[ARMADA-3251] Reenable support for quad precision

Completed definition for Sleef_quad in `quaddef.h` where it can be used by sleefqp.c Note quad precision is not available for aarch64, as it does not support quadmath library. Change-Id: Iac54d39a871498e390a260356313c3f3c35bda93
shibatch · Jun 5, 2024 · f485670 · f485670
1 parent ffde961
commit f485670
Show file tree

Hide file tree

Showing 7 changed files with 119 additions and 6 deletions.
diff --git a/Configure.cmake b/Configure.cmake
@@ -366,6 +366,13 @@ if (SLEEF_ENFORCE_FLOAT128 AND NOT COMPILER_SUPPORTS_FLOAT128)
   message(FATAL_ERROR "SLEEF_ENFORCE_FLOAT128 is specified and that feature is disabled or not supported by the compiler")
 endif()
 
+if(COMPILER_SUPPORTS_FLOAT128)
+  CHECK_C_SOURCE_COMPILES("
+  #include <quadmath.h>
+  int main() { __float128 r = 1;
+  }" COMPILER_SUPPORTS_QUADMATH)
+endif()
+
 # SSE2
 
 option(SLEEF_DISABLE_SSE2 "Disable SSE2" OFF)

diff --git a/src/common/quaddef.h b/src/common/quaddef.h
@@ -16,7 +16,7 @@
 #if !defined(Sleef_quad_DEFINED)
 #define Sleef_quad_DEFINED
 typedef struct { uint64_t x, y; } Sleef_uint64_2t;
-#if defined(SLEEF_FLOAT128_IS_IEEEQP)
+#if defined(SLEEF_FLOAT128_IS_IEEEQP) || defined(ENABLEFLOAT128)
 typedef __float128 Sleef_quad;
 #define SLEEF_QUAD_C(x) (x ## Q)
 #elif defined(SLEEF_LONGDOUBLE_IS_IEEEQP)
@@ -27,6 +27,51 @@ typedef Sleef_uint64_2t Sleef_quad;
 #endif
 #endif
 
+#if !defined(Sleef_quad1_DEFINED)
+#define Sleef_quad1_DEFINED
+typedef union {
+  struct {
+    Sleef_quad x;
+  };
+  Sleef_quad s[1];
+} Sleef_quad1;
+#endif
+
+#if !defined(Sleef_quad2_DEFINED)
+#define Sleef_quad2_DEFINED
+typedef union {
+  struct {
+    Sleef_quad x, y;
+  };
+  Sleef_quad s[2];
+} Sleef_quad2;
+#endif
+
+#if !defined(Sleef_quad4_DEFINED)
+#define Sleef_quad4_DEFINED
+typedef union {
+  struct {
+    Sleef_quad x, y, z, w;
+  };
+  Sleef_quad s[4];
+} Sleef_quad4;
+#endif
+
+#if !defined(Sleef_quad8_DEFINED)
+#define Sleef_quad8_DEFINED
+typedef union {
+  Sleef_quad s[8];
+} Sleef_quad8;
+#endif
+
+#if defined(__ARM_FEATURE_SVE) && !defined(Sleef_quadx_DEFINED)
+#define Sleef_quadx_DEFINED
+typedef union {
+  Sleef_quad s[32];
+} Sleef_quadx;
+#endif
+
+
 #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))

diff --git a/src/libm-tester/CMakeLists.txt b/src/libm-tester/CMakeLists.txt
@@ -426,25 +426,27 @@ if(LIB_MPFR AND NOT MINGW)
   if(COMPILER_SUPPORTS_LONG_DOUBLE)
     list(APPEND PRECISIONS ld)
   endif()
+  if(COMPILER_SUPPORTS_QUADMATH)
+    list(APPEND PRECISIONS qp)
+    set(LIBQUADMATH "-lquadmath")
+    set(ENABLEFLOAT128 PRIVATE ENABLEFLOAT128=1)
+  endif()
   foreach(P ${PRECISIONS})
     set(T "tester2${P}")
     add_executable(${T} tester2${P}.c testerutil.c)
-    target_compile_definitions(${T} PRIVATE USEMPFR=1 ${COMMON_TARGET_DEFINITIONS})
+    target_compile_definitions(${T} PRIVATE USEMPFR=1 ${ENABLEFLOAT128} ${COMMON_TARGET_DEFINITIONS})
     set_target_properties(${T} PROPERTIES ${COMMON_TARGET_PROPERTIES})
     if (FORCE_AAVPCS)
       target_compile_definitions(${T} PRIVATE ENABLE_AAVPCS=1)
     endif(FORCE_AAVPCS)
     if (MPFR_INCLUDE_DIR)
       target_include_directories(${T} PRIVATE ${MPFR_INCLUDE_DIR})
     endif()
-
-    target_link_libraries(${T} ${TARGET_LIBSLEEF} ${LIB_MPFR} ${LIBM} ${LIBGMP})
+    target_link_libraries(${T} ${TARGET_LIBSLEEF} ${LIBQUADMATH} ${LIB_MPFR} ${LIBM} ${LIBGMP})
     add_dependencies(${T} ${TARGET_HEADERS})
     add_dependencies(${T} ${TARGET_LIBSLEEF})
   endforeach()
 
-  # No test defined with tester2
-
   # Compile executable 'tester'
   add_host_executable(${TARGET_TESTER} tester.c testerutil.c)
   if (NOT CMAKE_CROSSCOMPILING)

diff --git a/src/libm-tester/tester2qp.c b/src/libm-tester/tester2qp.c
@@ -15,10 +15,12 @@
 #include <math.h>
 #include <quadmath.h>
 
+#ifdef ENABLE_SYS_getrandom
 #define _GNU_SOURCE
 #include <unistd.h>
 #include <sys/syscall.h>
 #include <linux/random.h>
+#endif
 
 #include "sleef.h"
 
@@ -94,22 +96,35 @@ Sleef_quad rnd() {
   case 0: return  INFINITY;
   case 1: return -INFINITY;
   }
+#ifdef ENABLE_SYS_getrandom
   syscall(SYS_getrandom, &c.u128, sizeof(c.u128), 0);
+#else
+  c.u128 = random() | ((__int128)random() << 31) | ((__int128)random() << (31*2)) | ((__int128)random() << (31*3)) | ((__int128)random() << (31*4));
+#endif
   return c.d;
+
 }
 
 Sleef_quad rnd_fr() {
   conv_t c;
   do {
+#ifdef ENABLE_SYS_getrandom
     syscall(SYS_getrandom, &c.u128, sizeof(c.u128), 0);
+#else
+    c.u128 = random() | ((__int128)random() << 31) | ((__int128)random() << (31*2)) | ((__int128)random() << (31*3)) | ((__int128)random() << (31*4));
+#endif
   } while(!isnumberq(c.d));
   return c.d;
 }
 
 Sleef_quad rnd_zo() {
   conv_t c;
   do {
+#ifdef ENABLE_SYS_getrandom
     syscall(SYS_getrandom, &c.u128, sizeof(c.u128), 0);
+#else
+    c.u128 = random() | ((__int128)random() << 31) | ((__int128)random() << (31*2)) | ((__int128)random() << (31*3)) | ((__int128)random() << (31*4));
+#endif
   } while(!isnumberq(c.d) || c.d < -1 || 1 < c.d);
   return c.d;
 }

diff --git a/src/libm/CMakeLists.txt b/src/libm/CMakeLists.txt
@@ -376,6 +376,13 @@ target_compile_definitions(${TARGET_LIBSLEEF}
   PRIVATE DORENAME=1 ${COMMON_TARGET_DEFINITIONS}
 )
 
+if(COMPILER_SUPPORTS_FLOAT128)
+  # TODO: Not supported for LLVM bitcode gen as it has a specific compilation flags
+  target_sources(${TARGET_LIBSLEEF} PRIVATE sleefqp.c)
+  target_compile_definitions(${TARGET_LIBSLEEF}
+    PRIVATE ENABLEFLOAT128=1 ${COMMON_TARGET_DEFINITIONS})
+endif()
+
 if(COMPILER_SUPPORTS_BUILTIN_MATH)
   target_compile_definitions(${TARGET_LIBSLEEF} PRIVATE ENABLE_BUILTIN_MATH=1)
 endif()

diff --git a/src/libm/sleeflibm_header.h.org.in b/src/libm/sleeflibm_header.h.org.in
@@ -164,6 +164,37 @@ typedef struct {
 } Sleef_longdouble2;
 #endif
 
+#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__))
+#define SLEEF_LONGDOUBLE_IS_IEEEQP
+#endif
+
+#if !defined(Sleef_quad_DEFINED)
+#define Sleef_quad_DEFINED
+typedef struct { uint64_t x, y; } Sleef_uint64_2t;
+#if defined(SLEEF_FLOAT128_IS_IEEEQP) || defined(ENABLEFLOAT128)
+typedef __float128 Sleef_quad;
+#define SLEEF_QUAD_C(x) (x ## Q)
+#elif defined(SLEEF_LONGDOUBLE_IS_IEEEQP)
+typedef long double Sleef_quad;
+#define SLEEF_QUAD_C(x) (x ## L)
+#else
+typedef Sleef_uint64_2t Sleef_quad;
+#endif
+#endif
+#if !defined(Sleef_quad2_DEFINED)
+#define Sleef_quad2_DEFINED
+typedef union {
+  struct {
+    Sleef_quad x, y;
+  };
+  Sleef_quad s[2];
+} Sleef_quad2;
+#endif
+
 #ifdef __cplusplus
 extern "C"
 {
@@ -326,3 +357,8 @@ SLEEF_PRAGMA_OMP_SIMD_SP SLEEF_IMPORT SLEEF_CONST float Sleef_erfcf_u15(float);
 
 SLEEF_IMPORT SLEEF_CONST Sleef_longdouble2 Sleef_sincospil_u05(long double);
 SLEEF_IMPORT SLEEF_CONST Sleef_longdouble2 Sleef_sincospil_u35(long double);
+
+#if defined(Sleef_quad2_DEFINED)
+SLEEF_IMPORT SLEEF_CONST Sleef_quad2 Sleef_sincospiq_u05(Sleef_quad);
+SLEEF_IMPORT SLEEF_CONST Sleef_quad2 Sleef_sincospiq_u35(Sleef_quad);
+#endif
diff --git a/src/libm/sleefqp.c b/src/libm/sleefqp.c
@@ -12,6 +12,7 @@
 #include <limits.h>
 
 #include "misc.h"
+#include "quaddef.h"
 
 #ifdef DORENAME
 #include "rename.h"