MSVC fixes:

1. Do not pass -O3 on MSVC (the need for such custom flags is only on mobile builds anyway).
2. Pass /arch:AVX* flags, actually enabling SIMD paths on MSVC
3. Fix the #if logic to detect AVX2+FMA on MSVX: /arch:AVX2 enables FMA but does not define __FMA__.

PiperOrigin-RevId: 339375591

3 files changed, 37 insertions, 15 deletions

diff --git a/ruy/BUILD b/ruy/BUILD
index 59218d4..2bdfd41 100644
--- a/ruy/BUILD
+++ b/ruy/BUILD
@@ -40,8 +40,11 @@ config_setting(
     values = {"cpu": "haswell"},
 )
 
+# MSVC toolchains define a different "cpu" value, which helps us as we need
+# to pass different flags on MSVC vs GCC-compatible toolchains to enable
+# x86 SIMD extensions.
 selects.config_setting_group(
-    name = "x86_64",
+    name = "x86_64_and_not_msvc",
     match_any = [
         ":x86_64_k8",
         ":x86_64_haswell",
@@ -68,12 +71,28 @@ config_setting(
 )
 
 config_setting(
-    name = "optimized",
+    name = "dbg_build",
     values = {
-        "compilation_mode": "opt",
+        "compilation_mode": "dbg",
     },
 )
 
+config_setting(
+    name = "fastbuild_build",
+    values = {
+        "compilation_mode": "fastbuild",
+    },
+)
+
+selects.config_setting_group(
+    name = "do_not_want_O3",
+    match_any = [
+        "@bazel_tools//src/conditions:windows_msvc",
+        ":dbg_build",
+        ":fastbuild_build",
+    ],
+)
+
 cc_library(
     name = "platform",
     hdrs = ["platform.h"],
diff --git a/ruy/build_defs.bzl b/ruy/build_defs.bzl
index e2fc325..836f47a 100644
--- a/ruy/build_defs.bzl
+++ b/ruy/build_defs.bzl
@@ -46,31 +46,34 @@ def ruy_copts_optimize():
         # optimizing for speed is the better compromise, so we override that.
         # Careful to keep debug builds debuggable, whence the select based
         # on the compilation mode.
-        "//ruy:optimized": ["-O3"],
-        "//conditions:default": [],
+        "//ruy:do_not_want_O3": [],
+        "//conditions:default": ["-O3"],
     })
 
 # Returns compiler flags to use for all ruy code.
 def ruy_copts():
     return ruy_copts_warnings() + ruy_copts_neon() + ruy_copts_optimize()
 
-def ruy_copts_avx512():
-    # In some clang-based toolchains, in the default compilation mode (not -c opt),
-    # heavy spillage in the AVX512 kernels results in stack frames > 50k. This issue does not exist
-    # in optimized builds (-c opt).
+def ruy_copts_avx():
     return select({
-        "//ruy:x86_64": ["$(STACK_FRAME_UNLIMITED)", "-mavx512f", "-mavx512vl", "-mavx512cd", "-mavx512bw", "-mavx512dq"],
+        "//ruy:x86_64_and_not_msvc": ["-mavx"],
+        "@bazel_tools//src/conditions:windows_msvc": ["/arch:AVX"],
         "//conditions:default": [],
     })
 
-def ruy_copts_avx():
+def ruy_copts_avx2_fma():
     return select({
-        "//ruy:x86_64": ["-mavx"],
+        "//ruy:x86_64_and_not_msvc": ["-mavx2", "-mfma"],
+        "@bazel_tools//src/conditions:windows_msvc": ["/arch:AVX2"],
         "//conditions:default": [],
     })
 
-def ruy_copts_avx2_fma():
+def ruy_copts_avx512():
+    # In some clang-based toolchains, in the default compilation mode (not -c opt),
+    # heavy spillage in the AVX512 kernels results in stack frames > 50k. This issue does not exist
+    # in optimized builds (-c opt).
     return select({
-        "//ruy:x86_64": ["-mavx2", "-mfma"],
+        "//ruy:x86_64_and_not_msvc": ["$(STACK_FRAME_UNLIMITED)", "-mavx512f", "-mavx512vl", "-mavx512cd", "-mavx512bw", "-mavx512dq"],
+        "@bazel_tools//src/conditions:windows_msvc": ["/arch:AVX512"],
         "//conditions:default": [],
     })
diff --git a/ruy/platform.h b/ruy/platform.h
index 942feac..ffffeb1 100644
--- a/ruy/platform.h
+++ b/ruy/platform.h
@@ -137,7 +137,7 @@ limitations under the License.
 #endif
 
 #if RUY_PLATFORM_X86_ENHANCEMENTS && RUY_PLATFORM_X86 && defined(__AVX2__) && \
-    defined(__FMA__)
+    (defined(__FMA__) || defined(_MSC_VER))
 #define RUY_PLATFORM_AVX2_FMA 1
 #else
 #define RUY_PLATFORM_AVX2_FMA 0