[libc][math][c23] Optimize fabsf16 on x86 with Clang #104869
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Works around optimizations introduced in LLVM 17 and 18 that slow down `fputil::abs<float16>()` on x86.
|
@llvm/pr-subscribers-libc Author: OverMighty (overmighty) ChangesWorks around optimizations introduced in LLVM 17 and 18 that slow down Full diff: https://github.com/llvm/llvm-project/pull/104869.diff 2 Files Affected:
diff --git a/libc/src/math/generic/CMakeLists.txt b/libc/src/math/generic/CMakeLists.txt
index 350072f4b9649..7fa86f17269f2 100644
--- a/libc/src/math/generic/CMakeLists.txt
+++ b/libc/src/math/generic/CMakeLists.txt
@@ -563,10 +563,12 @@ add_entrypoint_object(
HDRS
../fabsf16.h
DEPENDS
- libc.src.__support.macros.properties.types
libc.src.__support.FPUtil.basic_operations
+ libc.src.__support.FPUtil.fp_bits
libc.src.__support.macros.properties.architectures
libc.src.__support.macros.properties.compiler
+ libc.src.__support.macros.properties.cpu_features
+ libc.src.__support.macros.properties.types
COMPILE_OPTIONS
-O3
FLAGS
diff --git a/libc/src/math/generic/fabsf16.cpp b/libc/src/math/generic/fabsf16.cpp
index 02e11330db718..a86aa0cb00a73 100644
--- a/libc/src/math/generic/fabsf16.cpp
+++ b/libc/src/math/generic/fabsf16.cpp
@@ -8,19 +8,30 @@
#include "src/math/fabsf16.h"
#include "src/__support/FPUtil/BasicOperations.h"
+#include "src/__support/FPUtil/FPBits.h"
#include "src/__support/common.h"
#include "src/__support/macros/config.h"
#include "src/__support/macros/properties/architectures.h"
#include "src/__support/macros/properties/compiler.h"
+#include "src/__support/macros/properties/cpu_features.h"
namespace LIBC_NAMESPACE_DECL {
LLVM_LIBC_FUNCTION(float16, fabsf16, (float16 x)) {
- // For x86, GCC generates better code from the generic implementation.
- // https://godbolt.org/z/K9orM4hTa
#if defined(__LIBC_MISC_MATH_BASIC_OPS_OPT) && \
- !(defined(LIBC_TARGET_ARCH_IS_X86) && defined(LIBC_COMPILER_IS_GCC))
+ defined(LIBC_TARGET_CPU_HAS_FAST_FLOAT16_OPS)
return __builtin_fabsf16(x);
+#elif defined(LIBC_TARGET_ARCH_IS_X86) && defined(LIBC_COMPILER_IS_CLANG)
+ // Prevent Clang from generating calls to slow soft-float conversion
+ // functions on x86. See https://godbolt.org/z/hvo6jbnGz.
+
+ using FPBits = fputil::FPBits<float16>;
+ using StorageType = typename FPBits::StorageType;
+
+ static constexpr volatile StorageType ABS_MASK = FPBits::EXP_SIG_MASK;
+
+ return FPBits(static_cast<StorageType>(FPBits(x).uintval() & ABS_MASK))
+ .get_val();
#else
return fputil::abs(x);
#endif
|
lntue
approved these changes
Aug 19, 2024
|
Minimum latencies: Before:
After:
If we enabled the same code path for other compilers/targets: Before:
After:
|
|
arsenm
reviewed
Aug 20, 2024
| return __builtin_fabsf16(x); | ||
| #elif defined(LIBC_TARGET_ARCH_IS_X86) && defined(LIBC_COMPILER_IS_CLANG) | ||
| // Prevent Clang from generating calls to slow soft-float conversion | ||
| // functions on x86. See https://godbolt.org/z/hvo6jbnGz. |
There was a problem hiding this comment.
This is just a codegen bug. it's wrong to introduce a canonicalizing extend when legalizing fabs. Don't workaround it here?
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Works around optimizations introduced in LLVM 17 and 18 that slow down
fputil::abs<float16>()on x86.