Add static dispatch macros

CMakeLists.txt

@@ -109,6 +109,8 @@ set(TARGET_MKRENAME_GNUABI "mkrename_gnuabi")
 set(TARGET_MKMASKED_GNUABI "mkmasked_gnuabi")
 # Generates the helper executable file mkdisp needed to write the sleef header
 set(TARGET_MKDISP "mkdisp")
+# Generates the helper executable file mkstatic for generating the static dispatch macros
+set(TARGET_MKSTATIC "mkstatic")
 set(TARGET_MKALIAS "mkalias")
 # Generates static library common
 # Defined in src/common/CMakeLists.txt via command add_library

sleef-config.h.in

@@ -8,9 +8,30 @@
 
 // FEATURE DETECTION ***********************************************************
 
+#cmakedefine COMPILER_SUPPORTS_BUILTIN_MATH
 #cmakedefine COMPILER_SUPPORTS_LONG_DOUBLE
-#cmakedefine COMPILER_SUPPORTS_FLOAT128
+#cmakedefine COMPILER_SUPPORTS_OPENMP
+#cmakedefine COMPILER_SUPPORTS_SYS_GETRANDOM
+#cmakedefine COMPILER_SUPPORTS_WEAK_ALIASES
 
+#cmakedefine COMPILER_SUPPORTS_ADVSIMD
+#cmakedefine COMPILER_SUPPORTS_ADVSIMDNOFMA
+#cmakedefine COMPILER_SUPPORTS_AVX
+#cmakedefine COMPILER_SUPPORTS_AVX2
+#cmakedefine COMPILER_SUPPORTS_AVX2128
+#cmakedefine COMPILER_SUPPORTS_AVX512F
+#cmakedefine COMPILER_SUPPORTS_AVX512FNOFMA
+#cmakedefine COMPILER_SUPPORTS_FLOAT128
+#cmakedefine COMPILER_SUPPORTS_FMA4
+#cmakedefine COMPILER_SUPPORTS_NEON32
+#cmakedefine COMPILER_SUPPORTS_NEON32VFPV4
+#cmakedefine COMPILER_SUPPORTS_PURECFMA_SCALAR
+#cmakedefine COMPILER_SUPPORTS_PUREC_SCALAR
 #cmakedefine COMPILER_SUPPORTS_SSE2
+#cmakedefine COMPILER_SUPPORTS_SSE4
+#cmakedefine COMPILER_SUPPORTS_SVE
+#cmakedefine COMPILER_SUPPORTS_SVENOFMA
+#cmakedefine COMPILER_SUPPORTS_VSX
+#cmakedefine COMPILER_SUPPORTS_VSXNOFMA
 
 #endif // CONFIG_H

src/libm/CMakeLists.txt

@@ -21,6 +21,8 @@ foreach(SIMD ${SLEEF_HEADER_LIST})
   list(APPEND SLEEF_HEADER_COMMANDS COMMAND echo Generating sleef.h: ${TARGET_MKRENAME} ${HEADER_PARAMS_${SIMD}})
   list(APPEND SLEEF_HEADER_COMMANDS COMMAND $<TARGET_FILE:${TARGET_MKRENAME}> ${HEADER_PARAMS_${SIMD}} >> ${SLEEF_INCLUDE_HEADER})
 endforeach()
+list(APPEND SLEEF_HEADER_COMMANDS COMMAND echo Generating sleef.h: ${TARGET_MKSTATIC})
+list(APPEND SLEEF_HEADER_COMMANDS COMMAND $<TARGET_FILE:${TARGET_MKSTATIC}> >> ${SLEEF_INCLUDE_HEADER})
 
 if((MSVC OR MINGW AND WIN32) OR SLEEF_CLANG_ON_WINDOWS)
   string(REPLACE "/" "\\" sleef_footer_input_file "${SLEEF_ORG_FOOTER}")
@@ -35,6 +37,7 @@ add_custom_command(OUTPUT ${SLEEF_INCLUDE_HEADER}
     ${SLEEF_ORG_HEADER}
     ${SLEEF_ORG_FOOTER}
     ${TARGET_MKRENAME}
+    ${TARGET_MKSTATIC}
 )
 
 # --------------------------------------------------------------------
@@ -155,6 +158,13 @@ set_target_properties(
   PROPERTIES C_STANDARD 99
 )
 
+# --------------------------------------------------------------------
+# TARGET_MKDISP
+# --------------------------------------------------------------------
+# Helper executable: dispatcher for the vector extensions
+add_host_executable(${TARGET_MKSTATIC} mkstatic.c)
+target_include_directories(${TARGET_MKSTATIC} PRIVATE ${PROJECT_BINARY_DIR}/include)
+
 # --------------------------------------------------------------------
 # TARGET_LIBSLEEF
 # --------------------------------------------------------------------

src/libm/mkstatic.c

@@ -0,0 +1,150 @@
+//   Copyright Naoki Shibata and contributors 2010 - 2020.
+// Distributed under the Boost Software License, Version 1.0.
+//    (See accompanying file LICENSE.txt or copy at
+//          http://www.boost.org/LICENSE_1_0.txt)
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <string.h>
+#include <assert.h>
+
+#include "funcproto.h"
+#include <sleef-config.h>
+
+static const size_t vecSizes[] = {
+  0, 16, 32, 64,
+};
+
+static const struct variantInfo {
+  size_t vecSize;
+  const char* ppTest;
+  const char* suffix;
+} variantList[] = {
+
+  /* Arm */
+
+  #if defined(COMPILER_SUPPORTS_SVE)
+    { 0, "defined(__ARM_SVE)", "sve" },
+  #endif
+  #if defined(COMPILER_SUPPORTS_SVENOFMA)
+    { 0, "defined(__ARM_SVE)", "svenofma" },
+  #endif
+
+  #if defined(COMPILER_SUPPORTS_ADVSIMD)
+    /* Not sure what the feature check for FMA should look like */
+    { 16, "defined(__ARM_NEON) && defined(__aarch64__)", "advsimd" },
+  #endif
+  #if defined(COMPILER_SUPPORTS_ADVSIMDNOFMA)
+    { 16, "defined(__ARM_NEON) && defined(__aarch64__)", "advsimdnofma" },
+  #endif
+
+  #if defined(COMPILER_SUPPORTS_NEON32VFPV4)
+    { 16, "defined(__ARM_NEON) && (__ARM_NEON_FP >= 4)", "neonvfpv4" },
+  #endif
+  #if defined(COMPILER_SUPPORTS_NEON32VFPV4)
+    { 16, "defined(__ARM_NEON)", "neon" },
+  #endif
+
+  /* POWER */
+
+  #if defined(COMPILER_SUPPORTS_VSX)
+    /* Not sure what the feature check for FMA should look like */
+    { 16, "defined(__VSX__) && defined(__ARCH_PWR8) && defined(__LITTLE_ENDIAN__)", "vsx" },
+  #endif
+  #if defined(COMPILER_SUPPORTS_VSXNOFMA)
+    { 16, "defined(__VSX__) && defined(__ARCH_PWR8) && defined(__LITTLE_ENDIAN__)", "vsxnofma" },
+  #endif
+
+  /* x86 */
+
+  #if defined(COMPILER_SUPPORTS_AVX512F)
+    { 64, "defined(__AVX512F__) && defined(__FMA4__)", "avx512f" },
+  #endif
+  #if defined(COMPILER_SUPPORTS_AVX512FNOFMA)
+    { 64, "defined(__AVX512F__)", "avx512fnofma" },
+  #endif
+
+  #if defined(COMPILER_SUPPORTS_FMA4)
+    { 32, "defined(__FMA4__)", "fma4" },
+  #endif
+  #if defined(COMPILER_SUPPORTS_AVX2)
+    { 32, "defined(__AVX2__)", "avx2" },
+  #endif
+  #if defined(COMPILER_SUPPORTS_AVX)
+    { 32, "defined(__AVX__)", "avx" },
+  #endif
+
+  #if defined(COMPILER_SUPPORTS_AVX2128)
+    { 16, "defined(__AVX2__)", "avx2128" },
+  #endif
+  #if defined(COMPILER_SUPPORTS_SSE4)
+    { 16, "defined(__SSE4_1__)", "sse4" },
+  #endif
+  #if defined(COMPILER_SUPPORTS_SSE2)
+    { 16, "defined(__SSE2__)", "sse2" },
+  #endif
+
+  { 0, NULL, NULL }
+};
+
+void
+write_static_dispatch(FILE* fp, size_t vecSize) {
+  int first = 1;
+  for (const struct variantInfo* vi = variantList ; vi->suffix != NULL ; vi++) {
+    if (vi->vecSize != vecSize)
+      continue;
+
+    if (first) {
+      first = 0;
+      fprintf(fp, "#if %s\n", vi->ppTest);
+    } else {
+      fprintf(fp, "#elif %s\n", vi->ppTest);
+    }
+
+    for (funcSpec* func = funcList ; func->name != NULL ; func++) {
+      if (func->funcType >= 7) {
+        continue;
+      }
+
+      if (vecSize != 0) {
+        if (func->ulp > 0) {
+          fprintf(fp, "#  define Sleef_%sf%zu_u%02dstatic Sleef_%sf%zu_u%02d%s\n",
+                  func->name, vecSize / sizeof(float), func->ulp,
+                  func->name, vecSize / sizeof(float), func->ulp, vi->suffix);
+        } else {
+          fprintf(fp, "#  define Sleef_%sf%zu_static Sleef_%sf%zu_%s\n",
+                  func->name, vecSize / sizeof(float),
+                  func->name, vecSize / sizeof(float), vi->suffix);
+        }
+      } else {
+        if (func->ulp > 0) {
+          fprintf(fp, "#  define Sleef_%sfx_u%02dstatic Sleef_%sfx_u%02d%s\n",
+                  func->name, func->ulp,
+                  func->name, func->ulp, vi->suffix);
+        } else {
+          fprintf(fp, "#  define Sleef_%sfx_static Sleef_%sfx_%s\n",
+                  func->name,
+                  func->name, vi->suffix);
+        }
+      }
+    }
+  }
+
+  if (!first) {
+    fputs("#endif\n\n", fp);
+  }
+}
+
+int
+main(void) {
+  fprintf(stdout, "\n/* Begin static dispatch macros */\n\n");
+
+  for (size_t i = 0 ; i < (sizeof(vecSizes) / sizeof(vecSizes[0])) ; i++) {
+    write_static_dispatch(stdout, vecSizes[i]);
+  }
+
+  fprintf(stdout, "/* End static distpatch macros */\n\n");
+
+  return 0;
+}