[libc][NFC] Detect host CPU features using try_compile instead of try_run.

[gchatelet]

Summary

This implements the same behavior as D141997 but makes sure that the same detection mechanism is used between CMake and source code.

[gchatelet]

I have two design questions:

• Do we want to merge "architectures.h" and "cpu_features.h" since they usually are used together?
• Do we want to autogenerate the C files since they are not quite minimal?

[lntue]

In D142108#4065441, @gchatelet wrote:

I have two design questions:

• Do we want to merge "architectures.h" and "cpu_features.h" since they usually are used together?
• Do we want to autogenerate the C files since they are not quite minimal?

I think it is good to have "architectures.h" and "cpu_features.h" as 2 separate files, but maybe we can require user to only include one of them by like including "cpu_features.h" at the end of "architectures.h" (or the other way around).
About autogenerating those C files, for now I don't think we need it yet, as the number of CPU features are quite minimal and manageable. So for a few CPU features that we care right now, explicit check files are clearer IMO, and I don't expect them to be too many.

libc/cmake/modules/LLVMLibCCheckCpuFeatures.cmake

60

Look like SOURCES argument needs to be right after output directory.

[sivachandra]

Mostly LGTM but a few minor comments.

I don't see a CMakeLists.txt change and a corresponding Bazel change for the new header file.

In D142108#4065441, @gchatelet wrote:

• Do we want to merge "architectures.h" and "cpu_features.h" since they usually are used together?

I like the separation.

• Do we want to autogenerate the C files since they are not quite minimal?

I think the explicit listing is probably a better approach. As this get more complicated, the CMake scripting will get more complicated. I prefer simpler CMake scripting over increased verbosity/repetitiveness.

libc/cmake/modules/cpu_features/check_AVX2.c

1 ↗

(On Diff #490484)

Nit: These checker sources are including a .h which is marked C++. Also, we need a C++ compiler anyway for the building the libc anyway. So, why not make these files have a .cpp extension?

libc/src/__support/cpu_features.h

9

This comment seems incomplete; Perhaps, "This file **lists** target CPU features ..." ?

[gchatelet]

Address comments

[gchatelet]

Thx for the review.

I think the explicit listing is probably a better approach. As this get more complicated, the CMake scripting will get more complicated. I prefer simpler CMake scripting over increased verbosity/repetitiveness.

I fail to understand this comment. The first part seems to suggest to keep the patch as-is but AFAIU the second part suggests to use CMake scripting.

[sivachandra]

In D142108#4068362, @gchatelet wrote:

I think the explicit listing is probably a better approach. As this get more complicated, the CMake scripting will get more complicated. I prefer simpler CMake scripting over increased verbosity/repetitiveness.

I fail to understand this comment. The first part seems to suggest to keep the patch as-is but AFAIU the second part suggests to use CMake scripting.

No, the second part is kind of reasoning out why I prefer keeping the patch as is - to keep the CMake scripting simpler.

[gchatelet]

In D142108#4069139, @sivachandra wrote:

In D142108#4068362, @gchatelet wrote:

I think the explicit listing is probably a better approach. As this get more complicated, the CMake scripting will get more complicated. I prefer simpler CMake scripting over increased verbosity/repetitiveness.

I fail to understand this comment. The first part seems to suggest to keep the patch as-is but AFAIU the second part suggests to use CMake scripting.

No, the second part is kind of reasoning out why I prefer keeping the patch as is - to keep the CMake scripting simpler.

👍
Submitting then.

[gchatelet]

For a reason that I don't understand yet, the build bots are not detecting the cpu features.
https://lab.llvm.org/buildbot/#/builders/90/builds/44634/steps/4/logs/stdio

-- Set COMPILER_RESOURCE_DIR to /usr/lib/llvm-11/lib/clang/11.0.1 using --print-resource-dir
-- Building libc for x86_64 on linux
-- Set CPU features: 
-- Skipping libc entrypoint libc.src.stdlib.getenv.
-- Skipping libc entrypoint libc.src.stdio.fopen.

@sivachandra is there a way to log onto one of the build bots to try to reproduce the issue?
I haven't been able to reproduce it locally.

libc/cmake/modules/LLVMLibCCheckCpuFeatures.cmake

function(_intersection output_var list1 list2)
   foreach(element IN LISTS list1)
     if("${list2}" MATCHES "(^|;)${element}(;|$)")
       list(APPEND tmp "${element}")
     endif()
   endforeach()
   set(${output_var} ${tmp} PARENT_SCOPE)
 endfunction()
 
-# Generates a cpp file to introspect the compiler defined flags.
-function(_generate_check_code)
+set(AVAILABLE_CPU_FEATURES "")
+if(LIBC_CROSSBUILD)
+  # If we are doing a cross build, we will just assume that all CPU features
+  # are available.
+  set(AVAILABLE_CPU_FEATURES ${ALL_CPU_FEATURES})
+else()
+  # Try compile a C file to check if flag is supported.
+  set(CMAKE_TRY_COMPILE_TARGET_TYPE STATIC_LIBRARY)
   foreach(feature IN LISTS ALL_CPU_FEATURES)
-    set(DEFINITIONS
-        "${DEFINITIONS}
-#ifdef __${feature}__
-    \"${feature}\",
-#endif")
+    try_compile(
+      has_feature
+      ${CMAKE_CURRENT_BINARY_DIR}/cpu_features
+      SOURCES ${LIBC_SOURCE_DIR}/cmake/modules/cpu_features/check_${feature}.cpp
+      COMPILE_DEFINITIONS -I${LIBC_SOURCE_DIR} ${LIBC_COMPILE_OPTIONS_NATIVE}
+    )
+    if(has_feature)
+      list(APPEND AVAILABLE_CPU_FEATURES ${feature})
+    endif()
   endforeach()
-  configure_file(
-    "${LIBC_SOURCE_DIR}/cmake/modules/cpu_features/check_cpu_features.cpp.in"
-    "cpu_features/check_cpu_features.cpp" @ONLY)
-endfunction()
-_generate_check_code()
+endif()
 
-set(LIBC_CPU_FEATURES "" CACHE PATH "Host supported CPU features")
+set(LIBC_CPU_FEATURES ${AVAILABLE_CPU_FEATURES} CACHE STRING "Host supported CPU features")
 
-if(LIBC_CROSSBUILD)
-  _intersection(cpu_features "${ALL_CPU_FEATURES}" "${LIBC_CPU_FEATURES}")
-  if(NOT "${cpu_features}" STREQUAL "${LIBC_CPU_FEATURES}")
-    message(FATAL_ERROR "Unsupported CPU features: ${cpu_features}")
-  endif()
-  message(STATUS "Set CPU features: ${cpu_features}")
-  set(LIBC_CPU_FEATURES "${cpu_features}")
-else()
-  # Populates the LIBC_CPU_FEATURES list from host.
-  try_run(
-    run_result compile_result "${CMAKE_CURRENT_BINARY_DIR}/check_${feature}"
-    "${CMAKE_CURRENT_BINARY_DIR}/cpu_features/check_cpu_features.cpp"
-    COMPILE_DEFINITIONS ${LIBC_COMPILE_OPTIONS_NATIVE}
-    COMPILE_OUTPUT_VARIABLE compile_output
-    RUN_OUTPUT_VARIABLE run_output)
-  if("${run_result}" EQUAL 0)
-    message(STATUS "Set CPU features: ${run_output}")
-    set(LIBC_CPU_FEATURES "${run_output}")
-  elseif(NOT ${compile_result})
-    message(FATAL_ERROR "Failed to compile: ${compile_output}")
-  else()
-    message(FATAL_ERROR "Failed to run: ${run_output}")
-  endif()
+_intersection(cpu_features "${AVAILABLE_CPU_FEATURES}" "${LIBC_CPU_FEATURES}")
+  message(STATUS "Set CPU features: ${cpu_features}")
 endif()

libc/cmake/modules/cpu_features/check_AVX2.cpp

+#include "src/__support/cpu_features.h"
+
+#ifndef LIBC_TARGET_HAS_AVX2
+#error unsupported
+#endif

libc/cmake/modules/cpu_features/check_AVX512BW.cpp

+#include "src/__support/cpu_features.h"
+
+#ifndef LIBC_TARGET_HAS_AVX512BW
+#error unsupported
+#endif

libc/cmake/modules/cpu_features/check_AVX512F.cpp

+#include "src/__support/cpu_features.h"
+
+#ifndef LIBC_TARGET_HAS_AVX512F
+#error unsupported
+#endif

libc/cmake/modules/cpu_features/check_FMA.cpp

+#include "src/__support/cpu_features.h"
+
+#ifndef LIBC_TARGET_HAS_FMA
+#error unsupported
+#endif

libc/cmake/modules/cpu_features/check_SSE2.cpp

+#include "src/__support/cpu_features.h"
+
+#ifndef LIBC_TARGET_HAS_SSE2
+#error unsupported
+#endif

libc/cmake/modules/cpu_features/check_SSE4_2.cpp

+#include "src/__support/cpu_features.h"
+
+#ifndef LIBC_TARGET_HAS_SSE4_2
+#error unsupported
+#endif

libc/cmake/modules/cpu_features/check_cpu_features.cpp.in

-#include <stdio.h>
-#include <stdlib.h>
-
-// This file is instantiated by CMake.
-// DEFINITIONS below is replaced with a set of lines like so:
-//   #ifdef __SSE2__
-//     "SSE2",
-//   #endif
-//
-// This allows for introspection of compiler definitions.
-// The output of the program is a single line of semi colon separated feature
-// names.
-
-// MSVC is using a different set of preprocessor definitions for 
-// SSE and SSE2, see _M_IX86_FP in
-// https://docs.microsoft.com/en-us/cpp/preprocessor/predefined-macros
-
-int main(int, char **) {
-  const char *strings[] = {
-      @DEFINITIONS@
-      // If DEFINITIONS turns out to be empty, we don't want to list
-      // an empty array. So, we add an end of list marker.
-      "<end_of_feature_list>"
-  };
-  const size_t size = sizeof(strings) / sizeof(strings[0]);
-  for (size_t i = 0; i < size; ++i) {
-    if (i)
-      putchar(';');
-    fputs(strings[i], stdout);
-  }
-  return EXIT_SUCCESS;
-}

libc/src/__support/CMakeLists.txt

DEPENDS
     libc.src.__support.compiler_features
 )
 
 add_header_library(
   common
   HDRS
     architectures.h
     common.h
+    cpu_features.h
     endian.h
 )
 
 add_header_library(
   builtin_wrappers
   HDRS
     builtin_wrappers.h
   DEPENDS

libc/src/__support/FPUtil/FMA.h

 //===-- Common header for FMA implementations -------------------*- C++ -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_LIBC_SRC_SUPPORT_FPUTIL_FMA_H
 #define LLVM_LIBC_SRC_SUPPORT_FPUTIL_FMA_H
 
 #include "src/__support/architectures.h"
 #include "src/__support/common.h"
+#include "src/__support/cpu_features.h"
 
 #if defined(LIBC_TARGET_HAS_FMA)
 
 #if defined(LLVM_LIBC_ARCH_X86_64)
 #include "x86_64/FMA.h"
 #elif defined(LLVM_LIBC_ARCH_AARCH64)
 #include "aarch64/FMA.h"
 #endif

libc/src/__support/FPUtil/aarch64/FMA.h

 //===-- Aarch64 implementations of the fma function -------------*- C++ -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_LIBC_SRC_SUPPORT_FPUTIL_AARCH64_FMA_H
 #define LLVM_LIBC_SRC_SUPPORT_FPUTIL_AARCH64_FMA_H
 
 #include "src/__support/architectures.h"
+#include "src/__support/cpu_features.h"
 
 #if !defined(LLVM_LIBC_ARCH_AARCH64)
 #error "Invalid include"
 #endif
 
 #if !defined(LIBC_TARGET_HAS_FMA)
 #error "FMA instructions are not supported"
 #endif

libc/src/__support/FPUtil/multiply_add.h

 //===-- Common header for multiply-add implementations ----------*- C++ -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_LIBC_SRC_SUPPORT_FPUTIL_MULTIPLY_ADD_H
 #define LLVM_LIBC_SRC_SUPPORT_FPUTIL_MULTIPLY_ADD_H
 
 #include "src/__support/architectures.h"
 #include "src/__support/common.h"
+#include "src/__support/cpu_features.h"
 
 namespace __llvm_libc {
 namespace fputil {
 
 // Implement a simple wrapper for multiply-add operation:
 //   multiply_add(x, y, z) = x*y + z
 // which uses FMA instructions to speed up if available.
 

libc/src/__support/FPUtil/x86_64/FMA.h

 //===-- x86_64 implementations of the fma function --------------*- C++ -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_LIBC_SRC_SUPPORT_FPUTIL_X86_64_FMA_H
 #define LLVM_LIBC_SRC_SUPPORT_FPUTIL_X86_64_FMA_H
 
 #include "src/__support/architectures.h"
 #include "src/__support/common.h"
+#include "src/__support/cpu_features.h"
 
 #if !defined(LLVM_LIBC_ARCH_X86_64)
 #error "Invalid include"
 #endif
 
 #if !defined(LIBC_TARGET_HAS_FMA)
 #error "FMA instructions are not supported"
 #endif

libc/src/__support/architectures.h

 #if defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64)
 #define LLVM_LIBC_ARCH_AARCH64
 #endif
 
 #if (defined(LLVM_LIBC_ARCH_AARCH64) || defined(LLVM_LIBC_ARCH_ARM))
 #define LLVM_LIBC_ARCH_ANY_ARM
 #endif
 
-#if defined(LLVM_LIBC_ARCH_AARCH64)
-#define LIBC_TARGET_HAS_FMA
-#elif defined(LLVM_LIBC_ARCH_X86_64)
-#if (defined(__AVX2__) || defined(__FMA__))
-#define LIBC_TARGET_HAS_FMA
-#endif
-#endif
-
 #endif // LLVM_LIBC_SUPPORT_ARCHITECTURES_H

libc/src/__support/cpu_features.h

+//===-- Compile time cpu feature detection ----------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+// This file lists target cpu features by introspecting compiler enabled
+// preprocessor definitions.
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_LIBC_SRC_SUPPORT_CPU_FEATURES_H
+#define LLVM_LIBC_SRC_SUPPORT_CPU_FEATURES_H
+
+#if defined(__SSE2__)
+#define LIBC_TARGET_HAS_SSE2
+#endif
+
+#if defined(__SSE4_2__)
+#define LIBC_TARGET_HAS_SSE4_2
+#endif
+
+#if defined(__AVX__)
+#define LIBC_TARGET_HAS_AVX
+#endif
+
+#if defined(__AVX2__)
+#define LIBC_TARGET_HAS_AVX2
+#endif
+
+#if defined(__AVX512F__)
+#define LIBC_TARGET_HAS_AVX512F
+#endif
+
+#if defined(__AVX512BW__)
+#define LIBC_TARGET_HAS_AVX512BW
+#endif
+
+#if defined(__ARM_FEATURE_FMA) || defined(__AVX2__)
+#define LIBC_TARGET_HAS_FMA
+#endif
+
+#endif // LLVM_LIBC_SRC_SUPPORT_CPU_FEATURES_H

libc/src/math/generic/asinf.cpp

 //===-- Single-precision asin function ------------------------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #include "src/math/asinf.h"
 #include "src/__support/FPUtil/FEnvImpl.h"
 #include "src/__support/FPUtil/FPBits.h"
 #include "src/__support/FPUtil/PolyEval.h"
 #include "src/__support/FPUtil/except_value_utils.h"
 #include "src/__support/FPUtil/multiply_add.h"
 #include "src/__support/FPUtil/sqrt.h"
+#include "src/__support/cpu_features.h"
 
 #include <errno.h>
 
 #include "inv_trigf_utils.h"
 
 namespace __llvm_libc {
 
 static constexpr size_t N_EXCEPTS = 2;

libc/src/math/generic/cosf.cpp

 //===-- Single-precision cos function -------------------------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #include "src/math/cosf.h"
 #include "sincosf_utils.h"
 #include "src/__support/FPUtil/BasicOperations.h"
 #include "src/__support/FPUtil/FEnvImpl.h"
 #include "src/__support/FPUtil/FPBits.h"
 #include "src/__support/FPUtil/except_value_utils.h"
 #include "src/__support/FPUtil/multiply_add.h"
 #include "src/__support/common.h"
+#include "src/__support/cpu_features.h"
 
 #include <errno.h>
 
 namespace __llvm_libc {
 
 // Exceptional cases for cosf.
 static constexpr size_t N_EXCEPTS = 6;
 

libc/src/math/generic/expm1f.cpp

 #include "src/__support/FPUtil/BasicOperations.h"
 #include "src/__support/FPUtil/FEnvImpl.h"
 #include "src/__support/FPUtil/FMA.h"
 #include "src/__support/FPUtil/FPBits.h"
 #include "src/__support/FPUtil/PolyEval.h"
 #include "src/__support/FPUtil/multiply_add.h"
 #include "src/__support/FPUtil/nearest_integer.h"
 #include "src/__support/common.h"
+#include "src/__support/cpu_features.h"
 
 #include <errno.h>
 
 namespace __llvm_libc {
 
 LLVM_LIBC_FUNCTION(float, expm1f, (float x)) {
   using FPBits = typename fputil::FPBits<float>;
   FPBits xbits(x);

libc/src/math/generic/sincosf.cpp

 //===-- Single-precision sincos function ----------------------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #include "src/math/sincosf.h"
 #include "sincosf_utils.h"
 #include "src/__support/FPUtil/FEnvImpl.h"
 #include "src/__support/FPUtil/FPBits.h"
 #include "src/__support/FPUtil/multiply_add.h"
 #include "src/__support/common.h"
+#include "src/__support/cpu_features.h"
 
 #include <errno.h>
 
 namespace __llvm_libc {
 
 // Exceptional values
 static constexpr int N_EXCEPTS = 6;
 

libc/src/math/generic/sincosf_utils.h

 //===-- Collection of utils for sinf/cosf/sincosf ---------------*- C++ -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_LIBC_SRC_MATH_SINCOSF_UTILS_H
 #define LLVM_LIBC_SRC_MATH_SINCOSF_UTILS_H
 
 #include "src/__support/FPUtil/FPBits.h"
 #include "src/__support/FPUtil/PolyEval.h"
 #include "src/__support/common.h"
+#include "src/__support/cpu_features.h"
 
 #if defined(LIBC_TARGET_HAS_FMA)
 #include "range_reduction_fma.h"
 // using namespace __llvm_libc::fma;
 using __llvm_libc::fma::FAST_PASS_BOUND;
 using __llvm_libc::fma::large_range_reduction;
 using __llvm_libc::fma::small_range_reduction;
 #else

libc/src/math/generic/sinf.cpp

 //===-- Single-precision sin function -------------------------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #include "src/math/sinf.h"
 #include "sincosf_utils.h"
 #include "src/__support/FPUtil/BasicOperations.h"
 #include "src/__support/FPUtil/FEnvImpl.h"
 #include "src/__support/FPUtil/FPBits.h"
 #include "src/__support/FPUtil/PolyEval.h"
 #include "src/__support/FPUtil/multiply_add.h"
 #include "src/__support/common.h"
+#include "src/__support/cpu_features.h"
 
 #include <errno.h>
 
 #if defined(LIBC_TARGET_HAS_FMA)
 #include "range_reduction_fma.h"
 #else
 #include "range_reduction.h"
 #endif

libc/src/math/generic/tanf.cpp

 #include "sincosf_utils.h"
 #include "src/__support/FPUtil/FEnvImpl.h"
 #include "src/__support/FPUtil/FPBits.h"
 #include "src/__support/FPUtil/PolyEval.h"
 #include "src/__support/FPUtil/except_value_utils.h"
 #include "src/__support/FPUtil/multiply_add.h"
 #include "src/__support/FPUtil/nearest_integer.h"
 #include "src/__support/common.h"
+#include "src/__support/cpu_features.h"
 
 #include <errno.h>
 
 namespace __llvm_libc {
 
 // Exceptional cases for tanf.
 constexpr size_t N_EXCEPTS = 6;
 

libc/src/math/generic/tanhf.cpp

 //===-- Single-precision tanh function ------------------------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #include "src/math/tanhf.h"
 #include "src/__support/FPUtil/FPBits.h"
+#include "src/__support/cpu_features.h"
 #include "src/math/generic/explogxf.h"
 
 namespace __llvm_libc {
 
 LLVM_LIBC_FUNCTION(float, tanhf, (float x)) {
   using FPBits = typename fputil::FPBits<float>;
   FPBits xbits(x);
   bool sign = xbits.get_sign();

utils/bazel/llvm-project-overlay/libc/BUILD.bazel

libc_support_library(
     ],
 )
 
 libc_support_library(
     name = "__support_common",
     hdrs = [
         "src/__support/architectures.h",
         "src/__support/common.h",
+        "src/__support/cpu_features.h",
         "src/__support/endian.h",
     ],
 )
 
 libc_support_library(
     name = "__support_cpp_array",
     hdrs = ["src/__support/CPP/array.h"],
     deps = [":libc_root"],