(); + } +#endif } } // {asmtest} diff --git a/test/asmjit_test_opcode.cpp b/test/asmjit_test_opcode.cpp deleted file mode 100644 index 1cd67e2..0000000 --- a/test/asmjit_test_opcode.cpp +++ /dev/null @@ -1,115 +0,0 @@ -// AsmJit - Machine code generation for C++ -// -// * Official AsmJit Home Page: https://asmjit.com -// * Official Github Repository: https://github.com/asmjit/asmjit -// -// Copyright (c) 2008-2020 The AsmJit Authors -// -// This software is provided 'as-is', without any express or implied -// warranty. In no event will the authors be held liable for any damages -// arising from the use of this software. -// -// Permission is granted to anyone to use this software for any purpose, -// including commercial applications, and to alter it and redistribute it -// freely, subject to the following restrictions: -// -// 1. The origin of this software must not be misrepresented; you must not -// claim that you wrote the original software. If you use this software -// in a product, an acknowledgment in the product documentation would be -// appreciated but is not required. -// 2. Altered source versions must be plainly marked as such, and must not be -// misrepresented as being the original software. -// 3. This notice may not be removed or altered from any source distribution. - -// ---------------------------------------------------------------------------- -// This file is used to test opcodes generated by AsmJit. Output can be -// disassembled in your IDE or by your favorite disassembler. Instructions -// are grouped by category and then sorted alphabetically. -// ---------------------------------------------------------------------------- - -#include <asmjit/x86.h> -#include <stdio.h> -#include <stdlib.h> - -#include "cmdline.h" -#include "asmjit_test_opcode.h" - -using namespace asmjit; - -struct OpcodeDumpInfo { - uint32_t arch; - bool useRex1; - bool useRex2; -}; - -static const char* archToString(uint32_t arch) noexcept { - switch (arch & ~Environment::kArchBigEndianMask) { - case Environment::kArchX86 : return "X86"; - case Environment::kArchX64 : return "X64"; - case Environment::kArchARM : return "ARM"; - case Environment::kArchThumb : return "Thumb"; - case Environment::kArchAArch64 : return "AArch64"; - case Environment::kArchMIPS32_LE: return "MIPS32"; - case Environment::kArchMIPS64_LE: return "MIPS64"; - default: return "Unknown"; - } -} - -struct TestErrorHandler : public ErrorHandler { - virtual void handleError(Error err, const char* message, BaseEmitter* origin) { - (void)origin; - printf("ERROR 0x%08X: %s\n", err, message); - } -}; - -typedef void (*VoidFunc)(void); - -int main(int argc, char* argv[]) { - CmdLine cmdLine(argc, argv); - TestErrorHandler eh; - - OpcodeDumpInfo infoList[] = { - { Environment::kArchX86, false, false }, - { Environment::kArchX64, false, false }, - { Environment::kArchX64, false, true }, - { Environment::kArchX64, true , false }, - { Environment::kArchX64, true , true } - }; - - bool quiet = cmdLine.hasArg("--quiet"); - - for (uint32_t i = 0; i < ASMJIT_ARRAY_SIZE(infoList); i++) { - const OpcodeDumpInfo& info = infoList[i]; - - printf("Opcodes [ARCH=%s REX1=%s REX2=%s]\n", - archToString(info.arch), - info.useRex1 ? "true" : "false", - info.useRex2 ? "true" : "false"); - - CodeHolder code; - code.init(Environment(info.arch)); - code.setErrorHandler(&eh); - -#ifndef ASMJIT_NO_LOGGING - FileLogger logger(stdout); - logger.addFlags(FormatOptions::kFlagMachineCode); - if (!quiet) - code.setLogger(&logger); -#endif - - x86::Assembler a(&code); - asmtest::generateOpcodes(a.as<x86::Emitter>(), info.useRex1, info.useRex2); - - // If this is the host architecture the code generated can be executed - // for debugging purposes (the first instruction is ret anyway). - if (code.arch() == Environment::kArchHost) { - JitRuntime runtime; - VoidFunc p; - - Error err = runtime.add(&p, &code); - if (err == kErrorOk) p(); - } - } - - return 0; -} diff --git a/test/asmjit_test_opcode.h b/test/asmjit_test_opcode.h deleted file mode 100644 index 3768cb1..0000000 --- a/test/asmjit_test_opcode.h +++ /dev/null @@ -1,6055 +0,0 @@ -// AsmJit - Machine code generation for C++ -// -// * Official AsmJit Home Page: https://asmjit.com -// * Official Github Repository: https://github.com/asmjit/asmjit -// -// Copyright (c) 2008-2020 The AsmJit Authors -// -// This software is provided 'as-is', without any express or implied -// warranty. In no event will the authors be held liable for any damages -// arising from the use of this software. -// -// Permission is granted to anyone to use this software for any purpose, -// including commercial applications, and to alter it and redistribute it -// freely, subject to the following restrictions: -// -// 1. The origin of this software must not be misrepresented; you must not -// claim that you wrote the original software. If you use this software -// in a product, an acknowledgment in the product documentation would be -// appreciated but is not required. -// 2. Altered source versions must be plainly marked as such, and must not be -// misrepresented as being the original software. -// 3. This notice may not be removed or altered from any source distribution. - -#ifndef ASMJIT_TEST_OPCODE_H_INCLUDED -#define ASMJIT_TEST_OPCODE_H_INCLUDED - -#include <asmjit/x86.h> - -namespace asmtest { - -// Generate all instructions asmjit can emit. -static void generateOpcodes(asmjit::x86::Emitter* e, bool useRex1 = false, bool useRex2 = false) { - using namespace asmjit; - using namespace asmjit::x86; - - bool isX64 = e->is64Bit(); - - // Prevent a crash when the generated function is called to see the disassembly. - e->ret(); - - // All instructions use the following register that can be changed to see if - // `x86::Assembler` can properly encode all possible combinations. If the given - // `useRexRegs` argument is true the `A` version will in most cases contain - // a register having index 8 or greater to force REX prefix. - Gp gLoA = useRex1 ? r8b : al; - Gp gLoB = useRex2 ? r9b : bl; - - Gp gHiA = ah; - Gp gHiB = bh; - - Gp gwA = useRex1 ? r8w : ax; - Gp gwB = useRex2 ? r9w : bx; - - Gp gdA = useRex1 ? r8d : eax; - Gp gdB = useRex2 ? r9d : ebx; - Gp gdC = useRex2 ? r10d : ecx; - - Gp gzA = useRex1 ? r8.as<Gp>() : e->zax(); - Gp gzB = useRex2 ? r9.as<Gp>() : e->zbx(); - Gp gzC = useRex2 ? r10.as<Gp>() : e->zcx(); - Gp gzD = useRex2 ? r11.as<Gp>() : e->zdx(); - - KReg kA = k1; - KReg kB = k2; - KReg kC = k3; - - Mem anyptr_gpA = ptr(gzA); - Mem anyptr_gpB = ptr(gzB); - Mem anyptr_gpC = ptr(gzC); - Mem anyptr_gpD = ptr(gzD); - - Mem intptr_gpA = e->intptr_ptr(gzA); - Mem intptr_gpB = e->intptr_ptr(gzB); - - St stA = st0; - St stB = st7; - - Mm mmA = mm0; - Mm mmB = mm1; - - Xmm xmmA = useRex1 ? xmm8 : xmm0; - Xmm xmmB = useRex2 ? xmm9 : xmm1; - Xmm xmmC = useRex2 ? xmm10 : xmm2; - Xmm xmmD = useRex2 ? xmm11 : xmm3; - - Ymm ymmA = useRex1 ? ymm8 : ymm0; - Ymm ymmB = useRex2 ? ymm9 : ymm1; - Ymm ymmC = useRex2 ? ymm10 : ymm2; - Ymm ymmD = useRex2 ? ymm11 : ymm3; - - Zmm zmmA = useRex1 ? zmm8 : zmm0; - Zmm zmmB = useRex2 ? zmm9 : zmm1; - Zmm zmmC = useRex2 ? zmm10 : zmm2; - - Mem vx_ptr = ptr(gzB, xmmB); - Mem vy_ptr = ptr(gzB, ymmB); - Mem vz_ptr = ptr(gzB, zmmB); - - Label L; - - // Base. - e->adc(gLoA, 1); - e->adc(gLoB, 1); - e->adc(gHiA, 1); - e->adc(gHiB, 1); - e->adc(gwA, 1); - e->adc(gwB, 1); - e->adc(gdA, 1); - e->adc(gdB, 1); - e->adc(gzA, 1); - e->adc(gzA, gzB); - e->adc(gzA, intptr_gpB); - e->adc(intptr_gpA, 1); - e->adc(intptr_gpA, gzB); - e->add(gLoA, 1); - e->add(gLoB, 1); - e->add(gHiA, 1); - e->add(gHiB, 1); - e->add(gwA, 1); - e->add(gwB, 1); - e->add(gdA, 1); - e->add(gdB, 1); - e->add(gzA, 1); - e->add(gzA, gzB); - e->add(gzA, intptr_gpB); - e->add(intptr_gpA, 1); - e->add(intptr_gpA, gzB); - e->and_(gLoA, 1); - e->and_(gLoB, 1); - e->and_(gHiA, 1); - e->and_(gHiB, 1); - e->and_(gwA, 1); - e->and_(gwB, 1); - e->and_(gdA, 1); - e->and_(gdB, 1); - e->and_(gzA, 1); - e->and_(gzA, gzB); - e->and_(gzA, intptr_gpB); - e->and_(intptr_gpA, 1); - e->and_(intptr_gpA, gzB); - e->bswap(gzA); - e->bt(gdA, 1); - e->bt(gzA, 1); - e->bt(gdA, gdB); - e->bt(gzA, gzB); - e->bt(intptr_gpA, 1); - e->bt(anyptr_gpA, gdB); - e->bt(intptr_gpA, gzB); - e->btc(gdA, 1); - e->btc(gzA, 1); - e->btc(gdA, gdB); - e->btc(gzA, gzB); - e->btc(intptr_gpA, 1); - e->btc(anyptr_gpA, gdB); - e->btc(intptr_gpA, gzB); - e->btr(gdA, 1); - e->btr(gzA, 1); - e->btr(gdA, gdB); - e->btr(gzA, gzB); - e->btr(intptr_gpA, 1); - e->btr(anyptr_gpA, gdB); - e->btr(intptr_gpA, gzB); - e->bts(gdA, 1); - e->bts(gzA, 1); - e->bts(gdA, gdB); - e->bts(gzA, gzB); - e->bts(intptr_gpA, 1); - e->bts(anyptr_gpA, gdB); - e->bts(intptr_gpA, gzB); - e->call(gzA); - e->call(intptr_gpA); - e->cbw(); // Implicit AX <- Sign Extend AL. - e->cbw(ax); // Explicit AX <- Sign Extend AL. - e->cdq(); // Implicit EDX:EAX <- Sign Extend EAX. - e->cdq(edx, eax); // Explicit EDX:EAX <- Sign Extend EAX. - if (isX64) e->cdqe(); // Implicit RAX <- Sign Extend EAX. - if (isX64) e->cdqe(eax); // Explicit RAX <- Sign Extend EAX. - e->cwd(); // Implicit DX:AX <- Sign Extend AX. - e->cwd(dx, ax); // Explicit DX:AX <- Sign Extend AX. - e->cwde(); // Implicit EAX <- Sign Extend AX. - e->cwde(eax); // Explicit EAX <- Sign Extend AX. - if (isX64) e->cqo(); // Implicit RDX:RAX <- Sign Extend RAX. - if (isX64) e->cqo(rdx, rax); // Explicit RDX:RAX <- Sign Extend RAX. - e->clc(); - e->cld(); - e->cmc(); - e->cmp(gLoA, 1); - e->cmp(gLoB, 1); - e->cmp(gHiA, 1); - e->cmp(gHiB, 1); - e->cmp(gwA, 1); - e->cmp(gwB, 1); - e->cmp(gdA, 1); - e->cmp(gdB, 1); - e->cmp(gzA, 1); - e->cmp(gLoA, gLoB); - e->cmp(gHiA, gHiB); - e->cmp(gwA, gwB); - e->cmp(gdA, gdB); - e->cmp(gzA, gzB); - e->cmp(gdA, anyptr_gpB); - e->cmp(gzA, intptr_gpB); - e->cmp(intptr_gpA, 1); - e->cmp(anyptr_gpA, gdB); - e->cmp(intptr_gpA, gzB); - e->cmpxchg(gdA, gdB); // Implicit regA, regB, <EAX> - e->cmpxchg(gzA, gzB); // Implicit regA, regB, <ZAX> - e->cmpxchg(gdA, gdB, eax); // Explicit regA, regB, <EAX> - e->cmpxchg(gzA, gzB, e->zax()); // Explicit regA, regB, <ZAX> - e->cmpxchg(anyptr_gpA, gdB); // Implicit mem , regB, <EAX> - e->cmpxchg(anyptr_gpA, gzB); // Implicit mem , regB, <ZAX> - e->cmpxchg(anyptr_gpA, gdB, eax); // Explicit mem , regB, <EAX> - e->cmpxchg(anyptr_gpA, gzB, e->zax()); // Explicit mem , regB, <ZAX> - e->cmpxchg8b(anyptr_gpA); // Implicit mem , <EDX>, <EAX>, <ECX>, <EBX> - e->cmpxchg8b(anyptr_gpA, - x86::edx, x86::eax, - x86::ecx, x86::ebx); // Explicit mem , <EDX>, <EAX>, <ECX>, <EBX> - if (isX64) e->cmpxchg16b(anyptr_gpA); // Implicit mem , <RDX>, <RAX>, <RCX>, <RBX> - if (isX64) e->cmpxchg16b(anyptr_gpA, - x86::rdx, x86::rax, - x86::rcx, x86::rbx); // Explicit mem , <EDX>, <EAX>, <ECX>, <EBX> - e->cpuid(); // Implicit <EAX>, <EBX>, <ECX>, <EDX> - e->cpuid(eax, ebx, ecx, edx); // Explicit <EAX>, <EBX>, <ECX>, <EDX> - e->crc32(gdA, byte_ptr(gzB)); - e->crc32(gdA, word_ptr(gzB)); - e->crc32(gdA, dword_ptr(gzB)); - if (isX64) e->crc32(gdA, qword_ptr(gzB)); - if (isX64) e->crc32(gzA, qword_ptr(gzB)); - e->dec(gLoA); - e->dec(gHiA); - e->dec(gwA); - e->dec(gdA); - e->dec(gzA); - e->dec(intptr_gpA); - e->inc(gLoA); - e->inc(gwA); - e->inc(gdA); - e->inc(gzA); - e->inc(intptr_gpA); - e->int_(13); - e->int3(); - e->into(); - e->lea(gzA, intptr_gpB); - e->mov(gLoA, 1); - e->mov(gHiA, 1); - e->mov(gwA, 1); - e->mov(gdA, 1); - e->mov(gzA, 1); - e->mov(gLoA, gLoB); - e->mov(gHiA, gHiB); - e->mov(gwA, gwB); - e->mov(gdA, gdB); - e->mov(gzA, gzB); - e->mov(gLoA, anyptr_gpB); - e->mov(gwA, anyptr_gpB); - e->mov(gdA, anyptr_gpB); - e->mov(gzA, intptr_gpB); - e->mov(anyptr_gpA, gLoB); - e->mov(anyptr_gpA, gwB); - e->mov(anyptr_gpA, gdB); - e->mov(intptr_gpA, 1); - e->mov(intptr_gpA, gzB); - e->movsx(gzA, gLoB); - e->movsx(gzA, byte_ptr(gzB)); - e->movzx(gzA, gLoB); - e->movzx(gzA, byte_ptr(gzB)); - e->movbe(gzA, anyptr_gpB); - e->movbe(anyptr_gpA, gzB); - e->neg(gzA); - e->neg(intptr_gpA); - e->nop(); - e->not_(gzA); - e->not_(intptr_gpA); - e->or_(gLoA, 1); - e->or_(gLoB, 1); - e->or_(gHiA, 1); - e->or_(gHiB, 1); - e->or_(gwA, 1); - e->or_(gwB, 1); - e->or_(gdA, 1); - e->or_(gdB, 1); - e->or_(gzA, 1); - e->or_(gzA, gzB); - e->or_(gzA, intptr_gpB); - e->or_(intptr_gpA, 1); - e->or_(intptr_gpA, gzB); - e->pop(gzA); - e->pop(intptr_gpA); - if (!isX64) e->popa(); - if (!isX64) e->popad(); - e->popf(); - if (!isX64) e->popfd(); - if ( isX64) e->popfq(); - e->push(gzA); - e->push(intptr_gpA); - e->push(0); - if (!isX64) e->pusha(); - if (!isX64) e->pushad(); - e->pushf(); - if (!isX64) e->pushfd(); - if ( isX64) e->pushfq(); - e->rcl(gdA, 0); - e->rcl(gzA, 0); - e->rcl(gdA, 1); - e->rcl(gzA, 1); - e->rcl(gdA, cl); - e->rcl(gzA, cl); - e->rcl(intptr_gpA, 0); - e->rcl(intptr_gpA, 1); - e->rcl(intptr_gpA, cl); - e->rcr(gdA, 0); - e->rcr(gzA, 0); - e->rcr(gdA, 1); - e->rcr(gzA, 1); - e->rcr(gdA, cl); - e->rcr(gzA, cl); - e->rcr(intptr_gpA, 0); - e->rcr(intptr_gpA, 1); - e->rcr(intptr_gpA, cl); - e->rdtsc(); // Implicit <EDX:EAX> - e->rdtsc(edx, eax); // Explicit <EDX:EAX> - e->rdtscp(); // Implicit <EDX:EAX>, <ECX> - e->rdtscp(edx, eax, ecx); // Implicit <EDX:EAX>, <ECX> - e->ret(); - e->ret(0); - e->rol(gdA, 0); - e->rol(gzA, 0); - e->rol(gdA, 1); - e->rol(gzA, 1); - e->rol(gdA, cl); - e->rol(gzA, cl); - e->rol(intptr_gpA, 0); - e->rol(intptr_gpA, 1); - e->rol(intptr_gpA, cl); - e->ror(gdA, 0); - e->ror(gzA, 0); - e->ror(gdA, 1); - e->ror(gzA, 1); - e->ror(gdA, cl); - e->ror(gzA, cl); - e->ror(intptr_gpA, 0); - e->ror(intptr_gpA, 1); - e->ror(intptr_gpA, cl); - e->sbb(gLoA, 1); - e->sbb(gLoB, 1); - e->sbb(gHiA, 1); - e->sbb(gHiB, 1); - e->sbb(gwA, 1); - e->sbb(gwB, 1); - e->sbb(gdA, 1); - e->sbb(gdB, 1); - e->sbb(gzA, 1); - e->sbb(gzA, gzB); - e->sbb(gzA, intptr_gpB); - e->sbb(intptr_gpA, 1); - e->sbb(intptr_gpA, gzB); - e->sal(gdA, 0); - e->sal(gzA, 0); - e->sal(gdA, 1); - e->sal(gzA, 1); - e->sal(gdA, cl); - e->sal(gzA, cl); - e->sal(intptr_gpA, 0); - e->sal(intptr_gpA, 1); - e->sal(intptr_gpA, cl); - e->sar(gdA, 0); - e->sar(gzA, 0); - e->sar(gdA, 1); - e->sar(gzA, 1); - e->sar(gdA, cl); - e->sar(gzA, cl); - e->sar(intptr_gpA, 0); - e->sar(intptr_gpA, 1); - e->sar(intptr_gpA, cl); - e->shl(gdA, 0); - e->shl(gzA, 0); - e->shl(gdA, 1); - e->shl(gzA, 1); - e->shl(gdA, cl); - e->shl(gzA, cl); - e->shl(intptr_gpA, 0); - e->shl(intptr_gpA, 1); - e->shl(intptr_gpA, cl); - e->shr(gdA, 0); - e->shr(gzA, 0); - e->shr(gdA, 1); - e->shr(gzA, 1); - e->shr(gdA, cl); - e->shr(gzA, cl); - e->shr(intptr_gpA, 0); - e->shr(intptr_gpA, 1); - e->shr(intptr_gpA, cl); - e->shld(gdA, gdB, 0); - e->shld(gzA, gzB, 0); - e->shld(gdA, gdB, cl); - e->shld(gzA, gzB, cl); - e->shld(anyptr_gpA, gdB, 0); - e->shld(intptr_gpA, gzB, 0); - e->shld(anyptr_gpA, gdB, cl); - e->shld(intptr_gpA, gzB, cl); - e->shrd(gdA, gdB, 0); - e->shrd(gzA, gzB, 0); - e->shrd(gdA, gdB, cl); - e->shrd(gzA, gzB, cl); - e->shrd(anyptr_gpA, gdB, 0); - e->shrd(intptr_gpA, gzB, 0); - e->shrd(anyptr_gpA, gdB, cl); - e->shrd(intptr_gpA, gzB, cl); - e->stc(); - e->std(); - e->sti(); - e->sub(gLoA, 1); - e->sub(gLoB, 1); - e->sub(gHiA, 1); - e->sub(gHiB, 1); - e->sub(gwA, 1); - e->sub(gwB, 1); - e->sub(gdA, 1); - e->sub(gdB, 1); - e->sub(gzA, 1); - e->sub(gzA, gzB); - e->sub(gzA, intptr_gpB); - e->sub(intptr_gpA, 1); - e->sub(intptr_gpA, gzB); - e->swapgs(); - e->test(gzA, 1); - e->test(gzA, gzB); - e->test(intptr_gpA, 1); - e->test(intptr_gpA, gzB); - e->ud2(); - e->xadd(gzA, gzB); - e->xadd(intptr_gpA, gzB); - e->xchg(gzA, gzB); - e->xchg(intptr_gpA, gzB); - e->xchg(gzA, intptr_gpB); - e->xor_(gLoA, 1); - e->xor_(gLoB, 1); - e->xor_(gHiA, 1); - e->xor_(gHiB, 1); - e->xor_(gwA, 1); - e->xor_(gwB, 1); - e->xor_(gdA, 1); - e->xor_(gdB, 1); - e->xor_(gzA, 1); - e->xor_(gzA, gzB); - e->xor_(gzA, intptr_gpB); - e->xor_(intptr_gpA, 1); - e->xor_(intptr_gpA, gzB); - - // Special case - div|mul. - e->div(cl); // Implicit AH:AL <- AX * r8 - e->div(byte_ptr(gzA)); // Implicit AH:AL <- AX * m8 - e->div(ax, cl); // Explicit AH:AL <- AX * r8 - e->div(ax, anyptr_gpA); // Explicit AH:AL <- AX * m8 - - e->div(cx); // Implicit DX:AX <- DX:AX * r16 - e->div(word_ptr(gzA)); // Implicit DX:AX <- DX:AX * m16 - e->div(dx, ax, cx); // Explicit DX:AX <- DX:AX * r16 - e->div(dx, ax, anyptr_gpA); // Explicit DX:AX <- DX:AX * m16 - - e->div(ecx); // Implicit EDX:EAX <- EDX:EAX * r32 - e->div(dword_ptr(gzA)); // Implicit EDX:EAX <- EDX:EAX * m32 - e->div(edx, eax, ecx); // Explicit EDX:EAX <- EDX:EAX * r32 - e->div(edx, eax, anyptr_gpA); // Explicit EDX:EAX <- EDX:EAX * m32 - - if (isX64) e->div(rcx); // Implicit RDX|RAX <- RDX:RAX * r64 - if (isX64) e->div(qword_ptr(gzA)); // Implicit RDX|RAX <- RDX:RAX * m64 - if (isX64) e->div(rdx, rax, rcx); // Explicit RDX|RAX <- RDX:RAX * r64 - if (isX64) e->div(rdx, rax, anyptr_gpA); // Explicit RDX|RAX <- RDX:RAX * m64 - - e->idiv(cl); // Implicit AH:AL <- AX * r8 - e->idiv(byte_ptr(gzA)); // Implicit AH:AL <- AX * m8 - e->idiv(ax, cl); // Explicit AH:AL <- AX * r8 - e->idiv(ax, anyptr_gpA); // Explicit AH:AL <- AX * m8 - - e->idiv(cx); // Implicit DX:AX <- DX:AX * r16 - e->idiv(word_ptr(gzA)); // Implicit DX:AX <- DX:AX * m16 - e->idiv(dx, ax, cx); // Explicit DX:AX <- DX:AX * r16 - e->idiv(dx, ax, anyptr_gpA); // Explicit DX:AX <- DX:AX * m16 - - e->idiv(ecx); // Implicit EDX:EAX <- EDX:EAX * r32 - e->idiv(dword_ptr(gzA)); // Implicit EDX:EAX <- EDX:EAX * m32 - e->idiv(edx, eax, ecx); // Explicit EDX:EAX <- EDX:EAX * r32 - e->idiv(edx, eax, anyptr_gpA); // Explicit EDX:EAX <- EDX:EAX * m32 - - if (isX64) e->idiv(rcx); // Implicit RDX|RAX <- RDX:RAX * r64 - if (isX64) e->idiv(qword_ptr(gzA)); // Implicit RDX|RAX <- RDX:RAX * m64 - if (isX64) e->idiv(rdx, rax, rcx); // Explicit RDX|RAX <- RDX:RAX * r64 - if (isX64) e->idiv(rdx, rax, anyptr_gpA); // Explicit RDX|RAX <- RDX:RAX * m64 - - e->mul(cl); // Implicit AX <- AL * r8 - e->mul(byte_ptr(gzA)); // Implicit AX <- AL * m8 - e->mul(ax, cl); // Explicit AX <- AL * r8 - e->mul(ax, anyptr_gpA); // Explicit AX <- AL * m8 - - e->mul(cx); // Implicit DX:AX <- AX * r16 - e->mul(word_ptr(gzA)); // Implicit DX:AX <- AX * m16 - e->mul(dx, ax, cx); // Explicit DX:AX <- AX * r16 - e->mul(dx, ax, anyptr_gpA); // Explicit DX:AX <- AX * m16 - - e->mul(ecx); // Implicit EDX:EAX <- EAX * r32 - e->mul(dword_ptr(gzA)); // Implicit EDX:EAX <- EAX * m32 - e->mul(edx, eax, ecx); // Explicit EDX:EAX <- EAX * r32 - e->mul(edx, eax, anyptr_gpA); // Explicit EDX:EAX <- EAX * m32 - - if (isX64) e->mul(rcx); // Implicit RDX|RAX <- RAX * r64 - if (isX64) e->mul(qword_ptr(gzA)); // Implicit RDX|RAX <- RAX * m64 - if (isX64) e->mul(rdx, rax, rcx); // Explicit RDX|RAX <- RAX * r64 - if (isX64) e->mul(rdx, rax, anyptr_gpA); // Explicit RDX|RAX <- RAX * m64 - - e->imul(gdA); - e->imul(gzA); - e->imul(intptr_gpA); - e->imul(gdA, 1); - e->imul(gzA, 1); - e->imul(gdA, gdB); - e->imul(gzA, gzB); - e->imul(gdA, gdB, 1); - e->imul(gzA, gzB, 1); - e->imul(gdA, anyptr_gpB); - e->imul(gzA, intptr_gpB); - e->imul(gdA, anyptr_gpB, 1); - e->imul(gzA, intptr_gpB, 1); - - // Special case - zero-extend 32-bit immediate instead of sign-extend: - if (isX64) e->mov(gzA, uint32_t(0xFEEDFEED)); - if (isX64) e->and_(gzA, uint32_t(0xFEEDFEED)); - - // Special case - mov with absolute 32-bit address. - e->mov(al , ptr(0x01020304u)); - e->mov(ax , ptr(0x01020304u)); - e->mov(eax, ptr(0x01020304u)); - e->mov(ptr(0x01020304u), al ); - e->mov(ptr(0x01020304u), ax ); - e->mov(ptr(0x01020304u), eax); - - // Special case - mov with absolute 64-bit address. - if (isX64) e->mov(al , ptr(0x0102030405060708u)); - if (isX64) e->mov(ax , ptr(0x0102030405060708u)); - if (isX64) e->mov(eax, ptr(0x0102030405060708u)); - if (isX64) e->mov(rax, ptr(0x0102030405060708u)); - if (isX64) e->mov(ptr(0x0102030405060708u), al ); - if (isX64) e->mov(ptr(0x0102030405060708u), ax ); - if (isX64) e->mov(ptr(0x0102030405060708u), eax); - if (isX64) e->mov(ptr(0x0102030405060708u), rax); - - // Control registers. - e->nop(); - - e->mov(gzA, cr0); - e->mov(cr0, gzA); - if (isX64) e->mov(gzA, cr8); - if (isX64) e->mov(cr8, gzA); - - // Debug registers. - e->nop(); - - e->mov(gzA, dr0); - e->mov(dr0, gzA); - - // Segment registers. - e->nop(); - - if (!isX64) e->mov(es, ax); - if (!isX64) e->mov(es, bx); - if (!isX64) e->mov(ax, es); - if (!isX64) e->mov(bx, es); - - if (!isX64) e->mov(cs, ax); - if (!isX64) e->mov(cs, bx); - if (!isX64) e->mov(ax, cs); - if (!isX64) e->mov(bx, cs); - - if (!isX64) e->mov(ss, ax); - if (!isX64) e->mov(ss, bx); - if (!isX64) e->mov(ax, ss); - if (!isX64) e->mov(bx, ss); - - if (!isX64) e->mov(ds, ax); - if (!isX64) e->mov(ds, bx); - if (!isX64) e->mov(ax, ds); - if (!isX64) e->mov(bx, ds); - - e->mov(fs, ax); - e->mov(fs, bx); - e->mov(ax, fs); - e->mov(bx, fs); - - e->mov(gs, ax); - e->mov(gs, bx); - e->mov(ax, gs); - e->mov(bx, gs); - - // Instructions using REP prefix. - e->nop(); - - e->in(al, 0); - e->in(al, dx); - e->in(ax, 0); - e->in(ax, dx); - e->in(eax, 0); - e->in(eax, dx); - e->rep().ins(byte_ptr(e->zdi()), dx); - e->rep().ins(word_ptr(e->zdi()), dx); - e->rep().ins(dword_ptr(e->zdi()), dx); - - e->out(imm(0), al); - e->out(dx, al); - e->out(imm(0), ax); - e->out(dx, ax); - e->out(imm(0), eax); - e->out(dx, eax); - e->rep().outs(dx, byte_ptr(e->zsi())); - e->rep().outs(dx, word_ptr(e->zsi())); - e->rep().outs(dx, dword_ptr(e->zsi())); - - e->lodsb(); - e->lodsd(); - e->lodsw(); - e->rep().lodsb(); - e->rep().lodsd(); - e->rep().lodsw(); - if (isX64) e->rep().lodsq(); - - e->movsb(); - e->movsd(); - e->movsw(); - e->rep().movsb(); - e->rep().movsd(); - e->rep().movsw(); - if (isX64) e->rep().movsq(); - - e->stosb(); - e->stosd(); - e->stosw(); - e->rep().stosb(); - e->rep().stosd(); - e->rep().stosw(); - if (isX64) e->rep().stosq(); - - e->cmpsb(); - e->cmpsd(); - e->cmpsw(); - e->repz().cmpsb(); - e->repz().cmpsd(); - e->repz().cmpsw(); - if (isX64) e->repz().cmpsq(); - e->repnz().cmpsb(); - e->repnz().cmpsd(); - e->repnz().cmpsw(); - if (isX64) e->repnz().cmpsq(); - - e->scasb(); - e->scasd(); - e->scasw(); - e->repz().scasb(); - e->repz().scasd(); - e->repz().scasw(); - if (isX64) e->repz().scasq(); - e->repnz().scasb(); - e->repnz().scasd(); - e->repnz().scasw(); - if (isX64) e->repnz().scasq(); - - // Label...Jcc/Jecxz/Jmp. - e->nop(); - - L = e->newLabel(); - e->bind(L); - e->ja(L); - e->jae(L); - e->jb(L); - e->jbe(L); - e->jc(L); - e->je(L); - e->jg(L); - e->jge(L); - e->jl(L); - e->jle(L); - e->jna(L); - e->jnae(L); - e->jnb(L); - e->jnbe(L); - e->jnc(L); - e->jne(L); - e->jng(L); - e->jnge(L); - e->jnl(L); - e->jnle(L); - e->jno(L); - e->jnp(L); - e->jns(L); - e->jnz(L); - e->jo(L); - e->jp(L); - e->jpe(L); - e->jpo(L); - e->js(L); - e->jz(L); - e->jecxz(ecx, L); - e->jmp(L); - - // Jcc/Jecxz/Jmp...Label. - e->nop(); - - L = e->newLabel(); - e->ja(L); - e->jae(L); - e->jb(L); - e->jbe(L); - e->jc(L); - e->je(L); - e->jg(L); - e->jge(L); - e->jl(L); - e->jle(L); - e->jna(L); - e->jnae(L); - e->jnb(L); - e->jnbe(L); - e->jnc(L); - e->jne(L); - e->jng(L); - e->jnge(L); - e->jnl(L); - e->jnle(L); - e->jno(L); - e->jnp(L); - e->jns(L); - e->jnz(L); - e->jo(L); - e->jp(L); - e->jpe(L); - e->jpo(L); - e->js(L); - e->jz(L); - e->jecxz(ecx, L); - e->jmp(L); - e->bind(L); - - // FPU. - e->nop(); - - e->f2xm1(); - e->fabs(); - e->fadd(stA, stB); - e->fadd(stB, stA); - e->fadd(dword_ptr(gzA)); - e->fadd(qword_ptr(gzA)); - e->faddp(stB); - e->faddp(); - e->fbld(dword_ptr(gzA)); - e->fbstp(dword_ptr(gzA)); - e->fchs(); - e->fclex(); - e->fcom(stB); - e->fcom(); - e->fcom(dword_ptr(gzA)); - e->fcom(qword_ptr(gzA)); - e->fcomp(stB); - e->fcomp(); - e->fcomp(dword_ptr(gzA)); - e->fcomp(qword_ptr(gzA)); - e->fcompp(); - e->fcos(); - e->fdecstp(); - e->fdiv(stA, stB); - e->fdiv(stB, stA); - e->fdiv(dword_ptr(gzA)); - e->fdiv(qword_ptr(gzA)); - e->fdivp(stB); - e->fdivp(); - e->fdivr(stA, stB); - e->fdivr(stB, stA); - e->fdivr(dword_ptr(gzA)); - e->fdivr(qword_ptr(gzA)); - e->fdivrp(stB); - e->fdivrp(); - e->fiadd(dword_ptr(gzA)); - e->ficom(word_ptr(gzA)); - e->ficom(dword_ptr(gzA)); - e->ficomp(word_ptr(gzA)); - e->ficomp(dword_ptr(gzA)); - e->fidiv(word_ptr(gzA)); - e->fidiv(dword_ptr(gzA)); - e->fidivr(word_ptr(gzA)); - e->fidivr(dword_ptr(gzA)); - e->fild(word_ptr(gzA)); - e->fild(dword_ptr(gzA)); - e->fild(qword_ptr(gzA)); - e->fimul(word_ptr(gzA)); - e->fimul(dword_ptr(gzA)); - e->fincstp(); - e->finit(); - e->fninit(); - e->fisub(word_ptr(gzA)); - e->fisub(dword_ptr(gzA)); - e->fisubr(word_ptr(gzA)); - e->fisubr(dword_ptr(gzA)); - e->fist(word_ptr(gzA)); - e->fist(dword_ptr(gzA)); - e->fistp(word_ptr(gzA)); - e->fistp(dword_ptr(gzA)); - e->fistp(qword_ptr(gzA)); - e->fld(dword_ptr(gzA)); - e->fld(qword_ptr(gzA)); - e->fld(tword_ptr(gzA)); - e->fld1(); - e->fldl2t(); - e->fldl2e(); - e->fldpi(); - e->fldlg2(); - e->fldln2(); - e->fldz(); - e->fldcw(anyptr_gpA); - e->fldenv(anyptr_gpA); - e->fmul(stA, stB); - e->fmul(stB, stA); - e->fmul(dword_ptr(gzA)); - e->fmul(qword_ptr(gzA)); - e->fmulp(stB); - e->fmulp(); - e->fnclex(); - e->fnop(); - e->fnsave(anyptr_gpA); - e->fnstenv(anyptr_gpA); - e->fnstcw(anyptr_gpA); - e->fpatan(); - e->fprem(); - e->fprem1(); - e->fptan(); - e->frndint(); - e->frstor(anyptr_gpA); - e->fsave(anyptr_gpA); - e->fscale(); - e->fsin(); - e->fsincos(); - e->fsqrt(); - e->fst(dword_ptr(gzA)); - e->fst(qword_ptr(gzA)); - e->fstp(dword_ptr(gzA)); - e->fstp(qword_ptr(gzA)); - e->fstp(tword_ptr(gzA)); - e->fstcw(anyptr_gpA); - e->fstenv(anyptr_gpA); - e->fsub(stA, stB); - e->fsub(stB, stA); - e->fsub(dword_ptr(gzA)); - e->fsub(qword_ptr(gzA)); - e->fsubp(stB); - e->fsubp(); - e->fsubr(stA, stB); - e->fsubr(stB, stA); - e->fsubr(dword_ptr(gzA)); - e->fsubr(qword_ptr(gzA)); - e->fsubrp(stB); - e->fsubrp(); - e->ftst(); - e->fucom(stB); - e->fucom(); - e->fucom(stB); - e->fucomi(stB); - e->fucomip(stB); - e->fucomp(stB); - e->fucompp(); - e->fxam(); - e->fxtract(); - e->fyl2x(); - e->fyl2xp1(); - - // LAHF/SAHF - e->lahf(); // Implicit <AH> - e->lahf(ah); // Explicit <AH> - e->sahf(); // Implicit <AH> - e->sahf(ah); // Explicit <AH> - - // FXSR. - e->fxrstor(anyptr_gpA); - e->fxsave(anyptr_gpA); - - // XSAVE. - e->nop(); - - e->xgetbv(); // Implicit <EDX:EAX>, <ECX> - e->xgetbv(edx, eax, ecx); // Explicit <EDX:EAX>, <ECX> - - e->xsetbv(); // Implicit <EDX:EAX>, <ECX> - e->xsetbv(edx, eax, ecx); // Explicit <EDX:EAX>, <ECX> - - e->xrstor(anyptr_gpA); // Implicit <EDX:EAX> - e->xrstors(anyptr_gpA); // Implicit <EDX:EAX> - e->xsave(anyptr_gpA); // Implicit <EDX:EAX> - e->xsavec(anyptr_gpA); // Implicit <EDX:EAX> - e->xsaveopt(anyptr_gpA); // Implicit <EDX:EAX> - e->xsaves(anyptr_gpA); // Implicit <EDX:EAX> - - if (isX64) e->xrstor64(anyptr_gpA); // Implicit <EDX:EAX> - if (isX64) e->xrstors64(anyptr_gpA); // Implicit <EDX:EAX> - if (isX64) e->xsave64(anyptr_gpA); // Implicit <EDX:EAX> - if (isX64) e->xsavec64(anyptr_gpA); // Implicit <EDX:EAX> - if (isX64) e->xsaveopt64(anyptr_gpA); // Implicit <EDX:EAX> - if (isX64) e->xsaves64(anyptr_gpA); // Implicit <EDX:EAX> - - // POPCNT. - e->nop(); - - e->popcnt(gdA, gdB); - e->popcnt(gzA, gzB); - e->popcnt(gdA, anyptr_gpB); - e->popcnt(gzA, anyptr_gpB); - - // LZCNT. - e->nop(); - - e->lzcnt(gdA, gdB); - e->lzcnt(gzA, gzB); - e->lzcnt(gdA, anyptr_gpB); - e->lzcnt(gzA, anyptr_gpB); - - // BMI. - e->nop(); - - e->andn(gdA, gdB, gdC); - e->andn(gzA, gzB, gzC); - e->andn(gdA, gdB, anyptr_gpC); - e->andn(gzA, gzB, anyptr_gpC); - e->bextr(gdA, gdB, gdC); - e->bextr(gzA, gzB, gzC); - e->bextr(gdA, anyptr_gpB, gdC); - e->bextr(gzA, anyptr_gpB, gzC); - e->blsi(gdA, gdB); - e->blsi(gzA, gzB); - e->blsi(gdA, anyptr_gpB); - e->blsi(gzA, anyptr_gpB); - e->blsmsk(gdA, gdB); - e->blsmsk(gzA, gzB); - e->blsmsk(gdA, anyptr_gpB); - e->blsmsk(gzA, anyptr_gpB); - e->blsr(gdA, gdB); - e->blsr(gzA, gzB); - e->blsr(gdA, anyptr_gpB); - e->blsr(gzA, anyptr_gpB); - e->tzcnt(gdA, gdB); - e->tzcnt(gzA, gzB); - e->tzcnt(gdA, anyptr_gpB); - e->tzcnt(gzA, anyptr_gpB); - - // BMI2. - e->nop(); - - e->bzhi(gdA, gdB, gdC); - e->bzhi(gzA, gzB, gzC); - e->bzhi(gdA, anyptr_gpB, gdC); - e->bzhi(gzA, anyptr_gpB, gzC); - e->mulx(gdA, gdB, gdC); // Implicit gpA, gpB, gpC, <EDX> - e->mulx(gdA, gdB, gdC, edx); // Explicit gpA, gpB, gpC, <EDX> - e->mulx(gzA, gzB, gzC); // Implicit gpA, gpB, gpC, <EDX|RDX> - e->mulx(gzA, gzB, gzC, e->zdx()); // Explicit gpA, gpB, gpC, <EDX|RDX> - e->mulx(gdA, gdB, anyptr_gpC); // Implicit gpA, gpB, mem, <EDX> - e->mulx(gdA, gdB, anyptr_gpC, edx); // Explicit gpA, gpB, mem, <EDX> - e->mulx(gzA, gzB, anyptr_gpC); // Implicit gpA, gpB, mem, <EDX|RDX> - e->mulx(gzA, gzB, anyptr_gpC, e->zdx()); // Explicit gpA, gpB, mem, <EDX|RDX> - e->pdep(gdA, gdB, gdC); - e->pdep(gzA, gzB, gzC); - e->pdep(gdA, gdB, anyptr_gpC); - e->pdep(gzA, gzB, anyptr_gpC); - e->pext(gdA, gdB, gdC); - e->pext(gzA, gzB, gzC); - e->pext(gdA, gdB, anyptr_gpC); - e->pext(gzA, gzB, anyptr_gpC); - e->rorx(gdA, gdB, 0); - e->rorx(gzA, gzB, 0); - e->rorx(gdA, anyptr_gpB, 0); - e->rorx(gzA, anyptr_gpB, 0); - e->sarx(gdA, gdB, gdC); - e->sarx(gzA, gzB, gzC); - e->sarx(gdA, anyptr_gpB, gdC); - e->sarx(gzA, anyptr_gpB, gzC); - e->shlx(gdA, gdB, gdC); - e->shlx(gzA, gzB, gzC); - e->shlx(gdA, anyptr_gpB, gdC); - e->shlx(gzA, anyptr_gpB, gzC); - e->shrx(gdA, gdB, gdC); - e->shrx(gzA, gzB, gzC); - e->shrx(gdA, anyptr_gpB, gdC); - e->shrx(gzA, anyptr_gpB, gzC); - - // ADX. - e->nop(); - - e->adcx(gdA, gdB); - e->adcx(gzA, gzB); - e->adcx(gdA, anyptr_gpB); - e->adcx(gzA, anyptr_gpB); - e->adox(gdA, gdB); - e->adox(gzA, gzB); - e->adox(gdA, anyptr_gpB); - e->adox(gzA, anyptr_gpB); - - // TBM. - e->nop(); - - e->blcfill(gdA, gdB); - e->blcfill(gzA, gzB); - e->blcfill(gdA, anyptr_gpB); - e->blcfill(gzA, anyptr_gpB); - - e->blci(gdA, gdB); - e->blci(gzA, gzB); - e->blci(gdA, anyptr_gpB); - e->blci(gzA, anyptr_gpB); - - e->blcic(gdA, gdB); - e->blcic(gzA, gzB); - e->blcic(gdA, anyptr_gpB); - e->blcic(gzA, anyptr_gpB); - - e->blcmsk(gdA, gdB); - e->blcmsk(gzA, gzB); - e->blcmsk(gdA, anyptr_gpB); - e->blcmsk(gzA, anyptr_gpB); - - e->blcs(gdA, gdB); - e->blcs(gzA, gzB); - e->blcs(gdA, anyptr_gpB); - e->blcs(gzA, anyptr_gpB); - - e->blsfill(gdA, gdB); - e->blsfill(gzA, gzB); - e->blsfill(gdA, anyptr_gpB); - e->blsfill(gzA, anyptr_gpB); - - e->blsic(gdA, gdB); - e->blsic(gzA, gzB); - e->blsic(gdA, anyptr_gpB); - e->blsic(gzA, anyptr_gpB); - - e->t1mskc(gdA, gdB); - e->t1mskc(gzA, gzB); - e->t1mskc(gdA, anyptr_gpB); - e->t1mskc(gzA, anyptr_gpB); - - e->tzmsk(gdA, gdB); - e->tzmsk(gzA, gzB); - e->tzmsk(gdA, anyptr_gpB); - e->tzmsk(gzA, anyptr_gpB); - - // CLFLUSH / CLFLUSH_OPT. - e->nop(); - e->clflush(anyptr_gpA); - e->clflushopt(anyptr_gpA); - - // CLWB. - e->nop(); - e->clwb(anyptr_gpA); - - // CLZERO. - e->nop(); - e->clzero(); // Implicit <ds:[EAX|RAX]> - e->clzero(ptr(e->zax())); // Explicit <ds:[EAX|RAX]> - - // MONITOR[X] / MWAIT[X]. - e->nop(); - e->monitor(); // Implicit <ds:[EAX|RAX]>, <ECX>, <EDX> - e->monitorx(); // Implicit <ds:[EAX|RAX]>, <ECX>, <EDX> - e->mwait(); // Implicit <EAX>, <ECX> - e->mwaitx(); // Implicit <EAX>, <ECX>, <EBX> - - // PREFETCH / PREFETCHW / PREFETCHWT1. - e->nop(); - e->prefetch(anyptr_gpA); // 3DNOW. - e->prefetchnta(anyptr_gpA); // MMX+SSE. - e->prefetcht0(anyptr_gpA); // MMX+SSE. - e->prefetcht1(anyptr_gpA); // MMX+SSE. - e->prefetcht2(anyptr_gpA); // MMX+SSE. - e->prefetchw(anyptr_gpA); // PREFETCHW. - e->prefetchwt1(anyptr_gpA); // PREFETCHWT1. - - // RDRAND / RDSEED. - e->nop(); - - e->rdrand(gdA); - e->rdrand(gzA); - e->rdseed(gdA); - e->rdseed(gzA); - - // MMX/MMX2. - e->nop(); - - e->movd(anyptr_gpA, mmB); - e->movd(gdA, mmB); - e->movd(mmA, anyptr_gpB); - e->movd(mmA, gdB); - e->movq(mmA, mmB); - e->movq(anyptr_gpA, mmB); - e->movq(mmA, anyptr_gpB); - e->packuswb(mmA, mmB); - e->packuswb(mmA, anyptr_gpB); - e->paddb(mmA, mmB); - e->paddb(mmA, anyptr_gpB); - e->paddw(mmA, mmB); - e->paddw(mmA, anyptr_gpB); - e->paddd(mmA, mmB); - e->paddd(mmA, anyptr_gpB); - e->paddsb(mmA, mmB); - e->paddsb(mmA, anyptr_gpB); - e->paddsw(mmA, mmB); - e->paddsw(mmA, anyptr_gpB); - e->paddusb(mmA, mmB); - e->paddusb(mmA, anyptr_gpB); - e->paddusw(mmA, mmB); - e->paddusw(mmA, anyptr_gpB); - e->pand(mmA, mmB); - e->pand(mmA, anyptr_gpB); - e->pandn(mmA, mmB); - e->pandn(mmA, anyptr_gpB); - e->pcmpeqb(mmA, mmB); - e->pcmpeqb(mmA, anyptr_gpB); - e->pcmpeqw(mmA, mmB); - e->pcmpeqw(mmA, anyptr_gpB); - e->pcmpeqd(mmA, mmB); - e->pcmpeqd(mmA, anyptr_gpB); - e->pcmpgtb(mmA, mmB); - e->pcmpgtb(mmA, anyptr_gpB); - e->pcmpgtw(mmA, mmB); - e->pcmpgtw(mmA, anyptr_gpB); - e->pcmpgtd(mmA, mmB); - e->pcmpgtd(mmA, anyptr_gpB); - e->pmulhw(mmA, mmB); - e->pmulhw(mmA, anyptr_gpB); - e->pmullw(mmA, mmB); - e->pmullw(mmA, anyptr_gpB); - e->por(mmA, mmB); - e->por(mmA, anyptr_gpB); - e->pmaddwd(mmA, mmB); - e->pmaddwd(mmA, anyptr_gpB); - e->pslld(mmA, mmB); - e->pslld(mmA, anyptr_gpB); - e->pslld(mmA, 0); - e->psllq(mmA, mmB); - e->psllq(mmA, anyptr_gpB); - e->psllq(mmA, 0); - e->psllw(mmA, mmB); - e->psllw(mmA, anyptr_gpB); - e->psllw(mmA, 0); - e->psrad(mmA, mmB); - e->psrad(mmA, anyptr_gpB); - e->psrad(mmA, 0); - e->psraw(mmA, mmB); - e->psraw(mmA, anyptr_gpB); - e->psraw(mmA, 0); - e->psrld(mmA, mmB); - e->psrld(mmA, anyptr_gpB); - e->psrld(mmA, 0); - e->psrlq(mmA, mmB); - e->psrlq(mmA, anyptr_gpB); - e->psrlq(mmA, 0); - e->psrlw(mmA, mmB); - e->psrlw(mmA, anyptr_gpB); - e->psrlw(mmA, 0); - e->psubb(mmA, mmB); - e->psubb(mmA, anyptr_gpB); - e->psubw(mmA, mmB); - e->psubw(mmA, anyptr_gpB); - e->psubd(mmA, mmB); - e->psubd(mmA, anyptr_gpB); - e->psubsb(mmA, mmB); - e->psubsb(mmA, anyptr_gpB); - e->psubsw(mmA, mmB); - e->psubsw(mmA, anyptr_gpB); - e->psubusb(mmA, mmB); - e->psubusb(mmA, anyptr_gpB); - e->psubusw(mmA, mmB); - e->psubusw(mmA, anyptr_gpB); - e->punpckhbw(mmA, mmB); - e->punpckhbw(mmA, anyptr_gpB); - e->punpckhwd(mmA, mmB); - e->punpckhwd(mmA, anyptr_gpB); - e->punpckhdq(mmA, mmB); - e->punpckhdq(mmA, anyptr_gpB); - e->punpcklbw(mmA, mmB); - e->punpcklbw(mmA, anyptr_gpB); - e->punpcklwd(mmA, mmB); - e->punpcklwd(mmA, anyptr_gpB); - e->punpckldq(mmA, mmB); - e->punpckldq(mmA, anyptr_gpB); - e->pxor(mmA, mmB); - e->pxor(mmA, anyptr_gpB); - e->emms(); - - // 3DNOW. - e->nop(); - - e->pavgusb(mmA, mmB); - e->pavgusb(mmA, anyptr_gpB); - e->pf2id(mmA, mmB); - e->pf2id(mmA, anyptr_gpB); - e->pf2iw(mmA, mmB); - e->pf2iw(mmA, anyptr_gpB); - e->pfacc(mmA, mmB); - e->pfacc(mmA, anyptr_gpB); - e->pfadd(mmA, mmB); - e->pfadd(mmA, anyptr_gpB); - e->pfcmpeq(mmA, mmB); - e->pfcmpeq(mmA, anyptr_gpB); - e->pfcmpge(mmA, mmB); - e->pfcmpge(mmA, anyptr_gpB); - e->pfcmpgt(mmA, mmB); - e->pfcmpgt(mmA, anyptr_gpB); - e->pfmax(mmA, mmB); - e->pfmax(mmA, anyptr_gpB); - e->pfmin(mmA, mmB); - e->pfmin(mmA, anyptr_gpB); - e->pfmul(mmA, mmB); - e->pfmul(mmA, anyptr_gpB); - e->pfnacc(mmA, mmB); - e->pfnacc(mmA, anyptr_gpB); - e->pfpnacc(mmA, mmB); - e->pfpnacc(mmA, anyptr_gpB); - e->pfrcp(mmA, mmB); - e->pfrcp(mmA, anyptr_gpB); - e->pfrcpit1(mmA, mmB); - e->pfrcpit1(mmA, anyptr_gpB); - e->pfrcpit2(mmA, mmB); - e->pfrcpit2(mmA, anyptr_gpB); - e->pfrcpv(mmA, mmB); - e->pfrcpv(mmA, anyptr_gpB); - e->pfrsqit1(mmA, mmB); - e->pfrsqit1(mmA, anyptr_gpB); - e->pfrsqrt(mmA, mmB); - e->pfrsqrt(mmA, anyptr_gpB); - e->pfrsqrtv(mmA, mmB); - e->pfrsqrtv(mmA, anyptr_gpB); - e->pfsub(mmA, mmB); - e->pfsub(mmA, anyptr_gpB); - e->pfsubr(mmA, mmB); - e->pfsubr(mmA, anyptr_gpB); - e->pi2fd(mmA, mmB); - e->pi2fd(mmA, anyptr_gpB); - e->pi2fw(mmA, mmB); - e->pi2fw(mmA, anyptr_gpB); - e->pmulhrw(mmA, mmB); - e->pmulhrw(mmA, anyptr_gpB); - e->pswapd(mmA, mmB); - e->pswapd(mmA, anyptr_gpB); - e->femms(); - - // SSE. - e->nop(); - - e->addps(xmmA, xmmB); - e->addps(xmmA, anyptr_gpB); - e->addss(xmmA, xmmB); - e->addss(xmmA, anyptr_gpB); - e->andnps(xmmA, xmmB); - e->andnps(xmmA, anyptr_gpB); - e->andps(xmmA, xmmB); - e->andps(xmmA, anyptr_gpB); - e->cmpps(xmmA, xmmB, 0); - e->cmpps(xmmA, anyptr_gpB, 0); - e->cmpss(xmmA, xmmB, 0); - e->cmpss(xmmA, anyptr_gpB, 0); - e->comiss(xmmA, xmmB); - e->comiss(xmmA, anyptr_gpB); - e->cvtpi2ps(xmmA, mmB); - e->cvtpi2ps(xmmA, anyptr_gpB); - e->cvtps2pi(mmA, xmmB); - e->cvtps2pi(mmA, anyptr_gpB); - e->cvtsi2ss(xmmA, gdB); - e->cvtsi2ss(xmmA, gzB); - e->cvtsi2ss(xmmA, anyptr_gpB); - e->cvtss2si(gdA, xmmB); - e->cvtss2si(gzA, xmmB); - e->cvtss2si(gdA, anyptr_gpB); - e->cvtss2si(gzA, anyptr_gpB); - e->cvttps2pi(mmA, xmmB); - e->cvttps2pi(mmA, anyptr_gpB); - e->cvttss2si(gdA, xmmB); - e->cvttss2si(gzA, xmmB); - e->cvttss2si(gdA, anyptr_gpB); - e->cvttss2si(gzA, anyptr_gpB); - e->divps(xmmA, xmmB); - e->divps(xmmA, anyptr_gpB); - e->divss(xmmA, xmmB); - e->divss(xmmA, anyptr_gpB); - e->ldmxcsr(anyptr_gpA); - e->maskmovq(mmA, mmB); // Implicit mmA, mmB, <ds:[EDI|RDI]> - e->maskmovq(mmA, mmB, ptr(e->zdi())); // Explicit mmA, mmB, <ds:[EDI|RDI]> - e->maxps(xmmA, xmmB); - e->maxps(xmmA, anyptr_gpB); - e->maxss(xmmA, xmmB); - e->maxss(xmmA, anyptr_gpB); - e->minps(xmmA, xmmB); - e->minps(xmmA, anyptr_gpB); - e->minss(xmmA, xmmB); - e->minss(xmmA, anyptr_gpB); - e->movaps(xmmA, xmmB); - e->movaps(xmmA, anyptr_gpB); - e->movaps(anyptr_gpA, xmmB); - e->movd(anyptr_gpA, xmmB); - e->movd(gdA, xmmB); - e->movd(gzA, xmmB); - e->movd(xmmA, anyptr_gpB); - e->movd(xmmA, gdB); - e->movd(xmmA, gzB); - e->movq(mmA, mmB); - e->movq(xmmA, xmmB); - e->movq(anyptr_gpA, xmmB); - e->movq(xmmA, anyptr_gpB); - e->movntq(anyptr_gpA, mmB); - e->movhlps(xmmA, xmmB); - e->movhps(xmmA, anyptr_gpB); - e->movhps(anyptr_gpA, xmmB); - e->movlhps(xmmA, xmmB); - e->movlps(xmmA, anyptr_gpB); - e->movlps(anyptr_gpA, xmmB); - e->movntps(anyptr_gpA, xmmB); - e->movss(xmmA, anyptr_gpB); - e->movss(anyptr_gpA, xmmB); - e->movups(xmmA, xmmB); - e->movups(xmmA, anyptr_gpB); - e->movups(anyptr_gpA, xmmB); - e->mulps(xmmA, xmmB); - e->mulps(xmmA, anyptr_gpB); - e->mulss(xmmA, xmmB); - e->mulss(xmmA, anyptr_gpB); - e->orps(xmmA, xmmB); - e->orps(xmmA, anyptr_gpB); - e->pavgb(mmA, mmB); - e->pavgb(mmA, anyptr_gpB); - e->pavgw(mmA, mmB); - e->pavgw(mmA, anyptr_gpB); - e->pextrw(gdA, mmB, 0); - e->pextrw(gzA, mmB, 0); - e->pinsrw(mmA, gdB, 0); - e->pinsrw(mmA, gzB, 0); - e->pinsrw(mmA, anyptr_gpB, 0); - e->pmaxsw(mmA, mmB); - e->pmaxsw(mmA, anyptr_gpB); - e->pmaxub(mmA, mmB); - e->pmaxub(mmA, anyptr_gpB); - e->pminsw(mmA, mmB); - e->pminsw(mmA, anyptr_gpB); - e->pminub(mmA, mmB); - e->pminub(mmA, anyptr_gpB); - e->pmovmskb(gdA, mmB); - e->pmovmskb(gzA, mmB); - e->pmulhuw(mmA, mmB); - e->pmulhuw(mmA, anyptr_gpB); - e->psadbw(mmA, mmB); - e->psadbw(mmA, anyptr_gpB); - e->pshufw(mmA, mmB, 0); - e->pshufw(mmA, anyptr_gpB, 0); - e->rcpps(xmmA, xmmB); - e->rcpps(xmmA, anyptr_gpB); - e->rcpss(xmmA, xmmB); - e->rcpss(xmmA, anyptr_gpB); - e->psadbw(xmmA, xmmB); - e->psadbw(xmmA, anyptr_gpB); - e->rsqrtps(xmmA, xmmB); - e->rsqrtps(xmmA, anyptr_gpB); - e->rsqrtss(xmmA, xmmB); - e->rsqrtss(xmmA, anyptr_gpB); - e->sfence(); - e->shufps(xmmA, xmmB, 0); - e->shufps(xmmA, anyptr_gpB, 0); - e->sqrtps(xmmA, xmmB); - e->sqrtps(xmmA, anyptr_gpB); - e->sqrtss(xmmA, xmmB); - e->sqrtss(xmmA, anyptr_gpB); - e->stmxcsr(anyptr_gpA); - e->subps(xmmA, xmmB); - e->subps(xmmA, anyptr_gpB); - e->subss(xmmA, xmmB); - e->subss(xmmA, anyptr_gpB); - e->ucomiss(xmmA, xmmB); - e->ucomiss(xmmA, anyptr_gpB); - e->unpckhps(xmmA, xmmB); - e->unpckhps(xmmA, anyptr_gpB); - e->unpcklps(xmmA, xmmB); - e->unpcklps(xmmA, anyptr_gpB); - e->xorps(xmmA, xmmB); - e->xorps(xmmA, anyptr_gpB); - - // SSE2. - e->nop(); - - e->addpd(xmmA, xmmB); - e->addpd(xmmA, anyptr_gpB); - e->addsd(xmmA, xmmB); - e->addsd(xmmA, anyptr_gpB); - e->andnpd(xmmA, xmmB); - e->andnpd(xmmA, anyptr_gpB); - e->andpd(xmmA, xmmB); - e->andpd(xmmA, anyptr_gpB); - e->cmppd(xmmA, xmmB, 0); - e->cmppd(xmmA, anyptr_gpB, 0); - e->cmpsd(xmmA, xmmB, 0); - e->cmpsd(xmmA, anyptr_gpB, 0); - e->comisd(xmmA, xmmB); - e->comisd(xmmA, anyptr_gpB); - e->cvtdq2pd(xmmA, xmmB); - e->cvtdq2pd(xmmA, anyptr_gpB); - e->cvtdq2ps(xmmA, xmmB); - e->cvtdq2ps(xmmA, anyptr_gpB); - e->cvtpd2dq(xmmA, xmmB); - e->cvtpd2dq(xmmA, anyptr_gpB); - e->cvtpd2pi(mmA, xmmB); - e->cvtpd2pi(mmA, anyptr_gpB); - e->cvtpd2ps(xmmA, xmmB); - e->cvtpd2ps(xmmA, anyptr_gpB); - e->cvtpi2pd(xmmA, mmB); - e->cvtpi2pd(xmmA, anyptr_gpB); - e->cvtps2dq(xmmA, xmmB); - e->cvtps2dq(xmmA, anyptr_gpB); - e->cvtps2pd(xmmA, xmmB); - e->cvtps2pd(xmmA, anyptr_gpB); - e->cvtsd2si(gdA, xmmB); - e->cvtsd2si(gzA, xmmB); - e->cvtsd2si(gdA, anyptr_gpB); - e->cvtsd2si(gzA, anyptr_gpB); - e->cvtsd2ss(xmmA, xmmB); - e->cvtsd2ss(xmmA, anyptr_gpB); - e->cvtsi2sd(xmmA, gdB); - e->cvtsi2sd(xmmA, gzB); - e->cvtsi2sd(xmmA, anyptr_gpB); - e->cvtss2sd(xmmA, xmmB); - e->cvtss2sd(xmmA, anyptr_gpB); - e->cvtss2si(gdA, xmmB); - e->cvtss2si(gzA, xmmB); - e->cvtss2si(gdA, anyptr_gpB); - e->cvtss2si(gzA, anyptr_gpB); - e->cvttpd2pi(mmA, xmmB); - e->cvttpd2pi(mmA, anyptr_gpB); - e->cvttpd2dq(xmmA, xmmB); - e->cvttpd2dq(xmmA, anyptr_gpB); - e->cvttps2dq(xmmA, xmmB); - e->cvttps2dq(xmmA, anyptr_gpB); - e->cvttsd2si(gdA, xmmB); - e->cvttsd2si(gzA, xmmB); - e->cvttsd2si(gdA, anyptr_gpB); - e->cvttsd2si(gzA, anyptr_gpB); - e->divpd(xmmA, xmmB); - e->divpd(xmmA, anyptr_gpB); - e->divsd(xmmA, xmmB); - e->divsd(xmmA, anyptr_gpB); - e->lfence(); - e->maskmovdqu(xmmA, xmmB); // Implicit xmmA, xmmB, <ds:[EDI|RDI]> - e->maskmovdqu(xmmA, xmmB, ptr(e->zdi())); // Explicit xmmA, xmmB, <ds:[EDI|RDI]> - e->maxpd(xmmA, xmmB); - e->maxpd(xmmA, anyptr_gpB); - e->maxsd(xmmA, xmmB); - e->maxsd(xmmA, anyptr_gpB); - e->mfence(); - e->minpd(xmmA, xmmB); - e->minpd(xmmA, anyptr_gpB); - e->minsd(xmmA, xmmB); - e->minsd(xmmA, anyptr_gpB); - e->movdqa(xmmA, xmmB); - e->movdqa(xmmA, anyptr_gpB); - e->movdqa(anyptr_gpA, xmmB); - e->movdqu(xmmA, xmmB); - e->movdqu(xmmA, anyptr_gpB); - e->movdqu(anyptr_gpA, xmmB); - e->movmskps(gdA, xmmB); - e->movmskps(gzA, xmmB); - e->movmskpd(gdA, xmmB); - e->movmskpd(gzA, xmmB); - e->movsd(xmmA, xmmB); - e->movsd(xmmA, anyptr_gpB); - e->movsd(anyptr_gpA, xmmB); - e->movapd(xmmA, anyptr_gpB); - e->movapd(anyptr_gpA, xmmB); - e->movdq2q(mmA, xmmB); - e->movq2dq(xmmA, mmB); - e->movhpd(xmmA, anyptr_gpB); - e->movhpd(anyptr_gpA, xmmB); - e->movlpd(xmmA, anyptr_gpB); - e->movlpd(anyptr_gpA, xmmB); - e->movntdq(anyptr_gpA, xmmB); - e->movnti(anyptr_gpA, gdB); - e->movnti(anyptr_gpA, gzB); - e->movntpd(anyptr_gpA, xmmB); - e->movupd(xmmA, anyptr_gpB); - e->movupd(anyptr_gpA, xmmB); - e->mulpd(xmmA, xmmB); - e->mulpd(xmmA, anyptr_gpB); - e->mulsd(xmmA, xmmB); - e->mulsd(xmmA, anyptr_gpB); - e->orpd(xmmA, xmmB); - e->orpd(xmmA, anyptr_gpB); - e->packsswb(xmmA, xmmB); - e->packsswb(xmmA, anyptr_gpB); - e->packssdw(xmmA, xmmB); - e->packssdw(xmmA, anyptr_gpB); - e->packuswb(xmmA, xmmB); - e->packuswb(xmmA, anyptr_gpB); - e->paddb(xmmA, xmmB); - e->paddb(xmmA, anyptr_gpB); - e->paddw(xmmA, xmmB); - e->paddw(xmmA, anyptr_gpB); - e->paddd(xmmA, xmmB); - e->paddd(xmmA, anyptr_gpB); - e->paddq(mmA, mmB); - e->paddq(mmA, anyptr_gpB); - e->paddq(xmmA, xmmB); - e->paddq(xmmA, anyptr_gpB); - e->paddsb(xmmA, xmmB); - e->paddsb(xmmA, anyptr_gpB); - e->paddsw(xmmA, xmmB); - e->paddsw(xmmA, anyptr_gpB); - e->paddusb(xmmA, xmmB); - e->paddusb(xmmA, anyptr_gpB); - e->paddusw(xmmA, xmmB); - e->paddusw(xmmA, anyptr_gpB); - e->pand(xmmA, xmmB); - e->pand(xmmA, anyptr_gpB); - e->pandn(xmmA, xmmB); - e->pandn(xmmA, anyptr_gpB); - e->pause(); - e->pavgb(xmmA, xmmB); - e->pavgb(xmmA, anyptr_gpB); - e->pavgw(xmmA, xmmB); - e->pavgw(xmmA, anyptr_gpB); - e->pcmpeqb(xmmA, xmmB); - e->pcmpeqb(xmmA, anyptr_gpB); - e->pcmpeqw(xmmA, xmmB); - e->pcmpeqw(xmmA, anyptr_gpB); - e->pcmpeqd(xmmA, xmmB); - e->pcmpeqd(xmmA, anyptr_gpB); - e->pcmpgtb(xmmA, xmmB); - e->pcmpgtb(xmmA, anyptr_gpB); - e->pcmpgtw(xmmA, xmmB); - e->pcmpgtw(xmmA, anyptr_gpB); - e->pcmpgtd(xmmA, xmmB); - e->pcmpgtd(xmmA, anyptr_gpB); - e->pmaxsw(xmmA, xmmB); - e->pmaxsw(xmmA, anyptr_gpB); - e->pmaxub(xmmA, xmmB); - e->pmaxub(xmmA, anyptr_gpB); - e->pminsw(xmmA, xmmB); - e->pminsw(xmmA, anyptr_gpB); - e->pminub(xmmA, xmmB); - e->pminub(xmmA, anyptr_gpB); - e->pmovmskb(gdA, xmmB); - e->pmovmskb(gzA, xmmB); - e->pmulhw(xmmA, xmmB); - e->pmulhw(xmmA, anyptr_gpB); - e->pmulhuw(xmmA, xmmB); - e->pmulhuw(xmmA, anyptr_gpB); - e->pmullw(xmmA, xmmB); - e->pmullw(xmmA, anyptr_gpB); - e->pmuludq(mmA, mmB); - e->pmuludq(mmA, anyptr_gpB); - e->pmuludq(xmmA, xmmB); - e->pmuludq(xmmA, anyptr_gpB); - e->por(xmmA, xmmB); - e->por(xmmA, anyptr_gpB); - e->pslld(xmmA, xmmB); - e->pslld(xmmA, anyptr_gpB); - e->pslld(xmmA, 0); - e->psllq(xmmA, xmmB); - e->psllq(xmmA, anyptr_gpB); - e->psllq(xmmA, 0); - e->psllw(xmmA, xmmB); - e->psllw(xmmA, anyptr_gpB); - e->psllw(xmmA, 0); - e->pslldq(xmmA, 0); - e->psrad(xmmA, xmmB); - e->psrad(xmmA, anyptr_gpB); - e->psrad(xmmA, 0); - e->psraw(xmmA, xmmB); - e->psraw(xmmA, anyptr_gpB); - e->psraw(xmmA, 0); - e->psubb(xmmA, xmmB); - e->psubb(xmmA, anyptr_gpB); - e->psubw(xmmA, xmmB); - e->psubw(xmmA, anyptr_gpB); - e->psubd(xmmA, xmmB); - e->psubd(xmmA, anyptr_gpB); - e->psubq(mmA, mmB); - e->psubq(mmA, anyptr_gpB); - e->psubq(xmmA, xmmB); - e->psubq(xmmA, anyptr_gpB); - e->pmaddwd(xmmA, xmmB); - e->pmaddwd(xmmA, anyptr_gpB); - e->pshufd(xmmA, xmmB, 0); - e->pshufd(xmmA, anyptr_gpB, 0); - e->pshufhw(xmmA, xmmB, 0); - e->pshufhw(xmmA, anyptr_gpB, 0); - e->pshuflw(xmmA, xmmB, 0); - e->pshuflw(xmmA, anyptr_gpB, 0); - e->psrld(xmmA, xmmB); - e->psrld(xmmA, anyptr_gpB); - e->psrld(xmmA, 0); - e->psrlq(xmmA, xmmB); - e->psrlq(xmmA, anyptr_gpB); - e->psrlq(xmmA, 0); - e->psrldq(xmmA, 0); - e->psrlw(xmmA, xmmB); - e->psrlw(xmmA, anyptr_gpB); - e->psrlw(xmmA, 0); - e->psubsb(xmmA, xmmB); - e->psubsb(xmmA, anyptr_gpB); - e->psubsw(xmmA, xmmB); - e->psubsw(xmmA, anyptr_gpB); - e->psubusb(xmmA, xmmB); - e->psubusb(xmmA, anyptr_gpB); - e->psubusw(xmmA, xmmB); - e->psubusw(xmmA, anyptr_gpB); - e->punpckhbw(xmmA, xmmB); - e->punpckhbw(xmmA, anyptr_gpB); - e->punpckhwd(xmmA, xmmB); - e->punpckhwd(xmmA, anyptr_gpB); - e->punpckhdq(xmmA, xmmB); - e->punpckhdq(xmmA, anyptr_gpB); - e->punpckhqdq(xmmA, xmmB); - e->punpckhqdq(xmmA, anyptr_gpB); - e->punpcklbw(xmmA, xmmB); - e->punpcklbw(xmmA, anyptr_gpB); - e->punpcklwd(xmmA, xmmB); - e->punpcklwd(xmmA, anyptr_gpB); - e->punpckldq(xmmA, xmmB); - e->punpckldq(xmmA, anyptr_gpB); - e->punpcklqdq(xmmA, xmmB); - e->punpcklqdq(xmmA, anyptr_gpB); - e->pxor(xmmA, xmmB); - e->pxor(xmmA, anyptr_gpB); - e->sqrtpd(xmmA, xmmB); - e->sqrtpd(xmmA, anyptr_gpB); - e->sqrtsd(xmmA, xmmB); - e->sqrtsd(xmmA, anyptr_gpB); - e->subpd(xmmA, xmmB); - e->subpd(xmmA, anyptr_gpB); - e->subsd(xmmA, xmmB); - e->subsd(xmmA, anyptr_gpB); - e->ucomisd(xmmA, xmmB); - e->ucomisd(xmmA, anyptr_gpB); - e->unpckhpd(xmmA, xmmB); - e->unpckhpd(xmmA, anyptr_gpB); - e->unpcklpd(xmmA, xmmB); - e->unpcklpd(xmmA, anyptr_gpB); - e->xorpd(xmmA, xmmB); - e->xorpd(xmmA, anyptr_gpB); - - // SSE3. - e->nop(); - - e->addsubpd(xmmA, xmmB); - e->addsubpd(xmmA, anyptr_gpB); - e->addsubps(xmmA, xmmB); - e->addsubps(xmmA, anyptr_gpB); - e->fisttp(dword_ptr(gzA)); - e->haddpd(xmmA, xmmB); - e->haddpd(xmmA, anyptr_gpB); - e->haddps(xmmA, xmmB); - e->haddps(xmmA, anyptr_gpB); - e->hsubpd(xmmA, xmmB); - e->hsubpd(xmmA, anyptr_gpB); - e->hsubps(xmmA, xmmB); - e->hsubps(xmmA, anyptr_gpB); - e->lddqu(xmmA, anyptr_gpB); - e->movddup(xmmA, xmmB); - e->movddup(xmmA, anyptr_gpB); - e->movshdup(xmmA, xmmB); - e->movshdup(xmmA, anyptr_gpB); - e->movsldup(xmmA, xmmB); - e->movsldup(xmmA, anyptr_gpB); - - // SSSE3. - e->nop(); - - e->psignb(mmA, mmB); - e->psignb(mmA, anyptr_gpB); - e->psignb(xmmA, xmmB); - e->psignb(xmmA, anyptr_gpB); - e->psignw(mmA, mmB); - e->psignw(mmA, anyptr_gpB); - e->psignw(xmmA, xmmB); - e->psignw(xmmA, anyptr_gpB); - e->psignd(mmA, mmB); - e->psignd(mmA, anyptr_gpB); - e->psignd(xmmA, xmmB); - e->psignd(xmmA, anyptr_gpB); - e->phaddw(mmA, mmB); - e->phaddw(mmA, anyptr_gpB); - e->phaddw(xmmA, xmmB); - e->phaddw(xmmA, anyptr_gpB); - e->phaddd(mmA, mmB); - e->phaddd(mmA, anyptr_gpB); - e->phaddd(xmmA, xmmB); - e->phaddd(xmmA, anyptr_gpB); - e->phaddsw(mmA, mmB); - e->phaddsw(mmA, anyptr_gpB); - e->phaddsw(xmmA, xmmB); - e->phaddsw(xmmA, anyptr_gpB); - e->phsubw(mmA, mmB); - e->phsubw(mmA, anyptr_gpB); - e->phsubw(xmmA, xmmB); - e->phsubw(xmmA, anyptr_gpB); - e->phsubd(mmA, mmB); - e->phsubd(mmA, anyptr_gpB); - e->phsubd(xmmA, xmmB); - e->phsubd(xmmA, anyptr_gpB); - e->phsubsw(mmA, mmB); - e->phsubsw(mmA, anyptr_gpB); - e->phsubsw(xmmA, xmmB); - e->phsubsw(xmmA, anyptr_gpB); - e->pmaddubsw(mmA, mmB); - e->pmaddubsw(mmA, anyptr_gpB); - e->pmaddubsw(xmmA, xmmB); - e->pmaddubsw(xmmA, anyptr_gpB); - e->pabsb(mmA, mmB); - e->pabsb(mmA, anyptr_gpB); - e->pabsb(xmmA, xmmB); - e->pabsb(xmmA, anyptr_gpB); - e->pabsw(mmA, mmB); - e->pabsw(mmA, anyptr_gpB); - e->pabsw(xmmA, xmmB); - e->pabsw(xmmA, anyptr_gpB); - e->pabsd(mmA, mmB); - e->pabsd(mmA, anyptr_gpB); - e->pabsd(xmmA, xmmB); - e->pabsd(xmmA, anyptr_gpB); - e->pmulhrsw(mmA, mmB); - e->pmulhrsw(mmA, anyptr_gpB); - e->pmulhrsw(xmmA, xmmB); - e->pmulhrsw(xmmA, anyptr_gpB); - e->pshufb(mmA, mmB); - e->pshufb(mmA, anyptr_gpB); - e->pshufb(xmmA, xmmB); - e->pshufb(xmmA, anyptr_gpB); - e->palignr(mmA, mmB, 0); - e->palignr(mmA, anyptr_gpB, 0); - e->palignr(xmmA, xmmB, 0); - e->palignr(xmmA, anyptr_gpB, 0); - - // SSE4.1. - e->nop(); - - e->blendpd(xmmA, xmmB, 0); - e->blendpd(xmmA, anyptr_gpB, 0); - e->blendps(xmmA, xmmB, 0); - e->blendps(xmmA, anyptr_gpB, 0); - e->blendvpd(xmmA, xmmB); // Implicit xmmA, xmmB, <XMM0> - e->blendvpd(xmmA, xmmB, xmm0); // Explicit xmmA, xmmB, <XMM0> - e->blendvpd(xmmA, anyptr_gpB); // Implicit xmmA, mem , <XMM0> - e->blendvpd(xmmA, anyptr_gpB, xmm0); // Explicit xmmA, mem , <XMM0> - e->blendvps(xmmA, xmmB); // Implicit xmmA, xmmB, <XMM0> - e->blendvps(xmmA, xmmB, xmm0); // Explicit xmmA, xmmB, <XMM0> - e->blendvps(xmmA, anyptr_gpB); // Implicit xmmA, mem , <XMM0> - e->blendvps(xmmA, anyptr_gpB, xmm0); // Explicit xmmA, mem , <XMM0> - - e->dppd(xmmA, xmmB, 0); - e->dppd(xmmA, anyptr_gpB, 0); - e->dpps(xmmA, xmmB, 0); - e->dpps(xmmA, anyptr_gpB, 0); - e->extractps(gdA, xmmB, 0); - e->extractps(gzA, xmmB, 0); - e->extractps(anyptr_gpA, xmmB, 0); - e->insertps(xmmA, xmmB, 0); - e->insertps(xmmA, anyptr_gpB, 0); - e->movntdqa(xmmA, anyptr_gpB); - e->mpsadbw(xmmA, xmmB, 0); - e->mpsadbw(xmmA, anyptr_gpB, 0); - e->packusdw(xmmA, xmmB); - e->packusdw(xmmA, anyptr_gpB); - e->pblendvb(xmmA, xmmB); // Implicit xmmA, xmmB, <XMM0> - e->pblendvb(xmmA, xmmB, xmm0); // Explicit xmmA, xmmB, <XMM0> - e->pblendvb(xmmA, anyptr_gpB); // Implicit xmmA, mem, <XMM0> - e->pblendvb(xmmA, anyptr_gpB, xmm0); // Implicit xmmA, mem, <XMM0> - e->pblendw(xmmA, xmmB, 0); - e->pblendw(xmmA, anyptr_gpB, 0); - e->pcmpeqq(xmmA, xmmB); - e->pcmpeqq(xmmA, anyptr_gpB); - e->pextrb(gdA, xmmB, 0); - e->pextrb(gzA, xmmB, 0); - e->pextrb(anyptr_gpA, xmmB, 0); - e->pextrd(gdA, xmmB, 0); - e->pextrd(gzA, xmmB, 0); - e->pextrd(anyptr_gpA, xmmB, 0); - if (isX64) e->pextrq(gzA, xmmB, 0); - if (isX64) e->pextrq(anyptr_gpA, xmmB, 0); - e->pextrw(gdA, xmmB, 0); - e->pextrw(gzA, xmmB, 0); - e->pextrw(anyptr_gpA, xmmB, 0); - e->phminposuw(xmmA, xmmB); - e->phminposuw(xmmA, anyptr_gpB); - e->pinsrb(xmmA, gdB, 0); - e->pinsrb(xmmA, gzB, 0); - e->pinsrb(xmmA, anyptr_gpB, 0); - e->pinsrd(xmmA, gdB, 0); - e->pinsrd(xmmA, gzB, 0); - e->pinsrd(xmmA, anyptr_gpB, 0); - e->pinsrw(xmmA, gdB, 0); - e->pinsrw(xmmA, gzB, 0); - e->pinsrw(xmmA, anyptr_gpB, 0); - e->pmaxuw(xmmA, xmmB); - e->pmaxuw(xmmA, anyptr_gpB); - e->pmaxsb(xmmA, xmmB); - e->pmaxsb(xmmA, anyptr_gpB); - e->pmaxsd(xmmA, xmmB); - e->pmaxsd(xmmA, anyptr_gpB); - e->pmaxud(xmmA, xmmB); - e->pmaxud(xmmA, anyptr_gpB); - e->pminsb(xmmA, xmmB); - e->pminsb(xmmA, anyptr_gpB); - e->pminuw(xmmA, xmmB); - e->pminuw(xmmA, anyptr_gpB); - e->pminud(xmmA, xmmB); - e->pminud(xmmA, anyptr_gpB); - e->pminsd(xmmA, xmmB); - e->pminsd(xmmA, anyptr_gpB); - e->pmovsxbw(xmmA, xmmB); - e->pmovsxbw(xmmA, anyptr_gpB); - e->pmovsxbd(xmmA, xmmB); - e->pmovsxbd(xmmA, anyptr_gpB); - e->pmovsxbq(xmmA, xmmB); - e->pmovsxbq(xmmA, anyptr_gpB); - e->pmovsxwd(xmmA, xmmB); - e->pmovsxwd(xmmA, anyptr_gpB); - e->pmovsxwq(xmmA, xmmB); - e->pmovsxwq(xmmA, anyptr_gpB); - e->pmovsxdq(xmmA, xmmB); - e->pmovsxdq(xmmA, anyptr_gpB); - e->pmovzxbw(xmmA, xmmB); - e->pmovzxbw(xmmA, anyptr_gpB); - e->pmovzxbd(xmmA, xmmB); - e->pmovzxbd(xmmA, anyptr_gpB); - e->pmovzxbq(xmmA, xmmB); - e->pmovzxbq(xmmA, anyptr_gpB); - e->pmovzxwd(xmmA, xmmB); - e->pmovzxwd(xmmA, anyptr_gpB); - e->pmovzxwq(xmmA, xmmB); - e->pmovzxwq(xmmA, anyptr_gpB); - e->pmovzxdq(xmmA, xmmB); - e->pmovzxdq(xmmA, anyptr_gpB); - e->pmuldq(xmmA, xmmB); - e->pmuldq(xmmA, anyptr_gpB); - e->pmulld(xmmA, xmmB); - e->pmulld(xmmA, anyptr_gpB); - e->ptest(xmmA, xmmB); - e->ptest(xmmA, anyptr_gpB); - e->roundps(xmmA, xmmB, 0); - e->roundps(xmmA, anyptr_gpB, 0); - e->roundss(xmmA, xmmB, 0); - e->roundss(xmmA, anyptr_gpB, 0); - e->roundpd(xmmA, xmmB, 0); - e->roundpd(xmmA, anyptr_gpB, 0); - e->roundsd(xmmA, xmmB, 0); - e->roundsd(xmmA, anyptr_gpB, 0); - - // SSE4.2. - e->nop(); - - e->pcmpestri(xmmA, xmmB , imm(0)); // Implicit xmmA, xmmB, imm, <ECX>, <EAX>, <EDX> - e->pcmpestri(xmmA, xmmB , imm(0), ecx, eax, edx); // Explicit xmmA, xmmB, imm, <ECX>, <EAX>, <EDX> - e->pcmpestri(xmmA, anyptr_gpB, imm(0)); // Implicit xmmA, mem , imm, <ECX>, <EAX>, <EDX> - e->pcmpestri(xmmA, anyptr_gpB, imm(0), ecx, eax, edx); // Explicit xmmA, mem , imm, <ECX>, <EAX>, <EDX> - e->pcmpestrm(xmmA, xmmB , imm(0)); // Implicit xmmA, xmmB, imm, <XMM0>, <EAX>, <EDX> - e->pcmpestrm(xmmA, xmmB , imm(0), xmm0, eax, edx); // Explicit xmmA, xmmB, imm, <XMM0>, <EAX>, <EDX> - e->pcmpestrm(xmmA, anyptr_gpB, imm(0)); // Implicit xmmA, mem , imm, <XMM0>, <EAX>, <EDX> - e->pcmpestrm(xmmA, anyptr_gpB, imm(0), xmm0, eax, edx); // Explicit xmmA, mem , imm, <XMM0>, <EAX>, <EDX> - e->pcmpistri(xmmA, xmmB , imm(0)); // Implicit xmmA, xmmB, imm, <ECX> - e->pcmpistri(xmmA, xmmB , imm(0), ecx); // Explicit xmmA, xmmB, imm, <ECX> - e->pcmpistri(xmmA, anyptr_gpB, imm(0)); // Implicit xmmA, mem , imm, <ECX> - e->pcmpistri(xmmA, anyptr_gpB, imm(0), ecx); // Explicit xmmA, mem , imm, <ECX> - e->pcmpistrm(xmmA, xmmB , imm(0)); // Implicit xmmA, xmmB, imm, <XMM0> - e->pcmpistrm(xmmA, xmmB , imm(0), xmm0); // Explicit xmmA, xmmB, imm, <XMM0> - e->pcmpistrm(xmmA, anyptr_gpB, imm(0)); // Implicit xmmA, mem , imm, <XMM0> - e->pcmpistrm(xmmA, anyptr_gpB, imm(0), xmm0); // Explicit xmmA, mem , imm, <XMM0> - - e->pcmpgtq(xmmA, xmmB); - e->pcmpgtq(xmmA, anyptr_gpB); - - // SSE4A. - e->nop(); - - e->extrq(xmmA, xmmB); - e->extrq(xmmA, 0x1, 0x2); - e->extrq(xmmB, 0x1, 0x2); - e->insertq(xmmA, xmmB); - e->insertq(xmmA, xmmB, 0x1, 0x2); - e->movntsd(anyptr_gpA, xmmB); - e->movntss(anyptr_gpA, xmmB); - - // AESNI. - e->nop(); - - e->aesdec(xmmA, xmmB); - e->aesdec(xmmA, anyptr_gpB); - e->aesdeclast(xmmA, xmmB); - e->aesdeclast(xmmA, anyptr_gpB); - e->aesenc(xmmA, xmmB); - e->aesenc(xmmA, anyptr_gpB); - e->aesenclast(xmmA, xmmB); - e->aesenclast(xmmA, anyptr_gpB); - e->aesimc(xmmA, xmmB); - e->aesimc(xmmA, anyptr_gpB); - e->aeskeygenassist(xmmA, xmmB, 0); - e->aeskeygenassist(xmmA, anyptr_gpB, 0); - - // SHA. - e->nop(); - - e->sha1msg1(xmmA, xmmB); - e->sha1msg1(xmmA, anyptr_gpB); - e->sha1msg2(xmmA, xmmB); - e->sha1msg2(xmmA, anyptr_gpB); - e->sha1nexte(xmmA, xmmB); - e->sha1nexte(xmmA, anyptr_gpB); - e->sha1rnds4(xmmA, xmmB, 0); - e->sha1rnds4(xmmA, anyptr_gpB, 0); - e->sha256msg1(xmmA, xmmB); - e->sha256msg1(xmmA, anyptr_gpB); - e->sha256msg2(xmmA, xmmB); - e->sha256msg2(xmmA, anyptr_gpB); - e->sha256rnds2(xmmA, xmmB); // Implicit xmmA, xmmB, <XMM0> - e->sha256rnds2(xmmA, xmmB, xmm0); // Explicit xmmA, xmmB, <XMM0> - e->sha256rnds2(xmmA, anyptr_gpB); // Implicit xmmA, mem, <XMM0> - e->sha256rnds2(xmmA, anyptr_gpB, xmm0); // Explicit xmmA, mem, <XMM0> - - // PCLMULQDQ. - e->nop(); - - e->pclmulqdq(xmmA, xmmB, 0); - e->pclmulqdq(xmmA, anyptr_gpB, 0); - - // AVX. - e->nop(); - - e->vaddpd(xmmA, xmmB, xmmC); - e->vaddpd(xmmA, xmmB, anyptr_gpC); - e->vaddpd(ymmA, ymmB, ymmC); - e->vaddpd(ymmA, ymmB, anyptr_gpC); - e->vaddps(xmmA, xmmB, xmmC); - e->vaddps(xmmA, xmmB, anyptr_gpC); - e->vaddps(ymmA, ymmB, ymmC); - e->vaddps(ymmA, ymmB, anyptr_gpC); - e->vaddsd(xmmA, xmmB, xmmC); - e->vaddsd(xmmA, xmmB, anyptr_gpC); - e->vaddss(xmmA, xmmB, xmmC); - e->vaddss(xmmA, xmmB, anyptr_gpC); - e->vaddsubpd(xmmA, xmmB, xmmC); - e->vaddsubpd(xmmA, xmmB, anyptr_gpC); - e->vaddsubpd(ymmA, ymmB, ymmC); - e->vaddsubpd(ymmA, ymmB, anyptr_gpC); - e->vaddsubps(xmmA, xmmB, xmmC); - e->vaddsubps(xmmA, xmmB, anyptr_gpC); - e->vaddsubps(ymmA, ymmB, ymmC); - e->vaddsubps(ymmA, ymmB, anyptr_gpC); - e->vandpd(xmmA, xmmB, xmmC); - e->vandpd(xmmA, xmmB, anyptr_gpC); - e->vandpd(ymmA, ymmB, ymmC); - e->vandpd(ymmA, ymmB, anyptr_gpC); - e->vandps(xmmA, xmmB, xmmC); - e->vandps(xmmA, xmmB, anyptr_gpC); - e->vandps(ymmA, ymmB, ymmC); - e->vandps(ymmA, ymmB, anyptr_gpC); - e->vandnpd(xmmA, xmmB, xmmC); - e->vandnpd(xmmA, xmmB, anyptr_gpC); - e->vandnpd(ymmA, ymmB, ymmC); - e->vandnpd(ymmA, ymmB, anyptr_gpC); - e->vandnps(xmmA, xmmB, xmmC); - e->vandnps(xmmA, xmmB, anyptr_gpC); - e->vandnps(ymmA, ymmB, ymmC); - e->vandnps(ymmA, ymmB, anyptr_gpC); - e->vblendpd(xmmA, xmmB, xmmC, 0); - e->vblendpd(xmmA, xmmB, anyptr_gpC, 0); - e->vblendpd(ymmA, ymmB, ymmC, 0); - e->vblendpd(ymmA, ymmB, anyptr_gpC, 0); - e->vblendps(xmmA, xmmB, xmmC, 0); - e->vblendps(xmmA, xmmB, anyptr_gpC, 0); - e->vblendps(ymmA, ymmB, ymmC, 0); - e->vblendps(ymmA, ymmB, anyptr_gpC, 0); - e->vblendvpd(xmmA, xmmB, xmmC, xmmD); - e->vblendvpd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vblendvpd(ymmA, ymmB, ymmC, ymmD); - e->vblendvpd(ymmA, ymmB, anyptr_gpC, ymmD); - e->vbroadcastf128(ymmA, anyptr_gpB); - e->vbroadcastsd(ymmA, anyptr_gpB); - e->vbroadcastss(xmmA, anyptr_gpB); - e->vbroadcastss(ymmA, anyptr_gpB); - e->vcmppd(xmmA, xmmB, xmmC, 0); - e->vcmppd(xmmA, xmmB, anyptr_gpC, 0); - e->vcmppd(ymmA, ymmB, ymmC, 0); - e->vcmppd(ymmA, ymmB, anyptr_gpC, 0); - e->vcmpps(xmmA, xmmB, xmmC, 0); - e->vcmpps(xmmA, xmmB, anyptr_gpC, 0); - e->vcmpps(ymmA, ymmB, ymmC, 0); - e->vcmpps(ymmA, ymmB, anyptr_gpC, 0); - e->vcmpsd(xmmA, xmmB, xmmC, 0); - e->vcmpsd(xmmA, xmmB, anyptr_gpC, 0); - e->vcmpss(xmmA, xmmB, xmmC, 0); - e->vcmpss(xmmA, xmmB, anyptr_gpC, 0); - e->vcomisd(xmmA, xmmB); - e->vcomisd(xmmA, anyptr_gpB); - e->vcomiss(xmmA, xmmB); - e->vcomiss(xmmA, anyptr_gpB); - e->vcvtdq2pd(xmmA, xmmB); - e->vcvtdq2pd(xmmA, anyptr_gpB); - e->vcvtdq2pd(ymmA, xmmB); - e->vcvtdq2pd(ymmA, anyptr_gpB); - e->vcvtdq2ps(xmmA, xmmB); - e->vcvtdq2ps(xmmA, anyptr_gpB); - e->vcvtdq2ps(ymmA, ymmB); - e->vcvtdq2ps(ymmA, anyptr_gpB); - e->vcvtpd2dq(xmmA, xmmB); - e->vcvtpd2dq(xmmA, ymmB); - e->vcvtpd2dq(xmmA, anyptr_gpB); - e->vcvtpd2ps(xmmA, xmmB); - e->vcvtpd2ps(xmmA, ymmB); - e->vcvtpd2ps(xmmA, anyptr_gpB); - e->vcvtps2dq(xmmA, xmmB); - e->vcvtps2dq(xmmA, anyptr_gpB); - e->vcvtps2dq(ymmA, ymmB); - e->vcvtps2dq(ymmA, anyptr_gpB); - e->vcvtps2pd(xmmA, xmmB); - e->vcvtps2pd(xmmA, anyptr_gpB); - e->vcvtps2pd(ymmA, xmmB); - e->vcvtps2pd(ymmA, anyptr_gpB); - e->vcvtsd2si(gzA, xmmB); - e->vcvtsd2si(gzA, anyptr_gpB); - e->vcvtsd2ss(xmmA, xmmB, xmmC); - e->vcvtsd2ss(xmmA, xmmB, anyptr_gpC); - e->vcvtsi2sd(xmmA, xmmB, gzC); - e->vcvtsi2sd(xmmA, xmmB, anyptr_gpC); - e->vcvtsi2ss(xmmA, xmmB, gzC); - e->vcvtsi2ss(xmmA, xmmB, anyptr_gpC); - e->vcvtss2sd(xmmA, xmmB, xmmC); - e->vcvtss2sd(xmmA, xmmB, anyptr_gpC); - e->vcvtss2si(gzA, xmmB); - e->vcvtss2si(gzA, anyptr_gpB); - e->vcvttpd2dq(xmmA, xmmB); - e->vcvttpd2dq(xmmA, ymmB); - e->vcvttpd2dq(xmmA, anyptr_gpB); - e->vcvttps2dq(xmmA, xmmB); - e->vcvttps2dq(xmmA, anyptr_gpB); - e->vcvttps2dq(ymmA, ymmB); - e->vcvttps2dq(ymmA, anyptr_gpB); - e->vcvttsd2si(gzA, xmmB); - e->vcvttsd2si(gzA, anyptr_gpB); - e->vcvttss2si(gzA, xmmB); - e->vcvttss2si(gzA, anyptr_gpB); - e->vdivpd(xmmA, xmmB, xmmC); - e->vdivpd(xmmA, xmmB, anyptr_gpC); - e->vdivpd(ymmA, ymmB, ymmC); - e->vdivpd(ymmA, ymmB, anyptr_gpC); - e->vdivps(xmmA, xmmB, xmmC); - e->vdivps(xmmA, xmmB, anyptr_gpC); - e->vdivps(ymmA, ymmB, ymmC); - e->vdivps(ymmA, ymmB, anyptr_gpC); - e->vdivsd(xmmA, xmmB, xmmC); - e->vdivsd(xmmA, xmmB, anyptr_gpC); - e->vdivss(xmmA, xmmB, xmmC); - e->vdivss(xmmA, xmmB, anyptr_gpC); - e->vdppd(xmmA, xmmB, xmmC, 0); - e->vdppd(xmmA, xmmB, anyptr_gpC, 0); - e->vdpps(xmmA, xmmB, xmmC, 0); - e->vdpps(xmmA, xmmB, anyptr_gpC, 0); - e->vdpps(ymmA, ymmB, ymmC, 0); - e->vdpps(ymmA, ymmB, anyptr_gpC, 0); - e->vextractf128(xmmA, ymmB, 0); - e->vextractf128(anyptr_gpA, ymmB, 0); - e->vextractps(gzA, xmmB, 0); - e->vextractps(anyptr_gpA, xmmB, 0); - e->vhaddpd(xmmA, xmmB, xmmC); - e->vhaddpd(xmmA, xmmB, anyptr_gpC); - e->vhaddpd(ymmA, ymmB, ymmC); - e->vhaddpd(ymmA, ymmB, anyptr_gpC); - e->vhaddps(xmmA, xmmB, xmmC); - e->vhaddps(xmmA, xmmB, anyptr_gpC); - e->vhaddps(ymmA, ymmB, ymmC); - e->vhaddps(ymmA, ymmB, anyptr_gpC); - e->vhsubpd(xmmA, xmmB, xmmC); - e->vhsubpd(xmmA, xmmB, anyptr_gpC); - e->vhsubpd(ymmA, ymmB, ymmC); - e->vhsubpd(ymmA, ymmB, anyptr_gpC); - e->vhsubps(xmmA, xmmB, xmmC); - e->vhsubps(xmmA, xmmB, anyptr_gpC); - e->vhsubps(ymmA, ymmB, ymmC); - e->vhsubps(ymmA, ymmB, anyptr_gpC); - e->vinsertf128(ymmA, ymmB, xmmC, 0); - e->vinsertf128(ymmA, ymmB, anyptr_gpC, 0); - e->vinsertps(xmmA, xmmB, xmmC, 0); - e->vinsertps(xmmA, xmmB, anyptr_gpC, 0); - e->vlddqu(xmmA, anyptr_gpB); - e->vlddqu(ymmA, anyptr_gpB); - e->vldmxcsr(anyptr_gpA); - e->vmaskmovdqu(xmmA, xmmB); // Implicit xmmA, xmmB, <ds:[EDI|RDI]> - e->vmaskmovdqu(xmmA, xmmB, ptr(e->zdi())); // Explicit xmmA, xmmB, <ds:[EDI|RDI]> - e->vmaskmovps(xmmA, xmmB, anyptr_gpC); - e->vmaskmovps(ymmA, ymmB, anyptr_gpC); - e->vmaskmovps(anyptr_gpA, xmmB, xmmC); - e->vmaskmovps(anyptr_gpA, ymmB, ymmC); - e->vmaskmovpd(xmmA, xmmB, anyptr_gpC); - e->vmaskmovpd(ymmA, ymmB, anyptr_gpC); - e->vmaskmovpd(anyptr_gpA, xmmB, xmmC); - e->vmaskmovpd(anyptr_gpA, ymmB, ymmC); - e->vmaxpd(xmmA, xmmB, xmmC); - e->vmaxpd(xmmA, xmmB, anyptr_gpC); - e->vmaxpd(ymmA, ymmB, ymmC); - e->vmaxpd(ymmA, ymmB, anyptr_gpC); - e->vmaxps(xmmA, xmmB, xmmC); - e->vmaxps(xmmA, xmmB, anyptr_gpC); - e->vmaxps(ymmA, ymmB, ymmC); - e->vmaxps(ymmA, ymmB, anyptr_gpC); - e->vmaxsd(xmmA, xmmB, xmmC); - e->vmaxsd(xmmA, xmmB, anyptr_gpC); - e->vmaxss(xmmA, xmmB, xmmC); - e->vmaxss(xmmA, xmmB, anyptr_gpC); - e->vminpd(xmmA, xmmB, xmmC); - e->vminpd(xmmA, xmmB, anyptr_gpC); - e->vminpd(ymmA, ymmB, ymmC); - e->vminpd(ymmA, ymmB, anyptr_gpC); - e->vminps(xmmA, xmmB, xmmC); - e->vminps(xmmA, xmmB, anyptr_gpC); - e->vminps(ymmA, ymmB, ymmC); - e->vminps(ymmA, ymmB, anyptr_gpC); - e->vminsd(xmmA, xmmB, xmmC); - e->vminsd(xmmA, xmmB, anyptr_gpC); - e->vminss(xmmA, xmmB, xmmC); - e->vminss(xmmA, xmmB, anyptr_gpC); - e->vmovapd(xmmA, xmmB); - e->vmovapd(xmmA, anyptr_gpB); - e->vmovapd(anyptr_gpA, xmmB); - e->vmovapd(ymmA, ymmB); - e->vmovapd(ymmA, anyptr_gpB); - e->vmovapd(anyptr_gpA, ymmB); - e->vmovaps(xmmA, xmmB); - e->vmovaps(xmmA, anyptr_gpB); - e->vmovaps(anyptr_gpA, xmmB); - e->vmovaps(ymmA, ymmB); - e->vmovaps(ymmA, anyptr_gpB); - e->vmovaps(anyptr_gpA, ymmB); - e->vmovd(xmmA, gzB); - e->vmovd(xmmA, anyptr_gpB); - e->vmovd(gzA, xmmB); - e->vmovd(anyptr_gpA, xmmB); - e->vmovddup(xmmA, xmmB); - e->vmovddup(xmmA, anyptr_gpB); - e->vmovddup(ymmA, ymmB); - e->vmovddup(ymmA, anyptr_gpB); - e->vmovdqa(xmmA, xmmB); - e->vmovdqa(xmmA, anyptr_gpB); - e->vmovdqa(anyptr_gpA, xmmB); - e->vmovdqa(ymmA, ymmB); - e->vmovdqa(ymmA, anyptr_gpB); - e->vmovdqa(anyptr_gpA, ymmB); - e->vmovdqu(xmmA, xmmB); - e->vmovdqu(xmmA, anyptr_gpB); - e->vmovdqu(anyptr_gpA, xmmB); - e->vmovdqu(ymmA, ymmB); - e->vmovdqu(ymmA, anyptr_gpB); - e->vmovdqu(anyptr_gpA, ymmB); - e->vmovhlps(xmmA, xmmB, xmmC); - e->vmovhpd(xmmA, xmmB, anyptr_gpC); - e->vmovhpd(anyptr_gpA, xmmB); - e->vmovhps(xmmA, xmmB, anyptr_gpC); - e->vmovhps(anyptr_gpA, xmmB); - e->vmovlhps(xmmA, xmmB, xmmC); - e->vmovlpd(xmmA, xmmB, anyptr_gpC); - e->vmovlpd(anyptr_gpA, xmmB); - e->vmovlps(xmmA, xmmB, anyptr_gpC); - e->vmovlps(anyptr_gpA, xmmB); - e->vmovmskpd(gzA, xmmB); - e->vmovmskpd(gzA, ymmB); - e->vmovmskps(gzA, xmmB); - e->vmovmskps(gzA, ymmB); - e->vmovntdq(anyptr_gpA, xmmB); - e->vmovntdq(anyptr_gpA, ymmB); - e->vmovntdqa(xmmA, anyptr_gpB); - e->vmovntpd(anyptr_gpA, xmmB); - e->vmovntpd(anyptr_gpA, ymmB); - e->vmovntps(anyptr_gpA, xmmB); - e->vmovntps(anyptr_gpA, ymmB); - e->vmovsd(xmmA, xmmB, xmmC); - e->vmovsd(xmmA, anyptr_gpB); - e->vmovsd(anyptr_gpA, xmmB); - e->vmovshdup(xmmA, xmmB); - e->vmovshdup(xmmA, anyptr_gpB); - e->vmovshdup(ymmA, ymmB); - e->vmovshdup(ymmA, anyptr_gpB); - e->vmovsldup(xmmA, xmmB); - e->vmovsldup(xmmA, anyptr_gpB); - e->vmovsldup(ymmA, ymmB); - e->vmovsldup(ymmA, anyptr_gpB); - e->vmovss(xmmA, xmmB, xmmC); - e->vmovss(xmmA, anyptr_gpB); - e->vmovss(anyptr_gpA, xmmB); - e->vmovupd(xmmA, xmmB); - e->vmovupd(xmmA, anyptr_gpB); - e->vmovupd(anyptr_gpA, xmmB); - e->vmovupd(ymmA, ymmB); - e->vmovupd(ymmA, anyptr_gpB); - e->vmovupd(anyptr_gpA, ymmB); - e->vmovups(xmmA, xmmB); - e->vmovups(xmmA, anyptr_gpB); - e->vmovups(anyptr_gpA, xmmB); - e->vmovups(ymmA, ymmB); - e->vmovups(ymmA, anyptr_gpB); - e->vmovups(anyptr_gpA, ymmB); - e->vmpsadbw(xmmA, xmmB, xmmC, 0); - e->vmpsadbw(xmmA, xmmB, anyptr_gpC, 0); - e->vmulpd(xmmA, xmmB, xmmC); - e->vmulpd(xmmA, xmmB, anyptr_gpC); - e->vmulpd(ymmA, ymmB, ymmC); - e->vmulpd(ymmA, ymmB, anyptr_gpC); - e->vmulps(xmmA, xmmB, xmmC); - e->vmulps(xmmA, xmmB, anyptr_gpC); - e->vmulps(ymmA, ymmB, ymmC); - e->vmulps(ymmA, ymmB, anyptr_gpC); - e->vmulsd(xmmA, xmmB, xmmC); - e->vmulsd(xmmA, xmmB, anyptr_gpC); - e->vmulss(xmmA, xmmB, xmmC); - e->vmulss(xmmA, xmmB, anyptr_gpC); - e->vorpd(xmmA, xmmB, xmmC); - e->vorpd(xmmA, xmmB, anyptr_gpC); - e->vorpd(ymmA, ymmB, ymmC); - e->vorpd(ymmA, ymmB, anyptr_gpC); - e->vorps(xmmA, xmmB, xmmC); - e->vorps(xmmA, xmmB, anyptr_gpC); - e->vorps(ymmA, ymmB, ymmC); - e->vorps(ymmA, ymmB, anyptr_gpC); - e->vpabsb(xmmA, xmmB); - e->vpabsb(xmmA, anyptr_gpB); - e->vpabsd(xmmA, xmmB); - e->vpabsd(xmmA, anyptr_gpB); - e->vpabsw(xmmA, xmmB); - e->vpabsw(xmmA, anyptr_gpB); - e->vpackssdw(xmmA, xmmB, xmmC); - e->vpackssdw(xmmA, xmmB, anyptr_gpC); - e->vpacksswb(xmmA, xmmB, xmmC); - e->vpacksswb(xmmA, xmmB, anyptr_gpC); - e->vpackusdw(xmmA, xmmB, xmmC); - e->vpackusdw(xmmA, xmmB, anyptr_gpC); - e->vpackuswb(xmmA, xmmB, xmmC); - e->vpackuswb(xmmA, xmmB, anyptr_gpC); - e->vpaddb(xmmA, xmmB, xmmC); - e->vpaddb(xmmA, xmmB, anyptr_gpC); - e->vpaddd(xmmA, xmmB, xmmC); - e->vpaddd(xmmA, xmmB, anyptr_gpC); - e->vpaddq(xmmA, xmmB, xmmC); - e->vpaddq(xmmA, xmmB, anyptr_gpC); - e->vpaddw(xmmA, xmmB, xmmC); - e->vpaddw(xmmA, xmmB, anyptr_gpC); - e->vpaddsb(xmmA, xmmB, xmmC); - e->vpaddsb(xmmA, xmmB, anyptr_gpC); - e->vpaddsw(xmmA, xmmB, xmmC); - e->vpaddsw(xmmA, xmmB, anyptr_gpC); - e->vpaddusb(xmmA, xmmB, xmmC); - e->vpaddusb(xmmA, xmmB, anyptr_gpC); - e->vpaddusw(xmmA, xmmB, xmmC); - e->vpaddusw(xmmA, xmmB, anyptr_gpC); - e->vpalignr(xmmA, xmmB, xmmC, 0); - e->vpalignr(xmmA, xmmB, anyptr_gpC, 0); - e->vpand(xmmA, xmmB, xmmC); - e->vpand(xmmA, xmmB, anyptr_gpC); - e->vpandn(xmmA, xmmB, xmmC); - e->vpandn(xmmA, xmmB, anyptr_gpC); - e->vpavgb(xmmA, xmmB, xmmC); - e->vpavgb(xmmA, xmmB, anyptr_gpC); - e->vpavgw(xmmA, xmmB, xmmC); - e->vpavgw(xmmA, xmmB, anyptr_gpC); - e->vpblendvb(xmmA, xmmB, xmmC, xmmD); - e->vpblendvb(xmmA, xmmB, anyptr_gpC, xmmD); - e->vpblendw(xmmA, xmmB, xmmC, 0); - e->vpblendw(xmmA, xmmB, anyptr_gpC, 0); - e->vpcmpeqb(xmmA, xmmB, xmmC); - e->vpcmpeqb(xmmA, xmmB, anyptr_gpC); - e->vpcmpeqd(xmmA, xmmB, xmmC); - e->vpcmpeqd(xmmA, xmmB, anyptr_gpC); - e->vpcmpeqq(xmmA, xmmB, xmmC); - e->vpcmpeqq(xmmA, xmmB, anyptr_gpC); - e->vpcmpeqw(xmmA, xmmB, xmmC); - e->vpcmpeqw(xmmA, xmmB, anyptr_gpC); - e->vpcmpgtb(xmmA, xmmB, xmmC); - e->vpcmpgtb(xmmA, xmmB, anyptr_gpC); - e->vpcmpgtd(xmmA, xmmB, xmmC); - e->vpcmpgtd(xmmA, xmmB, anyptr_gpC); - e->vpcmpgtq(xmmA, xmmB, xmmC); - e->vpcmpgtq(xmmA, xmmB, anyptr_gpC); - e->vpcmpgtw(xmmA, xmmB, xmmC); - e->vpcmpgtw(xmmA, xmmB, anyptr_gpC); - e->vpcmpestri(xmmA, xmmB, 0); - e->vpcmpestri(xmmA, anyptr_gpB, 0); - e->vpcmpestrm(xmmA, xmmB, 0); - e->vpcmpestrm(xmmA, anyptr_gpB, 0); - e->vpcmpistri(xmmA, xmmB, 0); - e->vpcmpistri(xmmA, anyptr_gpB, 0); - e->vpcmpistrm(xmmA, xmmB, 0); - e->vpcmpistrm(xmmA, anyptr_gpB, 0); - e->vpermilpd(xmmA, xmmB, xmmC); - e->vpermilpd(xmmA, xmmB, anyptr_gpC); - e->vpermilpd(ymmA, ymmB, ymmC); - e->vpermilpd(ymmA, ymmB, anyptr_gpC); - e->vpermilpd(xmmA, xmmB, 0); - e->vpermilpd(xmmA, anyptr_gpB, 0); - e->vpermilpd(ymmA, ymmB, 0); - e->vpermilpd(ymmA, anyptr_gpB, 0); - e->vpermilps(xmmA, xmmB, xmmC); - e->vpermilps(xmmA, xmmB, anyptr_gpC); - e->vpermilps(ymmA, ymmB, ymmC); - e->vpermilps(ymmA, ymmB, anyptr_gpC); - e->vpermilps(xmmA, xmmB, 0); - e->vpermilps(xmmA, anyptr_gpB, 0); - e->vpermilps(ymmA, ymmB, 0); - e->vpermilps(ymmA, anyptr_gpB, 0); - e->vperm2f128(ymmA, ymmB, ymmC, 0); - e->vperm2f128(ymmA, ymmB, anyptr_gpC, 0); - e->vpextrb(gzA, xmmB, 0); - e->vpextrb(anyptr_gpA, xmmB, 0); - e->vpextrd(gzA, xmmB, 0); - e->vpextrd(anyptr_gpA, xmmB, 0); - if (isX64) e->vpextrq(gzA, xmmB, 0); - if (isX64) e->vpextrq(anyptr_gpA, xmmB, 0); - e->vpextrw(gzA, xmmB, 0); - e->vpextrw(anyptr_gpA, xmmB, 0); - e->vphaddd(xmmA, xmmB, xmmC); - e->vphaddd(xmmA, xmmB, anyptr_gpC); - e->vphaddsw(xmmA, xmmB, xmmC); - e->vphaddsw(xmmA, xmmB, anyptr_gpC); - e->vphaddw(xmmA, xmmB, xmmC); - e->vphaddw(xmmA, xmmB, anyptr_gpC); - e->vphminposuw(xmmA, xmmB); - e->vphminposuw(xmmA, anyptr_gpB); - e->vphsubd(xmmA, xmmB, xmmC); - e->vphsubd(xmmA, xmmB, anyptr_gpC); - e->vphsubsw(xmmA, xmmB, xmmC); - e->vphsubsw(xmmA, xmmB, anyptr_gpC); - e->vphsubw(xmmA, xmmB, xmmC); - e->vphsubw(xmmA, xmmB, anyptr_gpC); - e->vpinsrb(xmmA, xmmB, gzC, 0); - e->vpinsrb(xmmA, xmmB, anyptr_gpC, 0); - e->vpinsrd(xmmA, xmmB, gzC, 0); - e->vpinsrd(xmmA, xmmB, anyptr_gpC, 0); - e->vpinsrw(xmmA, xmmB, gzC, 0); - e->vpinsrw(xmmA, xmmB, anyptr_gpC, 0); - e->vpmaddubsw(xmmA, xmmB, xmmC); - e->vpmaddubsw(xmmA, xmmB, anyptr_gpC); - e->vpmaddwd(xmmA, xmmB, xmmC); - e->vpmaddwd(xmmA, xmmB, anyptr_gpC); - e->vpmaxsb(xmmA, xmmB, xmmC); - e->vpmaxsb(xmmA, xmmB, anyptr_gpC); - e->vpmaxsd(xmmA, xmmB, xmmC); - e->vpmaxsd(xmmA, xmmB, anyptr_gpC); - e->vpmaxsw(xmmA, xmmB, xmmC); - e->vpmaxsw(xmmA, xmmB, anyptr_gpC); - e->vpmaxub(xmmA, xmmB, xmmC); - e->vpmaxub(xmmA, xmmB, anyptr_gpC); - e->vpmaxud(xmmA, xmmB, xmmC); - e->vpmaxud(xmmA, xmmB, anyptr_gpC); - e->vpmaxuw(xmmA, xmmB, xmmC); - e->vpmaxuw(xmmA, xmmB, anyptr_gpC); - e->vpminsb(xmmA, xmmB, xmmC); - e->vpminsb(xmmA, xmmB, anyptr_gpC); - e->vpminsd(xmmA, xmmB, xmmC); - e->vpminsd(xmmA, xmmB, anyptr_gpC); - e->vpminsw(xmmA, xmmB, xmmC); - e->vpminsw(xmmA, xmmB, anyptr_gpC); - e->vpminub(xmmA, xmmB, xmmC); - e->vpminub(xmmA, xmmB, anyptr_gpC); - e->vpminud(xmmA, xmmB, xmmC); - e->vpminud(xmmA, xmmB, anyptr_gpC); - e->vpminuw(xmmA, xmmB, xmmC); - e->vpminuw(xmmA, xmmB, anyptr_gpC); - e->vpmovmskb(gzA, xmmB); - e->vpmovsxbd(xmmA, xmmB); - e->vpmovsxbd(xmmA, anyptr_gpB); - e->vpmovsxbq(xmmA, xmmB); - e->vpmovsxbq(xmmA, anyptr_gpB); - e->vpmovsxbw(xmmA, xmmB); - e->vpmovsxbw(xmmA, anyptr_gpB); - e->vpmovsxdq(xmmA, xmmB); - e->vpmovsxdq(xmmA, anyptr_gpB); - e->vpmovsxwd(xmmA, xmmB); - e->vpmovsxwd(xmmA, anyptr_gpB); - e->vpmovsxwq(xmmA, xmmB); - e->vpmovsxwq(xmmA, anyptr_gpB); - e->vpmovzxbd(xmmA, xmmB); - e->vpmovzxbd(xmmA, anyptr_gpB); - e->vpmovzxbq(xmmA, xmmB); - e->vpmovzxbq(xmmA, anyptr_gpB); - e->vpmovzxbw(xmmA, xmmB); - e->vpmovzxbw(xmmA, anyptr_gpB); - e->vpmovzxdq(xmmA, xmmB); - e->vpmovzxdq(xmmA, anyptr_gpB); - e->vpmovzxwd(xmmA, xmmB); - e->vpmovzxwd(xmmA, anyptr_gpB); - e->vpmovzxwq(xmmA, xmmB); - e->vpmovzxwq(xmmA, anyptr_gpB); - e->vpmuldq(xmmA, xmmB, xmmC); - e->vpmuldq(xmmA, xmmB, anyptr_gpC); - e->vpmulhrsw(xmmA, xmmB, xmmC); - e->vpmulhrsw(xmmA, xmmB, anyptr_gpC); - e->vpmulhuw(xmmA, xmmB, xmmC); - e->vpmulhuw(xmmA, xmmB, anyptr_gpC); - e->vpmulhw(xmmA, xmmB, xmmC); - e->vpmulhw(xmmA, xmmB, anyptr_gpC); - e->vpmulld(xmmA, xmmB, xmmC); - e->vpmulld(xmmA, xmmB, anyptr_gpC); - e->vpmullw(xmmA, xmmB, xmmC); - e->vpmullw(xmmA, xmmB, anyptr_gpC); - e->vpmuludq(xmmA, xmmB, xmmC); - e->vpmuludq(xmmA, xmmB, anyptr_gpC); - e->vpor(xmmA, xmmB, xmmC); - e->vpor(xmmA, xmmB, anyptr_gpC); - e->vpsadbw(xmmA, xmmB, xmmC); - e->vpsadbw(xmmA, xmmB, anyptr_gpC); - e->vpshufb(xmmA, xmmB, xmmC); - e->vpshufb(xmmA, xmmB, anyptr_gpC); - e->vpshufd(xmmA, xmmB, 0); - e->vpshufd(xmmA, anyptr_gpB, 0); - e->vpshufhw(xmmA, xmmB, 0); - e->vpshufhw(xmmA, anyptr_gpB, 0); - e->vpshuflw(xmmA, xmmB, 0); - e->vpshuflw(xmmA, anyptr_gpB, 0); - e->vpsignb(xmmA, xmmB, xmmC); - e->vpsignb(xmmA, xmmB, anyptr_gpC); - e->vpsignd(xmmA, xmmB, xmmC); - e->vpsignd(xmmA, xmmB, anyptr_gpC); - e->vpsignw(xmmA, xmmB, xmmC); - e->vpsignw(xmmA, xmmB, anyptr_gpC); - e->vpslld(xmmA, xmmB, xmmC); - e->vpslld(xmmA, xmmB, anyptr_gpC); - e->vpslld(xmmA, xmmB, 0); - e->vpslldq(xmmA, xmmB, 0); - e->vpsllq(xmmA, xmmB, xmmC); - e->vpsllq(xmmA, xmmB, anyptr_gpC); - e->vpsllq(xmmA, xmmB, 0); - e->vpsllw(xmmA, xmmB, xmmC); - e->vpsllw(xmmA, xmmB, anyptr_gpC); - e->vpsllw(xmmA, xmmB, 0); - e->vpsrad(xmmA, xmmB, xmmC); - e->vpsrad(xmmA, xmmB, anyptr_gpC); - e->vpsrad(xmmA, xmmB, 0); - e->vpsraw(xmmA, xmmB, xmmC); - e->vpsraw(xmmA, xmmB, anyptr_gpC); - e->vpsraw(xmmA, xmmB, 0); - e->vpsrld(xmmA, xmmB, xmmC); - e->vpsrld(xmmA, xmmB, anyptr_gpC); - e->vpsrld(xmmA, xmmB, 0); - e->vpsrldq(xmmA, xmmB, 0); - e->vpsrlq(xmmA, xmmB, xmmC); - e->vpsrlq(xmmA, xmmB, anyptr_gpC); - e->vpsrlq(xmmA, xmmB, 0); - e->vpsrlw(xmmA, xmmB, xmmC); - e->vpsrlw(xmmA, xmmB, anyptr_gpC); - e->vpsrlw(xmmA, xmmB, 0); - e->vpsubb(xmmA, xmmB, xmmC); - e->vpsubb(xmmA, xmmB, anyptr_gpC); - e->vpsubd(xmmA, xmmB, xmmC); - e->vpsubd(xmmA, xmmB, anyptr_gpC); - e->vpsubq(xmmA, xmmB, xmmC); - e->vpsubq(xmmA, xmmB, anyptr_gpC); - e->vpsubw(xmmA, xmmB, xmmC); - e->vpsubw(xmmA, xmmB, anyptr_gpC); - e->vpsubsb(xmmA, xmmB, xmmC); - e->vpsubsb(xmmA, xmmB, anyptr_gpC); - e->vpsubsw(xmmA, xmmB, xmmC); - e->vpsubsw(xmmA, xmmB, anyptr_gpC); - e->vpsubusb(xmmA, xmmB, xmmC); - e->vpsubusb(xmmA, xmmB, anyptr_gpC); - e->vpsubusw(xmmA, xmmB, xmmC); - e->vpsubusw(xmmA, xmmB, anyptr_gpC); - e->vptest(xmmA, xmmB); - e->vptest(xmmA, anyptr_gpB); - e->vptest(ymmA, ymmB); - e->vptest(ymmA, anyptr_gpB); - e->vpunpckhbw(xmmA, xmmB, xmmC); - e->vpunpckhbw(xmmA, xmmB, anyptr_gpC); - e->vpunpckhdq(xmmA, xmmB, xmmC); - e->vpunpckhdq(xmmA, xmmB, anyptr_gpC); - e->vpunpckhqdq(xmmA, xmmB, xmmC); - e->vpunpckhqdq(xmmA, xmmB, anyptr_gpC); - e->vpunpckhwd(xmmA, xmmB, xmmC); - e->vpunpckhwd(xmmA, xmmB, anyptr_gpC); - e->vpunpcklbw(xmmA, xmmB, xmmC); - e->vpunpcklbw(xmmA, xmmB, anyptr_gpC); - e->vpunpckldq(xmmA, xmmB, xmmC); - e->vpunpckldq(xmmA, xmmB, anyptr_gpC); - e->vpunpcklqdq(xmmA, xmmB, xmmC); - e->vpunpcklqdq(xmmA, xmmB, anyptr_gpC); - e->vpunpcklwd(xmmA, xmmB, xmmC); - e->vpunpcklwd(xmmA, xmmB, anyptr_gpC); - e->vpxor(xmmA, xmmB, xmmC); - e->vpxor(xmmA, xmmB, anyptr_gpC); - e->vrcpps(xmmA, xmmB); - e->vrcpps(xmmA, anyptr_gpB); - e->vrcpps(ymmA, ymmB); - e->vrcpps(ymmA, anyptr_gpB); - e->vrcpss(xmmA, xmmB, xmmC); - e->vrcpss(xmmA, xmmB, anyptr_gpC); - e->vrsqrtps(xmmA, xmmB); - e->vrsqrtps(xmmA, anyptr_gpB); - e->vrsqrtps(ymmA, ymmB); - e->vrsqrtps(ymmA, anyptr_gpB); - e->vrsqrtss(xmmA, xmmB, xmmC); - e->vrsqrtss(xmmA, xmmB, anyptr_gpC); - e->vroundpd(xmmA, xmmB, 0); - e->vroundpd(xmmA, anyptr_gpB, 0); - e->vroundpd(ymmA, ymmB, 0); - e->vroundpd(ymmA, anyptr_gpB, 0); - e->vroundps(xmmA, xmmB, 0); - e->vroundps(xmmA, anyptr_gpB, 0); - e->vroundps(ymmA, ymmB, 0); - e->vroundps(ymmA, anyptr_gpB, 0); - e->vroundsd(xmmA, xmmB, xmmC, 0); - e->vroundsd(xmmA, xmmB, anyptr_gpC, 0); - e->vroundss(xmmA, xmmB, xmmC, 0); - e->vroundss(xmmA, xmmB, anyptr_gpC, 0); - e->vshufpd(xmmA, xmmB, xmmC, 0); - e->vshufpd(xmmA, xmmB, anyptr_gpC, 0); - e->vshufpd(ymmA, ymmB, ymmC, 0); - e->vshufpd(ymmA, ymmB, anyptr_gpC, 0); - e->vshufps(xmmA, xmmB, xmmC, 0); - e->vshufps(xmmA, xmmB, anyptr_gpC, 0); - e->vshufps(ymmA, ymmB, ymmC, 0); - e->vshufps(ymmA, ymmB, anyptr_gpC, 0); - e->vsqrtpd(xmmA, xmmB); - e->vsqrtpd(xmmA, anyptr_gpB); - e->vsqrtpd(ymmA, ymmB); - e->vsqrtpd(ymmA, anyptr_gpB); - e->vsqrtps(xmmA, xmmB); - e->vsqrtps(xmmA, anyptr_gpB); - e->vsqrtps(ymmA, ymmB); - e->vsqrtps(ymmA, anyptr_gpB); - e->vsqrtsd(xmmA, xmmB, xmmC); - e->vsqrtsd(xmmA, xmmB, anyptr_gpC); - e->vsqrtss(xmmA, xmmB, xmmC); - e->vsqrtss(xmmA, xmmB, anyptr_gpC); - e->vstmxcsr(anyptr_gpA); - e->vsubpd(xmmA, xmmB, xmmC); - e->vsubpd(xmmA, xmmB, anyptr_gpC); - e->vsubpd(ymmA, ymmB, ymmC); - e->vsubpd(ymmA, ymmB, anyptr_gpC); - e->vsubps(xmmA, xmmB, xmmC); - e->vsubps(xmmA, xmmB, anyptr_gpC); - e->vsubps(ymmA, ymmB, ymmC); - e->vsubps(ymmA, ymmB, anyptr_gpC); - e->vsubsd(xmmA, xmmB, xmmC); - e->vsubsd(xmmA, xmmB, anyptr_gpC); - e->vsubss(xmmA, xmmB, xmmC); - e->vsubss(xmmA, xmmB, anyptr_gpC); - e->vtestps(xmmA, xmmB); - e->vtestps(xmmA, anyptr_gpB); - e->vtestps(ymmA, ymmB); - e->vtestps(ymmA, anyptr_gpB); - e->vtestpd(xmmA, xmmB); - e->vtestpd(xmmA, anyptr_gpB); - e->vtestpd(ymmA, ymmB); - e->vtestpd(ymmA, anyptr_gpB); - e->vucomisd(xmmA, xmmB); - e->vucomisd(xmmA, anyptr_gpB); - e->vucomiss(xmmA, xmmB); - e->vucomiss(xmmA, anyptr_gpB); - e->vunpckhpd(xmmA, xmmB, xmmC); - e->vunpckhpd(xmmA, xmmB, anyptr_gpC); - e->vunpckhpd(ymmA, ymmB, ymmC); - e->vunpckhpd(ymmA, ymmB, anyptr_gpC); - e->vunpckhps(xmmA, xmmB, xmmC); - e->vunpckhps(xmmA, xmmB, anyptr_gpC); - e->vunpckhps(ymmA, ymmB, ymmC); - e->vunpckhps(ymmA, ymmB, anyptr_gpC); - e->vunpcklpd(xmmA, xmmB, xmmC); - e->vunpcklpd(xmmA, xmmB, anyptr_gpC); - e->vunpcklpd(ymmA, ymmB, ymmC); - e->vunpcklpd(ymmA, ymmB, anyptr_gpC); - e->vunpcklps(xmmA, xmmB, xmmC); - e->vunpcklps(xmmA, xmmB, anyptr_gpC); - e->vunpcklps(ymmA, ymmB, ymmC); - e->vunpcklps(ymmA, ymmB, anyptr_gpC); - e->vxorpd(xmmA, xmmB, xmmC); - e->vxorpd(xmmA, xmmB, anyptr_gpC); - e->vxorpd(ymmA, ymmB, ymmC); - e->vxorpd(ymmA, ymmB, anyptr_gpC); - e->vxorps(xmmA, xmmB, xmmC); - e->vxorps(xmmA, xmmB, anyptr_gpC); - e->vxorps(ymmA, ymmB, ymmC); - e->vxorps(ymmA, ymmB, anyptr_gpC); - e->vzeroall(); - e->vex3().vzeroall(); - e->vzeroupper(); - e->vex3().vzeroupper(); - - // AVX+AESNI. - e->nop(); - - e->vaesdec(xmmA, xmmB, xmmC); - e->vaesdec(xmmA, xmmB, anyptr_gpC); - e->vaesdeclast(xmmA, xmmB, xmmC); - e->vaesdeclast(xmmA, xmmB, anyptr_gpC); - e->vaesenc(xmmA, xmmB, xmmC); - e->vaesenc(xmmA, xmmB, anyptr_gpC); - e->vaesenclast(xmmA, xmmB, xmmC); - e->vaesenclast(xmmA, xmmB, anyptr_gpC); - e->vaesimc(xmmA, xmmB); - e->vaesimc(xmmA, anyptr_gpB); - e->vaeskeygenassist(xmmA, xmmB, 0); - e->vaeskeygenassist(xmmA, anyptr_gpB, 0); - - // AVX+PCLMULQDQ. - e->nop(); - - e->vpclmulqdq(xmmA, xmmB, xmmC, 0); - e->vpclmulqdq(xmmA, xmmB, anyptr_gpC, 0); - - // AVX2. - e->nop(); - - e->vbroadcasti128(ymmA, anyptr_gpB); - e->vbroadcastsd(ymmA, xmmB); - e->vbroadcastss(xmmA, xmmB); - e->vbroadcastss(ymmA, xmmB); - e->vextracti128(xmmA, ymmB, 0); - e->vextracti128(anyptr_gpA, ymmB, 0); - e->vgatherdpd(xmmA, vx_ptr, xmmC); - e->vgatherdpd(ymmA, vx_ptr, ymmC); - e->vgatherdps(xmmA, vx_ptr, xmmC); - e->vgatherdps(ymmA, vy_ptr, ymmC); - e->vgatherqpd(xmmA, vx_ptr, xmmC); - e->vgatherqpd(ymmA, vy_ptr, ymmC); - e->vgatherqps(xmmA, vx_ptr, xmmC); - e->vgatherqps(xmmA, vy_ptr, xmmC); - e->vinserti128(ymmA, ymmB, xmmC, 0); - e->vinserti128(ymmA, ymmB, anyptr_gpC, 0); - e->vmovntdqa(ymmA, anyptr_gpB); - e->vmpsadbw(ymmA, ymmB, ymmC, 0); - e->vmpsadbw(ymmA, ymmB, anyptr_gpC, 0); - e->vpabsb(ymmA, ymmB); - e->vpabsb(ymmA, anyptr_gpB); - e->vpabsd(ymmA, ymmB); - e->vpabsd(ymmA, anyptr_gpB); - e->vpabsw(ymmA, ymmB); - e->vpabsw(ymmA, anyptr_gpB); - e->vpackssdw(ymmA, ymmB, ymmC); - e->vpackssdw(ymmA, ymmB, anyptr_gpC); - e->vpacksswb(ymmA, ymmB, ymmC); - e->vpacksswb(ymmA, ymmB, anyptr_gpC); - e->vpackusdw(ymmA, ymmB, ymmC); - e->vpackusdw(ymmA, ymmB, anyptr_gpC); - e->vpackuswb(ymmA, ymmB, ymmC); - e->vpackuswb(ymmA, ymmB, anyptr_gpC); - e->vpaddb(ymmA, ymmB, ymmC); - e->vpaddb(ymmA, ymmB, anyptr_gpC); - e->vpaddd(ymmA, ymmB, ymmC); - e->vpaddd(ymmA, ymmB, anyptr_gpC); - e->vpaddq(ymmA, ymmB, ymmC); - e->vpaddq(ymmA, ymmB, anyptr_gpC); - e->vpaddw(ymmA, ymmB, ymmC); - e->vpaddw(ymmA, ymmB, anyptr_gpC); - e->vpaddsb(ymmA, ymmB, ymmC); - e->vpaddsb(ymmA, ymmB, anyptr_gpC); - e->vpaddsw(ymmA, ymmB, ymmC); - e->vpaddsw(ymmA, ymmB, anyptr_gpC); - e->vpaddusb(ymmA, ymmB, ymmC); - e->vpaddusb(ymmA, ymmB, anyptr_gpC); - e->vpaddusw(ymmA, ymmB, ymmC); - e->vpaddusw(ymmA, ymmB, anyptr_gpC); - e->vpalignr(ymmA, ymmB, ymmC, 0); - e->vpalignr(ymmA, ymmB, anyptr_gpC, 0); - e->vpand(ymmA, ymmB, ymmC); - e->vpand(ymmA, ymmB, anyptr_gpC); - e->vpandn(ymmA, ymmB, ymmC); - e->vpandn(ymmA, ymmB, anyptr_gpC); - e->vpavgb(ymmA, ymmB, ymmC); - e->vpavgb(ymmA, ymmB, anyptr_gpC); - e->vpavgw(ymmA, ymmB, ymmC); - e->vpavgw(ymmA, ymmB, anyptr_gpC); - e->vpblendd(xmmA, xmmB, xmmC, 0); - e->vpblendd(xmmA, xmmB, anyptr_gpC, 0); - e->vpblendd(ymmA, ymmB, ymmC, 0); - e->vpblendd(ymmA, ymmB, anyptr_gpC, 0); - e->vpblendvb(ymmA, ymmB, ymmC, ymmD); - e->vpblendvb(ymmA, ymmB, anyptr_gpC, ymmD); - e->vpblendw(ymmA, ymmB, ymmC, 0); - e->vpblendw(ymmA, ymmB, anyptr_gpC, 0); - e->vpbroadcastb(xmmA, xmmB); - e->vpbroadcastb(xmmA, anyptr_gpB); - e->vpbroadcastb(ymmA, xmmB); - e->vpbroadcastb(ymmA, anyptr_gpB); - e->vpbroadcastd(xmmA, xmmB); - e->vpbroadcastd(xmmA, anyptr_gpB); - e->vpbroadcastd(ymmA, xmmB); - e->vpbroadcastd(ymmA, anyptr_gpB); - e->vpbroadcastq(xmmA, xmmB); - e->vpbroadcastq(xmmA, anyptr_gpB); - e->vpbroadcastq(ymmA, xmmB); - e->vpbroadcastq(ymmA, anyptr_gpB); - e->vpbroadcastw(xmmA, xmmB); - e->vpbroadcastw(xmmA, anyptr_gpB); - e->vpbroadcastw(ymmA, xmmB); - e->vpbroadcastw(ymmA, anyptr_gpB); - e->vpcmpeqb(ymmA, ymmB, ymmC); - e->vpcmpeqb(ymmA, ymmB, anyptr_gpC); - e->vpcmpeqd(ymmA, ymmB, ymmC); - e->vpcmpeqd(ymmA, ymmB, anyptr_gpC); - e->vpcmpeqq(ymmA, ymmB, ymmC); - e->vpcmpeqq(ymmA, ymmB, anyptr_gpC); - e->vpcmpeqw(ymmA, ymmB, ymmC); - e->vpcmpeqw(ymmA, ymmB, anyptr_gpC); - e->vpcmpgtb(ymmA, ymmB, ymmC); - e->vpcmpgtb(ymmA, ymmB, anyptr_gpC); - e->vpcmpgtd(ymmA, ymmB, ymmC); - e->vpcmpgtd(ymmA, ymmB, anyptr_gpC); - e->vpcmpgtq(ymmA, ymmB, ymmC); - e->vpcmpgtq(ymmA, ymmB, anyptr_gpC); - e->vpcmpgtw(ymmA, ymmB, ymmC); - e->vpcmpgtw(ymmA, ymmB, anyptr_gpC); - e->vperm2i128(ymmA, ymmB, ymmC, 0); - e->vperm2i128(ymmA, ymmB, anyptr_gpC, 0); - e->vpermd(ymmA, ymmB, ymmC); - e->vpermd(ymmA, ymmB, anyptr_gpC); - e->vpermps(ymmA, ymmB, ymmC); - e->vpermps(ymmA, ymmB, anyptr_gpC); - e->vpermpd(ymmA, ymmB, 0); - e->vpermpd(ymmA, anyptr_gpB, 0); - e->vpermq(ymmA, ymmB, 0); - e->vpermq(ymmA, anyptr_gpB, 0); - e->vpgatherdd(xmmA, vx_ptr, xmmC); - e->vpgatherdd(ymmA, vy_ptr, ymmC); - e->vpgatherdq(xmmA, vx_ptr, xmmC); - e->vpgatherdq(ymmA, vx_ptr, ymmC); - e->vpgatherqd(xmmA, vx_ptr, xmmC); - e->vpgatherqd(xmmA, vy_ptr, xmmC); - e->vpgatherqq(xmmA, vx_ptr, xmmC); - e->vpgatherqq(ymmA, vy_ptr, ymmC); - e->vpmovmskb(gzA, ymmB); - e->vpmovsxbd(ymmA, anyptr_gpB); - e->vpmovsxbd(ymmA, xmmB); - e->vpmovsxbq(ymmA, anyptr_gpB); - e->vpmovsxbq(ymmA, xmmB); - e->vpmovsxbw(ymmA, anyptr_gpB); - e->vpmovsxbw(ymmA, xmmB); - e->vpmovsxdq(ymmA, anyptr_gpB); - e->vpmovsxdq(ymmA, xmmB); - e->vpmovsxwd(ymmA, anyptr_gpB); - e->vpmovsxwd(ymmA, xmmB); - e->vpmovsxwq(ymmA, anyptr_gpB); - e->vpmovsxwq(ymmA, xmmB); - e->vpmovzxbd(ymmA, anyptr_gpB); - e->vpmovzxbd(ymmA, xmmB); - e->vpmovzxbq(ymmA, anyptr_gpB); - e->vpmovzxbq(ymmA, xmmB); - e->vpmovzxbw(ymmA, anyptr_gpB); - e->vpmovzxbw(ymmA, xmmB); - e->vpmovzxdq(ymmA, anyptr_gpB); - e->vpmovzxdq(ymmA, xmmB); - e->vpmovzxwd(ymmA, anyptr_gpB); - e->vpmovzxwd(ymmA, xmmB); - e->vpmovzxwq(ymmA, anyptr_gpB); - e->vpmovzxwq(ymmA, xmmB); - e->vpshufd(ymmA, anyptr_gpB, 0); - e->vpshufd(ymmA, ymmB, 0); - e->vpshufhw(ymmA, anyptr_gpB, 0); - e->vpshufhw(ymmA, ymmB, 0); - e->vpshuflw(ymmA, anyptr_gpB, 0); - e->vpshuflw(ymmA, ymmB, 0); - e->vpslld(ymmA, ymmB, 0); - e->vpslldq(ymmA, ymmB, 0); - e->vpsllq(ymmA, ymmB, 0); - e->vpsllw(ymmA, ymmB, 0); - e->vpsrad(ymmA, ymmB, 0); - e->vpsraw(ymmA, ymmB, 0); - e->vpsrld(ymmA, ymmB, 0); - e->vpsrldq(ymmA, ymmB, 0); - e->vpsrlq(ymmA, ymmB, 0); - e->vpsrlw(ymmA, ymmB, 0); - e->vphaddd(ymmA, ymmB, anyptr_gpC); - e->vphaddd(ymmA, ymmB, ymmC); - e->vphaddsw(ymmA, ymmB, anyptr_gpC); - e->vphaddsw(ymmA, ymmB, ymmC); - e->vphaddw(ymmA, ymmB, anyptr_gpC); - e->vphaddw(ymmA, ymmB, ymmC); - e->vphsubd(ymmA, ymmB, anyptr_gpC); - e->vphsubd(ymmA, ymmB, ymmC); - e->vphsubsw(ymmA, ymmB, anyptr_gpC); - e->vphsubsw(ymmA, ymmB, ymmC); - e->vphsubw(ymmA, ymmB, anyptr_gpC); - e->vphsubw(ymmA, ymmB, ymmC); - e->vpmaddubsw(ymmA, ymmB, anyptr_gpC); - e->vpmaddubsw(ymmA, ymmB, ymmC); - e->vpmaddwd(ymmA, ymmB, anyptr_gpC); - e->vpmaddwd(ymmA, ymmB, ymmC); - e->vpmaskmovd(anyptr_gpA, xmmB, xmmC); - e->vpmaskmovd(anyptr_gpA, ymmB, ymmC); - e->vpmaskmovd(xmmA, xmmB, anyptr_gpC); - e->vpmaskmovd(ymmA, ymmB, anyptr_gpC); - e->vpmaskmovq(anyptr_gpA, xmmB, xmmC); - e->vpmaskmovq(anyptr_gpA, ymmB, ymmC); - e->vpmaskmovq(xmmA, xmmB, anyptr_gpC); - e->vpmaskmovq(ymmA, ymmB, anyptr_gpC); - e->vpmaxsb(ymmA, ymmB, anyptr_gpC); - e->vpmaxsb(ymmA, ymmB, ymmC); - e->vpmaxsd(ymmA, ymmB, anyptr_gpC); - e->vpmaxsd(ymmA, ymmB, ymmC); - e->vpmaxsw(ymmA, ymmB, anyptr_gpC); - e->vpmaxsw(ymmA, ymmB, ymmC); - e->vpmaxub(ymmA, ymmB, anyptr_gpC); - e->vpmaxub(ymmA, ymmB, ymmC); - e->vpmaxud(ymmA, ymmB, anyptr_gpC); - e->vpmaxud(ymmA, ymmB, ymmC); - e->vpmaxuw(ymmA, ymmB, anyptr_gpC); - e->vpmaxuw(ymmA, ymmB, ymmC); - e->vpminsb(ymmA, ymmB, anyptr_gpC); - e->vpminsb(ymmA, ymmB, ymmC); - e->vpminsd(ymmA, ymmB, anyptr_gpC); - e->vpminsd(ymmA, ymmB, ymmC); - e->vpminsw(ymmA, ymmB, anyptr_gpC); - e->vpminsw(ymmA, ymmB, ymmC); - e->vpminub(ymmA, ymmB, anyptr_gpC); - e->vpminub(ymmA, ymmB, ymmC); - e->vpminud(ymmA, ymmB, anyptr_gpC); - e->vpminud(ymmA, ymmB, ymmC); - e->vpminuw(ymmA, ymmB, anyptr_gpC); - e->vpminuw(ymmA, ymmB, ymmC); - e->vpmuldq(ymmA, ymmB, anyptr_gpC); - e->vpmuldq(ymmA, ymmB, ymmC); - e->vpmulhrsw(ymmA, ymmB, anyptr_gpC); - e->vpmulhrsw(ymmA, ymmB, ymmC); - e->vpmulhuw(ymmA, ymmB, anyptr_gpC); - e->vpmulhuw(ymmA, ymmB, ymmC); - e->vpmulhw(ymmA, ymmB, anyptr_gpC); - e->vpmulhw(ymmA, ymmB, ymmC); - e->vpmulld(ymmA, ymmB, anyptr_gpC); - e->vpmulld(ymmA, ymmB, ymmC); - e->vpmullw(ymmA, ymmB, anyptr_gpC); - e->vpmullw(ymmA, ymmB, ymmC); - e->vpmuludq(ymmA, ymmB, anyptr_gpC); - e->vpmuludq(ymmA, ymmB, ymmC); - e->vpor(ymmA, ymmB, anyptr_gpC); - e->vpor(ymmA, ymmB, ymmC); - e->vpsadbw(ymmA, ymmB, anyptr_gpC); - e->vpsadbw(ymmA, ymmB, ymmC); - e->vpshufb(ymmA, ymmB, anyptr_gpC); - e->vpshufb(ymmA, ymmB, ymmC); - e->vpsignb(ymmA, ymmB, anyptr_gpC); - e->vpsignb(ymmA, ymmB, ymmC); - e->vpsignd(ymmA, ymmB, anyptr_gpC); - e->vpsignd(ymmA, ymmB, ymmC); - e->vpsignw(ymmA, ymmB, anyptr_gpC); - e->vpsignw(ymmA, ymmB, ymmC); - e->vpslld(ymmA, ymmB, anyptr_gpC); - e->vpslld(ymmA, ymmB, xmmC); - e->vpsllq(ymmA, ymmB, anyptr_gpC); - e->vpsllq(ymmA, ymmB, xmmC); - e->vpsllvd(xmmA, xmmB, anyptr_gpC); - e->vpsllvd(xmmA, xmmB, xmmC); - e->vpsllvd(ymmA, ymmB, anyptr_gpC); - e->vpsllvd(ymmA, ymmB, ymmC); - e->vpsllvq(xmmA, xmmB, anyptr_gpC); - e->vpsllvq(xmmA, xmmB, xmmC); - e->vpsllvq(ymmA, ymmB, anyptr_gpC); - e->vpsllvq(ymmA, ymmB, ymmC); - e->vpsllw(ymmA, ymmB, anyptr_gpC); - e->vpsllw(ymmA, ymmB, xmmC); - e->vpsrad(ymmA, ymmB, anyptr_gpC); - e->vpsrad(ymmA, ymmB, xmmC); - e->vpsravd(xmmA, xmmB, anyptr_gpC); - e->vpsravd(xmmA, xmmB, xmmC); - e->vpsravd(ymmA, ymmB, anyptr_gpC); - e->vpsravd(ymmA, ymmB, ymmC); - e->vpsraw(ymmA, ymmB, anyptr_gpC); - e->vpsraw(ymmA, ymmB, xmmC); - e->vpsrld(ymmA, ymmB, anyptr_gpC); - e->vpsrld(ymmA, ymmB, xmmC); - e->vpsrlq(ymmA, ymmB, anyptr_gpC); - e->vpsrlq(ymmA, ymmB, xmmC); - e->vpsrlvd(xmmA, xmmB, anyptr_gpC); - e->vpsrlvd(xmmA, xmmB, xmmC); - e->vpsrlvd(ymmA, ymmB, anyptr_gpC); - e->vpsrlvd(ymmA, ymmB, ymmC); - e->vpsrlvq(xmmA, xmmB, anyptr_gpC); - e->vpsrlvq(xmmA, xmmB, xmmC); - e->vpsrlvq(ymmA, ymmB, anyptr_gpC); - e->vpsrlvq(ymmA, ymmB, ymmC); - e->vpsrlw(ymmA, ymmB, anyptr_gpC); - e->vpsrlw(ymmA, ymmB, xmmC); - e->vpsubb(ymmA, ymmB, anyptr_gpC); - e->vpsubb(ymmA, ymmB, ymmC); - e->vpsubd(ymmA, ymmB, anyptr_gpC); - e->vpsubd(ymmA, ymmB, ymmC); - e->vpsubq(ymmA, ymmB, anyptr_gpC); - e->vpsubq(ymmA, ymmB, ymmC); - e->vpsubsb(ymmA, ymmB, anyptr_gpC); - e->vpsubsb(ymmA, ymmB, ymmC); - e->vpsubsw(ymmA, ymmB, anyptr_gpC); - e->vpsubsw(ymmA, ymmB, ymmC); - e->vpsubusb(ymmA, ymmB, anyptr_gpC); - e->vpsubusb(ymmA, ymmB, ymmC); - e->vpsubusw(ymmA, ymmB, anyptr_gpC); - e->vpsubusw(ymmA, ymmB, ymmC); - e->vpsubw(ymmA, ymmB, anyptr_gpC); - e->vpsubw(ymmA, ymmB, ymmC); - e->vpunpckhbw(ymmA, ymmB, anyptr_gpC); - e->vpunpckhbw(ymmA, ymmB, ymmC); - e->vpunpckhdq(ymmA, ymmB, anyptr_gpC); - e->vpunpckhdq(ymmA, ymmB, ymmC); - e->vpunpckhqdq(ymmA, ymmB, anyptr_gpC); - e->vpunpckhqdq(ymmA, ymmB, ymmC); - e->vpunpckhwd(ymmA, ymmB, anyptr_gpC); - e->vpunpckhwd(ymmA, ymmB, ymmC); - e->vpunpcklbw(ymmA, ymmB, anyptr_gpC); - e->vpunpcklbw(ymmA, ymmB, ymmC); - e->vpunpckldq(ymmA, ymmB, anyptr_gpC); - e->vpunpckldq(ymmA, ymmB, ymmC); - e->vpunpcklqdq(ymmA, ymmB, anyptr_gpC); - e->vpunpcklqdq(ymmA, ymmB, ymmC); - e->vpunpcklwd(ymmA, ymmB, anyptr_gpC); - e->vpunpcklwd(ymmA, ymmB, ymmC); - e->vpxor(ymmA, ymmB, anyptr_gpC); - e->vpxor(ymmA, ymmB, ymmC); - - // FMA. - e->nop(); - - e->vfmadd132pd(xmmA, xmmB, anyptr_gpC); - e->vfmadd132pd(xmmA, xmmB, xmmC); - e->vfmadd132pd(ymmA, ymmB, anyptr_gpC); - e->vfmadd132pd(ymmA, ymmB, ymmC); - e->vfmadd132ps(xmmA, xmmB, anyptr_gpC); - e->vfmadd132ps(xmmA, xmmB, xmmC); - e->vfmadd132ps(ymmA, ymmB, anyptr_gpC); - e->vfmadd132ps(ymmA, ymmB, ymmC); - e->vfmadd132sd(xmmA, xmmB, anyptr_gpC); - e->vfmadd132sd(xmmA, xmmB, xmmC); - e->vfmadd132ss(xmmA, xmmB, anyptr_gpC); - e->vfmadd132ss(xmmA, xmmB, xmmC); - e->vfmadd213pd(xmmA, xmmB, anyptr_gpC); - e->vfmadd213pd(xmmA, xmmB, xmmC); - e->vfmadd213pd(ymmA, ymmB, anyptr_gpC); - e->vfmadd213pd(ymmA, ymmB, ymmC); - e->vfmadd213ps(xmmA, xmmB, anyptr_gpC); - e->vfmadd213ps(xmmA, xmmB, xmmC); - e->vfmadd213ps(ymmA, ymmB, anyptr_gpC); - e->vfmadd213ps(ymmA, ymmB, ymmC); - e->vfmadd213sd(xmmA, xmmB, anyptr_gpC); - e->vfmadd213sd(xmmA, xmmB, xmmC); - e->vfmadd213ss(xmmA, xmmB, anyptr_gpC); - e->vfmadd213ss(xmmA, xmmB, xmmC); - e->vfmadd231pd(xmmA, xmmB, anyptr_gpC); - e->vfmadd231pd(xmmA, xmmB, xmmC); - e->vfmadd231pd(ymmA, ymmB, anyptr_gpC); - e->vfmadd231pd(ymmA, ymmB, ymmC); - e->vfmadd231ps(xmmA, xmmB, anyptr_gpC); - e->vfmadd231ps(xmmA, xmmB, xmmC); - e->vfmadd231ps(ymmA, ymmB, anyptr_gpC); - e->vfmadd231ps(ymmA, ymmB, ymmC); - e->vfmadd231sd(xmmA, xmmB, anyptr_gpC); - e->vfmadd231sd(xmmA, xmmB, xmmC); - e->vfmadd231ss(xmmA, xmmB, anyptr_gpC); - e->vfmadd231ss(xmmA, xmmB, xmmC); - e->vfmaddsub132pd(xmmA, xmmB, anyptr_gpC); - e->vfmaddsub132pd(xmmA, xmmB, xmmC); - e->vfmaddsub132pd(ymmA, ymmB, anyptr_gpC); - e->vfmaddsub132pd(ymmA, ymmB, ymmC); - e->vfmaddsub132ps(xmmA, xmmB, anyptr_gpC); - e->vfmaddsub132ps(xmmA, xmmB, xmmC); - e->vfmaddsub132ps(ymmA, ymmB, anyptr_gpC); - e->vfmaddsub132ps(ymmA, ymmB, ymmC); - e->vfmaddsub213pd(xmmA, xmmB, anyptr_gpC); - e->vfmaddsub213pd(xmmA, xmmB, xmmC); - e->vfmaddsub213pd(ymmA, ymmB, anyptr_gpC); - e->vfmaddsub213pd(ymmA, ymmB, ymmC); - e->vfmaddsub213ps(xmmA, xmmB, anyptr_gpC); - e->vfmaddsub213ps(xmmA, xmmB, xmmC); - e->vfmaddsub213ps(ymmA, ymmB, anyptr_gpC); - e->vfmaddsub213ps(ymmA, ymmB, ymmC); - e->vfmaddsub231pd(xmmA, xmmB, anyptr_gpC); - e->vfmaddsub231pd(xmmA, xmmB, xmmC); - e->vfmaddsub231pd(ymmA, ymmB, anyptr_gpC); - e->vfmaddsub231pd(ymmA, ymmB, ymmC); - e->vfmaddsub231ps(xmmA, xmmB, anyptr_gpC); - e->vfmaddsub231ps(xmmA, xmmB, xmmC); - e->vfmaddsub231ps(ymmA, ymmB, anyptr_gpC); - e->vfmaddsub231ps(ymmA, ymmB, ymmC); - e->vfmsub132pd(xmmA, xmmB, anyptr_gpC); - e->vfmsub132pd(xmmA, xmmB, xmmC); - e->vfmsub132pd(ymmA, ymmB, anyptr_gpC); - e->vfmsub132pd(ymmA, ymmB, ymmC); - e->vfmsub132ps(xmmA, xmmB, anyptr_gpC); - e->vfmsub132ps(xmmA, xmmB, xmmC); - e->vfmsub132ps(ymmA, ymmB, anyptr_gpC); - e->vfmsub132ps(ymmA, ymmB, ymmC); - e->vfmsub132sd(xmmA, xmmB, anyptr_gpC); - e->vfmsub132sd(xmmA, xmmB, xmmC); - e->vfmsub132ss(xmmA, xmmB, anyptr_gpC); - e->vfmsub132ss(xmmA, xmmB, xmmC); - e->vfmsub213pd(xmmA, xmmB, anyptr_gpC); - e->vfmsub213pd(xmmA, xmmB, xmmC); - e->vfmsub213pd(ymmA, ymmB, anyptr_gpC); - e->vfmsub213pd(ymmA, ymmB, ymmC); - e->vfmsub213ps(xmmA, xmmB, anyptr_gpC); - e->vfmsub213ps(xmmA, xmmB, xmmC); - e->vfmsub213ps(ymmA, ymmB, anyptr_gpC); - e->vfmsub213ps(ymmA, ymmB, ymmC); - e->vfmsub213sd(xmmA, xmmB, anyptr_gpC); - e->vfmsub213sd(xmmA, xmmB, xmmC); - e->vfmsub213ss(xmmA, xmmB, anyptr_gpC); - e->vfmsub213ss(xmmA, xmmB, xmmC); - e->vfmsub231pd(xmmA, xmmB, anyptr_gpC); - e->vfmsub231pd(xmmA, xmmB, xmmC); - e->vfmsub231pd(ymmA, ymmB, anyptr_gpC); - e->vfmsub231pd(ymmA, ymmB, ymmC); - e->vfmsub231ps(xmmA, xmmB, anyptr_gpC); - e->vfmsub231ps(xmmA, xmmB, xmmC); - e->vfmsub231ps(ymmA, ymmB, anyptr_gpC); - e->vfmsub231ps(ymmA, ymmB, ymmC); - e->vfmsub231sd(xmmA, xmmB, anyptr_gpC); - e->vfmsub231sd(xmmA, xmmB, xmmC); - e->vfmsub231ss(xmmA, xmmB, anyptr_gpC); - e->vfmsub231ss(xmmA, xmmB, xmmC); - e->vfmsubadd132pd(xmmA, xmmB, anyptr_gpC); - e->vfmsubadd132pd(xmmA, xmmB, xmmC); - e->vfmsubadd132pd(ymmA, ymmB, anyptr_gpC); - e->vfmsubadd132pd(ymmA, ymmB, ymmC); - e->vfmsubadd132ps(xmmA, xmmB, anyptr_gpC); - e->vfmsubadd132ps(xmmA, xmmB, xmmC); - e->vfmsubadd132ps(ymmA, ymmB, anyptr_gpC); - e->vfmsubadd132ps(ymmA, ymmB, ymmC); - e->vfmsubadd213pd(xmmA, xmmB, anyptr_gpC); - e->vfmsubadd213pd(xmmA, xmmB, xmmC); - e->vfmsubadd213pd(ymmA, ymmB, anyptr_gpC); - e->vfmsubadd213pd(ymmA, ymmB, ymmC); - e->vfmsubadd213ps(xmmA, xmmB, anyptr_gpC); - e->vfmsubadd213ps(xmmA, xmmB, xmmC); - e->vfmsubadd213ps(ymmA, ymmB, anyptr_gpC); - e->vfmsubadd213ps(ymmA, ymmB, ymmC); - e->vfmsubadd231pd(xmmA, xmmB, anyptr_gpC); - e->vfmsubadd231pd(xmmA, xmmB, xmmC); - e->vfmsubadd231pd(ymmA, ymmB, anyptr_gpC); - e->vfmsubadd231pd(ymmA, ymmB, ymmC); - e->vfmsubadd231ps(xmmA, xmmB, anyptr_gpC); - e->vfmsubadd231ps(xmmA, xmmB, xmmC); - e->vfmsubadd231ps(ymmA, ymmB, anyptr_gpC); - e->vfmsubadd231ps(ymmA, ymmB, ymmC); - e->vfnmadd132pd(xmmA, xmmB, anyptr_gpC); - e->vfnmadd132pd(xmmA, xmmB, xmmC); - e->vfnmadd132pd(ymmA, ymmB, anyptr_gpC); - e->vfnmadd132pd(ymmA, ymmB, ymmC); - e->vfnmadd132ps(xmmA, xmmB, anyptr_gpC); - e->vfnmadd132ps(xmmA, xmmB, xmmC); - e->vfnmadd132ps(ymmA, ymmB, anyptr_gpC); - e->vfnmadd132ps(ymmA, ymmB, ymmC); - e->vfnmadd132sd(xmmA, xmmB, anyptr_gpC); - e->vfnmadd132sd(xmmA, xmmB, xmmC); - e->vfnmadd132ss(xmmA, xmmB, anyptr_gpC); - e->vfnmadd132ss(xmmA, xmmB, xmmC); - e->vfnmadd213pd(xmmA, xmmB, anyptr_gpC); - e->vfnmadd213pd(xmmA, xmmB, xmmC); - e->vfnmadd213pd(ymmA, ymmB, anyptr_gpC); - e->vfnmadd213pd(ymmA, ymmB, ymmC); - e->vfnmadd213ps(xmmA, xmmB, anyptr_gpC); - e->vfnmadd213ps(xmmA, xmmB, xmmC); - e->vfnmadd213ps(ymmA, ymmB, anyptr_gpC); - e->vfnmadd213ps(ymmA, ymmB, ymmC); - e->vfnmadd213sd(xmmA, xmmB, anyptr_gpC); - e->vfnmadd213sd(xmmA, xmmB, xmmC); - e->vfnmadd213ss(xmmA, xmmB, anyptr_gpC); - e->vfnmadd213ss(xmmA, xmmB, xmmC); - e->vfnmadd231pd(xmmA, xmmB, anyptr_gpC); - e->vfnmadd231pd(xmmA, xmmB, xmmC); - e->vfnmadd231pd(ymmA, ymmB, anyptr_gpC); - e->vfnmadd231pd(ymmA, ymmB, ymmC); - e->vfnmadd231ps(xmmA, xmmB, anyptr_gpC); - e->vfnmadd231ps(xmmA, xmmB, xmmC); - e->vfnmadd231ps(ymmA, ymmB, anyptr_gpC); - e->vfnmadd231ps(ymmA, ymmB, ymmC); - e->vfnmadd231sd(xmmA, xmmB, anyptr_gpC); - e->vfnmadd231sd(xmmA, xmmB, xmmC); - e->vfnmadd231ss(xmmA, xmmB, anyptr_gpC); - e->vfnmadd231ss(xmmA, xmmB, xmmC); - e->vfnmsub132pd(xmmA, xmmB, anyptr_gpC); - e->vfnmsub132pd(xmmA, xmmB, xmmC); - e->vfnmsub132pd(ymmA, ymmB, anyptr_gpC); - e->vfnmsub132pd(ymmA, ymmB, ymmC); - e->vfnmsub132ps(xmmA, xmmB, anyptr_gpC); - e->vfnmsub132ps(xmmA, xmmB, xmmC); - e->vfnmsub132ps(ymmA, ymmB, anyptr_gpC); - e->vfnmsub132ps(ymmA, ymmB, ymmC); - e->vfnmsub132sd(xmmA, xmmB, anyptr_gpC); - e->vfnmsub132sd(xmmA, xmmB, xmmC); - e->vfnmsub132ss(xmmA, xmmB, anyptr_gpC); - e->vfnmsub132ss(xmmA, xmmB, xmmC); - e->vfnmsub213pd(xmmA, xmmB, anyptr_gpC); - e->vfnmsub213pd(xmmA, xmmB, xmmC); - e->vfnmsub213pd(ymmA, ymmB, anyptr_gpC); - e->vfnmsub213pd(ymmA, ymmB, ymmC); - e->vfnmsub213ps(xmmA, xmmB, anyptr_gpC); - e->vfnmsub213ps(xmmA, xmmB, xmmC); - e->vfnmsub213ps(ymmA, ymmB, anyptr_gpC); - e->vfnmsub213ps(ymmA, ymmB, ymmC); - e->vfnmsub213sd(xmmA, xmmB, anyptr_gpC); - e->vfnmsub213sd(xmmA, xmmB, xmmC); - e->vfnmsub213ss(xmmA, xmmB, anyptr_gpC); - e->vfnmsub213ss(xmmA, xmmB, xmmC); - e->vfnmsub231pd(xmmA, xmmB, anyptr_gpC); - e->vfnmsub231pd(xmmA, xmmB, xmmC); - e->vfnmsub231pd(ymmA, ymmB, anyptr_gpC); - e->vfnmsub231pd(ymmA, ymmB, ymmC); - e->vfnmsub231ps(xmmA, xmmB, anyptr_gpC); - e->vfnmsub231ps(xmmA, xmmB, xmmC); - e->vfnmsub231ps(ymmA, ymmB, anyptr_gpC); - e->vfnmsub231ps(ymmA, ymmB, ymmC); - e->vfnmsub231sd(xmmA, xmmB, anyptr_gpC); - e->vfnmsub231sd(xmmA, xmmB, xmmC); - e->vfnmsub231ss(xmmA, xmmB, anyptr_gpC); - e->vfnmsub231ss(xmmA, xmmB, xmmC); - - // FMA4. - e->nop(); - - e->vfmaddpd(xmmA, xmmB, xmmC, xmmD); - e->vfmaddpd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfmaddpd(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfmaddpd(ymmA, ymmB, ymmC, ymmD); - e->vfmaddpd(ymmA, ymmB, anyptr_gpC, ymmD); - e->vfmaddpd(ymmA, ymmB, ymmC, anyptr_gpD); - e->vfmaddps(xmmA, xmmB, xmmC, xmmD); - e->vfmaddps(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfmaddps(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfmaddps(ymmA, ymmB, ymmC, ymmD); - e->vfmaddps(ymmA, ymmB, anyptr_gpC, ymmD); - e->vfmaddps(ymmA, ymmB, ymmC, anyptr_gpD); - e->vfmaddsd(xmmA, xmmB, xmmC, xmmD); - e->vfmaddsd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfmaddsd(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfmaddss(xmmA, xmmB, xmmC, xmmD); - e->vfmaddss(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfmaddss(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfmaddsubpd(xmmA, xmmB, xmmC, xmmD); - e->vfmaddsubpd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfmaddsubpd(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfmaddsubpd(ymmA, ymmB, ymmC, ymmD); - e->vfmaddsubpd(ymmA, ymmB, anyptr_gpC, ymmD); - e->vfmaddsubpd(ymmA, ymmB, ymmC, anyptr_gpD); - e->vfmaddsubps(xmmA, xmmB, xmmC, xmmD); - e->vfmaddsubps(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfmaddsubps(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfmaddsubps(ymmA, ymmB, ymmC, ymmD); - e->vfmaddsubps(ymmA, ymmB, anyptr_gpC, ymmD); - e->vfmaddsubps(ymmA, ymmB, ymmC, anyptr_gpD); - e->vfmsubaddpd(xmmA, xmmB, xmmC, xmmD); - e->vfmsubaddpd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfmsubaddpd(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfmsubaddpd(ymmA, ymmB, ymmC, ymmD); - e->vfmsubaddpd(ymmA, ymmB, anyptr_gpC, ymmD); - e->vfmsubaddpd(ymmA, ymmB, ymmC, anyptr_gpD); - e->vfmsubaddps(xmmA, xmmB, xmmC, xmmD); - e->vfmsubaddps(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfmsubaddps(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfmsubaddps(ymmA, ymmB, ymmC, ymmD); - e->vfmsubaddps(ymmA, ymmB, anyptr_gpC, ymmD); - e->vfmsubaddps(ymmA, ymmB, ymmC, anyptr_gpD); - e->vfmsubpd(xmmA, xmmB, xmmC, xmmD); - e->vfmsubpd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfmsubpd(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfmsubpd(ymmA, ymmB, ymmC, ymmD); - e->vfmsubpd(ymmA, ymmB, anyptr_gpC, ymmD); - e->vfmsubpd(ymmA, ymmB, ymmC, anyptr_gpD); - e->vfmsubps(xmmA, xmmB, xmmC, xmmD); - e->vfmsubps(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfmsubps(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfmsubps(ymmA, ymmB, ymmC, ymmD); - e->vfmsubps(ymmA, ymmB, anyptr_gpC, ymmD); - e->vfmsubps(ymmA, ymmB, ymmC, anyptr_gpD); - e->vfmsubsd(xmmA, xmmB, xmmC, xmmD); - e->vfmsubsd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfmsubsd(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfmsubss(xmmA, xmmB, xmmC, xmmD); - e->vfmsubss(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfmsubss(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfnmaddpd(xmmA, xmmB, xmmC, xmmD); - e->vfnmaddpd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfnmaddpd(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfnmaddpd(ymmA, ymmB, ymmC, ymmD); - e->vfnmaddpd(ymmA, ymmB, anyptr_gpC, ymmD); - e->vfnmaddpd(ymmA, ymmB, ymmC, anyptr_gpD); - e->vfnmaddps(xmmA, xmmB, xmmC, xmmD); - e->vfnmaddps(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfnmaddps(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfnmaddps(ymmA, ymmB, ymmC, ymmD); - e->vfnmaddps(ymmA, ymmB, anyptr_gpC, ymmD); - e->vfnmaddps(ymmA, ymmB, ymmC, anyptr_gpD); - e->vfnmaddsd(xmmA, xmmB, xmmC, xmmD); - e->vfnmaddsd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfnmaddsd(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfnmaddss(xmmA, xmmB, xmmC, xmmD); - e->vfnmaddss(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfnmaddss(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfnmsubpd(xmmA, xmmB, xmmC, xmmD); - e->vfnmsubpd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfnmsubpd(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfnmsubpd(ymmA, ymmB, ymmC, ymmD); - e->vfnmsubpd(ymmA, ymmB, anyptr_gpC, ymmD); - e->vfnmsubpd(ymmA, ymmB, ymmC, anyptr_gpD); - e->vfnmsubps(xmmA, xmmB, xmmC, xmmD); - e->vfnmsubps(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfnmsubps(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfnmsubps(ymmA, ymmB, ymmC, ymmD); - e->vfnmsubps(ymmA, ymmB, anyptr_gpC, ymmD); - e->vfnmsubps(ymmA, ymmB, ymmC, anyptr_gpD); - e->vfnmsubsd(xmmA, xmmB, xmmC, xmmD); - e->vfnmsubsd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfnmsubsd(xmmA, xmmB, xmmC, anyptr_gpD); - e->vfnmsubss(xmmA, xmmB, xmmC, xmmD); - e->vfnmsubss(xmmA, xmmB, anyptr_gpC, xmmD); - e->vfnmsubss(xmmA, xmmB, xmmC, anyptr_gpD); - - // XOP. - e->nop(); - - e->vfrczpd(xmmA, xmmB); - e->vfrczpd(xmmA, anyptr_gpB); - e->vfrczpd(ymmA, ymmB); - e->vfrczpd(ymmA, anyptr_gpB); - e->vfrczps(xmmA, xmmB); - e->vfrczps(xmmA, anyptr_gpB); - e->vfrczps(ymmA, ymmB); - e->vfrczps(ymmA, anyptr_gpB); - e->vfrczsd(xmmA, xmmB); - e->vfrczsd(xmmA, anyptr_gpB); - e->vfrczss(xmmA, xmmB); - e->vfrczss(xmmA, anyptr_gpB); - e->vpcmov(xmmA, xmmB, xmmC, xmmD); - e->vpcmov(xmmA, xmmB, anyptr_gpC, xmmD); - e->vpcmov(xmmA, xmmB, xmmC, anyptr_gpD); - e->vpcmov(ymmA, ymmB, ymmC, ymmD); - e->vpcmov(ymmA, ymmB, anyptr_gpC, ymmD); - e->vpcmov(ymmA, ymmB, ymmC, anyptr_gpD); - e->vpcomb(xmmA, xmmB, xmmC, 0); - e->vpcomb(xmmA, xmmB, anyptr_gpC, 0); - e->vpcomd(xmmA, xmmB, xmmC, 0); - e->vpcomd(xmmA, xmmB, anyptr_gpC, 0); - e->vpcomq(xmmA, xmmB, xmmC, 0); - e->vpcomq(xmmA, xmmB, anyptr_gpC, 0); - e->vpcomw(xmmA, xmmB, xmmC, 0); - e->vpcomw(xmmA, xmmB, anyptr_gpC, 0); - e->vpcomub(xmmA, xmmB, xmmC, 0); - e->vpcomub(xmmA, xmmB, anyptr_gpC, 0); - e->vpcomud(xmmA, xmmB, xmmC, 0); - e->vpcomud(xmmA, xmmB, anyptr_gpC, 0); - e->vpcomuq(xmmA, xmmB, xmmC, 0); - e->vpcomuq(xmmA, xmmB, anyptr_gpC, 0); - e->vpcomuw(xmmA, xmmB, xmmC, 0); - e->vpcomuw(xmmA, xmmB, anyptr_gpC, 0); - e->vpermil2pd(xmmA, xmmB, xmmC, xmmD, 0); - e->vpermil2pd(xmmA, xmmB, anyptr_gpC, xmmD, 0); - e->vpermil2pd(xmmA, xmmB, xmmC, anyptr_gpD, 0); - e->vpermil2pd(ymmA, ymmB, ymmC, ymmD, 0); - e->vpermil2pd(ymmA, ymmB, anyptr_gpC, ymmD, 0); - e->vpermil2pd(ymmA, ymmB, ymmC, anyptr_gpD, 0); - e->vpermil2ps(xmmA, xmmB, xmmC, xmmD, 0); - e->vpermil2ps(xmmA, xmmB, anyptr_gpC, xmmD, 0); - e->vpermil2ps(xmmA, xmmB, xmmC, anyptr_gpD, 0); - e->vpermil2ps(ymmA, ymmB, ymmC, ymmD, 0); - e->vpermil2ps(ymmA, ymmB, anyptr_gpC, ymmD, 0); - e->vpermil2ps(ymmA, ymmB, ymmC, anyptr_gpD, 0); - e->vphaddbd(xmmA, xmmB); - e->vphaddbd(xmmA, anyptr_gpB); - e->vphaddbq(xmmA, xmmB); - e->vphaddbq(xmmA, anyptr_gpB); - e->vphaddbw(xmmA, xmmB); - e->vphaddbw(xmmA, anyptr_gpB); - e->vphadddq(xmmA, xmmB); - e->vphadddq(xmmA, anyptr_gpB); - e->vphaddwd(xmmA, xmmB); - e->vphaddwd(xmmA, anyptr_gpB); - e->vphaddwq(xmmA, xmmB); - e->vphaddwq(xmmA, anyptr_gpB); - e->vphaddubd(xmmA, xmmB); - e->vphaddubd(xmmA, anyptr_gpB); - e->vphaddubq(xmmA, xmmB); - e->vphaddubq(xmmA, anyptr_gpB); - e->vphaddubw(xmmA, xmmB); - e->vphaddubw(xmmA, anyptr_gpB); - e->vphaddudq(xmmA, xmmB); - e->vphaddudq(xmmA, anyptr_gpB); - e->vphadduwd(xmmA, xmmB); - e->vphadduwd(xmmA, anyptr_gpB); - e->vphadduwq(xmmA, xmmB); - e->vphadduwq(xmmA, anyptr_gpB); - e->vphsubbw(xmmA, xmmB); - e->vphsubbw(xmmA, anyptr_gpB); - e->vphsubdq(xmmA, xmmB); - e->vphsubdq(xmmA, anyptr_gpB); - e->vphsubwd(xmmA, xmmB); - e->vphsubwd(xmmA, anyptr_gpB); - e->vpmacsdd(xmmA, xmmB, xmmC, xmmD); - e->vpmacsdd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vpmacsdqh(xmmA, xmmB, xmmC, xmmD); - e->vpmacsdqh(xmmA, xmmB, anyptr_gpC, xmmD); - e->vpmacsdql(xmmA, xmmB, xmmC, xmmD); - e->vpmacsdql(xmmA, xmmB, anyptr_gpC, xmmD); - e->vpmacswd(xmmA, xmmB, xmmC, xmmD); - e->vpmacswd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vpmacsww(xmmA, xmmB, xmmC, xmmD); - e->vpmacsww(xmmA, xmmB, anyptr_gpC, xmmD); - e->vpmacssdd(xmmA, xmmB, xmmC, xmmD); - e->vpmacssdd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vpmacssdqh(xmmA, xmmB, xmmC, xmmD); - e->vpmacssdqh(xmmA, xmmB, anyptr_gpC, xmmD); - e->vpmacssdql(xmmA, xmmB, xmmC, xmmD); - e->vpmacssdql(xmmA, xmmB, anyptr_gpC, xmmD); - e->vpmacsswd(xmmA, xmmB, xmmC, xmmD); - e->vpmacsswd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vpmacssww(xmmA, xmmB, xmmC, xmmD); - e->vpmacssww(xmmA, xmmB, anyptr_gpC, xmmD); - e->vpmadcsswd(xmmA, xmmB, xmmC, xmmD); - e->vpmadcsswd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vpmadcswd(xmmA, xmmB, xmmC, xmmD); - e->vpmadcswd(xmmA, xmmB, anyptr_gpC, xmmD); - e->vpperm(xmmA, xmmB, xmmC, xmmD); - e->vpperm(xmmA, xmmB, anyptr_gpC, xmmD); - e->vpperm(xmmA, xmmB, xmmC, anyptr_gpD); - e->vprotb(xmmA, xmmB, xmmC); - e->vprotb(xmmA, anyptr_gpB, xmmC); - e->vprotb(xmmA, xmmB, anyptr_gpC); - e->vprotb(xmmA, xmmB, 0); - e->vprotb(xmmA, anyptr_gpB, 0); - e->vprotd(xmmA, xmmB, xmmC); - e->vprotd(xmmA, anyptr_gpB, xmmC); - e->vprotd(xmmA, xmmB, anyptr_gpC); - e->vprotd(xmmA, xmmB, 0); - e->vprotd(xmmA, anyptr_gpB, 0); - e->vprotq(xmmA, xmmB, xmmC); - e->vprotq(xmmA, anyptr_gpB, xmmC); - e->vprotq(xmmA, xmmB, anyptr_gpC); - e->vprotq(xmmA, xmmB, 0); - e->vprotq(xmmA, anyptr_gpB, 0); - e->vprotw(xmmA, xmmB, xmmC); - e->vprotw(xmmA, anyptr_gpB, xmmC); - e->vprotw(xmmA, xmmB, anyptr_gpC); - e->vprotw(xmmA, xmmB, 0); - e->vprotw(xmmA, anyptr_gpB, 0); - e->vpshab(xmmA, xmmB, xmmC); - e->vpshab(xmmA, anyptr_gpB, xmmC); - e->vpshab(xmmA, xmmB, anyptr_gpC); - e->vpshad(xmmA, xmmB, xmmC); - e->vpshad(xmmA, anyptr_gpB, xmmC); - e->vpshad(xmmA, xmmB, anyptr_gpC); - e->vpshaq(xmmA, xmmB, xmmC); - e->vpshaq(xmmA, anyptr_gpB, xmmC); - e->vpshaq(xmmA, xmmB, anyptr_gpC); - e->vpshaw(xmmA, xmmB, xmmC); - e->vpshaw(xmmA, anyptr_gpB, xmmC); - e->vpshaw(xmmA, xmmB, anyptr_gpC); - e->vpshlb(xmmA, xmmB, xmmC); - e->vpshlb(xmmA, anyptr_gpB, xmmC); - e->vpshlb(xmmA, xmmB, anyptr_gpC); - e->vpshld(xmmA, xmmB, xmmC); - e->vpshld(xmmA, anyptr_gpB, xmmC); - e->vpshld(xmmA, xmmB, anyptr_gpC); - e->vpshlq(xmmA, xmmB, xmmC); - e->vpshlq(xmmA, anyptr_gpB, xmmC); - e->vpshlq(xmmA, xmmB, anyptr_gpC); - e->vpshlw(xmmA, xmmB, xmmC); - e->vpshlw(xmmA, anyptr_gpB, xmmC); - e->vpshlw(xmmA, xmmB, anyptr_gpC); - - // F16C. - e->nop(); - - e->vcvtph2ps(xmmA, xmmB); - e->vcvtph2ps(xmmA, anyptr_gpB); - e->vcvtph2ps(ymmA, xmmB); - e->vcvtph2ps(ymmA, anyptr_gpB); - e->vcvtps2ph(xmmA, xmmB, 0); - e->vcvtps2ph(anyptr_gpA, xmmB, 0); - e->vcvtps2ph(xmmA, ymmB, 0); - e->vcvtps2ph(anyptr_gpA, ymmB, 0); - - // AVX512. - e->nop(); - - e->kaddb(kA, kB, kC); - e->kaddd(kA, kB, kC); - e->kaddq(kA, kB, kC); - e->kaddw(kA, kB, kC); - e->kandb(kA, kB, kC); - e->kandd(kA, kB, kC); - e->kandnb(kA, kB, kC); - e->kandnd(kA, kB, kC); - e->kandnq(kA, kB, kC); - e->kandnw(kA, kB, kC); - e->kandq(kA, kB, kC); - e->kandw(kA, kB, kC); - e->kmovb(kA, kB); - e->kmovb(kA, anyptr_gpB); - e->kmovb(kA, gdB); - if (isX64) e->kmovb(kA, gzB); - e->kmovb(anyptr_gpA, kB); - e->kmovb(gdA, kB); - if (isX64) e->kmovb(gzA, kB); - e->kmovd(kA, kB); - e->kmovd(kA, anyptr_gpB); - e->kmovd(kA, gdB); - if (isX64) e->kmovd(kA, gzB); - e->kmovd(anyptr_gpA, kB); - e->kmovd(gdA, kB); - if (isX64) e->kmovd(gzA, kB); - e->kmovq(kA, kB); - e->kmovq(kA, anyptr_gpB); - if (isX64) e->kmovq(kA, gzB); - e->kmovq(anyptr_gpA, kB); - if (isX64) e->kmovq(gzA, kB); - e->kmovw(kA, kB); - e->kmovw(kA, anyptr_gpB); - e->kmovw(kA, gdB); - if (isX64) e->kmovw(kA, gzB); - e->kmovw(anyptr_gpA, kB); - e->kmovw(gdA, kB); - if (isX64) e->kmovw(gzA, kB); - e->knotb(kA, kB); - e->knotd(kA, kB); - e->knotq(kA, kB); - e->knotw(kA, kB); - e->korb(kA, kB, kC); - e->kord(kA, kB, kC); - e->korq(kA, kB, kC); - e->kortestb(kA, kB); - e->kortestd(kA, kB); - e->kortestq(kA, kB); - e->kortestw(kA, kB); - e->korw(kA, kB, kC); - e->kshiftlb(kA, kB, 0); - e->kshiftld(kA, kB, 0); - e->kshiftlq(kA, kB, 0); - e->kshiftlw(kA, kB, 0); - e->kshiftrb(kA, kB, 0); - e->kshiftrd(kA, kB, 0); - e->kshiftrq(kA, kB, 0); - e->kshiftrw(kA, kB, 0); - e->ktestb(kA, kB); - e->ktestd(kA, kB); - e->ktestq(kA, kB); - e->ktestw(kA, kB); - e->kunpckbw(kA, kB, kC); - e->kunpckdq(kA, kB, kC); - e->kunpckwd(kA, kB, kC); - e->kxnorb(kA, kB, kC); - e->kxnord(kA, kB, kC); - e->kxnorq(kA, kB, kC); - e->kxnorw(kA, kB, kC); - e->kxorb(kA, kB, kC); - e->kxord(kA, kB, kC); - e->kxorq(kA, kB, kC); - e->kxorw(kA, kB, kC); - e->nop(); - - e->vaddpd(xmmA, xmmB, xmmC); - e->vaddpd(xmmA, xmmB, anyptr_gpC); - e->vaddpd(ymmA, ymmB, ymmC); - e->vaddpd(ymmA, ymmB, anyptr_gpC); - e->vaddpd(zmmA, zmmB, zmmC); - e->vaddpd(zmmA, zmmB, anyptr_gpC); - e->vaddps(xmmA, xmmB, xmmC); - e->vaddps(xmmA, xmmB, anyptr_gpC); - e->vaddps(ymmA, ymmB, ymmC); - e->vaddps(ymmA, ymmB, anyptr_gpC); - e->vaddps(zmmA, zmmB, zmmC); - e->vaddps(zmmA, zmmB, anyptr_gpC); - e->vaddsd(xmmA, xmmB, xmmC); - e->vaddsd(xmmA, xmmB, anyptr_gpC); - e->vaddss(xmmA, xmmB, xmmC); - e->vaddss(xmmA, xmmB, anyptr_gpC); - e->valignd(xmmA, xmmB, xmmC, 0); - e->valignd(xmmA, xmmB, anyptr_gpC, 0); - e->valignd(ymmA, ymmB, ymmC, 0); - e->valignd(ymmA, ymmB, anyptr_gpC, 0); - e->valignd(zmmA, zmmB, zmmC, 0); - e->valignd(zmmA, zmmB, anyptr_gpC, 0); - e->valignq(xmmA, xmmB, xmmC, 0); - e->valignq(xmmA, xmmB, anyptr_gpC, 0); - e->valignq(ymmA, ymmB, ymmC, 0); - e->valignq(ymmA, ymmB, anyptr_gpC, 0); - e->valignq(zmmA, zmmB, zmmC, 0); - e->valignq(zmmA, zmmB, anyptr_gpC, 0); - e->vandnpd(xmmA, xmmB, xmmC); - e->vandnpd(xmmA, xmmB, anyptr_gpC); - e->vandnpd(ymmA, ymmB, ymmC); - e->vandnpd(ymmA, ymmB, anyptr_gpC); - e->vandnpd(zmmA, zmmB, zmmC); - e->vandnpd(zmmA, zmmB, anyptr_gpC); - e->vandnps(xmmA, xmmB, xmmC); - e->vandnps(xmmA, xmmB, anyptr_gpC); - e->vandnps(ymmA, ymmB, ymmC); - e->vandnps(ymmA, ymmB, anyptr_gpC); - e->vandnps(zmmA, zmmB, zmmC); - e->vandnps(zmmA, zmmB, anyptr_gpC); - e->vandpd(xmmA, xmmB, xmmC); - e->vandpd(xmmA, xmmB, anyptr_gpC); - e->vandpd(ymmA, ymmB, ymmC); - e->vandpd(ymmA, ymmB, anyptr_gpC); - e->vandpd(zmmA, zmmB, zmmC); - e->vandpd(zmmA, zmmB, anyptr_gpC); - e->vandps(xmmA, xmmB, xmmC); - e->vandps(xmmA, xmmB, anyptr_gpC); - e->vandps(ymmA, ymmB, ymmC); - e->vandps(ymmA, ymmB, anyptr_gpC); - e->vandps(zmmA, zmmB, zmmC); - e->vandps(zmmA, zmmB, anyptr_gpC); - e->vblendmpd(xmmA, xmmB, xmmC); - e->vblendmpd(xmmA, xmmB, anyptr_gpC); - e->vblendmpd(ymmA, ymmB, ymmC); - e->vblendmpd(ymmA, ymmB, anyptr_gpC); - e->vblendmpd(zmmA, zmmB, zmmC); - e->vblendmpd(zmmA, zmmB, anyptr_gpC); - e->vblendmps(xmmA, xmmB, xmmC); - e->vblendmps(xmmA, xmmB, anyptr_gpC); - e->vblendmps(ymmA, ymmB, ymmC); - e->vblendmps(ymmA, ymmB, anyptr_gpC); - e->vblendmps(zmmA, zmmB, zmmC); - e->vblendmps(zmmA, zmmB, anyptr_gpC); - e->vbroadcastf32x2(ymmA, xmmB); - e->vbroadcastf32x2(ymmA, anyptr_gpB); - e->vbroadcastf32x2(zmmA, xmmB); - e->vbroadcastf32x2(zmmA, anyptr_gpB); - e->vbroadcastf32x4(ymmA, anyptr_gpB); - e->vbroadcastf32x4(zmmA, anyptr_gpB); - e->vbroadcastf32x8(zmmA, anyptr_gpB); - e->vbroadcastf64x2(ymmA, anyptr_gpB); - e->vbroadcastf64x2(zmmA, anyptr_gpB); - e->vbroadcastf64x4(zmmA, anyptr_gpB); - e->vbroadcasti32x2(xmmA, xmmB); - e->vbroadcasti32x2(xmmA, anyptr_gpB); - e->vbroadcasti32x2(ymmA, xmmB); - e->vbroadcasti32x2(ymmA, anyptr_gpB); - e->vbroadcasti32x2(zmmA, xmmB); - e->vbroadcasti32x2(zmmA, anyptr_gpB); - e->vbroadcasti32x4(ymmA, anyptr_gpB); - e->vbroadcasti32x4(zmmA, anyptr_gpB); - e->vbroadcasti32x8(zmmA, anyptr_gpB); - e->vbroadcasti64x2(ymmA, anyptr_gpB); - e->vbroadcasti64x2(zmmA, anyptr_gpB); - e->vbroadcasti64x4(zmmA, anyptr_gpB); - e->vbroadcastsd(ymmA, xmmB); - e->vbroadcastsd(ymmA, anyptr_gpB); - e->vbroadcastsd(zmmA, xmmB); - e->vbroadcastsd(zmmA, anyptr_gpB); - e->vbroadcastss(xmmA, xmmB); - e->vbroadcastss(xmmA, anyptr_gpB); - e->vbroadcastss(ymmA, xmmB); - e->vbroadcastss(ymmA, anyptr_gpB); - e->vbroadcastss(zmmA, xmmB); - e->vbroadcastss(zmmA, anyptr_gpB); - e->vcmppd(kA, xmmB, xmmC, 0); - e->vcmppd(kA, xmmB, anyptr_gpC, 0); - e->vcmppd(kA, ymmB, ymmC, 0); - e->vcmppd(kA, ymmB, anyptr_gpC, 0); - e->vcmppd(kA, zmmB, zmmC, 0); - e->vcmppd(kA, zmmB, anyptr_gpC, 0); - e->vcmpps(kA, xmmB, xmmC, 0); - e->vcmpps(kA, xmmB, anyptr_gpC, 0); - e->vcmpps(kA, ymmB, ymmC, 0); - e->vcmpps(kA, ymmB, anyptr_gpC, 0); - e->vcmpps(kA, zmmB, zmmC, 0); - e->vcmpps(kA, zmmB, anyptr_gpC, 0); - e->vcmpsd(kA, xmmB, xmmC, 0); - e->vcmpsd(kA, xmmB, anyptr_gpC, 0); - e->vcmpss(kA, xmmB, xmmC, 0); - e->vcmpss(kA, xmmB, anyptr_gpC, 0); - e->vcomisd(xmmA, xmmB); - e->vcomisd(xmmA, anyptr_gpB); - e->vcomiss(xmmA, xmmB); - e->vcomiss(xmmA, anyptr_gpB); - e->vcompresspd(xmmA, xmmB); - e->vcompresspd(anyptr_gpA, xmmB); - e->vcompresspd(ymmA, ymmB); - e->vcompresspd(anyptr_gpA, ymmB); - e->vcompresspd(zmmA, zmmB); - e->vcompresspd(anyptr_gpA, zmmB); - e->vcompressps(xmmA, xmmB); - e->vcompressps(anyptr_gpA, xmmB); - e->vcompressps(ymmA, ymmB); - e->vcompressps(anyptr_gpA, ymmB); - e->vcompressps(zmmA, zmmB); - e->vcompressps(anyptr_gpA, zmmB); - e->vcvtdq2pd(xmmA, xmmB); - e->vcvtdq2pd(xmmA, anyptr_gpB); - e->vcvtdq2pd(ymmA, xmmB); - e->vcvtdq2pd(ymmA, anyptr_gpB); - e->vcvtdq2pd(zmmA, ymmB); - e->vcvtdq2pd(zmmA, anyptr_gpB); - e->vcvtdq2ps(xmmA, xmmB); - e->vcvtdq2ps(xmmA, anyptr_gpB); - e->vcvtdq2ps(ymmA, ymmB); - e->vcvtdq2ps(ymmA, anyptr_gpB); - e->vcvtdq2ps(zmmA, zmmB); - e->vcvtdq2ps(zmmA, anyptr_gpB); - e->vcvtpd2dq(xmmA, xmmB); - e->vcvtpd2dq(xmmA, anyptr_gpB); - e->vcvtpd2dq(xmmA, ymmB); - e->vcvtpd2dq(xmmA, anyptr_gpB); - e->vcvtpd2dq(ymmA, zmmB); - e->vcvtpd2dq(ymmA, anyptr_gpB); - e->vcvtpd2qq(xmmA, xmmB); - e->vcvtpd2qq(xmmA, anyptr_gpB); - e->vcvtpd2qq(ymmA, ymmB); - e->vcvtpd2qq(ymmA, anyptr_gpB); - e->vcvtpd2qq(zmmA, zmmB); - e->vcvtpd2qq(zmmA, anyptr_gpB); - e->vcvtpd2udq(xmmA, xmmB); - e->vcvtpd2udq(xmmA, anyptr_gpB); - e->vcvtpd2udq(xmmA, ymmB); - e->vcvtpd2udq(xmmA, anyptr_gpB); - e->vcvtpd2udq(ymmA, zmmB); - e->vcvtpd2udq(ymmA, anyptr_gpB); - e->vcvtpd2uqq(xmmA, xmmB); - e->vcvtpd2uqq(xmmA, anyptr_gpB); - e->vcvtpd2uqq(ymmA, ymmB); - e->vcvtpd2uqq(ymmA, anyptr_gpB); - e->vcvtpd2uqq(zmmA, zmmB); - e->vcvtpd2uqq(zmmA, anyptr_gpB); - e->vcvtph2ps(xmmA, xmmB); - e->vcvtph2ps(xmmA, anyptr_gpB); - e->vcvtph2ps(ymmA, xmmB); - e->vcvtph2ps(ymmA, anyptr_gpB); - e->vcvtph2ps(zmmA, ymmB); - e->vcvtph2ps(zmmA, anyptr_gpB); - e->vcvtps2dq(xmmA, xmmB); - e->vcvtps2dq(xmmA, anyptr_gpB); - e->vcvtps2dq(ymmA, ymmB); - e->vcvtps2dq(ymmA, anyptr_gpB); - e->vcvtps2dq(zmmA, zmmB); - e->vcvtps2dq(zmmA, anyptr_gpB); - e->vcvtps2pd(xmmA, xmmB); - e->vcvtps2pd(xmmA, anyptr_gpB); - e->vcvtps2pd(ymmA, xmmB); - e->vcvtps2pd(ymmA, anyptr_gpB); - e->vcvtps2pd(zmmA, ymmB); - e->vcvtps2pd(zmmA, anyptr_gpB); - e->vcvtps2ph(xmmA, xmmB, 0); - e->vcvtps2ph(anyptr_gpA, xmmB, 0); - e->vcvtps2ph(xmmA, ymmB, 0); - e->vcvtps2ph(anyptr_gpA, ymmB, 0); - e->vcvtps2ph(ymmA, zmmB, 0); - e->vcvtps2ph(anyptr_gpA, zmmB, 0); - e->vcvtps2qq(xmmA, xmmB); - e->vcvtps2qq(xmmA, anyptr_gpB); - e->vcvtps2qq(ymmA, xmmB); - e->vcvtps2qq(ymmA, anyptr_gpB); - e->vcvtps2qq(zmmA, ymmB); - e->vcvtps2qq(zmmA, anyptr_gpB); - e->vcvtps2udq(xmmA, xmmB); - e->vcvtps2udq(xmmA, anyptr_gpB); - e->vcvtps2udq(ymmA, ymmB); - e->vcvtps2udq(ymmA, anyptr_gpB); - e->vcvtps2udq(zmmA, zmmB); - e->vcvtps2udq(zmmA, anyptr_gpB); - e->vcvtps2uqq(xmmA, xmmB); - e->vcvtps2uqq(xmmA, anyptr_gpB); - e->vcvtps2uqq(ymmA, xmmB); - e->vcvtps2uqq(ymmA, anyptr_gpB); - e->vcvtps2uqq(zmmA, ymmB); - e->vcvtps2uqq(zmmA, anyptr_gpB); - e->vcvtqq2pd(xmmA, xmmB); - e->vcvtqq2pd(xmmA, anyptr_gpB); - e->vcvtqq2pd(ymmA, ymmB); - e->vcvtqq2pd(ymmA, anyptr_gpB); - e->vcvtqq2pd(zmmA, zmmB); - e->vcvtqq2pd(zmmA, anyptr_gpB); - e->vcvtqq2ps(xmmA, xmmB); - e->vcvtqq2ps(xmmA, anyptr_gpB); - e->vcvtqq2ps(xmmA, ymmB); - e->vcvtqq2ps(xmmA, anyptr_gpB); - e->vcvtqq2ps(ymmA, zmmB); - e->vcvtqq2ps(ymmA, anyptr_gpB); - e->vcvtsd2si(gdA, xmmB); - e->vcvtsd2si(gdA, anyptr_gpB); - if (isX64) e->vcvtsd2si(gzA, xmmB); - if (isX64) e->vcvtsd2si(gzA, anyptr_gpB); - e->vcvtsd2ss(xmmA, xmmB, xmmC); - e->vcvtsd2ss(xmmA, xmmB, anyptr_gpC); - e->vcvtsd2usi(gdA, xmmB); - e->vcvtsd2usi(gdA, anyptr_gpB); - if (isX64) e->vcvtsd2usi(gzA, xmmB); - if (isX64) e->vcvtsd2usi(gzA, anyptr_gpB); - e->vcvtsi2sd(xmmA, xmmB, gdC); - e->vcvtsi2sd(xmmA, xmmB, dword_ptr(gzC)); - if (isX64) e->vcvtsi2sd(xmmA, xmmB, gzC); - if (isX64) e->vcvtsi2sd(xmmA, xmmB, qword_ptr(gzC)); - e->vcvtsi2ss(xmmA, xmmB, gdC); - e->vcvtsi2ss(xmmA, xmmB, dword_ptr(gzC)); - if (isX64) e->vcvtsi2ss(xmmA, xmmB, gzC); - if (isX64) e->vcvtsi2ss(xmmA, xmmB, qword_ptr(gzC)); - e->vcvtss2sd(xmmA, xmmB, xmmC); - e->vcvtss2sd(xmmA, xmmB, anyptr_gpC); - e->vcvtss2si(gdA, xmmB); - e->vcvtss2si(gdA, anyptr_gpB); - if (isX64) e->vcvtss2si(gzA, xmmB); - if (isX64) e->vcvtss2si(gzA, anyptr_gpB); - e->vcvtss2usi(gdA, xmmB); - e->vcvtss2usi(gdA, anyptr_gpB); - if (isX64) e->vcvtss2usi(gzA, xmmB); - if (isX64) e->vcvtss2usi(gzA, anyptr_gpB); - e->vcvttpd2dq(xmmA, xmmB); - e->vcvttpd2dq(xmmA, anyptr_gpB); - e->vcvttpd2dq(xmmA, ymmB); - e->vcvttpd2dq(xmmA, anyptr_gpB); - e->vcvttpd2dq(ymmA, zmmB); - e->vcvttpd2dq(ymmA, anyptr_gpB); - e->vcvttpd2qq(xmmA, xmmB); - e->vcvttpd2qq(xmmA, anyptr_gpB); - e->vcvttpd2qq(ymmA, ymmB); - e->vcvttpd2qq(ymmA, anyptr_gpB); - e->vcvttpd2qq(zmmA, zmmB); - e->vcvttpd2qq(zmmA, anyptr_gpB); - e->vcvttpd2udq(xmmA, xmmB); - e->vcvttpd2udq(xmmA, anyptr_gpB); - e->vcvttpd2udq(xmmA, ymmB); - e->vcvttpd2udq(xmmA, anyptr_gpB); - e->vcvttpd2udq(ymmA, zmmB); - e->vcvttpd2udq(ymmA, anyptr_gpB); - e->vcvttpd2uqq(xmmA, xmmB); - e->vcvttpd2uqq(xmmA, anyptr_gpB); - e->vcvttpd2uqq(ymmA, ymmB); - e->vcvttpd2uqq(ymmA, anyptr_gpB); - e->vcvttpd2uqq(zmmA, zmmB); - e->vcvttpd2uqq(zmmA, anyptr_gpB); - e->vcvttps2dq(xmmA, xmmB); - e->vcvttps2dq(xmmA, anyptr_gpB); - e->vcvttps2dq(ymmA, ymmB); - e->vcvttps2dq(ymmA, anyptr_gpB); - e->vcvttps2dq(zmmA, zmmB); - e->vcvttps2dq(zmmA, anyptr_gpB); - e->vcvttps2qq(xmmA, xmmB); - e->vcvttps2qq(xmmA, anyptr_gpB); - e->vcvttps2qq(ymmA, xmmB); - e->vcvttps2qq(ymmA, anyptr_gpB); - e->vcvttps2qq(zmmA, ymmB); - e->vcvttps2qq(zmmA, anyptr_gpB); - e->vcvttps2udq(xmmA, xmmB); - e->vcvttps2udq(xmmA, anyptr_gpB); - e->vcvttps2udq(ymmA, ymmB); - e->vcvttps2udq(ymmA, anyptr_gpB); - e->vcvttps2udq(zmmA, zmmB); - e->vcvttps2udq(zmmA, anyptr_gpB); - e->vcvttps2uqq(xmmA, xmmB); - e->vcvttps2uqq(xmmA, anyptr_gpB); - e->vcvttps2uqq(ymmA, xmmB); - e->vcvttps2uqq(ymmA, anyptr_gpB); - e->vcvttps2uqq(zmmA, ymmB); - e->vcvttps2uqq(zmmA, anyptr_gpB); - e->vcvttsd2si(gdA, xmmB); - e->vcvttsd2si(gdA, anyptr_gpB); - if (isX64) e->vcvttsd2si(gzA, xmmB); - if (isX64) e->vcvttsd2si(gzA, anyptr_gpB); - e->vcvttsd2usi(gdA, xmmB); - e->vcvttsd2usi(gdA, anyptr_gpB); - if (isX64) e->vcvttsd2usi(gzA, xmmB); - if (isX64) e->vcvttsd2usi(gzA, anyptr_gpB); - e->vcvttss2si(gdA, xmmB); - e->vcvttss2si(gdA, anyptr_gpB); - if (isX64) e->vcvttss2si(gzA, xmmB); - if (isX64) e->vcvttss2si(gzA, anyptr_gpB); - e->vcvttss2usi(gdA, xmmB); - e->vcvttss2usi(gdA, anyptr_gpB); - if (isX64) e->vcvttss2usi(gzA, xmmB); - if (isX64) e->vcvttss2usi(gzA, anyptr_gpB); - e->vcvtudq2pd(xmmA, xmmB); - e->vcvtudq2pd(xmmA, anyptr_gpB); - e->vcvtudq2pd(ymmA, xmmB); - e->vcvtudq2pd(ymmA, anyptr_gpB); - e->vcvtudq2pd(zmmA, ymmB); - e->vcvtudq2pd(zmmA, anyptr_gpB); - e->vcvtudq2ps(xmmA, xmmB); - e->vcvtudq2ps(xmmA, anyptr_gpB); - e->vcvtudq2ps(ymmA, ymmB); - e->vcvtudq2ps(ymmA, anyptr_gpB); - e->vcvtudq2ps(zmmA, zmmB); - e->vcvtudq2ps(zmmA, anyptr_gpB); - e->vcvtuqq2pd(xmmA, xmmB); - e->vcvtuqq2pd(xmmA, anyptr_gpB); - e->vcvtuqq2pd(ymmA, ymmB); - e->vcvtuqq2pd(ymmA, anyptr_gpB); - e->vcvtuqq2pd(zmmA, zmmB); - e->vcvtuqq2pd(zmmA, anyptr_gpB); - e->vcvtuqq2ps(xmmA, xmmB); - e->vcvtuqq2ps(xmmA, anyptr_gpB); - e->vcvtuqq2ps(xmmA, ymmB); - e->vcvtuqq2ps(xmmA, anyptr_gpB); - e->vcvtuqq2ps(ymmA, zmmB); - e->vcvtuqq2ps(ymmA, anyptr_gpB); - e->vcvtusi2sd(xmmA, xmmB, gdC); - e->vcvtusi2sd(xmmA, xmmB, dword_ptr(gzC)); - if (isX64) e->vcvtusi2sd(xmmA, xmmB, gzC); - if (isX64) e->vcvtusi2sd(xmmA, xmmB, qword_ptr(gzC)); - e->vcvtusi2ss(xmmA, xmmB, gdC); - e->vcvtusi2ss(xmmA, xmmB, dword_ptr(gzC)); - if (isX64) e->vcvtusi2ss(xmmA, xmmB, gzC); - if (isX64) e->vcvtusi2ss(xmmA, xmmB, qword_ptr(gzC)); - e->vdbpsadbw(xmmA, xmmB, xmmC, 0); - e->vdbpsadbw(xmmA, xmmB, anyptr_gpC, 0); - e->vdbpsadbw(ymmA, ymmB, ymmC, 0); - e->vdbpsadbw(ymmA, ymmB, anyptr_gpC, 0); - e->vdbpsadbw(zmmA, zmmB, zmmC, 0); - e->vdbpsadbw(zmmA, zmmB, anyptr_gpC, 0); - e->vdivpd(xmmA, xmmB, xmmC); - e->vdivpd(xmmA, xmmB, anyptr_gpC); - e->vdivpd(ymmA, ymmB, ymmC); - e->vdivpd(ymmA, ymmB, anyptr_gpC); - e->vdivpd(zmmA, zmmB, zmmC); - e->vdivpd(zmmA, zmmB, anyptr_gpC); - e->vdivps(xmmA, xmmB, xmmC); - e->vdivps(xmmA, xmmB, anyptr_gpC); - e->vdivps(ymmA, ymmB, ymmC); - e->vdivps(ymmA, ymmB, anyptr_gpC); - e->vdivps(zmmA, zmmB, zmmC); - e->vdivps(zmmA, zmmB, anyptr_gpC); - e->vdivsd(xmmA, xmmB, xmmC); - e->vdivsd(xmmA, xmmB, anyptr_gpC); - e->vdivss(xmmA, xmmB, xmmC); - e->vdivss(xmmA, xmmB, anyptr_gpC); - e->vexp2pd(zmmA, zmmB); - e->vexp2pd(zmmA, anyptr_gpB); - e->vexp2ps(zmmA, zmmB); - e->vexp2ps(zmmA, anyptr_gpB); - e->vexpandpd(xmmA, xmmB); - e->vexpandpd(xmmA, anyptr_gpB); - e->vexpandpd(ymmA, ymmB); - e->vexpandpd(ymmA, anyptr_gpB); - e->vexpandpd(zmmA, zmmB); - e->vexpandpd(zmmA, anyptr_gpB); - e->vexpandps(xmmA, xmmB); - e->vexpandps(xmmA, anyptr_gpB); - e->vexpandps(ymmA, ymmB); - e->vexpandps(ymmA, anyptr_gpB); - e->vexpandps(zmmA, zmmB); - e->vexpandps(zmmA, anyptr_gpB); - e->vextractf32x4(xmmA, ymmB, 0); - e->vextractf32x4(anyptr_gpA, ymmB, 0); - e->vextractf32x4(xmmA, zmmB, 0); - e->vextractf32x4(anyptr_gpA, zmmB, 0); - e->vextractf32x8(ymmA, zmmB, 0); - e->vextractf32x8(anyptr_gpA, zmmB, 0); - e->vextractf64x2(xmmA, ymmB, 0); - e->vextractf64x2(anyptr_gpA, ymmB, 0); - e->vextractf64x2(xmmA, zmmB, 0); - e->vextractf64x2(anyptr_gpA, zmmB, 0); - e->vextractf64x4(ymmA, zmmB, 0); - e->vextractf64x4(anyptr_gpA, zmmB, 0); - e->vextracti32x4(xmmA, ymmB, 0); - e->vextracti32x4(anyptr_gpA, ymmB, 0); - e->vextracti32x4(xmmA, zmmB, 0); - e->vextracti32x4(anyptr_gpA, zmmB, 0); - e->vextracti32x8(ymmA, zmmB, 0); - e->vextracti32x8(anyptr_gpA, zmmB, 0); - e->vextracti64x2(xmmA, ymmB, 0); - e->vextracti64x2(anyptr_gpA, ymmB, 0); - e->vextracti64x2(xmmA, zmmB, 0); - e->vextracti64x2(anyptr_gpA, zmmB, 0); - e->vextracti64x4(ymmA, zmmB, 0); - e->vextracti64x4(anyptr_gpA, zmmB, 0); - e->vextractps(gdA, xmmB, 0); - e->vextractps(gzA, xmmB, 0); - e->vextractps(anyptr_gpA, xmmB, 0); - e->vfixupimmpd(xmmA, xmmB, xmmC, 0); - e->vfixupimmpd(xmmA, xmmB, anyptr_gpC, 0); - e->vfixupimmpd(ymmA, ymmB, ymmC, 0); - e->vfixupimmpd(ymmA, ymmB, anyptr_gpC, 0); - e->vfixupimmpd(zmmA, zmmB, zmmC, 0); - e->vfixupimmpd(zmmA, zmmB, anyptr_gpC, 0); - e->vfixupimmps(xmmA, xmmB, xmmC, 0); - e->vfixupimmps(xmmA, xmmB, anyptr_gpC, 0); - e->vfixupimmps(ymmA, ymmB, ymmC, 0); - e->vfixupimmps(ymmA, ymmB, anyptr_gpC, 0); - e->vfixupimmps(zmmA, zmmB, zmmC, 0); - e->vfixupimmps(zmmA, zmmB, anyptr_gpC, 0); - e->vfixupimmsd(xmmA, xmmB, xmmC, 0); - e->vfixupimmsd(xmmA, xmmB, anyptr_gpC, 0); - e->vfixupimmss(xmmA, xmmB, xmmC, 0); - e->vfixupimmss(xmmA, xmmB, anyptr_gpC, 0); - e->vfmadd132pd(xmmA, xmmB, xmmC); - e->vfmadd132pd(xmmA, xmmB, anyptr_gpC); - e->vfmadd132pd(ymmA, ymmB, ymmC); - e->vfmadd132pd(ymmA, ymmB, anyptr_gpC); - e->vfmadd132pd(zmmA, zmmB, zmmC); - e->vfmadd132pd(zmmA, zmmB, anyptr_gpC); - e->vfmadd132ps(xmmA, xmmB, xmmC); - e->vfmadd132ps(xmmA, xmmB, anyptr_gpC); - e->vfmadd132ps(ymmA, ymmB, ymmC); - e->vfmadd132ps(ymmA, ymmB, anyptr_gpC); - e->vfmadd132ps(zmmA, zmmB, zmmC); - e->vfmadd132ps(zmmA, zmmB, anyptr_gpC); - e->vfmadd132sd(xmmA, xmmB, xmmC); - e->vfmadd132sd(xmmA, xmmB, anyptr_gpC); - e->vfmadd132ss(xmmA, xmmB, xmmC); - e->vfmadd132ss(xmmA, xmmB, anyptr_gpC); - e->vfmadd213pd(xmmA, xmmB, xmmC); - e->vfmadd213pd(xmmA, xmmB, anyptr_gpC); - e->vfmadd213pd(ymmA, ymmB, ymmC); - e->vfmadd213pd(ymmA, ymmB, anyptr_gpC); - e->vfmadd213pd(zmmA, zmmB, zmmC); - e->vfmadd213pd(zmmA, zmmB, anyptr_gpC); - e->vfmadd213ps(xmmA, xmmB, xmmC); - e->vfmadd213ps(xmmA, xmmB, anyptr_gpC); - e->vfmadd213ps(ymmA, ymmB, ymmC); - e->vfmadd213ps(ymmA, ymmB, anyptr_gpC); - e->vfmadd213ps(zmmA, zmmB, zmmC); - e->vfmadd213ps(zmmA, zmmB, anyptr_gpC); - e->vfmadd213sd(xmmA, xmmB, xmmC); - e->vfmadd213sd(xmmA, xmmB, anyptr_gpC); - e->vfmadd213ss(xmmA, xmmB, xmmC); - e->vfmadd213ss(xmmA, xmmB, anyptr_gpC); - e->vfmadd231pd(xmmA, xmmB, xmmC); - e->vfmadd231pd(xmmA, xmmB, anyptr_gpC); - e->vfmadd231pd(ymmA, ymmB, ymmC); - e->vfmadd231pd(ymmA, ymmB, anyptr_gpC); - e->vfmadd231pd(zmmA, zmmB, zmmC); - e->vfmadd231pd(zmmA, zmmB, anyptr_gpC); - e->vfmadd231ps(xmmA, xmmB, xmmC); - e->vfmadd231ps(xmmA, xmmB, anyptr_gpC); - e->vfmadd231ps(ymmA, ymmB, ymmC); - e->vfmadd231ps(ymmA, ymmB, anyptr_gpC); - e->vfmadd231ps(zmmA, zmmB, zmmC); - e->vfmadd231ps(zmmA, zmmB, anyptr_gpC); - e->vfmadd231sd(xmmA, xmmB, xmmC); - e->vfmadd231sd(xmmA, xmmB, anyptr_gpC); - e->vfmadd231ss(xmmA, xmmB, xmmC); - e->vfmadd231ss(xmmA, xmmB, anyptr_gpC); - e->vfmaddsub132pd(xmmA, xmmB, xmmC); - e->vfmaddsub132pd(xmmA, xmmB, anyptr_gpC); - e->vfmaddsub132pd(ymmA, ymmB, ymmC); - e->vfmaddsub132pd(ymmA, ymmB, anyptr_gpC); - e->vfmaddsub132pd(zmmA, zmmB, zmmC); - e->vfmaddsub132pd(zmmA, zmmB, anyptr_gpC); - e->vfmaddsub132ps(xmmA, xmmB, xmmC); - e->vfmaddsub132ps(xmmA, xmmB, anyptr_gpC); - e->vfmaddsub132ps(ymmA, ymmB, ymmC); - e->vfmaddsub132ps(ymmA, ymmB, anyptr_gpC); - e->vfmaddsub132ps(zmmA, zmmB, zmmC); - e->vfmaddsub132ps(zmmA, zmmB, anyptr_gpC); - e->vfmaddsub213pd(xmmA, xmmB, xmmC); - e->vfmaddsub213pd(xmmA, xmmB, anyptr_gpC); - e->vfmaddsub213pd(ymmA, ymmB, ymmC); - e->vfmaddsub213pd(ymmA, ymmB, anyptr_gpC); - e->vfmaddsub213pd(zmmA, zmmB, zmmC); - e->vfmaddsub213pd(zmmA, zmmB, anyptr_gpC); - e->vfmaddsub213ps(xmmA, xmmB, xmmC); - e->vfmaddsub213ps(xmmA, xmmB, anyptr_gpC); - e->vfmaddsub213ps(ymmA, ymmB, ymmC); - e->vfmaddsub213ps(ymmA, ymmB, anyptr_gpC); - e->vfmaddsub213ps(zmmA, zmmB, zmmC); - e->vfmaddsub213ps(zmmA, zmmB, anyptr_gpC); - e->vfmaddsub231pd(xmmA, xmmB, xmmC); - e->vfmaddsub231pd(xmmA, xmmB, anyptr_gpC); - e->vfmaddsub231pd(ymmA, ymmB, ymmC); - e->vfmaddsub231pd(ymmA, ymmB, anyptr_gpC); - e->vfmaddsub231pd(zmmA, zmmB, zmmC); - e->vfmaddsub231pd(zmmA, zmmB, anyptr_gpC); - e->vfmaddsub231ps(xmmA, xmmB, xmmC); - e->vfmaddsub231ps(xmmA, xmmB, anyptr_gpC); - e->vfmaddsub231ps(ymmA, ymmB, ymmC); - e->vfmaddsub231ps(ymmA, ymmB, anyptr_gpC); - e->vfmaddsub231ps(zmmA, zmmB, zmmC); - e->vfmaddsub231ps(zmmA, zmmB, anyptr_gpC); - e->vfmsub132pd(xmmA, xmmB, xmmC); - e->vfmsub132pd(xmmA, xmmB, anyptr_gpC); - e->vfmsub132pd(ymmA, ymmB, ymmC); - e->vfmsub132pd(ymmA, ymmB, anyptr_gpC); - e->vfmsub132pd(zmmA, zmmB, zmmC); - e->vfmsub132pd(zmmA, zmmB, anyptr_gpC); - e->vfmsub132ps(xmmA, xmmB, xmmC); - e->vfmsub132ps(xmmA, xmmB, anyptr_gpC); - e->vfmsub132ps(ymmA, ymmB, ymmC); - e->vfmsub132ps(ymmA, ymmB, anyptr_gpC); - e->vfmsub132ps(zmmA, zmmB, zmmC); - e->vfmsub132ps(zmmA, zmmB, anyptr_gpC); - e->vfmsub132sd(xmmA, xmmB, xmmC); - e->vfmsub132sd(xmmA, xmmB, anyptr_gpC); - e->vfmsub132ss(xmmA, xmmB, xmmC); - e->vfmsub132ss(xmmA, xmmB, anyptr_gpC); - e->vfmsub213pd(xmmA, xmmB, xmmC); - e->vfmsub213pd(xmmA, xmmB, anyptr_gpC); - e->vfmsub213pd(ymmA, ymmB, ymmC); - e->vfmsub213pd(ymmA, ymmB, anyptr_gpC); - e->vfmsub213pd(zmmA, zmmB, zmmC); - e->vfmsub213pd(zmmA, zmmB, anyptr_gpC); - e->vfmsub213ps(xmmA, xmmB, xmmC); - e->vfmsub213ps(xmmA, xmmB, anyptr_gpC); - e->vfmsub213ps(ymmA, ymmB, ymmC); - e->vfmsub213ps(ymmA, ymmB, anyptr_gpC); - e->vfmsub213ps(zmmA, zmmB, zmmC); - e->vfmsub213ps(zmmA, zmmB, anyptr_gpC); - e->vfmsub213sd(xmmA, xmmB, xmmC); - e->vfmsub213sd(xmmA, xmmB, anyptr_gpC); - e->vfmsub213ss(xmmA, xmmB, xmmC); - e->vfmsub213ss(xmmA, xmmB, anyptr_gpC); - e->vfmsub231pd(xmmA, xmmB, xmmC); - e->vfmsub231pd(xmmA, xmmB, anyptr_gpC); - e->vfmsub231pd(ymmA, ymmB, ymmC); - e->vfmsub231pd(ymmA, ymmB, anyptr_gpC); - e->vfmsub231pd(zmmA, zmmB, zmmC); - e->vfmsub231pd(zmmA, zmmB, anyptr_gpC); - e->vfmsub231ps(xmmA, xmmB, xmmC); - e->vfmsub231ps(xmmA, xmmB, anyptr_gpC); - e->vfmsub231ps(ymmA, ymmB, ymmC); - e->vfmsub231ps(ymmA, ymmB, anyptr_gpC); - e->vfmsub231ps(zmmA, zmmB, zmmC); - e->vfmsub231ps(zmmA, zmmB, anyptr_gpC); - e->vfmsub231sd(xmmA, xmmB, xmmC); - e->vfmsub231sd(xmmA, xmmB, anyptr_gpC); - e->vfmsub231ss(xmmA, xmmB, xmmC); - e->vfmsub231ss(xmmA, xmmB, anyptr_gpC); - e->vfmsubadd132pd(xmmA, xmmB, xmmC); - e->vfmsubadd132pd(xmmA, xmmB, anyptr_gpC); - e->vfmsubadd132pd(ymmA, ymmB, ymmC); - e->vfmsubadd132pd(ymmA, ymmB, anyptr_gpC); - e->vfmsubadd132pd(zmmA, zmmB, zmmC); - e->vfmsubadd132pd(zmmA, zmmB, anyptr_gpC); - e->vfmsubadd132ps(xmmA, xmmB, xmmC); - e->vfmsubadd132ps(xmmA, xmmB, anyptr_gpC); - e->vfmsubadd132ps(ymmA, ymmB, ymmC); - e->vfmsubadd132ps(ymmA, ymmB, anyptr_gpC); - e->vfmsubadd132ps(zmmA, zmmB, zmmC); - e->vfmsubadd132ps(zmmA, zmmB, anyptr_gpC); - e->vfmsubadd213pd(xmmA, xmmB, xmmC); - e->vfmsubadd213pd(xmmA, xmmB, anyptr_gpC); - e->vfmsubadd213pd(ymmA, ymmB, ymmC); - e->vfmsubadd213pd(ymmA, ymmB, anyptr_gpC); - e->vfmsubadd213pd(zmmA, zmmB, zmmC); - e->vfmsubadd213pd(zmmA, zmmB, anyptr_gpC); - e->vfmsubadd213ps(xmmA, xmmB, xmmC); - e->vfmsubadd213ps(xmmA, xmmB, anyptr_gpC); - e->vfmsubadd213ps(ymmA, ymmB, ymmC); - e->vfmsubadd213ps(ymmA, ymmB, anyptr_gpC); - e->vfmsubadd213ps(zmmA, zmmB, zmmC); - e->vfmsubadd213ps(zmmA, zmmB, anyptr_gpC); - e->vfmsubadd231pd(xmmA, xmmB, xmmC); - e->vfmsubadd231pd(xmmA, xmmB, anyptr_gpC); - e->vfmsubadd231pd(ymmA, ymmB, ymmC); - e->vfmsubadd231pd(ymmA, ymmB, anyptr_gpC); - e->vfmsubadd231pd(zmmA, zmmB, zmmC); - e->vfmsubadd231pd(zmmA, zmmB, anyptr_gpC); - e->vfmsubadd231ps(xmmA, xmmB, xmmC); - e->vfmsubadd231ps(xmmA, xmmB, anyptr_gpC); - e->vfmsubadd231ps(ymmA, ymmB, ymmC); - e->vfmsubadd231ps(ymmA, ymmB, anyptr_gpC); - e->vfmsubadd231ps(zmmA, zmmB, zmmC); - e->vfmsubadd231ps(zmmA, zmmB, anyptr_gpC); - e->vfnmadd132pd(xmmA, xmmB, xmmC); - e->vfnmadd132pd(xmmA, xmmB, anyptr_gpC); - e->vfnmadd132pd(ymmA, ymmB, ymmC); - e->vfnmadd132pd(ymmA, ymmB, anyptr_gpC); - e->vfnmadd132pd(zmmA, zmmB, zmmC); - e->vfnmadd132pd(zmmA, zmmB, anyptr_gpC); - e->vfnmadd132ps(xmmA, xmmB, xmmC); - e->vfnmadd132ps(xmmA, xmmB, anyptr_gpC); - e->vfnmadd132ps(ymmA, ymmB, ymmC); - e->vfnmadd132ps(ymmA, ymmB, anyptr_gpC); - e->vfnmadd132ps(zmmA, zmmB, zmmC); - e->vfnmadd132ps(zmmA, zmmB, anyptr_gpC); - e->vfnmadd132sd(xmmA, xmmB, xmmC); - e->vfnmadd132sd(xmmA, xmmB, anyptr_gpC); - e->vfnmadd132ss(xmmA, xmmB, xmmC); - e->vfnmadd132ss(xmmA, xmmB, anyptr_gpC); - e->vfnmadd213pd(xmmA, xmmB, xmmC); - e->vfnmadd213pd(xmmA, xmmB, anyptr_gpC); - e->vfnmadd213pd(ymmA, ymmB, ymmC); - e->vfnmadd213pd(ymmA, ymmB, anyptr_gpC); - e->vfnmadd213pd(zmmA, zmmB, zmmC); - e->vfnmadd213pd(zmmA, zmmB, anyptr_gpC); - e->vfnmadd213ps(xmmA, xmmB, xmmC); - e->vfnmadd213ps(xmmA, xmmB, anyptr_gpC); - e->vfnmadd213ps(ymmA, ymmB, ymmC); - e->vfnmadd213ps(ymmA, ymmB, anyptr_gpC); - e->vfnmadd213ps(zmmA, zmmB, zmmC); - e->vfnmadd213ps(zmmA, zmmB, anyptr_gpC); - e->vfnmadd213sd(xmmA, xmmB, xmmC); - e->vfnmadd213sd(xmmA, xmmB, anyptr_gpC); - e->vfnmadd213ss(xmmA, xmmB, xmmC); - e->vfnmadd213ss(xmmA, xmmB, anyptr_gpC); - e->vfnmadd231pd(xmmA, xmmB, xmmC); - e->vfnmadd231pd(xmmA, xmmB, anyptr_gpC); - e->vfnmadd231pd(ymmA, ymmB, ymmC); - e->vfnmadd231pd(ymmA, ymmB, anyptr_gpC); - e->vfnmadd231pd(zmmA, zmmB, zmmC); - e->vfnmadd231pd(zmmA, zmmB, anyptr_gpC); - e->vfnmadd231ps(xmmA, xmmB, xmmC); - e->vfnmadd231ps(xmmA, xmmB, anyptr_gpC); - e->vfnmadd231ps(ymmA, ymmB, ymmC); - e->vfnmadd231ps(ymmA, ymmB, anyptr_gpC); - e->vfnmadd231ps(zmmA, zmmB, zmmC); - e->vfnmadd231ps(zmmA, zmmB, anyptr_gpC); - e->vfnmadd231sd(xmmA, xmmB, xmmC); - e->vfnmadd231sd(xmmA, xmmB, anyptr_gpC); - e->vfnmadd231ss(xmmA, xmmB, xmmC); - e->vfnmadd231ss(xmmA, xmmB, anyptr_gpC); - e->vfnmsub132pd(xmmA, xmmB, xmmC); - e->vfnmsub132pd(xmmA, xmmB, anyptr_gpC); - e->vfnmsub132pd(ymmA, ymmB, ymmC); - e->vfnmsub132pd(ymmA, ymmB, anyptr_gpC); - e->vfnmsub132pd(zmmA, zmmB, zmmC); - e->vfnmsub132pd(zmmA, zmmB, anyptr_gpC); - e->vfnmsub132ps(xmmA, xmmB, xmmC); - e->vfnmsub132ps(xmmA, xmmB, anyptr_gpC); - e->vfnmsub132ps(ymmA, ymmB, ymmC); - e->vfnmsub132ps(ymmA, ymmB, anyptr_gpC); - e->vfnmsub132ps(zmmA, zmmB, zmmC); - e->vfnmsub132ps(zmmA, zmmB, anyptr_gpC); - e->vfnmsub132sd(xmmA, xmmB, xmmC); - e->vfnmsub132sd(xmmA, xmmB, anyptr_gpC); - e->vfnmsub132ss(xmmA, xmmB, xmmC); - e->vfnmsub132ss(xmmA, xmmB, anyptr_gpC); - e->vfnmsub213pd(xmmA, xmmB, xmmC); - e->vfnmsub213pd(xmmA, xmmB, anyptr_gpC); - e->vfnmsub213pd(ymmA, ymmB, ymmC); - e->vfnmsub213pd(ymmA, ymmB, anyptr_gpC); - e->vfnmsub213pd(zmmA, zmmB, zmmC); - e->vfnmsub213pd(zmmA, zmmB, anyptr_gpC); - e->vfnmsub213ps(xmmA, xmmB, xmmC); - e->vfnmsub213ps(xmmA, xmmB, anyptr_gpC); - e->vfnmsub213ps(ymmA, ymmB, ymmC); - e->vfnmsub213ps(ymmA, ymmB, anyptr_gpC); - e->vfnmsub213ps(zmmA, zmmB, zmmC); - e->vfnmsub213ps(zmmA, zmmB, anyptr_gpC); - e->vfnmsub213sd(xmmA, xmmB, xmmC); - e->vfnmsub213sd(xmmA, xmmB, anyptr_gpC); - e->vfnmsub213ss(xmmA, xmmB, xmmC); - e->vfnmsub213ss(xmmA, xmmB, anyptr_gpC); - e->vfnmsub231pd(xmmA, xmmB, xmmC); - e->vfnmsub231pd(xmmA, xmmB, anyptr_gpC); - e->vfnmsub231pd(ymmA, ymmB, ymmC); - e->vfnmsub231pd(ymmA, ymmB, anyptr_gpC); - e->vfnmsub231pd(zmmA, zmmB, zmmC); - e->vfnmsub231pd(zmmA, zmmB, anyptr_gpC); - e->vfnmsub231ps(xmmA, xmmB, xmmC); - e->vfnmsub231ps(xmmA, xmmB, anyptr_gpC); - e->vfnmsub231ps(ymmA, ymmB, ymmC); - e->vfnmsub231ps(ymmA, ymmB, anyptr_gpC); - e->vfnmsub231ps(zmmA, zmmB, zmmC); - e->vfnmsub231ps(zmmA, zmmB, anyptr_gpC); - e->vfnmsub231sd(xmmA, xmmB, xmmC); - e->vfnmsub231sd(xmmA, xmmB, anyptr_gpC); - e->vfnmsub231ss(xmmA, xmmB, xmmC); - e->vfnmsub231ss(xmmA, xmmB, anyptr_gpC); - e->vfpclasspd(kA, xmmB, 0); - e->vfpclasspd(kA, anyptr_gpB, 0); - e->vfpclasspd(kA, ymmB, 0); - e->vfpclasspd(kA, anyptr_gpB, 0); - e->vfpclasspd(kA, zmmB, 0); - e->vfpclasspd(kA, anyptr_gpB, 0); - e->vfpclassps(kA, xmmB, 0); - e->vfpclassps(kA, anyptr_gpB, 0); - e->vfpclassps(kA, ymmB, 0); - e->vfpclassps(kA, anyptr_gpB, 0); - e->vfpclassps(kA, zmmB, 0); - e->vfpclassps(kA, anyptr_gpB, 0); - e->vfpclasssd(kA, xmmB, 0); - e->vfpclasssd(kA, anyptr_gpB, 0); - e->vfpclassss(kA, xmmB, 0); - e->vfpclassss(kA, anyptr_gpB, 0); - e->vgatherdpd(xmmA, vx_ptr); - e->vgatherdpd(ymmA, vy_ptr); - e->vgatherdpd(zmmA, vz_ptr); - e->vgatherdps(xmmA, vx_ptr); - e->vgatherdps(ymmA, vy_ptr); - e->vgatherdps(zmmA, vz_ptr); - e->vgatherpf0dpd(vy_ptr); - e->vgatherpf0dps(vz_ptr); - e->vgatherpf0qpd(vz_ptr); - e->vgatherpf0qps(vz_ptr); - e->vgatherpf1dpd(vy_ptr); - e->vgatherpf1dps(vz_ptr); - e->vgatherpf1qpd(vz_ptr); - e->vgatherpf1qps(vz_ptr); - e->vgatherqpd(xmmA, vx_ptr); - e->vgatherqpd(ymmA, vy_ptr); - e->vgatherqpd(zmmA, vz_ptr); - e->vgatherqps(xmmA, vx_ptr); - e->vgatherqps(ymmA, vy_ptr); - e->vgatherqps(zmmA, vz_ptr); - e->vgetexppd(xmmA, xmmB); - e->vgetexppd(xmmA, anyptr_gpB); - e->vgetexppd(ymmA, ymmB); - e->vgetexppd(ymmA, anyptr_gpB); - e->vgetexppd(zmmA, zmmB); - e->vgetexppd(zmmA, anyptr_gpB); - e->vgetexpps(xmmA, xmmB); - e->vgetexpps(xmmA, anyptr_gpB); - e->vgetexpps(ymmA, ymmB); - e->vgetexpps(ymmA, anyptr_gpB); - e->vgetexpps(zmmA, zmmB); - e->vgetexpps(zmmA, anyptr_gpB); - e->vgetexpsd(xmmA, xmmB, xmmC); - e->vgetexpsd(xmmA, xmmB, anyptr_gpB); - e->vgetexpss(xmmA, xmmB, xmmC); - e->vgetexpss(xmmA, xmmB, anyptr_gpB); - e->vgetmantpd(xmmA, xmmB, 0); - e->vgetmantpd(xmmA, anyptr_gpB, 0); - e->vgetmantpd(ymmA, ymmB, 0); - e->vgetmantpd(ymmA, anyptr_gpB, 0); - e->vgetmantpd(zmmA, zmmB, 0); - e->vgetmantpd(zmmA, anyptr_gpB, 0); - e->vgetmantps(xmmA, xmmB, 0); - e->vgetmantps(xmmA, anyptr_gpB, 0); - e->vgetmantps(ymmA, ymmB, 0); - e->vgetmantps(ymmA, anyptr_gpB, 0); - e->vgetmantps(zmmA, zmmB, 0); - e->vgetmantps(zmmA, anyptr_gpB, 0); - e->vgetmantsd(xmmA, xmmB, xmmC, 0); - e->vgetmantsd(xmmA, xmmB, anyptr_gpB, 0); - e->vgetmantss(xmmA, xmmB, xmmC, 0); - e->vgetmantss(xmmA, xmmB, anyptr_gpB, 0); - e->vinsertf32x4(ymmA, ymmB, xmmC, 0); - e->vinsertf32x4(ymmA, ymmB, anyptr_gpC, 0); - e->vinsertf32x4(zmmA, zmmB, xmmC, 0); - e->vinsertf32x4(zmmA, zmmB, anyptr_gpC, 0); - e->vinsertf32x8(zmmA, zmmB, ymmC, 0); - e->vinsertf32x8(zmmA, zmmB, anyptr_gpC, 0); - e->vinsertf64x2(ymmA, ymmB, xmmC, 0); - e->vinsertf64x2(ymmA, ymmB, anyptr_gpC, 0); - e->vinsertf64x2(zmmA, zmmB, xmmC, 0); - e->vinsertf64x2(zmmA, zmmB, anyptr_gpC, 0); - e->vinsertf64x4(zmmA, zmmB, ymmC, 0); - e->vinsertf64x4(zmmA, zmmB, anyptr_gpC, 0); - e->vinserti32x4(ymmA, ymmB, xmmC, 0); - e->vinserti32x4(ymmA, ymmB, anyptr_gpC, 0); - e->vinserti32x4(zmmA, zmmB, xmmC, 0); - e->vinserti32x4(zmmA, zmmB, anyptr_gpC, 0); - e->vinserti32x8(zmmA, zmmB, ymmC, 0); - e->vinserti32x8(zmmA, zmmB, anyptr_gpC, 0); - e->vinserti64x2(ymmA, ymmB, xmmC, 0); - e->vinserti64x2(ymmA, ymmB, anyptr_gpC, 0); - e->vinserti64x2(zmmA, zmmB, xmmC, 0); - e->vinserti64x2(zmmA, zmmB, anyptr_gpC, 0); - e->vinserti64x4(zmmA, zmmB, ymmC, 0); - e->vinserti64x4(zmmA, zmmB, anyptr_gpC, 0); - e->vinsertps(xmmA, xmmB, xmmC, 0); - e->vinsertps(xmmA, xmmB, anyptr_gpC, 0); - e->vmaxpd(xmmA, xmmB, xmmC); - e->vmaxpd(xmmA, xmmB, anyptr_gpC); - e->vmaxpd(ymmA, ymmB, ymmC); - e->vmaxpd(ymmA, ymmB, anyptr_gpC); - e->vmaxpd(zmmA, zmmB, zmmC); - e->vmaxpd(zmmA, zmmB, anyptr_gpC); - e->vmaxps(xmmA, xmmB, xmmC); - e->vmaxps(xmmA, xmmB, anyptr_gpC); - e->vmaxps(ymmA, ymmB, ymmC); - e->vmaxps(ymmA, ymmB, anyptr_gpC); - e->vmaxps(zmmA, zmmB, zmmC); - e->vmaxps(zmmA, zmmB, anyptr_gpC); - e->vmaxsd(xmmA, xmmB, xmmC); - e->vmaxsd(xmmA, xmmB, anyptr_gpC); - e->vmaxss(xmmA, xmmB, xmmC); - e->vmaxss(xmmA, xmmB, anyptr_gpC); - e->vminpd(xmmA, xmmB, xmmC); - e->vminpd(xmmA, xmmB, anyptr_gpC); - e->vminpd(ymmA, ymmB, ymmC); - e->vminpd(ymmA, ymmB, anyptr_gpC); - e->vminpd(zmmA, zmmB, zmmC); - e->vminpd(zmmA, zmmB, anyptr_gpC); - e->vminps(xmmA, xmmB, xmmC); - e->vminps(xmmA, xmmB, anyptr_gpC); - e->vminps(ymmA, ymmB, ymmC); - e->vminps(ymmA, ymmB, anyptr_gpC); - e->vminps(zmmA, zmmB, zmmC); - e->vminps(zmmA, zmmB, anyptr_gpC); - e->vminsd(xmmA, xmmB, xmmC); - e->vminsd(xmmA, xmmB, anyptr_gpC); - e->vminss(xmmA, xmmB, xmmC); - e->vminss(xmmA, xmmB, anyptr_gpC); - e->vmovapd(xmmA, xmmB); - e->vmovapd(xmmA, anyptr_gpB); - e->vmovapd(xmmA, xmmB); - e->vmovapd(anyptr_gpA, xmmB); - e->vmovapd(ymmA, ymmB); - e->vmovapd(ymmA, anyptr_gpB); - e->vmovapd(ymmA, ymmB); - e->vmovapd(anyptr_gpA, ymmB); - e->vmovapd(zmmA, zmmB); - e->vmovapd(zmmA, anyptr_gpB); - e->vmovapd(zmmA, zmmB); - e->vmovapd(anyptr_gpA, zmmB); - e->vmovaps(xmmA, xmmB); - e->vmovaps(xmmA, anyptr_gpB); - e->vmovaps(xmmA, xmmB); - e->vmovaps(anyptr_gpA, xmmB); - e->vmovaps(ymmA, ymmB); - e->vmovaps(ymmA, anyptr_gpB); - e->vmovaps(ymmA, ymmB); - e->vmovaps(anyptr_gpA, ymmB); - e->vmovaps(zmmA, zmmB); - e->vmovaps(zmmA, anyptr_gpB); - e->vmovaps(zmmA, zmmB); - e->vmovaps(anyptr_gpA, zmmB); - e->vmovd(gdA, xmmB); - e->vmovd(gzA, xmmB); - e->vmovd(anyptr_gpA, xmmB); - e->vmovd(xmmA, gdB); - e->vmovd(xmmA, gzB); - e->vmovd(xmmA, anyptr_gpB); - e->vmovddup(xmmA, xmmB); - e->vmovddup(xmmA, anyptr_gpB); - e->vmovddup(ymmA, ymmB); - e->vmovddup(ymmA, anyptr_gpB); - e->vmovddup(zmmA, zmmB); - e->vmovddup(zmmA, anyptr_gpB); - e->vmovdqa32(xmmA, xmmB); - e->vmovdqa32(xmmA, anyptr_gpB); - e->vmovdqa32(xmmA, xmmB); - e->vmovdqa32(anyptr_gpA, xmmB); - e->vmovdqa32(ymmA, ymmB); - e->vmovdqa32(ymmA, anyptr_gpB); - e->vmovdqa32(ymmA, ymmB); - e->vmovdqa32(anyptr_gpA, ymmB); - e->vmovdqa32(zmmA, zmmB); - e->vmovdqa32(zmmA, anyptr_gpB); - e->vmovdqa32(zmmA, zmmB); - e->vmovdqa32(anyptr_gpA, zmmB); - e->vmovdqa64(xmmA, xmmB); - e->vmovdqa64(xmmA, anyptr_gpB); - e->vmovdqa64(xmmA, xmmB); - e->vmovdqa64(anyptr_gpA, xmmB); - e->vmovdqa64(ymmA, ymmB); - e->vmovdqa64(ymmA, anyptr_gpB); - e->vmovdqa64(ymmA, ymmB); - e->vmovdqa64(anyptr_gpA, ymmB); - e->vmovdqa64(zmmA, zmmB); - e->vmovdqa64(zmmA, anyptr_gpB); - e->vmovdqa64(zmmA, zmmB); - e->vmovdqa64(anyptr_gpA, zmmB); - e->vmovdqu16(xmmA, xmmB); - e->vmovdqu16(xmmA, anyptr_gpB); - e->vmovdqu16(xmmA, xmmB); - e->vmovdqu16(anyptr_gpA, xmmB); - e->vmovdqu16(ymmA, ymmB); - e->vmovdqu16(ymmA, anyptr_gpB); - e->vmovdqu16(ymmA, ymmB); - e->vmovdqu16(anyptr_gpA, ymmB); - e->vmovdqu16(zmmA, zmmB); - e->vmovdqu16(zmmA, anyptr_gpB); - e->vmovdqu16(zmmA, zmmB); - e->vmovdqu16(anyptr_gpA, zmmB); - e->vmovdqu32(xmmA, xmmB); - e->vmovdqu32(xmmA, anyptr_gpB); - e->vmovdqu32(xmmA, xmmB); - e->vmovdqu32(anyptr_gpA, xmmB); - e->vmovdqu32(ymmA, ymmB); - e->vmovdqu32(ymmA, anyptr_gpB); - e->vmovdqu32(ymmA, ymmB); - e->vmovdqu32(anyptr_gpA, ymmB); - e->vmovdqu32(zmmA, zmmB); - e->vmovdqu32(zmmA, anyptr_gpB); - e->vmovdqu32(zmmA, zmmB); - e->vmovdqu32(anyptr_gpA, zmmB); - e->vmovdqu64(xmmA, xmmB); - e->vmovdqu64(xmmA, anyptr_gpB); - e->vmovdqu64(xmmA, xmmB); - e->vmovdqu64(anyptr_gpA, xmmB); - e->vmovdqu64(ymmA, ymmB); - e->vmovdqu64(ymmA, anyptr_gpB); - e->vmovdqu64(ymmA, ymmB); - e->vmovdqu64(anyptr_gpA, ymmB); - e->vmovdqu64(zmmA, zmmB); - e->vmovdqu64(zmmA, anyptr_gpB); - e->vmovdqu64(zmmA, zmmB); - e->vmovdqu64(anyptr_gpA, zmmB); - e->vmovdqu8(xmmA, xmmB); - e->vmovdqu8(xmmA, anyptr_gpB); - e->vmovdqu8(xmmA, xmmB); - e->vmovdqu8(anyptr_gpA, xmmB); - e->vmovdqu8(ymmA, ymmB); - e->vmovdqu8(ymmA, anyptr_gpB); - e->vmovdqu8(ymmA, ymmB); - e->vmovdqu8(anyptr_gpA, ymmB); - e->vmovdqu8(zmmA, zmmB); - e->vmovdqu8(zmmA, anyptr_gpB); - e->vmovdqu8(zmmA, zmmB); - e->vmovdqu8(anyptr_gpA, zmmB); - e->vmovhlps(xmmA, xmmB, xmmC); - e->vmovhpd(anyptr_gpA, xmmB); - e->vmovhpd(xmmA, xmmB, anyptr_gpC); - e->vmovhps(anyptr_gpA, xmmB); - e->vmovhps(xmmA, xmmB, anyptr_gpC); - e->vmovlhps(xmmA, xmmB, xmmC); - e->vmovlpd(anyptr_gpA, xmmB); - e->vmovlpd(xmmA, xmmB, anyptr_gpC); - e->vmovlps(anyptr_gpA, xmmB); - e->vmovlps(xmmA, xmmB, anyptr_gpC); - e->vmovntdq(anyptr_gpA, xmmB); - e->vmovntdq(anyptr_gpA, ymmB); - e->vmovntdq(anyptr_gpA, zmmB); - e->vmovntdqa(xmmA, anyptr_gpB); - e->vmovntdqa(ymmA, anyptr_gpB); - e->vmovntdqa(zmmA, anyptr_gpB); - e->vmovntpd(anyptr_gpA, xmmB); - e->vmovntpd(anyptr_gpA, ymmB); - e->vmovntpd(anyptr_gpA, zmmB); - e->vmovntps(anyptr_gpA, xmmB); - e->vmovntps(anyptr_gpA, ymmB); - e->vmovntps(anyptr_gpA, zmmB); - if (isX64) e->vmovq(gzA, xmmB); - if (isX64) e->vmovq(xmmA, gzB); - e->vmovq(anyptr_gpA, xmmB); - e->vmovq(xmmA, anyptr_gpB); - e->vmovq(xmmA, xmmB); - e->vmovq(xmmA, anyptr_gpB); - e->vmovq(xmmA, xmmB); - e->vmovq(anyptr_gpA, xmmB); - e->vmovsd(anyptr_gpA, xmmB); - e->vmovsd(xmmA, anyptr_gpB); - e->vmovsd(xmmA, xmmB, xmmC); - e->vmovsd(xmmA, xmmB, xmmC); - e->vmovshdup(xmmA, xmmB); - e->vmovshdup(xmmA, anyptr_gpB); - e->vmovshdup(ymmA, ymmB); - e->vmovshdup(ymmA, anyptr_gpB); - e->vmovshdup(zmmA, zmmB); - e->vmovshdup(zmmA, anyptr_gpB); - e->vmovsldup(xmmA, xmmB); - e->vmovsldup(xmmA, anyptr_gpB); - e->vmovsldup(ymmA, ymmB); - e->vmovsldup(ymmA, anyptr_gpB); - e->vmovsldup(zmmA, zmmB); - e->vmovsldup(zmmA, anyptr_gpB); - e->vmovss(anyptr_gpA, xmmB); - e->vmovss(xmmA, anyptr_gpB); - e->vmovss(xmmA, xmmB, xmmC); - e->vmovss(xmmA, xmmB, xmmC); - e->vmovupd(xmmA, xmmB); - e->vmovupd(xmmA, anyptr_gpB); - e->vmovupd(xmmA, xmmB); - e->vmovupd(anyptr_gpA, xmmB); - e->vmovupd(ymmA, ymmB); - e->vmovupd(ymmA, anyptr_gpB); - e->vmovupd(ymmA, ymmB); - e->vmovupd(anyptr_gpA, ymmB); - e->vmovupd(zmmA, zmmB); - e->vmovupd(zmmA, anyptr_gpB); - e->vmovupd(zmmA, zmmB); - e->vmovupd(anyptr_gpA, zmmB); - e->vmovups(xmmA, xmmB); - e->vmovups(xmmA, anyptr_gpB); - e->vmovups(xmmA, xmmB); - e->vmovups(anyptr_gpA, xmmB); - e->vmovups(ymmA, ymmB); - e->vmovups(ymmA, anyptr_gpB); - e->vmovups(ymmA, ymmB); - e->vmovups(anyptr_gpA, ymmB); - e->vmovups(zmmA, zmmB); - e->vmovups(zmmA, anyptr_gpB); - e->vmovups(zmmA, zmmB); - e->vmovups(anyptr_gpA, zmmB); - e->vmulpd(xmmA, xmmB, xmmC); - e->vmulpd(xmmA, xmmB, anyptr_gpC); - e->vmulpd(ymmA, ymmB, ymmC); - e->vmulpd(ymmA, ymmB, anyptr_gpC); - e->vmulpd(zmmA, zmmB, zmmC); - e->vmulpd(zmmA, zmmB, anyptr_gpC); - e->vmulps(xmmA, xmmB, xmmC); - e->vmulps(xmmA, xmmB, anyptr_gpC); - e->vmulps(ymmA, ymmB, ymmC); - e->vmulps(ymmA, ymmB, anyptr_gpC); - e->vmulps(zmmA, zmmB, zmmC); - e->vmulps(zmmA, zmmB, anyptr_gpC); - e->vmulsd(xmmA, xmmB, xmmC); - e->vmulsd(xmmA, xmmB, anyptr_gpC); - e->vmulss(xmmA, xmmB, xmmC); - e->vmulss(xmmA, xmmB, anyptr_gpC); - e->vorpd(xmmA, xmmB, xmmC); - e->vorpd(xmmA, xmmB, anyptr_gpC); - e->vorpd(ymmA, ymmB, ymmC); - e->vorpd(ymmA, ymmB, anyptr_gpC); - e->vorpd(zmmA, zmmB, zmmC); - e->vorpd(zmmA, zmmB, anyptr_gpC); - e->vorps(xmmA, xmmB, xmmC); - e->vorps(xmmA, xmmB, anyptr_gpC); - e->vorps(ymmA, ymmB, ymmC); - e->vorps(ymmA, ymmB, anyptr_gpC); - e->vorps(zmmA, zmmB, zmmC); - e->vorps(zmmA, zmmB, anyptr_gpC); - e->vpabsb(xmmA, xmmB); - e->vpabsb(xmmA, anyptr_gpB); - e->vpabsb(ymmA, ymmB); - e->vpabsb(ymmA, anyptr_gpB); - e->vpabsb(zmmA, zmmB); - e->vpabsb(zmmA, anyptr_gpB); - e->vpabsd(xmmA, xmmB); - e->vpabsd(xmmA, anyptr_gpB); - e->vpabsd(ymmA, ymmB); - e->vpabsd(ymmA, anyptr_gpB); - e->vpabsd(zmmA, zmmB); - e->vpabsd(zmmA, anyptr_gpB); - e->vpabsq(xmmA, xmmB); - e->vpabsq(xmmA, anyptr_gpB); - e->vpabsq(ymmA, ymmB); - e->vpabsq(ymmA, anyptr_gpB); - e->vpabsq(zmmA, zmmB); - e->vpabsq(zmmA, anyptr_gpB); - e->vpabsw(xmmA, xmmB); - e->vpabsw(xmmA, anyptr_gpB); - e->vpabsw(ymmA, ymmB); - e->vpabsw(ymmA, anyptr_gpB); - e->vpabsw(zmmA, zmmB); - e->vpabsw(zmmA, anyptr_gpB); - e->vpackssdw(xmmA, xmmB, xmmC); - e->vpackssdw(xmmA, xmmB, anyptr_gpC); - e->vpackssdw(ymmA, ymmB, ymmC); - e->vpackssdw(ymmA, ymmB, anyptr_gpC); - e->vpackssdw(zmmA, zmmB, zmmC); - e->vpackssdw(zmmA, zmmB, anyptr_gpC); - e->vpacksswb(xmmA, xmmB, xmmC); - e->vpacksswb(xmmA, xmmB, anyptr_gpC); - e->vpacksswb(ymmA, ymmB, ymmC); - e->vpacksswb(ymmA, ymmB, anyptr_gpC); - e->vpacksswb(zmmA, zmmB, zmmC); - e->vpacksswb(zmmA, zmmB, anyptr_gpC); - e->vpackusdw(xmmA, xmmB, xmmC); - e->vpackusdw(xmmA, xmmB, anyptr_gpC); - e->vpackusdw(ymmA, ymmB, ymmC); - e->vpackusdw(ymmA, ymmB, anyptr_gpC); - e->vpackusdw(zmmA, zmmB, zmmC); - e->vpackusdw(zmmA, zmmB, anyptr_gpC); - e->vpackuswb(xmmA, xmmB, xmmC); - e->vpackuswb(xmmA, xmmB, anyptr_gpC); - e->vpackuswb(ymmA, ymmB, ymmC); - e->vpackuswb(ymmA, ymmB, anyptr_gpC); - e->vpackuswb(zmmA, zmmB, zmmC); - e->vpackuswb(zmmA, zmmB, anyptr_gpC); - e->vpaddb(xmmA, xmmB, xmmC); - e->vpaddb(xmmA, xmmB, anyptr_gpC); - e->vpaddb(ymmA, ymmB, ymmC); - e->vpaddb(ymmA, ymmB, anyptr_gpC); - e->vpaddb(zmmA, zmmB, zmmC); - e->vpaddb(zmmA, zmmB, anyptr_gpC); - e->vpaddd(xmmA, xmmB, xmmC); - e->vpaddd(xmmA, xmmB, anyptr_gpC); - e->vpaddd(ymmA, ymmB, ymmC); - e->vpaddd(ymmA, ymmB, anyptr_gpC); - e->vpaddd(zmmA, zmmB, zmmC); - e->vpaddd(zmmA, zmmB, anyptr_gpC); - e->vpaddq(xmmA, xmmB, xmmC); - e->vpaddq(xmmA, xmmB, anyptr_gpC); - e->vpaddq(ymmA, ymmB, ymmC); - e->vpaddq(ymmA, ymmB, anyptr_gpC); - e->vpaddq(zmmA, zmmB, zmmC); - e->vpaddq(zmmA, zmmB, anyptr_gpC); - e->vpaddsb(xmmA, xmmB, xmmC); - e->vpaddsb(xmmA, xmmB, anyptr_gpC); - e->vpaddsb(ymmA, ymmB, ymmC); - e->vpaddsb(ymmA, ymmB, anyptr_gpC); - e->vpaddsb(zmmA, zmmB, zmmC); - e->vpaddsb(zmmA, zmmB, anyptr_gpC); - e->vpaddsw(xmmA, xmmB, xmmC); - e->vpaddsw(xmmA, xmmB, anyptr_gpC); - e->vpaddsw(ymmA, ymmB, ymmC); - e->vpaddsw(ymmA, ymmB, anyptr_gpC); - e->vpaddsw(zmmA, zmmB, zmmC); - e->vpaddsw(zmmA, zmmB, anyptr_gpC); - e->vpaddusb(xmmA, xmmB, xmmC); - e->vpaddusb(xmmA, xmmB, anyptr_gpC); - e->vpaddusb(ymmA, ymmB, ymmC); - e->vpaddusb(ymmA, ymmB, anyptr_gpC); - e->vpaddusb(zmmA, zmmB, zmmC); - e->vpaddusb(zmmA, zmmB, anyptr_gpC); - e->vpaddusw(xmmA, xmmB, xmmC); - e->vpaddusw(xmmA, xmmB, anyptr_gpC); - e->vpaddusw(ymmA, ymmB, ymmC); - e->vpaddusw(ymmA, ymmB, anyptr_gpC); - e->vpaddusw(zmmA, zmmB, zmmC); - e->vpaddusw(zmmA, zmmB, anyptr_gpC); - e->vpaddw(xmmA, xmmB, xmmC); - e->vpaddw(xmmA, xmmB, anyptr_gpC); - e->vpaddw(ymmA, ymmB, ymmC); - e->vpaddw(ymmA, ymmB, anyptr_gpC); - e->vpaddw(zmmA, zmmB, zmmC); - e->vpaddw(zmmA, zmmB, anyptr_gpC); - e->vpalignr(xmmA, xmmB, xmmC, 0); - e->vpalignr(xmmA, xmmB, anyptr_gpC, 0); - e->vpalignr(ymmA, ymmB, ymmC, 0); - e->vpalignr(ymmA, ymmB, anyptr_gpC, 0); - e->vpalignr(zmmA, zmmB, zmmC, 0); - e->vpalignr(zmmA, zmmB, anyptr_gpC, 0); - e->vpandd(xmmA, xmmB, xmmC); - e->vpandd(xmmA, xmmB, anyptr_gpC); - e->vpandd(ymmA, ymmB, ymmC); - e->vpandd(ymmA, ymmB, anyptr_gpC); - e->vpandd(zmmA, zmmB, zmmC); - e->vpandd(zmmA, zmmB, anyptr_gpC); - e->vpandnd(xmmA, xmmB, xmmC); - e->vpandnd(xmmA, xmmB, anyptr_gpC); - e->vpandnd(ymmA, ymmB, ymmC); - e->vpandnd(ymmA, ymmB, anyptr_gpC); - e->vpandnd(zmmA, zmmB, zmmC); - e->vpandnd(zmmA, zmmB, anyptr_gpC); - e->vpandnq(xmmA, xmmB, xmmC); - e->vpandnq(xmmA, xmmB, anyptr_gpC); - e->vpandnq(ymmA, ymmB, ymmC); - e->vpandnq(ymmA, ymmB, anyptr_gpC); - e->vpandnq(zmmA, zmmB, zmmC); - e->vpandnq(zmmA, zmmB, anyptr_gpC); - e->vpandq(xmmA, xmmB, xmmC); - e->vpandq(xmmA, xmmB, anyptr_gpC); - e->vpandq(ymmA, ymmB, ymmC); - e->vpandq(ymmA, ymmB, anyptr_gpC); - e->vpandq(zmmA, zmmB, zmmC); - e->vpandq(zmmA, zmmB, anyptr_gpC); - e->vpavgb(xmmA, xmmB, xmmC); - e->vpavgb(xmmA, xmmB, anyptr_gpC); - e->vpavgb(ymmA, ymmB, ymmC); - e->vpavgb(ymmA, ymmB, anyptr_gpC); - e->vpavgb(zmmA, zmmB, zmmC); - e->vpavgb(zmmA, zmmB, anyptr_gpC); - e->vpavgw(xmmA, xmmB, xmmC); - e->vpavgw(xmmA, xmmB, anyptr_gpC); - e->vpavgw(ymmA, ymmB, ymmC); - e->vpavgw(ymmA, ymmB, anyptr_gpC); - e->vpavgw(zmmA, zmmB, zmmC); - e->vpavgw(zmmA, zmmB, anyptr_gpC); - e->vpblendmb(xmmA, xmmB, xmmC); - e->vpblendmb(xmmA, xmmB, anyptr_gpC); - e->vpblendmb(ymmA, ymmB, ymmC); - e->vpblendmb(ymmA, ymmB, anyptr_gpC); - e->vpblendmb(zmmA, zmmB, zmmC); - e->vpblendmb(zmmA, zmmB, anyptr_gpC); - e->vpblendmd(xmmA, xmmB, xmmC); - e->vpblendmd(xmmA, xmmB, anyptr_gpC); - e->vpblendmd(ymmA, ymmB, ymmC); - e->vpblendmd(ymmA, ymmB, anyptr_gpC); - e->vpblendmd(zmmA, zmmB, zmmC); - e->vpblendmd(zmmA, zmmB, anyptr_gpC); - e->vpblendmq(xmmA, xmmB, xmmC); - e->vpblendmq(xmmA, xmmB, anyptr_gpC); - e->vpblendmq(ymmA, ymmB, ymmC); - e->vpblendmq(ymmA, ymmB, anyptr_gpC); - e->vpblendmq(zmmA, zmmB, zmmC); - e->vpblendmq(zmmA, zmmB, anyptr_gpC); - e->vpblendmw(xmmA, xmmB, xmmC); - e->vpblendmw(xmmA, xmmB, anyptr_gpC); - e->vpblendmw(ymmA, ymmB, ymmC); - e->vpblendmw(ymmA, ymmB, anyptr_gpC); - e->vpblendmw(zmmA, zmmB, zmmC); - e->vpblendmw(zmmA, zmmB, anyptr_gpC); - e->vpbroadcastb(xmmA, gdB); - e->vpbroadcastb(xmmA, gzB); - e->vpbroadcastb(xmmA, xmmB); - e->vpbroadcastb(xmmA, anyptr_gpB); - e->vpbroadcastb(ymmA, gdB); - e->vpbroadcastb(ymmA, gzB); - e->vpbroadcastb(ymmA, xmmB); - e->vpbroadcastb(ymmA, anyptr_gpB); - e->vpbroadcastb(zmmA, gdB); - e->vpbroadcastb(zmmA, gzB); - e->vpbroadcastb(zmmA, xmmB); - e->vpbroadcastb(zmmA, anyptr_gpB); - e->vpbroadcastd(xmmA, gdB); - e->vpbroadcastd(xmmA, gzB); - e->vpbroadcastd(xmmA, xmmB); - e->vpbroadcastd(xmmA, anyptr_gpB); - e->vpbroadcastd(ymmA, gdB); - e->vpbroadcastd(ymmA, gzB); - e->vpbroadcastd(ymmA, xmmB); - e->vpbroadcastd(ymmA, anyptr_gpB); - e->vpbroadcastd(zmmA, gdB); - e->vpbroadcastd(zmmA, gzB); - e->vpbroadcastd(zmmA, xmmB); - e->vpbroadcastd(zmmA, anyptr_gpB); - e->vpbroadcastmb2q(xmmA, kB); - e->vpbroadcastmb2q(ymmA, kB); - e->vpbroadcastmb2q(zmmA, kB); - e->vpbroadcastmw2d(xmmA, kB); - e->vpbroadcastmw2d(ymmA, kB); - e->vpbroadcastmw2d(zmmA, kB); - if (isX64) e->vpbroadcastq(xmmA, gzB); - e->vpbroadcastq(xmmA, xmmB); - e->vpbroadcastq(xmmA, anyptr_gpB); - if (isX64) e->vpbroadcastq(ymmA, gzB); - e->vpbroadcastq(ymmA, xmmB); - e->vpbroadcastq(ymmA, anyptr_gpB); - if (isX64) e->vpbroadcastq(zmmA, gzB); - e->vpbroadcastq(zmmA, xmmB); - e->vpbroadcastq(zmmA, anyptr_gpB); - e->vpbroadcastw(xmmA, gdB); - e->vpbroadcastw(xmmA, gzB); - e->vpbroadcastw(xmmA, xmmB); - e->vpbroadcastw(xmmA, anyptr_gpB); - e->vpbroadcastw(ymmA, gdB); - e->vpbroadcastw(ymmA, gzB); - e->vpbroadcastw(ymmA, xmmB); - e->vpbroadcastw(ymmA, anyptr_gpB); - e->vpbroadcastw(zmmA, gdB); - e->vpbroadcastw(zmmA, gzB); - e->vpbroadcastw(zmmA, xmmB); - e->vpbroadcastw(zmmA, anyptr_gpB); - e->vpcmpb(kA, xmmB, xmmC, 0); - e->vpcmpb(kA, xmmB, anyptr_gpC, 0); - e->vpcmpb(kA, ymmB, ymmC, 0); - e->vpcmpb(kA, ymmB, anyptr_gpC, 0); - e->vpcmpb(kA, zmmB, zmmC, 0); - e->vpcmpb(kA, zmmB, anyptr_gpC, 0); - e->vpcmpd(kA, xmmB, xmmC, 0); - e->vpcmpd(kA, xmmB, anyptr_gpC, 0); - e->vpcmpd(kA, ymmB, ymmC, 0); - e->vpcmpd(kA, ymmB, anyptr_gpC, 0); - e->vpcmpd(kA, zmmB, zmmC, 0); - e->vpcmpd(kA, zmmB, anyptr_gpC, 0); - e->vpcmpeqb(kA, xmmB, xmmC); - e->vpcmpeqb(kA, xmmB, anyptr_gpC); - e->vpcmpeqb(kA, ymmB, ymmC); - e->vpcmpeqb(kA, ymmB, anyptr_gpC); - e->vpcmpeqb(kA, zmmB, zmmC); - e->vpcmpeqb(kA, zmmB, anyptr_gpC); - e->vpcmpeqd(kA, xmmB, xmmC); - e->vpcmpeqd(kA, xmmB, anyptr_gpC); - e->vpcmpeqd(kA, ymmB, ymmC); - e->vpcmpeqd(kA, ymmB, anyptr_gpC); - e->vpcmpeqd(kA, zmmB, zmmC); - e->vpcmpeqd(kA, zmmB, anyptr_gpC); - e->vpcmpeqq(kA, xmmB, xmmC); - e->vpcmpeqq(kA, xmmB, anyptr_gpC); - e->vpcmpeqq(kA, ymmB, ymmC); - e->vpcmpeqq(kA, ymmB, anyptr_gpC); - e->vpcmpeqq(kA, zmmB, zmmC); - e->vpcmpeqq(kA, zmmB, anyptr_gpC); - e->vpcmpeqw(kA, xmmB, xmmC); - e->vpcmpeqw(kA, xmmB, anyptr_gpC); - e->vpcmpeqw(kA, ymmB, ymmC); - e->vpcmpeqw(kA, ymmB, anyptr_gpC); - e->vpcmpeqw(kA, zmmB, zmmC); - e->vpcmpeqw(kA, zmmB, anyptr_gpC); - e->vpcmpgtb(kA, xmmB, xmmC); - e->vpcmpgtb(kA, xmmB, anyptr_gpC); - e->vpcmpgtb(kA, ymmB, ymmC); - e->vpcmpgtb(kA, ymmB, anyptr_gpC); - e->vpcmpgtb(kA, zmmB, zmmC); - e->vpcmpgtb(kA, zmmB, anyptr_gpC); - e->vpcmpgtd(kA, xmmB, xmmC); - e->vpcmpgtd(kA, xmmB, anyptr_gpC); - e->vpcmpgtd(kA, ymmB, ymmC); - e->vpcmpgtd(kA, ymmB, anyptr_gpC); - e->vpcmpgtd(kA, zmmB, zmmC); - e->vpcmpgtd(kA, zmmB, anyptr_gpC); - e->vpcmpgtq(kA, xmmB, xmmC); - e->vpcmpgtq(kA, xmmB, anyptr_gpC); - e->vpcmpgtq(kA, ymmB, ymmC); - e->vpcmpgtq(kA, ymmB, anyptr_gpC); - e->vpcmpgtq(kA, zmmB, zmmC); - e->vpcmpgtq(kA, zmmB, anyptr_gpC); - e->vpcmpgtw(kA, xmmB, xmmC); - e->vpcmpgtw(kA, xmmB, anyptr_gpC); - e->vpcmpgtw(kA, ymmB, ymmC); - e->vpcmpgtw(kA, ymmB, anyptr_gpC); - e->vpcmpgtw(kA, zmmB, zmmC); - e->vpcmpgtw(kA, zmmB, anyptr_gpC); - e->vpcmpq(kA, xmmB, xmmC, 0); - e->vpcmpq(kA, xmmB, anyptr_gpC, 0); - e->vpcmpq(kA, ymmB, ymmC, 0); - e->vpcmpq(kA, ymmB, anyptr_gpC, 0); - e->vpcmpq(kA, zmmB, zmmC, 0); - e->vpcmpq(kA, zmmB, anyptr_gpC, 0); - e->vpcmpub(kA, xmmB, xmmC, 0); - e->vpcmpub(kA, xmmB, anyptr_gpC, 0); - e->vpcmpub(kA, ymmB, ymmC, 0); - e->vpcmpub(kA, ymmB, anyptr_gpC, 0); - e->vpcmpub(kA, zmmB, zmmC, 0); - e->vpcmpub(kA, zmmB, anyptr_gpC, 0); - e->vpcmpud(kA, xmmB, xmmC, 0); - e->vpcmpud(kA, xmmB, anyptr_gpC, 0); - e->vpcmpud(kA, ymmB, ymmC, 0); - e->vpcmpud(kA, ymmB, anyptr_gpC, 0); - e->vpcmpud(kA, zmmB, zmmC, 0); - e->vpcmpud(kA, zmmB, anyptr_gpC, 0); - e->vpcmpuq(kA, xmmB, xmmC, 0); - e->vpcmpuq(kA, xmmB, anyptr_gpC, 0); - e->vpcmpuq(kA, ymmB, ymmC, 0); - e->vpcmpuq(kA, ymmB, anyptr_gpC, 0); - e->vpcmpuq(kA, zmmB, zmmC, 0); - e->vpcmpuq(kA, zmmB, anyptr_gpC, 0); - e->vpcmpuw(kA, xmmB, xmmC, 0); - e->vpcmpuw(kA, xmmB, anyptr_gpC, 0); - e->vpcmpuw(kA, ymmB, ymmC, 0); - e->vpcmpuw(kA, ymmB, anyptr_gpC, 0); - e->vpcmpuw(kA, zmmB, zmmC, 0); - e->vpcmpuw(kA, zmmB, anyptr_gpC, 0); - e->vpcmpw(kA, xmmB, xmmC, 0); - e->vpcmpw(kA, xmmB, anyptr_gpC, 0); - e->vpcmpw(kA, ymmB, ymmC, 0); - e->vpcmpw(kA, ymmB, anyptr_gpC, 0); - e->vpcmpw(kA, zmmB, zmmC, 0); - e->vpcmpw(kA, zmmB, anyptr_gpC, 0); - e->vpcompressd(xmmA, xmmB); - e->vpcompressd(anyptr_gpA, xmmB); - e->vpcompressd(ymmA, ymmB); - e->vpcompressd(anyptr_gpA, ymmB); - e->vpcompressd(zmmA, zmmB); - e->vpcompressd(anyptr_gpA, zmmB); - e->vpcompressq(xmmA, xmmB); - e->vpcompressq(anyptr_gpA, xmmB); - e->vpcompressq(ymmA, ymmB); - e->vpcompressq(anyptr_gpA, ymmB); - e->vpcompressq(zmmA, zmmB); - e->vpcompressq(anyptr_gpA, zmmB); - e->vpconflictd(xmmA, xmmB); - e->vpconflictd(xmmA, anyptr_gpB); - e->vpconflictd(ymmA, ymmB); - e->vpconflictd(ymmA, anyptr_gpB); - e->vpconflictd(zmmA, zmmB); - e->vpconflictd(zmmA, anyptr_gpB); - e->vpconflictq(xmmA, xmmB); - e->vpconflictq(xmmA, anyptr_gpB); - e->vpconflictq(ymmA, ymmB); - e->vpconflictq(ymmA, anyptr_gpB); - e->vpconflictq(zmmA, zmmB); - e->vpconflictq(zmmA, anyptr_gpB); - e->vpermb(xmmA, xmmB, xmmC); - e->vpermb(xmmA, xmmB, anyptr_gpC); - e->vpermb(ymmA, ymmB, ymmC); - e->vpermb(ymmA, ymmB, anyptr_gpC); - e->vpermb(zmmA, zmmB, zmmC); - e->vpermb(zmmA, zmmB, anyptr_gpC); - e->vpermd(ymmA, ymmB, ymmC); - e->vpermd(ymmA, ymmB, anyptr_gpC); - e->vpermd(zmmA, zmmB, zmmC); - e->vpermd(zmmA, zmmB, anyptr_gpC); - e->vpermi2b(xmmA, xmmB, xmmC); - e->vpermi2b(xmmA, xmmB, anyptr_gpC); - e->vpermi2b(ymmA, ymmB, ymmC); - e->vpermi2b(ymmA, ymmB, anyptr_gpC); - e->vpermi2b(zmmA, zmmB, zmmC); - e->vpermi2b(zmmA, zmmB, anyptr_gpC); - e->vpermi2d(xmmA, xmmB, xmmC); - e->vpermi2d(xmmA, xmmB, anyptr_gpC); - e->vpermi2d(ymmA, ymmB, ymmC); - e->vpermi2d(ymmA, ymmB, anyptr_gpC); - e->vpermi2d(zmmA, zmmB, zmmC); - e->vpermi2d(zmmA, zmmB, anyptr_gpC); - e->vpermi2pd(xmmA, xmmB, xmmC); - e->vpermi2pd(xmmA, xmmB, anyptr_gpC); - e->vpermi2pd(ymmA, ymmB, ymmC); - e->vpermi2pd(ymmA, ymmB, anyptr_gpC); - e->vpermi2pd(zmmA, zmmB, zmmC); - e->vpermi2pd(zmmA, zmmB, anyptr_gpC); - e->vpermi2ps(xmmA, xmmB, xmmC); - e->vpermi2ps(xmmA, xmmB, anyptr_gpC); - e->vpermi2ps(ymmA, ymmB, ymmC); - e->vpermi2ps(ymmA, ymmB, anyptr_gpC); - e->vpermi2ps(zmmA, zmmB, zmmC); - e->vpermi2ps(zmmA, zmmB, anyptr_gpC); - e->vpermi2q(xmmA, xmmB, xmmC); - e->vpermi2q(xmmA, xmmB, anyptr_gpC); - e->vpermi2q(ymmA, ymmB, ymmC); - e->vpermi2q(ymmA, ymmB, anyptr_gpC); - e->vpermi2q(zmmA, zmmB, zmmC); - e->vpermi2q(zmmA, zmmB, anyptr_gpC); - e->vpermi2w(xmmA, xmmB, xmmC); - e->vpermi2w(xmmA, xmmB, anyptr_gpC); - e->vpermi2w(ymmA, ymmB, ymmC); - e->vpermi2w(ymmA, ymmB, anyptr_gpC); - e->vpermi2w(zmmA, zmmB, zmmC); - e->vpermi2w(zmmA, zmmB, anyptr_gpC); - e->vpermilpd(xmmA, xmmB, xmmC); - e->vpermilpd(xmmA, xmmB, anyptr_gpC); - e->vpermilpd(ymmA, ymmB, ymmC); - e->vpermilpd(ymmA, ymmB, anyptr_gpC); - e->vpermilpd(zmmA, zmmB, zmmC); - e->vpermilpd(zmmA, zmmB, anyptr_gpC); - e->vpermilpd(xmmA, xmmB, 0); - e->vpermilpd(xmmA, anyptr_gpB, 0); - e->vpermilpd(ymmA, ymmB, 0); - e->vpermilpd(ymmA, anyptr_gpB, 0); - e->vpermilpd(zmmA, zmmB, 0); - e->vpermilpd(zmmA, anyptr_gpB, 0); - e->vpermilps(xmmA, xmmB, xmmC); - e->vpermilps(xmmA, xmmB, anyptr_gpC); - e->vpermilps(ymmA, ymmB, ymmC); - e->vpermilps(ymmA, ymmB, anyptr_gpC); - e->vpermilps(zmmA, zmmB, zmmC); - e->vpermilps(zmmA, zmmB, anyptr_gpC); - e->vpermilps(xmmA, xmmB, 0); - e->vpermilps(xmmA, anyptr_gpB, 0); - e->vpermilps(ymmA, ymmB, 0); - e->vpermilps(ymmA, anyptr_gpB, 0); - e->vpermilps(zmmA, zmmB, 0); - e->vpermilps(zmmA, anyptr_gpB, 0); - e->vpermq(ymmA, ymmB, ymmC); - e->vpermq(ymmA, ymmB, anyptr_gpC); - e->vpermq(zmmA, zmmB, zmmC); - e->vpermq(zmmA, zmmB, anyptr_gpC); - e->vpermq(ymmA, ymmB, 0); - e->vpermq(ymmA, anyptr_gpB, 0); - e->vpermq(zmmA, zmmB, 0); - e->vpermq(zmmA, anyptr_gpB, 0); - e->vpermt2b(xmmA, xmmB, xmmC); - e->vpermt2b(xmmA, xmmB, anyptr_gpC); - e->vpermt2b(ymmA, ymmB, ymmC); - e->vpermt2b(ymmA, ymmB, anyptr_gpC); - e->vpermt2b(zmmA, zmmB, zmmC); - e->vpermt2b(zmmA, zmmB, anyptr_gpC); - e->vpermt2d(xmmA, xmmB, xmmC); - e->vpermt2d(xmmA, xmmB, anyptr_gpC); - e->vpermt2d(ymmA, ymmB, ymmC); - e->vpermt2d(ymmA, ymmB, anyptr_gpC); - e->vpermt2d(zmmA, zmmB, zmmC); - e->vpermt2d(zmmA, zmmB, anyptr_gpC); - e->vpermt2pd(xmmA, xmmB, xmmC); - e->vpermt2pd(xmmA, xmmB, anyptr_gpC); - e->vpermt2pd(ymmA, ymmB, ymmC); - e->vpermt2pd(ymmA, ymmB, anyptr_gpC); - e->vpermt2pd(zmmA, zmmB, zmmC); - e->vpermt2pd(zmmA, zmmB, anyptr_gpC); - e->vpermt2ps(xmmA, xmmB, xmmC); - e->vpermt2ps(xmmA, xmmB, anyptr_gpC); - e->vpermt2ps(ymmA, ymmB, ymmC); - e->vpermt2ps(ymmA, ymmB, anyptr_gpC); - e->vpermt2ps(zmmA, zmmB, zmmC); - e->vpermt2ps(zmmA, zmmB, anyptr_gpC); - e->vpermt2q(xmmA, xmmB, xmmC); - e->vpermt2q(xmmA, xmmB, anyptr_gpC); - e->vpermt2q(ymmA, ymmB, ymmC); - e->vpermt2q(ymmA, ymmB, anyptr_gpC); - e->vpermt2q(zmmA, zmmB, zmmC); - e->vpermt2q(zmmA, zmmB, anyptr_gpC); - e->vpermt2w(xmmA, xmmB, xmmC); - e->vpermt2w(xmmA, xmmB, anyptr_gpC); - e->vpermt2w(ymmA, ymmB, ymmC); - e->vpermt2w(ymmA, ymmB, anyptr_gpC); - e->vpermt2w(zmmA, zmmB, zmmC); - e->vpermt2w(zmmA, zmmB, anyptr_gpC); - e->vpermw(xmmA, xmmB, xmmC); - e->vpermw(xmmA, xmmB, anyptr_gpC); - e->vpermw(ymmA, ymmB, ymmC); - e->vpermw(ymmA, ymmB, anyptr_gpC); - e->vpermw(zmmA, zmmB, zmmC); - e->vpermw(zmmA, zmmB, anyptr_gpC); - e->vpexpandd(xmmA, xmmB); - e->vpexpandd(xmmA, anyptr_gpB); - e->vpexpandd(ymmA, ymmB); - e->vpexpandd(ymmA, anyptr_gpB); - e->vpexpandd(zmmA, zmmB); - e->vpexpandd(zmmA, anyptr_gpB); - e->vpexpandq(xmmA, xmmB); - e->vpexpandq(xmmA, anyptr_gpB); - e->vpexpandq(ymmA, ymmB); - e->vpexpandq(ymmA, anyptr_gpB); - e->vpexpandq(zmmA, zmmB); - e->vpexpandq(zmmA, anyptr_gpB); - e->vpextrb(gdA, xmmB, 0); - e->vpextrb(anyptr_gpA, xmmB, 0); - e->vpextrb(gzA, xmmB, 0); - e->vpextrd(gdA, xmmB, 0); - e->vpextrd(anyptr_gpA, xmmB, 0); - if (isX64) e->vpextrd(gzA, xmmB, 0); - if (isX64) e->vpextrq(gzA, xmmB, 0); - e->vpextrq(anyptr_gpA, xmmB, 0); - e->vpextrw(gdA, xmmB, 0); - e->vpextrw(gzA, xmmB, 0); - e->vpextrw(gdA, xmmB, 0); - e->vpextrw(anyptr_gpA, xmmB, 0); - e->vpextrw(gzA, xmmB, 0); - e->vpgatherdd(xmmA, vx_ptr); - e->vpgatherdd(ymmA, vy_ptr); - e->vpgatherdd(zmmA, vz_ptr); - e->vpgatherdq(xmmA, vx_ptr); - e->vpgatherdq(ymmA, vy_ptr); - e->vpgatherdq(zmmA, vz_ptr); - e->vpgatherqd(xmmA, vx_ptr); - e->vpgatherqd(ymmA, vy_ptr); - e->vpgatherqd(zmmA, vz_ptr); - e->vpgatherqq(xmmA, vx_ptr); - e->vpgatherqq(ymmA, vy_ptr); - e->vpgatherqq(zmmA, vz_ptr); - e->vpinsrb(xmmA, xmmB, gdC, 0); - e->vpinsrb(xmmA, xmmB, anyptr_gpC, 0); - e->vpinsrb(xmmA, xmmB, gzC, 0); - e->vpinsrd(xmmA, xmmB, gdC, 0); - e->vpinsrd(xmmA, xmmB, anyptr_gpC, 0); - e->vpinsrd(xmmA, xmmB, gzC, 0); - if (isX64) e->vpinsrq(xmmA, xmmB, gzC, 0); - e->vpinsrq(xmmA, xmmB, anyptr_gpC, 0); - e->vpinsrw(xmmA, xmmB, gdC, 0); - e->vpinsrw(xmmA, xmmB, anyptr_gpC, 0); - e->vpinsrw(xmmA, xmmB, gzC, 0); - e->vplzcntd(xmmA, xmmB); - e->vplzcntd(xmmA, anyptr_gpB); - e->vplzcntd(ymmA, ymmB); - e->vplzcntd(ymmA, anyptr_gpB); - e->vplzcntd(zmmA, zmmB); - e->vplzcntd(zmmA, anyptr_gpB); - e->vplzcntq(xmmA, xmmB); - e->vplzcntq(xmmA, anyptr_gpB); - e->vplzcntq(ymmA, ymmB); - e->vplzcntq(ymmA, anyptr_gpB); - e->vplzcntq(zmmA, zmmB); - e->vplzcntq(zmmA, anyptr_gpB); - e->vpmadd52huq(xmmA, xmmB, xmmC); - e->vpmadd52huq(xmmA, xmmB, anyptr_gpC); - e->vpmadd52huq(ymmA, ymmB, ymmC); - e->vpmadd52huq(ymmA, ymmB, anyptr_gpC); - e->vpmadd52huq(zmmA, zmmB, zmmC); - e->vpmadd52huq(zmmA, zmmB, anyptr_gpC); - e->vpmadd52luq(xmmA, xmmB, xmmC); - e->vpmadd52luq(xmmA, xmmB, anyptr_gpC); - e->vpmadd52luq(ymmA, ymmB, ymmC); - e->vpmadd52luq(ymmA, ymmB, anyptr_gpC); - e->vpmadd52luq(zmmA, zmmB, zmmC); - e->vpmadd52luq(zmmA, zmmB, anyptr_gpC); - e->vpmaddubsw(xmmA, xmmB, xmmC); - e->vpmaddubsw(xmmA, xmmB, anyptr_gpC); - e->vpmaddubsw(ymmA, ymmB, ymmC); - e->vpmaddubsw(ymmA, ymmB, anyptr_gpC); - e->vpmaddubsw(zmmA, zmmB, zmmC); - e->vpmaddubsw(zmmA, zmmB, anyptr_gpC); - e->vpmaddwd(xmmA, xmmB, xmmC); - e->vpmaddwd(xmmA, xmmB, anyptr_gpC); - e->vpmaddwd(ymmA, ymmB, ymmC); - e->vpmaddwd(ymmA, ymmB, anyptr_gpC); - e->vpmaddwd(zmmA, zmmB, zmmC); - e->vpmaddwd(zmmA, zmmB, anyptr_gpC); - e->vpmaxsb(xmmA, xmmB, xmmC); - e->vpmaxsb(xmmA, xmmB, anyptr_gpC); - e->vpmaxsb(ymmA, ymmB, ymmC); - e->vpmaxsb(ymmA, ymmB, anyptr_gpC); - e->vpmaxsb(zmmA, zmmB, zmmC); - e->vpmaxsb(zmmA, zmmB, anyptr_gpC); - e->vpmaxsd(xmmA, xmmB, xmmC); - e->vpmaxsd(xmmA, xmmB, anyptr_gpC); - e->vpmaxsd(ymmA, ymmB, ymmC); - e->vpmaxsd(ymmA, ymmB, anyptr_gpC); - e->vpmaxsd(zmmA, zmmB, zmmC); - e->vpmaxsd(zmmA, zmmB, anyptr_gpC); - e->vpmaxsq(xmmA, xmmB, xmmC); - e->vpmaxsq(xmmA, xmmB, anyptr_gpC); - e->vpmaxsq(ymmA, ymmB, ymmC); - e->vpmaxsq(ymmA, ymmB, anyptr_gpC); - e->vpmaxsq(zmmA, zmmB, zmmC); - e->vpmaxsq(zmmA, zmmB, anyptr_gpC); - e->vpmaxsw(xmmA, xmmB, xmmC); - e->vpmaxsw(xmmA, xmmB, anyptr_gpC); - e->vpmaxsw(ymmA, ymmB, ymmC); - e->vpmaxsw(ymmA, ymmB, anyptr_gpC); - e->vpmaxsw(zmmA, zmmB, zmmC); - e->vpmaxsw(zmmA, zmmB, anyptr_gpC); - e->vpmaxub(xmmA, xmmB, xmmC); - e->vpmaxub(xmmA, xmmB, anyptr_gpC); - e->vpmaxub(ymmA, ymmB, ymmC); - e->vpmaxub(ymmA, ymmB, anyptr_gpC); - e->vpmaxub(zmmA, zmmB, zmmC); - e->vpmaxub(zmmA, zmmB, anyptr_gpC); - e->vpmaxud(xmmA, xmmB, xmmC); - e->vpmaxud(xmmA, xmmB, anyptr_gpC); - e->vpmaxud(ymmA, ymmB, ymmC); - e->vpmaxud(ymmA, ymmB, anyptr_gpC); - e->vpmaxud(zmmA, zmmB, zmmC); - e->vpmaxud(zmmA, zmmB, anyptr_gpC); - e->vpmaxuq(xmmA, xmmB, xmmC); - e->vpmaxuq(xmmA, xmmB, anyptr_gpC); - e->vpmaxuq(ymmA, ymmB, ymmC); - e->vpmaxuq(ymmA, ymmB, anyptr_gpC); - e->vpmaxuq(zmmA, zmmB, zmmC); - e->vpmaxuq(zmmA, zmmB, anyptr_gpC); - e->vpmaxuw(xmmA, xmmB, xmmC); - e->vpmaxuw(xmmA, xmmB, anyptr_gpC); - e->vpmaxuw(ymmA, ymmB, ymmC); - e->vpmaxuw(ymmA, ymmB, anyptr_gpC); - e->vpmaxuw(zmmA, zmmB, zmmC); - e->vpmaxuw(zmmA, zmmB, anyptr_gpC); - e->vpminsb(xmmA, xmmB, xmmC); - e->vpminsb(xmmA, xmmB, anyptr_gpC); - e->vpminsb(ymmA, ymmB, ymmC); - e->vpminsb(ymmA, ymmB, anyptr_gpC); - e->vpminsb(zmmA, zmmB, zmmC); - e->vpminsb(zmmA, zmmB, anyptr_gpC); - e->vpminsd(xmmA, xmmB, xmmC); - e->vpminsd(xmmA, xmmB, anyptr_gpC); - e->vpminsd(ymmA, ymmB, ymmC); - e->vpminsd(ymmA, ymmB, anyptr_gpC); - e->vpminsd(zmmA, zmmB, zmmC); - e->vpminsd(zmmA, zmmB, anyptr_gpC); - e->vpminsq(xmmA, xmmB, xmmC); - e->vpminsq(xmmA, xmmB, anyptr_gpC); - e->vpminsq(ymmA, ymmB, ymmC); - e->vpminsq(ymmA, ymmB, anyptr_gpC); - e->vpminsq(zmmA, zmmB, zmmC); - e->vpminsq(zmmA, zmmB, anyptr_gpC); - e->vpminsw(xmmA, xmmB, xmmC); - e->vpminsw(xmmA, xmmB, anyptr_gpC); - e->vpminsw(ymmA, ymmB, ymmC); - e->vpminsw(ymmA, ymmB, anyptr_gpC); - e->vpminsw(zmmA, zmmB, zmmC); - e->vpminsw(zmmA, zmmB, anyptr_gpC); - e->vpminub(xmmA, xmmB, xmmC); - e->vpminub(xmmA, xmmB, anyptr_gpC); - e->vpminub(ymmA, ymmB, ymmC); - e->vpminub(ymmA, ymmB, anyptr_gpC); - e->vpminub(zmmA, zmmB, zmmC); - e->vpminub(zmmA, zmmB, anyptr_gpC); - e->vpminud(xmmA, xmmB, xmmC); - e->vpminud(xmmA, xmmB, anyptr_gpC); - e->vpminud(ymmA, ymmB, ymmC); - e->vpminud(ymmA, ymmB, anyptr_gpC); - e->vpminud(zmmA, zmmB, zmmC); - e->vpminud(zmmA, zmmB, anyptr_gpC); - e->vpminuq(xmmA, xmmB, xmmC); - e->vpminuq(xmmA, xmmB, anyptr_gpC); - e->vpminuq(ymmA, ymmB, ymmC); - e->vpminuq(ymmA, ymmB, anyptr_gpC); - e->vpminuq(zmmA, zmmB, zmmC); - e->vpminuq(zmmA, zmmB, anyptr_gpC); - e->vpminuw(xmmA, xmmB, xmmC); - e->vpminuw(xmmA, xmmB, anyptr_gpC); - e->vpminuw(ymmA, ymmB, ymmC); - e->vpminuw(ymmA, ymmB, anyptr_gpC); - e->vpminuw(zmmA, zmmB, zmmC); - e->vpminuw(zmmA, zmmB, anyptr_gpC); - e->vpmovb2m(kA, xmmB); - e->vpmovb2m(kA, ymmB); - e->vpmovb2m(kA, zmmB); - e->vpmovd2m(kA, xmmB); - e->vpmovd2m(kA, ymmB); - e->vpmovd2m(kA, zmmB); - e->vpmovdb(xmmA, xmmB); - e->vpmovdb(anyptr_gpA, xmmB); - e->vpmovdb(xmmA, ymmB); - e->vpmovdb(anyptr_gpA, ymmB); - e->vpmovdb(xmmA, zmmB); - e->vpmovdb(anyptr_gpA, zmmB); - e->vpmovdw(xmmA, xmmB); - e->vpmovdw(anyptr_gpA, xmmB); - e->vpmovdw(xmmA, ymmB); - e->vpmovdw(anyptr_gpA, ymmB); - e->vpmovdw(ymmA, zmmB); - e->vpmovdw(anyptr_gpA, zmmB); - e->vpmovm2b(xmmA, kB); - e->vpmovm2b(ymmA, kB); - e->vpmovm2b(zmmA, kB); - e->vpmovm2d(xmmA, kB); - e->vpmovm2d(ymmA, kB); - e->vpmovm2d(zmmA, kB); - e->vpmovm2q(xmmA, kB); - e->vpmovm2q(ymmA, kB); - e->vpmovm2q(zmmA, kB); - e->vpmovm2w(xmmA, kB); - e->vpmovm2w(ymmA, kB); - e->vpmovm2w(zmmA, kB); - e->vpmovq2m(kA, xmmB); - e->vpmovq2m(kA, ymmB); - e->vpmovq2m(kA, zmmB); - e->vpmovqb(xmmA, xmmB); - e->vpmovqb(anyptr_gpA, xmmB); - e->vpmovqb(xmmA, ymmB); - e->vpmovqb(anyptr_gpA, ymmB); - e->vpmovqb(xmmA, zmmB); - e->vpmovqb(anyptr_gpA, zmmB); - e->vpmovqd(xmmA, xmmB); - e->vpmovqd(anyptr_gpA, xmmB); - e->vpmovqd(xmmA, ymmB); - e->vpmovqd(anyptr_gpA, ymmB); - e->vpmovqd(ymmA, zmmB); - e->vpmovqd(anyptr_gpA, zmmB); - e->vpmovqw(xmmA, xmmB); - e->vpmovqw(anyptr_gpA, xmmB); - e->vpmovqw(xmmA, ymmB); - e->vpmovqw(anyptr_gpA, ymmB); - e->vpmovqw(xmmA, zmmB); - e->vpmovqw(anyptr_gpA, zmmB); - e->vpmovsdb(xmmA, xmmB); - e->vpmovsdb(anyptr_gpA, xmmB); - e->vpmovsdb(xmmA, ymmB); - e->vpmovsdb(anyptr_gpA, ymmB); - e->vpmovsdb(xmmA, zmmB); - e->vpmovsdb(anyptr_gpA, zmmB); - e->vpmovsdw(xmmA, xmmB); - e->vpmovsdw(anyptr_gpA, xmmB); - e->vpmovsdw(xmmA, ymmB); - e->vpmovsdw(anyptr_gpA, ymmB); - e->vpmovsdw(ymmA, zmmB); - e->vpmovsdw(anyptr_gpA, zmmB); - e->vpmovsqb(xmmA, xmmB); - e->vpmovsqb(anyptr_gpA, xmmB); - e->vpmovsqb(xmmA, ymmB); - e->vpmovsqb(anyptr_gpA, ymmB); - e->vpmovsqb(xmmA, zmmB); - e->vpmovsqb(anyptr_gpA, zmmB); - e->vpmovsqd(xmmA, xmmB); - e->vpmovsqd(anyptr_gpA, xmmB); - e->vpmovsqd(xmmA, ymmB); - e->vpmovsqd(anyptr_gpA, ymmB); - e->vpmovsqd(ymmA, zmmB); - e->vpmovsqd(anyptr_gpA, zmmB); - e->vpmovsqw(xmmA, xmmB); - e->vpmovsqw(anyptr_gpA, xmmB); - e->vpmovsqw(xmmA, ymmB); - e->vpmovsqw(anyptr_gpA, ymmB); - e->vpmovsqw(xmmA, zmmB); - e->vpmovsqw(anyptr_gpA, zmmB); - e->vpmovswb(xmmA, xmmB); - e->vpmovswb(anyptr_gpA, xmmB); - e->vpmovswb(xmmA, ymmB); - e->vpmovswb(anyptr_gpA, ymmB); - e->vpmovswb(ymmA, zmmB); - e->vpmovswb(anyptr_gpA, zmmB); - e->vpmovsxbd(xmmA, xmmB); - e->vpmovsxbd(xmmA, anyptr_gpB); - e->vpmovsxbd(ymmA, xmmB); - e->vpmovsxbd(ymmA, anyptr_gpB); - e->vpmovsxbd(zmmA, xmmB); - e->vpmovsxbd(zmmA, anyptr_gpB); - e->vpmovsxbq(xmmA, xmmB); - e->vpmovsxbq(xmmA, anyptr_gpB); - e->vpmovsxbq(ymmA, xmmB); - e->vpmovsxbq(ymmA, anyptr_gpB); - e->vpmovsxbq(zmmA, xmmB); - e->vpmovsxbq(zmmA, anyptr_gpB); - e->vpmovsxbw(xmmA, xmmB); - e->vpmovsxbw(xmmA, anyptr_gpB); - e->vpmovsxbw(ymmA, xmmB); - e->vpmovsxbw(ymmA, anyptr_gpB); - e->vpmovsxbw(zmmA, ymmB); - e->vpmovsxbw(zmmA, anyptr_gpB); - e->vpmovsxdq(xmmA, xmmB); - e->vpmovsxdq(xmmA, anyptr_gpB); - e->vpmovsxdq(ymmA, xmmB); - e->vpmovsxdq(ymmA, anyptr_gpB); - e->vpmovsxdq(zmmA, ymmB); - e->vpmovsxdq(zmmA, anyptr_gpB); - e->vpmovsxwd(xmmA, xmmB); - e->vpmovsxwd(xmmA, anyptr_gpB); - e->vpmovsxwd(ymmA, xmmB); - e->vpmovsxwd(ymmA, anyptr_gpB); - e->vpmovsxwd(zmmA, ymmB); - e->vpmovsxwd(zmmA, anyptr_gpB); - e->vpmovsxwq(xmmA, xmmB); - e->vpmovsxwq(xmmA, anyptr_gpB); - e->vpmovsxwq(ymmA, xmmB); - e->vpmovsxwq(ymmA, anyptr_gpB); - e->vpmovsxwq(zmmA, xmmB); - e->vpmovsxwq(zmmA, anyptr_gpB); - e->vpmovusdb(xmmA, xmmB); - e->vpmovusdb(anyptr_gpA, xmmB); - e->vpmovusdb(xmmA, ymmB); - e->vpmovusdb(anyptr_gpA, ymmB); - e->vpmovusdb(xmmA, zmmB); - e->vpmovusdb(anyptr_gpA, zmmB); - e->vpmovusdw(xmmA, xmmB); - e->vpmovusdw(anyptr_gpA, xmmB); - e->vpmovusdw(xmmA, ymmB); - e->vpmovusdw(anyptr_gpA, ymmB); - e->vpmovusdw(ymmA, zmmB); - e->vpmovusdw(anyptr_gpA, zmmB); - e->vpmovusqb(xmmA, xmmB); - e->vpmovusqb(anyptr_gpA, xmmB); - e->vpmovusqb(xmmA, ymmB); - e->vpmovusqb(anyptr_gpA, ymmB); - e->vpmovusqb(xmmA, zmmB); - e->vpmovusqb(anyptr_gpA, zmmB); - e->vpmovusqd(xmmA, xmmB); - e->vpmovusqd(anyptr_gpA, xmmB); - e->vpmovusqd(xmmA, ymmB); - e->vpmovusqd(anyptr_gpA, ymmB); - e->vpmovusqd(ymmA, zmmB); - e->vpmovusqd(anyptr_gpA, zmmB); - e->vpmovusqw(xmmA, xmmB); - e->vpmovusqw(anyptr_gpA, xmmB); - e->vpmovusqw(xmmA, ymmB); - e->vpmovusqw(anyptr_gpA, ymmB); - e->vpmovusqw(xmmA, zmmB); - e->vpmovusqw(anyptr_gpA, zmmB); - e->vpmovuswb(xmmA, xmmB); - e->vpmovuswb(anyptr_gpA, xmmB); - e->vpmovuswb(xmmA, ymmB); - e->vpmovuswb(anyptr_gpA, ymmB); - e->vpmovuswb(ymmA, zmmB); - e->vpmovuswb(anyptr_gpA, zmmB); - e->vpmovw2m(kA, xmmB); - e->vpmovw2m(kA, ymmB); - e->vpmovw2m(kA, zmmB); - e->vpmovwb(xmmA, xmmB); - e->vpmovwb(anyptr_gpA, xmmB); - e->vpmovwb(xmmA, ymmB); - e->vpmovwb(anyptr_gpA, ymmB); - e->vpmovwb(ymmA, zmmB); - e->vpmovwb(anyptr_gpA, zmmB); - e->vpmovzxbd(xmmA, xmmB); - e->vpmovzxbd(xmmA, anyptr_gpB); - e->vpmovzxbd(ymmA, xmmB); - e->vpmovzxbd(ymmA, anyptr_gpB); - e->vpmovzxbd(zmmA, xmmB); - e->vpmovzxbd(zmmA, anyptr_gpB); - e->vpmovzxbq(xmmA, xmmB); - e->vpmovzxbq(xmmA, anyptr_gpB); - e->vpmovzxbq(ymmA, xmmB); - e->vpmovzxbq(ymmA, anyptr_gpB); - e->vpmovzxbq(zmmA, xmmB); - e->vpmovzxbq(zmmA, anyptr_gpB); - e->vpmovzxbw(xmmA, xmmB); - e->vpmovzxbw(xmmA, anyptr_gpB); - e->vpmovzxbw(ymmA, xmmB); - e->vpmovzxbw(ymmA, anyptr_gpB); - e->vpmovzxbw(zmmA, ymmB); - e->vpmovzxbw(zmmA, anyptr_gpB); - e->vpmovzxdq(xmmA, xmmB); - e->vpmovzxdq(xmmA, anyptr_gpB); - e->vpmovzxdq(ymmA, xmmB); - e->vpmovzxdq(ymmA, anyptr_gpB); - e->vpmovzxdq(zmmA, ymmB); - e->vpmovzxdq(zmmA, anyptr_gpB); - e->vpmovzxwd(xmmA, xmmB); - e->vpmovzxwd(xmmA, anyptr_gpB); - e->vpmovzxwd(ymmA, xmmB); - e->vpmovzxwd(ymmA, anyptr_gpB); - e->vpmovzxwd(zmmA, ymmB); - e->vpmovzxwd(zmmA, anyptr_gpB); - e->vpmovzxwq(xmmA, xmmB); - e->vpmovzxwq(xmmA, anyptr_gpB); - e->vpmovzxwq(ymmA, xmmB); - e->vpmovzxwq(ymmA, anyptr_gpB); - e->vpmovzxwq(zmmA, xmmB); - e->vpmovzxwq(zmmA, anyptr_gpB); - e->vpmuldq(xmmA, xmmB, xmmC); - e->vpmuldq(xmmA, xmmB, anyptr_gpC); - e->vpmuldq(ymmA, ymmB, ymmC); - e->vpmuldq(ymmA, ymmB, anyptr_gpC); - e->vpmuldq(zmmA, zmmB, zmmC); - e->vpmuldq(zmmA, zmmB, anyptr_gpC); - e->vpmulhrsw(xmmA, xmmB, xmmC); - e->vpmulhrsw(xmmA, xmmB, anyptr_gpC); - e->vpmulhrsw(ymmA, ymmB, ymmC); - e->vpmulhrsw(ymmA, ymmB, anyptr_gpC); - e->vpmulhrsw(zmmA, zmmB, zmmC); - e->vpmulhrsw(zmmA, zmmB, anyptr_gpC); - e->vpmulhuw(xmmA, xmmB, xmmC); - e->vpmulhuw(xmmA, xmmB, anyptr_gpC); - e->vpmulhuw(ymmA, ymmB, ymmC); - e->vpmulhuw(ymmA, ymmB, anyptr_gpC); - e->vpmulhuw(zmmA, zmmB, zmmC); - e->vpmulhuw(zmmA, zmmB, anyptr_gpC); - e->vpmulhw(xmmA, xmmB, xmmC); - e->vpmulhw(xmmA, xmmB, anyptr_gpC); - e->vpmulhw(ymmA, ymmB, ymmC); - e->vpmulhw(ymmA, ymmB, anyptr_gpC); - e->vpmulhw(zmmA, zmmB, zmmC); - e->vpmulhw(zmmA, zmmB, anyptr_gpC); - e->vpmulld(xmmA, xmmB, xmmC); - e->vpmulld(xmmA, xmmB, anyptr_gpC); - e->vpmulld(ymmA, ymmB, ymmC); - e->vpmulld(ymmA, ymmB, anyptr_gpC); - e->vpmulld(zmmA, zmmB, zmmC); - e->vpmulld(zmmA, zmmB, anyptr_gpC); - e->vpmullq(xmmA, xmmB, xmmC); - e->vpmullq(xmmA, xmmB, anyptr_gpC); - e->vpmullq(ymmA, ymmB, ymmC); - e->vpmullq(ymmA, ymmB, anyptr_gpC); - e->vpmullq(zmmA, zmmB, zmmC); - e->vpmullq(zmmA, zmmB, anyptr_gpC); - e->vpmullw(xmmA, xmmB, xmmC); - e->vpmullw(xmmA, xmmB, anyptr_gpC); - e->vpmullw(ymmA, ymmB, ymmC); - e->vpmullw(ymmA, ymmB, anyptr_gpC); - e->vpmullw(zmmA, zmmB, zmmC); - e->vpmullw(zmmA, zmmB, anyptr_gpC); - e->vpmultishiftqb(xmmA, xmmB, xmmC); - e->vpmultishiftqb(xmmA, xmmB, anyptr_gpC); - e->vpmultishiftqb(ymmA, ymmB, ymmC); - e->vpmultishiftqb(ymmA, ymmB, anyptr_gpC); - e->vpmultishiftqb(zmmA, zmmB, zmmC); - e->vpmultishiftqb(zmmA, zmmB, anyptr_gpC); - e->vpmuludq(xmmA, xmmB, xmmC); - e->vpmuludq(xmmA, xmmB, anyptr_gpC); - e->vpmuludq(ymmA, ymmB, ymmC); - e->vpmuludq(ymmA, ymmB, anyptr_gpC); - e->vpmuludq(zmmA, zmmB, zmmC); - e->vpmuludq(zmmA, zmmB, anyptr_gpC); - e->vpopcntd(zmmA, zmmB); - e->vpopcntd(zmmA, anyptr_gpB); - e->vpopcntq(zmmA, zmmB); - e->vpopcntq(zmmA, anyptr_gpB); - e->vpord(xmmA, xmmB, xmmC); - e->vpord(xmmA, xmmB, anyptr_gpC); - e->vpord(ymmA, ymmB, ymmC); - e->vpord(ymmA, ymmB, anyptr_gpC); - e->vpord(zmmA, zmmB, zmmC); - e->vpord(zmmA, zmmB, anyptr_gpC); - e->vporq(xmmA, xmmB, xmmC); - e->vporq(xmmA, xmmB, anyptr_gpC); - e->vporq(ymmA, ymmB, ymmC); - e->vporq(ymmA, ymmB, anyptr_gpC); - e->vporq(zmmA, zmmB, zmmC); - e->vporq(zmmA, zmmB, anyptr_gpC); - e->vprold(xmmA, xmmB, 0); - e->vprold(xmmA, anyptr_gpB, 0); - e->vprold(ymmA, ymmB, 0); - e->vprold(ymmA, anyptr_gpB, 0); - e->vprold(zmmA, zmmB, 0); - e->vprold(zmmA, anyptr_gpB, 0); - e->vprolq(xmmA, xmmB, 0); - e->vprolq(xmmA, anyptr_gpB, 0); - e->vprolq(ymmA, ymmB, 0); - e->vprolq(ymmA, anyptr_gpB, 0); - e->vprolq(zmmA, zmmB, 0); - e->vprolq(zmmA, anyptr_gpB, 0); - e->vprolvd(xmmA, xmmB, xmmC); - e->vprolvd(xmmA, xmmB, anyptr_gpC); - e->vprolvd(ymmA, ymmB, ymmC); - e->vprolvd(ymmA, ymmB, anyptr_gpC); - e->vprolvd(zmmA, zmmB, zmmC); - e->vprolvd(zmmA, zmmB, anyptr_gpC); - e->vprolvq(xmmA, xmmB, xmmC); - e->vprolvq(xmmA, xmmB, anyptr_gpC); - e->vprolvq(ymmA, ymmB, ymmC); - e->vprolvq(ymmA, ymmB, anyptr_gpC); - e->vprolvq(zmmA, zmmB, zmmC); - e->vprolvq(zmmA, zmmB, anyptr_gpC); - e->vprord(xmmA, xmmB, 0); - e->vprord(xmmA, anyptr_gpB, 0); - e->vprord(ymmA, ymmB, 0); - e->vprord(ymmA, anyptr_gpB, 0); - e->vprord(zmmA, zmmB, 0); - e->vprord(zmmA, anyptr_gpB, 0); - e->vprorq(xmmA, xmmB, 0); - e->vprorq(xmmA, anyptr_gpB, 0); - e->vprorq(ymmA, ymmB, 0); - e->vprorq(ymmA, anyptr_gpB, 0); - e->vprorq(zmmA, zmmB, 0); - e->vprorq(zmmA, anyptr_gpB, 0); - e->vprorvd(xmmA, xmmB, xmmC); - e->vprorvd(xmmA, xmmB, anyptr_gpC); - e->vprorvd(ymmA, ymmB, ymmC); - e->vprorvd(ymmA, ymmB, anyptr_gpC); - e->vprorvd(zmmA, zmmB, zmmC); - e->vprorvd(zmmA, zmmB, anyptr_gpC); - e->vprorvq(xmmA, xmmB, xmmC); - e->vprorvq(xmmA, xmmB, anyptr_gpC); - e->vprorvq(ymmA, ymmB, ymmC); - e->vprorvq(ymmA, ymmB, anyptr_gpC); - e->vprorvq(zmmA, zmmB, zmmC); - e->vprorvq(zmmA, zmmB, anyptr_gpC); - e->vpsadbw(xmmA, xmmB, xmmC); - e->vpsadbw(xmmA, xmmB, anyptr_gpC); - e->vpsadbw(ymmA, ymmB, ymmC); - e->vpsadbw(ymmA, ymmB, anyptr_gpC); - e->vpsadbw(zmmA, zmmB, zmmC); - e->vpsadbw(zmmA, zmmB, anyptr_gpC); - e->vpscatterdd(vx_ptr, xmmB); - e->vpscatterdd(vy_ptr, ymmB); - e->vpscatterdd(vz_ptr, zmmB); - e->vpscatterdq(vx_ptr, xmmB); - e->vpscatterdq(vy_ptr, ymmB); - e->vpscatterdq(vz_ptr, zmmB); - e->vpscatterqd(vx_ptr, xmmB); - e->vpscatterqd(vy_ptr, xmmB); - e->vpscatterqd(vz_ptr, ymmB); - e->vpscatterqq(vx_ptr, xmmB); - e->vpscatterqq(vy_ptr, ymmB); - e->vpscatterqq(vz_ptr, zmmB); - e->vpshufb(xmmA, xmmB, xmmC); - e->vpshufb(xmmA, xmmB, anyptr_gpC); - e->vpshufb(ymmA, ymmB, ymmC); - e->vpshufb(ymmA, ymmB, anyptr_gpC); - e->vpshufb(zmmA, zmmB, zmmC); - e->vpshufb(zmmA, zmmB, anyptr_gpC); - e->vpshufd(xmmA, xmmB, 0); - e->vpshufd(xmmA, anyptr_gpB, 0); - e->vpshufd(ymmA, ymmB, 0); - e->vpshufd(ymmA, anyptr_gpB, 0); - e->vpshufd(zmmA, zmmB, 0); - e->vpshufd(zmmA, anyptr_gpB, 0); - e->vpshufhw(xmmA, xmmB, 0); - e->vpshufhw(xmmA, anyptr_gpB, 0); - e->vpshufhw(ymmA, ymmB, 0); - e->vpshufhw(ymmA, anyptr_gpB, 0); - e->vpshufhw(zmmA, zmmB, 0); - e->vpshufhw(zmmA, anyptr_gpB, 0); - e->vpshuflw(xmmA, xmmB, 0); - e->vpshuflw(xmmA, anyptr_gpB, 0); - e->vpshuflw(ymmA, ymmB, 0); - e->vpshuflw(ymmA, anyptr_gpB, 0); - e->vpshuflw(zmmA, zmmB, 0); - e->vpshuflw(zmmA, anyptr_gpB, 0); - e->vpslld(xmmA, xmmB, xmmC); - e->vpslld(xmmA, xmmB, anyptr_gpC); - e->vpslld(xmmA, xmmB, 0); - e->vpslld(xmmA, anyptr_gpB, 0); - e->vpslld(ymmA, ymmB, xmmC); - e->vpslld(ymmA, ymmB, anyptr_gpC); - e->vpslld(ymmA, ymmB, 0); - e->vpslld(ymmA, anyptr_gpB, 0); - e->vpslld(zmmA, zmmB, xmmC); - e->vpslld(zmmA, zmmB, anyptr_gpC); - e->vpslld(zmmA, zmmB, 0); - e->vpslld(zmmA, anyptr_gpB, 0); - e->vpslldq(xmmA, xmmB, 0); - e->vpslldq(xmmA, anyptr_gpB, 0); - e->vpslldq(ymmA, ymmB, 0); - e->vpslldq(ymmA, anyptr_gpB, 0); - e->vpslldq(zmmA, zmmB, 0); - e->vpslldq(zmmA, anyptr_gpB, 0); - e->vpsllq(xmmA, xmmB, xmmC); - e->vpsllq(xmmA, xmmB, anyptr_gpC); - e->vpsllq(xmmA, xmmB, 0); - e->vpsllq(xmmA, anyptr_gpB, 0); - e->vpsllq(ymmA, ymmB, xmmC); - e->vpsllq(ymmA, ymmB, anyptr_gpC); - e->vpsllq(ymmA, ymmB, 0); - e->vpsllq(ymmA, anyptr_gpB, 0); - e->vpsllq(zmmA, zmmB, xmmC); - e->vpsllq(zmmA, zmmB, anyptr_gpC); - e->vpsllq(zmmA, zmmB, 0); - e->vpsllq(zmmA, anyptr_gpB, 0); - e->vpsllvd(xmmA, xmmB, xmmC); - e->vpsllvd(xmmA, xmmB, anyptr_gpC); - e->vpsllvd(ymmA, ymmB, ymmC); - e->vpsllvd(ymmA, ymmB, anyptr_gpC); - e->vpsllvd(zmmA, zmmB, zmmC); - e->vpsllvd(zmmA, zmmB, anyptr_gpC); - e->vpsllvq(xmmA, xmmB, xmmC); - e->vpsllvq(xmmA, xmmB, anyptr_gpC); - e->vpsllvq(ymmA, ymmB, ymmC); - e->vpsllvq(ymmA, ymmB, anyptr_gpC); - e->vpsllvq(zmmA, zmmB, zmmC); - e->vpsllvq(zmmA, zmmB, anyptr_gpC); - e->vpsllvw(xmmA, xmmB, xmmC); - e->vpsllvw(xmmA, xmmB, anyptr_gpC); - e->vpsllvw(ymmA, ymmB, ymmC); - e->vpsllvw(ymmA, ymmB, anyptr_gpC); - e->vpsllvw(zmmA, zmmB, zmmC); - e->vpsllvw(zmmA, zmmB, anyptr_gpC); - e->vpsllw(xmmA, xmmB, xmmC); - e->vpsllw(xmmA, xmmB, anyptr_gpC); - e->vpsllw(xmmA, xmmB, 0); - e->vpsllw(xmmA, anyptr_gpB, 0); - e->vpsllw(ymmA, ymmB, xmmC); - e->vpsllw(ymmA, ymmB, anyptr_gpC); - e->vpsllw(ymmA, ymmB, 0); - e->vpsllw(ymmA, anyptr_gpB, 0); - e->vpsllw(zmmA, zmmB, xmmC); - e->vpsllw(zmmA, zmmB, anyptr_gpC); - e->vpsllw(zmmA, zmmB, 0); - e->vpsllw(zmmA, anyptr_gpB, 0); - e->vpsrad(xmmA, xmmB, xmmC); - e->vpsrad(xmmA, xmmB, anyptr_gpC); - e->vpsrad(xmmA, xmmB, 0); - e->vpsrad(xmmA, anyptr_gpB, 0); - e->vpsrad(ymmA, ymmB, xmmC); - e->vpsrad(ymmA, ymmB, anyptr_gpC); - e->vpsrad(ymmA, ymmB, 0); - e->vpsrad(ymmA, anyptr_gpB, 0); - e->vpsrad(zmmA, zmmB, xmmC); - e->vpsrad(zmmA, zmmB, anyptr_gpC); - e->vpsrad(zmmA, zmmB, 0); - e->vpsrad(zmmA, anyptr_gpB, 0); - e->vpsraq(xmmA, xmmB, xmmC); - e->vpsraq(xmmA, xmmB, anyptr_gpC); - e->vpsraq(xmmA, xmmB, 0); - e->vpsraq(xmmA, anyptr_gpB, 0); - e->vpsraq(ymmA, ymmB, xmmC); - e->vpsraq(ymmA, ymmB, anyptr_gpC); - e->vpsraq(ymmA, ymmB, 0); - e->vpsraq(ymmA, anyptr_gpB, 0); - e->vpsraq(zmmA, zmmB, xmmC); - e->vpsraq(zmmA, zmmB, anyptr_gpC); - e->vpsraq(zmmA, zmmB, 0); - e->vpsraq(zmmA, anyptr_gpB, 0); - e->vpsravd(xmmA, xmmB, xmmC); - e->vpsravd(xmmA, xmmB, anyptr_gpC); - e->vpsravd(ymmA, ymmB, ymmC); - e->vpsravd(ymmA, ymmB, anyptr_gpC); - e->vpsravd(zmmA, zmmB, zmmC); - e->vpsravd(zmmA, zmmB, anyptr_gpC); - e->vpsravq(xmmA, xmmB, xmmC); - e->vpsravq(xmmA, xmmB, anyptr_gpC); - e->vpsravq(ymmA, ymmB, ymmC); - e->vpsravq(ymmA, ymmB, anyptr_gpC); - e->vpsravq(zmmA, zmmB, zmmC); - e->vpsravq(zmmA, zmmB, anyptr_gpC); - e->vpsravw(xmmA, xmmB, xmmC); - e->vpsravw(xmmA, xmmB, anyptr_gpC); - e->vpsravw(ymmA, ymmB, ymmC); - e->vpsravw(ymmA, ymmB, anyptr_gpC); - e->vpsravw(zmmA, zmmB, zmmC); - e->vpsravw(zmmA, zmmB, anyptr_gpC); - e->vpsraw(xmmA, xmmB, xmmC); - e->vpsraw(xmmA, xmmB, anyptr_gpC); - e->vpsraw(xmmA, xmmB, 0); - e->vpsraw(xmmA, anyptr_gpB, 0); - e->vpsraw(ymmA, ymmB, xmmC); - e->vpsraw(ymmA, ymmB, anyptr_gpC); - e->vpsraw(ymmA, ymmB, 0); - e->vpsraw(ymmA, anyptr_gpB, 0); - e->vpsraw(zmmA, zmmB, xmmC); - e->vpsraw(zmmA, zmmB, anyptr_gpC); - e->vpsraw(zmmA, zmmB, 0); - e->vpsraw(zmmA, anyptr_gpB, 0); - e->vpsrld(xmmA, xmmB, xmmC); - e->vpsrld(xmmA, xmmB, anyptr_gpC); - e->vpsrld(xmmA, xmmB, 0); - e->vpsrld(xmmA, anyptr_gpB, 0); - e->vpsrld(ymmA, ymmB, xmmC); - e->vpsrld(ymmA, ymmB, anyptr_gpC); - e->vpsrld(ymmA, ymmB, 0); - e->vpsrld(ymmA, anyptr_gpB, 0); - e->vpsrld(zmmA, zmmB, xmmC); - e->vpsrld(zmmA, zmmB, anyptr_gpC); - e->vpsrld(zmmA, zmmB, 0); - e->vpsrld(zmmA, anyptr_gpB, 0); - e->vpsrldq(xmmA, xmmB, 0); - e->vpsrldq(xmmA, anyptr_gpB, 0); - e->vpsrldq(ymmA, ymmB, 0); - e->vpsrldq(ymmA, anyptr_gpB, 0); - e->vpsrldq(zmmA, zmmB, 0); - e->vpsrldq(zmmA, anyptr_gpB, 0); - e->vpsrlq(xmmA, xmmB, xmmC); - e->vpsrlq(xmmA, xmmB, anyptr_gpC); - e->vpsrlq(xmmA, xmmB, 0); - e->vpsrlq(xmmA, anyptr_gpB, 0); - e->vpsrlq(ymmA, ymmB, xmmC); - e->vpsrlq(ymmA, ymmB, anyptr_gpC); - e->vpsrlq(ymmA, ymmB, 0); - e->vpsrlq(ymmA, anyptr_gpB, 0); - e->vpsrlq(zmmA, zmmB, xmmC); - e->vpsrlq(zmmA, zmmB, anyptr_gpC); - e->vpsrlq(zmmA, zmmB, 0); - e->vpsrlq(zmmA, anyptr_gpB, 0); - e->vpsrlvd(xmmA, xmmB, xmmC); - e->vpsrlvd(xmmA, xmmB, anyptr_gpC); - e->vpsrlvd(ymmA, ymmB, ymmC); - e->vpsrlvd(ymmA, ymmB, anyptr_gpC); - e->vpsrlvd(zmmA, zmmB, zmmC); - e->vpsrlvd(zmmA, zmmB, anyptr_gpC); - e->vpsrlvq(xmmA, xmmB, xmmC); - e->vpsrlvq(xmmA, xmmB, anyptr_gpC); - e->vpsrlvq(ymmA, ymmB, ymmC); - e->vpsrlvq(ymmA, ymmB, anyptr_gpC); - e->vpsrlvq(zmmA, zmmB, zmmC); - e->vpsrlvq(zmmA, zmmB, anyptr_gpC); - e->vpsrlvw(xmmA, xmmB, xmmC); - e->vpsrlvw(xmmA, xmmB, anyptr_gpC); - e->vpsrlvw(ymmA, ymmB, ymmC); - e->vpsrlvw(ymmA, ymmB, anyptr_gpC); - e->vpsrlvw(zmmA, zmmB, zmmC); - e->vpsrlvw(zmmA, zmmB, anyptr_gpC); - e->vpsrlw(xmmA, xmmB, xmmC); - e->vpsrlw(xmmA, xmmB, anyptr_gpC); - e->vpsrlw(xmmA, xmmB, 0); - e->vpsrlw(xmmA, anyptr_gpB, 0); - e->vpsrlw(ymmA, ymmB, xmmC); - e->vpsrlw(ymmA, ymmB, anyptr_gpC); - e->vpsrlw(ymmA, ymmB, 0); - e->vpsrlw(ymmA, anyptr_gpB, 0); - e->vpsrlw(zmmA, zmmB, xmmC); - e->vpsrlw(zmmA, zmmB, anyptr_gpC); - e->vpsrlw(zmmA, zmmB, 0); - e->vpsrlw(zmmA, anyptr_gpB, 0); - e->vpsubb(xmmA, xmmB, xmmC); - e->vpsubb(xmmA, xmmB, anyptr_gpC); - e->vpsubb(ymmA, ymmB, ymmC); - e->vpsubb(ymmA, ymmB, anyptr_gpC); - e->vpsubb(zmmA, zmmB, zmmC); - e->vpsubb(zmmA, zmmB, anyptr_gpC); - e->vpsubd(xmmA, xmmB, xmmC); - e->vpsubd(xmmA, xmmB, anyptr_gpC); - e->vpsubd(ymmA, ymmB, ymmC); - e->vpsubd(ymmA, ymmB, anyptr_gpC); - e->vpsubd(zmmA, zmmB, zmmC); - e->vpsubd(zmmA, zmmB, anyptr_gpC); - e->vpsubq(xmmA, xmmB, xmmC); - e->vpsubq(xmmA, xmmB, anyptr_gpC); - e->vpsubq(ymmA, ymmB, ymmC); - e->vpsubq(ymmA, ymmB, anyptr_gpC); - e->vpsubq(zmmA, zmmB, zmmC); - e->vpsubq(zmmA, zmmB, anyptr_gpC); - e->vpsubsb(xmmA, xmmB, xmmC); - e->vpsubsb(xmmA, xmmB, anyptr_gpC); - e->vpsubsb(ymmA, ymmB, ymmC); - e->vpsubsb(ymmA, ymmB, anyptr_gpC); - e->vpsubsb(zmmA, zmmB, zmmC); - e->vpsubsb(zmmA, zmmB, anyptr_gpC); - e->vpsubsw(xmmA, xmmB, xmmC); - e->vpsubsw(xmmA, xmmB, anyptr_gpC); - e->vpsubsw(ymmA, ymmB, ymmC); - e->vpsubsw(ymmA, ymmB, anyptr_gpC); - e->vpsubsw(zmmA, zmmB, zmmC); - e->vpsubsw(zmmA, zmmB, anyptr_gpC); - e->vpsubusb(xmmA, xmmB, xmmC); - e->vpsubusb(xmmA, xmmB, anyptr_gpC); - e->vpsubusb(ymmA, ymmB, ymmC); - e->vpsubusb(ymmA, ymmB, anyptr_gpC); - e->vpsubusb(zmmA, zmmB, zmmC); - e->vpsubusb(zmmA, zmmB, anyptr_gpC); - e->vpsubusw(xmmA, xmmB, xmmC); - e->vpsubusw(xmmA, xmmB, anyptr_gpC); - e->vpsubusw(ymmA, ymmB, ymmC); - e->vpsubusw(ymmA, ymmB, anyptr_gpC); - e->vpsubusw(zmmA, zmmB, zmmC); - e->vpsubusw(zmmA, zmmB, anyptr_gpC); - e->vpsubw(xmmA, xmmB, xmmC); - e->vpsubw(xmmA, xmmB, anyptr_gpC); - e->vpsubw(ymmA, ymmB, ymmC); - e->vpsubw(ymmA, ymmB, anyptr_gpC); - e->vpsubw(zmmA, zmmB, zmmC); - e->vpsubw(zmmA, zmmB, anyptr_gpC); - e->vpternlogd(xmmA, xmmB, xmmC, 0); - e->vpternlogd(xmmA, xmmB, anyptr_gpC, 0); - e->vpternlogd(ymmA, ymmB, ymmC, 0); - e->vpternlogd(ymmA, ymmB, anyptr_gpC, 0); - e->vpternlogd(zmmA, zmmB, zmmC, 0); - e->vpternlogd(zmmA, zmmB, anyptr_gpC, 0); - e->vpternlogq(xmmA, xmmB, xmmC, 0); - e->vpternlogq(xmmA, xmmB, anyptr_gpC, 0); - e->vpternlogq(ymmA, ymmB, ymmC, 0); - e->vpternlogq(ymmA, ymmB, anyptr_gpC, 0); - e->vpternlogq(zmmA, zmmB, zmmC, 0); - e->vpternlogq(zmmA, zmmB, anyptr_gpC, 0); - e->vptestmb(kA, xmmB, xmmC); - e->vptestmb(kA, xmmB, anyptr_gpC); - e->vptestmb(kA, ymmB, ymmC); - e->vptestmb(kA, ymmB, anyptr_gpC); - e->vptestmb(kA, zmmB, zmmC); - e->vptestmb(kA, zmmB, anyptr_gpC); - e->vptestmd(kA, xmmB, xmmC); - e->vptestmd(kA, xmmB, anyptr_gpC); - e->vptestmd(kA, ymmB, ymmC); - e->vptestmd(kA, ymmB, anyptr_gpC); - e->vptestmd(kA, zmmB, zmmC); - e->vptestmd(kA, zmmB, anyptr_gpC); - e->vptestmq(kA, xmmB, xmmC); - e->vptestmq(kA, xmmB, anyptr_gpC); - e->vptestmq(kA, ymmB, ymmC); - e->vptestmq(kA, ymmB, anyptr_gpC); - e->vptestmq(kA, zmmB, zmmC); - e->vptestmq(kA, zmmB, anyptr_gpC); - e->vptestmw(kA, xmmB, xmmC); - e->vptestmw(kA, xmmB, anyptr_gpC); - e->vptestmw(kA, ymmB, ymmC); - e->vptestmw(kA, ymmB, anyptr_gpC); - e->vptestmw(kA, zmmB, zmmC); - e->vptestmw(kA, zmmB, anyptr_gpC); - e->vptestnmb(kA, xmmB, xmmC); - e->vptestnmb(kA, xmmB, anyptr_gpC); - e->vptestnmb(kA, ymmB, ymmC); - e->vptestnmb(kA, ymmB, anyptr_gpC); - e->vptestnmb(kA, zmmB, zmmC); - e->vptestnmb(kA, zmmB, anyptr_gpC); - e->vptestnmd(kA, xmmB, xmmC); - e->vptestnmd(kA, xmmB, anyptr_gpC); - e->vptestnmd(kA, ymmB, ymmC); - e->vptestnmd(kA, ymmB, anyptr_gpC); - e->vptestnmd(kA, zmmB, zmmC); - e->vptestnmd(kA, zmmB, anyptr_gpC); - e->vptestnmq(kA, xmmB, xmmC); - e->vptestnmq(kA, xmmB, anyptr_gpC); - e->vptestnmq(kA, ymmB, ymmC); - e->vptestnmq(kA, ymmB, anyptr_gpC); - e->vptestnmq(kA, zmmB, zmmC); - e->vptestnmq(kA, zmmB, anyptr_gpC); - e->vptestnmw(kA, xmmB, xmmC); - e->vptestnmw(kA, xmmB, anyptr_gpC); - e->vptestnmw(kA, ymmB, ymmC); - e->vptestnmw(kA, ymmB, anyptr_gpC); - e->vptestnmw(kA, zmmB, zmmC); - e->vptestnmw(kA, zmmB, anyptr_gpC); - e->vpunpckhbw(xmmA, xmmB, xmmC); - e->vpunpckhbw(xmmA, xmmB, anyptr_gpC); - e->vpunpckhbw(ymmA, ymmB, ymmC); - e->vpunpckhbw(ymmA, ymmB, anyptr_gpC); - e->vpunpckhbw(zmmA, zmmB, zmmC); - e->vpunpckhbw(zmmA, zmmB, anyptr_gpC); - e->vpunpckhdq(xmmA, xmmB, xmmC); - e->vpunpckhdq(xmmA, xmmB, anyptr_gpC); - e->vpunpckhdq(ymmA, ymmB, ymmC); - e->vpunpckhdq(ymmA, ymmB, anyptr_gpC); - e->vpunpckhdq(zmmA, zmmB, zmmC); - e->vpunpckhdq(zmmA, zmmB, anyptr_gpC); - e->vpunpckhqdq(xmmA, xmmB, xmmC); - e->vpunpckhqdq(xmmA, xmmB, anyptr_gpC); - e->vpunpckhqdq(ymmA, ymmB, ymmC); - e->vpunpckhqdq(ymmA, ymmB, anyptr_gpC); - e->vpunpckhqdq(zmmA, zmmB, zmmC); - e->vpunpckhqdq(zmmA, zmmB, anyptr_gpC); - e->vpunpckhwd(xmmA, xmmB, xmmC); - e->vpunpckhwd(xmmA, xmmB, anyptr_gpC); - e->vpunpckhwd(ymmA, ymmB, ymmC); - e->vpunpckhwd(ymmA, ymmB, anyptr_gpC); - e->vpunpckhwd(zmmA, zmmB, zmmC); - e->vpunpckhwd(zmmA, zmmB, anyptr_gpC); - e->vpunpcklbw(xmmA, xmmB, xmmC); - e->vpunpcklbw(xmmA, xmmB, anyptr_gpC); - e->vpunpcklbw(ymmA, ymmB, ymmC); - e->vpunpcklbw(ymmA, ymmB, anyptr_gpC); - e->vpunpcklbw(zmmA, zmmB, zmmC); - e->vpunpcklbw(zmmA, zmmB, anyptr_gpC); - e->vpunpckldq(xmmA, xmmB, xmmC); - e->vpunpckldq(xmmA, xmmB, anyptr_gpC); - e->vpunpckldq(ymmA, ymmB, ymmC); - e->vpunpckldq(ymmA, ymmB, anyptr_gpC); - e->vpunpckldq(zmmA, zmmB, zmmC); - e->vpunpckldq(zmmA, zmmB, anyptr_gpC); - e->vpunpcklqdq(xmmA, xmmB, xmmC); - e->vpunpcklqdq(xmmA, xmmB, anyptr_gpC); - e->vpunpcklqdq(ymmA, ymmB, ymmC); - e->vpunpcklqdq(ymmA, ymmB, anyptr_gpC); - e->vpunpcklqdq(zmmA, zmmB, zmmC); - e->vpunpcklqdq(zmmA, zmmB, anyptr_gpC); - e->vpunpcklwd(xmmA, xmmB, xmmC); - e->vpunpcklwd(xmmA, xmmB, anyptr_gpC); - e->vpunpcklwd(ymmA, ymmB, ymmC); - e->vpunpcklwd(ymmA, ymmB, anyptr_gpC); - e->vpunpcklwd(zmmA, zmmB, zmmC); - e->vpunpcklwd(zmmA, zmmB, anyptr_gpC); - e->vpxord(xmmA, xmmB, xmmC); - e->vpxord(xmmA, xmmB, anyptr_gpC); - e->vpxord(ymmA, ymmB, ymmC); - e->vpxord(ymmA, ymmB, anyptr_gpC); - e->vpxord(zmmA, zmmB, zmmC); - e->vpxord(zmmA, zmmB, anyptr_gpC); - e->vpxorq(xmmA, xmmB, xmmC); - e->vpxorq(xmmA, xmmB, anyptr_gpC); - e->vpxorq(ymmA, ymmB, ymmC); - e->vpxorq(ymmA, ymmB, anyptr_gpC); - e->vpxorq(zmmA, zmmB, zmmC); - e->vpxorq(zmmA, zmmB, anyptr_gpC); - e->vrangepd(xmmA, xmmB, xmmC, 0); - e->vrangepd(xmmA, xmmB, anyptr_gpC, 0); - e->vrangepd(ymmA, ymmB, ymmC, 0); - e->vrangepd(ymmA, ymmB, anyptr_gpC, 0); - e->vrangepd(zmmA, zmmB, zmmC, 0); - e->vrangepd(zmmA, zmmB, anyptr_gpC, 0); - e->vrangeps(xmmA, xmmB, xmmC, 0); - e->vrangeps(xmmA, xmmB, anyptr_gpC, 0); - e->vrangeps(ymmA, ymmB, ymmC, 0); - e->vrangeps(ymmA, ymmB, anyptr_gpC, 0); - e->vrangeps(zmmA, zmmB, zmmC, 0); - e->vrangeps(zmmA, zmmB, anyptr_gpC, 0); - e->vrangesd(xmmA, xmmB, xmmC, 0); - e->vrangesd(xmmA, xmmB, anyptr_gpC, 0); - e->vrangess(xmmA, xmmB, xmmC, 0); - e->vrangess(xmmA, xmmB, anyptr_gpC, 0); - e->vrcp14pd(xmmA, xmmB); - e->vrcp14pd(xmmA, anyptr_gpB); - e->vrcp14pd(ymmA, ymmB); - e->vrcp14pd(ymmA, anyptr_gpB); - e->vrcp14pd(zmmA, zmmB); - e->vrcp14pd(zmmA, anyptr_gpB); - e->vrcp14ps(xmmA, xmmB); - e->vrcp14ps(xmmA, anyptr_gpB); - e->vrcp14ps(ymmA, ymmB); - e->vrcp14ps(ymmA, anyptr_gpB); - e->vrcp14ps(zmmA, zmmB); - e->vrcp14ps(zmmA, anyptr_gpB); - e->vrcp14sd(xmmA, xmmB, xmmC); - e->vrcp14sd(xmmA, xmmB, anyptr_gpC); - e->vrcp14ss(xmmA, xmmB, xmmC); - e->vrcp14ss(xmmA, xmmB, anyptr_gpC); - e->vrcp28pd(zmmA, zmmB); - e->vrcp28pd(zmmA, anyptr_gpB); - e->vrcp28ps(zmmA, zmmB); - e->vrcp28ps(zmmA, anyptr_gpB); - e->vrcp28sd(xmmA, xmmB, xmmC); - e->vrcp28sd(xmmA, xmmB, anyptr_gpC); - e->vrcp28ss(xmmA, xmmB, xmmC); - e->vrcp28ss(xmmA, xmmB, anyptr_gpC); - e->vreducepd(xmmA, xmmB, 0); - e->vreducepd(xmmA, anyptr_gpB, 0); - e->vreducepd(ymmA, ymmB, 0); - e->vreducepd(ymmA, anyptr_gpB, 0); - e->vreducepd(zmmA, zmmB, 0); - e->vreducepd(zmmA, anyptr_gpB, 0); - e->vreduceps(xmmA, xmmB, 0); - e->vreduceps(xmmA, anyptr_gpB, 0); - e->vreduceps(ymmA, ymmB, 0); - e->vreduceps(ymmA, anyptr_gpB, 0); - e->vreduceps(zmmA, zmmB, 0); - e->vreduceps(zmmA, anyptr_gpB, 0); - e->vreducesd(xmmA, xmmB, xmmC, 0); - e->vreducesd(xmmA, xmmB, anyptr_gpC, 0); - e->vreducess(xmmA, xmmB, xmmC, 0); - e->vreducess(xmmA, xmmB, anyptr_gpC, 0); - e->vrndscalepd(xmmA, xmmB, 0); - e->vrndscalepd(xmmA, anyptr_gpB, 0); - e->vrndscalepd(ymmA, ymmB, 0); - e->vrndscalepd(ymmA, anyptr_gpB, 0); - e->vrndscalepd(zmmA, zmmB, 0); - e->vrndscalepd(zmmA, anyptr_gpB, 0); - e->vrndscaleps(xmmA, xmmB, 0); - e->vrndscaleps(xmmA, anyptr_gpB, 0); - e->vrndscaleps(ymmA, ymmB, 0); - e->vrndscaleps(ymmA, anyptr_gpB, 0); - e->vrndscaleps(zmmA, zmmB, 0); - e->vrndscaleps(zmmA, anyptr_gpB, 0); - e->vrndscalesd(xmmA, xmmB, xmmC, 0); - e->vrndscalesd(xmmA, xmmB, anyptr_gpC, 0); - e->vrndscaless(xmmA, xmmB, xmmC, 0); - e->vrndscaless(xmmA, xmmB, anyptr_gpC, 0); - e->vrsqrt14pd(xmmA, xmmB); - e->vrsqrt14pd(xmmA, anyptr_gpB); - e->vrsqrt14pd(ymmA, ymmB); - e->vrsqrt14pd(ymmA, anyptr_gpB); - e->vrsqrt14pd(zmmA, zmmB); - e->vrsqrt14pd(zmmA, anyptr_gpB); - e->vrsqrt14ps(xmmA, xmmB); - e->vrsqrt14ps(xmmA, anyptr_gpB); - e->vrsqrt14ps(ymmA, ymmB); - e->vrsqrt14ps(ymmA, anyptr_gpB); - e->vrsqrt14ps(zmmA, zmmB); - e->vrsqrt14ps(zmmA, anyptr_gpB); - e->vrsqrt14sd(xmmA, xmmB, xmmC); - e->vrsqrt14sd(xmmA, xmmB, anyptr_gpC); - e->vrsqrt14ss(xmmA, xmmB, xmmC); - e->vrsqrt14ss(xmmA, xmmB, anyptr_gpC); - e->vrsqrt28pd(zmmA, zmmB); - e->vrsqrt28pd(zmmA, anyptr_gpB); - e->vrsqrt28ps(zmmA, zmmB); - e->vrsqrt28ps(zmmA, anyptr_gpB); - e->vrsqrt28sd(xmmA, xmmB, xmmC); - e->vrsqrt28sd(xmmA, xmmB, anyptr_gpC); - e->vrsqrt28ss(xmmA, xmmB, xmmC); - e->vrsqrt28ss(xmmA, xmmB, anyptr_gpC); - e->vscalefpd(xmmA, xmmB, xmmC); - e->vscalefpd(xmmA, xmmB, anyptr_gpC); - e->vscalefpd(ymmA, ymmB, ymmC); - e->vscalefpd(ymmA, ymmB, anyptr_gpC); - e->vscalefpd(zmmA, zmmB, zmmC); - e->vscalefpd(zmmA, zmmB, anyptr_gpC); - e->vscalefps(xmmA, xmmB, xmmC); - e->vscalefps(xmmA, xmmB, anyptr_gpC); - e->vscalefps(ymmA, ymmB, ymmC); - e->vscalefps(ymmA, ymmB, anyptr_gpC); - e->vscalefps(zmmA, zmmB, zmmC); - e->vscalefps(zmmA, zmmB, anyptr_gpC); - e->vscalefsd(xmmA, xmmB, xmmC); - e->vscalefsd(xmmA, xmmB, anyptr_gpC); - e->vscalefss(xmmA, xmmB, xmmC); - e->vscalefss(xmmA, xmmB, anyptr_gpC); - e->vscatterdpd(vx_ptr, xmmB); - e->vscatterdpd(vx_ptr, ymmB); - e->vscatterdpd(vy_ptr, zmmB); - e->vscatterdps(vx_ptr, xmmB); - e->vscatterdps(vy_ptr, ymmB); - e->vscatterdps(vz_ptr, zmmB); - e->vscatterpf0dpd(vy_ptr); - e->vscatterpf0dps(vz_ptr); - e->vscatterpf0qpd(vz_ptr); - e->vscatterpf0qps(vz_ptr); - e->vscatterpf1dpd(vy_ptr); - e->vscatterpf1dps(vz_ptr); - e->vscatterpf1qpd(vz_ptr); - e->vscatterpf1qps(vz_ptr); - e->vscatterqpd(vx_ptr, xmmB); - e->vscatterqpd(vy_ptr, ymmB); - e->vscatterqpd(vz_ptr, zmmB); - e->vscatterqps(vx_ptr, xmmB); - e->vscatterqps(vy_ptr, xmmB); - e->vscatterqps(vz_ptr, ymmB); - e->vshuff32x4(ymmA, ymmB, ymmC, 0); - e->vshuff32x4(ymmA, ymmB, anyptr_gpC, 0); - e->vshuff32x4(zmmA, zmmB, zmmC, 0); - e->vshuff32x4(zmmA, zmmB, anyptr_gpC, 0); - e->vshuff64x2(ymmA, ymmB, ymmC, 0); - e->vshuff64x2(ymmA, ymmB, anyptr_gpC, 0); - e->vshuff64x2(zmmA, zmmB, zmmC, 0); - e->vshuff64x2(zmmA, zmmB, anyptr_gpC, 0); - e->vshufi32x4(ymmA, ymmB, ymmC, 0); - e->vshufi32x4(ymmA, ymmB, anyptr_gpC, 0); - e->vshufi32x4(zmmA, zmmB, zmmC, 0); - e->vshufi32x4(zmmA, zmmB, anyptr_gpC, 0); - e->vshufi64x2(ymmA, ymmB, ymmC, 0); - e->vshufi64x2(ymmA, ymmB, anyptr_gpC, 0); - e->vshufi64x2(zmmA, zmmB, zmmC, 0); - e->vshufi64x2(zmmA, zmmB, anyptr_gpC, 0); - e->vshufpd(xmmA, xmmB, xmmC, 0); - e->vshufpd(xmmA, xmmB, anyptr_gpC, 0); - e->vshufpd(ymmA, ymmB, ymmC, 0); - e->vshufpd(ymmA, ymmB, anyptr_gpC, 0); - e->vshufpd(zmmA, zmmB, zmmC, 0); - e->vshufpd(zmmA, zmmB, anyptr_gpC, 0); - e->vshufps(xmmA, xmmB, xmmC, 0); - e->vshufps(xmmA, xmmB, anyptr_gpC, 0); - e->vshufps(ymmA, ymmB, ymmC, 0); - e->vshufps(ymmA, ymmB, anyptr_gpC, 0); - e->vshufps(zmmA, zmmB, zmmC, 0); - e->vshufps(zmmA, zmmB, anyptr_gpC, 0); - e->vsqrtpd(xmmA, xmmB); - e->vsqrtpd(xmmA, anyptr_gpB); - e->vsqrtpd(ymmA, ymmB); - e->vsqrtpd(ymmA, anyptr_gpB); - e->vsqrtpd(zmmA, zmmB); - e->vsqrtpd(zmmA, anyptr_gpB); - e->vsqrtps(xmmA, xmmB); - e->vsqrtps(xmmA, anyptr_gpB); - e->vsqrtps(ymmA, ymmB); - e->vsqrtps(ymmA, anyptr_gpB); - e->vsqrtps(zmmA, zmmB); - e->vsqrtps(zmmA, anyptr_gpB); - e->vsqrtsd(xmmA, xmmB, xmmC); - e->vsqrtsd(xmmA, xmmB, anyptr_gpC); - e->vsqrtss(xmmA, xmmB, xmmC); - e->vsqrtss(xmmA, xmmB, anyptr_gpC); - e->vsubpd(xmmA, xmmB, xmmC); - e->vsubpd(xmmA, xmmB, anyptr_gpC); - e->vsubpd(ymmA, ymmB, ymmC); - e->vsubpd(ymmA, ymmB, anyptr_gpC); - e->vsubpd(zmmA, zmmB, zmmC); - e->vsubpd(zmmA, zmmB, anyptr_gpC); - e->vsubps(xmmA, xmmB, xmmC); - e->vsubps(xmmA, xmmB, anyptr_gpC); - e->vsubps(ymmA, ymmB, ymmC); - e->vsubps(ymmA, ymmB, anyptr_gpC); - e->vsubps(zmmA, zmmB, zmmC); - e->vsubps(zmmA, zmmB, anyptr_gpC); - e->vsubsd(xmmA, xmmB, xmmC); - e->vsubsd(xmmA, xmmB, anyptr_gpC); - e->vsubss(xmmA, xmmB, xmmC); - e->vsubss(xmmA, xmmB, anyptr_gpC); - e->vucomisd(xmmA, xmmB); - e->vucomisd(xmmA, anyptr_gpB); - e->vucomiss(xmmA, xmmB); - e->vucomiss(xmmA, anyptr_gpB); - e->vunpckhpd(xmmA, xmmB, xmmC); - e->vunpckhpd(xmmA, xmmB, anyptr_gpC); - e->vunpckhpd(ymmA, ymmB, ymmC); - e->vunpckhpd(ymmA, ymmB, anyptr_gpC); - e->vunpckhpd(zmmA, zmmB, zmmC); - e->vunpckhpd(zmmA, zmmB, anyptr_gpC); - e->vunpckhps(xmmA, xmmB, xmmC); - e->vunpckhps(xmmA, xmmB, anyptr_gpC); - e->vunpckhps(ymmA, ymmB, ymmC); - e->vunpckhps(ymmA, ymmB, anyptr_gpC); - e->vunpckhps(zmmA, zmmB, zmmC); - e->vunpckhps(zmmA, zmmB, anyptr_gpC); - e->vunpcklpd(xmmA, xmmB, xmmC); - e->vunpcklpd(xmmA, xmmB, anyptr_gpC); - e->vunpcklpd(ymmA, ymmB, ymmC); - e->vunpcklpd(ymmA, ymmB, anyptr_gpC); - e->vunpcklpd(zmmA, zmmB, zmmC); - e->vunpcklpd(zmmA, zmmB, anyptr_gpC); - e->vunpcklps(xmmA, xmmB, xmmC); - e->vunpcklps(xmmA, xmmB, anyptr_gpC); - e->vunpcklps(ymmA, ymmB, ymmC); - e->vunpcklps(ymmA, ymmB, anyptr_gpC); - e->vunpcklps(zmmA, zmmB, zmmC); - e->vunpcklps(zmmA, zmmB, anyptr_gpC); - e->vxorpd(xmmA, xmmB, xmmC); - e->vxorpd(xmmA, xmmB, anyptr_gpC); - e->vxorpd(ymmA, ymmB, ymmC); - e->vxorpd(ymmA, ymmB, anyptr_gpC); - e->vxorpd(zmmA, zmmB, zmmC); - e->vxorpd(zmmA, zmmB, anyptr_gpC); - e->vxorps(xmmA, xmmB, xmmC); - e->vxorps(xmmA, xmmB, anyptr_gpC); - e->vxorps(ymmA, ymmB, ymmC); - e->vxorps(ymmA, ymmB, anyptr_gpC); - e->vxorps(zmmA, zmmB, zmmC); - e->vxorps(zmmA, zmmB, anyptr_gpC); - - // Mark the end. - e->nop(); - e->nop(); - e->nop(); - e->nop(); -} - -} // {asmtest} - -#endif // ASMJIT_TEST_OPCODE_H_INCLUDED diff --git a/test/asmjit_test_perf.cpp b/test/asmjit_test_perf.cpp new file mode 100644 index 0000000..a1b638e --- /dev/null +++ b/test/asmjit_test_perf.cpp @@ -0,0 +1,69 @@ +// AsmJit - Machine code generation for C++ +// +// * Official AsmJit Home Page: https://asmjit.com +// * Official Github Repository: https://github.com/asmjit/asmjit +// +// Copyright (c) 2008-2020 The AsmJit Authors +// +// This software is provided 'as-is', without any express or implied +// warranty. In no event will the authors be held liable for any damages +// arising from the use of this software. +// +// Permission is granted to anyone to use this software for any purpose, +// including commercial applications, and to alter it and redistribute it +// freely, subject to the following restrictions: +// +// 1. The origin of this software must not be misrepresented; you must not +// claim that you wrote the original software. If you use this software +// in a product, an acknowledgment in the product documentation would be +// appreciated but is not required. +// 2. Altered source versions must be plainly marked as such, and must not be +// misrepresented as being the original software. +// 3. This notice may not be removed or altered from any source distribution. + +#include <asmjit/core.h> +#include <stdio.h> +#include <stdlib.h> +#include <string.h> + +#include "cmdline.h" + +using namespace asmjit; + +#if defined(ASMJIT_BUILD_X86) +void benchmarkX86Emitters(uint32_t numIterations, bool testX86, bool testX64) noexcept; +#endif + +int main(int argc, char* argv[]) { + CmdLine cmdLine(argc, argv); + uint32_t numIterations = 20000; + + printf("AsmJit Performance Suite v%u.%u.%u:\n\n", + unsigned((ASMJIT_LIBRARY_VERSION >> 16) ), + unsigned((ASMJIT_LIBRARY_VERSION >> 8) & 0xFF), + unsigned((ASMJIT_LIBRARY_VERSION ) & 0xFF)); + + printf("Usage:\n"); + printf(" --help Show usage only\n"); + printf(" --quick Decrease the number of iterations to make tests quicker\n"); + printf(" --arch=<ARCH> Select architecture to run ('all' by default)\n"); + printf("\n"); + + if (cmdLine.hasArg("--help")) + return 0; + + if (cmdLine.hasArg("--quick")) + numIterations = 1000; + + const char* arch = cmdLine.valueOf("--arch", "all"); + +#if defined(ASMJIT_BUILD_X86) + bool testX86 = strcmp(arch, "all") == 0 || strcmp(arch, "x86") == 0; + bool testX64 = strcmp(arch, "all") == 0 || strcmp(arch, "x64") == 0; + + if (testX86 || testX64) + benchmarkX86Emitters(numIterations, testX86, testX64); +#endif + + return 0; +} diff --git a/test/asmjit_test_perf.h b/test/asmjit_test_perf.h new file mode 100644 index 0000000..565c18a --- /dev/null +++ b/test/asmjit_test_perf.h @@ -0,0 +1,81 @@ +// AsmJit - Machine code generation for C++ +// +// * Official AsmJit Home Page: https://asmjit.com +// * Official Github Repository: https://github.com/asmjit/asmjit +// +// Copyright (c) 2008-2020 The AsmJit Authors +// +// This software is provided 'as-is', without any express or implied +// warranty. In no event will the authors be held liable for any damages +// arising from the use of this software. +// +// Permission is granted to anyone to use this software for any purpose, +// including commercial applications, and to alter it and redistribute it +// freely, subject to the following restrictions: +// +// 1. The origin of this software must not be misrepresented; you must not +// claim that you wrote the original software. If you use this software +// in a product, an acknowledgment in the product documentation would be +// appreciated but is not required. +// 2. Altered source versions must be plainly marked as such, and must not be +// misrepresented as being the original software. +// 3. This notice may not be removed or altered from any source distribution. + +#ifndef ASMJIT_TEST_PERF_H_INCLUDED +#define ASMJIT_TEST_PERF_H_INCLUDED + +#include <asmjit/core.h> +#include "performancetimer.h" + +class MyErrorHandler : public asmjit::ErrorHandler { + void handleError(asmjit::Error err, const char* message, asmjit::BaseEmitter* origin) { + (void)err; + (void)origin; + printf("ERROR: %s\n", message); + abort(); + } +}; + +template<typename EmitterT, typename FuncT> +static void bench(asmjit::CodeHolder& code, uint32_t arch, uint32_t numIterations, const char* testName, const FuncT& func) noexcept { + EmitterT emitter; + MyErrorHandler eh; + + const char* archName = + arch == asmjit::Environment::kArchX86 ? "X86" : + arch == asmjit::Environment::kArchX64 ? "X64" : "???"; + + const char* emitterName = + emitter.isAssembler() ? "Assembler" : + emitter.isCompiler() ? "Compiler" : + emitter.isBuilder() ? "Builder" : "Unknown"; + + uint64_t codeSize = 0; + asmjit::Environment env(arch); + + PerformanceTimer timer; + double duration = std::numeric_limits<double>::infinity(); + + for (uint32_t r = 0; r < numIterations; r++) { + codeSize = 0; + code.init(env); + code.setErrorHandler(&eh); + code.attach(&emitter); + + timer.start(); + func(emitter); + timer.stop(); + + codeSize += code.codeSize(); + + code.reset(); + duration = asmjit::Support::min(duration, timer.duration()); + } + + printf(" [%s] %-9s %-16s | CodeSize:%5llu [B] | Time:%8.4f [ms]", archName, emitterName, testName, (unsigned long long)codeSize, duration); + if (codeSize) + printf(" | Speed:%8.3f [MB/s]", mbps(duration, codeSize)); + printf("\n"); +} + +#endif // ASMJIT_TEST_PERF_H_INCLUDED diff --git a/test/asmjit_test_perf_x86.cpp b/test/asmjit_test_perf_x86.cpp new file mode 100644 index 0000000..cfbaa1f --- /dev/null +++ b/test/asmjit_test_perf_x86.cpp @@ -0,0 +1,5049 @@ +// AsmJit - Machine code generation for C++ +// +// * Official AsmJit Home Page: https://asmjit.com +// * Official Github Repository: https://github.com/asmjit/asmjit +// +// Copyright (c) 2008-2020 The AsmJit Authors +// +// This software is provided 'as-is', without any express or implied +// warranty. In no event will the authors be held liable for any damages +// arising from the use of this software. +// +// Permission is granted to anyone to use this software for any purpose, +// including commercial applications, and to alter it and redistribute it +// freely, subject to the following restrictions: +// +// 1. The origin of this software must not be misrepresented; you must not +// claim that you wrote the original software. If you use this software +// in a product, an acknowledgment in the product documentation would be +// appreciated but is not required. +// 2. Altered source versions must be plainly marked as such, and must not be +// misrepresented as being the original software. +// 3. This notice may not be removed or altered from any source distribution. + +#include <asmjit/core.h> + +#ifdef ASMJIT_BUILD_X86 +#include <asmjit/x86.h> + +#include <limits> +#include <stdio.h> +#include <string.h> + +#include "asmjit_test_misc.h" +#include "asmjit_test_perf.h" + +using namespace asmjit; + +enum class InstForm { + kReg, + kMem +}; + +// Generates a long sequence of GP instructions. +template<typename Emitter> +static void generateGpSequenceInternal( + Emitter& cc, + InstForm form, + const x86::Gp& a, const x86::Gp& b, const x86::Gp& c, const x86::Gp& d) { + + cc.mov(a, 0xAAAAAAAA); + cc.mov(b, 0xBBBBBBBB); + cc.mov(c, 0xCCCCCCCC); + cc.mov(d, 0xFFFFFFFF); + + if (form == InstForm::kReg) { + cc.adc(a, b); + cc.adc(b, c); + cc.adc(c, d); + cc.add(a, b); + cc.add(b, c); + cc.add(c, d); + cc.and_(a, b); + cc.and_(b, c); + cc.and_(c, d); + cc.bsf(a, b); + cc.bsf(b, c); + cc.bsf(c, d); + cc.bsr(a, b); + cc.bsr(b, c); + cc.bsr(c, d); + cc.bswap(a); + cc.bswap(b); + cc.bswap(c); + cc.bt(a, b); + cc.bt(b, c); + cc.bt(c, d); + cc.btc(a, b); + cc.btc(b, c); + cc.btc(c, d); + cc.btr(a, b); + cc.btr(b, c); + cc.btr(c, d); + cc.bts(a, b); + cc.bts(b, c); + cc.bts(c, d); + cc.cmp(a, b); + cc.cmovc(a, b); + cc.cmp(b, c); + cc.cmovc(b, c); + cc.cmp(c, d); + cc.cmovc(c, d); + cc.dec(a); + cc.dec(b); + cc.dec(c); + cc.imul(a, b); + cc.imul(b, c); + cc.imul(c, d); + cc.movsx(a, b.r8Lo()); + cc.movsx(b, c.r8Lo()); + cc.movsx(c, d.r8Lo()); + cc.movzx(a, b.r8Lo()); + cc.movzx(b, c.r8Lo()); + cc.movzx(c, d.r8Lo()); + cc.neg(a); + cc.neg(b); + cc.neg(c); + cc.not_(a); + cc.not_(b); + cc.not_(c); + cc.or_(a, b); + cc.or_(b, c); + cc.or_(c, d); + cc.sbb(a, b); + cc.sbb(b, c); + cc.sbb(c, d); + cc.sub(a, b); + cc.sub(b, c); + cc.sub(c, d); + cc.test(a, b); + cc.test(b, c); + cc.test(c, d); + cc.xchg(a, b); + cc.xchg(b, c); + cc.xchg(c, d); + cc.xor_(a, b); + cc.xor_(b, c); + cc.xor_(c, d); + + cc.rcl(a, c.r8Lo()); + cc.rcl(b, c.r8Lo()); + cc.rcl(d, c.r8Lo()); + cc.rcr(a, c.r8Lo()); + cc.rcr(b, c.r8Lo()); + cc.rcr(d, c.r8Lo()); + cc.rol(a, c.r8Lo()); + cc.rol(b, c.r8Lo()); + cc.rol(d, c.r8Lo()); + cc.ror(a, c.r8Lo()); + cc.ror(b, c.r8Lo()); + cc.ror(d, c.r8Lo()); + cc.shl(a, c.r8Lo()); + cc.shl(b, c.r8Lo()); + cc.shl(d, c.r8Lo()); + cc.shr(a, c.r8Lo()); + cc.shr(b, c.r8Lo()); + cc.shr(d, c.r8Lo()); + cc.sar(a, c.r8Lo()); + cc.sar(b, c.r8Lo()); + cc.sar(d, c.r8Lo()); + cc.shld(a, b, c.r8Lo()); + cc.shld(b, d, c.r8Lo()); + cc.shld(d, a, c.r8Lo()); + cc.shrd(a, b, c.r8Lo()); + cc.shrd(b, d, c.r8Lo()); + cc.shrd(d, a, c.r8Lo()); + + cc.adcx(a, b); + cc.adox(a, b); + cc.adcx(b, c); + cc.adox(b, c); + cc.adcx(c, d); + cc.adox(c, d); + cc.andn(a, b, c); + cc.andn(b, c, d); + cc.andn(c, d, a); + cc.bextr(a, b, c); + cc.bextr(b, c, d); + cc.bextr(c, d, a); + cc.blsi(a, b); + cc.blsi(b, c); + cc.blsi(c, d); + cc.blsmsk(a, b); + cc.blsmsk(b, c); + cc.blsmsk(c, d); + cc.blsr(a, b); + cc.blsr(b, c); + cc.blsr(c, d); + cc.bzhi(a, b, c); + cc.bzhi(b, c, d); + cc.bzhi(c, d, a); + cc.lzcnt(a, b); + cc.lzcnt(b, c); + cc.lzcnt(c, d); + cc.pdep(a, b, c); + cc.pdep(b, c, d); + cc.pdep(c, d, a); + cc.pext(a, b, c); + cc.pext(b, c, d); + cc.pext(c, d, a); + cc.popcnt(a, b); + cc.popcnt(b, c); + cc.popcnt(c, d); + cc.rorx(a, b, 8); + cc.rorx(b, c, 8); + cc.rorx(c, d, 8); + cc.sarx(a, b, c); + cc.sarx(b, c, d); + cc.sarx(c, d, a); + cc.shlx(a, b, c); + cc.shlx(b, c, d); + cc.shlx(c, d, a); + cc.shrx(a, b, c); + cc.shrx(b, c, d); + cc.shrx(c, d, a); + cc.tzcnt(a, b); + cc.tzcnt(b, c); + cc.tzcnt(c, d); + } + else { + uint32_t regSize = cc.registerSize(); + x86::Mem m = x86::ptr(c, 0, regSize); + x86::Mem m8 = x86::byte_ptr(c); + + cc.adc(a, m); + cc.adc(b, m); + cc.adc(c, m); + cc.add(a, m); + cc.add(b, m); + cc.add(c, m); + cc.and_(a, m); + cc.and_(b, m); + cc.and_(c, m); + cc.bsf(a, m); + cc.bsf(b, m); + cc.bsf(c, m); + cc.bsr(a, m); + cc.bsr(b, m); + cc.bsr(c, m); + cc.bt(m, a); + cc.bt(m, b); + cc.bt(m, c); + cc.btc(m, a); + cc.btc(m, b); + cc.btc(m, c); + cc.btr(m, a); + cc.btr(m, b); + cc.btr(m, c); + cc.bts(m, a); + cc.bts(m, b); + cc.bts(m, c); + cc.cmp(a, m); + cc.cmovc(a, m); + cc.cmp(b, m); + cc.cmovc(b, m); + cc.cmp(c, m); + cc.cmovc(c, m); + cc.dec(m); + cc.movsx(a, m8); + cc.movsx(b, m8); + cc.movsx(c, m8); + cc.movzx(a, m8); + cc.movzx(b, m8); + cc.movzx(c, m8); + cc.neg(m); + cc.not_(m); + cc.or_(a, m); + cc.or_(b, m); + cc.or_(c, m); + cc.sbb(a, m); + cc.sbb(b, m); + cc.sbb(c, m); + cc.sub(a, m); + cc.sub(b, m); + cc.sub(c, m); + cc.test(m, a); + cc.test(m, b); + cc.test(m, c); + cc.xchg(a, m); + cc.xchg(b, m); + cc.xchg(c, m); + cc.xor_(a, m); + cc.xor_(b, m); + cc.xor_(c, m); + + cc.rcl(m, c.r8Lo()); + cc.rcr(m, c.r8Lo()); + cc.rol(m, c.r8Lo()); + cc.ror(m, c.r8Lo()); + cc.shl(m, c.r8Lo()); + cc.shr(m, c.r8Lo()); + cc.sar(m, c.r8Lo()); + cc.shld(m, b, c.r8Lo()); + cc.shld(m, d, c.r8Lo()); + cc.shld(m, a, c.r8Lo()); + cc.shrd(m, b, c.r8Lo()); + cc.shrd(m, d, c.r8Lo()); + cc.shrd(m, a, c.r8Lo()); + + cc.adcx(a, m); + cc.adox(a, m); + cc.adcx(b, m); + cc.adox(b, m); + cc.adcx(c, m); + cc.adox(c, m); + cc.andn(a, b, m); + cc.andn(b, c, m); + cc.andn(c, d, m); + cc.bextr(a, m, c); + cc.bextr(b, m, d); + cc.bextr(c, m, a); + cc.blsi(a, m); + cc.blsi(b, m); + cc.blsi(c, m); + cc.blsmsk(a, m); + cc.blsmsk(b, m); + cc.blsmsk(c, m); + cc.blsr(a, m); + cc.blsr(b, m); + cc.blsr(c, m); + cc.bzhi(a, m, c); + cc.bzhi(b, m, d); + cc.bzhi(c, m, a); + cc.lzcnt(a, m); + cc.lzcnt(b, m); + cc.lzcnt(c, m); + cc.pdep(a, b, m); + cc.pdep(b, c, m); + cc.pdep(c, d, m); + cc.pext(a, b, m); + cc.pext(b, c, m); + cc.pext(c, d, m); + cc.popcnt(a, m); + cc.popcnt(b, m); + cc.popcnt(c, m); + cc.rorx(a, m, 8); + cc.rorx(b, m, 8); + cc.rorx(c, m, 8); + cc.sarx(a, m, c); + cc.sarx(b, m, d); + cc.sarx(c, m, a); + cc.shlx(a, m, c); + cc.shlx(b, m, d); + cc.shlx(c, m, a); + cc.shrx(a, m, c); + cc.shrx(b, m, d); + cc.shrx(c, m, a); + cc.tzcnt(a, m); + cc.tzcnt(b, m); + cc.tzcnt(c, m); + } +} + +static void generateGpSequence(BaseEmitter& emitter, InstForm form, bool emitPrologEpilog) { + using namespace asmjit::x86; + + if (emitter.isAssembler()) { + Assembler& cc = *emitter.as<Assembler>(); + + x86::Gp a = cc.zax(); + x86::Gp b = cc.zbx(); + x86::Gp c = cc.zcx(); + x86::Gp d = cc.zdx(); + + if (emitPrologEpilog) { + FuncDetail func; + func.init(FuncSignatureT<void, void*, const void*, size_t>(CallConv::kIdHost), cc.environment()); + + FuncFrame frame; + frame.init(func); + frame.addDirtyRegs(a, b, c, d); + frame.finalize(); + + cc.emitProlog(frame); + generateGpSequenceInternal(cc, form, a, b, c, d); + cc.emitEpilog(frame); + } + else { + generateGpSequenceInternal(cc, form, a, b, c, d); + } + } +#ifndef ASMJIT_NO_BUILDER + else if (emitter.isBuilder()) { + Builder& cc = *emitter.as<Builder>(); + + x86::Gp a = cc.zax(); + x86::Gp b = cc.zbx(); + x86::Gp c = cc.zcx(); + x86::Gp d = cc.zdx(); + + if (emitPrologEpilog) { + FuncDetail func; + func.init(FuncSignatureT<void, void*, const void*, size_t>(CallConv::kIdHost), cc.environment()); + + FuncFrame frame; + frame.init(func); + frame.addDirtyRegs(a, b, c, d); + frame.finalize(); + + cc.emitProlog(frame); + generateGpSequenceInternal(cc, form, a, b, c, d); + cc.emitEpilog(frame); + } + else { + generateGpSequenceInternal(cc, form, a, b, c, d); + } + } +#endif +#ifndef ASMJIT_NO_COMPILER + else if (emitter.isCompiler()) { + Compiler& cc = *emitter.as<Compiler>(); + + Gp a = cc.newIntPtr("a"); + Gp b = cc.newIntPtr("b"); + Gp c = cc.newIntPtr("c"); + Gp d = cc.newIntPtr("d"); + + cc.addFunc(FuncSignatureT<void>(CallConv::kIdHost)); + generateGpSequenceInternal(cc, form, a, b, c, d); + cc.endFunc(); + } +#endif +} + +// Generates a long sequence of SSE instructions using only registers. +template<typename Emitter> +static void generateSseSequenceInternal( + Emitter& cc, + InstForm form, + const x86::Gp& gp, + const x86::Xmm& xmmA, const x86::Xmm& xmmB, const x86::Xmm& xmmC, const x86::Xmm& xmmD) { + + x86::Gp gpd = gp.r32(); + x86::Gp gpq = gp.r64(); + x86::Gp gpz = cc.is32Bit() ? gpd : gpq; + + cc.xor_(gpd, gpd); + cc.xorps(xmmA, xmmA); + cc.xorps(xmmB, xmmB); + cc.xorps(xmmC, xmmC); + cc.xorps(xmmD, xmmD); + + if (form == InstForm::kReg) { + // SSE. + cc.addps(xmmA, xmmB); + cc.addss(xmmA, xmmB); + cc.andnps(xmmA, xmmB); + cc.andps(xmmA, xmmB); + cc.cmpps(xmmA, xmmB, 0); + cc.cmpss(xmmA, xmmB, 0); + cc.comiss(xmmA, xmmB); + cc.cvtsi2ss(xmmA, gpd); + cc.cvtsi2ss(xmmA, gpz); + cc.cvtss2si(gpd, xmmB); + cc.cvtss2si(gpz, xmmB); + cc.cvttss2si(gpd, xmmB); + cc.cvttss2si(gpz, xmmB); + cc.divps(xmmA, xmmB); + cc.divss(xmmA, xmmB); + cc.maxps(xmmA, xmmB); + cc.maxss(xmmA, xmmB); + cc.minps(xmmA, xmmB); + cc.minss(xmmA, xmmB); + cc.movaps(xmmA, xmmB); + cc.movd(gpd, xmmB); + cc.movd(xmmA, gpd); + cc.movq(xmmA, xmmB); + cc.movhlps(xmmA, xmmB); + cc.movlhps(xmmA, xmmB); + cc.movups(xmmA, xmmB); + cc.mulps(xmmA, xmmB); + cc.mulss(xmmA, xmmB); + cc.orps(xmmA, xmmB); + cc.rcpps(xmmA, xmmB); + cc.rcpss(xmmA, xmmB); + cc.psadbw(xmmA, xmmB); + cc.rsqrtps(xmmA, xmmB); + cc.rsqrtss(xmmA, xmmB); + cc.sfence(); + cc.shufps(xmmA, xmmB, 0); + cc.sqrtps(xmmA, xmmB); + cc.sqrtss(xmmA, xmmB); + cc.subps(xmmA, xmmB); + cc.subss(xmmA, xmmB); + cc.ucomiss(xmmA, xmmB); + cc.unpckhps(xmmA, xmmB); + cc.unpcklps(xmmA, xmmB); + cc.xorps(xmmA, xmmB); + + // SSE2. + cc.addpd(xmmA, xmmB); + cc.addsd(xmmA, xmmB); + cc.andnpd(xmmA, xmmB); + cc.andpd(xmmA, xmmB); + cc.cmppd(xmmA, xmmB, 0); + cc.cmpsd(xmmA, xmmB, 0); + cc.comisd(xmmA, xmmB); + cc.cvtdq2pd(xmmA, xmmB); + cc.cvtdq2ps(xmmA, xmmB); + cc.cvtpd2dq(xmmA, xmmB); + cc.cvtpd2ps(xmmA, xmmB); + cc.cvtps2dq(xmmA, xmmB); + cc.cvtps2pd(xmmA, xmmB); + cc.cvtsd2si(gpd, xmmB); + cc.cvtsd2si(gpz, xmmB); + cc.cvtsd2ss(xmmA, xmmB); + cc.cvtsi2sd(xmmA, gpd); + cc.cvtsi2sd(xmmA, gpz); + cc.cvtss2sd(xmmA, xmmB); + cc.cvtss2si(gpd, xmmB); + cc.cvtss2si(gpz, xmmB); + cc.cvttpd2dq(xmmA, xmmB); + cc.cvttps2dq(xmmA, xmmB); + cc.cvttsd2si(gpd, xmmB); + cc.cvttsd2si(gpz, xmmB); + cc.divpd(xmmA, xmmB); + cc.divsd(xmmA, xmmB); + cc.maxpd(xmmA, xmmB); + cc.maxsd(xmmA, xmmB); + cc.minpd(xmmA, xmmB); + cc.minsd(xmmA, xmmB); + cc.movdqa(xmmA, xmmB); + cc.movdqu(xmmA, xmmB); + cc.movmskps(gpd, xmmB); + cc.movmskpd(gpd, xmmB); + cc.movsd(xmmA, xmmB); + cc.mulpd(xmmA, xmmB); + cc.mulsd(xmmA, xmmB); + cc.orpd(xmmA, xmmB); + cc.packsswb(xmmA, xmmB); + cc.packssdw(xmmA, xmmB); + cc.packuswb(xmmA, xmmB); + cc.paddb(xmmA, xmmB); + cc.paddw(xmmA, xmmB); + cc.paddd(xmmA, xmmB); + cc.paddq(xmmA, xmmB); + cc.paddsb(xmmA, xmmB); + cc.paddsw(xmmA, xmmB); + cc.paddusb(xmmA, xmmB); + cc.paddusw(xmmA, xmmB); + cc.pand(xmmA, xmmB); + cc.pandn(xmmA, xmmB); + cc.pavgb(xmmA, xmmB); + cc.pavgw(xmmA, xmmB); + cc.pcmpeqb(xmmA, xmmB); + cc.pcmpeqw(xmmA, xmmB); + cc.pcmpeqd(xmmA, xmmB); + cc.pcmpgtb(xmmA, xmmB); + cc.pcmpgtw(xmmA, xmmB); + cc.pcmpgtd(xmmA, xmmB); + cc.pmaxsw(xmmA, xmmB); + cc.pmaxub(xmmA, xmmB); + cc.pminsw(xmmA, xmmB); + cc.pminub(xmmA, xmmB); + cc.pmovmskb(gpd, xmmB); + cc.pmulhw(xmmA, xmmB); + cc.pmulhuw(xmmA, xmmB); + cc.pmullw(xmmA, xmmB); + cc.pmuludq(xmmA, xmmB); + cc.por(xmmA, xmmB); + cc.pslld(xmmA, xmmB); + cc.pslld(xmmA, 0); + cc.psllq(xmmA, xmmB); + cc.psllq(xmmA, 0); + cc.psllw(xmmA, xmmB); + cc.psllw(xmmA, 0); + cc.pslldq(xmmA, 0); + cc.psrad(xmmA, xmmB); + cc.psrad(xmmA, 0); + cc.psraw(xmmA, xmmB); + cc.psraw(xmmA, 0); + cc.psubb(xmmA, xmmB); + cc.psubw(xmmA, xmmB); + cc.psubd(xmmA, xmmB); + cc.psubq(xmmA, xmmB); + cc.pmaddwd(xmmA, xmmB); + cc.pshufd(xmmA, xmmB, 0); + cc.pshufhw(xmmA, xmmB, 0); + cc.pshuflw(xmmA, xmmB, 0); + cc.psrld(xmmA, xmmB); + cc.psrld(xmmA, 0); + cc.psrlq(xmmA, xmmB); + cc.psrlq(xmmA, 0); + cc.psrldq(xmmA, 0); + cc.psrlw(xmmA, xmmB); + cc.psrlw(xmmA, 0); + cc.psubsb(xmmA, xmmB); + cc.psubsw(xmmA, xmmB); + cc.psubusb(xmmA, xmmB); + cc.psubusw(xmmA, xmmB); + cc.punpckhbw(xmmA, xmmB); + cc.punpckhwd(xmmA, xmmB); + cc.punpckhdq(xmmA, xmmB); + cc.punpckhqdq(xmmA, xmmB); + cc.punpcklbw(xmmA, xmmB); + cc.punpcklwd(xmmA, xmmB); + cc.punpckldq(xmmA, xmmB); + cc.punpcklqdq(xmmA, xmmB); + cc.pxor(xmmA, xmmB); + cc.sqrtpd(xmmA, xmmB); + cc.sqrtsd(xmmA, xmmB); + cc.subpd(xmmA, xmmB); + cc.subsd(xmmA, xmmB); + cc.ucomisd(xmmA, xmmB); + cc.unpckhpd(xmmA, xmmB); + cc.unpcklpd(xmmA, xmmB); + cc.xorpd(xmmA, xmmB); + + // SSE3. + cc.addsubpd(xmmA, xmmB); + cc.addsubps(xmmA, xmmB); + cc.haddpd(xmmA, xmmB); + cc.haddps(xmmA, xmmB); + cc.hsubpd(xmmA, xmmB); + cc.hsubps(xmmA, xmmB); + cc.movddup(xmmA, xmmB); + cc.movshdup(xmmA, xmmB); + cc.movsldup(xmmA, xmmB); + + // SSSE3. + cc.psignb(xmmA, xmmB); + cc.psignw(xmmA, xmmB); + cc.psignd(xmmA, xmmB); + cc.phaddw(xmmA, xmmB); + cc.phaddd(xmmA, xmmB); + cc.phaddsw(xmmA, xmmB); + cc.phsubw(xmmA, xmmB); + cc.phsubd(xmmA, xmmB); + cc.phsubsw(xmmA, xmmB); + cc.pmaddubsw(xmmA, xmmB); + cc.pabsb(xmmA, xmmB); + cc.pabsw(xmmA, xmmB); + cc.pabsd(xmmA, xmmB); + cc.pmulhrsw(xmmA, xmmB); + cc.pshufb(xmmA, xmmB); + cc.palignr(xmmA, xmmB, 0); + + // SSE4.1. + cc.blendpd(xmmA, xmmB, 0); + cc.blendps(xmmA, xmmB, 0); + cc.blendvpd(xmmA, xmmB, xmmA); + cc.blendvps(xmmA, xmmB, xmmA); + + cc.dppd(xmmA, xmmB, 0); + cc.dpps(xmmA, xmmB, 0); + cc.extractps(gpd, xmmB, 0); + cc.insertps(xmmA, xmmB, 0); + cc.mpsadbw(xmmA, xmmB, 0); + cc.packusdw(xmmA, xmmB); + cc.pblendvb(xmmA, xmmB, xmmA); + cc.pblendw(xmmA, xmmB, 0); + cc.pcmpeqq(xmmA, xmmB); + cc.pextrb(gpd, xmmB, 0); + cc.pextrd(gpd, xmmB, 0); + if (cc.is64Bit()) cc.pextrq(gpq, xmmB, 0); + cc.pextrw(gpd, xmmB, 0); + cc.phminposuw(xmmA, xmmB); + cc.pinsrb(xmmA, gpd, 0); + cc.pinsrd(xmmA, gpd, 0); + cc.pinsrw(xmmA, gpd, 0); + cc.pmaxuw(xmmA, xmmB); + cc.pmaxsb(xmmA, xmmB); + cc.pmaxsd(xmmA, xmmB); + cc.pmaxud(xmmA, xmmB); + cc.pminsb(xmmA, xmmB); + cc.pminuw(xmmA, xmmB); + cc.pminud(xmmA, xmmB); + cc.pminsd(xmmA, xmmB); + cc.pmovsxbw(xmmA, xmmB); + cc.pmovsxbd(xmmA, xmmB); + cc.pmovsxbq(xmmA, xmmB); + cc.pmovsxwd(xmmA, xmmB); + cc.pmovsxwq(xmmA, xmmB); + cc.pmovsxdq(xmmA, xmmB); + cc.pmovzxbw(xmmA, xmmB); + cc.pmovzxbd(xmmA, xmmB); + cc.pmovzxbq(xmmA, xmmB); + cc.pmovzxwd(xmmA, xmmB); + cc.pmovzxwq(xmmA, xmmB); + cc.pmovzxdq(xmmA, xmmB); + cc.pmuldq(xmmA, xmmB); + cc.pmulld(xmmA, xmmB); + cc.ptest(xmmA, xmmB); + cc.roundps(xmmA, xmmB, 0); + cc.roundss(xmmA, xmmB, 0); + cc.roundpd(xmmA, xmmB, 0); + cc.roundsd(xmmA, xmmB, 0); + } + else { + x86::Mem m = x86::ptr(gpz); + + cc.addps(xmmA, m); + cc.addss(xmmA, m); + cc.andnps(xmmA, m); + cc.andps(xmmA, m); + cc.cmpps(xmmA, m, 0); + cc.cmpss(xmmA, m, 0); + cc.comiss(xmmA, m); + cc.cvtpi2ps(xmmA, m); + cc.cvtsi2ss(xmmA, m); + cc.cvtss2si(gpd, m); + if (cc.is64Bit()) cc.cvtss2si(gpq, m); + cc.cvttss2si(gpd, m); + if (cc.is64Bit()) cc.cvttss2si(gpq, m); + cc.divps(xmmA, m); + cc.divss(xmmA, m); + cc.maxps(xmmA, m); + cc.maxss(xmmA, m); + cc.minps(xmmA, m); + cc.minss(xmmA, m); + cc.movaps(xmmA, m); + cc.movaps(m, xmmB); + cc.movd(m, xmmB); + cc.movd(xmmA, m); + cc.movq(m, xmmB); + cc.movq(xmmA, m); + cc.movhps(xmmA, m); + cc.movhps(m, xmmB); + cc.movlps(xmmA, m); + cc.movlps(m, xmmB); + cc.movntps(m, xmmB); + cc.movss(xmmA, m); + cc.movss(m, xmmB); + cc.movups(xmmA, m); + cc.movups(m, xmmB); + cc.mulps(xmmA, m); + cc.mulss(xmmA, m); + cc.orps(xmmA, m); + cc.rcpps(xmmA, m); + cc.rcpss(xmmA, m); + cc.psadbw(xmmA, m); + cc.rsqrtps(xmmA, m); + cc.rsqrtss(xmmA, m); + cc.shufps(xmmA, m, 0); + cc.sqrtps(xmmA, m); + cc.sqrtss(xmmA, m); + cc.stmxcsr(m); + cc.subps(xmmA, m); + cc.subss(xmmA, m); + cc.ucomiss(xmmA, m); + cc.unpckhps(xmmA, m); + cc.unpcklps(xmmA, m); + cc.xorps(xmmA, m); + + // SSE2. + cc.addpd(xmmA, m); + cc.addsd(xmmA, m); + cc.andnpd(xmmA, m); + cc.andpd(xmmA, m); + cc.cmppd(xmmA, m, 0); + cc.cmpsd(xmmA, m, 0); + cc.comisd(xmmA, m); + cc.cvtdq2pd(xmmA, m); + cc.cvtdq2ps(xmmA, m); + cc.cvtpd2dq(xmmA, m); + cc.cvtpd2ps(xmmA, m); + cc.cvtpi2pd(xmmA, m); + cc.cvtps2dq(xmmA, m); + cc.cvtps2pd(xmmA, m); + cc.cvtsd2si(gpd, m); + if (cc.is64Bit()) cc.cvtsd2si(gpq, m); + cc.cvtsd2ss(xmmA, m); + cc.cvtsi2sd(xmmA, m); + cc.cvtss2sd(xmmA, m); + cc.cvtss2si(gpd, m); + if (cc.is64Bit()) cc.cvtss2si(gpq, m); + cc.cvttpd2dq(xmmA, m); + cc.cvttps2dq(xmmA, m); + cc.cvttsd2si(gpd, m); + if (cc.is64Bit()) cc.cvttsd2si(gpq, m); + cc.divpd(xmmA, m); + cc.divsd(xmmA, m); + cc.maxpd(xmmA, m); + cc.maxsd(xmmA, m); + cc.minpd(xmmA, m); + cc.minsd(xmmA, m); + cc.movdqa(xmmA, m); + cc.movdqa(m, xmmB); + cc.movdqu(xmmA, m); + cc.movdqu(m, xmmB); + cc.movsd(xmmA, m); + cc.movsd(m, xmmB); + cc.movapd(xmmA, m); + cc.movapd(m, xmmB); + cc.movhpd(xmmA, m); + cc.movhpd(m, xmmB); + cc.movlpd(xmmA, m); + cc.movlpd(m, xmmB); + cc.movntdq(m, xmmB); + cc.movntpd(m, xmmB); + cc.movupd(xmmA, m); + cc.movupd(m, xmmB); + cc.mulpd(xmmA, m); + cc.mulsd(xmmA, m); + cc.orpd(xmmA, m); + cc.packsswb(xmmA, m); + cc.packssdw(xmmA, m); + cc.packuswb(xmmA, m); + cc.paddb(xmmA, m); + cc.paddw(xmmA, m); + cc.paddd(xmmA, m); + cc.paddq(xmmA, m); + cc.paddsb(xmmA, m); + cc.paddsw(xmmA, m); + cc.paddusb(xmmA, m); + cc.paddusw(xmmA, m); + cc.pand(xmmA, m); + cc.pandn(xmmA, m); + cc.pavgb(xmmA, m); + cc.pavgw(xmmA, m); + cc.pcmpeqb(xmmA, m); + cc.pcmpeqw(xmmA, m); + cc.pcmpeqd(xmmA, m); + cc.pcmpgtb(xmmA, m); + cc.pcmpgtw(xmmA, m); + cc.pcmpgtd(xmmA, m); + cc.pmaxsw(xmmA, m); + cc.pmaxub(xmmA, m); + cc.pminsw(xmmA, m); + cc.pminub(xmmA, m); + cc.pmulhw(xmmA, m); + cc.pmulhuw(xmmA, m); + cc.pmullw(xmmA, m); + cc.pmuludq(xmmA, m); + cc.por(xmmA, m); + cc.pslld(xmmA, m); + cc.psllq(xmmA, m); + cc.psllw(xmmA, m); + cc.psrad(xmmA, m); + cc.psraw(xmmA, m); + cc.psubb(xmmA, m); + cc.psubw(xmmA, m); + cc.psubd(xmmA, m); + cc.psubq(xmmA, m); + cc.pmaddwd(xmmA, m); + cc.pshufd(xmmA, m, 0); + cc.pshufhw(xmmA, m, 0); + cc.pshuflw(xmmA, m, 0); + cc.psrld(xmmA, m); + cc.psrlq(xmmA, m); + cc.psrlw(xmmA, m); + cc.psubsb(xmmA, m); + cc.psubsw(xmmA, m); + cc.psubusb(xmmA, m); + cc.psubusw(xmmA, m); + cc.punpckhbw(xmmA, m); + cc.punpckhwd(xmmA, m); + cc.punpckhdq(xmmA, m); + cc.punpckhqdq(xmmA, m); + cc.punpcklbw(xmmA, m); + cc.punpcklwd(xmmA, m); + cc.punpckldq(xmmA, m); + cc.punpcklqdq(xmmA, m); + cc.pxor(xmmA, m); + cc.sqrtpd(xmmA, m); + cc.sqrtsd(xmmA, m); + cc.subpd(xmmA, m); + cc.subsd(xmmA, m); + cc.ucomisd(xmmA, m); + cc.unpckhpd(xmmA, m); + cc.unpcklpd(xmmA, m); + cc.xorpd(xmmA, m); + + // SSE3. + cc.addsubpd(xmmA, m); + cc.addsubps(xmmA, m); + cc.haddpd(xmmA, m); + cc.haddps(xmmA, m); + cc.hsubpd(xmmA, m); + cc.hsubps(xmmA, m); + cc.lddqu(xmmA, m); + cc.movddup(xmmA, m); + cc.movshdup(xmmA, m); + cc.movsldup(xmmA, m); + + // SSSE3. + cc.psignb(xmmA, m); + cc.psignw(xmmA, m); + cc.psignd(xmmA, m); + cc.phaddw(xmmA, m); + cc.phaddd(xmmA, m); + cc.phaddsw(xmmA, m); + cc.phsubw(xmmA, m); + cc.phsubd(xmmA, m); + cc.phsubsw(xmmA, m); + cc.pmaddubsw(xmmA, m); + cc.pabsb(xmmA, m); + cc.pabsw(xmmA, m); + cc.pabsd(xmmA, m); + cc.pmulhrsw(xmmA, m); + cc.pshufb(xmmA, m); + cc.palignr(xmmA, m, 0); + + // SSE4.1. + cc.blendpd(xmmA, m, 0); + cc.blendps(xmmA, m, 0); + cc.blendvpd(xmmA, m, xmmA); + cc.blendvps(xmmA, m, xmmA); + + cc.dppd(xmmA, m, 0); + cc.dpps(xmmA, m, 0); + cc.extractps(m, xmmB, 0); + cc.insertps(xmmA, m, 0); + cc.movntdqa(xmmA, m); + cc.mpsadbw(xmmA, m, 0); + cc.packusdw(xmmA, m); + cc.pblendvb(xmmA, m, xmmA); + cc.pblendw(xmmA, m, 0); + cc.pcmpeqq(xmmA, m); + cc.pextrb(m, xmmB, 0); + cc.pextrd(m, xmmB, 0); + if (cc.is64Bit()) cc.pextrq(m, xmmB, 0); + cc.pextrw(m, xmmB, 0); + cc.phminposuw(xmmA, m); + cc.pinsrb(xmmA, m, 0); + cc.pinsrd(xmmA, m, 0); + cc.pinsrw(xmmA, m, 0); + cc.pmaxuw(xmmA, m); + cc.pmaxsb(xmmA, m); + cc.pmaxsd(xmmA, m); + cc.pmaxud(xmmA, m); + cc.pminsb(xmmA, m); + cc.pminuw(xmmA, m); + cc.pminud(xmmA, m); + cc.pminsd(xmmA, m); + cc.pmovsxbw(xmmA, m); + cc.pmovsxbd(xmmA, m); + cc.pmovsxbq(xmmA, m); + cc.pmovsxwd(xmmA, m); + cc.pmovsxwq(xmmA, m); + cc.pmovsxdq(xmmA, m); + cc.pmovzxbw(xmmA, m); + cc.pmovzxbd(xmmA, m); + cc.pmovzxbq(xmmA, m); + cc.pmovzxwd(xmmA, m); + cc.pmovzxwq(xmmA, m); + cc.pmovzxdq(xmmA, m); + cc.pmuldq(xmmA, m); + cc.pmulld(xmmA, m); + cc.ptest(xmmA, m); + cc.roundps(xmmA, m, 0); + cc.roundss(xmmA, m, 0); + cc.roundpd(xmmA, m, 0); + cc.roundsd(xmmA, m, 0); + + // SSE4.2. + cc.pcmpgtq(xmmA, m); + } +} + +static void generateSseSequence(BaseEmitter& emitter, InstForm form, bool emitPrologEpilog) { + using namespace asmjit::x86; + + if (emitter.isAssembler()) { + Assembler& cc = *emitter.as<Assembler>(); + + if (emitPrologEpilog) { + FuncDetail func; + func.init(FuncSignatureT<void, void*, const void*, size_t>(CallConv::kIdHost), cc.environment()); + + FuncFrame frame; + frame.init(func); + frame.addDirtyRegs(eax, xmm0, xmm1, xmm2, xmm3); + frame.finalize(); + + cc.emitProlog(frame); + generateSseSequenceInternal(cc, form, eax, xmm0, xmm1, xmm2, xmm3); + cc.emitEpilog(frame); + } + else { + generateSseSequenceInternal(cc, form, eax, xmm0, xmm1, xmm2, xmm3); + } + } +#ifndef ASMJIT_NO_BUILDER + else if (emitter.isBuilder()) { + Builder& cc = *emitter.as<Builder>(); + + if (emitPrologEpilog) { + FuncDetail func; + func.init(FuncSignatureT<void, void*, const void*, size_t>(CallConv::kIdHost), cc.environment()); + + FuncFrame frame; + frame.init(func); + frame.addDirtyRegs(eax, xmm0, xmm1, xmm2, xmm3); + frame.finalize(); + + cc.emitProlog(frame); + generateSseSequenceInternal(cc, form, eax, xmm0, xmm1, xmm2, xmm3); + cc.emitEpilog(frame); + } + else { + generateSseSequenceInternal(cc, form, eax, xmm0, xmm1, xmm2, xmm3); + } + } +#endif +#ifndef ASMJIT_NO_COMPILER + else if (emitter.isCompiler()) { + Compiler& cc = *emitter.as<Compiler>(); + + Gp gp = cc.newGpz("gp"); + Xmm a = cc.newXmm("a"); + Xmm b = cc.newXmm("b"); + Xmm c = cc.newXmm("c"); + Xmm d = cc.newXmm("d"); + + cc.addFunc(FuncSignatureT<void>(CallConv::kIdHost)); + generateSseSequenceInternal(cc, form, gp, a, b, c, d); + cc.endFunc(); + } +#endif +} + +// Generates a long sequence of AVX instructions. +template<typename Emitter> +static void generateAvxSequenceInternal( + Emitter& cc, + InstForm form, + const x86::Gp& gp, + const x86::Vec& vecA, const x86::Vec& vecB, const x86::Vec& vecC, const x86::Vec& vecD) { + + x86::Gp gpd = gp.r32(); + x86::Gp gpq = gp.r64(); + x86::Gp gpz = cc.is32Bit() ? gpd : gpq; + + x86::Xmm xmmA = vecA.xmm(); + x86::Xmm xmmB = vecB.xmm(); + x86::Xmm xmmC = vecC.xmm(); + x86::Xmm xmmD = vecD.xmm(); + + x86::Ymm ymmA = vecA.ymm(); + x86::Ymm ymmB = vecB.ymm(); + x86::Ymm ymmC = vecC.ymm(); + x86::Ymm ymmD = vecD.ymm(); + + cc.xor_(gpd, gpd); + cc.vxorps(xmmA, xmmA, xmmA); + cc.vxorps(xmmB, xmmB, xmmB); + cc.vxorps(xmmC, xmmC, xmmC); + cc.vxorps(xmmD, xmmD, xmmD); + + if (form == InstForm::kReg) { + cc.vaddpd(xmmA, xmmB, xmmC); + cc.vaddpd(ymmA, ymmB, ymmC); + cc.vaddps(xmmA, xmmB, xmmC); + cc.vaddps(ymmA, ymmB, ymmC); + cc.vaddsd(xmmA, xmmB, xmmC); + cc.vaddss(xmmA, xmmB, xmmC); + cc.vaddsubpd(xmmA, xmmB, xmmC); + cc.vaddsubpd(ymmA, ymmB, ymmC); + cc.vaddsubps(xmmA, xmmB, xmmC); + cc.vaddsubps(ymmA, ymmB, ymmC); + cc.vandpd(xmmA, xmmB, xmmC); + cc.vandpd(ymmA, ymmB, ymmC); + cc.vandps(xmmA, xmmB, xmmC); + cc.vandps(ymmA, ymmB, ymmC); + cc.vandnpd(xmmA, xmmB, xmmC); + cc.vandnpd(ymmA, ymmB, ymmC); + cc.vandnps(xmmA, xmmB, xmmC); + cc.vandnps(ymmA, ymmB, ymmC); + cc.vblendpd(xmmA, xmmB, xmmC, 0); + cc.vblendpd(ymmA, ymmB, ymmC, 0); + cc.vblendps(xmmA, xmmB, xmmC, 0); + cc.vblendps(ymmA, ymmB, ymmC, 0); + cc.vblendvpd(xmmA, xmmB, xmmC, xmmA); + cc.vblendvpd(ymmA, ymmB, ymmC, ymmA); + cc.vcmppd(xmmA, xmmB, xmmC, 0); + cc.vcmppd(ymmA, ymmB, ymmC, 0); + cc.vcmpps(xmmA, xmmB, xmmC, 0); + cc.vcmpps(ymmA, ymmB, ymmC, 0); + cc.vcmpsd(xmmA, xmmB, xmmC, 0); + cc.vcmpss(xmmA, xmmB, xmmC, 0); + cc.vcomisd(xmmA, xmmB); + cc.vcomiss(xmmA, xmmB); + cc.vcvtdq2pd(xmmA, xmmB); + cc.vcvtdq2pd(ymmA, xmmB); + cc.vcvtdq2ps(xmmA, xmmB); + cc.vcvtdq2ps(ymmA, ymmB); + cc.vcvtpd2dq(xmmA, xmmB); + cc.vcvtpd2dq(xmmA, ymmB); + cc.vcvtpd2ps(xmmA, xmmB); + cc.vcvtpd2ps(xmmA, ymmB); + cc.vcvtps2dq(xmmA, xmmB); + cc.vcvtps2dq(ymmA, ymmB); + cc.vcvtps2pd(xmmA, xmmB); + cc.vcvtps2pd(ymmA, xmmB); + cc.vcvtsd2si(gpd, xmmB); + cc.vcvtsd2si(gpz, xmmB); + cc.vcvtsd2ss(xmmA, xmmB, xmmC); + cc.vcvtsi2sd(xmmA, xmmB, gpd); + cc.vcvtsi2sd(xmmA, xmmB, gpz); + cc.vcvtsi2ss(xmmA, xmmB, gpd); + cc.vcvtsi2ss(xmmA, xmmB, gpz); + cc.vcvtss2sd(xmmA, xmmB, xmmC); + cc.vcvtss2si(gpd, xmmB); + cc.vcvttpd2dq(xmmA, xmmB); + cc.vcvttpd2dq(xmmA, ymmB); + cc.vcvttps2dq(xmmA, xmmB); + cc.vcvttps2dq(ymmA, ymmB); + cc.vcvttsd2si(gpd, xmmB); + cc.vcvttss2si(gpz, xmmB); + cc.vdivpd(xmmA, xmmB, xmmC); + cc.vdivpd(ymmA, ymmB, ymmC); + cc.vdivps(xmmA, xmmB, xmmC); + cc.vdivps(ymmA, ymmB, ymmC); + cc.vdivsd(xmmA, xmmB, xmmC); + cc.vdivss(xmmA, xmmB, xmmC); + cc.vdppd(xmmA, xmmB, xmmC, 0); + cc.vdpps(xmmA, xmmB, xmmC, 0); + cc.vdpps(ymmA, ymmB, ymmC, 0); + cc.vextractf128(xmmA, ymmB, 0); + cc.vextractps(gpd, xmmB, 0); + cc.vhaddpd(xmmA, xmmB, xmmC); + cc.vhaddpd(ymmA, ymmB, ymmC); + cc.vhaddps(xmmA, xmmB, xmmC); + cc.vhaddps(ymmA, ymmB, ymmC); + cc.vhsubpd(xmmA, xmmB, xmmC); + cc.vhsubpd(ymmA, ymmB, ymmC); + cc.vhsubps(xmmA, xmmB, xmmC); + cc.vhsubps(ymmA, ymmB, ymmC); + cc.vinsertf128(ymmA, ymmB, xmmC, 0); + cc.vinsertps(xmmA, xmmB, xmmC, 0); + cc.vmaxpd(xmmA, xmmB, xmmC); + cc.vmaxpd(ymmA, ymmB, ymmC); + cc.vmaxps(xmmA, xmmB, xmmC); + cc.vmaxps(ymmA, ymmB, ymmC); + cc.vmaxsd(xmmA, xmmB, xmmC); + cc.vmaxss(xmmA, xmmB, xmmC); + cc.vminpd(xmmA, xmmB, xmmC); + cc.vminpd(ymmA, ymmB, ymmC); + cc.vminps(xmmA, xmmB, xmmC); + cc.vminps(ymmA, ymmB, ymmC); + cc.vminsd(xmmA, xmmB, xmmC); + cc.vminss(xmmA, xmmB, xmmC); + cc.vmovapd(xmmA, xmmB); + cc.vmovapd(ymmA, ymmB); + cc.vmovaps(xmmA, xmmB); + cc.vmovaps(ymmA, ymmB); + cc.vmovd(xmmA, gpd); + cc.vmovd(gpd, xmmB); + cc.vmovddup(xmmA, xmmB); + cc.vmovddup(ymmA, ymmB); + cc.vmovdqa(xmmA, xmmB); + cc.vmovdqa(ymmA, ymmB); + cc.vmovdqu(xmmA, xmmB); + cc.vmovdqu(ymmA, ymmB); + cc.vmovhlps(xmmA, xmmB, xmmC); + cc.vmovlhps(xmmA, xmmB, xmmC); + cc.vmovmskpd(gpd, xmmB); + cc.vmovmskpd(gpd, ymmB); + cc.vmovmskps(gpd, xmmB); + cc.vmovmskps(gpd, ymmB); + cc.vmovsd(xmmA, xmmB, xmmC); + cc.vmovshdup(xmmA, xmmB); + cc.vmovshdup(ymmA, ymmB); + cc.vmovsldup(xmmA, xmmB); + cc.vmovsldup(ymmA, ymmB); + cc.vmovss(xmmA, xmmB, xmmC); + cc.vmovupd(xmmA, xmmB); + cc.vmovupd(ymmA, ymmB); + cc.vmovups(xmmA, xmmB); + cc.vmovups(ymmA, ymmB); + cc.vmpsadbw(xmmA, xmmB, xmmC, 0); + cc.vmulpd(xmmA, xmmB, xmmC); + cc.vmulpd(ymmA, ymmB, ymmC); + cc.vmulps(xmmA, xmmB, xmmC); + cc.vmulps(ymmA, ymmB, ymmC); + cc.vmulsd(xmmA, xmmB, xmmC); + cc.vmulss(xmmA, xmmB, xmmC); + cc.vorpd(xmmA, xmmB, xmmC); + cc.vorpd(ymmA, ymmB, ymmC); + cc.vorps(xmmA, xmmB, xmmC); + cc.vorps(ymmA, ymmB, ymmC); + cc.vpabsb(xmmA, xmmB); + cc.vpabsd(xmmA, xmmB); + cc.vpabsw(xmmA, xmmB); + cc.vpackssdw(xmmA, xmmB, xmmC); + cc.vpacksswb(xmmA, xmmB, xmmC); + cc.vpackusdw(xmmA, xmmB, xmmC); + cc.vpackuswb(xmmA, xmmB, xmmC); + cc.vpaddb(xmmA, xmmB, xmmC); + cc.vpaddd(xmmA, xmmB, xmmC); + cc.vpaddq(xmmA, xmmB, xmmC); + cc.vpaddw(xmmA, xmmB, xmmC); + cc.vpaddsb(xmmA, xmmB, xmmC); + cc.vpaddsw(xmmA, xmmB, xmmC); + cc.vpaddusb(xmmA, xmmB, xmmC); + cc.vpaddusw(xmmA, xmmB, xmmC); + cc.vpalignr(xmmA, xmmB, xmmC, 0); + cc.vpand(xmmA, xmmB, xmmC); + cc.vpandn(xmmA, xmmB, xmmC); + cc.vpavgb(xmmA, xmmB, xmmC); + cc.vpavgw(xmmA, xmmB, xmmC); + cc.vpblendvb(xmmA, xmmB, xmmC, xmmA); + cc.vpblendw(xmmA, xmmB, xmmC, 0); + cc.vpcmpeqb(xmmA, xmmB, xmmC); + cc.vpcmpeqd(xmmA, xmmB, xmmC); + cc.vpcmpeqq(xmmA, xmmB, xmmC); + cc.vpcmpeqw(xmmA, xmmB, xmmC); + cc.vpcmpgtb(xmmA, xmmB, xmmC); + cc.vpcmpgtd(xmmA, xmmB, xmmC); + cc.vpcmpgtq(xmmA, xmmB, xmmC); + cc.vpcmpgtw(xmmA, xmmB, xmmC); + cc.vpermilpd(xmmA, xmmB, xmmC); + cc.vpermilpd(ymmA, ymmB, ymmC); + cc.vpermilpd(xmmA, xmmB, 0); + cc.vpermilpd(ymmA, ymmB, 0); + cc.vpermilps(xmmA, xmmB, xmmC); + cc.vpermilps(ymmA, ymmB, ymmC); + cc.vpermilps(xmmA, xmmB, 0); + cc.vpermilps(ymmA, ymmB, 0); + cc.vperm2f128(ymmA, ymmB, ymmC, 0); + cc.vpextrb(gpd, xmmB, 0); + cc.vpextrd(gpd, xmmB, 0); + if (cc.is64Bit()) cc.vpextrq(gpq, xmmB, 0); + cc.vpextrw(gpd, xmmB, 0); + cc.vphaddd(xmmA, xmmB, xmmC); + cc.vphaddsw(xmmA, xmmB, xmmC); + cc.vphaddw(xmmA, xmmB, xmmC); + cc.vphminposuw(xmmA, xmmB); + cc.vphsubd(xmmA, xmmB, xmmC); + cc.vphsubsw(xmmA, xmmB, xmmC); + cc.vphsubw(xmmA, xmmB, xmmC); + cc.vpinsrb(xmmA, xmmB, gpd, 0); + cc.vpinsrd(xmmA, xmmB, gpd, 0); + cc.vpinsrw(xmmA, xmmB, gpd, 0); + cc.vpmaddubsw(xmmA, xmmB, xmmC); + cc.vpmaddwd(xmmA, xmmB, xmmC); + cc.vpmaxsb(xmmA, xmmB, xmmC); + cc.vpmaxsd(xmmA, xmmB, xmmC); + cc.vpmaxsw(xmmA, xmmB, xmmC); + cc.vpmaxub(xmmA, xmmB, xmmC); + cc.vpmaxud(xmmA, xmmB, xmmC); + cc.vpmaxuw(xmmA, xmmB, xmmC); + cc.vpminsb(xmmA, xmmB, xmmC); + cc.vpminsd(xmmA, xmmB, xmmC); + cc.vpminsw(xmmA, xmmB, xmmC); + cc.vpminub(xmmA, xmmB, xmmC); + cc.vpminud(xmmA, xmmB, xmmC); + cc.vpminuw(xmmA, xmmB, xmmC); + cc.vpmovmskb(gpd, xmmB); + cc.vpmovsxbd(xmmA, xmmB); + cc.vpmovsxbq(xmmA, xmmB); + cc.vpmovsxbw(xmmA, xmmB); + cc.vpmovsxdq(xmmA, xmmB); + cc.vpmovsxwd(xmmA, xmmB); + cc.vpmovsxwq(xmmA, xmmB); + cc.vpmovzxbd(xmmA, xmmB); + cc.vpmovzxbq(xmmA, xmmB); + cc.vpmovzxbw(xmmA, xmmB); + cc.vpmovzxdq(xmmA, xmmB); + cc.vpmovzxwd(xmmA, xmmB); + cc.vpmovzxwq(xmmA, xmmB); + cc.vpmuldq(xmmA, xmmB, xmmC); + cc.vpmulhrsw(xmmA, xmmB, xmmC); + cc.vpmulhuw(xmmA, xmmB, xmmC); + cc.vpmulhw(xmmA, xmmB, xmmC); + cc.vpmulld(xmmA, xmmB, xmmC); + cc.vpmullw(xmmA, xmmB, xmmC); + cc.vpmuludq(xmmA, xmmB, xmmC); + cc.vpor(xmmA, xmmB, xmmC); + cc.vpsadbw(xmmA, xmmB, xmmC); + cc.vpshufb(xmmA, xmmB, xmmC); + cc.vpshufd(xmmA, xmmB, 0); + cc.vpshufhw(xmmA, xmmB, 0); + cc.vpshuflw(xmmA, xmmB, 0); + cc.vpsignb(xmmA, xmmB, xmmC); + cc.vpsignd(xmmA, xmmB, xmmC); + cc.vpsignw(xmmA, xmmB, xmmC); + cc.vpslld(xmmA, xmmB, xmmC); + cc.vpslld(xmmA, xmmB, 0); + cc.vpslldq(xmmA, xmmB, 0); + cc.vpsllq(xmmA, xmmB, xmmC); + cc.vpsllq(xmmA, xmmB, 0); + cc.vpsllw(xmmA, xmmB, xmmC); + cc.vpsllw(xmmA, xmmB, 0); + cc.vpsrad(xmmA, xmmB, xmmC); + cc.vpsrad(xmmA, xmmB, 0); + cc.vpsraw(xmmA, xmmB, xmmC); + cc.vpsraw(xmmA, xmmB, 0); + cc.vpsrld(xmmA, xmmB, xmmC); + cc.vpsrld(xmmA, xmmB, 0); + cc.vpsrldq(xmmA, xmmB, 0); + cc.vpsrlq(xmmA, xmmB, xmmC); + cc.vpsrlq(xmmA, xmmB, 0); + cc.vpsrlw(xmmA, xmmB, xmmC); + cc.vpsrlw(xmmA, xmmB, 0); + cc.vpsubb(xmmA, xmmB, xmmC); + cc.vpsubd(xmmA, xmmB, xmmC); + cc.vpsubq(xmmA, xmmB, xmmC); + cc.vpsubw(xmmA, xmmB, xmmC); + cc.vpsubsb(xmmA, xmmB, xmmC); + cc.vpsubsw(xmmA, xmmB, xmmC); + cc.vpsubusb(xmmA, xmmB, xmmC); + cc.vpsubusw(xmmA, xmmB, xmmC); + cc.vptest(xmmA, xmmB); + cc.vptest(ymmA, ymmB); + cc.vpunpckhbw(xmmA, xmmB, xmmC); + cc.vpunpckhdq(xmmA, xmmB, xmmC); + cc.vpunpckhqdq(xmmA, xmmB, xmmC); + cc.vpunpckhwd(xmmA, xmmB, xmmC); + cc.vpunpcklbw(xmmA, xmmB, xmmC); + cc.vpunpckldq(xmmA, xmmB, xmmC); + cc.vpunpcklqdq(xmmA, xmmB, xmmC); + cc.vpunpcklwd(xmmA, xmmB, xmmC); + cc.vpxor(xmmA, xmmB, xmmC); + cc.vrcpps(xmmA, xmmB); + cc.vrcpps(ymmA, ymmB); + cc.vrcpss(xmmA, xmmB, xmmC); + cc.vrsqrtps(xmmA, xmmB); + cc.vrsqrtps(ymmA, ymmB); + cc.vrsqrtss(xmmA, xmmB, xmmC); + cc.vroundpd(xmmA, xmmB, 0); + cc.vroundpd(ymmA, ymmB, 0); + cc.vroundps(xmmA, xmmB, 0); + cc.vroundps(ymmA, ymmB, 0); + cc.vroundsd(xmmA, xmmB, xmmC, 0); + cc.vroundss(xmmA, xmmB, xmmC, 0); + cc.vshufpd(xmmA, xmmB, xmmC, 0); + cc.vshufpd(ymmA, ymmB, ymmC, 0); + cc.vshufps(xmmA, xmmB, xmmC, 0); + cc.vshufps(ymmA, ymmB, ymmC, 0); + cc.vsqrtpd(xmmA, xmmB); + cc.vsqrtpd(ymmA, ymmB); + cc.vsqrtps(xmmA, xmmB); + cc.vsqrtps(ymmA, ymmB); + cc.vsqrtsd(xmmA, xmmB, xmmC); + cc.vsqrtss(xmmA, xmmB, xmmC); + cc.vsubpd(xmmA, xmmB, xmmC); + cc.vsubpd(ymmA, ymmB, ymmC); + cc.vsubps(xmmA, xmmB, xmmC); + cc.vsubps(ymmA, ymmB, ymmC); + cc.vsubsd(xmmA, xmmB, xmmC); + cc.vsubss(xmmA, xmmB, xmmC); + cc.vtestps(xmmA, xmmB); + cc.vtestps(ymmA, ymmB); + cc.vtestpd(xmmA, xmmB); + cc.vtestpd(ymmA, ymmB); + cc.vucomisd(xmmA, xmmB); + cc.vucomiss(xmmA, xmmB); + cc.vunpckhpd(xmmA, xmmB, xmmC); + cc.vunpckhpd(ymmA, ymmB, ymmC); + cc.vunpckhps(xmmA, xmmB, xmmC); + cc.vunpckhps(ymmA, ymmB, ymmC); + cc.vunpcklpd(xmmA, xmmB, xmmC); + cc.vunpcklpd(ymmA, ymmB, ymmC); + cc.vunpcklps(xmmA, xmmB, xmmC); + cc.vunpcklps(ymmA, ymmB, ymmC); + cc.vxorpd(xmmA, xmmB, xmmC); + cc.vxorpd(ymmA, ymmB, ymmC); + cc.vxorps(xmmA, xmmB, xmmC); + cc.vxorps(ymmA, ymmB, ymmC); + + // AVX+AESNI. + cc.vaesdec(xmmA, xmmB, xmmC); + cc.vaesdeclast(xmmA, xmmB, xmmC); + cc.vaesenc(xmmA, xmmB, xmmC); + cc.vaesenclast(xmmA, xmmB, xmmC); + cc.vaesimc(xmmA, xmmB); + cc.vaeskeygenassist(xmmA, xmmB, 0); + + // AVX+PCLMULQDQ. + cc.vpclmulqdq(xmmA, xmmB, xmmC, 0); + + // AVX2. + cc.vbroadcastsd(ymmA, xmmB); + cc.vbroadcastss(xmmA, xmmB); + cc.vbroadcastss(ymmA, xmmB); + cc.vextracti128(xmmA, ymmB, 0); + cc.vinserti128(ymmA, ymmB, xmmC, 0); + cc.vmpsadbw(ymmA, ymmB, ymmC, 0); + cc.vpabsb(ymmA, ymmB); + cc.vpabsd(ymmA, ymmB); + cc.vpabsw(ymmA, ymmB); + cc.vpackssdw(ymmA, ymmB, ymmC); + cc.vpacksswb(ymmA, ymmB, ymmC); + cc.vpackusdw(ymmA, ymmB, ymmC); + cc.vpackuswb(ymmA, ymmB, ymmC); + cc.vpaddb(ymmA, ymmB, ymmC); + cc.vpaddd(ymmA, ymmB, ymmC); + cc.vpaddq(ymmA, ymmB, ymmC); + cc.vpaddw(ymmA, ymmB, ymmC); + cc.vpaddsb(ymmA, ymmB, ymmC); + cc.vpaddsw(ymmA, ymmB, ymmC); + cc.vpaddusb(ymmA, ymmB, ymmC); + cc.vpaddusw(ymmA, ymmB, ymmC); + cc.vpalignr(ymmA, ymmB, ymmC, 0); + cc.vpand(ymmA, ymmB, ymmC); + cc.vpandn(ymmA, ymmB, ymmC); + cc.vpavgb(ymmA, ymmB, ymmC); + cc.vpavgw(ymmA, ymmB, ymmC); + cc.vpblendd(xmmA, xmmB, xmmC, 0); + cc.vpblendd(ymmA, ymmB, ymmC, 0); + cc.vpblendvb(ymmA, ymmB, ymmC, ymmA); + cc.vpblendw(ymmA, ymmB, ymmC, 0); + cc.vpbroadcastb(xmmA, xmmB); + cc.vpbroadcastb(ymmA, xmmB); + cc.vpbroadcastd(xmmA, xmmB); + cc.vpbroadcastd(ymmA, xmmB); + cc.vpbroadcastq(xmmA, xmmB); + cc.vpbroadcastq(ymmA, xmmB); + cc.vpbroadcastw(xmmA, xmmB); + cc.vpbroadcastw(ymmA, xmmB); + cc.vpcmpeqb(ymmA, ymmB, ymmC); + cc.vpcmpeqd(ymmA, ymmB, ymmC); + cc.vpcmpeqq(ymmA, ymmB, ymmC); + cc.vpcmpeqw(ymmA, ymmB, ymmC); + cc.vpcmpgtb(ymmA, ymmB, ymmC); + cc.vpcmpgtd(ymmA, ymmB, ymmC); + cc.vpcmpgtq(ymmA, ymmB, ymmC); + cc.vpcmpgtw(ymmA, ymmB, ymmC); + cc.vperm2i128(ymmA, ymmB, ymmC, 0); + cc.vpermd(ymmA, ymmB, ymmC); + cc.vpermps(ymmA, ymmB, ymmC); + cc.vpermpd(ymmA, ymmB, 0); + cc.vpermq(ymmA, ymmB, 0); + cc.vpmovmskb(gpd, ymmB); + cc.vpmovsxbd(ymmA, xmmB); + cc.vpmovsxbq(ymmA, xmmB); + cc.vpmovsxbw(ymmA, xmmB); + cc.vpmovsxdq(ymmA, xmmB); + cc.vpmovsxwd(ymmA, xmmB); + cc.vpmovsxwq(ymmA, xmmB); + cc.vpmovzxbd(ymmA, xmmB); + cc.vpmovzxbq(ymmA, xmmB); + cc.vpmovzxbw(ymmA, xmmB); + cc.vpmovzxdq(ymmA, xmmB); + cc.vpmovzxwd(ymmA, xmmB); + cc.vpmovzxwq(ymmA, xmmB); + cc.vpshufd(ymmA, ymmB, 0); + cc.vpshufhw(ymmA, ymmB, 0); + cc.vpshuflw(ymmA, ymmB, 0); + cc.vpslld(ymmA, ymmB, 0); + cc.vpslldq(ymmA, ymmB, 0); + cc.vpsllq(ymmA, ymmB, 0); + cc.vpsllw(ymmA, ymmB, 0); + cc.vpsrad(ymmA, ymmB, 0); + cc.vpsraw(ymmA, ymmB, 0); + cc.vpsrld(ymmA, ymmB, 0); + cc.vpsrldq(ymmA, ymmB, 0); + cc.vpsrlq(ymmA, ymmB, 0); + cc.vpsrlw(ymmA, ymmB, 0); + cc.vphaddd(ymmA, ymmB, ymmC); + cc.vphaddsw(ymmA, ymmB, ymmC); + cc.vphaddw(ymmA, ymmB, ymmC); + cc.vphsubd(ymmA, ymmB, ymmC); + cc.vphsubsw(ymmA, ymmB, ymmC); + cc.vphsubw(ymmA, ymmB, ymmC); + cc.vpmaddubsw(ymmA, ymmB, ymmC); + cc.vpmaddwd(ymmA, ymmB, ymmC); + cc.vpmaxsb(ymmA, ymmB, ymmC); + cc.vpmaxsd(ymmA, ymmB, ymmC); + cc.vpmaxsw(ymmA, ymmB, ymmC); + cc.vpmaxub(ymmA, ymmB, ymmC); + cc.vpmaxud(ymmA, ymmB, ymmC); + cc.vpmaxuw(ymmA, ymmB, ymmC); + cc.vpminsb(ymmA, ymmB, ymmC); + cc.vpminsd(ymmA, ymmB, ymmC); + cc.vpminsw(ymmA, ymmB, ymmC); + cc.vpminub(ymmA, ymmB, ymmC); + cc.vpminud(ymmA, ymmB, ymmC); + cc.vpminuw(ymmA, ymmB, ymmC); + cc.vpmuldq(ymmA, ymmB, ymmC); + cc.vpmulhrsw(ymmA, ymmB, ymmC); + cc.vpmulhuw(ymmA, ymmB, ymmC); + cc.vpmulhw(ymmA, ymmB, ymmC); + cc.vpmulld(ymmA, ymmB, ymmC); + cc.vpmullw(ymmA, ymmB, ymmC); + cc.vpmuludq(ymmA, ymmB, ymmC); + cc.vpor(ymmA, ymmB, ymmC); + cc.vpsadbw(ymmA, ymmB, ymmC); + cc.vpshufb(ymmA, ymmB, ymmC); + cc.vpsignb(ymmA, ymmB, ymmC); + cc.vpsignd(ymmA, ymmB, ymmC); + cc.vpsignw(ymmA, ymmB, ymmC); + cc.vpslld(ymmA, ymmB, xmmC); + cc.vpsllq(ymmA, ymmB, xmmC); + cc.vpsllvd(xmmA, xmmB, xmmC); + cc.vpsllvd(ymmA, ymmB, ymmC); + cc.vpsllvq(xmmA, xmmB, xmmC); + cc.vpsllvq(ymmA, ymmB, ymmC); + cc.vpsllw(ymmA, ymmB, xmmC); + cc.vpsrad(ymmA, ymmB, xmmC); + cc.vpsravd(xmmA, xmmB, xmmC); + cc.vpsravd(ymmA, ymmB, ymmC); + cc.vpsraw(ymmA, ymmB, xmmC); + cc.vpsrld(ymmA, ymmB, xmmC); + cc.vpsrlq(ymmA, ymmB, xmmC); + cc.vpsrlvd(xmmA, xmmB, xmmC); + cc.vpsrlvd(ymmA, ymmB, ymmC); + cc.vpsrlvq(xmmA, xmmB, xmmC); + cc.vpsrlvq(ymmA, ymmB, ymmC); + cc.vpsrlw(ymmA, ymmB, xmmC); + cc.vpsubb(ymmA, ymmB, ymmC); + cc.vpsubd(ymmA, ymmB, ymmC); + cc.vpsubq(ymmA, ymmB, ymmC); + cc.vpsubsb(ymmA, ymmB, ymmC); + cc.vpsubsw(ymmA, ymmB, ymmC); + cc.vpsubusb(ymmA, ymmB, ymmC); + cc.vpsubusw(ymmA, ymmB, ymmC); + cc.vpsubw(ymmA, ymmB, ymmC); + cc.vpunpckhbw(ymmA, ymmB, ymmC); + cc.vpunpckhdq(ymmA, ymmB, ymmC); + cc.vpunpckhqdq(ymmA, ymmB, ymmC); + cc.vpunpckhwd(ymmA, ymmB, ymmC); + cc.vpunpcklbw(ymmA, ymmB, ymmC); + cc.vpunpckldq(ymmA, ymmB, ymmC); + cc.vpunpcklqdq(ymmA, ymmB, ymmC); + cc.vpunpcklwd(ymmA, ymmB, ymmC); + cc.vpxor(ymmA, ymmB, ymmC); + + // FMA. + cc.vfmadd132pd(xmmA, xmmB, xmmC); + cc.vfmadd132pd(ymmA, ymmB, ymmC); + cc.vfmadd132ps(xmmA, xmmB, xmmC); + cc.vfmadd132ps(ymmA, ymmB, ymmC); + cc.vfmadd132sd(xmmA, xmmB, xmmC); + cc.vfmadd132ss(xmmA, xmmB, xmmC); + cc.vfmadd213pd(xmmA, xmmB, xmmC); + cc.vfmadd213pd(ymmA, ymmB, ymmC); + cc.vfmadd213ps(xmmA, xmmB, xmmC); + cc.vfmadd213ps(ymmA, ymmB, ymmC); + cc.vfmadd213sd(xmmA, xmmB, xmmC); + cc.vfmadd213ss(xmmA, xmmB, xmmC); + cc.vfmadd231pd(xmmA, xmmB, xmmC); + cc.vfmadd231pd(ymmA, ymmB, ymmC); + cc.vfmadd231ps(xmmA, xmmB, xmmC); + cc.vfmadd231ps(ymmA, ymmB, ymmC); + cc.vfmadd231sd(xmmA, xmmB, xmmC); + cc.vfmadd231ss(xmmA, xmmB, xmmC); + cc.vfmaddsub132pd(xmmA, xmmB, xmmC); + cc.vfmaddsub132pd(ymmA, ymmB, ymmC); + cc.vfmaddsub132ps(xmmA, xmmB, xmmC); + cc.vfmaddsub132ps(ymmA, ymmB, ymmC); + cc.vfmaddsub213pd(xmmA, xmmB, xmmC); + cc.vfmaddsub213pd(ymmA, ymmB, ymmC); + cc.vfmaddsub213ps(xmmA, xmmB, xmmC); + cc.vfmaddsub213ps(ymmA, ymmB, ymmC); + cc.vfmaddsub231pd(xmmA, xmmB, xmmC); + cc.vfmaddsub231pd(ymmA, ymmB, ymmC); + cc.vfmaddsub231ps(xmmA, xmmB, xmmC); + cc.vfmaddsub231ps(ymmA, ymmB, ymmC); + cc.vfmsub132pd(xmmA, xmmB, xmmC); + cc.vfmsub132pd(ymmA, ymmB, ymmC); + cc.vfmsub132ps(xmmA, xmmB, xmmC); + cc.vfmsub132ps(ymmA, ymmB, ymmC); + cc.vfmsub132sd(xmmA, xmmB, xmmC); + cc.vfmsub132ss(xmmA, xmmB, xmmC); + cc.vfmsub213pd(xmmA, xmmB, xmmC); + cc.vfmsub213pd(ymmA, ymmB, ymmC); + cc.vfmsub213ps(xmmA, xmmB, xmmC); + cc.vfmsub213ps(ymmA, ymmB, ymmC); + cc.vfmsub213sd(xmmA, xmmB, xmmC); + cc.vfmsub213ss(xmmA, xmmB, xmmC); + cc.vfmsub231pd(xmmA, xmmB, xmmC); + cc.vfmsub231pd(ymmA, ymmB, ymmC); + cc.vfmsub231ps(xmmA, xmmB, xmmC); + cc.vfmsub231ps(ymmA, ymmB, ymmC); + cc.vfmsub231sd(xmmA, xmmB, xmmC); + cc.vfmsub231ss(xmmA, xmmB, xmmC); + cc.vfmsubadd132pd(xmmA, xmmB, xmmC); + cc.vfmsubadd132pd(ymmA, ymmB, ymmC); + cc.vfmsubadd132ps(xmmA, xmmB, xmmC); + cc.vfmsubadd132ps(ymmA, ymmB, ymmC); + cc.vfmsubadd213pd(xmmA, xmmB, xmmC); + cc.vfmsubadd213pd(ymmA, ymmB, ymmC); + cc.vfmsubadd213ps(xmmA, xmmB, xmmC); + cc.vfmsubadd213ps(ymmA, ymmB, ymmC); + cc.vfmsubadd231pd(xmmA, xmmB, xmmC); + cc.vfmsubadd231pd(ymmA, ymmB, ymmC); + cc.vfmsubadd231ps(xmmA, xmmB, xmmC); + cc.vfmsubadd231ps(ymmA, ymmB, ymmC); + cc.vfnmadd132pd(xmmA, xmmB, xmmC); + cc.vfnmadd132pd(ymmA, ymmB, ymmC); + cc.vfnmadd132ps(xmmA, xmmB, xmmC); + cc.vfnmadd132ps(ymmA, ymmB, ymmC); + cc.vfnmadd132sd(xmmA, xmmB, xmmC); + cc.vfnmadd132ss(xmmA, xmmB, xmmC); + cc.vfnmadd213pd(xmmA, xmmB, xmmC); + cc.vfnmadd213pd(ymmA, ymmB, ymmC); + cc.vfnmadd213ps(xmmA, xmmB, xmmC); + cc.vfnmadd213ps(ymmA, ymmB, ymmC); + cc.vfnmadd213sd(xmmA, xmmB, xmmC); + cc.vfnmadd213ss(xmmA, xmmB, xmmC); + cc.vfnmadd231pd(xmmA, xmmB, xmmC); + cc.vfnmadd231pd(ymmA, ymmB, ymmC); + cc.vfnmadd231ps(xmmA, xmmB, xmmC); + cc.vfnmadd231ps(ymmA, ymmB, ymmC); + cc.vfnmadd231sd(xmmA, xmmB, xmmC); + cc.vfnmadd231ss(xmmA, xmmB, xmmC); + cc.vfnmsub132pd(xmmA, xmmB, xmmC); + cc.vfnmsub132pd(ymmA, ymmB, ymmC); + cc.vfnmsub132ps(xmmA, xmmB, xmmC); + cc.vfnmsub132ps(ymmA, ymmB, ymmC); + cc.vfnmsub132sd(xmmA, xmmB, xmmC); + cc.vfnmsub132ss(xmmA, xmmB, xmmC); + cc.vfnmsub213pd(xmmA, xmmB, xmmC); + cc.vfnmsub213pd(ymmA, ymmB, ymmC); + cc.vfnmsub213ps(xmmA, xmmB, xmmC); + cc.vfnmsub213ps(ymmA, ymmB, ymmC); + cc.vfnmsub213sd(xmmA, xmmB, xmmC); + cc.vfnmsub213ss(xmmA, xmmB, xmmC); + cc.vfnmsub231pd(xmmA, xmmB, xmmC); + cc.vfnmsub231pd(ymmA, ymmB, ymmC); + cc.vfnmsub231ps(xmmA, xmmB, xmmC); + cc.vfnmsub231ps(ymmA, ymmB, ymmC); + cc.vfnmsub231sd(xmmA, xmmB, xmmC); + cc.vfnmsub231ss(xmmA, xmmB, xmmC); + } + else { + x86::Mem m = x86::ptr(gpz); + x86::Mem m128 = x86::xmmword_ptr(gpz); + x86::Mem m256 = x86::xmmword_ptr(gpz); + x86::Mem vx_ptr = x86::ptr(gpz, xmmD); + x86::Mem vy_ptr = x86::ptr(gpz, ymmD); + + cc.vaddpd(xmmA, xmmB, m); + cc.vaddpd(ymmA, ymmB, m); + cc.vaddps(xmmA, xmmB, m); + cc.vaddps(ymmA, ymmB, m); + cc.vaddsd(xmmA, xmmB, m); + cc.vaddss(xmmA, xmmB, m); + cc.vaddsubpd(xmmA, xmmB, m); + cc.vaddsubpd(ymmA, ymmB, m); + cc.vaddsubps(xmmA, xmmB, m); + cc.vaddsubps(ymmA, ymmB, m); + cc.vandpd(xmmA, xmmB, m); + cc.vandpd(ymmA, ymmB, m); + cc.vandps(xmmA, xmmB, m); + cc.vandps(ymmA, ymmB, m); + cc.vandnpd(xmmA, xmmB, m); + cc.vandnpd(ymmA, ymmB, m); + cc.vandnps(xmmA, xmmB, m); + cc.vandnps(ymmA, ymmB, m); + cc.vblendpd(xmmA, xmmB, m, 0); + cc.vblendpd(ymmA, ymmB, m, 0); + cc.vblendps(xmmA, xmmB, m, 0); + cc.vblendps(ymmA, ymmB, m, 0); + cc.vblendvpd(xmmA, xmmB, m, xmmA); + cc.vblendvpd(ymmA, ymmB, m, ymmA); + cc.vbroadcastf128(ymmA, m); + cc.vbroadcastsd(ymmA, m); + cc.vbroadcastss(xmmA, m); + cc.vbroadcastss(ymmA, m); + cc.vcmppd(xmmA, xmmB, m, 0); + cc.vcmppd(ymmA, ymmB, m, 0); + cc.vcmpps(xmmA, xmmB, m, 0); + cc.vcmpps(ymmA, ymmB, m, 0); + cc.vcmpsd(xmmA, xmmB, m, 0); + cc.vcmpss(xmmA, xmmB, m, 0); + cc.vcomisd(xmmA, m); + cc.vcomiss(xmmA, m); + cc.vcvtdq2pd(xmmA, m); + cc.vcvtdq2pd(ymmA, m); + cc.vcvtdq2ps(xmmA, m); + cc.vcvtdq2ps(ymmA, m); + cc.vcvtpd2dq(xmmA, m128); + cc.vcvtpd2dq(xmmA, m256); + cc.vcvtpd2ps(xmmA, m128); + cc.vcvtpd2ps(xmmA, m256); + cc.vcvtps2dq(xmmA, m); + cc.vcvtps2dq(ymmA, m); + cc.vcvtps2pd(xmmA, m); + cc.vcvtps2pd(ymmA, m); + cc.vcvtsd2si(gpd, m); + cc.vcvtsd2ss(xmmA, xmmB, m); + cc.vcvtsi2sd(xmmA, xmmB, m); + cc.vcvtsi2ss(xmmA, xmmB, m); + cc.vcvtss2sd(xmmA, xmmB, m); + cc.vcvtss2si(gpd, m); + cc.vcvttpd2dq(xmmA, m128); + cc.vcvttpd2dq(xmmA, m256); + cc.vcvttps2dq(xmmA, m); + cc.vcvttps2dq(ymmA, m); + cc.vcvttsd2si(gpd, m); + cc.vcvttss2si(gpd, m); + cc.vdivpd(xmmA, xmmB, m); + cc.vdivpd(ymmA, ymmB, m); + cc.vdivps(xmmA, xmmB, m); + cc.vdivps(ymmA, ymmB, m); + cc.vdivsd(xmmA, xmmB, m); + cc.vdivss(xmmA, xmmB, m); + cc.vdppd(xmmA, xmmB, m, 0); + cc.vdpps(xmmA, xmmB, m, 0); + cc.vdpps(ymmA, ymmB, m, 0); + cc.vextractf128(m, ymmB, 0); + cc.vextractps(m, xmmB, 0); + cc.vhaddpd(xmmA, xmmB, m); + cc.vhaddpd(ymmA, ymmB, m); + cc.vhaddps(xmmA, xmmB, m); + cc.vhaddps(ymmA, ymmB, m); + cc.vhsubpd(xmmA, xmmB, m); + cc.vhsubpd(ymmA, ymmB, m); + cc.vhsubps(xmmA, xmmB, m); + cc.vhsubps(ymmA, ymmB, m); + cc.vinsertf128(ymmA, ymmB, m, 0); + cc.vinsertps(xmmA, xmmB, m, 0); + cc.vlddqu(xmmA, m); + cc.vlddqu(ymmA, m); + cc.vmaskmovps(xmmA, xmmB, m); + cc.vmaskmovps(ymmA, ymmB, m); + cc.vmaskmovps(m, xmmB, xmmC); + cc.vmaskmovps(m, ymmB, ymmC); + cc.vmaskmovpd(xmmA, xmmB, m); + cc.vmaskmovpd(ymmA, ymmB, m); + cc.vmaskmovpd(m, xmmB, xmmC); + cc.vmaskmovpd(m, ymmB, ymmC); + cc.vmaxpd(xmmA, xmmB, m); + cc.vmaxpd(ymmA, ymmB, m); + cc.vmaxps(xmmA, xmmB, m); + cc.vmaxps(ymmA, ymmB, m); + cc.vmaxsd(xmmA, xmmB, m); + cc.vmaxss(xmmA, xmmB, m); + cc.vminpd(xmmA, xmmB, m); + cc.vminpd(ymmA, ymmB, m); + cc.vminps(xmmA, xmmB, m); + cc.vminps(ymmA, ymmB, m); + cc.vminsd(xmmA, xmmB, m); + cc.vminss(xmmA, xmmB, m); + cc.vmovapd(xmmA, m); + cc.vmovapd(m, xmmB); + cc.vmovapd(ymmA, m); + cc.vmovapd(m, ymmB); + cc.vmovaps(xmmA, m); + cc.vmovaps(m, xmmB); + cc.vmovaps(ymmA, m); + cc.vmovaps(m, ymmB); + cc.vmovd(xmmA, m); + cc.vmovd(m, xmmB); + cc.vmovddup(xmmA, m); + cc.vmovddup(ymmA, m); + cc.vmovdqa(xmmA, m); + cc.vmovdqa(m, xmmB); + cc.vmovdqa(ymmA, m); + cc.vmovdqa(m, ymmB); + cc.vmovdqu(xmmA, m); + cc.vmovdqu(m, xmmB); + cc.vmovdqu(ymmA, m); + cc.vmovdqu(m, ymmB); + cc.vmovhpd(xmmA, xmmB, m); + cc.vmovhps(xmmA, xmmB, m); + cc.vmovhps(m, xmmB); + cc.vmovlpd(xmmA, xmmB, m); + cc.vmovlpd(m, xmmB); + cc.vmovlps(xmmA, xmmB, m); + cc.vmovlps(m, xmmB); + cc.vmovntdq(m, xmmB); + cc.vmovntdq(m, ymmB); + cc.vmovntdqa(xmmA, m); + cc.vmovntpd(m, xmmB); + cc.vmovntpd(m, ymmB); + cc.vmovntps(m, xmmB); + cc.vmovntps(m, ymmB); + cc.vmovsd(xmmA, m); + cc.vmovsd(m, xmmB); + cc.vmovshdup(xmmA, m); + cc.vmovshdup(ymmA, m); + cc.vmovsldup(xmmA, m); + cc.vmovsldup(ymmA, m); + cc.vmovss(xmmA, m); + cc.vmovss(m, xmmB); + cc.vmovupd(xmmA, m); + cc.vmovupd(m, xmmB); + cc.vmovupd(ymmA, m); + cc.vmovupd(m, ymmB); + cc.vmovups(xmmA, m); + cc.vmovups(m, xmmB); + cc.vmovups(ymmA, m); + cc.vmovups(m, ymmB); + cc.vmpsadbw(xmmA, xmmB, m, 0); + cc.vmulpd(xmmA, xmmB, m); + cc.vmulpd(ymmA, ymmB, m); + cc.vmulps(xmmA, xmmB, m); + cc.vmulps(ymmA, ymmB, m); + cc.vmulsd(xmmA, xmmB, m); + cc.vmulss(xmmA, xmmB, m); + cc.vorpd(xmmA, xmmB, m); + cc.vorpd(ymmA, ymmB, m); + cc.vorps(xmmA, xmmB, m); + cc.vorps(ymmA, ymmB, m); + cc.vpabsb(xmmA, m); + cc.vpabsd(xmmA, m); + cc.vpabsw(xmmA, m); + cc.vpackssdw(xmmA, xmmB, m); + cc.vpacksswb(xmmA, xmmB, m); + cc.vpackusdw(xmmA, xmmB, m); + cc.vpackuswb(xmmA, xmmB, m); + cc.vpaddb(xmmA, xmmB, m); + cc.vpaddd(xmmA, xmmB, m); + cc.vpaddq(xmmA, xmmB, m); + cc.vpaddw(xmmA, xmmB, m); + cc.vpaddsb(xmmA, xmmB, m); + cc.vpaddsw(xmmA, xmmB, m); + cc.vpaddusb(xmmA, xmmB, m); + cc.vpaddusw(xmmA, xmmB, m); + cc.vpalignr(xmmA, xmmB, m, 0); + cc.vpand(xmmA, xmmB, m); + cc.vpandn(xmmA, xmmB, m); + cc.vpavgb(xmmA, xmmB, m); + cc.vpavgw(xmmA, xmmB, m); + cc.vpblendvb(xmmA, xmmB, m, xmmA); + cc.vpblendw(xmmA, xmmB, m, 0); + cc.vpcmpeqb(xmmA, xmmB, m); + cc.vpcmpeqd(xmmA, xmmB, m); + cc.vpcmpeqq(xmmA, xmmB, m); + cc.vpcmpeqw(xmmA, xmmB, m); + cc.vpcmpgtb(xmmA, xmmB, m); + cc.vpcmpgtd(xmmA, xmmB, m); + cc.vpcmpgtq(xmmA, xmmB, m); + cc.vpcmpgtw(xmmA, xmmB, m); + cc.vpermilpd(xmmA, xmmB, m); + cc.vpermilpd(ymmA, ymmB, m); + cc.vpermilpd(xmmA, m, 0); + cc.vpermilpd(ymmA, m, 0); + cc.vpermilps(xmmA, xmmB, m); + cc.vpermilps(ymmA, ymmB, m); + cc.vpermilps(xmmA, m, 0); + cc.vpermilps(ymmA, m, 0); + cc.vperm2f128(ymmA, ymmB, m, 0); + cc.vpextrb(m, xmmB, 0); + cc.vpextrd(m, xmmB, 0); + if (cc.is64Bit()) cc.vpextrq(m, xmmB, 0); + cc.vpextrw(m, xmmB, 0); + cc.vphaddd(xmmA, xmmB, m); + cc.vphaddsw(xmmA, xmmB, m); + cc.vphaddw(xmmA, xmmB, m); + cc.vphminposuw(xmmA, m); + cc.vphsubd(xmmA, xmmB, m); + cc.vphsubsw(xmmA, xmmB, m); + cc.vphsubw(xmmA, xmmB, m); + cc.vpinsrb(xmmA, xmmB, m, 0); + cc.vpinsrd(xmmA, xmmB, m, 0); + cc.vpinsrw(xmmA, xmmB, m, 0); + cc.vpmaddubsw(xmmA, xmmB, m); + cc.vpmaddwd(xmmA, xmmB, m); + cc.vpmaxsb(xmmA, xmmB, m); + cc.vpmaxsd(xmmA, xmmB, m); + cc.vpmaxsw(xmmA, xmmB, m); + cc.vpmaxub(xmmA, xmmB, m); + cc.vpmaxud(xmmA, xmmB, m); + cc.vpmaxuw(xmmA, xmmB, m); + cc.vpminsb(xmmA, xmmB, m); + cc.vpminsd(xmmA, xmmB, m); + cc.vpminsw(xmmA, xmmB, m); + cc.vpminub(xmmA, xmmB, m); + cc.vpminud(xmmA, xmmB, m); + cc.vpminuw(xmmA, xmmB, m); + cc.vpmovsxbd(xmmA, m); + cc.vpmovsxbq(xmmA, m); + cc.vpmovsxbw(xmmA, m); + cc.vpmovsxdq(xmmA, m); + cc.vpmovsxwd(xmmA, m); + cc.vpmovsxwq(xmmA, m); + cc.vpmovzxbd(xmmA, m); + cc.vpmovzxbq(xmmA, m); + cc.vpmovzxbw(xmmA, m); + cc.vpmovzxdq(xmmA, m); + cc.vpmovzxwd(xmmA, m); + cc.vpmovzxwq(xmmA, m); + cc.vpmuldq(xmmA, xmmB, m); + cc.vpmulhrsw(xmmA, xmmB, m); + cc.vpmulhuw(xmmA, xmmB, m); + cc.vpmulhw(xmmA, xmmB, m); + cc.vpmulld(xmmA, xmmB, m); + cc.vpmullw(xmmA, xmmB, m); + cc.vpmuludq(xmmA, xmmB, m); + cc.vpor(xmmA, xmmB, m); + cc.vpsadbw(xmmA, xmmB, m); + cc.vpshufb(xmmA, xmmB, m); + cc.vpshufd(xmmA, m, 0); + cc.vpshufhw(xmmA, m, 0); + cc.vpshuflw(xmmA, m, 0); + cc.vpsignb(xmmA, xmmB, m); + cc.vpsignd(xmmA, xmmB, m); + cc.vpsignw(xmmA, xmmB, m); + cc.vpslld(xmmA, xmmB, m); + cc.vpsllq(xmmA, xmmB, m); + cc.vpsllw(xmmA, xmmB, m); + cc.vpsrad(xmmA, xmmB, m); + cc.vpsraw(xmmA, xmmB, m); + cc.vpsrld(xmmA, xmmB, m); + cc.vpsrlq(xmmA, xmmB, m); + cc.vpsrlw(xmmA, xmmB, m); + cc.vpsubb(xmmA, xmmB, m); + cc.vpsubd(xmmA, xmmB, m); + cc.vpsubq(xmmA, xmmB, m); + cc.vpsubw(xmmA, xmmB, m); + cc.vpsubsb(xmmA, xmmB, m); + cc.vpsubsw(xmmA, xmmB, m); + cc.vpsubusb(xmmA, xmmB, m); + cc.vpsubusw(xmmA, xmmB, m); + cc.vptest(xmmA, m); + cc.vptest(ymmA, m); + cc.vpunpckhbw(xmmA, xmmB, m); + cc.vpunpckhdq(xmmA, xmmB, m); + cc.vpunpckhqdq(xmmA, xmmB, m); + cc.vpunpckhwd(xmmA, xmmB, m); + cc.vpunpcklbw(xmmA, xmmB, m); + cc.vpunpckldq(xmmA, xmmB, m); + cc.vpunpcklqdq(xmmA, xmmB, m); + cc.vpunpcklwd(xmmA, xmmB, m); + cc.vpxor(xmmA, xmmB, m); + cc.vrcpps(xmmA, m); + cc.vrcpps(ymmA, m); + cc.vrcpss(xmmA, xmmB, m); + cc.vrsqrtps(xmmA, m); + cc.vrsqrtps(ymmA, m); + cc.vrsqrtss(xmmA, xmmB, m); + cc.vroundpd(xmmA, m, 0); + cc.vroundpd(ymmA, m, 0); + cc.vroundps(xmmA, m, 0); + cc.vroundps(ymmA, m, 0); + cc.vroundsd(xmmA, xmmB, m, 0); + cc.vroundss(xmmA, xmmB, m, 0); + cc.vshufpd(xmmA, xmmB, m, 0); + cc.vshufpd(ymmA, ymmB, m, 0); + cc.vshufps(xmmA, xmmB, m, 0); + cc.vshufps(ymmA, ymmB, m, 0); + cc.vsqrtpd(xmmA, m); + cc.vsqrtpd(ymmA, m); + cc.vsqrtps(xmmA, m); + cc.vsqrtps(ymmA, m); + cc.vsqrtsd(xmmA, xmmB, m); + cc.vsqrtss(xmmA, xmmB, m); + cc.vsubpd(xmmA, xmmB, m); + cc.vsubpd(ymmA, ymmB, m); + cc.vsubps(xmmA, xmmB, m); + cc.vsubps(ymmA, ymmB, m); + cc.vsubsd(xmmA, xmmB, m); + cc.vsubss(xmmA, xmmB, m); + cc.vtestps(xmmA, m); + cc.vtestps(ymmA, m); + cc.vtestpd(xmmA, m); + cc.vtestpd(ymmA, m); + cc.vucomisd(xmmA, m); + cc.vucomiss(xmmA, m); + cc.vunpckhpd(xmmA, xmmB, m); + cc.vunpckhpd(ymmA, ymmB, m); + cc.vunpckhps(xmmA, xmmB, m); + cc.vunpckhps(ymmA, ymmB, m); + cc.vunpcklpd(xmmA, xmmB, m); + cc.vunpcklpd(ymmA, ymmB, m); + cc.vunpcklps(xmmA, xmmB, m); + cc.vunpcklps(ymmA, ymmB, m); + cc.vxorpd(xmmA, xmmB, m); + cc.vxorpd(ymmA, ymmB, m); + cc.vxorps(xmmA, xmmB, m); + cc.vxorps(ymmA, ymmB, m); + + // AVX+AESNI. + cc.vaesdec(xmmA, xmmB, m); + cc.vaesdeclast(xmmA, xmmB, m); + cc.vaesenc(xmmA, xmmB, m); + cc.vaesenclast(xmmA, xmmB, m); + cc.vaesimc(xmmA, m); + cc.vaeskeygenassist(xmmA, m, 0); + + // AVX+PCLMULQDQ. + cc.vpclmulqdq(xmmA, xmmB, m, 0); + + // AVX2. + cc.vbroadcasti128(ymmA, m); + cc.vextracti128(m, ymmB, 0); + cc.vgatherdpd(xmmA, vx_ptr, xmmC); + cc.vgatherdpd(ymmA, vx_ptr, ymmC); + cc.vgatherdps(xmmA, vx_ptr, xmmC); + cc.vgatherdps(ymmA, vy_ptr, ymmC); + cc.vgatherqpd(xmmA, vx_ptr, xmmC); + cc.vgatherqpd(ymmA, vy_ptr, ymmC); + cc.vgatherqps(xmmA, vx_ptr, xmmC); + cc.vgatherqps(xmmA, vy_ptr, xmmC); + cc.vinserti128(ymmA, ymmB, m, 0); + cc.vmovntdqa(ymmA, m); + cc.vmpsadbw(ymmA, ymmB, m, 0); + cc.vpabsb(ymmA, m); + cc.vpabsd(ymmA, m); + cc.vpabsw(ymmA, m); + cc.vpackssdw(ymmA, ymmB, m); + cc.vpacksswb(ymmA, ymmB, m); + cc.vpackusdw(ymmA, ymmB, m); + cc.vpackuswb(ymmA, ymmB, m); + cc.vpaddb(ymmA, ymmB, m); + cc.vpaddd(ymmA, ymmB, m); + cc.vpaddq(ymmA, ymmB, m); + cc.vpaddw(ymmA, ymmB, m); + cc.vpaddsb(ymmA, ymmB, m); + cc.vpaddsw(ymmA, ymmB, m); + cc.vpaddusb(ymmA, ymmB, m); + cc.vpaddusw(ymmA, ymmB, m); + cc.vpalignr(ymmA, ymmB, m, 0); + cc.vpand(ymmA, ymmB, m); + cc.vpandn(ymmA, ymmB, m); + cc.vpavgb(ymmA, ymmB, m); + cc.vpavgw(ymmA, ymmB, m); + cc.vpblendd(xmmA, xmmB, m, 0); + cc.vpblendd(ymmA, ymmB, m, 0); + cc.vpblendvb(ymmA, ymmB, m, ymmA); + cc.vpblendw(ymmA, ymmB, m, 0); + cc.vpbroadcastb(xmmA, m); + cc.vpbroadcastb(ymmA, m); + cc.vpbroadcastd(xmmA, m); + cc.vpbroadcastd(ymmA, m); + cc.vpbroadcastq(xmmA, m); + cc.vpbroadcastq(ymmA, m); + cc.vpbroadcastw(xmmA, m); + cc.vpbroadcastw(ymmA, m); + cc.vpcmpeqb(ymmA, ymmB, m); + cc.vpcmpeqd(ymmA, ymmB, m); + cc.vpcmpeqq(ymmA, ymmB, m); + cc.vpcmpeqw(ymmA, ymmB, m); + cc.vpcmpgtb(ymmA, ymmB, m); + cc.vpcmpgtd(ymmA, ymmB, m); + cc.vpcmpgtq(ymmA, ymmB, m); + cc.vpcmpgtw(ymmA, ymmB, m); + cc.vperm2i128(ymmA, ymmB, m, 0); + cc.vpermd(ymmA, ymmB, m); + cc.vpermps(ymmA, ymmB, m); + cc.vpermpd(ymmA, m, 0); + cc.vpermq(ymmA, m, 0); + cc.vpgatherdd(xmmA, vx_ptr, xmmC); + cc.vpgatherdd(ymmA, vy_ptr, ymmC); + cc.vpgatherdq(xmmA, vx_ptr, xmmC); + cc.vpgatherdq(ymmA, vx_ptr, ymmC); + cc.vpgatherqd(xmmA, vx_ptr, xmmC); + cc.vpgatherqd(xmmA, vy_ptr, xmmC); + cc.vpgatherqq(xmmA, vx_ptr, xmmC); + cc.vpgatherqq(ymmA, vy_ptr, ymmC); + cc.vpmovsxbd(ymmA, m); + cc.vpmovsxbq(ymmA, m); + cc.vpmovsxbw(ymmA, m); + cc.vpmovsxdq(ymmA, m); + cc.vpmovsxwd(ymmA, m); + cc.vpmovsxwq(ymmA, m); + cc.vpmovzxbd(ymmA, m); + cc.vpmovzxbq(ymmA, m); + cc.vpmovzxbw(ymmA, m); + cc.vpmovzxdq(ymmA, m); + cc.vpmovzxwd(ymmA, m); + cc.vpmovzxwq(ymmA, m); + cc.vpshufd(ymmA, m, 0); + cc.vpshufhw(ymmA, m, 0); + cc.vpshuflw(ymmA, m, 0); + cc.vphaddd(ymmA, ymmB, m); + cc.vphaddsw(ymmA, ymmB, m); + cc.vphaddw(ymmA, ymmB, m); + cc.vphsubd(ymmA, ymmB, m); + cc.vphsubsw(ymmA, ymmB, m); + cc.vphsubw(ymmA, ymmB, m); + cc.vpmaddubsw(ymmA, ymmB, m); + cc.vpmaddwd(ymmA, ymmB, m); + cc.vpmaskmovd(m, xmmB, xmmC); + cc.vpmaskmovd(m, ymmB, ymmC); + cc.vpmaskmovd(xmmA, xmmB, m); + cc.vpmaskmovd(ymmA, ymmB, m); + cc.vpmaskmovq(m, xmmB, xmmC); + cc.vpmaskmovq(m, ymmB, ymmC); + cc.vpmaskmovq(xmmA, xmmB, m); + cc.vpmaskmovq(ymmA, ymmB, m); + cc.vpmaxsb(ymmA, ymmB, m); + cc.vpmaxsd(ymmA, ymmB, m); + cc.vpmaxsw(ymmA, ymmB, m); + cc.vpmaxub(ymmA, ymmB, m); + cc.vpmaxud(ymmA, ymmB, m); + cc.vpmaxuw(ymmA, ymmB, m); + cc.vpminsb(ymmA, ymmB, m); + cc.vpminsd(ymmA, ymmB, m); + cc.vpminsw(ymmA, ymmB, m); + cc.vpminub(ymmA, ymmB, m); + cc.vpminud(ymmA, ymmB, m); + cc.vpminuw(ymmA, ymmB, m); + cc.vpmuldq(ymmA, ymmB, m); + cc.vpmulhrsw(ymmA, ymmB, m); + cc.vpmulhuw(ymmA, ymmB, m); + cc.vpmulhw(ymmA, ymmB, m); + cc.vpmulld(ymmA, ymmB, m); + cc.vpmullw(ymmA, ymmB, m); + cc.vpmuludq(ymmA, ymmB, m); + cc.vpor(ymmA, ymmB, m); + cc.vpsadbw(ymmA, ymmB, m); + cc.vpshufb(ymmA, ymmB, m); + cc.vpsignb(ymmA, ymmB, m); + cc.vpsignd(ymmA, ymmB, m); + cc.vpsignw(ymmA, ymmB, m); + cc.vpslld(ymmA, ymmB, m); + cc.vpsllq(ymmA, ymmB, m); + cc.vpsllvd(xmmA, xmmB, m); + cc.vpsllvd(ymmA, ymmB, m); + cc.vpsllvq(xmmA, xmmB, m); + cc.vpsllvq(ymmA, ymmB, m); + cc.vpsllw(ymmA, ymmB, m); + cc.vpsrad(ymmA, ymmB, m); + cc.vpsravd(xmmA, xmmB, m); + cc.vpsravd(ymmA, ymmB, m); + cc.vpsraw(ymmA, ymmB, m); + cc.vpsrld(ymmA, ymmB, m); + cc.vpsrlq(ymmA, ymmB, m); + cc.vpsrlvd(xmmA, xmmB, m); + cc.vpsrlvd(ymmA, ymmB, m); + cc.vpsrlvq(xmmA, xmmB, m); + cc.vpsrlvq(ymmA, ymmB, m); + cc.vpsrlw(ymmA, ymmB, m); + cc.vpsubb(ymmA, ymmB, m); + cc.vpsubd(ymmA, ymmB, m); + cc.vpsubq(ymmA, ymmB, m); + cc.vpsubsb(ymmA, ymmB, m); + cc.vpsubsw(ymmA, ymmB, m); + cc.vpsubusb(ymmA, ymmB, m); + cc.vpsubusw(ymmA, ymmB, m); + cc.vpsubw(ymmA, ymmB, m); + cc.vpunpckhbw(ymmA, ymmB, m); + cc.vpunpckhdq(ymmA, ymmB, m); + cc.vpunpckhqdq(ymmA, ymmB, m); + cc.vpunpckhwd(ymmA, ymmB, m); + cc.vpunpcklbw(ymmA, ymmB, m); + cc.vpunpckldq(ymmA, ymmB, m); + cc.vpunpcklqdq(ymmA, ymmB, m); + cc.vpunpcklwd(ymmA, ymmB, m); + cc.vpxor(ymmA, ymmB, m); + } +} + +static void generateAvxSequence(BaseEmitter& emitter, InstForm form, bool emitPrologEpilog) { + using namespace asmjit::x86; + + if (emitter.isAssembler()) { + Assembler& cc = *emitter.as<Assembler>(); + + if (emitPrologEpilog) { + FuncDetail func; + func.init(FuncSignatureT<void, void*, const void*, size_t>(CallConv::kIdHost), cc.environment()); + + FuncFrame frame; + frame.init(func); + frame.addDirtyRegs(eax, ymm0, ymm1, ymm2, ymm3); + frame.finalize(); + + cc.emitProlog(frame); + generateAvxSequenceInternal(cc, form, eax, ymm0, ymm1, ymm2, ymm3); + cc.emitEpilog(frame); + } + else { + generateAvxSequenceInternal(cc, form, eax, ymm0, ymm1, ymm2, ymm3); + } + } +#ifndef ASMJIT_NO_BUILDER + else if (emitter.isBuilder()) { + Builder& cc = *emitter.as<Builder>(); + + if (emitPrologEpilog) { + FuncDetail func; + func.init(FuncSignatureT<void, void*, const void*, size_t>(CallConv::kIdHost), cc.environment()); + + FuncFrame frame; + frame.init(func); + frame.addDirtyRegs(eax, ymm0, ymm1, ymm2, ymm3); + frame.finalize(); + + cc.emitProlog(frame); + generateAvxSequenceInternal(cc, form, eax, ymm0, ymm1, ymm2, ymm3); + cc.emitEpilog(frame); + } + else { + generateAvxSequenceInternal(cc, form, eax, ymm0, ymm1, ymm2, ymm3); + } + } +#endif +#ifndef ASMJIT_NO_COMPILER + else if (emitter.isCompiler()) { + Compiler& cc = *emitter.as<Compiler>(); + + Gp gp = cc.newGpz("gp"); + Ymm a = cc.newYmm("a"); + Ymm b = cc.newYmm("b"); + Ymm c = cc.newYmm("c"); + Ymm d = cc.newYmm("d"); + + cc.addFunc(FuncSignatureT<void>(CallConv::kIdHost)); + generateAvxSequenceInternal(cc, form, gp, a, b, c, d); + cc.endFunc(); + } +#endif +} + +// Generates a long sequence of AVX512 instructions. +template<typename Emitter> +static void generateAvx512SequenceInternal( + Emitter& cc, + InstForm form, + const x86::Gp& gp, + const x86::KReg& kA, const x86::KReg& kB, const x86::KReg& kC, + const x86::Vec& vecA, const x86::Vec& vecB, const x86::Vec& vecC, const x86::Vec& vecD) { + + x86::Gp gpd = gp.r32(); + x86::Gp gpq = gp.r64(); + x86::Gp gpz = cc.is32Bit() ? gpd : gpq; + + x86::Xmm xmmA = vecA.xmm(); + x86::Xmm xmmB = vecB.xmm(); + x86::Xmm xmmC = vecC.xmm(); + x86::Xmm xmmD = vecD.xmm(); + + x86::Ymm ymmA = vecA.ymm(); + x86::Ymm ymmB = vecB.ymm(); + x86::Ymm ymmC = vecC.ymm(); + x86::Ymm ymmD = vecD.ymm(); + + x86::Zmm zmmA = vecA.zmm(); + x86::Zmm zmmB = vecB.zmm(); + x86::Zmm zmmC = vecC.zmm(); + x86::Zmm zmmD = vecD.zmm(); + + cc.xor_(gpd, gpd); + cc.vxorps(xmmA, xmmA, xmmA); + cc.vxorps(xmmB, xmmB, xmmB); + cc.vxorps(xmmC, xmmC, xmmC); + cc.vxorps(xmmD, xmmD, xmmD); + + if (form == InstForm::kReg) { + cc.kaddb(kA, kB, kC); + cc.kaddd(kA, kB, kC); + cc.kaddq(kA, kB, kC); + cc.kaddw(kA, kB, kC); + cc.kandb(kA, kB, kC); + cc.kandd(kA, kB, kC); + cc.kandnb(kA, kB, kC); + cc.kandnd(kA, kB, kC); + cc.kandnq(kA, kB, kC); + cc.kandnw(kA, kB, kC); + cc.kandq(kA, kB, kC); + cc.kandw(kA, kB, kC); + cc.kmovb(kA, kB); + cc.kmovb(kA, gpd); + cc.kmovb(gpd, kB); + cc.kmovd(kA, kB); + cc.kmovd(kA, gpd); + cc.kmovd(gpd, kB); + cc.kmovq(kA, kB); + if (cc.is64Bit()) cc.kmovq(kA, gpq); + if (cc.is64Bit()) cc.kmovq(gpq, kB); + cc.kmovw(kA, kB); + cc.kmovw(kA, gpd); + cc.kmovw(gpd, kB); + cc.knotb(kA, kB); + cc.knotd(kA, kB); + cc.knotq(kA, kB); + cc.knotw(kA, kB); + cc.korb(kA, kB, kC); + cc.kord(kA, kB, kC); + cc.korq(kA, kB, kC); + cc.kortestb(kA, kB); + cc.kortestd(kA, kB); + cc.kortestq(kA, kB); + cc.kortestw(kA, kB); + cc.korw(kA, kB, kC); + cc.kshiftlb(kA, kB, 0); + cc.kshiftld(kA, kB, 0); + cc.kshiftlq(kA, kB, 0); + cc.kshiftlw(kA, kB, 0); + cc.kshiftrb(kA, kB, 0); + cc.kshiftrd(kA, kB, 0); + cc.kshiftrq(kA, kB, 0); + cc.kshiftrw(kA, kB, 0); + cc.ktestb(kA, kB); + cc.ktestd(kA, kB); + cc.ktestq(kA, kB); + cc.ktestw(kA, kB); + cc.kunpckbw(kA, kB, kC); + cc.kunpckdq(kA, kB, kC); + cc.kunpckwd(kA, kB, kC); + cc.kxnorb(kA, kB, kC); + cc.kxnord(kA, kB, kC); + cc.kxnorq(kA, kB, kC); + cc.kxnorw(kA, kB, kC); + cc.kxorb(kA, kB, kC); + cc.kxord(kA, kB, kC); + cc.kxorq(kA, kB, kC); + cc.kxorw(kA, kB, kC); + cc.nop(); + + cc.evex().vaddpd(xmmA, xmmB, xmmC); + cc.evex().vaddpd(ymmA, ymmB, ymmC); + cc.evex().vaddpd(zmmA, zmmB, zmmC); + cc.evex().vaddps(xmmA, xmmB, xmmC); + cc.evex().vaddps(ymmA, ymmB, ymmC); + cc.evex().vaddps(zmmA, zmmB, zmmC); + cc.evex().vaddsd(xmmA, xmmB, xmmC); + cc.evex().vaddss(xmmA, xmmB, xmmC); + cc.evex().valignd(xmmA, xmmB, xmmC, 0); + cc.evex().valignd(ymmA, ymmB, ymmC, 0); + cc.evex().valignd(zmmA, zmmB, zmmC, 0); + cc.evex().valignq(xmmA, xmmB, xmmC, 0); + cc.evex().valignq(ymmA, ymmB, ymmC, 0); + cc.evex().valignq(zmmA, zmmB, zmmC, 0); + cc.evex().vandnpd(xmmA, xmmB, xmmC); + cc.evex().vandnpd(ymmA, ymmB, ymmC); + cc.evex().vandnpd(zmmA, zmmB, zmmC); + cc.evex().vandnps(xmmA, xmmB, xmmC); + cc.evex().vandnps(ymmA, ymmB, ymmC); + cc.evex().vandnps(zmmA, zmmB, zmmC); + cc.evex().vandpd(xmmA, xmmB, xmmC); + cc.evex().vandpd(ymmA, ymmB, ymmC); + cc.evex().vandpd(zmmA, zmmB, zmmC); + cc.evex().vandps(xmmA, xmmB, xmmC); + cc.evex().vandps(ymmA, ymmB, ymmC); + cc.evex().vandps(zmmA, zmmB, zmmC); + cc.evex().vblendmpd(xmmA, xmmB, xmmC); + cc.evex().vblendmpd(ymmA, ymmB, ymmC); + cc.evex().vblendmpd(zmmA, zmmB, zmmC); + cc.evex().vblendmps(xmmA, xmmB, xmmC); + cc.evex().vblendmps(ymmA, ymmB, ymmC); + cc.evex().vblendmps(zmmA, zmmB, zmmC); + cc.evex().vbroadcastf32x2(ymmA, xmmB); + cc.evex().vbroadcastf32x2(zmmA, xmmB); + cc.evex().vbroadcasti32x2(xmmA, xmmB); + cc.evex().vbroadcasti32x2(ymmA, xmmB); + cc.evex().vbroadcasti32x2(zmmA, xmmB); + cc.evex().vbroadcastsd(ymmA, xmmB); + cc.evex().vbroadcastsd(zmmA, xmmB); + cc.evex().vbroadcastss(xmmA, xmmB); + cc.evex().vbroadcastss(ymmA, xmmB); + cc.evex().vbroadcastss(zmmA, xmmB); + cc.evex().vcmppd(kA, xmmB, xmmC, 0); + cc.evex().vcmppd(kA, ymmB, ymmC, 0); + cc.evex().vcmppd(kA, zmmB, zmmC, 0); + cc.evex().vcmpps(kA, xmmB, xmmC, 0); + cc.evex().vcmpps(kA, ymmB, ymmC, 0); + cc.evex().vcmpps(kA, zmmB, zmmC, 0); + cc.evex().vcmpsd(kA, xmmB, xmmC, 0); + cc.evex().vcmpss(kA, xmmB, xmmC, 0); + cc.evex().vcomisd(xmmA, xmmB); + cc.evex().vcomiss(xmmA, xmmB); + cc.evex().vcompresspd(xmmA, xmmB); + cc.evex().vcompresspd(ymmA, ymmB); + cc.evex().vcompresspd(zmmA, zmmB); + cc.evex().vcompressps(xmmA, xmmB); + cc.evex().vcompressps(ymmA, ymmB); + cc.evex().vcompressps(zmmA, zmmB); + cc.evex().vcvtdq2pd(xmmA, xmmB); + cc.evex().vcvtdq2pd(ymmA, xmmB); + cc.evex().vcvtdq2pd(zmmA, ymmB); + cc.evex().vcvtdq2ps(xmmA, xmmB); + cc.evex().vcvtdq2ps(ymmA, ymmB); + cc.evex().vcvtdq2ps(zmmA, zmmB); + cc.evex().vcvtpd2dq(xmmA, xmmB); + cc.evex().vcvtpd2dq(xmmA, ymmB); + cc.evex().vcvtpd2dq(ymmA, zmmB); + cc.evex().vcvtpd2qq(xmmA, xmmB); + cc.evex().vcvtpd2qq(ymmA, ymmB); + cc.evex().vcvtpd2qq(zmmA, zmmB); + cc.evex().vcvtpd2udq(xmmA, xmmB); + cc.evex().vcvtpd2udq(xmmA, ymmB); + cc.evex().vcvtpd2udq(ymmA, zmmB); + cc.evex().vcvtpd2uqq(xmmA, xmmB); + cc.evex().vcvtpd2uqq(ymmA, ymmB); + cc.evex().vcvtpd2uqq(zmmA, zmmB); + cc.evex().vcvtph2ps(xmmA, xmmB); + cc.evex().vcvtph2ps(ymmA, xmmB); + cc.evex().vcvtph2ps(zmmA, ymmB); + cc.evex().vcvtps2dq(xmmA, xmmB); + cc.evex().vcvtps2dq(ymmA, ymmB); + cc.evex().vcvtps2dq(zmmA, zmmB); + cc.evex().vcvtps2pd(xmmA, xmmB); + cc.evex().vcvtps2pd(ymmA, xmmB); + cc.evex().vcvtps2pd(zmmA, ymmB); + cc.evex().vcvtps2ph(xmmA, xmmB, 0); + cc.evex().vcvtps2ph(xmmA, ymmB, 0); + cc.evex().vcvtps2ph(ymmA, zmmB, 0); + cc.evex().vcvtps2qq(xmmA, xmmB); + cc.evex().vcvtps2qq(ymmA, xmmB); + cc.evex().vcvtps2qq(zmmA, ymmB); + cc.evex().vcvtps2udq(xmmA, xmmB); + cc.evex().vcvtps2udq(ymmA, ymmB); + cc.evex().vcvtps2udq(zmmA, zmmB); + cc.evex().vcvtps2uqq(xmmA, xmmB); + cc.evex().vcvtps2uqq(ymmA, xmmB); + cc.evex().vcvtps2uqq(zmmA, ymmB); + cc.evex().vcvtqq2pd(xmmA, xmmB); + cc.evex().vcvtqq2pd(ymmA, ymmB); + cc.evex().vcvtqq2pd(zmmA, zmmB); + cc.evex().vcvtqq2ps(xmmA, xmmB); + cc.evex().vcvtqq2ps(xmmA, ymmB); + cc.evex().vcvtqq2ps(ymmA, zmmB); + cc.evex().vcvtsd2si(gpd, xmmB); + if (cc.is64Bit()) cc.evex().vcvtsd2si(gpq, xmmB); + cc.evex().vcvtsd2ss(xmmA, xmmB, xmmC); + cc.evex().vcvtsd2usi(gpd, xmmB); + if (cc.is64Bit()) cc.evex().vcvtsd2usi(gpq, xmmB); + cc.evex().vcvtsi2sd(xmmA, xmmB, gpd); + if (cc.is64Bit()) cc.evex().vcvtsi2sd(xmmA, xmmB, gpq); + cc.evex().vcvtsi2ss(xmmA, xmmB, gpd); + if (cc.is64Bit()) cc.evex().vcvtsi2ss(xmmA, xmmB, gpq); + cc.evex().vcvtss2sd(xmmA, xmmB, xmmC); + cc.evex().vcvtss2si(gpd, xmmB); + if (cc.is64Bit()) cc.evex().vcvtss2si(gpq, xmmB); + cc.evex().vcvtss2usi(gpd, xmmB); + if (cc.is64Bit()) cc.evex().vcvtss2usi(gpq, xmmB); + cc.evex().vcvttpd2dq(xmmA, xmmB); + cc.evex().vcvttpd2dq(xmmA, ymmB); + cc.evex().vcvttpd2dq(ymmA, zmmB); + cc.evex().vcvttpd2qq(xmmA, xmmB); + cc.evex().vcvttpd2qq(ymmA, ymmB); + cc.evex().vcvttpd2qq(zmmA, zmmB); + cc.evex().vcvttpd2udq(xmmA, xmmB); + cc.evex().vcvttpd2udq(xmmA, ymmB); + cc.evex().vcvttpd2udq(ymmA, zmmB); + cc.evex().vcvttpd2uqq(xmmA, xmmB); + cc.evex().vcvttpd2uqq(ymmA, ymmB); + cc.evex().vcvttpd2uqq(zmmA, zmmB); + cc.evex().vcvttps2dq(xmmA, xmmB); + cc.evex().vcvttps2dq(ymmA, ymmB); + cc.evex().vcvttps2dq(zmmA, zmmB); + cc.evex().vcvttps2qq(xmmA, xmmB); + cc.evex().vcvttps2qq(ymmA, xmmB); + cc.evex().vcvttps2qq(zmmA, ymmB); + cc.evex().vcvttps2udq(xmmA, xmmB); + cc.evex().vcvttps2udq(ymmA, ymmB); + cc.evex().vcvttps2udq(zmmA, zmmB); + cc.evex().vcvttps2uqq(xmmA, xmmB); + cc.evex().vcvttps2uqq(ymmA, xmmB); + cc.evex().vcvttps2uqq(zmmA, ymmB); + cc.evex().vcvttsd2si(gpd, xmmB); + if (cc.is64Bit()) cc.evex().vcvttsd2si(gpq, xmmB); + cc.evex().vcvttsd2usi(gpd, xmmB); + if (cc.is64Bit()) cc.evex().vcvttsd2usi(gpq, xmmB); + cc.evex().vcvttss2si(gpd, xmmB); + if (cc.is64Bit()) cc.evex().vcvttss2si(gpq, xmmB); + cc.evex().vcvttss2usi(gpd, xmmB); + if (cc.is64Bit()) cc.evex().vcvttss2usi(gpq, xmmB); + cc.evex().vcvtudq2pd(xmmA, xmmB); + cc.evex().vcvtudq2pd(ymmA, xmmB); + cc.evex().vcvtudq2pd(zmmA, ymmB); + cc.evex().vcvtudq2ps(xmmA, xmmB); + cc.evex().vcvtudq2ps(ymmA, ymmB); + cc.evex().vcvtudq2ps(zmmA, zmmB); + cc.evex().vcvtuqq2pd(xmmA, xmmB); + cc.evex().vcvtuqq2pd(ymmA, ymmB); + cc.evex().vcvtuqq2pd(zmmA, zmmB); + cc.evex().vcvtuqq2ps(xmmA, xmmB); + cc.evex().vcvtuqq2ps(xmmA, ymmB); + cc.evex().vcvtuqq2ps(ymmA, zmmB); + cc.evex().vcvtusi2sd(xmmA, xmmB, gpd); + if (cc.is64Bit()) cc.evex().vcvtusi2sd(xmmA, xmmB, gpq); + cc.evex().vcvtusi2ss(xmmA, xmmB, gpd); + if (cc.is64Bit()) cc.evex().vcvtusi2ss(xmmA, xmmB, gpq); + cc.evex().vdbpsadbw(xmmA, xmmB, xmmC, 0); + cc.evex().vdbpsadbw(ymmA, ymmB, ymmC, 0); + cc.evex().vdbpsadbw(zmmA, zmmB, zmmC, 0); + cc.evex().vdivpd(xmmA, xmmB, xmmC); + cc.evex().vdivpd(ymmA, ymmB, ymmC); + cc.evex().vdivpd(zmmA, zmmB, zmmC); + cc.evex().vdivps(xmmA, xmmB, xmmC); + cc.evex().vdivps(ymmA, ymmB, ymmC); + cc.evex().vdivps(zmmA, zmmB, zmmC); + cc.evex().vdivsd(xmmA, xmmB, xmmC); + cc.evex().vdivss(xmmA, xmmB, xmmC); + cc.evex().vexp2pd(zmmA, zmmB); + cc.evex().vexp2ps(zmmA, zmmB); + cc.evex().vexpandpd(xmmA, xmmB); + cc.evex().vexpandpd(ymmA, ymmB); + cc.evex().vexpandpd(zmmA, zmmB); + cc.evex().vexpandps(xmmA, xmmB); + cc.evex().vexpandps(ymmA, ymmB); + cc.evex().vexpandps(zmmA, zmmB); + cc.evex().vextractf32x4(xmmA, ymmB, 0); + cc.evex().vextractf32x4(xmmA, zmmB, 0); + cc.evex().vextractf32x8(ymmA, zmmB, 0); + cc.evex().vextractf64x2(xmmA, ymmB, 0); + cc.evex().vextractf64x2(xmmA, zmmB, 0); + cc.evex().vextractf64x4(ymmA, zmmB, 0); + cc.evex().vextracti32x4(xmmA, ymmB, 0); + cc.evex().vextracti32x4(xmmA, zmmB, 0); + cc.evex().vextracti32x8(ymmA, zmmB, 0); + cc.evex().vextracti64x2(xmmA, ymmB, 0); + cc.evex().vextracti64x2(xmmA, zmmB, 0); + cc.evex().vextracti64x4(ymmA, zmmB, 0); + cc.evex().vextractps(gpd, xmmB, 0); + cc.evex().vfixupimmpd(xmmA, xmmB, xmmC, 0); + cc.evex().vfixupimmpd(ymmA, ymmB, ymmC, 0); + cc.evex().vfixupimmpd(zmmA, zmmB, zmmC, 0); + cc.evex().vfixupimmps(xmmA, xmmB, xmmC, 0); + cc.evex().vfixupimmps(ymmA, ymmB, ymmC, 0); + cc.evex().vfixupimmps(zmmA, zmmB, zmmC, 0); + cc.evex().vfixupimmsd(xmmA, xmmB, xmmC, 0); + cc.evex().vfixupimmss(xmmA, xmmB, xmmC, 0); + cc.evex().vfmadd132pd(xmmA, xmmB, xmmC); + cc.evex().vfmadd132pd(ymmA, ymmB, ymmC); + cc.evex().vfmadd132pd(zmmA, zmmB, zmmC); + cc.evex().vfmadd132ps(xmmA, xmmB, xmmC); + cc.evex().vfmadd132ps(ymmA, ymmB, ymmC); + cc.evex().vfmadd132ps(zmmA, zmmB, zmmC); + cc.evex().vfmadd132sd(xmmA, xmmB, xmmC); + cc.evex().vfmadd132ss(xmmA, xmmB, xmmC); + cc.evex().vfmadd213pd(xmmA, xmmB, xmmC); + cc.evex().vfmadd213pd(ymmA, ymmB, ymmC); + cc.evex().vfmadd213pd(zmmA, zmmB, zmmC); + cc.evex().vfmadd213ps(xmmA, xmmB, xmmC); + cc.evex().vfmadd213ps(ymmA, ymmB, ymmC); + cc.evex().vfmadd213ps(zmmA, zmmB, zmmC); + cc.evex().vfmadd213sd(xmmA, xmmB, xmmC); + cc.evex().vfmadd213ss(xmmA, xmmB, xmmC); + cc.evex().vfmadd231pd(xmmA, xmmB, xmmC); + cc.evex().vfmadd231pd(ymmA, ymmB, ymmC); + cc.evex().vfmadd231pd(zmmA, zmmB, zmmC); + cc.evex().vfmadd231ps(xmmA, xmmB, xmmC); + cc.evex().vfmadd231ps(ymmA, ymmB, ymmC); + cc.evex().vfmadd231ps(zmmA, zmmB, zmmC); + cc.evex().vfmadd231sd(xmmA, xmmB, xmmC); + cc.evex().vfmadd231ss(xmmA, xmmB, xmmC); + cc.evex().vfmaddsub132pd(xmmA, xmmB, xmmC); + cc.evex().vfmaddsub132pd(ymmA, ymmB, ymmC); + cc.evex().vfmaddsub132pd(zmmA, zmmB, zmmC); + cc.evex().vfmaddsub132ps(xmmA, xmmB, xmmC); + cc.evex().vfmaddsub132ps(ymmA, ymmB, ymmC); + cc.evex().vfmaddsub132ps(zmmA, zmmB, zmmC); + cc.evex().vfmaddsub213pd(xmmA, xmmB, xmmC); + cc.evex().vfmaddsub213pd(ymmA, ymmB, ymmC); + cc.evex().vfmaddsub213pd(zmmA, zmmB, zmmC); + cc.evex().vfmaddsub213ps(xmmA, xmmB, xmmC); + cc.evex().vfmaddsub213ps(ymmA, ymmB, ymmC); + cc.evex().vfmaddsub213ps(zmmA, zmmB, zmmC); + cc.evex().vfmaddsub231pd(xmmA, xmmB, xmmC); + cc.evex().vfmaddsub231pd(ymmA, ymmB, ymmC); + cc.evex().vfmaddsub231pd(zmmA, zmmB, zmmC); + cc.evex().vfmaddsub231ps(xmmA, xmmB, xmmC); + cc.evex().vfmaddsub231ps(ymmA, ymmB, ymmC); + cc.evex().vfmaddsub231ps(zmmA, zmmB, zmmC); + cc.evex().vfmsub132pd(xmmA, xmmB, xmmC); + cc.evex().vfmsub132pd(ymmA, ymmB, ymmC); + cc.evex().vfmsub132pd(zmmA, zmmB, zmmC); + cc.evex().vfmsub132ps(xmmA, xmmB, xmmC); + cc.evex().vfmsub132ps(ymmA, ymmB, ymmC); + cc.evex().vfmsub132ps(zmmA, zmmB, zmmC); + cc.evex().vfmsub132sd(xmmA, xmmB, xmmC); + cc.evex().vfmsub132ss(xmmA, xmmB, xmmC); + cc.evex().vfmsub213pd(xmmA, xmmB, xmmC); + cc.evex().vfmsub213pd(ymmA, ymmB, ymmC); + cc.evex().vfmsub213pd(zmmA, zmmB, zmmC); + cc.evex().vfmsub213ps(xmmA, xmmB, xmmC); + cc.evex().vfmsub213ps(ymmA, ymmB, ymmC); + cc.evex().vfmsub213ps(zmmA, zmmB, zmmC); + cc.evex().vfmsub213sd(xmmA, xmmB, xmmC); + cc.evex().vfmsub213ss(xmmA, xmmB, xmmC); + cc.evex().vfmsub231pd(xmmA, xmmB, xmmC); + cc.evex().vfmsub231pd(ymmA, ymmB, ymmC); + cc.evex().vfmsub231pd(zmmA, zmmB, zmmC); + cc.evex().vfmsub231ps(xmmA, xmmB, xmmC); + cc.evex().vfmsub231ps(ymmA, ymmB, ymmC); + cc.evex().vfmsub231ps(zmmA, zmmB, zmmC); + cc.evex().vfmsub231sd(xmmA, xmmB, xmmC); + cc.evex().vfmsub231ss(xmmA, xmmB, xmmC); + cc.evex().vfmsubadd132pd(xmmA, xmmB, xmmC); + cc.evex().vfmsubadd132pd(ymmA, ymmB, ymmC); + cc.evex().vfmsubadd132pd(zmmA, zmmB, zmmC); + cc.evex().vfmsubadd132ps(xmmA, xmmB, xmmC); + cc.evex().vfmsubadd132ps(ymmA, ymmB, ymmC); + cc.evex().vfmsubadd132ps(zmmA, zmmB, zmmC); + cc.evex().vfmsubadd213pd(xmmA, xmmB, xmmC); + cc.evex().vfmsubadd213pd(ymmA, ymmB, ymmC); + cc.evex().vfmsubadd213pd(zmmA, zmmB, zmmC); + cc.evex().vfmsubadd213ps(xmmA, xmmB, xmmC); + cc.evex().vfmsubadd213ps(ymmA, ymmB, ymmC); + cc.evex().vfmsubadd213ps(zmmA, zmmB, zmmC); + cc.evex().vfmsubadd231pd(xmmA, xmmB, xmmC); + cc.evex().vfmsubadd231pd(ymmA, ymmB, ymmC); + cc.evex().vfmsubadd231pd(zmmA, zmmB, zmmC); + cc.evex().vfmsubadd231ps(xmmA, xmmB, xmmC); + cc.evex().vfmsubadd231ps(ymmA, ymmB, ymmC); + cc.evex().vfmsubadd231ps(zmmA, zmmB, zmmC); + cc.evex().vfnmadd132pd(xmmA, xmmB, xmmC); + cc.evex().vfnmadd132pd(ymmA, ymmB, ymmC); + cc.evex().vfnmadd132pd(zmmA, zmmB, zmmC); + cc.evex().vfnmadd132ps(xmmA, xmmB, xmmC); + cc.evex().vfnmadd132ps(ymmA, ymmB, ymmC); + cc.evex().vfnmadd132ps(zmmA, zmmB, zmmC); + cc.evex().vfnmadd132sd(xmmA, xmmB, xmmC); + cc.evex().vfnmadd132ss(xmmA, xmmB, xmmC); + cc.evex().vfnmadd213pd(xmmA, xmmB, xmmC); + cc.evex().vfnmadd213pd(ymmA, ymmB, ymmC); + cc.evex().vfnmadd213pd(zmmA, zmmB, zmmC); + cc.evex().vfnmadd213ps(xmmA, xmmB, xmmC); + cc.evex().vfnmadd213ps(ymmA, ymmB, ymmC); + cc.evex().vfnmadd213ps(zmmA, zmmB, zmmC); + cc.evex().vfnmadd213sd(xmmA, xmmB, xmmC); + cc.evex().vfnmadd213ss(xmmA, xmmB, xmmC); + cc.evex().vfnmadd231pd(xmmA, xmmB, xmmC); + cc.evex().vfnmadd231pd(ymmA, ymmB, ymmC); + cc.evex().vfnmadd231pd(zmmA, zmmB, zmmC); + cc.evex().vfnmadd231ps(xmmA, xmmB, xmmC); + cc.evex().vfnmadd231ps(ymmA, ymmB, ymmC); + cc.evex().vfnmadd231ps(zmmA, zmmB, zmmC); + cc.evex().vfnmadd231sd(xmmA, xmmB, xmmC); + cc.evex().vfnmadd231ss(xmmA, xmmB, xmmC); + cc.evex().vfnmsub132pd(xmmA, xmmB, xmmC); + cc.evex().vfnmsub132pd(ymmA, ymmB, ymmC); + cc.evex().vfnmsub132pd(zmmA, zmmB, zmmC); + cc.evex().vfnmsub132ps(xmmA, xmmB, xmmC); + cc.evex().vfnmsub132ps(ymmA, ymmB, ymmC); + cc.evex().vfnmsub132ps(zmmA, zmmB, zmmC); + cc.evex().vfnmsub132sd(xmmA, xmmB, xmmC); + cc.evex().vfnmsub132ss(xmmA, xmmB, xmmC); + cc.evex().vfnmsub213pd(xmmA, xmmB, xmmC); + cc.evex().vfnmsub213pd(ymmA, ymmB, ymmC); + cc.evex().vfnmsub213pd(zmmA, zmmB, zmmC); + cc.evex().vfnmsub213ps(xmmA, xmmB, xmmC); + cc.evex().vfnmsub213ps(ymmA, ymmB, ymmC); + cc.evex().vfnmsub213ps(zmmA, zmmB, zmmC); + cc.evex().vfnmsub213sd(xmmA, xmmB, xmmC); + cc.evex().vfnmsub213ss(xmmA, xmmB, xmmC); + cc.evex().vfnmsub231pd(xmmA, xmmB, xmmC); + cc.evex().vfnmsub231pd(ymmA, ymmB, ymmC); + cc.evex().vfnmsub231pd(zmmA, zmmB, zmmC); + cc.evex().vfnmsub231ps(xmmA, xmmB, xmmC); + cc.evex().vfnmsub231ps(ymmA, ymmB, ymmC); + cc.evex().vfnmsub231ps(zmmA, zmmB, zmmC); + cc.evex().vfnmsub231sd(xmmA, xmmB, xmmC); + cc.evex().vfnmsub231ss(xmmA, xmmB, xmmC); + cc.evex().vfpclasspd(kA, xmmB, 0); + cc.evex().vfpclasspd(kA, ymmB, 0); + cc.evex().vfpclasspd(kA, zmmB, 0); + cc.evex().vfpclassps(kA, xmmB, 0); + cc.evex().vfpclassps(kA, ymmB, 0); + cc.evex().vfpclassps(kA, zmmB, 0); + cc.evex().vfpclasssd(kA, xmmB, 0); + cc.evex().vfpclassss(kA, xmmB, 0); + cc.evex().vgetexppd(xmmA, xmmB); + cc.evex().vgetexppd(ymmA, ymmB); + cc.evex().vgetexppd(zmmA, zmmB); + cc.evex().vgetexpps(xmmA, xmmB); + cc.evex().vgetexpps(ymmA, ymmB); + cc.evex().vgetexpps(zmmA, zmmB); + cc.evex().vgetexpsd(xmmA, xmmB, xmmC); + cc.evex().vgetexpss(xmmA, xmmB, xmmC); + cc.evex().vgetmantpd(xmmA, xmmB, 0); + cc.evex().vgetmantpd(ymmA, ymmB, 0); + cc.evex().vgetmantpd(zmmA, zmmB, 0); + cc.evex().vgetmantps(xmmA, xmmB, 0); + cc.evex().vgetmantps(ymmA, ymmB, 0); + cc.evex().vgetmantps(zmmA, zmmB, 0); + cc.evex().vgetmantsd(xmmA, xmmB, xmmC, 0); + cc.evex().vgetmantss(xmmA, xmmB, xmmC, 0); + cc.evex().vinsertf32x4(ymmA, ymmB, xmmC, 0); + cc.evex().vinsertf32x4(zmmA, zmmB, xmmC, 0); + cc.evex().vinsertf32x8(zmmA, zmmB, ymmC, 0); + cc.evex().vinsertf64x2(ymmA, ymmB, xmmC, 0); + cc.evex().vinsertf64x2(zmmA, zmmB, xmmC, 0); + cc.evex().vinsertf64x4(zmmA, zmmB, ymmC, 0); + cc.evex().vinserti32x4(ymmA, ymmB, xmmC, 0); + cc.evex().vinserti32x4(zmmA, zmmB, xmmC, 0); + cc.evex().vinserti32x8(zmmA, zmmB, ymmC, 0); + cc.evex().vinserti64x2(ymmA, ymmB, xmmC, 0); + cc.evex().vinserti64x2(zmmA, zmmB, xmmC, 0); + cc.evex().vinserti64x4(zmmA, zmmB, ymmC, 0); + cc.evex().vinsertps(xmmA, xmmB, xmmC, 0); + cc.evex().vmaxpd(xmmA, xmmB, xmmC); + cc.evex().vmaxpd(ymmA, ymmB, ymmC); + cc.evex().vmaxpd(zmmA, zmmB, zmmC); + cc.evex().vmaxps(xmmA, xmmB, xmmC); + cc.evex().vmaxps(ymmA, ymmB, ymmC); + cc.evex().vmaxps(zmmA, zmmB, zmmC); + cc.evex().vmaxsd(xmmA, xmmB, xmmC); + cc.evex().vmaxss(xmmA, xmmB, xmmC); + cc.evex().vminpd(xmmA, xmmB, xmmC); + cc.evex().vminpd(ymmA, ymmB, ymmC); + cc.evex().vminpd(zmmA, zmmB, zmmC); + cc.evex().vminps(xmmA, xmmB, xmmC); + cc.evex().vminps(ymmA, ymmB, ymmC); + cc.evex().vminps(zmmA, zmmB, zmmC); + cc.evex().vminsd(xmmA, xmmB, xmmC); + cc.evex().vminss(xmmA, xmmB, xmmC); + cc.evex().vmovapd(xmmA, xmmB); + cc.evex().vmovapd(xmmA, xmmB); + cc.evex().vmovapd(ymmA, ymmB); + cc.evex().vmovapd(ymmA, ymmB); + cc.evex().vmovapd(zmmA, zmmB); + cc.evex().vmovapd(zmmA, zmmB); + cc.evex().vmovaps(xmmA, xmmB); + cc.evex().vmovaps(xmmA, xmmB); + cc.evex().vmovaps(ymmA, ymmB); + cc.evex().vmovaps(ymmA, ymmB); + cc.evex().vmovaps(zmmA, zmmB); + cc.evex().vmovaps(zmmA, zmmB); + cc.evex().vmovd(gpd, xmmB); + cc.evex().vmovd(xmmA, gpd); + cc.evex().vmovddup(xmmA, xmmB); + cc.evex().vmovddup(ymmA, ymmB); + cc.evex().vmovddup(zmmA, zmmB); + cc.evex().vmovdqa32(xmmA, xmmB); + cc.evex().vmovdqa32(xmmA, xmmB); + cc.evex().vmovdqa32(ymmA, ymmB); + cc.evex().vmovdqa32(ymmA, ymmB); + cc.evex().vmovdqa32(zmmA, zmmB); + cc.evex().vmovdqa32(zmmA, zmmB); + cc.evex().vmovdqa64(xmmA, xmmB); + cc.evex().vmovdqa64(xmmA, xmmB); + cc.evex().vmovdqa64(ymmA, ymmB); + cc.evex().vmovdqa64(ymmA, ymmB); + cc.evex().vmovdqa64(zmmA, zmmB); + cc.evex().vmovdqa64(zmmA, zmmB); + cc.evex().vmovdqu16(xmmA, xmmB); + cc.evex().vmovdqu16(xmmA, xmmB); + cc.evex().vmovdqu16(ymmA, ymmB); + cc.evex().vmovdqu16(ymmA, ymmB); + cc.evex().vmovdqu16(zmmA, zmmB); + cc.evex().vmovdqu16(zmmA, zmmB); + cc.evex().vmovdqu32(xmmA, xmmB); + cc.evex().vmovdqu32(xmmA, xmmB); + cc.evex().vmovdqu32(ymmA, ymmB); + cc.evex().vmovdqu32(ymmA, ymmB); + cc.evex().vmovdqu32(zmmA, zmmB); + cc.evex().vmovdqu32(zmmA, zmmB); + cc.evex().vmovdqu64(xmmA, xmmB); + cc.evex().vmovdqu64(xmmA, xmmB); + cc.evex().vmovdqu64(ymmA, ymmB); + cc.evex().vmovdqu64(ymmA, ymmB); + cc.evex().vmovdqu64(zmmA, zmmB); + cc.evex().vmovdqu64(zmmA, zmmB); + cc.evex().vmovdqu8(xmmA, xmmB); + cc.evex().vmovdqu8(xmmA, xmmB); + cc.evex().vmovdqu8(ymmA, ymmB); + cc.evex().vmovdqu8(ymmA, ymmB); + cc.evex().vmovdqu8(zmmA, zmmB); + cc.evex().vmovdqu8(zmmA, zmmB); + cc.evex().vmovhlps(xmmA, xmmB, xmmC); + if (cc.is64Bit()) cc.evex().vmovq(gpq, xmmB); + if (cc.is64Bit()) cc.evex().vmovq(xmmA, gpq); + cc.evex().vmovq(xmmA, xmmB); + cc.evex().vmovsd(xmmA, xmmB, xmmC); + cc.evex().vmovshdup(xmmA, xmmB); + cc.evex().vmovshdup(ymmA, ymmB); + cc.evex().vmovshdup(zmmA, zmmB); + cc.evex().vmovsldup(xmmA, xmmB); + cc.evex().vmovsldup(ymmA, ymmB); + cc.evex().vmovsldup(zmmA, zmmB); + cc.evex().vmovss(xmmA, xmmB, xmmC); + cc.evex().vmovupd(xmmA, xmmB); + cc.evex().vmovupd(xmmA, xmmB); + cc.evex().vmovupd(ymmA, ymmB); + cc.evex().vmovupd(ymmA, ymmB); + cc.evex().vmovupd(zmmA, zmmB); + cc.evex().vmovupd(zmmA, zmmB); + cc.evex().vmovups(xmmA, xmmB); + cc.evex().vmovups(xmmA, xmmB); + cc.evex().vmovups(ymmA, ymmB); + cc.evex().vmovups(ymmA, ymmB); + cc.evex().vmovups(zmmA, zmmB); + cc.evex().vmovups(zmmA, zmmB); + cc.evex().vmulpd(xmmA, xmmB, xmmC); + cc.evex().vmulpd(ymmA, ymmB, ymmC); + cc.evex().vmulpd(zmmA, zmmB, zmmC); + cc.evex().vmulps(xmmA, xmmB, xmmC); + cc.evex().vmulps(ymmA, ymmB, ymmC); + cc.evex().vmulps(zmmA, zmmB, zmmC); + cc.evex().vmulsd(xmmA, xmmB, xmmC); + cc.evex().vmulss(xmmA, xmmB, xmmC); + cc.evex().vorpd(xmmA, xmmB, xmmC); + cc.evex().vorpd(ymmA, ymmB, ymmC); + cc.evex().vorpd(zmmA, zmmB, zmmC); + cc.evex().vorps(xmmA, xmmB, xmmC); + cc.evex().vorps(ymmA, ymmB, ymmC); + cc.evex().vorps(zmmA, zmmB, zmmC); + cc.evex().vpabsb(xmmA, xmmB); + cc.evex().vpabsb(ymmA, ymmB); + cc.evex().vpabsb(zmmA, zmmB); + cc.evex().vpabsd(xmmA, xmmB); + cc.evex().vpabsd(ymmA, ymmB); + cc.evex().vpabsd(zmmA, zmmB); + cc.evex().vpabsq(xmmA, xmmB); + cc.evex().vpabsq(ymmA, ymmB); + cc.evex().vpabsq(zmmA, zmmB); + cc.evex().vpabsw(xmmA, xmmB); + cc.evex().vpabsw(ymmA, ymmB); + cc.evex().vpabsw(zmmA, zmmB); + cc.evex().vpackssdw(xmmA, xmmB, xmmC); + cc.evex().vpackssdw(ymmA, ymmB, ymmC); + cc.evex().vpackssdw(zmmA, zmmB, zmmC); + cc.evex().vpacksswb(xmmA, xmmB, xmmC); + cc.evex().vpacksswb(ymmA, ymmB, ymmC); + cc.evex().vpacksswb(zmmA, zmmB, zmmC); + cc.evex().vpackusdw(xmmA, xmmB, xmmC); + cc.evex().vpackusdw(ymmA, ymmB, ymmC); + cc.evex().vpackusdw(zmmA, zmmB, zmmC); + cc.evex().vpackuswb(xmmA, xmmB, xmmC); + cc.evex().vpackuswb(ymmA, ymmB, ymmC); + cc.evex().vpackuswb(zmmA, zmmB, zmmC); + cc.evex().vpaddb(xmmA, xmmB, xmmC); + cc.evex().vpaddb(ymmA, ymmB, ymmC); + cc.evex().vpaddb(zmmA, zmmB, zmmC); + cc.evex().vpaddd(xmmA, xmmB, xmmC); + cc.evex().vpaddd(ymmA, ymmB, ymmC); + cc.evex().vpaddd(zmmA, zmmB, zmmC); + cc.evex().vpaddq(xmmA, xmmB, xmmC); + cc.evex().vpaddq(ymmA, ymmB, ymmC); + cc.evex().vpaddq(zmmA, zmmB, zmmC); + cc.evex().vpaddsb(xmmA, xmmB, xmmC); + cc.evex().vpaddsb(ymmA, ymmB, ymmC); + cc.evex().vpaddsb(zmmA, zmmB, zmmC); + cc.evex().vpaddsw(xmmA, xmmB, xmmC); + cc.evex().vpaddsw(ymmA, ymmB, ymmC); + cc.evex().vpaddsw(zmmA, zmmB, zmmC); + cc.evex().vpaddusb(xmmA, xmmB, xmmC); + cc.evex().vpaddusb(ymmA, ymmB, ymmC); + cc.evex().vpaddusb(zmmA, zmmB, zmmC); + cc.evex().vpaddusw(xmmA, xmmB, xmmC); + cc.evex().vpaddusw(ymmA, ymmB, ymmC); + cc.evex().vpaddusw(zmmA, zmmB, zmmC); + cc.evex().vpaddw(xmmA, xmmB, xmmC); + cc.evex().vpaddw(ymmA, ymmB, ymmC); + cc.evex().vpaddw(zmmA, zmmB, zmmC); + cc.evex().vpalignr(xmmA, xmmB, xmmC, 0); + cc.evex().vpalignr(ymmA, ymmB, ymmC, 0); + cc.evex().vpalignr(zmmA, zmmB, zmmC, 0); + cc.evex().vpandd(xmmA, xmmB, xmmC); + cc.evex().vpandd(ymmA, ymmB, ymmC); + cc.evex().vpandd(zmmA, zmmB, zmmC); + cc.evex().vpandnd(xmmA, xmmB, xmmC); + cc.evex().vpandnd(ymmA, ymmB, ymmC); + cc.evex().vpandnd(zmmA, zmmB, zmmC); + cc.evex().vpandnq(xmmA, xmmB, xmmC); + cc.evex().vpandnq(ymmA, ymmB, ymmC); + cc.evex().vpandnq(zmmA, zmmB, zmmC); + cc.evex().vpandq(xmmA, xmmB, xmmC); + cc.evex().vpandq(ymmA, ymmB, ymmC); + cc.evex().vpandq(zmmA, zmmB, zmmC); + cc.evex().vpavgb(xmmA, xmmB, xmmC); + cc.evex().vpavgb(ymmA, ymmB, ymmC); + cc.evex().vpavgb(zmmA, zmmB, zmmC); + cc.evex().vpavgw(xmmA, xmmB, xmmC); + cc.evex().vpavgw(ymmA, ymmB, ymmC); + cc.evex().vpavgw(zmmA, zmmB, zmmC); + cc.evex().vpblendmb(xmmA, xmmB, xmmC); + cc.evex().vpblendmb(ymmA, ymmB, ymmC); + cc.evex().vpblendmb(zmmA, zmmB, zmmC); + cc.evex().vpblendmd(xmmA, xmmB, xmmC); + cc.evex().vpblendmd(ymmA, ymmB, ymmC); + cc.evex().vpblendmd(zmmA, zmmB, zmmC); + cc.evex().vpblendmq(xmmA, xmmB, xmmC); + cc.evex().vpblendmq(ymmA, ymmB, ymmC); + cc.evex().vpblendmq(zmmA, zmmB, zmmC); + cc.evex().vpblendmw(xmmA, xmmB, xmmC); + cc.evex().vpblendmw(ymmA, ymmB, ymmC); + cc.evex().vpblendmw(zmmA, zmmB, zmmC); + cc.evex().vpbroadcastb(xmmA, gpd); + cc.evex().vpbroadcastb(xmmA, xmmB); + cc.evex().vpbroadcastb(ymmA, gpd); + cc.evex().vpbroadcastb(ymmA, xmmB); + cc.evex().vpbroadcastb(zmmA, gpd); + cc.evex().vpbroadcastb(zmmA, xmmB); + cc.evex().vpbroadcastd(xmmA, gpd); + cc.evex().vpbroadcastd(xmmA, xmmB); + cc.evex().vpbroadcastd(ymmA, gpd); + cc.evex().vpbroadcastd(ymmA, xmmB); + cc.evex().vpbroadcastd(zmmA, gpd); + cc.evex().vpbroadcastd(zmmA, xmmB); + cc.evex().vpbroadcastmb2q(xmmA, kB); + cc.evex().vpbroadcastmb2q(ymmA, kB); + cc.evex().vpbroadcastmb2q(zmmA, kB); + cc.evex().vpbroadcastmw2d(xmmA, kB); + cc.evex().vpbroadcastmw2d(ymmA, kB); + cc.evex().vpbroadcastmw2d(zmmA, kB); + if (cc.is64Bit()) cc.evex().vpbroadcastq(xmmA, gpq); + cc.evex().vpbroadcastq(xmmA, xmmB); + if (cc.is64Bit()) cc.evex().vpbroadcastq(ymmA, gpq); + cc.evex().vpbroadcastq(ymmA, xmmB); + if (cc.is64Bit()) cc.evex().vpbroadcastq(zmmA, gpq); + cc.evex().vpbroadcastq(zmmA, xmmB); + cc.evex().vpbroadcastw(xmmA, gpd); + cc.evex().vpbroadcastw(xmmA, xmmB); + cc.evex().vpbroadcastw(ymmA, gpd); + cc.evex().vpbroadcastw(ymmA, xmmB); + cc.evex().vpbroadcastw(zmmA, gpd); + cc.evex().vpbroadcastw(zmmA, xmmB); + cc.evex().vpcmpb(kA, xmmB, xmmC, 0); + cc.evex().vpcmpb(kA, ymmB, ymmC, 0); + cc.evex().vpcmpb(kA, zmmB, zmmC, 0); + cc.evex().vpcmpd(kA, xmmB, xmmC, 0); + cc.evex().vpcmpd(kA, ymmB, ymmC, 0); + cc.evex().vpcmpd(kA, zmmB, zmmC, 0); + cc.evex().vpcmpeqb(kA, xmmB, xmmC); + cc.evex().vpcmpeqb(kA, ymmB, ymmC); + cc.evex().vpcmpeqb(kA, zmmB, zmmC); + cc.evex().vpcmpeqd(kA, xmmB, xmmC); + cc.evex().vpcmpeqd(kA, ymmB, ymmC); + cc.evex().vpcmpeqd(kA, zmmB, zmmC); + cc.evex().vpcmpeqq(kA, xmmB, xmmC); + cc.evex().vpcmpeqq(kA, ymmB, ymmC); + cc.evex().vpcmpeqq(kA, zmmB, zmmC); + cc.evex().vpcmpeqw(kA, xmmB, xmmC); + cc.evex().vpcmpeqw(kA, ymmB, ymmC); + cc.evex().vpcmpeqw(kA, zmmB, zmmC); + cc.evex().vpcmpgtb(kA, xmmB, xmmC); + cc.evex().vpcmpgtb(kA, ymmB, ymmC); + cc.evex().vpcmpgtb(kA, zmmB, zmmC); + cc.evex().vpcmpgtd(kA, xmmB, xmmC); + cc.evex().vpcmpgtd(kA, ymmB, ymmC); + cc.evex().vpcmpgtd(kA, zmmB, zmmC); + cc.evex().vpcmpgtq(kA, xmmB, xmmC); + cc.evex().vpcmpgtq(kA, ymmB, ymmC); + cc.evex().vpcmpgtq(kA, zmmB, zmmC); + cc.evex().vpcmpgtw(kA, xmmB, xmmC); + cc.evex().vpcmpgtw(kA, ymmB, ymmC); + cc.evex().vpcmpgtw(kA, zmmB, zmmC); + cc.evex().vpcmpq(kA, xmmB, xmmC, 0); + cc.evex().vpcmpq(kA, ymmB, ymmC, 0); + cc.evex().vpcmpq(kA, zmmB, zmmC, 0); + cc.evex().vpcmpub(kA, xmmB, xmmC, 0); + cc.evex().vpcmpub(kA, ymmB, ymmC, 0); + cc.evex().vpcmpub(kA, zmmB, zmmC, 0); + cc.evex().vpcmpud(kA, xmmB, xmmC, 0); + cc.evex().vpcmpud(kA, ymmB, ymmC, 0); + cc.evex().vpcmpud(kA, zmmB, zmmC, 0); + cc.evex().vpcmpuq(kA, xmmB, xmmC, 0); + cc.evex().vpcmpuq(kA, ymmB, ymmC, 0); + cc.evex().vpcmpuq(kA, zmmB, zmmC, 0); + cc.evex().vpcmpuw(kA, xmmB, xmmC, 0); + cc.evex().vpcmpuw(kA, ymmB, ymmC, 0); + cc.evex().vpcmpuw(kA, zmmB, zmmC, 0); + cc.evex().vpcmpw(kA, xmmB, xmmC, 0); + cc.evex().vpcmpw(kA, ymmB, ymmC, 0); + cc.evex().vpcmpw(kA, zmmB, zmmC, 0); + cc.evex().vpcompressd(xmmA, xmmB); + cc.evex().vpcompressd(ymmA, ymmB); + cc.evex().vpcompressd(zmmA, zmmB); + cc.evex().vpcompressq(xmmA, xmmB); + cc.evex().vpcompressq(ymmA, ymmB); + cc.evex().vpcompressq(zmmA, zmmB); + cc.evex().vpconflictd(xmmA, xmmB); + cc.evex().vpconflictd(ymmA, ymmB); + cc.evex().vpconflictd(zmmA, zmmB); + cc.evex().vpconflictq(xmmA, xmmB); + cc.evex().vpconflictq(ymmA, ymmB); + cc.evex().vpconflictq(zmmA, zmmB); + cc.evex().vpermb(xmmA, xmmB, xmmC); + cc.evex().vpermb(ymmA, ymmB, ymmC); + cc.evex().vpermb(zmmA, zmmB, zmmC); + cc.evex().vpermd(ymmA, ymmB, ymmC); + cc.evex().vpermd(zmmA, zmmB, zmmC); + cc.evex().vpermi2b(xmmA, xmmB, xmmC); + cc.evex().vpermi2b(ymmA, ymmB, ymmC); + cc.evex().vpermi2b(zmmA, zmmB, zmmC); + cc.evex().vpermi2d(xmmA, xmmB, xmmC); + cc.evex().vpermi2d(ymmA, ymmB, ymmC); + cc.evex().vpermi2d(zmmA, zmmB, zmmC); + cc.evex().vpermi2pd(xmmA, xmmB, xmmC); + cc.evex().vpermi2pd(ymmA, ymmB, ymmC); + cc.evex().vpermi2pd(zmmA, zmmB, zmmC); + cc.evex().vpermi2ps(xmmA, xmmB, xmmC); + cc.evex().vpermi2ps(ymmA, ymmB, ymmC); + cc.evex().vpermi2ps(zmmA, zmmB, zmmC); + cc.evex().vpermi2q(xmmA, xmmB, xmmC); + cc.evex().vpermi2q(ymmA, ymmB, ymmC); + cc.evex().vpermi2q(zmmA, zmmB, zmmC); + cc.evex().vpermi2w(xmmA, xmmB, xmmC); + cc.evex().vpermi2w(ymmA, ymmB, ymmC); + cc.evex().vpermi2w(zmmA, zmmB, zmmC); + cc.evex().vpermilpd(xmmA, xmmB, xmmC); + cc.evex().vpermilpd(ymmA, ymmB, ymmC); + cc.evex().vpermilpd(zmmA, zmmB, zmmC); + cc.evex().vpermilpd(xmmA, xmmB, 0); + cc.evex().vpermilpd(ymmA, ymmB, 0); + cc.evex().vpermilpd(zmmA, zmmB, 0); + cc.evex().vpermilps(xmmA, xmmB, xmmC); + cc.evex().vpermilps(ymmA, ymmB, ymmC); + cc.evex().vpermilps(zmmA, zmmB, zmmC); + cc.evex().vpermilps(xmmA, xmmB, 0); + cc.evex().vpermilps(ymmA, ymmB, 0); + cc.evex().vpermilps(zmmA, zmmB, 0); + cc.evex().vpermq(ymmA, ymmB, ymmC); + cc.evex().vpermq(zmmA, zmmB, zmmC); + cc.evex().vpermq(ymmA, ymmB, 0); + cc.evex().vpermq(zmmA, zmmB, 0); + cc.evex().vpermt2b(xmmA, xmmB, xmmC); + cc.evex().vpermt2b(ymmA, ymmB, ymmC); + cc.evex().vpermt2b(zmmA, zmmB, zmmC); + cc.evex().vpermt2d(xmmA, xmmB, xmmC); + cc.evex().vpermt2d(ymmA, ymmB, ymmC); + cc.evex().vpermt2d(zmmA, zmmB, zmmC); + cc.evex().vpermt2pd(xmmA, xmmB, xmmC); + cc.evex().vpermt2pd(ymmA, ymmB, ymmC); + cc.evex().vpermt2pd(zmmA, zmmB, zmmC); + cc.evex().vpermt2ps(xmmA, xmmB, xmmC); + cc.evex().vpermt2ps(ymmA, ymmB, ymmC); + cc.evex().vpermt2ps(zmmA, zmmB, zmmC); + cc.evex().vpermt2q(xmmA, xmmB, xmmC); + cc.evex().vpermt2q(ymmA, ymmB, ymmC); + cc.evex().vpermt2q(zmmA, zmmB, zmmC); + cc.evex().vpermt2w(xmmA, xmmB, xmmC); + cc.evex().vpermt2w(ymmA, ymmB, ymmC); + cc.evex().vpermt2w(zmmA, zmmB, zmmC); + cc.evex().vpermw(xmmA, xmmB, xmmC); + cc.evex().vpermw(ymmA, ymmB, ymmC); + cc.evex().vpermw(zmmA, zmmB, zmmC); + cc.evex().vpexpandd(xmmA, xmmB); + cc.evex().vpexpandd(ymmA, ymmB); + cc.evex().vpexpandd(zmmA, zmmB); + cc.evex().vpexpandq(xmmA, xmmB); + cc.evex().vpexpandq(ymmA, ymmB); + cc.evex().vpexpandq(zmmA, zmmB); + cc.evex().vpextrb(gpd, xmmB, 0); + cc.evex().vpextrd(gpd, xmmB, 0); + if (cc.is64Bit()) cc.evex().vpextrq(gpq, xmmB, 0); + cc.evex().vpextrw(gpd, xmmB, 0); + cc.evex().vpinsrb(xmmA, xmmB, gpd, 0); + cc.evex().vpinsrd(xmmA, xmmB, gpd, 0); + if (cc.is64Bit()) cc.evex().vpinsrq(xmmA, xmmB, gpq, 0); + cc.evex().vpinsrw(xmmA, xmmB, gpd, 0); + cc.evex().vplzcntd(xmmA, xmmB); + cc.evex().vplzcntd(ymmA, ymmB); + cc.evex().vplzcntd(zmmA, zmmB); + cc.evex().vplzcntq(xmmA, xmmB); + cc.evex().vplzcntq(ymmA, ymmB); + cc.evex().vplzcntq(zmmA, zmmB); + cc.evex().vpmadd52huq(xmmA, xmmB, xmmC); + cc.evex().vpmadd52huq(ymmA, ymmB, ymmC); + cc.evex().vpmadd52huq(zmmA, zmmB, zmmC); + cc.evex().vpmadd52luq(xmmA, xmmB, xmmC); + cc.evex().vpmadd52luq(ymmA, ymmB, ymmC); + cc.evex().vpmadd52luq(zmmA, zmmB, zmmC); + cc.evex().vpmaddubsw(xmmA, xmmB, xmmC); + cc.evex().vpmaddubsw(ymmA, ymmB, ymmC); + cc.evex().vpmaddubsw(zmmA, zmmB, zmmC); + cc.evex().vpmaddwd(xmmA, xmmB, xmmC); + cc.evex().vpmaddwd(ymmA, ymmB, ymmC); + cc.evex().vpmaddwd(zmmA, zmmB, zmmC); + cc.evex().vpmaxsb(xmmA, xmmB, xmmC); + cc.evex().vpmaxsb(ymmA, ymmB, ymmC); + cc.evex().vpmaxsb(zmmA, zmmB, zmmC); + cc.evex().vpmaxsd(xmmA, xmmB, xmmC); + cc.evex().vpmaxsd(ymmA, ymmB, ymmC); + cc.evex().vpmaxsd(zmmA, zmmB, zmmC); + cc.evex().vpmaxsq(xmmA, xmmB, xmmC); + cc.evex().vpmaxsq(ymmA, ymmB, ymmC); + cc.evex().vpmaxsq(zmmA, zmmB, zmmC); + cc.evex().vpmaxsw(xmmA, xmmB, xmmC); + cc.evex().vpmaxsw(ymmA, ymmB, ymmC); + cc.evex().vpmaxsw(zmmA, zmmB, zmmC); + cc.evex().vpmaxub(xmmA, xmmB, xmmC); + cc.evex().vpmaxub(ymmA, ymmB, ymmC); + cc.evex().vpmaxub(zmmA, zmmB, zmmC); + cc.evex().vpmaxud(xmmA, xmmB, xmmC); + cc.evex().vpmaxud(ymmA, ymmB, ymmC); + cc.evex().vpmaxud(zmmA, zmmB, zmmC); + cc.evex().vpmaxuq(xmmA, xmmB, xmmC); + cc.evex().vpmaxuq(ymmA, ymmB, ymmC); + cc.evex().vpmaxuq(zmmA, zmmB, zmmC); + cc.evex().vpmaxuw(xmmA, xmmB, xmmC); + cc.evex().vpmaxuw(ymmA, ymmB, ymmC); + cc.evex().vpmaxuw(zmmA, zmmB, zmmC); + cc.evex().vpminsb(xmmA, xmmB, xmmC); + cc.evex().vpminsb(ymmA, ymmB, ymmC); + cc.evex().vpminsb(zmmA, zmmB, zmmC); + cc.evex().vpminsd(xmmA, xmmB, xmmC); + cc.evex().vpminsd(ymmA, ymmB, ymmC); + cc.evex().vpminsd(zmmA, zmmB, zmmC); + cc.evex().vpminsq(xmmA, xmmB, xmmC); + cc.evex().vpminsq(ymmA, ymmB, ymmC); + cc.evex().vpminsq(zmmA, zmmB, zmmC); + cc.evex().vpminsw(xmmA, xmmB, xmmC); + cc.evex().vpminsw(ymmA, ymmB, ymmC); + cc.evex().vpminsw(zmmA, zmmB, zmmC); + cc.evex().vpminub(xmmA, xmmB, xmmC); + cc.evex().vpminub(ymmA, ymmB, ymmC); + cc.evex().vpminub(zmmA, zmmB, zmmC); + cc.evex().vpminud(xmmA, xmmB, xmmC); + cc.evex().vpminud(ymmA, ymmB, ymmC); + cc.evex().vpminud(zmmA, zmmB, zmmC); + cc.evex().vpminuq(xmmA, xmmB, xmmC); + cc.evex().vpminuq(ymmA, ymmB, ymmC); + cc.evex().vpminuq(zmmA, zmmB, zmmC); + cc.evex().vpminuw(xmmA, xmmB, xmmC); + cc.evex().vpminuw(ymmA, ymmB, ymmC); + cc.evex().vpminuw(zmmA, zmmB, zmmC); + cc.evex().vpmovb2m(kA, xmmB); + cc.evex().vpmovb2m(kA, ymmB); + cc.evex().vpmovb2m(kA, zmmB); + cc.evex().vpmovd2m(kA, xmmB); + cc.evex().vpmovd2m(kA, ymmB); + cc.evex().vpmovd2m(kA, zmmB); + cc.evex().vpmovdb(xmmA, xmmB); + cc.evex().vpmovdb(xmmA, ymmB); + cc.evex().vpmovdb(xmmA, zmmB); + cc.evex().vpmovdw(xmmA, xmmB); + cc.evex().vpmovdw(xmmA, ymmB); + cc.evex().vpmovdw(ymmA, zmmB); + cc.evex().vpmovm2b(xmmA, kB); + cc.evex().vpmovm2b(ymmA, kB); + cc.evex().vpmovm2b(zmmA, kB); + cc.evex().vpmovm2d(xmmA, kB); + cc.evex().vpmovm2d(ymmA, kB); + cc.evex().vpmovm2d(zmmA, kB); + cc.evex().vpmovm2q(xmmA, kB); + cc.evex().vpmovm2q(ymmA, kB); + cc.evex().vpmovm2q(zmmA, kB); + cc.evex().vpmovm2w(xmmA, kB); + cc.evex().vpmovm2w(ymmA, kB); + cc.evex().vpmovm2w(zmmA, kB); + cc.evex().vpmovq2m(kA, xmmB); + cc.evex().vpmovq2m(kA, ymmB); + cc.evex().vpmovq2m(kA, zmmB); + cc.evex().vpmovqb(xmmA, xmmB); + cc.evex().vpmovqb(xmmA, ymmB); + cc.evex().vpmovqb(xmmA, zmmB); + cc.evex().vpmovqd(xmmA, xmmB); + cc.evex().vpmovqd(xmmA, ymmB); + cc.evex().vpmovqd(ymmA, zmmB); + cc.evex().vpmovqw(xmmA, xmmB); + cc.evex().vpmovqw(xmmA, ymmB); + cc.evex().vpmovqw(xmmA, zmmB); + cc.evex().vpmovsdb(xmmA, xmmB); + cc.evex().vpmovsdb(xmmA, ymmB); + cc.evex().vpmovsdb(xmmA, zmmB); + cc.evex().vpmovsdw(xmmA, xmmB); + cc.evex().vpmovsdw(xmmA, ymmB); + cc.evex().vpmovsdw(ymmA, zmmB); + cc.evex().vpmovsqb(xmmA, xmmB); + cc.evex().vpmovsqb(xmmA, ymmB); + cc.evex().vpmovsqb(xmmA, zmmB); + cc.evex().vpmovsqd(xmmA, xmmB); + cc.evex().vpmovsqd(xmmA, ymmB); + cc.evex().vpmovsqd(ymmA, zmmB); + cc.evex().vpmovsqw(xmmA, xmmB); + cc.evex().vpmovsqw(xmmA, ymmB); + cc.evex().vpmovsqw(xmmA, zmmB); + cc.evex().vpmovswb(xmmA, xmmB); + cc.evex().vpmovswb(xmmA, ymmB); + cc.evex().vpmovswb(ymmA, zmmB); + cc.evex().vpmovsxbd(xmmA, xmmB); + cc.evex().vpmovsxbd(ymmA, xmmB); + cc.evex().vpmovsxbd(zmmA, xmmB); + cc.evex().vpmovsxbq(xmmA, xmmB); + cc.evex().vpmovsxbq(ymmA, xmmB); + cc.evex().vpmovsxbq(zmmA, xmmB); + cc.evex().vpmovsxbw(xmmA, xmmB); + cc.evex().vpmovsxbw(ymmA, xmmB); + cc.evex().vpmovsxbw(zmmA, ymmB); + cc.evex().vpmovsxdq(xmmA, xmmB); + cc.evex().vpmovsxdq(ymmA, xmmB); + cc.evex().vpmovsxdq(zmmA, ymmB); + cc.evex().vpmovsxwd(xmmA, xmmB); + cc.evex().vpmovsxwd(ymmA, xmmB); + cc.evex().vpmovsxwd(zmmA, ymmB); + cc.evex().vpmovsxwq(xmmA, xmmB); + cc.evex().vpmovsxwq(ymmA, xmmB); + cc.evex().vpmovsxwq(zmmA, xmmB); + cc.evex().vpmovusdb(xmmA, xmmB); + cc.evex().vpmovusdb(xmmA, ymmB); + cc.evex().vpmovusdb(xmmA, zmmB); + cc.evex().vpmovusdw(xmmA, xmmB); + cc.evex().vpmovusdw(xmmA, ymmB); + cc.evex().vpmovusdw(ymmA, zmmB); + cc.evex().vpmovusqb(xmmA, xmmB); + cc.evex().vpmovusqb(xmmA, ymmB); + cc.evex().vpmovusqb(xmmA, zmmB); + cc.evex().vpmovusqd(xmmA, xmmB); + cc.evex().vpmovusqd(xmmA, ymmB); + cc.evex().vpmovusqd(ymmA, zmmB); + cc.evex().vpmovusqw(xmmA, xmmB); + cc.evex().vpmovusqw(xmmA, ymmB); + cc.evex().vpmovusqw(xmmA, zmmB); + cc.evex().vpmovuswb(xmmA, xmmB); + cc.evex().vpmovuswb(xmmA, ymmB); + cc.evex().vpmovuswb(ymmA, zmmB); + cc.evex().vpmovw2m(kA, xmmB); + cc.evex().vpmovw2m(kA, ymmB); + cc.evex().vpmovw2m(kA, zmmB); + cc.evex().vpmovwb(xmmA, xmmB); + cc.evex().vpmovwb(xmmA, ymmB); + cc.evex().vpmovwb(ymmA, zmmB); + cc.evex().vpmovzxbd(xmmA, xmmB); + cc.evex().vpmovzxbd(ymmA, xmmB); + cc.evex().vpmovzxbd(zmmA, xmmB); + cc.evex().vpmovzxbq(xmmA, xmmB); + cc.evex().vpmovzxbq(ymmA, xmmB); + cc.evex().vpmovzxbq(zmmA, xmmB); + cc.evex().vpmovzxbw(xmmA, xmmB); + cc.evex().vpmovzxbw(ymmA, xmmB); + cc.evex().vpmovzxbw(zmmA, ymmB); + cc.evex().vpmovzxdq(xmmA, xmmB); + cc.evex().vpmovzxdq(ymmA, xmmB); + cc.evex().vpmovzxdq(zmmA, ymmB); + cc.evex().vpmovzxwd(xmmA, xmmB); + cc.evex().vpmovzxwd(ymmA, xmmB); + cc.evex().vpmovzxwd(zmmA, ymmB); + cc.evex().vpmovzxwq(xmmA, xmmB); + cc.evex().vpmovzxwq(ymmA, xmmB); + cc.evex().vpmovzxwq(zmmA, xmmB); + cc.evex().vpmuldq(xmmA, xmmB, xmmC); + cc.evex().vpmuldq(ymmA, ymmB, ymmC); + cc.evex().vpmuldq(zmmA, zmmB, zmmC); + cc.evex().vpmulhrsw(xmmA, xmmB, xmmC); + cc.evex().vpmulhrsw(ymmA, ymmB, ymmC); + cc.evex().vpmulhrsw(zmmA, zmmB, zmmC); + cc.evex().vpmulhuw(xmmA, xmmB, xmmC); + cc.evex().vpmulhuw(ymmA, ymmB, ymmC); + cc.evex().vpmulhuw(zmmA, zmmB, zmmC); + cc.evex().vpmulhw(xmmA, xmmB, xmmC); + cc.evex().vpmulhw(ymmA, ymmB, ymmC); + cc.evex().vpmulhw(zmmA, zmmB, zmmC); + cc.evex().vpmulld(xmmA, xmmB, xmmC); + cc.evex().vpmulld(ymmA, ymmB, ymmC); + cc.evex().vpmulld(zmmA, zmmB, zmmC); + cc.evex().vpmullq(xmmA, xmmB, xmmC); + cc.evex().vpmullq(ymmA, ymmB, ymmC); + cc.evex().vpmullq(zmmA, zmmB, zmmC); + cc.evex().vpmullw(xmmA, xmmB, xmmC); + cc.evex().vpmullw(ymmA, ymmB, ymmC); + cc.evex().vpmullw(zmmA, zmmB, zmmC); + cc.evex().vpmultishiftqb(xmmA, xmmB, xmmC); + cc.evex().vpmultishiftqb(ymmA, ymmB, ymmC); + cc.evex().vpmultishiftqb(zmmA, zmmB, zmmC); + cc.evex().vpmuludq(xmmA, xmmB, xmmC); + cc.evex().vpmuludq(ymmA, ymmB, ymmC); + cc.evex().vpmuludq(zmmA, zmmB, zmmC); + cc.evex().vpopcntd(zmmA, zmmB); + cc.evex().vpopcntq(zmmA, zmmB); + cc.evex().vpord(xmmA, xmmB, xmmC); + cc.evex().vpord(ymmA, ymmB, ymmC); + cc.evex().vpord(zmmA, zmmB, zmmC); + cc.evex().vporq(xmmA, xmmB, xmmC); + cc.evex().vporq(ymmA, ymmB, ymmC); + cc.evex().vporq(zmmA, zmmB, zmmC); + cc.evex().vprold(xmmA, xmmB, 0); + cc.evex().vprold(ymmA, ymmB, 0); + cc.evex().vprold(zmmA, zmmB, 0); + cc.evex().vprolq(xmmA, xmmB, 0); + cc.evex().vprolq(ymmA, ymmB, 0); + cc.evex().vprolq(zmmA, zmmB, 0); + cc.evex().vprolvd(xmmA, xmmB, xmmC); + cc.evex().vprolvd(ymmA, ymmB, ymmC); + cc.evex().vprolvd(zmmA, zmmB, zmmC); + cc.evex().vprolvq(xmmA, xmmB, xmmC); + cc.evex().vprolvq(ymmA, ymmB, ymmC); + cc.evex().vprolvq(zmmA, zmmB, zmmC); + cc.evex().vprord(xmmA, xmmB, 0); + cc.evex().vprord(ymmA, ymmB, 0); + cc.evex().vprord(zmmA, zmmB, 0); + cc.evex().vprorq(xmmA, xmmB, 0); + cc.evex().vprorq(ymmA, ymmB, 0); + cc.evex().vprorq(zmmA, zmmB, 0); + cc.evex().vprorvd(xmmA, xmmB, xmmC); + cc.evex().vprorvd(ymmA, ymmB, ymmC); + cc.evex().vprorvd(zmmA, zmmB, zmmC); + cc.evex().vprorvq(xmmA, xmmB, xmmC); + cc.evex().vprorvq(ymmA, ymmB, ymmC); + cc.evex().vprorvq(zmmA, zmmB, zmmC); + cc.evex().vpsadbw(xmmA, xmmB, xmmC); + cc.evex().vpsadbw(ymmA, ymmB, ymmC); + cc.evex().vpsadbw(zmmA, zmmB, zmmC); + cc.evex().vpshufb(xmmA, xmmB, xmmC); + cc.evex().vpshufb(ymmA, ymmB, ymmC); + cc.evex().vpshufb(zmmA, zmmB, zmmC); + cc.evex().vpshufd(xmmA, xmmB, 0); + cc.evex().vpshufd(ymmA, ymmB, 0); + cc.evex().vpshufd(zmmA, zmmB, 0); + cc.evex().vpshufhw(xmmA, xmmB, 0); + cc.evex().vpshufhw(ymmA, ymmB, 0); + cc.evex().vpshufhw(zmmA, zmmB, 0); + cc.evex().vpshuflw(xmmA, xmmB, 0); + cc.evex().vpshuflw(ymmA, ymmB, 0); + cc.evex().vpshuflw(zmmA, zmmB, 0); + cc.evex().vpslld(xmmA, xmmB, xmmC); + cc.evex().vpslld(xmmA, xmmB, 0); + cc.evex().vpslld(ymmA, ymmB, xmmC); + cc.evex().vpslld(ymmA, ymmB, 0); + cc.evex().vpslld(zmmA, zmmB, xmmC); + cc.evex().vpslld(zmmA, zmmB, 0); + cc.evex().vpslldq(xmmA, xmmB, 0); + cc.evex().vpslldq(ymmA, ymmB, 0); + cc.evex().vpslldq(zmmA, zmmB, 0); + cc.evex().vpsllq(xmmA, xmmB, xmmC); + cc.evex().vpsllq(xmmA, xmmB, 0); + cc.evex().vpsllq(ymmA, ymmB, xmmC); + cc.evex().vpsllq(ymmA, ymmB, 0); + cc.evex().vpsllq(zmmA, zmmB, xmmC); + cc.evex().vpsllq(zmmA, zmmB, 0); + cc.evex().vpsllvd(xmmA, xmmB, xmmC); + cc.evex().vpsllvd(ymmA, ymmB, ymmC); + cc.evex().vpsllvd(zmmA, zmmB, zmmC); + cc.evex().vpsllvq(xmmA, xmmB, xmmC); + cc.evex().vpsllvq(ymmA, ymmB, ymmC); + cc.evex().vpsllvq(zmmA, zmmB, zmmC); + cc.evex().vpsllvw(xmmA, xmmB, xmmC); + cc.evex().vpsllvw(ymmA, ymmB, ymmC); + cc.evex().vpsllvw(zmmA, zmmB, zmmC); + cc.evex().vpsllw(xmmA, xmmB, xmmC); + cc.evex().vpsllw(xmmA, xmmB, 0); + cc.evex().vpsllw(ymmA, ymmB, xmmC); + cc.evex().vpsllw(ymmA, ymmB, 0); + cc.evex().vpsllw(zmmA, zmmB, xmmC); + cc.evex().vpsllw(zmmA, zmmB, 0); + cc.evex().vpsrad(xmmA, xmmB, xmmC); + cc.evex().vpsrad(xmmA, xmmB, 0); + cc.evex().vpsrad(ymmA, ymmB, xmmC); + cc.evex().vpsrad(ymmA, ymmB, 0); + cc.evex().vpsrad(zmmA, zmmB, xmmC); + cc.evex().vpsrad(zmmA, zmmB, 0); + cc.evex().vpsraq(xmmA, xmmB, xmmC); + cc.evex().vpsraq(xmmA, xmmB, 0); + cc.evex().vpsraq(ymmA, ymmB, xmmC); + cc.evex().vpsraq(ymmA, ymmB, 0); + cc.evex().vpsraq(zmmA, zmmB, xmmC); + cc.evex().vpsraq(zmmA, zmmB, 0); + cc.evex().vpsravd(xmmA, xmmB, xmmC); + cc.evex().vpsravd(ymmA, ymmB, ymmC); + cc.evex().vpsravd(zmmA, zmmB, zmmC); + cc.evex().vpsravq(xmmA, xmmB, xmmC); + cc.evex().vpsravq(ymmA, ymmB, ymmC); + cc.evex().vpsravq(zmmA, zmmB, zmmC); + cc.evex().vpsravw(xmmA, xmmB, xmmC); + cc.evex().vpsravw(ymmA, ymmB, ymmC); + cc.evex().vpsravw(zmmA, zmmB, zmmC); + cc.evex().vpsraw(xmmA, xmmB, xmmC); + cc.evex().vpsraw(xmmA, xmmB, 0); + cc.evex().vpsraw(ymmA, ymmB, xmmC); + cc.evex().vpsraw(ymmA, ymmB, 0); + cc.evex().vpsraw(zmmA, zmmB, xmmC); + cc.evex().vpsraw(zmmA, zmmB, 0); + cc.evex().vpsrld(xmmA, xmmB, xmmC); + cc.evex().vpsrld(xmmA, xmmB, 0); + cc.evex().vpsrld(ymmA, ymmB, xmmC); + cc.evex().vpsrld(ymmA, ymmB, 0); + cc.evex().vpsrld(zmmA, zmmB, xmmC); + cc.evex().vpsrld(zmmA, zmmB, 0); + cc.evex().vpsrldq(xmmA, xmmB, 0); + cc.evex().vpsrldq(ymmA, ymmB, 0); + cc.evex().vpsrldq(zmmA, zmmB, 0); + cc.evex().vpsrlq(xmmA, xmmB, xmmC); + cc.evex().vpsrlq(xmmA, xmmB, 0); + cc.evex().vpsrlq(ymmA, ymmB, xmmC); + cc.evex().vpsrlq(ymmA, ymmB, 0); + cc.evex().vpsrlq(zmmA, zmmB, xmmC); + cc.evex().vpsrlq(zmmA, zmmB, 0); + cc.evex().vpsrlvd(xmmA, xmmB, xmmC); + cc.evex().vpsrlvd(ymmA, ymmB, ymmC); + cc.evex().vpsrlvd(zmmA, zmmB, zmmC); + cc.evex().vpsrlvq(xmmA, xmmB, xmmC); + cc.evex().vpsrlvq(ymmA, ymmB, ymmC); + cc.evex().vpsrlvq(zmmA, zmmB, zmmC); + cc.evex().vpsrlvw(xmmA, xmmB, xmmC); + cc.evex().vpsrlvw(ymmA, ymmB, ymmC); + cc.evex().vpsrlvw(zmmA, zmmB, zmmC); + cc.evex().vpsrlw(xmmA, xmmB, xmmC); + cc.evex().vpsrlw(xmmA, xmmB, 0); + cc.evex().vpsrlw(ymmA, ymmB, xmmC); + cc.evex().vpsrlw(ymmA, ymmB, 0); + cc.evex().vpsrlw(zmmA, zmmB, xmmC); + cc.evex().vpsrlw(zmmA, zmmB, 0); + cc.evex().vpsubb(xmmA, xmmB, xmmC); + cc.evex().vpsubb(ymmA, ymmB, ymmC); + cc.evex().vpsubb(zmmA, zmmB, zmmC); + cc.evex().vpsubd(xmmA, xmmB, xmmC); + cc.evex().vpsubd(ymmA, ymmB, ymmC); + cc.evex().vpsubd(zmmA, zmmB, zmmC); + cc.evex().vpsubq(xmmA, xmmB, xmmC); + cc.evex().vpsubq(ymmA, ymmB, ymmC); + cc.evex().vpsubq(zmmA, zmmB, zmmC); + cc.evex().vpsubsb(xmmA, xmmB, xmmC); + cc.evex().vpsubsb(ymmA, ymmB, ymmC); + cc.evex().vpsubsb(zmmA, zmmB, zmmC); + cc.evex().vpsubsw(xmmA, xmmB, xmmC); + cc.evex().vpsubsw(ymmA, ymmB, ymmC); + cc.evex().vpsubsw(zmmA, zmmB, zmmC); + cc.evex().vpsubusb(xmmA, xmmB, xmmC); + cc.evex().vpsubusb(ymmA, ymmB, ymmC); + cc.evex().vpsubusb(zmmA, zmmB, zmmC); + cc.evex().vpsubusw(xmmA, xmmB, xmmC); + cc.evex().vpsubusw(ymmA, ymmB, ymmC); + cc.evex().vpsubusw(zmmA, zmmB, zmmC); + cc.evex().vpsubw(xmmA, xmmB, xmmC); + cc.evex().vpsubw(ymmA, ymmB, ymmC); + cc.evex().vpsubw(zmmA, zmmB, zmmC); + cc.evex().vpternlogd(xmmA, xmmB, xmmC, 0); + cc.evex().vpternlogd(ymmA, ymmB, ymmC, 0); + cc.evex().vpternlogd(zmmA, zmmB, zmmC, 0); + cc.evex().vpternlogq(xmmA, xmmB, xmmC, 0); + cc.evex().vpternlogq(ymmA, ymmB, ymmC, 0); + cc.evex().vpternlogq(zmmA, zmmB, zmmC, 0); + cc.evex().vptestmb(kA, xmmB, xmmC); + cc.evex().vptestmb(kA, ymmB, ymmC); + cc.evex().vptestmb(kA, zmmB, zmmC); + cc.evex().vptestmd(kA, xmmB, xmmC); + cc.evex().vptestmd(kA, ymmB, ymmC); + cc.evex().vptestmd(kA, zmmB, zmmC); + cc.evex().vptestmq(kA, xmmB, xmmC); + cc.evex().vptestmq(kA, ymmB, ymmC); + cc.evex().vptestmq(kA, zmmB, zmmC); + cc.evex().vptestmw(kA, xmmB, xmmC); + cc.evex().vptestmw(kA, ymmB, ymmC); + cc.evex().vptestmw(kA, zmmB, zmmC); + cc.evex().vptestnmb(kA, xmmB, xmmC); + cc.evex().vptestnmb(kA, ymmB, ymmC); + cc.evex().vptestnmb(kA, zmmB, zmmC); + cc.evex().vptestnmd(kA, xmmB, xmmC); + cc.evex().vptestnmd(kA, ymmB, ymmC); + cc.evex().vptestnmd(kA, zmmB, zmmC); + cc.evex().vptestnmq(kA, xmmB, xmmC); + cc.evex().vptestnmq(kA, ymmB, ymmC); + cc.evex().vptestnmq(kA, zmmB, zmmC); + cc.evex().vptestnmw(kA, xmmB, xmmC); + cc.evex().vptestnmw(kA, ymmB, ymmC); + cc.evex().vptestnmw(kA, zmmB, zmmC); + cc.evex().vpunpckhbw(xmmA, xmmB, xmmC); + cc.evex().vpunpckhbw(ymmA, ymmB, ymmC); + cc.evex().vpunpckhbw(zmmA, zmmB, zmmC); + cc.evex().vpunpckhdq(xmmA, xmmB, xmmC); + cc.evex().vpunpckhdq(ymmA, ymmB, ymmC); + cc.evex().vpunpckhdq(zmmA, zmmB, zmmC); + cc.evex().vpunpckhqdq(xmmA, xmmB, xmmC); + cc.evex().vpunpckhqdq(ymmA, ymmB, ymmC); + cc.evex().vpunpckhqdq(zmmA, zmmB, zmmC); + cc.evex().vpunpckhwd(xmmA, xmmB, xmmC); + cc.evex().vpunpckhwd(ymmA, ymmB, ymmC); + cc.evex().vpunpckhwd(zmmA, zmmB, zmmC); + cc.evex().vpunpcklbw(xmmA, xmmB, xmmC); + cc.evex().vpunpcklbw(ymmA, ymmB, ymmC); + cc.evex().vpunpcklbw(zmmA, zmmB, zmmC); + cc.evex().vpunpckldq(xmmA, xmmB, xmmC); + cc.evex().vpunpckldq(ymmA, ymmB, ymmC); + cc.evex().vpunpckldq(zmmA, zmmB, zmmC); + cc.evex().vpunpcklqdq(xmmA, xmmB, xmmC); + cc.evex().vpunpcklqdq(ymmA, ymmB, ymmC); + cc.evex().vpunpcklqdq(zmmA, zmmB, zmmC); + cc.evex().vpunpcklwd(xmmA, xmmB, xmmC); + cc.evex().vpunpcklwd(ymmA, ymmB, ymmC); + cc.evex().vpunpcklwd(zmmA, zmmB, zmmC); + cc.evex().vpxord(xmmA, xmmB, xmmC); + cc.evex().vpxord(ymmA, ymmB, ymmC); + cc.evex().vpxord(zmmA, zmmB, zmmC); + cc.evex().vpxorq(xmmA, xmmB, xmmC); + cc.evex().vpxorq(ymmA, ymmB, ymmC); + cc.evex().vpxorq(zmmA, zmmB, zmmC); + cc.evex().vrangepd(xmmA, xmmB, xmmC, 0); + cc.evex().vrangepd(ymmA, ymmB, ymmC, 0); + cc.evex().vrangepd(zmmA, zmmB, zmmC, 0); + cc.evex().vrangeps(xmmA, xmmB, xmmC, 0); + cc.evex().vrangeps(ymmA, ymmB, ymmC, 0); + cc.evex().vrangeps(zmmA, zmmB, zmmC, 0); + cc.evex().vrangesd(xmmA, xmmB, xmmC, 0); + cc.evex().vrangess(xmmA, xmmB, xmmC, 0); + cc.evex().vrcp14pd(xmmA, xmmB); + cc.evex().vrcp14pd(ymmA, ymmB); + cc.evex().vrcp14pd(zmmA, zmmB); + cc.evex().vrcp14ps(xmmA, xmmB); + cc.evex().vrcp14ps(ymmA, ymmB); + cc.evex().vrcp14ps(zmmA, zmmB); + cc.evex().vrcp14sd(xmmA, xmmB, xmmC); + cc.evex().vrcp14ss(xmmA, xmmB, xmmC); + cc.evex().vrcp28pd(zmmA, zmmB); + cc.evex().vrcp28ps(zmmA, zmmB); + cc.evex().vrcp28sd(xmmA, xmmB, xmmC); + cc.evex().vrcp28ss(xmmA, xmmB, xmmC); + cc.evex().vreducepd(xmmA, xmmB, 0); + cc.evex().vreducepd(ymmA, ymmB, 0); + cc.evex().vreducepd(zmmA, zmmB, 0); + cc.evex().vreduceps(xmmA, xmmB, 0); + cc.evex().vreduceps(ymmA, ymmB, 0); + cc.evex().vreduceps(zmmA, zmmB, 0); + cc.evex().vreducesd(xmmA, xmmB, xmmC, 0); + cc.evex().vreducess(xmmA, xmmB, xmmC, 0); + cc.evex().vrndscalepd(xmmA, xmmB, 0); + cc.evex().vrndscalepd(ymmA, ymmB, 0); + cc.evex().vrndscalepd(zmmA, zmmB, 0); + cc.evex().vrndscaleps(xmmA, xmmB, 0); + cc.evex().vrndscaleps(ymmA, ymmB, 0); + cc.evex().vrndscaleps(zmmA, zmmB, 0); + cc.evex().vrndscalesd(xmmA, xmmB, xmmC, 0); + cc.evex().vrndscaless(xmmA, xmmB, xmmC, 0); + cc.evex().vrsqrt14pd(xmmA, xmmB); + cc.evex().vrsqrt14pd(ymmA, ymmB); + cc.evex().vrsqrt14pd(zmmA, zmmB); + cc.evex().vrsqrt14ps(xmmA, xmmB); + cc.evex().vrsqrt14ps(ymmA, ymmB); + cc.evex().vrsqrt14ps(zmmA, zmmB); + cc.evex().vrsqrt14sd(xmmA, xmmB, xmmC); + cc.evex().vrsqrt14ss(xmmA, xmmB, xmmC); + cc.evex().vrsqrt28pd(zmmA, zmmB); + cc.evex().vrsqrt28ps(zmmA, zmmB); + cc.evex().vrsqrt28sd(xmmA, xmmB, xmmC); + cc.evex().vrsqrt28ss(xmmA, xmmB, xmmC); + cc.evex().vscalefpd(xmmA, xmmB, xmmC); + cc.evex().vscalefpd(ymmA, ymmB, ymmC); + cc.evex().vscalefpd(zmmA, zmmB, zmmC); + cc.evex().vscalefps(xmmA, xmmB, xmmC); + cc.evex().vscalefps(ymmA, ymmB, ymmC); + cc.evex().vscalefps(zmmA, zmmB, zmmC); + cc.evex().vscalefsd(xmmA, xmmB, xmmC); + cc.evex().vscalefss(xmmA, xmmB, xmmC); + cc.evex().vshuff32x4(ymmA, ymmB, ymmC, 0); + cc.evex().vshuff32x4(zmmA, zmmB, zmmC, 0); + cc.evex().vshuff64x2(ymmA, ymmB, ymmC, 0); + cc.evex().vshuff64x2(zmmA, zmmB, zmmC, 0); + cc.evex().vshufi32x4(ymmA, ymmB, ymmC, 0); + cc.evex().vshufi32x4(zmmA, zmmB, zmmC, 0); + cc.evex().vshufi64x2(ymmA, ymmB, ymmC, 0); + cc.evex().vshufi64x2(zmmA, zmmB, zmmC, 0); + cc.evex().vshufpd(xmmA, xmmB, xmmC, 0); + cc.evex().vshufpd(ymmA, ymmB, ymmC, 0); + cc.evex().vshufpd(zmmA, zmmB, zmmC, 0); + cc.evex().vshufps(xmmA, xmmB, xmmC, 0); + cc.evex().vshufps(ymmA, ymmB, ymmC, 0); + cc.evex().vshufps(zmmA, zmmB, zmmC, 0); + cc.evex().vsqrtpd(xmmA, xmmB); + cc.evex().vsqrtpd(ymmA, ymmB); + cc.evex().vsqrtpd(zmmA, zmmB); + cc.evex().vsqrtps(xmmA, xmmB); + cc.evex().vsqrtps(ymmA, ymmB); + cc.evex().vsqrtps(zmmA, zmmB); + cc.evex().vsqrtsd(xmmA, xmmB, xmmC); + cc.evex().vsqrtss(xmmA, xmmB, xmmC); + cc.evex().vsubpd(xmmA, xmmB, xmmC); + cc.evex().vsubpd(ymmA, ymmB, ymmC); + cc.evex().vsubpd(zmmA, zmmB, zmmC); + cc.evex().vsubps(xmmA, xmmB, xmmC); + cc.evex().vsubps(ymmA, ymmB, ymmC); + cc.evex().vsubps(zmmA, zmmB, zmmC); + cc.evex().vsubsd(xmmA, xmmB, xmmC); + cc.evex().vsubss(xmmA, xmmB, xmmC); + cc.evex().vucomisd(xmmA, xmmB); + cc.evex().vucomiss(xmmA, xmmB); + cc.evex().vunpckhpd(xmmA, xmmB, xmmC); + cc.evex().vunpckhpd(ymmA, ymmB, ymmC); + cc.evex().vunpckhpd(zmmA, zmmB, zmmC); + cc.evex().vunpckhps(xmmA, xmmB, xmmC); + cc.evex().vunpckhps(ymmA, ymmB, ymmC); + cc.evex().vunpckhps(zmmA, zmmB, zmmC); + cc.evex().vunpcklpd(xmmA, xmmB, xmmC); + cc.evex().vunpcklpd(ymmA, ymmB, ymmC); + cc.evex().vunpcklpd(zmmA, zmmB, zmmC); + cc.evex().vunpcklps(xmmA, xmmB, xmmC); + cc.evex().vunpcklps(ymmA, ymmB, ymmC); + cc.evex().vunpcklps(zmmA, zmmB, zmmC); + cc.evex().vxorpd(xmmA, xmmB, xmmC); + cc.evex().vxorpd(ymmA, ymmB, ymmC); + cc.evex().vxorpd(zmmA, zmmB, zmmC); + cc.evex().vxorps(xmmA, xmmB, xmmC); + cc.evex().vxorps(ymmA, ymmB, ymmC); + cc.evex().vxorps(zmmA, zmmB, zmmC); + } + else { + x86::Mem m = x86::ptr(gpz); + x86::Mem m32 = x86::dword_ptr(gpz); + x86::Mem m64 = x86::qword_ptr(gpz); + x86::Mem m128 = x86::xmmword_ptr(gpz); + x86::Mem m256 = x86::ymmword_ptr(gpz); + x86::Mem m512 = x86::zmmword_ptr(gpz); + x86::Mem vx_ptr = x86::ptr(gpz, xmmD); + x86::Mem vy_ptr = x86::ptr(gpz, ymmD); + x86::Mem vz_ptr = x86::ptr(gpz, zmmD); + + cc.kmovb(kA, m); + cc.kmovb(m, kB); + cc.kmovd(kA, m); + cc.kmovd(m, kB); + cc.kmovq(kA, m); + cc.kmovq(m, kB); + cc.kmovw(kA, m); + cc.kmovw(m, kB); + + cc.evex().vaddpd(xmmA, xmmB, m); + cc.evex().vaddpd(ymmA, ymmB, m); + cc.evex().vaddpd(zmmA, zmmB, m); + cc.evex().vaddps(xmmA, xmmB, m); + cc.evex().vaddps(ymmA, ymmB, m); + cc.evex().vaddps(zmmA, zmmB, m); + cc.evex().vaddsd(xmmA, xmmB, m); + cc.evex().vaddss(xmmA, xmmB, m); + cc.evex().valignd(xmmA, xmmB, m, 0); + cc.evex().valignd(ymmA, ymmB, m, 0); + cc.evex().valignd(zmmA, zmmB, m, 0); + cc.evex().valignq(xmmA, xmmB, m, 0); + cc.evex().valignq(ymmA, ymmB, m, 0); + cc.evex().valignq(zmmA, zmmB, m, 0); + cc.evex().vandnpd(xmmA, xmmB, m); + cc.evex().vandnpd(ymmA, ymmB, m); + cc.evex().vandnpd(zmmA, zmmB, m); + cc.evex().vandnps(xmmA, xmmB, m); + cc.evex().vandnps(ymmA, ymmB, m); + cc.evex().vandnps(zmmA, zmmB, m); + cc.evex().vandpd(xmmA, xmmB, m); + cc.evex().vandpd(ymmA, ymmB, m); + cc.evex().vandpd(zmmA, zmmB, m); + cc.evex().vandps(xmmA, xmmB, m); + cc.evex().vandps(ymmA, ymmB, m); + cc.evex().vandps(zmmA, zmmB, m); + cc.evex().vblendmpd(xmmA, xmmB, m); + cc.evex().vblendmpd(ymmA, ymmB, m); + cc.evex().vblendmpd(zmmA, zmmB, m); + cc.evex().vblendmps(xmmA, xmmB, m); + cc.evex().vblendmps(ymmA, ymmB, m); + cc.evex().vblendmps(zmmA, zmmB, m); + cc.evex().vbroadcastf32x2(ymmA, m); + cc.evex().vbroadcastf32x2(zmmA, m); + cc.evex().vbroadcastf32x4(ymmA, m); + cc.evex().vbroadcastf32x4(zmmA, m); + cc.evex().vbroadcastf32x8(zmmA, m); + cc.evex().vbroadcastf64x2(ymmA, m); + cc.evex().vbroadcastf64x2(zmmA, m); + cc.evex().vbroadcastf64x4(zmmA, m); + cc.evex().vbroadcasti32x2(xmmA, m); + cc.evex().vbroadcasti32x2(ymmA, m); + cc.evex().vbroadcasti32x2(zmmA, m); + cc.evex().vbroadcasti32x4(ymmA, m); + cc.evex().vbroadcasti32x4(zmmA, m); + cc.evex().vbroadcasti32x8(zmmA, m); + cc.evex().vbroadcasti64x2(ymmA, m); + cc.evex().vbroadcasti64x2(zmmA, m); + cc.evex().vbroadcasti64x4(zmmA, m); + cc.evex().vbroadcastsd(ymmA, m); + cc.evex().vbroadcastsd(zmmA, m); + cc.evex().vbroadcastss(xmmA, m); + cc.evex().vbroadcastss(ymmA, m); + cc.evex().vbroadcastss(zmmA, m); + cc.evex().vcmppd(kA, xmmB, m, 0); + cc.evex().vcmppd(kA, ymmB, m, 0); + cc.evex().vcmppd(kA, zmmB, m, 0); + cc.evex().vcmpps(kA, xmmB, m, 0); + cc.evex().vcmpps(kA, ymmB, m, 0); + cc.evex().vcmpps(kA, zmmB, m, 0); + cc.evex().vcmpsd(kA, xmmB, m, 0); + cc.evex().vcmpss(kA, xmmB, m, 0); + cc.evex().vcomisd(xmmA, m); + cc.evex().vcomiss(xmmA, m); + cc.evex().vcompresspd(m, xmmB); + cc.evex().vcompresspd(m, ymmB); + cc.evex().vcompresspd(m, zmmB); + cc.evex().vcompressps(m, xmmB); + cc.evex().vcompressps(m, ymmB); + cc.evex().vcompressps(m, zmmB); + cc.evex().vcvtdq2pd(xmmA, m); + cc.evex().vcvtdq2pd(ymmA, m); + cc.evex().vcvtdq2pd(zmmA, m); + cc.evex().vcvtdq2ps(xmmA, m); + cc.evex().vcvtdq2ps(ymmA, m); + cc.evex().vcvtdq2ps(zmmA, m); + cc.evex().vcvtpd2dq(xmmA, m128); + cc.evex().vcvtpd2dq(xmmA, m256); + cc.evex().vcvtpd2dq(ymmA, m512); + cc.evex().vcvtpd2qq(xmmA, m); + cc.evex().vcvtpd2qq(ymmA, m); + cc.evex().vcvtpd2qq(zmmA, m); + cc.evex().vcvtpd2udq(xmmA, m128); + cc.evex().vcvtpd2udq(xmmA, m256); + cc.evex().vcvtpd2udq(ymmA, m512); + cc.evex().vcvtpd2uqq(xmmA, m); + cc.evex().vcvtpd2uqq(ymmA, m); + cc.evex().vcvtpd2uqq(zmmA, m); + cc.evex().vcvtph2ps(xmmA, m); + cc.evex().vcvtph2ps(ymmA, m); + cc.evex().vcvtph2ps(zmmA, m); + cc.evex().vcvtps2dq(xmmA, m); + cc.evex().vcvtps2dq(ymmA, m); + cc.evex().vcvtps2dq(zmmA, m); + cc.evex().vcvtps2pd(xmmA, m); + cc.evex().vcvtps2pd(ymmA, m); + cc.evex().vcvtps2pd(zmmA, m); + cc.evex().vcvtps2ph(m, xmmB, 0); + cc.evex().vcvtps2ph(m, ymmB, 0); + cc.evex().vcvtps2ph(m, zmmB, 0); + cc.evex().vcvtps2qq(xmmA, m); + cc.evex().vcvtps2qq(ymmA, m); + cc.evex().vcvtps2qq(zmmA, m); + cc.evex().vcvtps2udq(xmmA, m); + cc.evex().vcvtps2udq(ymmA, m); + cc.evex().vcvtps2udq(zmmA, m); + cc.evex().vcvtps2uqq(xmmA, m); + cc.evex().vcvtps2uqq(ymmA, m); + cc.evex().vcvtps2uqq(zmmA, m); + cc.evex().vcvtqq2pd(xmmA, m); + cc.evex().vcvtqq2pd(ymmA, m); + cc.evex().vcvtqq2pd(zmmA, m); + cc.evex().vcvtqq2ps(xmmA, m128); + cc.evex().vcvtqq2ps(xmmA, m256); + cc.evex().vcvtqq2ps(ymmA, m512); + cc.evex().vcvtsd2si(gpd, m); + if (cc.is64Bit()) cc.evex().vcvtsd2si(gpq, m); + cc.evex().vcvtsd2ss(xmmA, xmmB, m); + cc.evex().vcvtsd2usi(gpd, m); + if (cc.is64Bit()) cc.evex().vcvtsd2usi(gpq, m); + cc.evex().vcvtsi2sd(xmmA, xmmB, m32); + if (cc.is64Bit()) cc.evex().vcvtsi2sd(xmmA, xmmB, m64); + cc.evex().vcvtsi2ss(xmmA, xmmB, m32); + if (cc.is64Bit()) cc.evex().vcvtsi2ss(xmmA, xmmB, m64); + cc.evex().vcvtss2sd(xmmA, xmmB, m); + cc.evex().vcvtss2si(gpd, m); + if (cc.is64Bit()) cc.evex().vcvtss2si(gpq, m); + cc.evex().vcvtss2usi(gpd, m); + if (cc.is64Bit()) cc.evex().vcvtss2usi(gpq, m); + cc.evex().vcvttpd2dq(xmmA, m128); + cc.evex().vcvttpd2dq(xmmA, m256); + cc.evex().vcvttpd2dq(ymmA, m512); + cc.evex().vcvttpd2qq(xmmA, m); + cc.evex().vcvttpd2qq(ymmA, m); + cc.evex().vcvttpd2qq(zmmA, m); + cc.evex().vcvttpd2udq(xmmA, m128); + cc.evex().vcvttpd2udq(xmmA, m256); + cc.evex().vcvttpd2udq(ymmA, m512); + cc.evex().vcvttpd2uqq(xmmA, m); + cc.evex().vcvttpd2uqq(ymmA, m); + cc.evex().vcvttpd2uqq(zmmA, m); + cc.evex().vcvttps2dq(xmmA, m); + cc.evex().vcvttps2dq(ymmA, m); + cc.evex().vcvttps2dq(zmmA, m); + cc.evex().vcvttps2qq(xmmA, m); + cc.evex().vcvttps2qq(ymmA, m); + cc.evex().vcvttps2qq(zmmA, m); + cc.evex().vcvttps2udq(xmmA, m); + cc.evex().vcvttps2udq(ymmA, m); + cc.evex().vcvttps2udq(zmmA, m); + cc.evex().vcvttps2uqq(xmmA, m); + cc.evex().vcvttps2uqq(ymmA, m); + cc.evex().vcvttps2uqq(zmmA, m); + cc.evex().vcvttsd2si(gpd, m); + if (cc.is64Bit()) cc.evex().vcvttsd2si(gpq, m); + cc.evex().vcvttsd2usi(gpd, m); + if (cc.is64Bit()) cc.evex().vcvttsd2usi(gpq, m); + cc.evex().vcvttss2si(gpd, m); + if (cc.is64Bit()) cc.evex().vcvttss2si(gpq, m); + cc.evex().vcvttss2usi(gpd, m); + if (cc.is64Bit()) cc.evex().vcvttss2usi(gpq, m); + cc.evex().vcvtudq2pd(xmmA, m); + cc.evex().vcvtudq2pd(ymmA, m); + cc.evex().vcvtudq2pd(zmmA, m); + cc.evex().vcvtudq2ps(xmmA, m); + cc.evex().vcvtudq2ps(ymmA, m); + cc.evex().vcvtudq2ps(zmmA, m); + cc.evex().vcvtuqq2pd(xmmA, m); + cc.evex().vcvtuqq2pd(ymmA, m); + cc.evex().vcvtuqq2pd(zmmA, m); + cc.evex().vcvtuqq2ps(xmmA, m128); + cc.evex().vcvtuqq2ps(xmmA, m256); + cc.evex().vcvtuqq2ps(ymmA, m512); + cc.evex().vcvtusi2sd(xmmA, xmmB, m32); + if (cc.is64Bit()) cc.evex().vcvtusi2sd(xmmA, xmmB, m64); + cc.evex().vcvtusi2ss(xmmA, xmmB, m32); + if (cc.is64Bit()) cc.evex().vcvtusi2ss(xmmA, xmmB, m64); + cc.evex().vdbpsadbw(xmmA, xmmB, m, 0); + cc.evex().vdbpsadbw(ymmA, ymmB, m, 0); + cc.evex().vdbpsadbw(zmmA, zmmB, m, 0); + cc.evex().vdivpd(xmmA, xmmB, m); + cc.evex().vdivpd(ymmA, ymmB, m); + cc.evex().vdivpd(zmmA, zmmB, m); + cc.evex().vdivps(xmmA, xmmB, m); + cc.evex().vdivps(ymmA, ymmB, m); + cc.evex().vdivps(zmmA, zmmB, m); + cc.evex().vdivsd(xmmA, xmmB, m); + cc.evex().vdivss(xmmA, xmmB, m); + cc.evex().vexp2pd(zmmA, m); + cc.evex().vexp2ps(zmmA, m); + cc.evex().vexpandpd(xmmA, m); + cc.evex().vexpandpd(ymmA, m); + cc.evex().vexpandpd(zmmA, m); + cc.evex().vexpandps(xmmA, m); + cc.evex().vexpandps(ymmA, m); + cc.evex().vexpandps(zmmA, m); + cc.evex().vextractf32x4(m, ymmB, 0); + cc.evex().vextractf32x4(m, zmmB, 0); + cc.evex().vextractf32x8(m, zmmB, 0); + cc.evex().vextractf64x2(m, ymmB, 0); + cc.evex().vextractf64x2(m, zmmB, 0); + cc.evex().vextractf64x4(m, zmmB, 0); + cc.evex().vextracti32x4(m, ymmB, 0); + cc.evex().vextracti32x4(m, zmmB, 0); + cc.evex().vextracti32x8(m, zmmB, 0); + cc.evex().vextracti64x2(m, ymmB, 0); + cc.evex().vextracti64x2(m, zmmB, 0); + cc.evex().vextracti64x4(m, zmmB, 0); + cc.evex().vextractps(m, xmmB, 0); + cc.evex().vfixupimmpd(xmmA, xmmB, m, 0); + cc.evex().vfixupimmpd(ymmA, ymmB, m, 0); + cc.evex().vfixupimmpd(zmmA, zmmB, m, 0); + cc.evex().vfixupimmps(xmmA, xmmB, m, 0); + cc.evex().vfixupimmps(ymmA, ymmB, m, 0); + cc.evex().vfixupimmps(zmmA, zmmB, m, 0); + cc.evex().vfixupimmsd(xmmA, xmmB, m, 0); + cc.evex().vfixupimmss(xmmA, xmmB, m, 0); + cc.evex().vfmadd132pd(xmmA, xmmB, m); + cc.evex().vfmadd132pd(ymmA, ymmB, m); + cc.evex().vfmadd132pd(zmmA, zmmB, m); + cc.evex().vfmadd132ps(xmmA, xmmB, m); + cc.evex().vfmadd132ps(ymmA, ymmB, m); + cc.evex().vfmadd132ps(zmmA, zmmB, m); + cc.evex().vfmadd132sd(xmmA, xmmB, m); + cc.evex().vfmadd132ss(xmmA, xmmB, m); + cc.evex().vfmadd213pd(xmmA, xmmB, m); + cc.evex().vfmadd213pd(ymmA, ymmB, m); + cc.evex().vfmadd213pd(zmmA, zmmB, m); + cc.evex().vfmadd213ps(xmmA, xmmB, m); + cc.evex().vfmadd213ps(ymmA, ymmB, m); + cc.evex().vfmadd213ps(zmmA, zmmB, m); + cc.evex().vfmadd213sd(xmmA, xmmB, m); + cc.evex().vfmadd213ss(xmmA, xmmB, m); + cc.evex().vfmadd231pd(xmmA, xmmB, m); + cc.evex().vfmadd231pd(ymmA, ymmB, m); + cc.evex().vfmadd231pd(zmmA, zmmB, m); + cc.evex().vfmadd231ps(xmmA, xmmB, m); + cc.evex().vfmadd231ps(ymmA, ymmB, m); + cc.evex().vfmadd231ps(zmmA, zmmB, m); + cc.evex().vfmadd231sd(xmmA, xmmB, m); + cc.evex().vfmadd231ss(xmmA, xmmB, m); + cc.evex().vfmaddsub132pd(xmmA, xmmB, m); + cc.evex().vfmaddsub132pd(ymmA, ymmB, m); + cc.evex().vfmaddsub132pd(zmmA, zmmB, m); + cc.evex().vfmaddsub132ps(xmmA, xmmB, m); + cc.evex().vfmaddsub132ps(ymmA, ymmB, m); + cc.evex().vfmaddsub132ps(zmmA, zmmB, m); + cc.evex().vfmaddsub213pd(xmmA, xmmB, m); + cc.evex().vfmaddsub213pd(ymmA, ymmB, m); + cc.evex().vfmaddsub213pd(zmmA, zmmB, m); + cc.evex().vfmaddsub213ps(xmmA, xmmB, m); + cc.evex().vfmaddsub213ps(ymmA, ymmB, m); + cc.evex().vfmaddsub213ps(zmmA, zmmB, m); + cc.evex().vfmaddsub231pd(xmmA, xmmB, m); + cc.evex().vfmaddsub231pd(ymmA, ymmB, m); + cc.evex().vfmaddsub231pd(zmmA, zmmB, m); + cc.evex().vfmaddsub231ps(xmmA, xmmB, m); + cc.evex().vfmaddsub231ps(ymmA, ymmB, m); + cc.evex().vfmaddsub231ps(zmmA, zmmB, m); + cc.evex().vfmsub132pd(xmmA, xmmB, m); + cc.evex().vfmsub132pd(ymmA, ymmB, m); + cc.evex().vfmsub132pd(zmmA, zmmB, m); + cc.evex().vfmsub132ps(xmmA, xmmB, m); + cc.evex().vfmsub132ps(ymmA, ymmB, m); + cc.evex().vfmsub132ps(zmmA, zmmB, m); + cc.evex().vfmsub132sd(xmmA, xmmB, m); + cc.evex().vfmsub132ss(xmmA, xmmB, m); + cc.evex().vfmsub213pd(xmmA, xmmB, m); + cc.evex().vfmsub213pd(ymmA, ymmB, m); + cc.evex().vfmsub213pd(zmmA, zmmB, m); + cc.evex().vfmsub213ps(xmmA, xmmB, m); + cc.evex().vfmsub213ps(ymmA, ymmB, m); + cc.evex().vfmsub213ps(zmmA, zmmB, m); + cc.evex().vfmsub213sd(xmmA, xmmB, m); + cc.evex().vfmsub213ss(xmmA, xmmB, m); + cc.evex().vfmsub231pd(xmmA, xmmB, m); + cc.evex().vfmsub231pd(ymmA, ymmB, m); + cc.evex().vfmsub231pd(zmmA, zmmB, m); + cc.evex().vfmsub231ps(xmmA, xmmB, m); + cc.evex().vfmsub231ps(ymmA, ymmB, m); + cc.evex().vfmsub231ps(zmmA, zmmB, m); + cc.evex().vfmsub231sd(xmmA, xmmB, m); + cc.evex().vfmsub231ss(xmmA, xmmB, m); + cc.evex().vfmsubadd132pd(xmmA, xmmB, m); + cc.evex().vfmsubadd132pd(ymmA, ymmB, m); + cc.evex().vfmsubadd132pd(zmmA, zmmB, m); + cc.evex().vfmsubadd132ps(xmmA, xmmB, m); + cc.evex().vfmsubadd132ps(ymmA, ymmB, m); + cc.evex().vfmsubadd132ps(zmmA, zmmB, m); + cc.evex().vfmsubadd213pd(xmmA, xmmB, m); + cc.evex().vfmsubadd213pd(ymmA, ymmB, m); + cc.evex().vfmsubadd213pd(zmmA, zmmB, m); + cc.evex().vfmsubadd213ps(xmmA, xmmB, m); + cc.evex().vfmsubadd213ps(ymmA, ymmB, m); + cc.evex().vfmsubadd213ps(zmmA, zmmB, m); + cc.evex().vfmsubadd231pd(xmmA, xmmB, m); + cc.evex().vfmsubadd231pd(ymmA, ymmB, m); + cc.evex().vfmsubadd231pd(zmmA, zmmB, m); + cc.evex().vfmsubadd231ps(xmmA, xmmB, m); + cc.evex().vfmsubadd231ps(ymmA, ymmB, m); + cc.evex().vfmsubadd231ps(zmmA, zmmB, m); + cc.evex().vfnmadd132pd(xmmA, xmmB, m); + cc.evex().vfnmadd132pd(ymmA, ymmB, m); + cc.evex().vfnmadd132pd(zmmA, zmmB, m); + cc.evex().vfnmadd132ps(xmmA, xmmB, m); + cc.evex().vfnmadd132ps(ymmA, ymmB, m); + cc.evex().vfnmadd132ps(zmmA, zmmB, m); + cc.evex().vfnmadd132sd(xmmA, xmmB, m); + cc.evex().vfnmadd132ss(xmmA, xmmB, m); + cc.evex().vfnmadd213pd(xmmA, xmmB, m); + cc.evex().vfnmadd213pd(ymmA, ymmB, m); + cc.evex().vfnmadd213pd(zmmA, zmmB, m); + cc.evex().vfnmadd213ps(xmmA, xmmB, m); + cc.evex().vfnmadd213ps(ymmA, ymmB, m); + cc.evex().vfnmadd213ps(zmmA, zmmB, m); + cc.evex().vfnmadd213sd(xmmA, xmmB, m); + cc.evex().vfnmadd213ss(xmmA, xmmB, m); + cc.evex().vfnmadd231pd(xmmA, xmmB, m); + cc.evex().vfnmadd231pd(ymmA, ymmB, m); + cc.evex().vfnmadd231pd(zmmA, zmmB, m); + cc.evex().vfnmadd231ps(xmmA, xmmB, m); + cc.evex().vfnmadd231ps(ymmA, ymmB, m); + cc.evex().vfnmadd231ps(zmmA, zmmB, m); + cc.evex().vfnmadd231sd(xmmA, xmmB, m); + cc.evex().vfnmadd231ss(xmmA, xmmB, m); + cc.evex().vfnmsub132pd(xmmA, xmmB, m); + cc.evex().vfnmsub132pd(ymmA, ymmB, m); + cc.evex().vfnmsub132pd(zmmA, zmmB, m); + cc.evex().vfnmsub132ps(xmmA, xmmB, m); + cc.evex().vfnmsub132ps(ymmA, ymmB, m); + cc.evex().vfnmsub132ps(zmmA, zmmB, m); + cc.evex().vfnmsub132sd(xmmA, xmmB, m); + cc.evex().vfnmsub132ss(xmmA, xmmB, m); + cc.evex().vfnmsub213pd(xmmA, xmmB, m); + cc.evex().vfnmsub213pd(ymmA, ymmB, m); + cc.evex().vfnmsub213pd(zmmA, zmmB, m); + cc.evex().vfnmsub213ps(xmmA, xmmB, m); + cc.evex().vfnmsub213ps(ymmA, ymmB, m); + cc.evex().vfnmsub213ps(zmmA, zmmB, m); + cc.evex().vfnmsub213sd(xmmA, xmmB, m); + cc.evex().vfnmsub213ss(xmmA, xmmB, m); + cc.evex().vfnmsub231pd(xmmA, xmmB, m); + cc.evex().vfnmsub231pd(ymmA, ymmB, m); + cc.evex().vfnmsub231pd(zmmA, zmmB, m); + cc.evex().vfnmsub231ps(xmmA, xmmB, m); + cc.evex().vfnmsub231ps(ymmA, ymmB, m); + cc.evex().vfnmsub231ps(zmmA, zmmB, m); + cc.evex().vfnmsub231sd(xmmA, xmmB, m); + cc.evex().vfnmsub231ss(xmmA, xmmB, m); + cc.evex().vfpclasspd(kA, m128, 0); + cc.evex().vfpclasspd(kA, m256, 0); + cc.evex().vfpclasspd(kA, m512, 0); + cc.evex().vfpclassps(kA, m128, 0); + cc.evex().vfpclassps(kA, m256, 0); + cc.evex().vfpclassps(kA, m512, 0); + cc.evex().vfpclasssd(kA, m, 0); + cc.evex().vfpclassss(kA, m, 0); + cc.evex().k(kA).vgatherdpd(xmmA, vx_ptr); + cc.evex().k(kA).vgatherdpd(ymmA, vx_ptr); + cc.evex().k(kA).vgatherdpd(zmmA, vy_ptr); + cc.evex().k(kA).vgatherdps(xmmA, vx_ptr); + cc.evex().k(kA).vgatherdps(ymmA, vy_ptr); + cc.evex().k(kA).vgatherdps(zmmA, vz_ptr); + cc.evex().k(kA).vgatherpf0dpd(vy_ptr); + cc.evex().k(kA).vgatherpf0dps(vz_ptr); + cc.evex().k(kA).vgatherpf0qpd(vz_ptr); + cc.evex().k(kA).vgatherpf0qps(vz_ptr); + cc.evex().k(kA).vgatherpf1dpd(vy_ptr); + cc.evex().k(kA).vgatherpf1dps(vz_ptr); + cc.evex().k(kA).vgatherpf1qpd(vz_ptr); + cc.evex().k(kA).vgatherpf1qps(vz_ptr); + cc.evex().k(kA).vgatherqpd(xmmA, vx_ptr); + cc.evex().k(kA).vgatherqpd(ymmA, vy_ptr); + cc.evex().k(kA).vgatherqpd(zmmA, vz_ptr); + cc.evex().k(kA).vgatherqps(xmmA, vx_ptr); + cc.evex().k(kA).vgatherqps(xmmA, vy_ptr); + cc.evex().k(kA).vgatherqps(ymmA, vz_ptr); + cc.evex().vgetexppd(xmmA, m); + cc.evex().vgetexppd(ymmA, m); + cc.evex().vgetexppd(zmmA, m); + cc.evex().vgetexpps(xmmA, m); + cc.evex().vgetexpps(ymmA, m); + cc.evex().vgetexpps(zmmA, m); + cc.evex().vgetexpsd(xmmA, xmmB, m); + cc.evex().vgetexpss(xmmA, xmmB, m); + cc.evex().vgetmantpd(xmmA, m, 0); + cc.evex().vgetmantpd(ymmA, m, 0); + cc.evex().vgetmantpd(zmmA, m, 0); + cc.evex().vgetmantps(xmmA, m, 0); + cc.evex().vgetmantps(ymmA, m, 0); + cc.evex().vgetmantps(zmmA, m, 0); + cc.evex().vgetmantsd(xmmA, xmmB, m, 0); + cc.evex().vgetmantss(xmmA, xmmB, m, 0); + cc.evex().vinsertf32x4(ymmA, ymmB, m, 0); + cc.evex().vinsertf32x4(zmmA, zmmB, m, 0); + cc.evex().vinsertf32x8(zmmA, zmmB, m, 0); + cc.evex().vinsertf64x2(ymmA, ymmB, m, 0); + cc.evex().vinsertf64x2(zmmA, zmmB, m, 0); + cc.evex().vinsertf64x4(zmmA, zmmB, m, 0); + cc.evex().vinserti32x4(ymmA, ymmB, m, 0); + cc.evex().vinserti32x4(zmmA, zmmB, m, 0); + cc.evex().vinserti32x8(zmmA, zmmB, m, 0); + cc.evex().vinserti64x2(ymmA, ymmB, m, 0); + cc.evex().vinserti64x2(zmmA, zmmB, m, 0); + cc.evex().vinserti64x4(zmmA, zmmB, m, 0); + cc.evex().vinsertps(xmmA, xmmB, m, 0); + cc.evex().vmaxpd(xmmA, xmmB, m); + cc.evex().vmaxpd(ymmA, ymmB, m); + cc.evex().vmaxpd(zmmA, zmmB, m); + cc.evex().vmaxps(xmmA, xmmB, m); + cc.evex().vmaxps(ymmA, ymmB, m); + cc.evex().vmaxps(zmmA, zmmB, m); + cc.evex().vmaxsd(xmmA, xmmB, m); + cc.evex().vmaxss(xmmA, xmmB, m); + cc.evex().vminpd(xmmA, xmmB, m); + cc.evex().vminpd(ymmA, ymmB, m); + cc.evex().vminpd(zmmA, zmmB, m); + cc.evex().vminps(xmmA, xmmB, m); + cc.evex().vminps(ymmA, ymmB, m); + cc.evex().vminps(zmmA, zmmB, m); + cc.evex().vminsd(xmmA, xmmB, m); + cc.evex().vminss(xmmA, xmmB, m); + cc.evex().vmovapd(xmmA, m); + cc.evex().vmovapd(m, xmmB); + cc.evex().vmovapd(ymmA, m); + cc.evex().vmovapd(m, ymmB); + cc.evex().vmovapd(zmmA, m); + cc.evex().vmovapd(m, zmmB); + cc.evex().vmovaps(xmmA, m); + cc.evex().vmovaps(m, xmmB); + cc.evex().vmovaps(ymmA, m); + cc.evex().vmovaps(m, ymmB); + cc.evex().vmovaps(zmmA, m); + cc.evex().vmovaps(m, zmmB); + cc.evex().vmovd(m, xmmB); + cc.evex().vmovd(xmmA, m); + cc.evex().vmovddup(xmmA, m); + cc.evex().vmovddup(ymmA, m); + cc.evex().vmovddup(zmmA, m); + cc.evex().vmovdqa32(xmmA, m); + cc.evex().vmovdqa32(m, xmmB); + cc.evex().vmovdqa32(ymmA, m); + cc.evex().vmovdqa32(m, ymmB); + cc.evex().vmovdqa32(zmmA, m); + cc.evex().vmovdqa32(m, zmmB); + cc.evex().vmovdqa64(xmmA, m); + cc.evex().vmovdqa64(m, xmmB); + cc.evex().vmovdqa64(ymmA, m); + cc.evex().vmovdqa64(m, ymmB); + cc.evex().vmovdqa64(zmmA, m); + cc.evex().vmovdqa64(m, zmmB); + cc.evex().vmovdqu16(xmmA, m); + cc.evex().vmovdqu16(m, xmmB); + cc.evex().vmovdqu16(ymmA, m); + cc.evex().vmovdqu16(m, ymmB); + cc.evex().vmovdqu16(zmmA, m); + cc.evex().vmovdqu16(m, zmmB); + cc.evex().vmovdqu32(xmmA, m); + cc.evex().vmovdqu32(m, xmmB); + cc.evex().vmovdqu32(ymmA, m); + cc.evex().vmovdqu32(m, ymmB); + cc.evex().vmovdqu32(zmmA, m); + cc.evex().vmovdqu32(m, zmmB); + cc.evex().vmovdqu64(xmmA, m); + cc.evex().vmovdqu64(m, xmmB); + cc.evex().vmovdqu64(ymmA, m); + cc.evex().vmovdqu64(m, ymmB); + cc.evex().vmovdqu64(zmmA, m); + cc.evex().vmovdqu64(m, zmmB); + cc.evex().vmovdqu8(xmmA, m); + cc.evex().vmovdqu8(m, xmmB); + cc.evex().vmovdqu8(ymmA, m); + cc.evex().vmovdqu8(m, ymmB); + cc.evex().vmovdqu8(zmmA, m); + cc.evex().vmovdqu8(m, zmmB); + cc.evex().vmovhpd(m, xmmB); + cc.evex().vmovhpd(xmmA, xmmB, m); + cc.evex().vmovhps(m, xmmB); + cc.evex().vmovhps(xmmA, xmmB, m); + cc.evex().vmovlpd(m, xmmB); + cc.evex().vmovlpd(xmmA, xmmB, m); + cc.evex().vmovlps(m, xmmB); + cc.evex().vmovlps(xmmA, xmmB, m); + cc.evex().vmovntdq(m, xmmB); + cc.evex().vmovntdq(m, ymmB); + cc.evex().vmovntdq(m, zmmB); + cc.evex().vmovntdqa(xmmA, m); + cc.evex().vmovntdqa(ymmA, m); + cc.evex().vmovntdqa(zmmA, m); + cc.evex().vmovntpd(m, xmmB); + cc.evex().vmovntpd(m, ymmB); + cc.evex().vmovntpd(m, zmmB); + cc.evex().vmovntps(m, xmmB); + cc.evex().vmovntps(m, ymmB); + cc.evex().vmovntps(m, zmmB); + cc.evex().vmovq(m, xmmB); + cc.evex().vmovq(xmmA, m); + cc.evex().vmovq(xmmA, m); + cc.evex().vmovq(m, xmmB); + cc.evex().vmovsd(m, xmmB); + cc.evex().vmovsd(xmmA, m); + cc.evex().vmovshdup(xmmA, m); + cc.evex().vmovshdup(ymmA, m); + cc.evex().vmovshdup(zmmA, m); + cc.evex().vmovsldup(xmmA, m); + cc.evex().vmovsldup(ymmA, m); + cc.evex().vmovsldup(zmmA, m); + cc.evex().vmovss(m, xmmB); + cc.evex().vmovss(xmmA, m); + cc.evex().vmovupd(xmmA, m); + cc.evex().vmovupd(m, xmmB); + cc.evex().vmovupd(ymmA, m); + cc.evex().vmovupd(m, ymmB); + cc.evex().vmovupd(zmmA, m); + cc.evex().vmovupd(m, zmmB); + cc.evex().vmovups(xmmA, m); + cc.evex().vmovups(m, xmmB); + cc.evex().vmovups(ymmA, m); + cc.evex().vmovups(m, ymmB); + cc.evex().vmovups(zmmA, m); + cc.evex().vmovups(m, zmmB); + cc.evex().vmulpd(xmmA, xmmB, m); + cc.evex().vmulpd(ymmA, ymmB, m); + cc.evex().vmulpd(zmmA, zmmB, m); + cc.evex().vmulps(xmmA, xmmB, m); + cc.evex().vmulps(ymmA, ymmB, m); + cc.evex().vmulps(zmmA, zmmB, m); + cc.evex().vmulsd(xmmA, xmmB, m); + cc.evex().vmulss(xmmA, xmmB, m); + cc.evex().vorpd(xmmA, xmmB, m); + cc.evex().vorpd(ymmA, ymmB, m); + cc.evex().vorpd(zmmA, zmmB, m); + cc.evex().vorps(xmmA, xmmB, m); + cc.evex().vorps(ymmA, ymmB, m); + cc.evex().vorps(zmmA, zmmB, m); + cc.evex().vpabsb(xmmA, m); + cc.evex().vpabsb(ymmA, m); + cc.evex().vpabsb(zmmA, m); + cc.evex().vpabsd(xmmA, m); + cc.evex().vpabsd(ymmA, m); + cc.evex().vpabsd(zmmA, m); + cc.evex().vpabsq(xmmA, m); + cc.evex().vpabsq(ymmA, m); + cc.evex().vpabsq(zmmA, m); + cc.evex().vpabsw(xmmA, m); + cc.evex().vpabsw(ymmA, m); + cc.evex().vpabsw(zmmA, m); + cc.evex().vpackssdw(xmmA, xmmB, m); + cc.evex().vpackssdw(ymmA, ymmB, m); + cc.evex().vpackssdw(zmmA, zmmB, m); + cc.evex().vpacksswb(xmmA, xmmB, m); + cc.evex().vpacksswb(ymmA, ymmB, m); + cc.evex().vpacksswb(zmmA, zmmB, m); + cc.evex().vpackusdw(xmmA, xmmB, m); + cc.evex().vpackusdw(ymmA, ymmB, m); + cc.evex().vpackusdw(zmmA, zmmB, m); + cc.evex().vpackuswb(xmmA, xmmB, m); + cc.evex().vpackuswb(ymmA, ymmB, m); + cc.evex().vpackuswb(zmmA, zmmB, m); + cc.evex().vpaddb(xmmA, xmmB, m); + cc.evex().vpaddb(ymmA, ymmB, m); + cc.evex().vpaddb(zmmA, zmmB, m); + cc.evex().vpaddd(xmmA, xmmB, m); + cc.evex().vpaddd(ymmA, ymmB, m); + cc.evex().vpaddd(zmmA, zmmB, m); + cc.evex().vpaddq(xmmA, xmmB, m); + cc.evex().vpaddq(ymmA, ymmB, m); + cc.evex().vpaddq(zmmA, zmmB, m); + cc.evex().vpaddsb(xmmA, xmmB, m); + cc.evex().vpaddsb(ymmA, ymmB, m); + cc.evex().vpaddsb(zmmA, zmmB, m); + cc.evex().vpaddsw(xmmA, xmmB, m); + cc.evex().vpaddsw(ymmA, ymmB, m); + cc.evex().vpaddsw(zmmA, zmmB, m); + cc.evex().vpaddusb(xmmA, xmmB, m); + cc.evex().vpaddusb(ymmA, ymmB, m); + cc.evex().vpaddusb(zmmA, zmmB, m); + cc.evex().vpaddusw(xmmA, xmmB, m); + cc.evex().vpaddusw(ymmA, ymmB, m); + cc.evex().vpaddusw(zmmA, zmmB, m); + cc.evex().vpaddw(xmmA, xmmB, m); + cc.evex().vpaddw(ymmA, ymmB, m); + cc.evex().vpaddw(zmmA, zmmB, m); + cc.evex().vpalignr(xmmA, xmmB, m, 0); + cc.evex().vpalignr(ymmA, ymmB, m, 0); + cc.evex().vpalignr(zmmA, zmmB, m, 0); + cc.evex().vpandd(xmmA, xmmB, m); + cc.evex().vpandd(ymmA, ymmB, m); + cc.evex().vpandd(zmmA, zmmB, m); + cc.evex().vpandnd(xmmA, xmmB, m); + cc.evex().vpandnd(ymmA, ymmB, m); + cc.evex().vpandnd(zmmA, zmmB, m); + cc.evex().vpandnq(xmmA, xmmB, m); + cc.evex().vpandnq(ymmA, ymmB, m); + cc.evex().vpandnq(zmmA, zmmB, m); + cc.evex().vpandq(xmmA, xmmB, m); + cc.evex().vpandq(ymmA, ymmB, m); + cc.evex().vpandq(zmmA, zmmB, m); + cc.evex().vpavgb(xmmA, xmmB, m); + cc.evex().vpavgb(ymmA, ymmB, m); + cc.evex().vpavgb(zmmA, zmmB, m); + cc.evex().vpavgw(xmmA, xmmB, m); + cc.evex().vpavgw(ymmA, ymmB, m); + cc.evex().vpavgw(zmmA, zmmB, m); + cc.evex().vpblendmb(xmmA, xmmB, m); + cc.evex().vpblendmb(ymmA, ymmB, m); + cc.evex().vpblendmb(zmmA, zmmB, m); + cc.evex().vpblendmd(xmmA, xmmB, m); + cc.evex().vpblendmd(ymmA, ymmB, m); + cc.evex().vpblendmd(zmmA, zmmB, m); + cc.evex().vpblendmq(xmmA, xmmB, m); + cc.evex().vpblendmq(ymmA, ymmB, m); + cc.evex().vpblendmq(zmmA, zmmB, m); + cc.evex().vpblendmw(xmmA, xmmB, m); + cc.evex().vpblendmw(ymmA, ymmB, m); + cc.evex().vpblendmw(zmmA, zmmB, m); + cc.evex().vpbroadcastb(xmmA, m); + cc.evex().vpbroadcastb(ymmA, m); + cc.evex().vpbroadcastb(zmmA, m); + cc.evex().vpbroadcastd(xmmA, m); + cc.evex().vpbroadcastd(ymmA, m); + cc.evex().vpbroadcastd(zmmA, m); + cc.evex().vpbroadcastq(xmmA, m); + cc.evex().vpbroadcastq(ymmA, m); + cc.evex().vpbroadcastq(zmmA, m); + cc.evex().vpbroadcastw(xmmA, m); + cc.evex().vpbroadcastw(ymmA, m); + cc.evex().vpbroadcastw(zmmA, m); + cc.evex().vpcmpb(kA, xmmB, m, 0); + cc.evex().vpcmpb(kA, ymmB, m, 0); + cc.evex().vpcmpb(kA, zmmB, m, 0); + cc.evex().vpcmpd(kA, xmmB, m, 0); + cc.evex().vpcmpd(kA, ymmB, m, 0); + cc.evex().vpcmpd(kA, zmmB, m, 0); + cc.evex().vpcmpeqb(kA, xmmB, m); + cc.evex().vpcmpeqb(kA, ymmB, m); + cc.evex().vpcmpeqb(kA, zmmB, m); + cc.evex().vpcmpeqd(kA, xmmB, m); + cc.evex().vpcmpeqd(kA, ymmB, m); + cc.evex().vpcmpeqd(kA, zmmB, m); + cc.evex().vpcmpeqq(kA, xmmB, m); + cc.evex().vpcmpeqq(kA, ymmB, m); + cc.evex().vpcmpeqq(kA, zmmB, m); + cc.evex().vpcmpeqw(kA, xmmB, m); + cc.evex().vpcmpeqw(kA, ymmB, m); + cc.evex().vpcmpeqw(kA, zmmB, m); + cc.evex().vpcmpgtb(kA, xmmB, m); + cc.evex().vpcmpgtb(kA, ymmB, m); + cc.evex().vpcmpgtb(kA, zmmB, m); + cc.evex().vpcmpgtd(kA, xmmB, m); + cc.evex().vpcmpgtd(kA, ymmB, m); + cc.evex().vpcmpgtd(kA, zmmB, m); + cc.evex().vpcmpgtq(kA, xmmB, m); + cc.evex().vpcmpgtq(kA, ymmB, m); + cc.evex().vpcmpgtq(kA, zmmB, m); + cc.evex().vpcmpgtw(kA, xmmB, m); + cc.evex().vpcmpgtw(kA, ymmB, m); + cc.evex().vpcmpgtw(kA, zmmB, m); + cc.evex().vpcmpq(kA, xmmB, m, 0); + cc.evex().vpcmpq(kA, ymmB, m, 0); + cc.evex().vpcmpq(kA, zmmB, m, 0); + cc.evex().vpcmpub(kA, xmmB, m, 0); + cc.evex().vpcmpub(kA, ymmB, m, 0); + cc.evex().vpcmpub(kA, zmmB, m, 0); + cc.evex().vpcmpud(kA, xmmB, m, 0); + cc.evex().vpcmpud(kA, ymmB, m, 0); + cc.evex().vpcmpud(kA, zmmB, m, 0); + cc.evex().vpcmpuq(kA, xmmB, m, 0); + cc.evex().vpcmpuq(kA, ymmB, m, 0); + cc.evex().vpcmpuq(kA, zmmB, m, 0); + cc.evex().vpcmpuw(kA, xmmB, m, 0); + cc.evex().vpcmpuw(kA, ymmB, m, 0); + cc.evex().vpcmpuw(kA, zmmB, m, 0); + cc.evex().vpcmpw(kA, xmmB, m, 0); + cc.evex().vpcmpw(kA, ymmB, m, 0); + cc.evex().vpcmpw(kA, zmmB, m, 0); + cc.evex().vpcompressd(m, xmmB); + cc.evex().vpcompressd(m, ymmB); + cc.evex().vpcompressd(m, zmmB); + cc.evex().vpcompressq(m, xmmB); + cc.evex().vpcompressq(m, ymmB); + cc.evex().vpcompressq(m, zmmB); + cc.evex().vpconflictd(xmmA, m); + cc.evex().vpconflictd(ymmA, m); + cc.evex().vpconflictd(zmmA, m); + cc.evex().vpconflictq(xmmA, m); + cc.evex().vpconflictq(ymmA, m); + cc.evex().vpconflictq(zmmA, m); + cc.evex().vpermb(xmmA, xmmB, m); + cc.evex().vpermb(ymmA, ymmB, m); + cc.evex().vpermb(zmmA, zmmB, m); + cc.evex().vpermd(ymmA, ymmB, m); + cc.evex().vpermd(zmmA, zmmB, m); + cc.evex().vpermi2b(xmmA, xmmB, m); + cc.evex().vpermi2b(ymmA, ymmB, m); + cc.evex().vpermi2b(zmmA, zmmB, m); + cc.evex().vpermi2d(xmmA, xmmB, m); + cc.evex().vpermi2d(ymmA, ymmB, m); + cc.evex().vpermi2d(zmmA, zmmB, m); + cc.evex().vpermi2pd(xmmA, xmmB, m); + cc.evex().vpermi2pd(ymmA, ymmB, m); + cc.evex().vpermi2pd(zmmA, zmmB, m); + cc.evex().vpermi2ps(xmmA, xmmB, m); + cc.evex().vpermi2ps(ymmA, ymmB, m); + cc.evex().vpermi2ps(zmmA, zmmB, m); + cc.evex().vpermi2q(xmmA, xmmB, m); + cc.evex().vpermi2q(ymmA, ymmB, m); + cc.evex().vpermi2q(zmmA, zmmB, m); + cc.evex().vpermi2w(xmmA, xmmB, m); + cc.evex().vpermi2w(ymmA, ymmB, m); + cc.evex().vpermi2w(zmmA, zmmB, m); + cc.evex().vpermilpd(xmmA, xmmB, m); + cc.evex().vpermilpd(ymmA, ymmB, m); + cc.evex().vpermilpd(zmmA, zmmB, m); + cc.evex().vpermilpd(xmmA, m, 0); + cc.evex().vpermilpd(ymmA, m, 0); + cc.evex().vpermilpd(zmmA, m, 0); + cc.evex().vpermilps(xmmA, xmmB, m); + cc.evex().vpermilps(ymmA, ymmB, m); + cc.evex().vpermilps(zmmA, zmmB, m); + cc.evex().vpermilps(xmmA, m, 0); + cc.evex().vpermilps(ymmA, m, 0); + cc.evex().vpermilps(zmmA, m, 0); + cc.evex().vpermq(ymmA, ymmB, m); + cc.evex().vpermq(zmmA, zmmB, m); + cc.evex().vpermq(ymmA, m, 0); + cc.evex().vpermq(zmmA, m, 0); + cc.evex().vpermt2b(xmmA, xmmB, m); + cc.evex().vpermt2b(ymmA, ymmB, m); + cc.evex().vpermt2b(zmmA, zmmB, m); + cc.evex().vpermt2d(xmmA, xmmB, m); + cc.evex().vpermt2d(ymmA, ymmB, m); + cc.evex().vpermt2d(zmmA, zmmB, m); + cc.evex().vpermt2pd(xmmA, xmmB, m); + cc.evex().vpermt2pd(ymmA, ymmB, m); + cc.evex().vpermt2pd(zmmA, zmmB, m); + cc.evex().vpermt2ps(xmmA, xmmB, m); + cc.evex().vpermt2ps(ymmA, ymmB, m); + cc.evex().vpermt2ps(zmmA, zmmB, m); + cc.evex().vpermt2q(xmmA, xmmB, m); + cc.evex().vpermt2q(ymmA, ymmB, m); + cc.evex().vpermt2q(zmmA, zmmB, m); + cc.evex().vpermt2w(xmmA, xmmB, m); + cc.evex().vpermt2w(ymmA, ymmB, m); + cc.evex().vpermt2w(zmmA, zmmB, m); + cc.evex().vpermw(xmmA, xmmB, m); + cc.evex().vpermw(ymmA, ymmB, m); + cc.evex().vpermw(zmmA, zmmB, m); + cc.evex().vpexpandd(xmmA, m); + cc.evex().vpexpandd(ymmA, m); + cc.evex().vpexpandd(zmmA, m); + cc.evex().vpexpandq(xmmA, m); + cc.evex().vpexpandq(ymmA, m); + cc.evex().vpexpandq(zmmA, m); + cc.evex().vpextrb(m, xmmB, 0); + cc.evex().vpextrd(m, xmmB, 0); + if (cc.is64Bit()) cc.evex().vpextrq(m, xmmB, 0); + cc.evex().vpextrw(m, xmmB, 0); + cc.evex().k(kA).vpgatherdd(xmmA, vx_ptr); + cc.evex().k(kA).vpgatherdd(ymmA, vy_ptr); + cc.evex().k(kA).vpgatherdd(zmmA, vz_ptr); + cc.evex().k(kA).vpgatherdq(xmmA, vx_ptr); + cc.evex().k(kA).vpgatherdq(ymmA, vx_ptr); + cc.evex().k(kA).vpgatherdq(zmmA, vy_ptr); + cc.evex().k(kA).vpgatherqd(xmmA, vx_ptr); + cc.evex().k(kA).vpgatherqd(xmmA, vy_ptr); + cc.evex().k(kA).vpgatherqd(ymmA, vz_ptr); + cc.evex().k(kA).vpgatherqq(xmmA, vx_ptr); + cc.evex().k(kA).vpgatherqq(ymmA, vy_ptr); + cc.evex().k(kA).vpgatherqq(zmmA, vz_ptr); + cc.evex().vpinsrb(xmmA, xmmB, m, 0); + cc.evex().vpinsrd(xmmA, xmmB, m, 0); + if (cc.is64Bit()) cc.evex().vpinsrq(xmmA, xmmB, m, 0); + cc.evex().vpinsrw(xmmA, xmmB, m, 0); + cc.evex().vplzcntd(xmmA, m); + cc.evex().vplzcntd(ymmA, m); + cc.evex().vplzcntd(zmmA, m); + cc.evex().vplzcntq(xmmA, m); + cc.evex().vplzcntq(ymmA, m); + cc.evex().vplzcntq(zmmA, m); + cc.evex().vpmadd52huq(xmmA, xmmB, m); + cc.evex().vpmadd52huq(ymmA, ymmB, m); + cc.evex().vpmadd52huq(zmmA, zmmB, m); + cc.evex().vpmadd52luq(xmmA, xmmB, m); + cc.evex().vpmadd52luq(ymmA, ymmB, m); + cc.evex().vpmadd52luq(zmmA, zmmB, m); + cc.evex().vpmaddubsw(xmmA, xmmB, m); + cc.evex().vpmaddubsw(ymmA, ymmB, m); + cc.evex().vpmaddubsw(zmmA, zmmB, m); + cc.evex().vpmaddwd(xmmA, xmmB, m); + cc.evex().vpmaddwd(ymmA, ymmB, m); + cc.evex().vpmaddwd(zmmA, zmmB, m); + cc.evex().vpmaxsb(xmmA, xmmB, m); + cc.evex().vpmaxsb(ymmA, ymmB, m); + cc.evex().vpmaxsb(zmmA, zmmB, m); + cc.evex().vpmaxsd(xmmA, xmmB, m); + cc.evex().vpmaxsd(ymmA, ymmB, m); + cc.evex().vpmaxsd(zmmA, zmmB, m); + cc.evex().vpmaxsq(xmmA, xmmB, m); + cc.evex().vpmaxsq(ymmA, ymmB, m); + cc.evex().vpmaxsq(zmmA, zmmB, m); + cc.evex().vpmaxsw(xmmA, xmmB, m); + cc.evex().vpmaxsw(ymmA, ymmB, m); + cc.evex().vpmaxsw(zmmA, zmmB, m); + cc.evex().vpmaxub(xmmA, xmmB, m); + cc.evex().vpmaxub(ymmA, ymmB, m); + cc.evex().vpmaxub(zmmA, zmmB, m); + cc.evex().vpmaxud(xmmA, xmmB, m); + cc.evex().vpmaxud(ymmA, ymmB, m); + cc.evex().vpmaxud(zmmA, zmmB, m); + cc.evex().vpmaxuq(xmmA, xmmB, m); + cc.evex().vpmaxuq(ymmA, ymmB, m); + cc.evex().vpmaxuq(zmmA, zmmB, m); + cc.evex().vpmaxuw(xmmA, xmmB, m); + cc.evex().vpmaxuw(ymmA, ymmB, m); + cc.evex().vpmaxuw(zmmA, zmmB, m); + cc.evex().vpminsb(xmmA, xmmB, m); + cc.evex().vpminsb(ymmA, ymmB, m); + cc.evex().vpminsb(zmmA, zmmB, m); + cc.evex().vpminsd(xmmA, xmmB, m); + cc.evex().vpminsd(ymmA, ymmB, m); + cc.evex().vpminsd(zmmA, zmmB, m); + cc.evex().vpminsq(xmmA, xmmB, m); + cc.evex().vpminsq(ymmA, ymmB, m); + cc.evex().vpminsq(zmmA, zmmB, m); + cc.evex().vpminsw(xmmA, xmmB, m); + cc.evex().vpminsw(ymmA, ymmB, m); + cc.evex().vpminsw(zmmA, zmmB, m); + cc.evex().vpminub(xmmA, xmmB, m); + cc.evex().vpminub(ymmA, ymmB, m); + cc.evex().vpminub(zmmA, zmmB, m); + cc.evex().vpminud(xmmA, xmmB, m); + cc.evex().vpminud(ymmA, ymmB, m); + cc.evex().vpminud(zmmA, zmmB, m); + cc.evex().vpminuq(xmmA, xmmB, m); + cc.evex().vpminuq(ymmA, ymmB, m); + cc.evex().vpminuq(zmmA, zmmB, m); + cc.evex().vpminuw(xmmA, xmmB, m); + cc.evex().vpminuw(ymmA, ymmB, m); + cc.evex().vpminuw(zmmA, zmmB, m); + cc.evex().vpmovdb(m, xmmB); + cc.evex().vpmovdb(m, ymmB); + cc.evex().vpmovdb(m, zmmB); + cc.evex().vpmovdw(m, xmmB); + cc.evex().vpmovdw(m, ymmB); + cc.evex().vpmovdw(m, zmmB); + cc.evex().vpmovqb(m, xmmB); + cc.evex().vpmovqb(m, ymmB); + cc.evex().vpmovqb(m, zmmB); + cc.evex().vpmovqd(m, xmmB); + cc.evex().vpmovqd(m, ymmB); + cc.evex().vpmovqd(m, zmmB); + cc.evex().vpmovqw(m, xmmB); + cc.evex().vpmovqw(m, ymmB); + cc.evex().vpmovqw(m, zmmB); + cc.evex().vpmovsdb(m, xmmB); + cc.evex().vpmovsdb(m, ymmB); + cc.evex().vpmovsdb(m, zmmB); + cc.evex().vpmovsdw(m, xmmB); + cc.evex().vpmovsdw(m, ymmB); + cc.evex().vpmovsdw(m, zmmB); + cc.evex().vpmovsqb(m, xmmB); + cc.evex().vpmovsqb(m, ymmB); + cc.evex().vpmovsqb(m, zmmB); + cc.evex().vpmovsqd(m, xmmB); + cc.evex().vpmovsqd(m, ymmB); + cc.evex().vpmovsqd(m, zmmB); + cc.evex().vpmovsqw(m, xmmB); + cc.evex().vpmovsqw(m, ymmB); + cc.evex().vpmovsqw(m, zmmB); + cc.evex().vpmovswb(m, xmmB); + cc.evex().vpmovswb(m, ymmB); + cc.evex().vpmovswb(m, zmmB); + cc.evex().vpmovsxbd(xmmA, m); + cc.evex().vpmovsxbd(ymmA, m); + cc.evex().vpmovsxbd(zmmA, m); + cc.evex().vpmovsxbq(xmmA, m); + cc.evex().vpmovsxbq(ymmA, m); + cc.evex().vpmovsxbq(zmmA, m); + cc.evex().vpmovsxbw(xmmA, m); + cc.evex().vpmovsxbw(ymmA, m); + cc.evex().vpmovsxbw(zmmA, m); + cc.evex().vpmovsxdq(xmmA, m); + cc.evex().vpmovsxdq(ymmA, m); + cc.evex().vpmovsxdq(zmmA, m); + cc.evex().vpmovsxwd(xmmA, m); + cc.evex().vpmovsxwd(ymmA, m); + cc.evex().vpmovsxwd(zmmA, m); + cc.evex().vpmovsxwq(xmmA, m); + cc.evex().vpmovsxwq(ymmA, m); + cc.evex().vpmovsxwq(zmmA, m); + cc.evex().vpmovusdb(m, xmmB); + cc.evex().vpmovusdb(m, ymmB); + cc.evex().vpmovusdb(m, zmmB); + cc.evex().vpmovusdw(m, xmmB); + cc.evex().vpmovusdw(m, ymmB); + cc.evex().vpmovusdw(m, zmmB); + cc.evex().vpmovusqb(m, xmmB); + cc.evex().vpmovusqb(m, ymmB); + cc.evex().vpmovusqb(m, zmmB); + cc.evex().vpmovusqd(m, xmmB); + cc.evex().vpmovusqd(m, ymmB); + cc.evex().vpmovusqd(m, zmmB); + cc.evex().vpmovusqw(m, xmmB); + cc.evex().vpmovusqw(m, ymmB); + cc.evex().vpmovusqw(m, zmmB); + cc.evex().vpmovuswb(m, xmmB); + cc.evex().vpmovuswb(m, ymmB); + cc.evex().vpmovuswb(m, zmmB); + cc.evex().vpmovwb(m, xmmB); + cc.evex().vpmovwb(m, ymmB); + cc.evex().vpmovwb(m, zmmB); + cc.evex().vpmovzxbd(xmmA, m); + cc.evex().vpmovzxbd(ymmA, m); + cc.evex().vpmovzxbd(zmmA, m); + cc.evex().vpmovzxbq(xmmA, m); + cc.evex().vpmovzxbq(ymmA, m); + cc.evex().vpmovzxbq(zmmA, m); + cc.evex().vpmovzxbw(xmmA, m); + cc.evex().vpmovzxbw(ymmA, m); + cc.evex().vpmovzxbw(zmmA, m); + cc.evex().vpmovzxdq(xmmA, m); + cc.evex().vpmovzxdq(ymmA, m); + cc.evex().vpmovzxdq(zmmA, m); + cc.evex().vpmovzxwd(xmmA, m); + cc.evex().vpmovzxwd(ymmA, m); + cc.evex().vpmovzxwd(zmmA, m); + cc.evex().vpmovzxwq(xmmA, m); + cc.evex().vpmovzxwq(ymmA, m); + cc.evex().vpmovzxwq(zmmA, m); + cc.evex().vpmuldq(xmmA, xmmB, m); + cc.evex().vpmuldq(ymmA, ymmB, m); + cc.evex().vpmuldq(zmmA, zmmB, m); + cc.evex().vpmulhrsw(xmmA, xmmB, m); + cc.evex().vpmulhrsw(ymmA, ymmB, m); + cc.evex().vpmulhrsw(zmmA, zmmB, m); + cc.evex().vpmulhuw(xmmA, xmmB, m); + cc.evex().vpmulhuw(ymmA, ymmB, m); + cc.evex().vpmulhuw(zmmA, zmmB, m); + cc.evex().vpmulhw(xmmA, xmmB, m); + cc.evex().vpmulhw(ymmA, ymmB, m); + cc.evex().vpmulhw(zmmA, zmmB, m); + cc.evex().vpmulld(xmmA, xmmB, m); + cc.evex().vpmulld(ymmA, ymmB, m); + cc.evex().vpmulld(zmmA, zmmB, m); + cc.evex().vpmullq(xmmA, xmmB, m); + cc.evex().vpmullq(ymmA, ymmB, m); + cc.evex().vpmullq(zmmA, zmmB, m); + cc.evex().vpmullw(xmmA, xmmB, m); + cc.evex().vpmullw(ymmA, ymmB, m); + cc.evex().vpmullw(zmmA, zmmB, m); + cc.evex().vpmultishiftqb(xmmA, xmmB, m); + cc.evex().vpmultishiftqb(ymmA, ymmB, m); + cc.evex().vpmultishiftqb(zmmA, zmmB, m); + cc.evex().vpmuludq(xmmA, xmmB, m); + cc.evex().vpmuludq(ymmA, ymmB, m); + cc.evex().vpmuludq(zmmA, zmmB, m); + cc.evex().vpopcntd(zmmA, m); + cc.evex().vpopcntq(zmmA, m); + cc.evex().vpord(xmmA, xmmB, m); + cc.evex().vpord(ymmA, ymmB, m); + cc.evex().vpord(zmmA, zmmB, m); + cc.evex().vporq(xmmA, xmmB, m); + cc.evex().vporq(ymmA, ymmB, m); + cc.evex().vporq(zmmA, zmmB, m); + cc.evex().vprold(xmmA, m, 0); + cc.evex().vprold(ymmA, m, 0); + cc.evex().vprold(zmmA, m, 0); + cc.evex().vprolq(xmmA, m, 0); + cc.evex().vprolq(ymmA, m, 0); + cc.evex().vprolq(zmmA, m, 0); + cc.evex().vprolvd(xmmA, xmmB, m); + cc.evex().vprolvd(ymmA, ymmB, m); + cc.evex().vprolvd(zmmA, zmmB, m); + cc.evex().vprolvq(xmmA, xmmB, m); + cc.evex().vprolvq(ymmA, ymmB, m); + cc.evex().vprolvq(zmmA, zmmB, m); + cc.evex().vprord(xmmA, m, 0); + cc.evex().vprord(ymmA, m, 0); + cc.evex().vprord(zmmA, m, 0); + cc.evex().vprorq(xmmA, m, 0); + cc.evex().vprorq(ymmA, m, 0); + cc.evex().vprorq(zmmA, m, 0); + cc.evex().vprorvd(xmmA, xmmB, m); + cc.evex().vprorvd(ymmA, ymmB, m); + cc.evex().vprorvd(zmmA, zmmB, m); + cc.evex().vprorvq(xmmA, xmmB, m); + cc.evex().vprorvq(ymmA, ymmB, m); + cc.evex().vprorvq(zmmA, zmmB, m); + cc.evex().vpsadbw(xmmA, xmmB, m); + cc.evex().vpsadbw(ymmA, ymmB, m); + cc.evex().vpsadbw(zmmA, zmmB, m); + cc.evex().k(kA).vpscatterdd(vx_ptr, xmmB); + cc.evex().k(kA).vpscatterdd(vy_ptr, ymmB); + cc.evex().k(kA).vpscatterdd(vz_ptr, zmmB); + cc.evex().k(kA).vpscatterdq(vx_ptr, xmmB); + cc.evex().k(kA).vpscatterdq(vx_ptr, ymmB); + cc.evex().k(kA).vpscatterdq(vy_ptr, zmmB); + cc.evex().k(kA).vpscatterqd(vx_ptr, xmmB); + cc.evex().k(kA).vpscatterqd(vy_ptr, xmmB); + cc.evex().k(kA).vpscatterqd(vz_ptr, ymmB); + cc.evex().k(kA).vpscatterqq(vx_ptr, xmmB); + cc.evex().k(kA).vpscatterqq(vy_ptr, ymmB); + cc.evex().k(kA).vpscatterqq(vz_ptr, zmmB); + cc.evex().vpshufb(xmmA, xmmB, m); + cc.evex().vpshufb(ymmA, ymmB, m); + cc.evex().vpshufb(zmmA, zmmB, m); + cc.evex().vpshufd(xmmA, m, 0); + cc.evex().vpshufd(ymmA, m, 0); + cc.evex().vpshufd(zmmA, m, 0); + cc.evex().vpshufhw(xmmA, m, 0); + cc.evex().vpshufhw(ymmA, m, 0); + cc.evex().vpshufhw(zmmA, m, 0); + cc.evex().vpshuflw(xmmA, m, 0); + cc.evex().vpshuflw(ymmA, m, 0); + cc.evex().vpshuflw(zmmA, m, 0); + cc.evex().vpslld(xmmA, xmmB, m); + cc.evex().vpslld(xmmA, m, 0); + cc.evex().vpslld(ymmA, ymmB, m); + cc.evex().vpslld(ymmA, m, 0); + cc.evex().vpslld(zmmA, zmmB, m); + cc.evex().vpslld(zmmA, m, 0); + cc.evex().vpslldq(xmmA, m, 0); + cc.evex().vpslldq(ymmA, m, 0); + cc.evex().vpslldq(zmmA, m, 0); + cc.evex().vpsllq(xmmA, xmmB, m); + cc.evex().vpsllq(xmmA, m, 0); + cc.evex().vpsllq(ymmA, ymmB, m); + cc.evex().vpsllq(ymmA, m, 0); + cc.evex().vpsllq(zmmA, zmmB, m); + cc.evex().vpsllq(zmmA, m, 0); + cc.evex().vpsllvd(xmmA, xmmB, m); + cc.evex().vpsllvd(ymmA, ymmB, m); + cc.evex().vpsllvd(zmmA, zmmB, m); + cc.evex().vpsllvq(xmmA, xmmB, m); + cc.evex().vpsllvq(ymmA, ymmB, m); + cc.evex().vpsllvq(zmmA, zmmB, m); + cc.evex().vpsllvw(xmmA, xmmB, m); + cc.evex().vpsllvw(ymmA, ymmB, m); + cc.evex().vpsllvw(zmmA, zmmB, m); + cc.evex().vpsllw(xmmA, xmmB, m); + cc.evex().vpsllw(xmmA, m, 0); + cc.evex().vpsllw(ymmA, ymmB, m); + cc.evex().vpsllw(ymmA, m, 0); + cc.evex().vpsllw(zmmA, zmmB, m); + cc.evex().vpsllw(zmmA, m, 0); + cc.evex().vpsrad(xmmA, xmmB, m); + cc.evex().vpsrad(xmmA, m, 0); + cc.evex().vpsrad(ymmA, ymmB, m); + cc.evex().vpsrad(ymmA, m, 0); + cc.evex().vpsrad(zmmA, zmmB, m); + cc.evex().vpsrad(zmmA, m, 0); + cc.evex().vpsraq(xmmA, xmmB, m); + cc.evex().vpsraq(xmmA, m, 0); + cc.evex().vpsraq(ymmA, ymmB, m); + cc.evex().vpsraq(ymmA, m, 0); + cc.evex().vpsraq(zmmA, zmmB, m); + cc.evex().vpsraq(zmmA, m, 0); + cc.evex().vpsravd(xmmA, xmmB, m); + cc.evex().vpsravd(ymmA, ymmB, m); + cc.evex().vpsravd(zmmA, zmmB, m); + cc.evex().vpsravq(xmmA, xmmB, m); + cc.evex().vpsravq(ymmA, ymmB, m); + cc.evex().vpsravq(zmmA, zmmB, m); + cc.evex().vpsravw(xmmA, xmmB, m); + cc.evex().vpsravw(ymmA, ymmB, m); + cc.evex().vpsravw(zmmA, zmmB, m); + cc.evex().vpsraw(xmmA, xmmB, m); + cc.evex().vpsraw(xmmA, m, 0); + cc.evex().vpsraw(ymmA, ymmB, m); + cc.evex().vpsraw(ymmA, m, 0); + cc.evex().vpsraw(zmmA, zmmB, m); + cc.evex().vpsraw(zmmA, m, 0); + cc.evex().vpsrld(xmmA, xmmB, m); + cc.evex().vpsrld(xmmA, m, 0); + cc.evex().vpsrld(ymmA, ymmB, m); + cc.evex().vpsrld(ymmA, m, 0); + cc.evex().vpsrld(zmmA, zmmB, m); + cc.evex().vpsrld(zmmA, m, 0); + cc.evex().vpsrldq(xmmA, m, 0); + cc.evex().vpsrldq(ymmA, m, 0); + cc.evex().vpsrldq(zmmA, m, 0); + cc.evex().vpsrlq(xmmA, xmmB, m); + cc.evex().vpsrlq(xmmA, m, 0); + cc.evex().vpsrlq(ymmA, ymmB, m); + cc.evex().vpsrlq(ymmA, m, 0); + cc.evex().vpsrlq(zmmA, zmmB, m); + cc.evex().vpsrlq(zmmA, m, 0); + cc.evex().vpsrlvd(xmmA, xmmB, m); + cc.evex().vpsrlvd(ymmA, ymmB, m); + cc.evex().vpsrlvd(zmmA, zmmB, m); + cc.evex().vpsrlvq(xmmA, xmmB, m); + cc.evex().vpsrlvq(ymmA, ymmB, m); + cc.evex().vpsrlvq(zmmA, zmmB, m); + cc.evex().vpsrlvw(xmmA, xmmB, m); + cc.evex().vpsrlvw(ymmA, ymmB, m); + cc.evex().vpsrlvw(zmmA, zmmB, m); + cc.evex().vpsrlw(xmmA, xmmB, m); + cc.evex().vpsrlw(xmmA, m, 0); + cc.evex().vpsrlw(ymmA, ymmB, m); + cc.evex().vpsrlw(ymmA, m, 0); + cc.evex().vpsrlw(zmmA, zmmB, m); + cc.evex().vpsrlw(zmmA, m, 0); + cc.evex().vpsubb(xmmA, xmmB, m); + cc.evex().vpsubb(ymmA, ymmB, m); + cc.evex().vpsubb(zmmA, zmmB, m); + cc.evex().vpsubd(xmmA, xmmB, m); + cc.evex().vpsubd(ymmA, ymmB, m); + cc.evex().vpsubd(zmmA, zmmB, m); + cc.evex().vpsubq(xmmA, xmmB, m); + cc.evex().vpsubq(ymmA, ymmB, m); + cc.evex().vpsubq(zmmA, zmmB, m); + cc.evex().vpsubsb(xmmA, xmmB, m); + cc.evex().vpsubsb(ymmA, ymmB, m); + cc.evex().vpsubsb(zmmA, zmmB, m); + cc.evex().vpsubsw(xmmA, xmmB, m); + cc.evex().vpsubsw(ymmA, ymmB, m); + cc.evex().vpsubsw(zmmA, zmmB, m); + cc.evex().vpsubusb(xmmA, xmmB, m); + cc.evex().vpsubusb(ymmA, ymmB, m); + cc.evex().vpsubusb(zmmA, zmmB, m); + cc.evex().vpsubusw(xmmA, xmmB, m); + cc.evex().vpsubusw(ymmA, ymmB, m); + cc.evex().vpsubusw(zmmA, zmmB, m); + cc.evex().vpsubw(xmmA, xmmB, m); + cc.evex().vpsubw(ymmA, ymmB, m); + cc.evex().vpsubw(zmmA, zmmB, m); + cc.evex().vpternlogd(xmmA, xmmB, m, 0); + cc.evex().vpternlogd(ymmA, ymmB, m, 0); + cc.evex().vpternlogd(zmmA, zmmB, m, 0); + cc.evex().vpternlogq(xmmA, xmmB, m, 0); + cc.evex().vpternlogq(ymmA, ymmB, m, 0); + cc.evex().vpternlogq(zmmA, zmmB, m, 0); + cc.evex().vptestmb(kA, xmmB, m); + cc.evex().vptestmb(kA, ymmB, m); + cc.evex().vptestmb(kA, zmmB, m); + cc.evex().vptestmd(kA, xmmB, m); + cc.evex().vptestmd(kA, ymmB, m); + cc.evex().vptestmd(kA, zmmB, m); + cc.evex().vptestmq(kA, xmmB, m); + cc.evex().vptestmq(kA, ymmB, m); + cc.evex().vptestmq(kA, zmmB, m); + cc.evex().vptestmw(kA, xmmB, m); + cc.evex().vptestmw(kA, ymmB, m); + cc.evex().vptestmw(kA, zmmB, m); + cc.evex().vptestnmb(kA, xmmB, m); + cc.evex().vptestnmb(kA, ymmB, m); + cc.evex().vptestnmb(kA, zmmB, m); + cc.evex().vptestnmd(kA, xmmB, m); + cc.evex().vptestnmd(kA, ymmB, m); + cc.evex().vptestnmd(kA, zmmB, m); + cc.evex().vptestnmq(kA, xmmB, m); + cc.evex().vptestnmq(kA, ymmB, m); + cc.evex().vptestnmq(kA, zmmB, m); + cc.evex().vptestnmw(kA, xmmB, m); + cc.evex().vptestnmw(kA, ymmB, m); + cc.evex().vptestnmw(kA, zmmB, m); + cc.evex().vpunpckhbw(xmmA, xmmB, m); + cc.evex().vpunpckhbw(ymmA, ymmB, m); + cc.evex().vpunpckhbw(zmmA, zmmB, m); + cc.evex().vpunpckhdq(xmmA, xmmB, m); + cc.evex().vpunpckhdq(ymmA, ymmB, m); + cc.evex().vpunpckhdq(zmmA, zmmB, m); + cc.evex().vpunpckhqdq(xmmA, xmmB, m); + cc.evex().vpunpckhqdq(ymmA, ymmB, m); + cc.evex().vpunpckhqdq(zmmA, zmmB, m); + cc.evex().vpunpckhwd(xmmA, xmmB, m); + cc.evex().vpunpckhwd(ymmA, ymmB, m); + cc.evex().vpunpckhwd(zmmA, zmmB, m); + cc.evex().vpunpcklbw(xmmA, xmmB, m); + cc.evex().vpunpcklbw(ymmA, ymmB, m); + cc.evex().vpunpcklbw(zmmA, zmmB, m); + cc.evex().vpunpckldq(xmmA, xmmB, m); + cc.evex().vpunpckldq(ymmA, ymmB, m); + cc.evex().vpunpckldq(zmmA, zmmB, m); + cc.evex().vpunpcklqdq(xmmA, xmmB, m); + cc.evex().vpunpcklqdq(ymmA, ymmB, m); + cc.evex().vpunpcklqdq(zmmA, zmmB, m); + cc.evex().vpunpcklwd(xmmA, xmmB, m); + cc.evex().vpunpcklwd(ymmA, ymmB, m); + cc.evex().vpunpcklwd(zmmA, zmmB, m); + cc.evex().vpxord(xmmA, xmmB, m); + cc.evex().vpxord(ymmA, ymmB, m); + cc.evex().vpxord(zmmA, zmmB, m); + cc.evex().vpxorq(xmmA, xmmB, m); + cc.evex().vpxorq(ymmA, ymmB, m); + cc.evex().vpxorq(zmmA, zmmB, m); + cc.evex().vrangepd(xmmA, xmmB, m, 0); + cc.evex().vrangepd(ymmA, ymmB, m, 0); + cc.evex().vrangepd(zmmA, zmmB, m, 0); + cc.evex().vrangeps(xmmA, xmmB, m, 0); + cc.evex().vrangeps(ymmA, ymmB, m, 0); + cc.evex().vrangeps(zmmA, zmmB, m, 0); + cc.evex().vrangesd(xmmA, xmmB, m, 0); + cc.evex().vrangess(xmmA, xmmB, m, 0); + cc.evex().vrcp14pd(xmmA, m); + cc.evex().vrcp14pd(ymmA, m); + cc.evex().vrcp14pd(zmmA, m); + cc.evex().vrcp14ps(xmmA, m); + cc.evex().vrcp14ps(ymmA, m); + cc.evex().vrcp14ps(zmmA, m); + cc.evex().vrcp14sd(xmmA, xmmB, m); + cc.evex().vrcp14ss(xmmA, xmmB, m); + cc.evex().vrcp28pd(zmmA, m); + cc.evex().vrcp28ps(zmmA, m); + cc.evex().vrcp28sd(xmmA, xmmB, m); + cc.evex().vrcp28ss(xmmA, xmmB, m); + cc.evex().vreducepd(xmmA, m, 0); + cc.evex().vreducepd(ymmA, m, 0); + cc.evex().vreducepd(zmmA, m, 0); + cc.evex().vreduceps(xmmA, m, 0); + cc.evex().vreduceps(ymmA, m, 0); + cc.evex().vreduceps(zmmA, m, 0); + cc.evex().vreducesd(xmmA, xmmB, m, 0); + cc.evex().vreducess(xmmA, xmmB, m, 0); + cc.evex().vrndscalepd(xmmA, m, 0); + cc.evex().vrndscalepd(ymmA, m, 0); + cc.evex().vrndscalepd(zmmA, m, 0); + cc.evex().vrndscaleps(xmmA, m, 0); + cc.evex().vrndscaleps(ymmA, m, 0); + cc.evex().vrndscaleps(zmmA, m, 0); + cc.evex().vrndscalesd(xmmA, xmmB, m, 0); + cc.evex().vrndscaless(xmmA, xmmB, m, 0); + cc.evex().vrsqrt14pd(xmmA, m); + cc.evex().vrsqrt14pd(ymmA, m); + cc.evex().vrsqrt14pd(zmmA, m); + cc.evex().vrsqrt14ps(xmmA, m); + cc.evex().vrsqrt14ps(ymmA, m); + cc.evex().vrsqrt14ps(zmmA, m); + cc.evex().vrsqrt14sd(xmmA, xmmB, m); + cc.evex().vrsqrt14ss(xmmA, xmmB, m); + cc.evex().vrsqrt28pd(zmmA, m); + cc.evex().vrsqrt28ps(zmmA, m); + cc.evex().vrsqrt28sd(xmmA, xmmB, m); + cc.evex().vrsqrt28ss(xmmA, xmmB, m); + cc.evex().vscalefpd(xmmA, xmmB, m); + cc.evex().vscalefpd(ymmA, ymmB, m); + cc.evex().vscalefpd(zmmA, zmmB, m); + cc.evex().vscalefps(xmmA, xmmB, m); + cc.evex().vscalefps(ymmA, ymmB, m); + cc.evex().vscalefps(zmmA, zmmB, m); + cc.evex().vscalefsd(xmmA, xmmB, m); + cc.evex().vscalefss(xmmA, xmmB, m); + cc.evex().k(kA).vscatterdpd(vx_ptr, xmmB); + cc.evex().k(kA).vscatterdpd(vx_ptr, ymmB); + cc.evex().k(kA).vscatterdpd(vy_ptr, zmmB); + cc.evex().k(kA).vscatterdps(vx_ptr, xmmB); + cc.evex().k(kA).vscatterdps(vy_ptr, ymmB); + cc.evex().k(kA).vscatterdps(vz_ptr, zmmB); + cc.evex().k(kA).vscatterpf0dpd(vy_ptr); + cc.evex().k(kA).vscatterpf0dps(vz_ptr); + cc.evex().k(kA).vscatterpf0qpd(vz_ptr); + cc.evex().k(kA).vscatterpf0qps(vz_ptr); + cc.evex().k(kA).vscatterpf1dpd(vy_ptr); + cc.evex().k(kA).vscatterpf1dps(vz_ptr); + cc.evex().k(kA).vscatterpf1qpd(vz_ptr); + cc.evex().k(kA).vscatterpf1qps(vz_ptr); + cc.evex().k(kA).vscatterqpd(vx_ptr, xmmB); + cc.evex().k(kA).vscatterqpd(vy_ptr, ymmB); + cc.evex().k(kA).vscatterqpd(vz_ptr, zmmB); + cc.evex().k(kA).vscatterqps(vx_ptr, xmmB); + cc.evex().k(kA).vscatterqps(vy_ptr, xmmB); + cc.evex().k(kA).vscatterqps(vz_ptr, ymmB); + cc.evex().vshuff32x4(ymmA, ymmB, m, 0); + cc.evex().vshuff32x4(zmmA, zmmB, m, 0); + cc.evex().vshuff64x2(ymmA, ymmB, m, 0); + cc.evex().vshuff64x2(zmmA, zmmB, m, 0); + cc.evex().vshufi32x4(ymmA, ymmB, m, 0); + cc.evex().vshufi32x4(zmmA, zmmB, m, 0); + cc.evex().vshufi64x2(ymmA, ymmB, m, 0); + cc.evex().vshufi64x2(zmmA, zmmB, m, 0); + cc.evex().vshufpd(xmmA, xmmB, m, 0); + cc.evex().vshufpd(ymmA, ymmB, m, 0); + cc.evex().vshufpd(zmmA, zmmB, m, 0); + cc.evex().vshufps(xmmA, xmmB, m, 0); + cc.evex().vshufps(ymmA, ymmB, m, 0); + cc.evex().vshufps(zmmA, zmmB, m, 0); + cc.evex().vsqrtpd(xmmA, m); + cc.evex().vsqrtpd(ymmA, m); + cc.evex().vsqrtpd(zmmA, m); + cc.evex().vsqrtps(xmmA, m); + cc.evex().vsqrtps(ymmA, m); + cc.evex().vsqrtps(zmmA, m); + cc.evex().vsqrtsd(xmmA, xmmB, m); + cc.evex().vsqrtss(xmmA, xmmB, m); + cc.evex().vsubpd(xmmA, xmmB, m); + cc.evex().vsubpd(ymmA, ymmB, m); + cc.evex().vsubpd(zmmA, zmmB, m); + cc.evex().vsubps(xmmA, xmmB, m); + cc.evex().vsubps(ymmA, ymmB, m); + cc.evex().vsubps(zmmA, zmmB, m); + cc.evex().vsubsd(xmmA, xmmB, m); + cc.evex().vsubss(xmmA, xmmB, m); + cc.evex().vucomisd(xmmA, m); + cc.evex().vucomiss(xmmA, m); + cc.evex().vunpckhpd(xmmA, xmmB, m); + cc.evex().vunpckhpd(ymmA, ymmB, m); + cc.evex().vunpckhpd(zmmA, zmmB, m); + cc.evex().vunpckhps(xmmA, xmmB, m); + cc.evex().vunpckhps(ymmA, ymmB, m); + cc.evex().vunpckhps(zmmA, zmmB, m); + cc.evex().vunpcklpd(xmmA, xmmB, m); + cc.evex().vunpcklpd(ymmA, ymmB, m); + cc.evex().vunpcklpd(zmmA, zmmB, m); + cc.evex().vunpcklps(xmmA, xmmB, m); + cc.evex().vunpcklps(ymmA, ymmB, m); + cc.evex().vunpcklps(zmmA, zmmB, m); + cc.evex().vxorpd(xmmA, xmmB, m); + cc.evex().vxorpd(ymmA, ymmB, m); + cc.evex().vxorpd(zmmA, zmmB, m); + cc.evex().vxorps(xmmA, xmmB, m); + cc.evex().vxorps(ymmA, ymmB, m); + cc.evex().vxorps(zmmA, zmmB, m); + } +} + +static void generateAvx512Sequence(BaseEmitter& emitter, InstForm form, bool emitPrologEpilog) { + using namespace asmjit::x86; + + if (emitter.isAssembler()) { + Assembler& cc = *emitter.as<Assembler>(); + + if (emitPrologEpilog) { + FuncDetail func; + func.init(FuncSignatureT<void, void*, const void*, size_t>(CallConv::kIdHost), cc.environment()); + + FuncFrame frame; + frame.init(func); + frame.addDirtyRegs(eax, k1, k2, k3, zmm0, zmm1, zmm2, zmm3); + frame.finalize(); + + cc.emitProlog(frame); + generateAvx512SequenceInternal(cc, form, eax, k1, k2, k3, zmm0, zmm1, zmm2, zmm3); + cc.emitEpilog(frame); + } + else { + generateAvx512SequenceInternal(cc, form, eax, k1, k2, k3, zmm0, zmm1, zmm2, zmm3); + } + } +#ifndef ASMJIT_NO_BUILDER + else if (emitter.isBuilder()) { + Builder& cc = *emitter.as<Builder>(); + + if (emitPrologEpilog) { + FuncDetail func; + func.init(FuncSignatureT<void, void*, const void*, size_t>(CallConv::kIdHost), cc.environment()); + + FuncFrame frame; + frame.init(func); + frame.addDirtyRegs(eax, k1, k2, k3, zmm0, zmm1, zmm2, zmm3); + frame.finalize(); + + cc.emitProlog(frame); + generateAvx512SequenceInternal(cc, form, eax, k1, k2, k3, zmm0, zmm1, zmm2, zmm3); + cc.emitEpilog(frame); + } + else { + generateAvx512SequenceInternal(cc, form, eax, k1, k2, k3, zmm0, zmm1, zmm2, zmm3); + } + } +#endif +#ifndef ASMJIT_NO_COMPILER + else if (emitter.isCompiler()) { + Compiler& cc = *emitter.as<Compiler>(); + + Gp gp = cc.newGpz("gp"); + Zmm vecA = cc.newZmm("vecA"); + Zmm vecB = cc.newZmm("vecB"); + Zmm vecC = cc.newZmm("vecC"); + Zmm vecD = cc.newZmm("vecD"); + + KReg kA = cc.newKq("kA"); + KReg kB = cc.newKq("kB"); + KReg kC = cc.newKq("kC"); + + cc.addFunc(FuncSignatureT<void>(CallConv::kIdHost)); + generateAvx512SequenceInternal(cc, form, gp, kA, kB, kC, vecA, vecB, vecC, vecD); + cc.endFunc(); + } +#endif +} + +template<typename EmitterFn> +static void benchmarkX86Function(uint32_t arch, uint32_t numIterations, const char* description, const EmitterFn& emitterFn) noexcept { + CodeHolder code; + printf("%s:\n", description); + + bench<x86::Assembler>(code, arch, numIterations, "[raw]", [&](x86::Assembler& cc) { + emitterFn(cc, false); + }); + + bench<x86::Assembler>(code, arch, numIterations, "[validated]", [&](x86::Assembler& cc) { + cc.addValidationOptions(BaseEmitter::kValidationOptionAssembler); + emitterFn(cc, false); + }); + + bench<x86::Assembler>(code, arch, numIterations, "[prolog/epilog]", [&](x86::Assembler& cc) { + cc.addValidationOptions(BaseEmitter::kValidationOptionAssembler); + emitterFn(cc, true); + }); + +#ifndef ASMJIT_NO_BUILDER + bench<x86::Builder>(code, arch, numIterations, "[no-asm]", [&](x86::Builder& cc) { + emitterFn(cc, false); + }); + + bench<x86::Builder>(code, arch, numIterations, "[finalized]", [&](x86::Builder& cc) { + emitterFn(cc, false); + cc.finalize(); + }); + + bench<x86::Builder>(code, arch, numIterations, "[prolog/epilog]", [&](x86::Builder& cc) { + emitterFn(cc, true); + cc.finalize(); + }); +#endif + +#ifndef ASMJIT_NO_COMPILER + bench<x86::Compiler>(code, arch, numIterations, "[no-asm]", [&](x86::Compiler& cc) { + emitterFn(cc, true); + }); + + bench<x86::Compiler>(code, arch, numIterations, "[finalized]", [&](x86::Compiler& cc) { + emitterFn(cc, true); + cc.finalize(); + }); +#endif + + printf("\n"); +} + +void benchmarkX86Emitters(uint32_t numIterations, bool testX86, bool testX64) { + uint32_t i = 0; + uint32_t n = 0; + + uint32_t archs[2] {}; + + if (testX86) archs[n++] = Environment::kArchX86; + if (testX64) archs[n++] = Environment::kArchX64; + + for (i = 0; i < n; i++) { + static const char description[] = "GpSequence<Reg> (Sequence of GP instructions - reg-only)"; + benchmarkX86Function(archs[i], numIterations, description, [](BaseEmitter& emitter, bool emitPrologEpilog) { + generateGpSequence(emitter, InstForm::kReg, emitPrologEpilog); + }); + } + + for (i = 0; i < n; i++) { + static const char description[] = "GpSequence<Mem> (Sequence of GP instructions - reg/mem)"; + benchmarkX86Function(archs[i], numIterations, description, [](BaseEmitter& emitter, bool emitPrologEpilog) { + generateGpSequence(emitter, InstForm::kMem, emitPrologEpilog); + }); + } + + for (i = 0; i < n; i++) { + static const char description[] = "SseSequence<Reg> (sequence of SSE+ instructions - reg-only)"; + benchmarkX86Function(archs[i], numIterations, description, [](BaseEmitter& emitter, bool emitPrologEpilog) { + generateSseSequence(emitter, InstForm::kReg, emitPrologEpilog); + }); + } + + for (i = 0; i < n; i++) { + static const char description[] = "SseSequence<Mem> (sequence of SSE+ instructions - reg/mem)"; + benchmarkX86Function(archs[i], numIterations, description, [](BaseEmitter& emitter, bool emitPrologEpilog) { + generateSseSequence(emitter, InstForm::kMem, emitPrologEpilog); + }); + } + + for (i = 0; i < n; i++) { + static const char description[] = "AvxSequence<Reg> (sequence of AVX+ instructions - reg-only)"; + benchmarkX86Function(archs[i], numIterations, description, [](BaseEmitter& emitter, bool emitPrologEpilog) { + generateAvxSequence(emitter, InstForm::kReg, emitPrologEpilog); + }); + } + + for (i = 0; i < n; i++) { + static const char description[] = "AvxSequence<Mem> (sequence of AVX+ instructions - reg/mem)"; + benchmarkX86Function(archs[i], numIterations, description, [](BaseEmitter& emitter, bool emitPrologEpilog) { + generateAvxSequence(emitter, InstForm::kMem, emitPrologEpilog); + }); + } + + for (i = 0; i < n; i++) { + static const char description[] = "Avx512Sequence<Reg> (sequence of AVX512+ instructions - reg-only)"; + benchmarkX86Function(archs[i], numIterations, description, [](BaseEmitter& emitter, bool emitPrologEpilog) { + generateAvx512Sequence(emitter, InstForm::kReg, emitPrologEpilog); + }); + } + + for (i = 0; i < n; i++) { + static const char description[] = "Avx512Sequence<Mem> (sequence of AVX512+ instructions - reg/mem)"; + benchmarkX86Function(archs[i], numIterations, description, [](BaseEmitter& emitter, bool emitPrologEpilog) { + generateAvx512Sequence(emitter, InstForm::kMem, emitPrologEpilog); + }); + } + + for (i = 0; i < n; i++) { + static const char description[] = "SseAlphaBlend (alpha-blend function with labels and jumps)"; + benchmarkX86Function(archs[i], numIterations, description, [](BaseEmitter& emitter, bool emitPrologEpilog) { + asmtest::generateSseAlphaBlend(emitter, emitPrologEpilog); + }); + } +} +#endif diff --git a/test/cmdline.h b/test/cmdline.h index caafa6c..effedd1 100644 --- a/test/cmdline.h +++ b/test/cmdline.h @@ -21,8 +21,8 @@ // misrepresented as being the original software. // 3. This notice may not be removed or altered from any source distribution. -#ifndef ASMJIT_TEST_CMDLINE_H_INCLUDED -#define ASMJIT_TEST_CMDLINE_H_INCLUDED +#ifndef CMDLINE_H_INCLUDED +#define CMDLINE_H_INCLUDED #include <stdint.h> #include <stdlib.h> @@ -80,4 +80,4 @@ public: } }; -#endif // ASMJIT_TEST_CMDLINE_H_INCLUDED +#endif // CMDLINE_H_INCLUDED diff --git a/test/performancetimer.h b/test/performancetimer.h new file mode 100644 index 0000000..ebbaca3 --- /dev/null +++ b/test/performancetimer.h @@ -0,0 +1,59 @@ +// AsmJit - Machine code generation for C++ +// +// * Official AsmJit Home Page: https://asmjit.com +// * Official Github Repository: https://github.com/asmjit/asmjit +// +// Copyright (c) 2008-2020 The AsmJit Authors +// +// This software is provided 'as-is', without any express or implied +// warranty. In no event will the authors be held liable for any damages +// arising from the use of this software. +// +// Permission is granted to anyone to use this software for any purpose, +// including commercial applications, and to alter it and redistribute it +// freely, subject to the following restrictions: +// +// 1. The origin of this software must not be misrepresented; you must not +// claim that you wrote the original software. If you use this software +// in a product, an acknowledgment in the product documentation would be +// appreciated but is not required. +// 2. Altered source versions must be plainly marked as such, and must not be +// misrepresented as being the original software. +// 3. This notice may not be removed or altered from any source distribution. + +#ifndef PERFORMANCETIMER_H_INCLUDED +#define PERFORMANCETIMER_H_INCLUDED + +#include <asmjit/core.h> +#include <chrono> + +class PerformanceTimer { +public: + typedef std::chrono::high_resolution_clock::time_point TimePoint; + + TimePoint _startTime {}; + TimePoint _endTime {}; + + inline void start() { + _startTime = std::chrono::high_resolution_clock::now(); + } + + inline void stop() { + _endTime = std::chrono::high_resolution_clock::now(); + } + + inline double duration() const { + std::chrono::duration<double> elapsed = _endTime - _startTime; + return elapsed.count() * 1000; + } +}; + +static inline double mbps(double duration, uint64_t outputSize) noexcept { + if (duration == 0) + return 0.0; + + double bytesTotal = double(outputSize); + return (bytesTotal * 1000) / (duration * 1024 * 1024); +} + +#endif // PERFORMANCETIMER_H_INCLUDED |