Fix detection of AVX512 support for Apple clang compiler

diff --git a/client/Makefile b/client/Makefile
index 5cdaebb0a1c752cd6672639247adfe0cd5874df2..fd2ae29017525657003fa32f6a4ef92d1971632a 100644 (file)
--- a/client/Makefile
+++ b/client/Makefile
@@ -184,14 +184,21 @@ MULTIARCHOBJS = $(MULTIARCHSRCS:%.c=$(OBJDIR)/%_NOSIMD.o) \
                        $(MULTIARCHSRCS:%.c=$(OBJDIR)/%_AVX.o) \
                        $(MULTIARCHSRCS:%.c=$(OBJDIR)/%_AVX2.o)
 
                        $(MULTIARCHSRCS:%.c=$(OBJDIR)/%_AVX.o) \
                        $(MULTIARCHSRCS:%.c=$(OBJDIR)/%_AVX2.o)
 

-GCC_GTEQ_490 := $(shell expr `gcc --version | awk '/gcc/{print $$NF;}' | sed -e 's/\.\([0-9][0-9]\)/\1/g' -e 's/\.\([0-9]\)/0\1/g' -e 's/^[0-9]\{3,4\}$$/&00/'` \>= 40900)
+GCC_VERSION := $(shell gcc --version | awk '/gcc/{print $$NF;}' | sed -e 's/\.\([0-9][0-9]\)/\1/g' -e 's/\.\([0-9]\)/0\1/g' -e 's/^[0-9]\{3,4\}$$/&00/')
+CLANG_VERSION := $(shell gcc --version | awk '/Apple LLVM version/{print $$4;}' | sed -e 's/\.\([0-9][0-9]\)/\1/g' -e 's/\.\([0-9]\)/0\1/g' -e 's/^[0-9]\{3,4\}$$/&00/')
+ifneq ($(CLANG_VERSION), )
+       SUPPORTS_AVX512 :=  $(shell [ $(CLANG_VERSION) -ge 80000 ] && echo "True" )
+endif
+ifneq ($(GCC_VERSION), )
+       SUPPORTS_AVX512 :=  $(shell [ $(GCC_VERSION) -ge 40900 ] && echo "True" )
+endif

 HARD_SWITCH_NOSIMD = -mno-mmx -mno-sse2 -mno-avx -mno-avx2
 HARD_SWITCH_MMX = -mmmx -mno-sse2 -mno-avx -mno-avx2
 HARD_SWITCH_SSE2 = -mmmx -msse2 -mno-avx -mno-avx2
 HARD_SWITCH_AVX = -mmmx -msse2 -mavx -mno-avx2
 HARD_SWITCH_AVX2 = -mmmx -msse2 -mavx -mavx2
 HARD_SWITCH_AVX512 = -mmmx -msse2 -mavx -mavx2 -mavx512f
 HARD_SWITCH_NOSIMD = -mno-mmx -mno-sse2 -mno-avx -mno-avx2
 HARD_SWITCH_MMX = -mmmx -mno-sse2 -mno-avx -mno-avx2
 HARD_SWITCH_SSE2 = -mmmx -msse2 -mno-avx -mno-avx2
 HARD_SWITCH_AVX = -mmmx -msse2 -mavx -mno-avx2
 HARD_SWITCH_AVX2 = -mmmx -msse2 -mavx -mavx2
 HARD_SWITCH_AVX512 = -mmmx -msse2 -mavx -mavx2 -mavx512f

-ifeq "$(GCC_GTEQ_490)" "1"
+ifeq "$(SUPPORTS_AVX512)" "True"

        HARD_SWITCH_NOSIMD += -mno-avx512f
        HARD_SWITCH_MMX += -mno-avx512f
        HARD_SWITCH_SSE2 += -mno-avx512f
        HARD_SWITCH_NOSIMD += -mno-avx512f
        HARD_SWITCH_MMX += -mno-avx512f
        HARD_SWITCH_SSE2 += -mno-avx512f

diff --git a/client/cmdhfmfhard.c b/client/cmdhfmfhard.c
index fd5fbbf9516c5bef5343a86245875cb3a888ebc3..0153541e187e8ea46119231ebf9b3e1ba9157763 100644 (file)
--- a/client/cmdhfmfhard.c
+++ b/client/cmdhfmfhard.c
@@ -72,7 +72,7 @@ static float brute_force_per_second;
 
 static void get_SIMD_instruction_set(char* instruction_set) {
 #if defined (__i386__) || defined (__x86_64__) 
 
 static void get_SIMD_instruction_set(char* instruction_set) {
 #if defined (__i386__) || defined (__x86_64__) 

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) strcpy(instruction_set, "AVX512F");
        else if (__builtin_cpu_supports("avx2")) strcpy(instruction_set, "AVX2");
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) strcpy(instruction_set, "AVX512F");
        else if (__builtin_cpu_supports("avx2")) strcpy(instruction_set, "AVX2");

diff --git a/client/hardnested/hardnested_bf_core.c b/client/hardnested/hardnested_bf_core.c
index 2388f6f5416bf0c2e54a00f89cecc2b83e321d8e..5e81c2ba80c157ffac8492426d4040133c2c0a43 100644 (file)
--- a/client/hardnested/hardnested_bf_core.c
+++ b/client/hardnested/hardnested_bf_core.c
@@ -551,7 +551,7 @@ bitslice_test_nonces_t *bitslice_test_nonces_function_p = &bitslice_test_nonces_
 // determine the available instruction set at runtime and call the correct function
 const uint64_t crack_states_bitsliced_dispatch(uint32_t cuid, uint8_t *best_first_bytes, statelist_t *p, uint32_t *keys_found, uint64_t *num_keys_tested, uint32_t nonces_to_bruteforce, uint8_t *bf_test_nonce_2nd_byte, noncelist_t *nonces) {
 #if defined (__i386__) || defined (__x86_64__)
 // determine the available instruction set at runtime and call the correct function
 const uint64_t crack_states_bitsliced_dispatch(uint32_t cuid, uint8_t *best_first_bytes, statelist_t *p, uint32_t *keys_found, uint64_t *num_keys_tested, uint32_t nonces_to_bruteforce, uint8_t *bf_test_nonce_2nd_byte, noncelist_t *nonces) {
 #if defined (__i386__) || defined (__x86_64__)

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2) 
                if (__builtin_cpu_supports("avx512f")) crack_states_bitsliced_function_p = &crack_states_bitsliced_AVX512;
                else if (__builtin_cpu_supports("avx2")) crack_states_bitsliced_function_p = &crack_states_bitsliced_AVX2;
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2) 
                if (__builtin_cpu_supports("avx512f")) crack_states_bitsliced_function_p = &crack_states_bitsliced_AVX512;
                else if (__builtin_cpu_supports("avx2")) crack_states_bitsliced_function_p = &crack_states_bitsliced_AVX2;


@@ -572,7 +572,7 @@ const uint64_t crack_states_bitsliced_dispatch(uint32_t cuid, uint8_t *best_firs
 
 void bitslice_test_nonces_dispatch(uint32_t nonces_to_bruteforce, uint32_t *bf_test_nonce, uint8_t *bf_test_nonce_par) {
 #if defined (__i386__) || defined (__x86_64__) 
 
 void bitslice_test_nonces_dispatch(uint32_t nonces_to_bruteforce, uint32_t *bf_test_nonce, uint8_t *bf_test_nonce_par) {
 #if defined (__i386__) || defined (__x86_64__) 

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
                if (__builtin_cpu_supports("avx512f")) bitslice_test_nonces_function_p = &bitslice_test_nonces_AVX512;
                else if (__builtin_cpu_supports("avx2")) bitslice_test_nonces_function_p = &bitslice_test_nonces_AVX2;
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
                if (__builtin_cpu_supports("avx512f")) bitslice_test_nonces_function_p = &bitslice_test_nonces_AVX512;
                else if (__builtin_cpu_supports("avx2")) bitslice_test_nonces_function_p = &bitslice_test_nonces_AVX2;

diff --git a/client/hardnested/hardnested_bitarray_core.c b/client/hardnested/hardnested_bitarray_core.c
index 5615d00606c2096d793a859c1afb34a882f453cb..aca4f149d958887e96b8cf16e924f1b2e5faf766 100644 (file)
--- a/client/hardnested/hardnested_bitarray_core.c
+++ b/client/hardnested/hardnested_bitarray_core.c
@@ -319,7 +319,7 @@ count_bitarray_AND4_t *count_bitarray_AND4_function_p = &count_bitarray_AND4_dis
 // determine the available instruction set at runtime and call the correct function
 uint32_t *malloc_bitarray_dispatch(uint32_t x) {
 #if defined (__i386__) || defined (__x86_64__) 
 // determine the available instruction set at runtime and call the correct function
 uint32_t *malloc_bitarray_dispatch(uint32_t x) {
 #if defined (__i386__) || defined (__x86_64__) 

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) malloc_bitarray_function_p = &malloc_bitarray_AVX512;
        else if (__builtin_cpu_supports("avx2")) malloc_bitarray_function_p = &malloc_bitarray_AVX2;
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) malloc_bitarray_function_p = &malloc_bitarray_AVX512;
        else if (__builtin_cpu_supports("avx2")) malloc_bitarray_function_p = &malloc_bitarray_AVX2;


@@ -340,7 +340,7 @@ uint32_t *malloc_bitarray_dispatch(uint32_t x) {
 
 void free_bitarray_dispatch(uint32_t *x) {
 #if defined (__i386__) || defined (__x86_64__)
 
 void free_bitarray_dispatch(uint32_t *x) {
 #if defined (__i386__) || defined (__x86_64__)

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) free_bitarray_function_p = &free_bitarray_AVX512;
        else if (__builtin_cpu_supports("avx2")) free_bitarray_function_p = &free_bitarray_AVX2;
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) free_bitarray_function_p = &free_bitarray_AVX512;
        else if (__builtin_cpu_supports("avx2")) free_bitarray_function_p = &free_bitarray_AVX2;


@@ -361,7 +361,7 @@ void free_bitarray_dispatch(uint32_t *x) {
 
 uint32_t bitcount_dispatch(uint32_t a) {
 #if defined (__i386__) || defined (__x86_64__) 
 
 uint32_t bitcount_dispatch(uint32_t a) {
 #if defined (__i386__) || defined (__x86_64__) 

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) bitcount_function_p = &bitcount_AVX512;
        else if (__builtin_cpu_supports("avx2")) bitcount_function_p = &bitcount_AVX2;
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) bitcount_function_p = &bitcount_AVX512;
        else if (__builtin_cpu_supports("avx2")) bitcount_function_p = &bitcount_AVX2;


@@ -382,7 +382,7 @@ uint32_t bitcount_dispatch(uint32_t a) {
 
 uint32_t count_states_dispatch(uint32_t *bitarray) {
 #if defined (__i386__) || defined (__x86_64__) 
 
 uint32_t count_states_dispatch(uint32_t *bitarray) {
 #if defined (__i386__) || defined (__x86_64__) 

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) count_states_function_p = &count_states_AVX512;
        else if (__builtin_cpu_supports("avx2")) count_states_function_p = &count_states_AVX2;
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) count_states_function_p = &count_states_AVX512;
        else if (__builtin_cpu_supports("avx2")) count_states_function_p = &count_states_AVX2;


@@ -403,7 +403,7 @@ uint32_t count_states_dispatch(uint32_t *bitarray) {
 
 void bitarray_AND_dispatch(uint32_t *A, uint32_t *B) {
 #if defined (__i386__) || defined (__x86_64__) 
 
 void bitarray_AND_dispatch(uint32_t *A, uint32_t *B) {
 #if defined (__i386__) || defined (__x86_64__) 

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) bitarray_AND_function_p = &bitarray_AND_AVX512;
        else if (__builtin_cpu_supports("avx2")) bitarray_AND_function_p = &bitarray_AND_AVX2;
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) bitarray_AND_function_p = &bitarray_AND_AVX512;
        else if (__builtin_cpu_supports("avx2")) bitarray_AND_function_p = &bitarray_AND_AVX2;


@@ -424,7 +424,7 @@ void bitarray_AND_dispatch(uint32_t *A, uint32_t *B) {
 
 void bitarray_low20_AND_dispatch(uint32_t *A, uint32_t *B) {
 #if defined (__i386__) || defined (__x86_64__)
 
 void bitarray_low20_AND_dispatch(uint32_t *A, uint32_t *B) {
 #if defined (__i386__) || defined (__x86_64__)

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) bitarray_low20_AND_function_p = &bitarray_low20_AND_AVX512;
        else if (__builtin_cpu_supports("avx2")) bitarray_low20_AND_function_p = &bitarray_low20_AND_AVX2;
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) bitarray_low20_AND_function_p = &bitarray_low20_AND_AVX512;
        else if (__builtin_cpu_supports("avx2")) bitarray_low20_AND_function_p = &bitarray_low20_AND_AVX2;


@@ -445,7 +445,7 @@ void bitarray_low20_AND_dispatch(uint32_t *A, uint32_t *B) {
 
 uint32_t count_bitarray_AND_dispatch(uint32_t *A, uint32_t *B) {
 #if defined (__i386__) || defined (__x86_64__) 
 
 uint32_t count_bitarray_AND_dispatch(uint32_t *A, uint32_t *B) {
 #if defined (__i386__) || defined (__x86_64__) 

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) count_bitarray_AND_function_p = &count_bitarray_AND_AVX512;
        else if (__builtin_cpu_supports("avx2")) count_bitarray_AND_function_p = &count_bitarray_AND_AVX2;
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) count_bitarray_AND_function_p = &count_bitarray_AND_AVX512;
        else if (__builtin_cpu_supports("avx2")) count_bitarray_AND_function_p = &count_bitarray_AND_AVX2;


@@ -466,7 +466,7 @@ uint32_t count_bitarray_AND_dispatch(uint32_t *A, uint32_t *B) {
 
 uint32_t count_bitarray_low20_AND_dispatch(uint32_t *A, uint32_t *B) {
 #if defined (__i386__) || defined (__x86_64__) 
 
 uint32_t count_bitarray_low20_AND_dispatch(uint32_t *A, uint32_t *B) {
 #if defined (__i386__) || defined (__x86_64__) 

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) count_bitarray_low20_AND_function_p = &count_bitarray_low20_AND_AVX512;
        else if (__builtin_cpu_supports("avx2")) count_bitarray_low20_AND_function_p = &count_bitarray_low20_AND_AVX2;
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) count_bitarray_low20_AND_function_p = &count_bitarray_low20_AND_AVX512;
        else if (__builtin_cpu_supports("avx2")) count_bitarray_low20_AND_function_p = &count_bitarray_low20_AND_AVX2;


@@ -487,7 +487,7 @@ uint32_t count_bitarray_low20_AND_dispatch(uint32_t *A, uint32_t *B) {
 
 void bitarray_AND4_dispatch(uint32_t *A, uint32_t *B, uint32_t *C, uint32_t *D) {
 #if defined (__i386__) || defined (__x86_64__) 
 
 void bitarray_AND4_dispatch(uint32_t *A, uint32_t *B, uint32_t *C, uint32_t *D) {
 #if defined (__i386__) || defined (__x86_64__) 

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) bitarray_AND4_function_p = &bitarray_AND4_AVX512;
        else if (__builtin_cpu_supports("avx2")) bitarray_AND4_function_p = &bitarray_AND4_AVX2;
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) bitarray_AND4_function_p = &bitarray_AND4_AVX512;
        else if (__builtin_cpu_supports("avx2")) bitarray_AND4_function_p = &bitarray_AND4_AVX2;


@@ -508,7 +508,7 @@ void bitarray_AND4_dispatch(uint32_t *A, uint32_t *B, uint32_t *C, uint32_t *D)
 
 void bitarray_OR_dispatch(uint32_t *A, uint32_t *B) {
 #if defined (__i386__) || defined (__x86_64__) 
 
 void bitarray_OR_dispatch(uint32_t *A, uint32_t *B) {
 #if defined (__i386__) || defined (__x86_64__) 

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) bitarray_OR_function_p = &bitarray_OR_AVX512;
        else if (__builtin_cpu_supports("avx2")) bitarray_OR_function_p = &bitarray_OR_AVX2;
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) bitarray_OR_function_p = &bitarray_OR_AVX512;
        else if (__builtin_cpu_supports("avx2")) bitarray_OR_function_p = &bitarray_OR_AVX2;


@@ -529,7 +529,7 @@ void bitarray_OR_dispatch(uint32_t *A, uint32_t *B) {
 
 uint32_t count_bitarray_AND2_dispatch(uint32_t *A, uint32_t *B) {
 #if defined (__i386__) || defined (__x86_64__) 
 
 uint32_t count_bitarray_AND2_dispatch(uint32_t *A, uint32_t *B) {
 #if defined (__i386__) || defined (__x86_64__) 

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) count_bitarray_AND2_function_p = &count_bitarray_AND2_AVX512;
        else if (__builtin_cpu_supports("avx2")) count_bitarray_AND2_function_p = &count_bitarray_AND2_AVX2;
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) count_bitarray_AND2_function_p = &count_bitarray_AND2_AVX512;
        else if (__builtin_cpu_supports("avx2")) count_bitarray_AND2_function_p = &count_bitarray_AND2_AVX2;


@@ -550,7 +550,7 @@ uint32_t count_bitarray_AND2_dispatch(uint32_t *A, uint32_t *B) {
 
 uint32_t count_bitarray_AND3_dispatch(uint32_t *A, uint32_t *B, uint32_t *C) {
 #if defined (__i386__) || defined (__x86_64__) 
 
 uint32_t count_bitarray_AND3_dispatch(uint32_t *A, uint32_t *B, uint32_t *C) {
 #if defined (__i386__) || defined (__x86_64__) 

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) count_bitarray_AND3_function_p = &count_bitarray_AND3_AVX512;
        else if (__builtin_cpu_supports("avx2")) count_bitarray_AND3_function_p = &count_bitarray_AND3_AVX2;
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) count_bitarray_AND3_function_p = &count_bitarray_AND3_AVX512;
        else if (__builtin_cpu_supports("avx2")) count_bitarray_AND3_function_p = &count_bitarray_AND3_AVX2;


@@ -571,7 +571,7 @@ uint32_t count_bitarray_AND3_dispatch(uint32_t *A, uint32_t *B, uint32_t *C) {
 
 uint32_t count_bitarray_AND4_dispatch(uint32_t *A, uint32_t *B, uint32_t *C, uint32_t *D) {
 #if defined (__i386__) || defined (__x86_64__) 
 
 uint32_t count_bitarray_AND4_dispatch(uint32_t *A, uint32_t *B, uint32_t *C, uint32_t *D) {
 #if defined (__i386__) || defined (__x86_64__) 

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
+       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8 || __clang_major__ == 8 && __clang_minor__ >= 1))

                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) count_bitarray_AND4_function_p = &count_bitarray_AND4_AVX512;
        else if (__builtin_cpu_supports("avx2")) count_bitarray_AND4_function_p = &count_bitarray_AND4_AVX2;
                #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
        if (__builtin_cpu_supports("avx512f")) count_bitarray_AND4_function_p = &count_bitarray_AND4_AVX512;
        else if (__builtin_cpu_supports("avx2")) count_bitarray_AND4_function_p = &count_bitarray_AND4_AVX2;