add necessary modifications to CMake system to enable building on ARM, add arm_neon.h intrinsic header to intrinsics.h

CMakeLists.txt

@@ -175,7 +175,7 @@ else()
         string(REGEX REPLACE "-O[^ ]*" "" CMAKE_CXX_FLAGS_${CONFIG} "${CMAKE_CXX_FLAGS_${CONFIG}}")
     endforeach ()
 
-    if (CMAKE_COMPILER_IS_GNUCC)
+    if (ARCH_IA32 OR ARCH_X86_64 AND CMAKE_COMPILER_IS_GNUCC)
         message(STATUS "gcc version ${CMAKE_C_COMPILER_VERSION}")
         # If gcc doesn't recognise the host cpu, then mtune=native becomes
         # generic, which isn't very good in some cases. march=native looks at
@@ -281,10 +281,14 @@ else()
 endif()
 
 CHECK_INCLUDE_FILES(unistd.h HAVE_UNISTD_H)
-CHECK_INCLUDE_FILES(intrin.h HAVE_C_INTRIN_H)
+if (ARCH_IA32 OR ARCH_X86_64)
-CHECK_INCLUDE_FILE_CXX(intrin.h HAVE_CXX_INTRIN_H)
+  CHECK_INCLUDE_FILES(intrin.h HAVE_C_INTRIN_H)
-CHECK_INCLUDE_FILES(x86intrin.h HAVE_C_X86INTRIN_H)
+  CHECK_INCLUDE_FILE_CXX(intrin.h HAVE_CXX_INTRIN_H)
-CHECK_INCLUDE_FILE_CXX(x86intrin.h HAVE_CXX_X86INTRIN_H)
+  CHECK_INCLUDE_FILES(x86intrin.h HAVE_C_X86INTRIN_H)
+  CHECK_INCLUDE_FILE_CXX(x86intrin.h HAVE_CXX_X86INTRIN_H)
+elseif (ARCH_ARM32 OR ARCH_AARCH64)
+  CHECK_INCLUDE_FILE_CXX(arm_neon.h HAVE_C_ARM_NEON_H)
+endif()
 
 CHECK_FUNCTION_EXISTS(posix_memalign HAVE_POSIX_MEMALIGN)
 CHECK_FUNCTION_EXISTS(_aligned_malloc HAVE__ALIGNED_MALLOC)

cmake/arch.cmake

@@ -6,7 +6,10 @@ if (HAVE_C_X86INTRIN_H)
     set (INTRIN_INC_H "x86intrin.h")
 elseif (HAVE_C_INTRIN_H)
     set (INTRIN_INC_H "intrin.h")
-else ()
+elseif (HAVE_C_ARM_NEON_H)
+    set (INTRIN_INC_H "arm_neon.h")
+    set (FAT_RUNTIME OFF)
+else()
     message (FATAL_ERROR "No intrinsics header found")
 endif ()
 
@@ -29,15 +32,16 @@ else (NOT FAT_RUNTIME)
     set (CMAKE_REQUIRED_FLAGS "${CMAKE_C_FLAGS} ${EXTRA_C_FLAGS} ${ARCH_C_FLAGS}")
 endif ()
 
-# ensure we have the minimum of SSSE3 - call a SSSE3 intrinsic
+if (ARCH_IA32 OR ARCH_X86_64)
-CHECK_C_SOURCE_COMPILES("#include <${INTRIN_INC_H}>
+    # ensure we have the minimum of SSSE3 - call a SSSE3 intrinsic
+    CHECK_C_SOURCE_COMPILES("#include <${INTRIN_INC_H}>
 int main() {
     __m128i a = _mm_set1_epi8(1);
     (void)_mm_shuffle_epi8(a, a);
 }" HAVE_SSSE3)
 
-# now look for AVX2
+    # now look for AVX2
-CHECK_C_SOURCE_COMPILES("#include <${INTRIN_INC_H}>
+    CHECK_C_SOURCE_COMPILES("#include <${INTRIN_INC_H}>
 #if !defined(__AVX2__)
 #error no avx2
 #endif
@@ -47,8 +51,8 @@ int main(){
     (void)_mm256_xor_si256(z, z);
 }" HAVE_AVX2)
 
-# and now for AVX512
+    # and now for AVX512
-CHECK_C_SOURCE_COMPILES("#include <${INTRIN_INC_H}>
+    CHECK_C_SOURCE_COMPILES("#include <${INTRIN_INC_H}>
 #if !defined(__AVX512BW__)
 #error no avx512bw
 #endif
@@ -58,8 +62,8 @@ int main(){
     (void)_mm512_abs_epi8(z);
 }" HAVE_AVX512)
 
-# and now for AVX512VBMI
+    # and now for AVX512VBMI
-CHECK_C_SOURCE_COMPILES("#include <${INTRIN_INC_H}>
+    CHECK_C_SOURCE_COMPILES("#include <${INTRIN_INC_H}>
 #if !defined(__AVX512VBMI__)
 #error no avx512vbmi
 #endif
@@ -70,26 +74,38 @@ int main(){
     (void)_mm512_permutexvar_epi8(idx, a);
 }" HAVE_AVX512VBMI)
 
+elseif (ARCH_ARM32 OR ARCH_AARCH64)
+    CHECK_C_SOURCE_COMPILES("#include <${INTRIN_INC_H}>
+int main() {
+    int32x4_t a = vdupq_n_s32(1);
+}" HAVE_NEON)
+else ()
+    message (FATAL_ERROR "Unsupported architecture")
+endif ()
+
 if (FAT_RUNTIME)
-    if (NOT HAVE_SSSE3)
+    if ((ARCH_IA32 OR ARCH_X86_64) AND NOT HAVE_SSSE3)
         message(FATAL_ERROR "SSSE3 support required to build fat runtime")
     endif ()
-    if (NOT HAVE_AVX2)
+    if ((ARCH_IA32 OR ARCH_X86_64) AND NOT HAVE_AVX2)
         message(FATAL_ERROR "AVX2 support required to build fat runtime")
     endif ()
-    if (BUILD_AVX512 AND NOT HAVE_AVX512)
+    if ((ARCH_IA32 OR ARCH_X86_64) AND BUILD_AVX512 AND NOT HAVE_AVX512)
         message(FATAL_ERROR "AVX512 support requested but not supported")
     endif ()
 else (NOT FAT_RUNTIME)
-    if (NOT HAVE_AVX2)
+    if ((ARCH_IA32 OR ARCH_X86_64) AND NOT HAVE_AVX2)
         message(STATUS "Building without AVX2 support")
     endif ()
-    if (NOT HAVE_AVX512)
+    if ((ARCH_IA32 OR ARCH_X86_64) AND NOT HAVE_AVX512)
         message(STATUS "Building without AVX512 support")
     endif ()
-    if (NOT HAVE_SSSE3)
+    if ((ARCH_IA32 OR ARCH_X86_64) AND NOT HAVE_SSSE3)
         message(FATAL_ERROR "A minimum of SSSE3 compiler support is required")
     endif ()
+    if ((ARCH_ARM32 OR ARCH_AARCH64) AND NOT HAVE_NEON)
+        message(FATAL_ERROR "NEON support required for ARM support")
+    endif ()
 endif ()
 
 unset (CMAKE_REQUIRED_FLAGS)

cmake/config.h.in

@@ -15,6 +15,12 @@
 /* "Define if building for EM64T" */
 #cmakedefine ARCH_X86_64
 
+/* "Define if building for ARM32" */
+#cmakedefine ARCH_ARM32
+
+/* "Define if building for AARCH64" */
+#cmakedefine ARCH_AARCH64
+
 /* internal build, switch on dump support. */
 #cmakedefine DUMP_SUPPORT
 
@@ -45,6 +51,9 @@
 /* C compiler has intrin.h */
 #cmakedefine HAVE_C_INTRIN_H
 
+/* C compiler has arm_neon.h */
+#cmakedefine HAVE_C_ARM_NEON_H
+
 /* Define to 1 if you have the declaration of `pthread_setaffinity_np', and to
    0 if you don't. */
 #cmakedefine HAVE_DECL_PTHREAD_SETAFFINITY_NP

cmake/platform.cmake

@@ -5,10 +5,10 @@ CHECK_C_SOURCE_COMPILES("#if !(defined(__x86_64__) || defined(_M_X64))\n#error n
 
 CHECK_C_SOURCE_COMPILES("#if !(defined(__i386__) || defined(_M_IX86))\n#error not 32bit\n#endif\nint main(void) { return 0; }" ARCH_IA32)
 
-CHECK_C_SOURCE_COMPILES("#if !defined(__aarch64__)\n#error not 64bit\n#endif\nint main(void) { return 0; }" ARCH_ARM64)
+CHECK_C_SOURCE_COMPILES("#if !defined(__aarch64__)\n#error not 64bit\n#endif\nint main(void) { return 0; }" ARCH_AARCH64)
 CHECK_C_SOURCE_COMPILES("#if !(defined(__arm__) && !defined(__aarch64__))\n#error not 32bit\n#endif\nint main(void) { return 0; }" ARCH_ARM32)
 
-if (DEFINED(ARCH_X86_64) OR DEFINED(ARCH_ARM64))
+if (DEFINED(ARCH_X86_64) OR DEFINED(ARCH_AARCH64))
   set(ARCH_64_BIT TRUE)
 else()
   set(ARCH_32_BIT TRUE)

src/util/intrinsics.h

@@ -45,6 +45,10 @@
 # endif
 #endif
 
+#if defined(HAVE_C_ARM_NEON_H)
+#  define USE_ARM_NEON_H
+#endif
+
 #ifdef __cplusplus
 # if defined(HAVE_CXX_INTRIN_H)
 #  define USE_INTRIN_H
@@ -59,6 +63,8 @@
 #include <x86intrin.h>
 #elif defined(USE_INTRIN_H)
 #include <intrin.h>
+#elif defined(USE_ARM_NEON_H)
+#include <arm_neon.h>
 #else
 #error no intrinsics file
 #endif