Robust detection of compiler support for IEEE binary128 (#643)

Co-authored-by: Naoki Shibata <shibatch.sf.net@gmail.com>

Author: shibatch

CMakeLists.txt

@@ -7,13 +7,15 @@ project(SLEEF VERSION ${SLEEF_VERSION} LANGUAGES C CXX)
 
 set(SLEEF_SOVERSION ${SLEEF_VERSION_MAJOR})
 
+set(CMAKE_CXX_STANDARD 20)
+
 # Options
 
 option(SLEEF_BUILD_STATIC_TEST_BINS "Build statically linked test executables" OFF)
 option(SLEEF_ENABLE_LTO "Enable LTO on GCC or ThinLTO on clang" OFF)
 option(SLEEF_BUILD_LIBM "libsleef will be built." ON)
 option(SLEEF_BUILD_DFT "libsleefdft will be built." OFF)
-option(SLEEF_BUILD_QUAD "libsleefquad will be built." OFF)
+option(SLEEF_BUILD_QUAD "libsleefquad will be built." ON)
 option(SLEEF_BUILD_GNUABI_LIBS "libsleefgnuabi will be built." ON)
 option(SLEEF_BUILD_SCALAR_LIB "libsleefscalar will be built." OFF)
 option(SLEEF_BUILD_TESTS "Tests will be built." ON)
@@ -27,28 +29,66 @@ option(SLEEF_SHOW_ERROR_LOG "Show cmake error log." OFF)
 option(SLEEF_ASAN "Enable address sanitizing on all targets." OFF)
 
 option(SLEEF_ENABLE_TESTER "Enable testing libm with tester" OFF)
-option(SLEEF_ENABLE_TESTER4 "Enable testing with tester4" ON)
-
 option(SLEEF_ENFORCE_TESTER "Build fails if tester is not available" OFF)
+
 option(SLEEF_ENFORCE_TESTER3 "Build fails if tester3 is not built" OFF)
+
+option(SLEEF_ENABLE_TESTER4 "Enable testing with tester4" ON)
 option(SLEEF_ENFORCE_TESTER4 "Build fails if tester4 is not available" OFF)
 
 option(SLEEF_ENABLE_ALTDIV  "Enable alternative division method (aarch64 only)" OFF)
 option(SLEEF_ENABLE_ALTSQRT "Enable alternative sqrt method (aarch64 only)" OFF)
 
 option(SLEEF_DISABLE_FFTW "Disable testing the DFT library with FFTW" OFF)
 option(SLEEF_DISABLE_MPFR "Disable testing with the MPFR library" OFF)
-option(SLEEF_DISABLE_SSL "Disable testing with the SSL library" OFF)
 option(SLEEF_ENABLE_TLFLOAT "Enable use of TLFloat library" ON)
 
+option(SLEEF_DISABLE_SSL "Disable testing with the SSL library" OFF)
+set(OPENSSL_EXTRA_LIBRARIES "" CACHE STRING "Extra libraries for openssl")
+
 option(SLEEF_ENABLE_CUDA "Enable CUDA" OFF)
 
 option(SLEEF_BUILD_WITH_LIBM "build libsleef with libm, can turn off on Windows to solve mutiple math functions issue." ON)
 
-#
+option(SLEEF_DISABLE_LONG_DOUBLE "Disable long double" OFF)
+option(SLEEF_ENFORCE_LONG_DOUBLE "Build fails if long double is not supported by the compiler" OFF)
+
+option(SLEEF_DISABLE_FLOAT128 "Disable float128" OFF)
+option(SLEEF_ENFORCE_FLOAT128 "Build fails if float128 is not supported by the compiler" OFF)
+
+option(SLEEF_DISABLE_SSE2 "Disable SSE2" OFF)
+option(SLEEF_ENFORCE_SSE2 "Build fails if SSE2 is not supported by the compiler" OFF)
+option(SLEEF_DISABLE_SSE4 "Disable SSE4" OFF)
+option(SLEEF_ENFORCE_SSE4 "Build fails if SSE4 is not supported by the compiler" OFF)
+option(SLEEF_ENFORCE_AVX "Disable AVX" OFF)
+option(SLEEF_ENFORCE_AVX "Build fails if AVX is not supported by the compiler" OFF)
+option(SLEEF_DISABLE_FMA4 "Disable FMA4" OFF)
+option(SLEEF_ENFORCE_FMA4 "Build fails if FMA4 is not supported by the compiler" OFF)
+option(SLEEF_DISABLE_AVX2 "Disable AVX2" OFF)
+option(SLEEF_ENFORCE_AVX2 "Build fails if AVX2 is not supported by the compiler" OFF)
+option(SLEEF_DISABLE_AVX512F "Disable AVX512F" OFF)
+option(SLEEF_ENFORCE_AVX512F "Build fails if AVX512F is not supported by the compiler" OFF)
+option(SLEEF_DISABLE_SVE "Disable SVE" OFF)
+option(SLEEF_ENFORCE_SVE "Build fails if SVE is not supported by the compiler" OFF)
+option(SLEEF_DISABLE_VSX "Disable VSX" OFF)
+option(SLEEF_ENFORCE_VSX "Build fails if VSX is not supported by the compiler" OFF)
+option(SLEEF_DISABLE_VSX3 "Disable VSX3" OFF)
+option(SLEEF_ENFORCE_VSX3 "Build fails if VSX3 is not supported by the compiler" OFF)
+option(SLEEF_DISABLE_VXE "Disable VXE" OFF)
+option(SLEEF_ENFORCE_VXE "Build fails if VXE is not supported by the compiler" OFF)
+option(SLEEF_DISABLE_VXE2 "Disable VXE2" OFF)
+option(SLEEF_ENFORCE_VXE2 "Build fails if VXE2 is not supported by the compiler" OFF)
+option(SLEEF_DISABLE_RVVM1 "Disable RVVM1" OFF)
+option(SLEEF_ENFORCE_RVVM1 "Build fails if RVVM1 is not supported by the compiler" OFF)
+option(SLEEF_DISABLE_RVVM2 "Disable RVVM2" OFF)
+option(SLEEF_ENFORCE_RVVM2 "Build fails if RVVM2 is not supported by the compiler" OFF)
+
+option(SLEEF_ENFORCE_CUDA "Build fails if CUDA is not supported" OFF)
+
+option(SLEEF_DISABLE_OPENMP "Disable OPENMP" OFF)
+option(SLEEF_ENFORCE_OPENMP "Build fails if OPENMP is not supported by the compiler" OFF)
 
-set(SLEEF_ENABLE_CXX ON)
-set(CMAKE_CXX_STANDARD 20)
+#
 
 if (SLEEF_BUILD_BENCH_REF)
   if (NOT SLEEF_BUILD_BENCH)
@@ -147,8 +187,8 @@ set(COSTOVERRIDE_AVX2 2)
 set(COSTOVERRIDE_AVX 2)
 set(COSTOVERRIDE_NEON32 2)
 set(COSTOVERRIDE_NEON32VFPV4 2)
-set(COSTOVERRIDE_SVE 20)
-set(COSTOVERRIDE_SVENOFMA 20)
+set(COSTOVERRIDE_SVE 10)
+set(COSTOVERRIDE_SVENOFMA 10)
 set(COSTOVERRIDE_RVVM1 10)
 set(COSTOVERRIDE_RVVM1NOFMA 10)
 set(COSTOVERRIDE_RVVM2 20)
@@ -157,10 +197,7 @@ set(COSTOVERRIDE_RVVM2NOFMA 20)
 #
 
 enable_testing()
-
-if (SLEEF_ENABLE_CXX)
-  enable_language(CXX)
-endif()
+enable_language(CXX)
 
 if (SLEEF_ENABLE_CUDA)
   enable_language(CUDA)
@@ -222,6 +259,7 @@ include(Configure.cmake)
 configure_file(
   ${PROJECT_SOURCE_DIR}/sleef-config.h.in
   ${PROJECT_BINARY_DIR}/include/sleef-config.h @ONLY)
+include_directories(AFTER "${PROJECT_BINARY_DIR}/include")
 
 # We like to have a documented index of all targets in the project. The
 # variables listed below carry the names of the targets defined throughout

Configure.cmake

@@ -1,5 +1,6 @@
 include(CheckCCompilerFlag)
 include(CheckCSourceCompiles)
+include(CheckCXXSourceCompiles)
 include(CheckTypeSize)
 include(CheckLanguage)
 
@@ -12,7 +13,6 @@ if (SLEEF_BUILD_STATIC_TEST_BINS)
 endif()
 
 if (NOT SLEEF_DISABLE_SSL)
-  set(OPENSSL_EXTRA_LIBRARIES "" CACHE STRING "Extra libraries for openssl")
   if (NOT CMAKE_CROSSCOMPILING AND NOT SLEEF_FORCE_FIND_PACKAGE_SSL)
     if (SLEEF_BUILD_STATIC_TEST_BINS)
       set(OPENSSL_USE_STATIC_LIBS TRUE)
@@ -223,7 +223,7 @@ if(NOT CLANG_EXE_PATH)
     set(CLANG_EXE_PATH ${CMAKE_C_COMPILER})
   else()
     # Else we may find clang on the path?
-    find_program(CLANG_EXE_PATH NAMES clang "clang-11" "clang-10" "clang-9" "clang-8" "clang-7" "clang-6.0" "clang-5.0" "clang-4.0" "clang-3.9")
+    find_program(CLANG_EXE_PATH NAMES clang "clang-25" "clang-24" "clang-23" "clang-22" "clang-21" "clang-20" "clang-19" "clang-18" "clang-17")
   endif()
 endif()
 
@@ -410,9 +410,6 @@ endif()
 
 # Long double
 
-option(SLEEF_DISABLE_LONG_DOUBLE "Disable long double" OFF)
-option(SLEEF_ENFORCE_LONG_DOUBLE "Build fails if long double is not supported by the compiler" OFF)
-
 if(NOT SLEEF_DISABLE_LONG_DOUBLE)
   CHECK_TYPE_SIZE("long double" LD_SIZE)
   if(LD_SIZE GREATER "9")
@@ -433,9 +430,6 @@ endif()
 
 # float128
 
-option(SLEEF_DISABLE_FLOAT128 "Disable float128" OFF)
-option(SLEEF_ENFORCE_FLOAT128 "Build fails if float128 is not supported by the compiler" OFF)
-
 if(NOT SLEEF_DISABLE_FLOAT128)
   CHECK_C_SOURCE_COMPILES("
   int main() { __float128 r = 1;
@@ -455,10 +449,37 @@ if(COMPILER_SUPPORTS_FLOAT128)
   }" COMPILER_SUPPORTS_QUADMATH)
 endif()
 
-# SSE2
+if(COMPILER_SUPPORTS_FLOAT128)
+  if (CMAKE_CXX_COMPILER_TARGET)
+    set(CMAKE_REQUIRED_FLAGS "--target=${CMAKE_CXX_COMPILER_TARGET}")
+  endif()
+  CHECK_CXX_SOURCE_COMPILES("
+#include <bit>
+struct s { long long x, y; };
+int main(int argc, char **argv) {
+  constexpr s a = std::bit_cast<s>(__float128(0.1234)*__float128(56.789));
+  static_assert((a.x ^ a.y) == 0xc7d695c93a4e2b71LL);
+  __float128 i = argc;
+  return (int)i;
+}
+" SLEEF_FLOAT128_IS_IEEEQP)
+  set(CMAKE_REQUIRED_FLAGS)
+endif()
+
+if (CMAKE_CXX_COMPILER_TARGET)
+  set(CMAKE_REQUIRED_FLAGS "--target=${CMAKE_CXX_COMPILER_TARGET}")
+endif()
+CHECK_CXX_SOURCE_COMPILES("
+#include <bit>
+struct s { long long x, y; };
+int main(void) {
+  constexpr s a = std::bit_cast<s>((long double)0.1234*(long double)56.789);
+  static_assert((a.x ^ a.y) == 0xc7d695c93a4e2b71LL);
+}
+" SLEEF_LONGDOUBLE_IS_IEEEQP)
+set(CMAKE_REQUIRED_FLAGS)
 
-option(SLEEF_DISABLE_SSE2 "Disable SSE2" OFF)
-option(SLEEF_ENFORCE_SSE2 "Build fails if SSE2 is not supported by the compiler" OFF)
+# SSE2
 
 if(SLEEF_ARCH_X86 AND NOT SLEEF_DISABLE_SSE2)
   string (REPLACE ";" " " CMAKE_REQUIRED_FLAGS "${FLAGS_ENABLE_SSE2}")
@@ -479,9 +500,6 @@ endif()
 
 # SSE 4.1
 
-option(SLEEF_DISABLE_SSE4 "Disable SSE4" OFF)
-option(SLEEF_ENFORCE_SSE4 "Build fails if SSE4 is not supported by the compiler" OFF)
-
 if(SLEEF_ARCH_X86 AND NOT SLEEF_DISABLE_SSE4)
   string (REPLACE ";" " " CMAKE_REQUIRED_FLAGS "${FLAGS_ENABLE_SSE4}")
   CHECK_C_SOURCE_COMPILES("
@@ -501,9 +519,6 @@ endif()
 
 # AVX
 
-option(SLEEF_ENFORCE_AVX "Disable AVX" OFF)
-option(SLEEF_ENFORCE_AVX "Build fails if AVX is not supported by the compiler" OFF)
-
 if(SLEEF_ARCH_X86 AND NOT SLEEF_DISABLE_AVX)
   string (REPLACE ";" " " CMAKE_REQUIRED_FLAGS "${FLAGS_ENABLE_AVX}")
   CHECK_C_SOURCE_COMPILES("
@@ -523,9 +538,6 @@ endif()
 
 # FMA4
 
-option(SLEEF_DISABLE_FMA4 "Disable FMA4" OFF)
-option(SLEEF_ENFORCE_FMA4 "Build fails if FMA4 is not supported by the compiler" OFF)
-
 if(SLEEF_ARCH_X86 AND NOT SLEEF_DISABLE_FMA4)
   string (REPLACE ";" " " CMAKE_REQUIRED_FLAGS "${FLAGS_ENABLE_FMA4}")
   CHECK_C_SOURCE_COMPILES("
@@ -545,9 +557,6 @@ endif()
 
 # AVX2
 
-option(SLEEF_DISABLE_AVX2 "Disable AVX2" OFF)
-option(SLEEF_ENFORCE_AVX2 "Build fails if AVX2 is not supported by the compiler" OFF)
-
 if(SLEEF_ARCH_X86 AND NOT SLEEF_DISABLE_AVX2)
   string (REPLACE ";" " " CMAKE_REQUIRED_FLAGS "${FLAGS_ENABLE_AVX2}")
   CHECK_C_SOURCE_COMPILES("
@@ -572,8 +581,6 @@ endif()
 
 # AVX512F
 
-option(SLEEF_DISABLE_AVX512F "Disable AVX512F" OFF)
-option(SLEEF_ENFORCE_AVX512F "Build fails if AVX512F is not supported by the compiler" OFF)
 
 if(SLEEF_ARCH_X86 AND NOT SLEEF_DISABLE_AVX512F)
   string (REPLACE ";" " " CMAKE_REQUIRED_FLAGS "${FLAGS_ENABLE_AVX512F}")
@@ -604,9 +611,6 @@ endif()
 
 # SVE
 
-option(SLEEF_DISABLE_SVE "Disable SVE" OFF)
-option(SLEEF_ENFORCE_SVE "Build fails if SVE is not supported by the compiler" OFF)
-
 # Darwin does not support SVE yet (see issue #474),
 # therefore we disable SVE on Darwin systems.
 if(SLEEF_ARCH_AARCH64 AND NOT SLEEF_DISABLE_SVE AND NOT CMAKE_SYSTEM_NAME STREQUAL "Darwin")
@@ -628,9 +632,6 @@ endif()
 
 # VSX
 
-option(SLEEF_DISABLE_VSX "Disable VSX" OFF)
-option(SLEEF_ENFORCE_VSX "Build fails if VSX is not supported by the compiler" OFF)
-
 if(SLEEF_ARCH_PPC64 AND NOT SLEEF_DISABLE_VSX)
   string (REPLACE ";" " " CMAKE_REQUIRED_FLAGS "${FLAGS_ENABLE_VSX}")
   CHECK_C_SOURCE_COMPILES("
@@ -658,9 +659,6 @@ endif()
 
 # VSX3
 
-option(SLEEF_DISABLE_VSX3 "Disable VSX3" OFF)
-option(SLEEF_ENFORCE_VSX3 "Build fails if VSX3 is not supported by the compiler" OFF)
-
 if(SLEEF_ARCH_PPC64 AND NOT SLEEF_DISABLE_VSX3)
   string (REPLACE ";" " " CMAKE_REQUIRED_FLAGS "${FLAGS_ENABLE_VSX3}")
   CHECK_C_SOURCE_COMPILES("
@@ -687,9 +685,6 @@ endif()
 
 # IBM Z
 
-option(SLEEF_DISABLE_VXE "Disable VXE" OFF)
-option(SLEEF_ENFORCE_VXE "Build fails if VXE is not supported by the compiler" OFF)
-
 if(SLEEF_ARCH_S390X AND NOT SLEEF_DISABLE_VXE)
   string (REPLACE ";" " " CMAKE_REQUIRED_FLAGS "${FLAGS_ENABLE_VXE}")
   CHECK_C_SOURCE_COMPILES("
@@ -711,9 +706,6 @@ endif()
 
 #
 
-option(SLEEF_DISABLE_VXE2 "Disable VXE2" OFF)
-option(SLEEF_ENFORCE_VXE2 "Build fails if VXE2 is not supported by the compiler" OFF)
-
 if(SLEEF_ARCH_S390X AND NOT SLEEF_DISABLE_VXE2)
   string (REPLACE ";" " " CMAKE_REQUIRED_FLAGS "${FLAGS_ENABLE_VXE2}")
   CHECK_C_SOURCE_COMPILES("
@@ -735,9 +727,6 @@ endif()
 
 # RVVM1
 
-option(SLEEF_DISABLE_RVVM1 "Disable RVVM1" OFF)
-option(SLEEF_ENFORCE_RVVM1 "Build fails if RVVM1 is not supported by the compiler" OFF)
-
 if(SLEEF_ARCH_RISCV64 AND NOT SLEEF_DISABLE_RVVM1)
   string (REPLACE ";" " " CMAKE_REQUIRED_FLAGS "${FLAGS_ENABLE_RVVM1}")
   CHECK_C_SOURCE_COMPILES("
@@ -771,9 +760,6 @@ endif()
 
 # RVVM2
 
-option(SLEEF_DISABLE_RVVM2 "Disable RVVM2" OFF)
-option(SLEEF_ENFORCE_RVVM2 "Build fails if RVVM2 is not supported by the compiler" OFF)
-
 if(SLEEF_ARCH_RISCV64 AND NOT SLEEF_DISABLE_RVVM2)
   string (REPLACE ";" " " CMAKE_REQUIRED_FLAGS "${FLAGS_ENABLE_RVVM2}")
   CHECK_C_SOURCE_COMPILES("
@@ -807,17 +793,12 @@ endif()
 
 # CUDA
 
-option(SLEEF_ENFORCE_CUDA "Build fails if CUDA is not supported" OFF)
-
 if (SLEEF_ENFORCE_CUDA AND NOT CMAKE_CUDA_COMPILER)
   message(FATAL_ERROR "SLEEF_ENFORCE_CUDA is specified and that feature is disabled or not supported by the compiler")
 endif()
 
 # OpenMP
 
-option(SLEEF_DISABLE_OPENMP "Disable OPENMP" OFF)
-option(SLEEF_ENFORCE_OPENMP "Build fails if OPENMP is not supported by the compiler" OFF)
-
 if(NOT SLEEF_DISABLE_OPENMP)
   find_package(OpenMP)
   # Check if compilation with OpenMP really succeeds

sleef-config.h.in

@@ -6,6 +6,11 @@
 #define SLEEF_VERSION_MAJOR @SLEEF_VERSION_MAJOR@
 #define SLEEF_VERSION_MINOR @SLEEF_VERSION_MINOR@
 
+#cmakedefine SLEEF_FLOAT128_IS_IEEEQP
+#cmakedefine SLEEF_LONGDOUBLE_IS_IEEEQP
+
+#ifndef SLEEF_STATIC_LIBS
 #cmakedefine SLEEF_STATIC_LIBS
+#endif
 
 #endif // SLEEF_CONFIG_H

src/common/quaddef.h

@@ -5,20 +5,11 @@
 
 #if !defined(SLEEF_GENHEADER)
 
-#if (defined(__SIZEOF_FLOAT128__) && __SIZEOF_FLOAT128__ == 16) || (defined(__linux__) && defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))) || (defined(__PPC64__) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ >= 8)
-#define SLEEF_FLOAT128_IS_IEEEQP
-#endif
-
-#if !defined(SLEEF_FLOAT128_IS_IEEEQP) && defined(__SIZEOF_LONG_DOUBLE__) && __SIZEOF_LONG_DOUBLE__ == 16 && (defined(__aarch64__) || defined(__zarch__) || defined(__riscv))
-#define SLEEF_LONGDOUBLE_IS_IEEEQP
-#endif
+#include "sleef-config.h"
 
 #if !defined(Sleef_quad_DEFINED)
 #define Sleef_quad_DEFINED
 typedef struct { uint64_t x, y; } Sleef_uint64_2t;
-#ifdef _AIX
-#undef  SLEEF_FLOAT128_IS_IEEEQP
-#endif
 #if defined(SLEEF_FLOAT128_IS_IEEEQP) || defined(ENABLEFLOAT128)
 typedef __float128 Sleef_quad;
 #define SLEEF_QUAD_C(x) (x ## Q)
@@ -77,14 +68,6 @@ typedef union {
 
 #else // #if !defined(SLEEF_GENHEADER)
 
-SLEEFSHARPif !defined(SLEEFXXX__NVCC__) && ((defined(SLEEFXXX__SIZEOF_FLOAT128__) && SLEEFXXX__SIZEOF_FLOAT128__ == 16) || (defined(SLEEFXXX__linux__) && defined(SLEEFXXX__GNUC__) && (defined(SLEEFXXX__i386__) || defined(SLEEFXXX__x86_64__))) || (defined(SLEEFXXX__PPC64__) && defined(SLEEFXXX__GNUC__) && !defined(SLEEFXXX__clang__) && SLEEFXXX__GNUC__ >= 8))
-SLEEFSHARPdefine SLEEFXXXSLEEF_FLOAT128_IS_IEEEQP
-SLEEFSHARPendif
-
-SLEEFSHARPif !defined(SLEEFXXXSLEEF_FLOAT128_IS_IEEEQP) && !defined(SLEEFXXX__NVCC__) && defined(SLEEFXXX__SIZEOF_LONG_DOUBLE__) && SLEEFXXX__SIZEOF_LONG_DOUBLE__ == 16 && (defined(SLEEFXXX__aarch64__) || defined(SLEEFXXX__zarch__) || defined(SLEEFXXX__riscv))
-SLEEFSHARPdefine SLEEFXXXSLEEF_LONGDOUBLE_IS_IEEEQP
-SLEEFSHARPendif
-
 SLEEFSHARPif !defined(SLEEFXXXSleef_quad_DEFINED)
 SLEEFSHARPdefine SLEEFXXXSleef_quad_DEFINED
 typedef struct { uint64_t x, y; } Sleef_uint64_2t;