From: Hans Kristian Rosbach <hk-git@circlestorm.org>
Date: Wed, 6 Mar 2019 09:39:32 +0000 (+0100)
Subject: Default sse2 to be available on x86_64 arch.
X-Git-Tag: 1.9.9-b1~513
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=17b7471d240d633aa0a63791becf4aab3069625a;p=thirdparty%2Fzlib-ng.git

Default sse2 to be available on x86_64 arch.
Allow x86 arch to force sse2 availability.
It still depends on sse2 intrinsics being detected and optimizations enabled.
---

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 647ad19e6..a0be580e2 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -49,6 +49,7 @@ include(CheckFunctionExists)
 include(CheckIncludeFile)
 include(CheckCSourceCompiles)
 include(CheckCSourceRuns)
+include(CMakeDependentOption)
 include(FeatureSummary)
 
 # make sure we use an appropriate BUILD_TYPE by default, "Release" to be exact
@@ -476,6 +477,11 @@ check_c_source_compiles(
     MFPU_NEON_AVAILABLE FAIL_REGEX "not supported")
 set(CMAKE_REQUIRED_FLAGS)
 
+# FORCE_SSE2 option will only be shown if HAVE_SSE2_INTRIN is true
+if("${ARCH}" MATCHES "i[3-6]86")
+    cmake_dependent_option(FORCE_SSE2 "Always assume CPU is SSE2 capable" OFF "HAVE_SSE2_INTRIN" OFF)
+endif()
+
 #
 # Enable deflate_medium at level 4-6
 #
@@ -507,9 +513,6 @@ set(ARCHDIR "arch/generic")
 if("${ARCH}" MATCHES "x86_64" OR "${ARCH}" MATCHES "AMD64" OR "${ARCH}" MATCHES "i[3-6]86")
     set(ARCHDIR "arch/x86")
     add_definitions(-DUNALIGNED_OK)
-    if("${ARCH}" MATCHES "x86_64" OR "${ARCH}" MATCHES "AMD64")
-      add_definitions(-DX86_NOCHECK_SSE2)
-    endif()
     add_feature_info(SSE2 1 "Support the SSE2 instruction set, using \"${SSE2FLAG}\"")
 elseif("${ARCH}" MATCHES "arm")
     set(ARCHDIR "arch/arm")
@@ -579,6 +582,10 @@ if(WITH_OPTIM)
             set(ZLIB_ARCH_SRCS ${ZLIB_ARCH_SRCS} ${ARCHDIR}/fill_window_sse.c)
             if(NOT ${ARCH} MATCHES "x86_64")
                 add_intrinsics_option("${SSE2FLAG}")
+                add_feature_info(FORCE_SSE2 FORCE_SSE2 "Assume CPU is SSE2 capable")
+                if(FORCE_SSE2)
+                    add_definitions(-DX86_NOCHECK_SSE2)
+                endif()
             endif()
         endif()
         if(HAVE_PCLMULQDQ_INTRIN)
diff --git a/configure b/configure
index 13942343c..4f9328696 100755
--- a/configure
+++ b/configure
@@ -97,6 +97,7 @@ with_msan=0
 with_fuzzers=0
 floatabi=
 native=0
+forcesse2=0
 sse2flag="-msse2"
 sse4flag="-msse4"
 sse42flag="-msse4.2"
@@ -144,6 +145,7 @@ case "$1" in
       echo '    [--without-optimizations]   Compiles without support for optional instruction sets' | tee -a configure.log
       echo '    [--without-new-strategies]  Compiles without using new additional deflate strategies' | tee -a configure.log
       echo '    [--acle] [--neon]           Compiles with additional instruction set enabled' | tee -a configure.log
+      echo '    [--force-sse2]              Assume SSE2 instructions are always available (disabled by default on x86, enabled on x86_64)' | tee -a configure.log
       echo '    [--with-sanitizers]         Build with address sanitizer and all supported sanitizers other than memory sanitizer (disabled by default)' | tee -a configure.log
       echo '    [--with-msan]               Build with memory sanitizer (disabled by default)' | tee -a configure.log
       echo '    [--with-fuzzers]            Build test/fuzz (disabled by default)' | tee -a configure.log
@@ -167,6 +169,7 @@ case "$1" in
     -6* | --64) build64=1; shift ;;
     --without-acle) buildacle=0; shift ;;
     --without-neon) buildneon=0; shift ;;
+    --force-sse2) forcesse2=1; shift ;;
     -n | --native) native=1; shift ;;
     -a*=* | --archs=*) ARCHS=`echo $1 | sed 's/.*=//'`; shift ;;
     --sysconfdir=*) echo "ignored option: --sysconfdir" | tee -a configure.log; shift ;;
@@ -973,6 +976,11 @@ case "${ARCH}" in
                 ARCH_STATIC_OBJS="${ARCH_STATIC_OBJS} fill_window_sse.o"
                 ARCH_SHARED_OBJS="${ARCH_SHARED_OBJS} fill_window_sse.lo"
 
+                if test $forcesse2 -eq 1; then
+                    CFLAGS="${CFLAGS} -DX86_NOCHECK_SSE2"
+                    SFLAGS="${SFLAGS} -DX86_NOCHECK_SSE2"
+                fi
+
                 # Enable deflate_quick at level 1?
                 # requires SSE2: code uses fill_window_sse
                 if test $without_new_strategies -eq 0; then
@@ -1008,8 +1016,8 @@ case "${ARCH}" in
     x86_64)
         ARCHDIR=arch/x86
 
-        CFLAGS="${CFLAGS} -DX86_NOCHECK_SSE2 -DUNALIGNED_OK"
-        SFLAGS="${SFLAGS} -DX86_NOCHECK_SSE2 -DUNALIGNED_OK"
+        CFLAGS="${CFLAGS} -DUNALIGNED_OK"
+        SFLAGS="${SFLAGS} -DUNALIGNED_OK"
 
         # Enable arch-specific optimizations?
         if test $without_optimizations -eq 0; then
diff --git a/functable.c b/functable.c
index b4f1d769e..449f66ccb 100644
--- a/functable.c
+++ b/functable.c
@@ -73,7 +73,7 @@ ZLIB_INTERNAL void fill_window_stub(deflate_state *s) {
     functable.fill_window=&fill_window_c;
 
     #ifdef X86_SSE2
-    # ifndef X86_NOCHECK_SSE2
+    # if !defined(__x86_64__) && !defined(_M_X64) && !defined(X86_NOCHECK_SSE2)
     if (x86_cpu_has_sse2)
     # endif
         functable.fill_window=&fill_window_sse;