Feature/add arm support

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 +281,16 @@ else()
 endif()
 
 CHECK_INCLUDE_FILES(unistd.h HAVE_UNISTD_H)
-CHECK_INCLUDE_FILES(intrin.h HAVE_C_INTRIN_H)
-CHECK_INCLUDE_FILE_CXX(intrin.h HAVE_CXX_INTRIN_H)
-CHECK_INCLUDE_FILES(x86intrin.h HAVE_C_X86INTRIN_H)
-CHECK_INCLUDE_FILE_CXX(x86intrin.h HAVE_CXX_X86INTRIN_H)
+if (ARCH_IA32 OR ARCH_X86_64)
+  CHECK_INCLUDE_FILES(intrin.h HAVE_C_INTRIN_H)
+  CHECK_INCLUDE_FILE_CXX(intrin.h HAVE_CXX_INTRIN_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)
+  set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -flax-vector-conversions")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -flax-vector-conversions")
+endif()
 
 CHECK_FUNCTION_EXISTS(posix_memalign HAVE_POSIX_MEMALIGN)
 CHECK_FUNCTION_EXISTS(_aligned_malloc HAVE__ALIGNED_MALLOC)
@@ -564,11 +570,22 @@ install(FILES ${hs_HEADERS} DESTINATION "${CMAKE_INSTALL_INCLUDEDIR}/hs")
 set (hs_exec_common_SRCS
     src/alloc.c
     src/scratch.c
-    src/util/cpuid_flags.c
-    src/util/cpuid_flags.h
+    src/util/arch/common/cpuid_flags.h
     src/util/multibit.c
     )
 
+if (ARCH_IA32 OR ARCH_X86_64)
+set (hs_exec_common_SRCS
+    ${hs_exec_common_SRCS}
+    src/util/arch/x86/cpuid_flags.c
+    )
+else (ARCH_ARM32 OR ARCH_AARCH64)
+set (hs_exec_common_SRCS
+    ${hs_exec_common_SRCS}
+    src/util/arch/arm/cpuid_flags.c
+    )
+endif ()
+
 set (hs_exec_SRCS
     ${hs_HEADERS}
     src/hs_version.h
@@ -694,7 +711,6 @@ set (hs_exec_SRCS
     src/util/exhaust.h
     src/util/fatbit.h
     src/util/join.h
-    src/util/masked_move.h
     src/util/multibit.h
     src/util/multibit.c
     src/util/multibit_compress.h
@@ -716,7 +732,8 @@ set (hs_exec_SRCS
 
 set (hs_exec_avx2_SRCS
     src/fdr/teddy_avx2.c
-    src/util/masked_move.c
+    src/util/arch/x86/masked_move.c
+    src/util/arch/x86/masked_move.h
 )
 
 

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
-CHECK_C_SOURCE_COMPILES("#include <${INTRIN_INC_H}>
+if (ARCH_IA32 OR ARCH_X86_64)
+    # ensure we have the minimum of SSSE3 - call a SSSE3 intrinsic
+    CHECK_C_SOURCE_COMPILES("#include <${INTRIN_INC_H}>
 int main() {
     __m128i a = _mm_set1_epi8(1);
     (void)_mm_shuffle_epi8(a, a);
 }" HAVE_SSSE3)
 
-# now look for AVX2
-CHECK_C_SOURCE_COMPILES("#include <${INTRIN_INC_H}>
+    # now look for AVX2
+    CHECK_C_SOURCE_COMPILES("#include <${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
-CHECK_C_SOURCE_COMPILES("#include <${INTRIN_INC_H}>
+    # and now for AVX512
+    CHECK_C_SOURCE_COMPILES("#include <${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
-CHECK_C_SOURCE_COMPILES("#include <${INTRIN_INC_H}>
+    # and now for AVX512VBMI
+    CHECK_C_SOURCE_COMPILES("#include <${INTRIN_INC_H}>
 #if !defined(__AVX512VBMI__)
 #error no avx512vbmi
 #endif
@@ -70,26 +74,39 @@ 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);
+    (void)a;
+}" 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

@@ -1,9 +1,15 @@
 # determine the target arch
 
 # really only interested in the preprocessor here
-CHECK_C_SOURCE_COMPILES("#if !(defined(__x86_64__) || defined(_M_X64))\n#error not 64bit\n#endif\nint main(void) { return 0; }" ARCH_64_BIT)
+CHECK_C_SOURCE_COMPILES("#if !(defined(__x86_64__) || defined(_M_X64))\n#error not 64bit\n#endif\nint main(void) { return 0; }" ARCH_X86_64)
 
-CHECK_C_SOURCE_COMPILES("#if !(defined(__i386__) || defined(_M_IX86))\n#error not 64bit\n#endif\nint main(void) { return 0; }" ARCH_32_BIT)
+CHECK_C_SOURCE_COMPILES("#if !(defined(__i386__) || defined(_M_IX86))\n#error not 32bit\n#endif\nint main(void) { return 0; }" ARCH_IA32)
 
-set(ARCH_X86_64 ${ARCH_64_BIT})
-set(ARCH_IA32 ${ARCH_32_BIT})
+CHECK_C_SOURCE_COMPILES("#if !defined(__ARM_ARCH_ISA_A64)\n#error not 64bit\n#endif\nint main(void) { return 0; }" ARCH_AARCH64)
+CHECK_C_SOURCE_COMPILES("#if !defined(__ARM_ARCH_ISA_ARM)\n#error not 32bit\n#endif\nint main(void) { return 0; }" ARCH_ARM32)
+
+if (ARCH_X86_64 OR ARCH_AARCH64)
+  set(ARCH_64_BIT TRUE)
+else()
+  set(ARCH_32_BIT TRUE)
+endif()

src/crc32.c

@@ -30,7 +30,6 @@
 #include "config.h"
 #include "ue2common.h"
 #include "util/arch.h"
-#include "util/intrinsics.h"
 
 #if !defined(HAVE_SSE42)
 
@@ -579,53 +578,7 @@ u32 crc32c_sb8_64_bit(u32 running_crc, const unsigned char* p_buf,
 }
 
 #else // HAVE_SSE42
-
-#ifdef ARCH_64_BIT
-#define CRC_WORD 8
-#define CRC_TYPE u64a
-#define CRC_FUNC _mm_crc32_u64
-#else
-#define CRC_WORD 4
-#define CRC_TYPE u32
-#define CRC_FUNC _mm_crc32_u32
-#endif
-
-/*
- * Use the crc32 instruction from SSE4.2 to compute our checksum - same
- * polynomial as the above function.
- */
-static really_inline
-u32 crc32c_sse42(u32 running_crc, const unsigned char* p_buf,
-                      const size_t length) {
-    u32 crc = running_crc;
-
-    // Process byte-by-byte until p_buf is aligned
-
-    const unsigned char *aligned_buf = ROUNDUP_PTR(p_buf, CRC_WORD);
-    size_t init_bytes = aligned_buf - p_buf;
-    size_t running_length = ((length - init_bytes)/CRC_WORD)*CRC_WORD;
-    size_t end_bytes = length - init_bytes - running_length;
-
-    while (p_buf < aligned_buf) {
-        crc = _mm_crc32_u8(crc, *p_buf++);
-    }
-
-    // Main aligned loop, processes a word at a time.
-
-    for (size_t li = 0; li < running_length/CRC_WORD; li++) {
-        CRC_TYPE block = *(const CRC_TYPE *)p_buf;
-        crc = CRC_FUNC(crc, block);
-        p_buf += CRC_WORD;
-    }
-
-    // Remaining bytes
-
-    for(size_t li = 0; li < end_bytes; li++) {
-        crc = _mm_crc32_u8(crc, *p_buf++);
-    }
-
-    return crc;
-}
+#include "util/arch/x86/crc32.h"
 #endif
 
 #ifdef VERIFY_ASSERTION

src/database.h

@@ -51,6 +51,7 @@ extern "C"
 // CPU type is the low 6 bits (we can't need more than 64, surely!)
 
 #define HS_PLATFORM_INTEL           1
+#define HS_PLATFORM_ARM             2
 #define HS_PLATFORM_CPU_MASK        0x3F
 
 #define HS_PLATFORM_NOAVX2          (4<<13)

src/dispatcher.c

@@ -30,7 +30,9 @@
 #include "hs_common.h"
 #include "hs_runtime.h"
 #include "ue2common.h"
-#include "util/cpuid_inline.h"
+#if defined(ARCH_X86_64)
+#include "util/arch/x86/cpuid_inline.h"
+#endif
 #include "util/join.h"
 
 #if defined(DISABLE_AVX512_DISPATCH)

src/fdr/fdr.c

@@ -36,6 +36,7 @@
 #include "teddy.h"
 #include "teddy_internal.h"
 #include "util/arch.h"
+#include "util/bitutils.h"
 #include "util/simd_utils.h"
 #include "util/uniform_ops.h"
 
@@ -119,20 +120,6 @@ const ALIGN_CL_DIRECTIVE u8 zone_or_mask[ITER_BYTES+1][ITER_BYTES] = {
       0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00 }
 };
 
-/* compilers don't reliably synthesize the 32-bit ANDN instruction here,
- * so we force its generation.
- */
-static really_inline
-u64a andn(const u32 a, const u8 *b) {
-    u64a r;
-#if defined(HAVE_BMI) && !defined(NO_ASM)
-    __asm__ ("andn\t%2,%1,%k0" : "=r"(r) : "r"(a), "m"(*(const u32 *)b));
-#else
-    r = unaligned_load_u32(b) & ~a;
-#endif
-    return r;
-}
-
 /* generates an initial state mask based on the last byte-ish of history rather
  * than being all accepting. If there is no history to consider, the state is
  * generated based on the minimum length of each bucket in order to prevent
@@ -739,6 +726,7 @@ hwlm_error_t fdr_engine_exec(const struct FDR *fdr,
     assert(ISALIGNED_CL(confBase));
     struct zone zones[ZONE_MAX];
     assert(fdr->domain > 8 && fdr->domain < 16);
+    memset(zones, 0, sizeof(zones));
 
     size_t numZone = prepareZones(a->buf, a->len,
                                   a->buf_history + a->len_history,