Clean up SSE4.2 detection

zlib-ng · Apr 15, 2023 · b1aafe5 · b1aafe5
1 parent 7a60d93
commit b1aafe5
Show file tree

Hide file tree

Showing 7 changed files with 17 additions and 53 deletions.
diff --git a/.github/workflows/cmake.yml b/.github/workflows/cmake.yml
@@ -105,12 +105,12 @@ jobs:
             cmake-args: -DWITH_SSE2=OFF -DWITH_SANITIZER=Undefined
             codecov: ubuntu_gcc_no_sse2
 
-          - name: Ubuntu GCC No SSE4 UBSAN
+          - name: Ubuntu GCC No SSE4.2 UBSAN
             os: ubuntu-latest
             compiler: gcc
             cxx-compiler: g++
-            cmake-args: -DWITH_SSE4=OFF -DWITH_SANITIZER=Undefined
-            codecov: ubuntu_gcc_no_sse4
+            cmake-args: -DWITH_SSE42=OFF -DWITH_SANITIZER=Undefined
+            codecov: ubuntu_gcc_no_sse42
 
           - name: Ubuntu GCC No PCLMULQDQ UBSAN
             os: ubuntu-latest

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -810,9 +810,8 @@ if(WITH_OPTIM)
                 if(HAVE_SSE42CRC_INTRIN)
                     add_definitions(-DX86_SSE42_CRC_INTRIN)
                 endif()
-            endif()
-            if(NOT HAVE_SSE42CRC_INLINE_ASM AND NOT HAVE_SSE42CRC_INTRIN AND NOT HAVE_SSE42CMPSTR_INTRIN)
-                set(WITH_SSE4 OFF)
+            else()
+                set(WITH_SSE42 OFF)
             endif()
         endif()
         if(WITH_SSE2)

diff --git a/arch/x86/insert_string_sse42.c b/arch/x86/insert_string_sse42.c
@@ -41,9 +41,9 @@
 #define HASH_CALC_VAR       h
 #define HASH_CALC_VAR_INIT  uint32_t h = 0
 
-#define UPDATE_HASH         update_hash_sse4
-#define INSERT_STRING       insert_string_sse4
-#define QUICK_INSERT_STRING quick_insert_string_sse4
+#define UPDATE_HASH         update_hash_sse42
+#define INSERT_STRING       insert_string_sse42
+#define QUICK_INSERT_STRING quick_insert_string_sse42
 
 #ifdef X86_SSE42
 #  include "../../insert_string_tpl.h"

diff --git a/cmake/detect-intrinsics.cmake b/cmake/detect-intrinsics.cmake
@@ -447,7 +447,7 @@ macro(check_sse42_intrinsics)
             set(SSE42FLAG "-msse4.2")
         endif()
     endif()
-    # Check whether compiler supports SSE4 CRC inline asm
+    # Check whether compiler supports SSE4.2 CRC inline asm
     set(CMAKE_REQUIRED_FLAGS "${SSE42FLAG} ${NATIVEFLAG}")
     check_c_source_compile_or_run(
         "int main(void) {
@@ -461,7 +461,7 @@ macro(check_sse42_intrinsics)
         }"
         HAVE_SSE42CRC_INLINE_ASM
     )
-    # Check whether compiler supports SSE4 CRC intrinsics
+    # Check whether compiler supports SSE4.2 CRC intrinsics
     check_c_source_compile_or_run(
         "#include <immintrin.h>
         int main(void) {
@@ -477,19 +477,6 @@ macro(check_sse42_intrinsics)
         }"
         HAVE_SSE42CRC_INTRIN
     )
-    # Check whether compiler supports SSE4.2 compare string intrinsics
-    check_c_source_compile_or_run(
-        "#include <immintrin.h>
-        int main(void) {
-            unsigned char a[64] = { 0 };
-            unsigned char b[64] = { 0 };
-            __m128i xmm_src0, xmm_src1;
-            xmm_src0 = _mm_loadu_si128((__m128i *)(char *)a);
-            xmm_src1 = _mm_loadu_si128((__m128i *)(char *)b);
-            return _mm_cmpestri(xmm_src0, 16, xmm_src1, 16, 0);
-        }"
-        HAVE_SSE42CMPSTR_INTRIN
-    )
     set(CMAKE_REQUIRED_FLAGS)
 endmacro()
 

diff --git a/configure b/configure
@@ -262,7 +262,6 @@ if test $native -eq 1; then
   avx2flag=""
   sse2flag=""
   ssse3flag=""
-  sse4flag=""
   sse42flag=""
   pclmulflag=""
   vpclmulflag=""
@@ -1415,7 +1414,7 @@ EOF
 }
 
 check_sse42_intrinsics() {
-    # Check whether compiler supports SSE4 CRC inline asm
+    # Check whether compiler supports SSE4.2 CRC inline asm
     cat > $test.c << EOF
 int main(void) {
     unsigned val = 0, h = 0;
@@ -1448,27 +1447,6 @@ EOF
         echo "Checking for SSE4.2 CRC intrinsics ... No." | tee -a configure.log
         HAVE_SSE42CRC_INTRIN=0
     fi
-
-    # Check whether compiler supports SSE4.2 compare string intrinsics
-    cat > $test.c << EOF
-#include <immintrin.h>
-int main(void)
-{
-    unsigned char a[64] = { 0 };
-    unsigned char b[64] = { 0 };
-    __m128i xmm_src0, xmm_src1;
-    xmm_src0 = _mm_loadu_si128((__m128i *)(char *)a);
-    xmm_src1 = _mm_loadu_si128((__m128i *)(char *)b);
-    return _mm_cmpestri(xmm_src0, 16, xmm_src1, 16, 0);
-}
-EOF
-    if try ${CC} ${CFLAGS} ${sse42flag} $test.c; then
-        echo "Checking for SSE4.2 compare string intrinsics ... Yes." | tee -a configure.log
-        HAVE_SSE42CMPSTR_INTRIN=1
-    else
-        echo "Checking for SSE4.2 compare string intrinsics ... No." | tee -a configure.log
-        HAVE_SSE42CMPSTR_INTRIN=0
-    fi
 }
 
 check_ssse3_intrinsics() {

diff --git a/cpu_features.h b/cpu_features.h
@@ -181,7 +181,7 @@ extern uint32_t compare256_power9(const uint8_t *src0, const uint8_t *src1);
 /* insert_string */
 extern void insert_string_c(deflate_state *const s, const uint32_t str, uint32_t count);
 #ifdef X86_SSE42
-extern void insert_string_sse4(deflate_state *const s, const uint32_t str, uint32_t count);
+extern void insert_string_sse42(deflate_state *const s, const uint32_t str, uint32_t count);
 #elif defined(ARM_ACLE)
 extern void insert_string_acle(deflate_state *const s, const uint32_t str, uint32_t count);
 #endif
@@ -235,7 +235,7 @@ extern uint32_t longest_match_slow_power9(deflate_state *const s, Pos cur_match)
 /* quick_insert_string */
 extern Pos quick_insert_string_c(deflate_state *const s, const uint32_t str);
 #ifdef X86_SSE42
-extern Pos quick_insert_string_sse4(deflate_state *const s, const uint32_t str);
+extern Pos quick_insert_string_sse42(deflate_state *const s, const uint32_t str);
 #elif defined(ARM_ACLE)
 extern Pos quick_insert_string_acle(deflate_state *const s, const uint32_t str);
 #endif
@@ -261,7 +261,7 @@ extern void slide_hash_avx2(deflate_state *s);
 /* update_hash */
 extern uint32_t update_hash_c(deflate_state *const s, uint32_t h, uint32_t val);
 #ifdef X86_SSE42
-extern uint32_t update_hash_sse4(deflate_state *const s, uint32_t h, uint32_t val);
+extern uint32_t update_hash_sse42(deflate_state *const s, uint32_t h, uint32_t val);
 #elif defined(ARM_ACLE)
 extern uint32_t update_hash_acle(deflate_state *const s, uint32_t h, uint32_t val);
 #endif

diff --git a/functable.c b/functable.c
@@ -87,9 +87,9 @@ static void init_functable(void) {
 #ifdef X86_SSE42
     if (cf.x86.has_sse42) {
         ft.adler32_fold_copy = &adler32_fold_copy_sse42;
-        ft.insert_string = &insert_string_sse4;
-        ft.quick_insert_string = &quick_insert_string_sse4;
-        ft.update_hash = &update_hash_sse4;
+        ft.insert_string = &insert_string_sse42;
+        ft.quick_insert_string = &quick_insert_string_sse42;
+        ft.update_hash = &update_hash_sse42;
     }
 #endif
     // X86 - PCLMUL