Set the stack alignment to 64 bytes for x86 microkernel builds.

This makes register spilling a _lot_ faster in some cases. PiperOrigin-RevId: 719315622
google · Jan 24, 2025 · dc1f822 · dc1f822
1 parent a108468
commit dc1f822
Show file tree

Hide file tree

Showing 3 changed files with 48 additions and 0 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -965,6 +965,13 @@ IF(XNNPACK_TARGET_PROCESSOR MATCHES "^x86(_64)?$")
       SET_PROPERTY(SOURCE ${ALL_AVX512AMX_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -clang:-mf16c -clang:-mfma -clang:-mavx512f -clang:-mavx512cd -clang:-mavx512bw -clang:-mavx512dq -clang:-mavx512vl -clang:-mavx512vnni -clang:-mgfni -clang:-mamx-tile -clang:-mamx-int8 ")
     ENDIF()
   ELSE()
+    IF(CMAKE_C_COMPILER_ID STREQUAL "GNU")
+      SET_PROPERTY(SOURCE ${ALL_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -mpreferred-stack-boundary=6 -fomit-frame-pointer ")
+    ELSEIF(CMAKE_C_COMPILER_ID STREQUAL "Clang")
+      SET_PROPERTY(SOURCE ${ALL_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -mstack-alignment=64 -fomit-frame-pointer ")
+    ELSE()
+      SET_PROPERTY(SOURCE ${ALL_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -fomit-frame-pointer ")
+    ENDIF()
     SET_PROPERTY(SOURCE ${ALL_SSE_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -msse -mno-sse2 ")
     SET_PROPERTY(SOURCE ${ALL_SSE2_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -msse2 -mno-sse3 ")
     SET_PROPERTY(SOURCE ${ALL_SSSE3_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -mssse3 -mno-sse4.1 ")

diff --git a/build_config/BUILD.bazel b/build_config/BUILD.bazel
@@ -262,6 +262,20 @@ config_setting(
     },
 )
 
+config_setting(
+    name = "clang",
+    flag_values = {
+        "@bazel_tools//tools/cpp:compiler": "clang",
+    },
+)
+
+config_setting(
+    name = "gcc",
+    flag_values = {
+        "@bazel_tools//tools/cpp:compiler": "gcc",
+    },
+)
+
 ######################## Composite build configurations #######################
 
 selects.config_setting_group(
@@ -334,3 +348,19 @@ selects.config_setting_group(
     name = "riscv",
     match_any = [":linux_riscv64"],
 )
+
+selects.config_setting_group(
+    name = "x86_clang",
+    match_all =  = [
+        ":x86",
+        ":clang",
+    ],
+)
+
+selects.config_setting_group(
+    name = "x86_gcc",
+    match_all =  = [
+        ":x86",
+        ":gcc",
+    ],
+)
diff --git a/build_params.bzl b/build_params.bzl
@@ -209,6 +209,17 @@ def xnnpack_cc_library_for_arch(arch, name, deps = None, **kwargs):
         else:
             kwargs[key] = val
 
+    # Add extra copts.
+    copts = select({
+        "//build_config:x86_clang": ["-mstack-algiment=64", "-fomit-frame-pointer"],
+        "//build_config:x86_gcc": ["-mpreferred-stack-boundary=6", "-fomit-frame-pointer"],
+        "//conditions:default": [],
+    })
+    if "copts" in kwargs:
+        kwargs["copts"] += copts
+    else:
+        kwargs["copts"] = copts
+
     xnnpack_cc_library(
         name = name,
         target_compatible_with = xnnpack_select_if(