Merge pull request #212 from VectorCamp/bugfix/fix-simde-build

SIMDe on Clang needs SIMDE_NO_CHECK_IMMEDIATE_CONSTANT defined and other SIMDe related fixes now that SIMDe is part of the CI pipeline.

Some issue with SIMDe on x86 still remains because of an upstream bug:

https://github.com/simd-everywhere/simde/issues/1119

Similarly SIMDe native with clang on Arm also poses a non-high priority build failure:

https://buildbot-ci.vectorcamp.gr/#/builders/129/builds/11

Possibly a SIMDe issue as well, need to investigate but will merge this PR as these are non-blockers.

cmake/archdetect.cmake

@@ -44,7 +44,7 @@ if (USE_CPU_NATIVE)
         endif()
     elseif (CMAKE_COMPILER_IS_CLANG)
         if (ARCH_IA32 OR ARCH_X86_64)
-            set(GNUCC_ARCH x86_64_v2)
+            set(GNUCC_ARCH x86-64-v2)
             set(TUNE_FLAG generic)
         elseif(ARCH_AARCH64)
             if (BUILD_SVE2_BITPERM)
@@ -68,8 +68,22 @@ if (USE_CPU_NATIVE)
     endif()
 else()
     if (SIMDE_BACKEND)
+        if (ARCH_IA32 OR ARCH_X86_64)
+            set(GNUCC_ARCH x86-64-v2)
+            set(TUNE_FLAG generic)
+        elseif(ARCH_AARCH64)
+            set(GNUCC_ARCH armv8-a)
+            set(TUNE_FLAG generic)
+        elseif(ARCH_ARM32)
+            set(GNUCC_ARCH armv7a)
+            set(TUNE_FLAG generic)
+        elseif(ARCH_PPC64EL)
+            set(GNUCC_ARCH power8)
+            set(TUNE_FLAG power8)
+        else()
             set(GNUCC_ARCH native)
-        set(TUNE_FLAG native)
+            set(TUNE_FLAG generic)
+        endif()
     elseif (ARCH_IA32 OR ARCH_X86_64)
         set(GNUCC_ARCH native)
         set(TUNE_FLAG generic)

cmake/osdetection.cmake

@@ -7,16 +7,15 @@ if(CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
 endif(CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
 
 option(FAT_RUNTIME "Build a library that supports multiple microarchitectures" OFF)
-message("Checking Fat Runtime Requirements...")
+if (FAT_RUNTIME)
-if (FAT_RUNTIME AND NOT LINUX)
+    message("Checking Fat Runtime Requirements...")
+    if (NOT LINUX)
         message(FATAL_ERROR "Fat runtime is only supported on Linux OS")
-endif()
+    else()
-
+        if (USE_CPU_NATIVE AND FAT_RUNTIME)
-if (USE_CPU_NATIVE AND FAT_RUNTIME)
             message(FATAL_ERROR "Fat runtime is not compatible with Native CPU detection")
-endif()
+        endif()
 
-if (FAT_RUNTIME AND LINUX)
         if (NOT (ARCH_IA32 OR ARCH_X86_64 OR ARCH_AARCH64))
             message(FATAL_ERROR "Fat runtime is only supported on Intel and Aarch64 architectures")
         else()
@@ -32,9 +31,8 @@ if (FAT_RUNTIME AND LINUX)
                 endif()
             endif()
         endif()
+    endif()
     if (NOT RELEASE_BUILD)
         message(FATAL_ERROR "Fat runtime is only built on Release builds")
     endif()
 endif ()
-
-

cmake/simde.cmake

@@ -1,10 +1,16 @@
-include_directories(${PROJECT_SOURCE_DIR}/simde/simde)
+LIST(APPEND CMAKE_REQUIRED_INCLUDES ${PROJECT_SOURCE_DIR}/simde)
 
-CHECK_INCLUDE_FILES("simde/x86/sse4.2.h" SIMDE_SSE42_H_FOUND)
+CHECK_INCLUDE_FILES(simde/x86/sse4.2.h SIMDE_SSE42_H_FOUND)
 
 if (SIMDE_SSE42_H_FOUND)
   set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DVS_SIMDE_BACKEND")
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DVS_SIMDE_BACKEND")
+  include_directories(${PROJECT_SOURCE_DIR}/simde)
+
+  if (CMAKE_COMPILER_IS_CLANG)
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DSIMDE_NO_CHECK_IMMEDIATE_CONSTANT")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DSIMDE_NO_CHECK_IMMEDIATE_CONSTANT")
+  endif()
 
   if (SIMDE_NATIVE)
     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DVS_SIMDE_NATIVE -DSIMDE_ENABLE_OPENMP -fopenmp-simd")

src/util/arch/arm/bitutils.h

@@ -203,6 +203,11 @@ u64a pext64_impl(u64a x, u64a mask) {
     return pext64_impl_c(x, mask);
 }
 
+static really_inline
+u64a pdep64_impl(u64a x, u64a mask) {
+    return pdep64_impl_c(x, mask);
+}
+
 /* compilers don't reliably synthesize the 32-bit ANDN instruction here,
  * so we force its generation.
  */

src/util/arch/ppc64el/bitutils.h

@@ -201,7 +201,7 @@ u64a pext64_impl(u64a x, u64a mask) {
 }
 
 static really_inline
-u64a pdep64(u64a x, u64a mask) {
+u64a pdep64_impl(u64a x, u64a mask) {
     return pdep64_impl_c(x, mask);
 }
 

src/util/arch/x86/bitutils.h

@@ -282,9 +282,14 @@ u64a pext64_impl(u64a x, u64a mask) {
 
 #if defined(HAVE_BMI2) && defined(ARCH_64_BIT)
 static really_inline
-u64a pdep64(u64a x, u64a mask) {
+u64a pdep64_impl(u64a x, u64a mask) {
     return _pdep_u64(x, mask);
 }
+#else
+static really_inline
+u64a pdep64_impl(u64a x, u64a mask) {
+    return pdep64_impl_c(x, mask);
+}
 #endif
 
 /* compilers don't reliably synthesize the 32-bit ANDN instruction here,

src/util/bitutils.h

@@ -78,6 +78,7 @@
 #define rank_in_mask64_impl rank_in_mask64_impl_c
 #define pext32_impl pext32_impl_c
 #define pext64_impl pext64_impl_c
+#define pdep64_impl pdep64_impl_c
 #endif
 
 static really_inline
@@ -207,6 +208,11 @@ u64a pext64(u64a x, u64a mask) {
     return pext64_impl(x, mask);
 }
 
+static really_inline
+u64a pdep64(u64a x, u64a mask) {
+    return pdep64_impl(x, mask);
+}
+
 /* compilers don't reliably synthesize the 32-bit ANDN instruction here,
  * so we force its generation.
  */