From: Cameron Cawley <ccawley2011@gmail.com>
Date: Thu, 22 Feb 2024 12:09:51 +0000 (+0000)
Subject: Remove mixing of SSE2 and SSSE3 chunkset code
X-Git-Tag: 2.2.0~56
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=d745a9bc24e523853de36a4206d8e04acd3c67c7;p=thirdparty%2Fzlib-ng.git

Remove mixing of SSE2 and SSSE3 chunkset code
---

diff --git a/arch/x86/chunkset_ssse3.c b/arch/x86/chunkset_ssse3.c
index 7ff16a4f..722ecd3d 100644
--- a/arch/x86/chunkset_ssse3.c
+++ b/arch/x86/chunkset_ssse3.c
@@ -4,9 +4,6 @@
 
 #include "zbuild.h"
 
-/* This requires SSE2 support. While it's implicit with SSSE3, we can minimize
- * code size by sharing the chunkcopy functions, which will certainly compile
- * to identical machine code */
 #if defined(X86_SSSE3)
 #include <immintrin.h>
 #include "../generic/chunk_permute_table.h"
@@ -19,8 +16,6 @@ typedef __m128i chunk_t;
 #define HAVE_CHUNKMEMSET_4
 #define HAVE_CHUNKMEMSET_8
 #define HAVE_CHUNK_MAG
-#define HAVE_CHUNKCOPY
-#define HAVE_CHUNKUNROLL
 
 static const lut_rem_pair perm_idx_lut[13] = {
     {0, 1},      /* 3 */
@@ -83,14 +78,11 @@ static inline chunk_t GET_CHUNK_MAG(uint8_t *buf, uint32_t *chunk_rem, uint32_t
     return ret_vec;
 }
 
-extern uint8_t* chunkcopy_sse2(uint8_t *out, uint8_t const *from, unsigned len);
-extern uint8_t* chunkunroll_sse2(uint8_t *out, unsigned *dist, unsigned *len);
-
 #define CHUNKSIZE        chunksize_ssse3
 #define CHUNKMEMSET      chunkmemset_ssse3
 #define CHUNKMEMSET_SAFE chunkmemset_safe_ssse3
-#define CHUNKCOPY        chunkcopy_sse2
-#define CHUNKUNROLL      chunkunroll_sse2
+#define CHUNKCOPY        chunkcopy_ssse3
+#define CHUNKUNROLL      chunkunroll_ssse3
 
 #include "chunkset_tpl.h"