[libc][math] Optimize nearest integer functions using builtins when available #98376
Conversation
|
cc @lntue |
|
@llvm/pr-subscribers-libc Author: OverMighty (overmighty) ChangesPatch is 34.78 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/98376.diff 27 Files Affected:
diff --git a/libc/cmake/modules/CheckCompilerFeatures.cmake b/libc/cmake/modules/CheckCompilerFeatures.cmake
index d84c07b35d2d7..83822892c8096 100644
--- a/libc/cmake/modules/CheckCompilerFeatures.cmake
+++ b/libc/cmake/modules/CheckCompilerFeatures.cmake
@@ -2,7 +2,15 @@
# Compiler features definition and flags
# ------------------------------------------------------------------------------
-set(ALL_COMPILER_FEATURES "float16" "float128" "fixed_point")
+set(
+ ALL_COMPILER_FEATURES
+ "builtin_ceil_floor_trunc"
+ "builtin_round"
+ "builtin_roundeven"
+ "float16"
+ "float128"
+ "fixed_point"
+)
# Making sure ALL_COMPILER_FEATURES is sorted.
list(SORT ALL_COMPILER_FEATURES)
@@ -39,11 +47,19 @@ endfunction()
set(AVAILABLE_COMPILER_FEATURES "")
# Try compile a C file to check if flag is supported.
-set(CMAKE_TRY_COMPILE_TARGET_TYPE STATIC_LIBRARY)
foreach(feature IN LISTS ALL_COMPILER_FEATURES)
+ set(CMAKE_TRY_COMPILE_TARGET_TYPE STATIC_LIBRARY)
set(compile_options ${LIBC_COMPILE_OPTIONS_NATIVE})
if(${feature} STREQUAL "fixed_point")
list(APPEND compile_options "-ffixed-point")
+ elseif(${feature} MATCHES "^builtin_")
+ set(compile_options ${LIBC_COMPILE_OPTIONS_DEFAULT})
+ # The compiler might handle calls to rounding builtins by generating calls
+ # to the respective libc math functions, in which case we cannot use these
+ # builtins in our implementations of these functions. We check that this is
+ # not the case by trying to link an executable, since linking would fail due
+ # to unresolved references if calls to libc functions were generated.
+ set(CMAKE_TRY_COMPILE_TARGET_TYPE EXECUTABLE)
endif()
try_compile(
@@ -60,6 +76,12 @@ foreach(feature IN LISTS ALL_COMPILER_FEATURES)
set(LIBC_TYPES_HAS_FLOAT128 TRUE)
elseif(${feature} STREQUAL "fixed_point")
set(LIBC_COMPILER_HAS_FIXED_POINT TRUE)
+ elseif(${feature} STREQUAL "builtin_ceil_floor_trunc")
+ set(LIBC_COMPILER_HAS_BUILTIN_CEIL_FLOOR_TRUNC TRUE)
+ elseif(${feature} STREQUAL "builtin_round")
+ set(LIBC_COMPILER_HAS_BUILTIN_ROUND TRUE)
+ elseif(${feature} STREQUAL "builtin_roundeven")
+ set(LIBC_COMPILER_HAS_BUILTIN_ROUNDEVEN TRUE)
endif()
endif()
endforeach()
diff --git a/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake b/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake
index c5e7dfe8abd0f..855d69d2a0fc9 100644
--- a/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake
+++ b/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake
@@ -4,7 +4,7 @@ function(_get_compile_options_from_flags output_var)
if(LIBC_TARGET_ARCHITECTURE_IS_RISCV64 OR(LIBC_CPU_FEATURES MATCHES "FMA"))
check_flag(ADD_FMA_FLAG ${FMA_OPT_FLAG} ${ARGN})
endif()
- check_flag(ADD_SSE4_2_FLAG ${ROUND_OPT_FLAG} ${ARGN})
+ check_flag(ADD_ROUND_OPT_FLAG ${ROUND_OPT_FLAG} ${ARGN})
check_flag(ADD_EXPLICIT_SIMD_OPT_FLAG ${EXPLICIT_SIMD_OPT_FLAG} ${ARGN})
if(LLVM_COMPILER_IS_GCC_COMPATIBLE)
@@ -16,8 +16,22 @@ function(_get_compile_options_from_flags output_var)
list(APPEND compile_options "-D__LIBC_RISCV_USE_FMA")
endif()
endif()
- if(ADD_SSE4_2_FLAG)
- list(APPEND compile_options "-msse4.2")
+ if(ADD_ROUND_OPT_FLAG)
+ if(LIBC_TARGET_ARCHITECTURE_IS_X86)
+ # ROUND_OPT_FLAG is only enabled if SSE4.2 is detected, not just SSE4.1,
+ # because there was code to check for SSE4.2 already, and few CPUs only
+ # have SSE4.1.
+ list(APPEND compile_options "-msse4.2")
+ endif()
+ if(LIBC_COMPILER_HAS_BUILTIN_CEIL_FLOOR_TRUNC)
+ list(APPEND compile_options "-D__LIBC_USE_BUILTIN_CEIL_FLOOR_TRUNC")
+ endif()
+ if(LIBC_COMPILER_HAS_BUILTIN_ROUND)
+ list(APPEND compile_options "-D__LIBC_USE_BUILTIN_ROUND")
+ endif()
+ if(LIBC_COMPILER_HAS_BUILTIN_ROUNDEVEN)
+ list(APPEND compile_options "-D__LIBC_USE_BUILTIN_ROUNDEVEN")
+ endif()
endif()
if(ADD_EXPLICIT_SIMD_OPT_FLAG)
list(APPEND compile_options "-D__LIBC_EXPLICIT_SIMD_OPT")
diff --git a/libc/cmake/modules/LLVMLibCFlagRules.cmake b/libc/cmake/modules/LLVMLibCFlagRules.cmake
index 18e36dfde5cc1..eca7ba8d183e6 100644
--- a/libc/cmake/modules/LLVMLibCFlagRules.cmake
+++ b/libc/cmake/modules/LLVMLibCFlagRules.cmake
@@ -277,6 +277,7 @@ if(NOT(LIBC_TARGET_ARCHITECTURE_IS_X86 AND (LIBC_CPU_FEATURES MATCHES "SSE2")))
endif()
# Skip ROUND_OPT flag for targets that don't support SSE 4.2.
-if(NOT(LIBC_TARGET_ARCHITECTURE_IS_X86 AND (LIBC_CPU_FEATURES MATCHES "SSE4_2")))
+if(NOT((LIBC_TARGET_ARCHITECTURE_IS_X86 AND (LIBC_CPU_FEATURES MATCHES "SSE4_2")) OR
+ LIBC_TARGET_ARCHITECTURE_IS_AARCH64))
set(SKIP_FLAG_EXPANSION_ROUND_OPT TRUE)
endif()
diff --git a/libc/cmake/modules/compiler_features/check_builtin_ceil_floor_trunc.cpp b/libc/cmake/modules/compiler_features/check_builtin_ceil_floor_trunc.cpp
new file mode 100644
index 0000000000000..031dd9376f3c1
--- /dev/null
+++ b/libc/cmake/modules/compiler_features/check_builtin_ceil_floor_trunc.cpp
@@ -0,0 +1,9 @@
+float try_builtin_ceilf(float x) { return __builtin_ceilf(x); }
+float try_builtin_floorf(float x) { return __builtin_ceilf(x); }
+float try_builtin_truncf(float x) { return __builtin_truncf(x); }
+
+double try_builtin_ceil(double x) { return __builtin_ceil(x); }
+double try_builtin_floor(double x) { return __builtin_ceil(x); }
+double try_builtin_trunc(double x) { return __builtin_trunc(x); }
+
+int main() {}
diff --git a/libc/cmake/modules/compiler_features/check_builtin_round.cpp b/libc/cmake/modules/compiler_features/check_builtin_round.cpp
new file mode 100644
index 0000000000000..8c3065c2de06a
--- /dev/null
+++ b/libc/cmake/modules/compiler_features/check_builtin_round.cpp
@@ -0,0 +1,5 @@
+float try_builtin_roundf(float x) { return __builtin_roundf(x); }
+
+double try_builtin_round(double x) { return __builtin_round(x); }
+
+int main() {}
diff --git a/libc/cmake/modules/compiler_features/check_builtin_roundeven.cpp b/libc/cmake/modules/compiler_features/check_builtin_roundeven.cpp
new file mode 100644
index 0000000000000..2480abae84c36
--- /dev/null
+++ b/libc/cmake/modules/compiler_features/check_builtin_roundeven.cpp
@@ -0,0 +1,5 @@
+float try_builtin_roundevenf(float x) { return __builtin_roundevenf(x); }
+
+double try_builtin_roundeven(double x) { return __builtin_roundeven(x); }
+
+int main() {}
diff --git a/libc/config/linux/aarch64/entrypoints.txt b/libc/config/linux/aarch64/entrypoints.txt
index 030c3d3a99a02..515c472ef309e 100644
--- a/libc/config/linux/aarch64/entrypoints.txt
+++ b/libc/config/linux/aarch64/entrypoints.txt
@@ -489,6 +489,7 @@ set(TARGET_LIBM_ENTRYPOINTS
libc.src.math.rintl
libc.src.math.round
libc.src.math.roundf
+ libc.src.math.roundevenf
libc.src.math.roundl
libc.src.math.scalbn
libc.src.math.scalbnf
diff --git a/libc/src/__support/FPUtil/NearestIntegerOperations.h b/libc/src/__support/FPUtil/NearestIntegerOperations.h
index cff32938229d0..a9a0a97eebb5c 100644
--- a/libc/src/__support/FPUtil/NearestIntegerOperations.h
+++ b/libc/src/__support/FPUtil/NearestIntegerOperations.h
@@ -75,15 +75,17 @@ LIBC_INLINE T ceil(T x) {
}
uint32_t trim_size = FPBits<T>::FRACTION_LEN - exponent;
- StorageType trunc_mantissa =
- static_cast<StorageType>((bits.get_mantissa() >> trim_size) << trim_size);
- bits.set_mantissa(trunc_mantissa);
- T trunc_value = bits.get_val();
+ StorageType x_u = bits.uintval();
+ StorageType trunc_u =
+ static_cast<StorageType>((x_u >> trim_size) << trim_size);
// If x is already an integer, return it.
- if (trunc_value == x)
+ if (trunc_u == x_u)
return x;
+ bits.set_uintval(trunc_u);
+ T trunc_value = bits.get_val();
+
// If x is negative, the ceil operation is equivalent to the trunc operation.
if (is_neg)
return trunc_value;
@@ -130,15 +132,17 @@ LIBC_INLINE T round(T x) {
uint32_t trim_size = FPBits<T>::FRACTION_LEN - exponent;
bool half_bit_set =
bool(bits.get_mantissa() & (StorageType(1) << (trim_size - 1)));
- StorageType trunc_mantissa =
- static_cast<StorageType>((bits.get_mantissa() >> trim_size) << trim_size);
- bits.set_mantissa(trunc_mantissa);
- T trunc_value = bits.get_val();
+ StorageType x_u = bits.uintval();
+ StorageType trunc_u =
+ static_cast<StorageType>((x_u >> trim_size) << trim_size);
// If x is already an integer, return it.
- if (trunc_value == x)
+ if (trunc_u == x_u)
return x;
+ bits.set_uintval(trunc_u);
+ T trunc_value = bits.get_val();
+
if (!half_bit_set) {
// Franctional part is less than 0.5 so round value is the
// same as the trunc value.
@@ -188,16 +192,17 @@ round_using_specific_rounding_mode(T x, int rnd) {
}
uint32_t trim_size = FPBits<T>::FRACTION_LEN - exponent;
- FPBits<T> new_bits = bits;
- StorageType trunc_mantissa =
- static_cast<StorageType>((bits.get_mantissa() >> trim_size) << trim_size);
- new_bits.set_mantissa(trunc_mantissa);
- T trunc_value = new_bits.get_val();
+ StorageType x_u = bits.uintval();
+ StorageType trunc_u =
+ static_cast<StorageType>((x_u >> trim_size) << trim_size);
// If x is already an integer, return it.
- if (trunc_value == x)
+ if (trunc_u == x_u)
return x;
+ FPBits<T> new_bits(trunc_u);
+ T trunc_value = new_bits.get_val();
+
StorageType trim_value =
bits.get_mantissa() &
static_cast<StorageType>(((StorageType(1) << trim_size) - 1));
diff --git a/libc/src/math/generic/CMakeLists.txt b/libc/src/math/generic/CMakeLists.txt
index 5e920307d39de..915fc076826f9 100644
--- a/libc/src/math/generic/CMakeLists.txt
+++ b/libc/src/math/generic/CMakeLists.txt
@@ -70,6 +70,8 @@ add_entrypoint_object(
-O3
DEPENDS
libc.src.__support.FPUtil.nearest_integer_operations
+ FLAGS
+ ROUND_OPT
)
add_entrypoint_object(
@@ -82,6 +84,8 @@ add_entrypoint_object(
-O3
DEPENDS
libc.src.__support.FPUtil.nearest_integer_operations
+ FLAGS
+ ROUND_OPT
)
add_entrypoint_object(
@@ -107,6 +111,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.macros.properties.types
libc.src.__support.FPUtil.nearest_integer_operations
+ libc.src.__support.macros.properties.architectures
+ FLAGS
+ ROUND_OPT
)
add_entrypoint_object(
@@ -455,6 +462,8 @@ add_entrypoint_object(
-O3
DEPENDS
libc.src.__support.FPUtil.nearest_integer_operations
+ FLAGS
+ ROUND_OPT
)
add_entrypoint_object(
@@ -467,6 +476,8 @@ add_entrypoint_object(
-O3
DEPENDS
libc.src.__support.FPUtil.nearest_integer_operations
+ FLAGS
+ ROUND_OPT
)
add_entrypoint_object(
@@ -492,6 +503,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.macros.properties.types
libc.src.__support.FPUtil.nearest_integer_operations
+ libc.src.__support.macros.properties.architectures
+ FLAGS
+ ROUND_OPT
)
add_entrypoint_object(
@@ -517,6 +531,8 @@ add_entrypoint_object(
-O3
DEPENDS
libc.src.__support.FPUtil.nearest_integer_operations
+ FLAGS
+ ROUND_OPT
)
add_entrypoint_object(
@@ -529,6 +545,8 @@ add_entrypoint_object(
-O3
DEPENDS
libc.src.__support.FPUtil.nearest_integer_operations
+ FLAGS
+ ROUND_OPT
)
add_entrypoint_object(
@@ -554,6 +572,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.macros.properties.types
libc.src.__support.FPUtil.nearest_integer_operations
+ libc.src.__support.macros.properties.architectures
+ FLAGS
+ ROUND_OPT
)
add_entrypoint_object(
@@ -579,6 +600,8 @@ add_entrypoint_object(
-O3
DEPENDS
libc.src.__support.FPUtil.nearest_integer_operations
+ FLAGS
+ ROUND_OPT
)
add_entrypoint_object(
@@ -591,6 +614,8 @@ add_entrypoint_object(
-O3
DEPENDS
libc.src.__support.FPUtil.nearest_integer_operations
+ FLAGS
+ ROUND_OPT
)
add_entrypoint_object(
@@ -616,6 +641,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.macros.properties.types
libc.src.__support.FPUtil.nearest_integer_operations
+ libc.src.__support.macros.properties.architectures
+ FLAGS
+ ROUND_OPT
)
add_entrypoint_object(
@@ -641,6 +669,8 @@ add_entrypoint_object(
-O3
DEPENDS
libc.src.__support.FPUtil.nearest_integer_operations
+ FLAGS
+ ROUND_OPT
)
add_entrypoint_object(
@@ -653,6 +683,8 @@ add_entrypoint_object(
-O3
DEPENDS
libc.src.__support.FPUtil.nearest_integer_operations
+ FLAGS
+ ROUND_OPT
)
add_entrypoint_object(
@@ -678,6 +710,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.macros.properties.types
libc.src.__support.FPUtil.nearest_integer_operations
+ libc.src.__support.macros.properties.architectures
+ FLAGS
+ ROUND_OPT
)
add_entrypoint_object(
diff --git a/libc/src/math/generic/ceil.cpp b/libc/src/math/generic/ceil.cpp
index efd0f246a9b90..63da803033e22 100644
--- a/libc/src/math/generic/ceil.cpp
+++ b/libc/src/math/generic/ceil.cpp
@@ -12,6 +12,12 @@
namespace LIBC_NAMESPACE {
-LLVM_LIBC_FUNCTION(double, ceil, (double x)) { return fputil::ceil(x); }
+LLVM_LIBC_FUNCTION(double, ceil, (double x)) {
+#ifdef __LIBC_USE_BUILTIN_CEIL_FLOOR_TRUNC
+ return __builtin_ceil(x);
+#else
+ return fputil::ceil(x);
+#endif
+}
} // namespace LIBC_NAMESPACE
diff --git a/libc/src/math/generic/ceilf.cpp b/libc/src/math/generic/ceilf.cpp
index d49b34242da4f..51ef68f1dd871 100644
--- a/libc/src/math/generic/ceilf.cpp
+++ b/libc/src/math/generic/ceilf.cpp
@@ -12,6 +12,12 @@
namespace LIBC_NAMESPACE {
-LLVM_LIBC_FUNCTION(float, ceilf, (float x)) { return fputil::ceil(x); }
+LLVM_LIBC_FUNCTION(float, ceilf, (float x)) {
+#ifdef __LIBC_USE_BUILTIN_CEIL_FLOOR_TRUNC
+ return __builtin_ceilf(x);
+#else
+ return fputil::ceil(x);
+#endif
+}
} // namespace LIBC_NAMESPACE
diff --git a/libc/src/math/generic/ceilf16.cpp b/libc/src/math/generic/ceilf16.cpp
index 205d7428f66e6..ee584c25a4ae9 100644
--- a/libc/src/math/generic/ceilf16.cpp
+++ b/libc/src/math/generic/ceilf16.cpp
@@ -9,9 +9,17 @@
#include "src/math/ceilf16.h"
#include "src/__support/FPUtil/NearestIntegerOperations.h"
#include "src/__support/common.h"
+#include "src/__support/macros/properties/architectures.h"
namespace LIBC_NAMESPACE {
-LLVM_LIBC_FUNCTION(float16, ceilf16, (float16 x)) { return fputil::ceil(x); }
+LLVM_LIBC_FUNCTION(float16, ceilf16, (float16 x)) {
+#if defined(__LIBC_USE_BUILTIN_CEIL_FLOOR_TRUNC) && \
+ defined(LIBC_TARGET_ARCH_IS_AARCH64)
+ return static_cast<float16>(__builtin_ceilf(x));
+#else
+ return fputil::ceil(x);
+#endif
+}
} // namespace LIBC_NAMESPACE
diff --git a/libc/src/math/generic/floor.cpp b/libc/src/math/generic/floor.cpp
index 60386f0c9cf81..bb58ca6a35402 100644
--- a/libc/src/math/generic/floor.cpp
+++ b/libc/src/math/generic/floor.cpp
@@ -12,6 +12,12 @@
namespace LIBC_NAMESPACE {
-LLVM_LIBC_FUNCTION(double, floor, (double x)) { return fputil::floor(x); }
+LLVM_LIBC_FUNCTION(double, floor, (double x)) {
+#ifdef __LIBC_USE_BUILTIN_CEIL_FLOOR_TRUNC
+ return __builtin_floor(x);
+#else
+ return fputil::floor(x);
+#endif
+}
} // namespace LIBC_NAMESPACE
diff --git a/libc/src/math/generic/floorf.cpp b/libc/src/math/generic/floorf.cpp
index 85666688685dc..459f338d897be 100644
--- a/libc/src/math/generic/floorf.cpp
+++ b/libc/src/math/generic/floorf.cpp
@@ -12,6 +12,12 @@
namespace LIBC_NAMESPACE {
-LLVM_LIBC_FUNCTION(float, floorf, (float x)) { return fputil::floor(x); }
+LLVM_LIBC_FUNCTION(float, floorf, (float x)) {
+#ifdef __LIBC_USE_BUILTIN_CEIL_FLOOR_TRUNC
+ return __builtin_floorf(x);
+#else
+ return fputil::floor(x);
+#endif
+}
} // namespace LIBC_NAMESPACE
diff --git a/libc/src/math/generic/floorf16.cpp b/libc/src/math/generic/floorf16.cpp
index db0b326c0e5f6..6d8c497946c84 100644
--- a/libc/src/math/generic/floorf16.cpp
+++ b/libc/src/math/generic/floorf16.cpp
@@ -9,9 +9,17 @@
#include "src/math/floorf16.h"
#include "src/__support/FPUtil/NearestIntegerOperations.h"
#include "src/__support/common.h"
+#include "src/__support/macros/properties/architectures.h"
namespace LIBC_NAMESPACE {
-LLVM_LIBC_FUNCTION(float16, floorf16, (float16 x)) { return fputil::floor(x); }
+LLVM_LIBC_FUNCTION(float16, floorf16, (float16 x)) {
+#if defined(__LIBC_USE_BUILTIN_CEIL_FLOOR_TRUNC) && \
+ defined(LIBC_TARGET_ARCH_IS_AARCH64)
+ return static_cast<float16>(__builtin_floorf(x));
+#else
+ return fputil::floor(x);
+#endif
+}
} // namespace LIBC_NAMESPACE
diff --git a/libc/src/math/generic/round.cpp b/libc/src/math/generic/round.cpp
index ca8f19f35f7fe..d873524ad9f42 100644
--- a/libc/src/math/generic/round.cpp
+++ b/libc/src/math/generic/round.cpp
@@ -12,6 +12,12 @@
namespace LIBC_NAMESPACE {
-LLVM_LIBC_FUNCTION(double, round, (double x)) { return fputil::round(x); }
+LLVM_LIBC_FUNCTION(double, round, (double x)) {
+#ifdef __LIBC_USE_BUILTIN_ROUND
+ return __builtin_round(x);
+#else
+ return fputil::round(x);
+#endif
+}
} // namespace LIBC_NAMESPACE
diff --git a/libc/src/math/generic/roundeven.cpp b/libc/src/math/generic/roundeven.cpp
index 5f2adf9b5fce6..76409d526e208 100644
--- a/libc/src/math/generic/roundeven.cpp
+++ b/libc/src/math/generic/roundeven.cpp
@@ -13,7 +13,11 @@
namespace LIBC_NAMESPACE {
LLVM_LIBC_FUNCTION(double, roundeven, (double x)) {
+#ifdef __LIBC_USE_BUILTIN_ROUNDEVEN
+ return __builtin_roundeven(x);
+#else
return fputil::round_using_specific_rounding_mode(x, FP_INT_TONEAREST);
+#endif
}
} // namespace LIBC_NAMESPACE
diff --git a/libc/src/math/generic/roundevenf.cpp b/libc/src/math/generic/roundevenf.cpp
index 353bec74ecf02..22538272bedbd 100644
--- a/libc/src/math/generic/roundevenf.cpp
+++ b/libc/src/math/generic/roundevenf.cpp
@@ -13,7 +13,11 @@
namespace LIBC_NAMESPACE {
LLVM_LIBC_FUNCTION(float, roundevenf, (float x)) {
+#ifdef __LIBC_USE_BUILTIN_ROUNDEVEN
+ return __builtin_roundevenf(x);
+#else
return fputil::round_using_specific_rounding_mode(x, FP_INT_TONEAREST);
+#endif
}
} // namespace LIBC_NAMESPACE
diff --git a/libc/src/math/generic/roundevenf16.cpp b/libc/src/math/generic/roundevenf16.cpp
index 9ecf79ce6f6c2..90c75a10d3ddb 100644
--- a/libc/src/math/generic/roundevenf16.cpp
+++ b/libc/src/math/generic/roundevenf16.cpp
@@ -9,11 +9,17 @@
#include "src/math/roundevenf16.h"
#include "src/__support/FPUtil/NearestIntegerOperations.h"
#include "src/__support/common.h"
+#include "src/__support/macros/properties/architectures.h"
namespace LIBC_NAMESPACE {
LLVM_LIBC_FUNCTION(float16, roundevenf16, (float16 x)) {
+#if defined(__LIBC_USE_BUILTIN_ROUNDEVEN) && \
+ defined(LIBC_TARGET_ARCH_IS_AARCH64)
+ return static_cast<float16>(__builtin_roundevenf(x));
+#else
return fputil::round_using_specific_rounding_mode(x, FP_INT_TONEAREST);
+#endif
}
} // namespace LIBC_NAMESPACE
diff --git a/libc/src/math/generic/roundf.cpp b/libc/src/math/generic/roundf.cpp
index 9627390ea8b8d..8b3add7cb9e2d 100644
--- a/libc/src/math/generic/roundf.cpp
+++ b/libc/src/math/generic/roundf.cpp
@@ -12,6 +12,12 @@
namespace LIBC_NAMESPACE {
-LLVM_LIBC_FUNCTION(float, roundf, (float x)) { return fputil::round(x); }
+LLVM_LIBC_FUNCTION(float, roundf, (float x)) {
+#ifdef __LIBC_USE_BUILTIN_ROUND
+ return __builtin_roundf(x);
+#else
+ return fputil::round(x);
+#endif
+}
} // namespace LIBC_NAMESPACE
diff --git a/libc/src/math/generic/roundf16.cpp b/libc/src/math/generic/roundf16.cpp
index 75a255d7798d5..fca0194ec5dbb 100644
--- a/libc/src/math/generic/roundf16.cpp
+++ b/libc/src/math/generic/roundf16.cpp
@@ -9,9 +9,16 @@
#include "src/math/roundf16.h"
#include "src/__support/FPUtil/NearestIntegerOperations.h"
#include "src/__support/common.h"
+#include "src/__support/macros/properties/architectures.h"
namespace LIBC_NAMESPACE {
-LLVM_LIBC_FUNCTION(float16, roundf16, (float16 x)) { return fputil::round(x); }
+LLVM_LIBC_FUNCTION(float16, roundf16, (float16 x)) {
+#if defined(__LIBC_USE_BUILTIN_ROUND) && defined(LIBC_TARGET_ARCH_IS_AARCH64)
+ return static_cast<float16>(__builtin_roundf(x));
+#else
+ return fputil::round(x);
+#endif
+}
} // namespace LIBC_NAMESPACE
diff --git a/libc/src/math/generic/trunc.cpp b/libc/src/math/generic/trunc.cpp
index d171ab1f092fd..5761565646c36 100644
--- a/libc/src/math/generic/trunc.cpp
+++ b/libc/src/math/generic/trunc.cpp
@@ -12,6 +12,12 @@
namespace L...
[truncated]
|
| if(NOT((LIBC_TARGET_ARCHITECTURE_IS_X86 AND (LIBC_CPU_FEATURES MATCHES "SSE4_2")) OR | ||
| LIBC_TARGET_ARCHITECTURE_IS_AARCH64)) |
There was a problem hiding this comment.
It would be nice if we had a CMake formatter. I copied the style from the lines above, but the OR line is longer than 80 chars (just like above).
| libc.src.__support.macros.properties.types | ||
| libc.src.__support.FPUtil.nearest_integer_operations | ||
| libc.src.__support.macros.properties.architectures |
There was a problem hiding this comment.
There are CMakeLists.txt files where libc.src.__support.macros* appear before libc.src.__support.FPUtil.* and files where it appears after. I assumed that it was because the header's dependencies are listed first and then the implementation file's, not because the ordering is random.
libc/src/math/generic/ceilf16.cpp
Outdated
| #if defined(__LIBC_USE_BUILTIN_CEIL_FLOOR_TRUNC) && \ | ||
| defined(LIBC_TARGET_ARCH_IS_AARCH64) |
There was a problem hiding this comment.
Probably would be faster than the generic implementation on x86-64 with AVX-512 FP16, but I don't have a CPU that supports it. It could also be interesting to try on a Zen 2 or later CPU, because F16C instructions have lower latencies on those: https://uops.info/table.html?search=vcvtph2ps%20(xmm%2C%20xmm)&cb_lat=on&cb_tp=on&cb_uops=on&cb_ports=on&cb_ADLP=on&cb_ADLE=on&cb_ZENp=on&cb_ZEN2=on&cb_measurements=on&cb_doc=on&cb_others=on.
On my i7-13700H however, using F16C instructions is slower than using the generic functions.
Edit: builtins are actually faster but only with GCC and -march=native.
libc/test/src/math/performance_testing/nearest_integer_funcs_perf.cpp
Outdated
Show resolved
Hide resolved
| #ifdef __LIBC_USE_BUILTIN_CEIL_FLOOR_TRUNC | ||
| return __builtin_ceilf(x); | ||
| #else | ||
| return fputil::ceil(x); |
There was a problem hiding this comment.
Can we not just make fputil::ceil() use the builtin? That way internal usage gets the more optimal version.
There was a problem hiding this comment.
I talked about this with Tue yesterday. We were initially talking about doing this in FPUtil, but yesterday he suggested having the builtin/generic switch in the entrypoints directly because fputil::{ceil,floor,round,trunc} are only used in the respective entrypoints (actually fputill::trunc is used in unit tests for modf, but nowhere in libc/src/).
libc/src/math/generic/ceilf.cpp
Outdated
| @@ -12,6 +12,12 @@ | |||
|
|
|||
| namespace LIBC_NAMESPACE { | |||
|
|
|||
| LLVM_LIBC_FUNCTION(float, ceilf, (float x)) { return fputil::ceil(x); } | |||
| LLVM_LIBC_FUNCTION(float, ceilf, (float x)) { | |||
| #ifdef __LIBC_USE_BUILTIN_CEIL_FLOOR_TRUNC | |||
There was a problem hiding this comment.
We shouldn't use double underscores for the name.
There was a problem hiding this comment.
libc/cmake/modules/LLVMLibCCompileOptionRules.cmake defines __LIBC_RISCV_USE_FMA and __LIBC_EXPLICIT_SIMD_OPT. I guessed that a macro defined through compile options should have __.
overmighty
force-pushed
the
libc-math-rounding-builtins
branch
from
34ffd03 to
11e4992
Compare
overmighty
changed the title
|
|
overmighty
force-pushed
the
libc-math-rounding-builtins
branch
from
11e4992 to
50f1237
Compare
| float try_builtin_ceilf(float x) { return __builtin_ceilf(x); } | ||
| float try_builtin_floorf(float x) { return __builtin_ceilf(x); } |
There was a problem hiding this comment.
Oops, try_builtin_floorf checks for __builtin_ceilf.
|
Before:
|
|
After:
|
… when available
Fix check_builtin_ceil_floor_trunc.cpp not checking for __builtin_floor{,f}.
… when available
Optimize rint{,f,f16} using __builtin_rint{,f} when available.
… when available Nit: sort roundeven after round.
… when available Fix builtin detection succeeding even when calls to functions provided by the system libc are generated.
… when available Delete inline assembly implementations for AArch64.
… when available Fix GCC warning "variable ‘result’ set but not used".
… when available Fix builtin detection on Android due to missing _start being an error with LLD from Android NDK r26d.
overmighty
force-pushed
the
libc-math-rounding-builtins
branch
from
911d1ef to
7f212a3
Compare
|
Rebased to fix the merge conflicts. I still get these "undefined hidden symbol" linking errors I've been getting since the switch to I change MPFRUtils to use diff --git a/libc/utils/MPFRWrapper/MPFRUtils.cpp b/libc/utils/MPFRWrapper/MPFRUtils.cpp
index b67a9da40bd7..3f1758b20519 100644
--- a/libc/utils/MPFRWrapper/MPFRUtils.cpp
+++ b/libc/utils/MPFRWrapper/MPFRUtils.cpp
@@ -23,7 +23,7 @@
template <typename T> using FPBits = LIBC_NAMESPACE::fputil::FPBits<T>;
-namespace LIBC_NAMESPACE_DECL {
+namespace LIBC_NAMESPACE {
namespace testing {
namespace mpfr {
diff --git a/libc/utils/MPFRWrapper/MPFRUtils.h b/libc/utils/MPFRWrapper/MPFRUtils.h
index 28390af9ee6d..9b0002209bed 100644
--- a/libc/utils/MPFRWrapper/MPFRUtils.h
+++ b/libc/utils/MPFRWrapper/MPFRUtils.h
@@ -16,7 +16,7 @@
#include <stdint.h>
-namespace LIBC_NAMESPACE_DECL {
+namespace LIBC_NAMESPACE {
namespace testing {
namespace mpfr { |
…vailable (#98376) Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags: Differential Revision: https://phabricator.intern.facebook.com/D60251690
No description provided.