From: Nathan Moinvaziri <nathan@nathanm.com>
Date: Sun, 27 Mar 2022 20:18:03 +0000 (-0700)
Subject: Allow SSE2 and AVX2 functions with -DWITH_UNALIGNED=OFF. Even though they use unalign... 
X-Git-Tag: 2.1.0-beta1~294
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=81633227813d4b275a2f5a6178b3af3e683db180;p=thirdparty%2Fzlib-ng.git

Allow SSE2 and AVX2 functions with -DWITH_UNALIGNED=OFF. Even though they use unaligned loads, they don't result in undefined behavior.
---

diff --git a/arch/x86/compare256_avx2.c b/arch/x86/compare256_avx2.c
index e25fa93eb..1318a0e33 100644
--- a/arch/x86/compare256_avx2.c
+++ b/arch/x86/compare256_avx2.c
@@ -14,8 +14,7 @@
 #  include <nmmintrin.h>
 #endif
 
-/* AVX2 unaligned intrinsic comparison */
-static inline uint32_t compare256_unaligned_avx2_static(const uint8_t *src0, const uint8_t *src1) {
+static inline uint32_t compare256_avx2_static(const uint8_t *src0, const uint8_t *src1) {
     uint32_t len = 0;
 
     do {
@@ -46,18 +45,18 @@ static inline uint32_t compare256_unaligned_avx2_static(const uint8_t *src0, con
     return 256;
 }
 
-Z_INTERNAL uint32_t compare256_unaligned_avx2(const uint8_t *src0, const uint8_t *src1) {
-    return compare256_unaligned_avx2_static(src0, src1);
+Z_INTERNAL uint32_t compare256_avx2(const uint8_t *src0, const uint8_t *src1) {
+    return compare256_avx2_static(src0, src1);
 }
 
-#define LONGEST_MATCH       longest_match_unaligned_avx2
-#define COMPARE256          compare256_unaligned_avx2_static
+#define LONGEST_MATCH       longest_match_avx2
+#define COMPARE256          compare256_avx2_static
 
 #include "match_tpl.h"
 
 #define LONGEST_MATCH_SLOW
-#define LONGEST_MATCH       longest_match_slow_unaligned_avx2
-#define COMPARE256          compare256_unaligned_avx2_static
+#define LONGEST_MATCH       longest_match_slow_avx2
+#define COMPARE256          compare256_avx2_static
 
 #include "match_tpl.h"
 
diff --git a/arch/x86/compare256_sse2.c b/arch/x86/compare256_sse2.c
index bd5d62cf7..aad4bd240 100644
--- a/arch/x86/compare256_sse2.c
+++ b/arch/x86/compare256_sse2.c
@@ -11,7 +11,7 @@
 
 #include <emmintrin.h>
 
-static inline uint32_t compare256_unaligned_sse2_static(const uint8_t *src0, const uint8_t *src1) {
+static inline uint32_t compare256_sse2_static(const uint8_t *src0, const uint8_t *src1) {
     uint32_t len = 0;
     int align_offset = ((uintptr_t)src0) & 15;
     const uint8_t *end0 = src0 + 256;
@@ -78,18 +78,18 @@ static inline uint32_t compare256_unaligned_sse2_static(const uint8_t *src0, con
     return 256;
 }
 
-Z_INTERNAL uint32_t compare256_unaligned_sse2(const uint8_t *src0, const uint8_t *src1) {
-    return compare256_unaligned_sse2_static(src0, src1);
+Z_INTERNAL uint32_t compare256_sse2(const uint8_t *src0, const uint8_t *src1) {
+    return compare256_sse2_static(src0, src1);
 }
 
-#define LONGEST_MATCH       longest_match_unaligned_sse2
-#define COMPARE256          compare256_unaligned_sse2_static
+#define LONGEST_MATCH       longest_match_sse2
+#define COMPARE256          compare256_sse2_static
 
 #include "match_tpl.h"
 
 #define LONGEST_MATCH_SLOW
-#define LONGEST_MATCH       longest_match_slow_unaligned_sse2
-#define COMPARE256          compare256_unaligned_sse2_static
+#define LONGEST_MATCH       longest_match_slow_sse2
+#define COMPARE256          compare256_sse2_static
 
 #include "match_tpl.h"
 
diff --git a/cpu_features.h b/cpu_features.h
index 7cc74a97a..560767217 100644
--- a/cpu_features.h
+++ b/cpu_features.h
@@ -120,12 +120,12 @@ extern uint32_t compare256_unaligned_32(const uint8_t *src0, const uint8_t *src1
 #if defined(UNALIGNED64_OK) && defined(HAVE_BUILTIN_CTZLL)
 extern uint32_t compare256_unaligned_64(const uint8_t *src0, const uint8_t *src1);
 #endif
+#endif
 #if defined(X86_SSE2) && defined(HAVE_BUILTIN_CTZ)
-extern uint32_t compare256_unaligned_sse2(const uint8_t *src0, const uint8_t *src1);
+extern uint32_t compare256_sse2(const uint8_t *src0, const uint8_t *src1);
 #endif
 #if defined(X86_AVX2) && defined(HAVE_BUILTIN_CTZ)
-extern uint32_t compare256_unaligned_avx2(const uint8_t *src0, const uint8_t *src1);
-#endif
+extern uint32_t compare256_avx2(const uint8_t *src0, const uint8_t *src1);
 #endif
 
 #ifdef DEFLATE_H_
@@ -147,12 +147,12 @@ extern uint32_t longest_match_unaligned_32(deflate_state *const s, Pos cur_match
 #if defined(UNALIGNED64_OK) && defined(HAVE_BUILTIN_CTZLL)
 extern uint32_t longest_match_unaligned_64(deflate_state *const s, Pos cur_match);
 #endif
+#endif
 #if defined(X86_SSE2) && defined(HAVE_BUILTIN_CTZ)
-extern uint32_t longest_match_unaligned_sse2(deflate_state *const s, Pos cur_match);
+extern uint32_t longest_match_sse2(deflate_state *const s, Pos cur_match);
 #endif
 #if defined(X86_AVX2) && defined(HAVE_BUILTIN_CTZ)
-extern uint32_t longest_match_unaligned_avx2(deflate_state *const s, Pos cur_match);
-#endif
+extern uint32_t longest_match_avx2(deflate_state *const s, Pos cur_match);
 #endif
 
 /* longest_match_slow */
@@ -163,12 +163,12 @@ extern uint32_t longest_match_slow_unaligned_32(deflate_state *const s, Pos cur_
 #ifdef UNALIGNED64_OK
 extern uint32_t longest_match_slow_unaligned_64(deflate_state *const s, Pos cur_match);
 #endif
+#endif
 #if defined(X86_SSE2) && defined(HAVE_BUILTIN_CTZ)
-extern uint32_t longest_match_slow_unaligned_sse2(deflate_state *const s, Pos cur_match);
+extern uint32_t longest_match_slow_sse2(deflate_state *const s, Pos cur_match);
 #endif
 #if defined(X86_AVX2) && defined(HAVE_BUILTIN_CTZ)
-extern uint32_t longest_match_slow_unaligned_avx2(deflate_state *const s, Pos cur_match);
-#endif
+extern uint32_t longest_match_slow_avx2(deflate_state *const s, Pos cur_match);
 #endif
 
 /* quick_insert_string */
diff --git a/functable.c b/functable.c
index 3e23c54e5..ffb0d3f47 100644
--- a/functable.c
+++ b/functable.c
@@ -106,17 +106,17 @@ Z_INTERNAL uint32_t longest_match_stub(deflate_state *const s, Pos cur_match) {
 #  else
     functable.longest_match = &longest_match_unaligned_16;
 #  endif
-#  if defined(X86_SSE2) && defined(HAVE_BUILTIN_CTZ)
-    if (x86_cpu_has_sse2)
-        functable.longest_match = &longest_match_unaligned_sse2;
-#  endif
-#  if defined(X86_AVX2) && defined(HAVE_BUILTIN_CTZ)
-    if (x86_cpu_has_avx2)
-        functable.longest_match = &longest_match_unaligned_avx2;
-#  endif
 #else
     functable.longest_match = &longest_match_c;
 #endif
+#if defined(X86_SSE2) && defined(HAVE_BUILTIN_CTZ)
+    if (x86_cpu_has_sse2)
+        functable.longest_match = &longest_match_sse2;
+#endif
+#if defined(X86_AVX2) && defined(HAVE_BUILTIN_CTZ)
+    if (x86_cpu_has_avx2)
+        functable.longest_match = &longest_match_avx2;
+#endif
 
     return functable.longest_match(s, cur_match);
 }
@@ -131,17 +131,17 @@ Z_INTERNAL uint32_t longest_match_slow_stub(deflate_state *const s, Pos cur_matc
 #  else
     functable.longest_match_slow = &longest_match_slow_unaligned_16;
 #  endif
-#  if defined(X86_SSE2) && defined(HAVE_BUILTIN_CTZ)
-    if (x86_cpu_has_sse2)
-        functable.longest_match_slow = &longest_match_slow_unaligned_sse2;
-#  endif
-#  if defined(X86_AVX2) && defined(HAVE_BUILTIN_CTZ)
-    if (x86_cpu_has_avx2)
-        functable.longest_match_slow = &longest_match_slow_unaligned_avx2;
-#  endif
 #else
     functable.longest_match_slow = &longest_match_slow_c;
 #endif
+#if defined(X86_SSE2) && defined(HAVE_BUILTIN_CTZ)
+    if (x86_cpu_has_sse2)
+        functable.longest_match_slow = &longest_match_slow_sse2;
+#endif
+#if defined(X86_AVX2) && defined(HAVE_BUILTIN_CTZ)
+    if (x86_cpu_has_avx2)
+        functable.longest_match_slow = &longest_match_slow_avx2;
+#endif
 
     return functable.longest_match_slow(s, cur_match);
 }
@@ -391,17 +391,17 @@ Z_INTERNAL uint32_t compare256_stub(const uint8_t *src0, const uint8_t *src1) {
 #  else
     functable.compare256 = &compare256_unaligned_16;
 #  endif
-#  if defined(X86_SSE2) && defined(HAVE_BUILTIN_CTZ)
-    if (x86_cpu_has_sse2)
-        functable.compare256 = &compare256_unaligned_sse2;
-#  endif
-#  if defined(X86_AVX2) && defined(HAVE_BUILTIN_CTZ)
-    if (x86_cpu_has_avx2)
-        functable.compare256 = &compare256_unaligned_avx2;
-#  endif
 #else
     functable.compare256 = &compare256_c;
 #endif
+#if defined(X86_SSE2) && defined(HAVE_BUILTIN_CTZ)
+    if (x86_cpu_has_sse2)
+        functable.compare256 = &compare256_sse2;
+#endif
+#if defined(X86_AVX2) && defined(HAVE_BUILTIN_CTZ)
+    if (x86_cpu_has_avx2)
+        functable.compare256 = &compare256_avx2;
+#endif
 
     return functable.compare256(src0, src1);
 }
diff --git a/test/benchmarks/benchmark_compare256.cc b/test/benchmarks/benchmark_compare256.cc
index bdfdd6bf9..e690e8eff 100644
--- a/test/benchmarks/benchmark_compare256.cc
+++ b/test/benchmarks/benchmark_compare256.cc
@@ -69,10 +69,10 @@ BENCHMARK_COMPARE256(unaligned_32, compare256_unaligned_32, 1);
 #if defined(UNALIGNED64_OK) && defined(HAVE_BUILTIN_CTZLL)
 BENCHMARK_COMPARE256(unaligned_64, compare256_unaligned_64, 1);
 #endif
+#endif
 #if defined(X86_SSE2) && defined(HAVE_BUILTIN_CTZ)
-BENCHMARK_COMPARE256(unaligned_sse2, compare256_unaligned_sse2, x86_cpu_has_sse2);
+BENCHMARK_COMPARE256(sse2, compare256_sse2, x86_cpu_has_sse2);
 #endif
 #if defined(X86_AVX2) && defined(HAVE_BUILTIN_CTZ)
-BENCHMARK_COMPARE256(unaligned_avx2, compare256_unaligned_avx2, x86_cpu_has_avx2);
-#endif
+BENCHMARK_COMPARE256(avx2, compare256_avx2, x86_cpu_has_avx2);
 #endif
diff --git a/test/test_compare256.cc b/test/test_compare256.cc
index 2c2f3aaa8..c252cfada 100644
--- a/test/test_compare256.cc
+++ b/test/test_compare256.cc
@@ -65,10 +65,10 @@ TEST_COMPARE256(unaligned_32, compare256_unaligned_32, 1)
 #if defined(UNALIGNED64_OK) && defined(HAVE_BUILTIN_CTZLL)
 TEST_COMPARE256(unaligned_64, compare256_unaligned_64, 1)
 #endif
+#endif
 #if defined(X86_SSE2) && defined(HAVE_BUILTIN_CTZ)
-TEST_COMPARE256(unaligned_sse2, compare256_unaligned_sse2, x86_cpu_has_sse2)
+TEST_COMPARE256(sse2, compare256_sse2, x86_cpu_has_sse2)
 #endif
 #if defined(X86_AVX2) && defined(HAVE_BUILTIN_CTZ)
-TEST_COMPARE256(unaligned_avx2, compare256_unaligned_avx2, x86_cpu_has_avx2)
-#endif
+TEST_COMPARE256(avx2, compare256_avx2, x86_cpu_has_avx2)
 #endif