diff options
| author | Henrik Gramner <gramner@twoorioles.com> | 2021-08-23 20:22:01 +0300 |
|---|---|---|
| committer | Henrik Gramner <gramner@twoorioles.com> | 2021-08-31 12:47:42 +0300 |
| commit | 64c65de3faf1fa5e0272858203a9590f4ca31927 (patch) | |
| tree | 9dbb39ddf653a0c04fb36540319665ed75239f76 | |
| parent | 94a0e5ec13febf376d9bf29ed8d0bb693c690c5c (diff) | |
x86: Automatically convert SSE asm to AVX when compiling for AVX targets
Requires meson 0.51 or newer (older versions will just keep the SSE).
| -rw-r--r-- | .gitlab-ci.yml | 15 | ||||
| -rw-r--r-- | meson.build | 3 | ||||
| -rw-r--r-- | src/ext/x86/x86inc.asm | 7 |
3 files changed, 24 insertions, 1 deletions
diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml index 3a566dc..c0029a5 100644 --- a/.gitlab-ci.yml +++ b/.gitlab-ci.yml @@ -178,6 +178,21 @@ build-debian-bitdepth-16: -Dbitdepths=16 - ninja -C build +build-debian-avx: + extends: .debian-amd64-common + tags: + - docker + - avx2 + - amd64 + variables: + CFLAGS: '-mavx' + script: + - meson build --buildtype debug + --werror + - ninja -C build + - cd build + - time meson test -v --suite checkasm + build-win32: extends: .debian-amd64-common script: diff --git a/meson.build b/meson.build index abbb28b..96ac31e 100644 --- a/meson.build +++ b/meson.build @@ -375,6 +375,9 @@ if host_machine.cpu_family().startswith('x86') cdata_asm.set10('ARCH_X86_64', host_machine.cpu_family() == 'x86_64') cdata_asm.set10('ARCH_X86_32', host_machine.cpu_family() == 'x86') cdata_asm.set10('PIC', true) + + # Convert SSE asm into (128-bit) AVX when compiler flags are set to use AVX instructions + cdata_asm.set10('FORCE_VEX_ENCODING', cc.get_define('__AVX__') != '') endif cdata.set10('ARCH_PPC64LE', host_machine.cpu() == 'ppc64le') diff --git a/src/ext/x86/x86inc.asm b/src/ext/x86/x86inc.asm index d7db10e..c9f2d63 100644 --- a/src/ext/x86/x86inc.asm +++ b/src/ext/x86/x86inc.asm @@ -79,6 +79,11 @@ %define mangle(x) x %endif +; Use VEX-encoding even in non-AVX functions +%ifndef FORCE_VEX_ENCODING + %define FORCE_VEX_ENCODING 0 +%endif + %macro SECTION_RODATA 0-1 16 %ifidn __OUTPUT_FORMAT__,win32 SECTION .rdata align=%1 @@ -1008,7 +1013,7 @@ BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae, %endmacro %macro INIT_XMM 0-1+ - %assign avx_enabled 0 + %assign avx_enabled FORCE_VEX_ENCODING %define RESET_MM_PERMUTATION INIT_XMM %1 %define mmsize 16 %define mova movdqa |