Improve SSE version detection.

Add SSE4 detection.
Fix setting of some opt flags in Makefile.in.

Makefile.in

@@ -158,18 +158,18 @@ $(RABINOBJS) $(BSDIFFOBJS) $(LZPOBJS) $(DELTA2OBJS) @LIBBSCWRAPOBJ@ $(SKEINOBJS)
 $(SKEIN_BLOCK_OBJ) @SHA256ASM_OBJS@ @SHA256_OBJS@ $(KECCAK_OBJS) $(KECCAK_OBJS_ASM) \
 $(TRANSP_OBJS) $(CRYPTO_OBJS) $(ZLIB_OBJS) $(BZLIB_OBJS)
 
-DEBUG_LINK = g++ -m64 -pthread -msse3 @LIBBSCGEN_OPT@
+DEBUG_LINK = g++ -pthread @LIBBSCGEN_OPT@ @EXTRA_OPT_FLAGS@
-DEBUG_COMPILE = gcc -m64 -g -msse3 -c
+DEBUG_COMPILE = gcc -g -c @EXTRA_OPT_FLAGS@
-DEBUG_COMPILE_cpp = g++ -m64 -g -msse3 -c
+DEBUG_COMPILE_cpp = g++ -g -c @EXTRA_OPT_FLAGS@
 DEBUG_VEC_FLAGS = 
 DEBUG_LOOP_OPTFLAGS = 
 DEBUG_GEN_OPT = -O -fno-omit-frame-pointer @LIBBSCGEN_OPT@
 DEBUG_CPPFLAGS = $(COMMON_CPPFLAGS)
 DEBUG_FPTR_FLAG = 
 
-RELEASE_LINK = g++ -m64 -pthread -msse3 @LIBBSCGEN_OPT@ @M64_FLAG@
+RELEASE_LINK = g++ -pthread @LIBBSCGEN_OPT@ @EXTRA_OPT_FLAGS@
-RELEASE_COMPILE = gcc -m64 -msse3 -c @M64_FLAG@
+RELEASE_COMPILE = gcc -c @EXTRA_OPT_FLAGS@
-RELEASE_COMPILE_cpp = g++ -m64 -msse3 -c @M64_FLAG@
+RELEASE_COMPILE_cpp = g++ -c @EXTRA_OPT_FLAGS@
 RELEASE_VEC_FLAGS = $(COMMON_VEC_FLAGS)
 RELEASE_LOOP_OPTFLAGS = $(COMMON_LOOP_OPTFLAGS)
 RELEASE_CPPFLAGS = $(COMMON_CPPFLAGS) -DNDEBUG

config

@@ -50,9 +50,11 @@ yasm=yasm
 keccak_srcs=
 keccak_hdrs=
 keccak_srcs_asm=
-m64_flag=
+extra_opt_flags=
 zlib_prefix=
 bzlib_prefix=
+sse_detect=1
+default_sse="-msse2"
 
 rm -rf ./buildtmp
 mkdir ./buildtmp
@@ -98,7 +100,7 @@ then
 
 	# If m64 compilation succeeds we assume platform to be 64-bit capable but
 	# explicit flag is reqd.
-	m64_flag="-m64"
+	extra_opt_flags="-m64"
 fi
 rm -f tst tst.c
 
@@ -138,6 +140,9 @@ do
 	--use-key256)
 		keylen='-DKEYLEN=32'
 	;;
+	--no-sse-check)
+		sse_detect=0
+	;;
 	--help) usage $0;;
 	*)
 		echo "Unrecognized option: ${arg1}"
@@ -183,6 +188,33 @@ then
 	exit 1
 fi
 
+# SSE Detection
+if [ $sse_detect -eq 1 ]
+then
+	gcc -o sse_level ./utils/sse_level.c ./utils/cpuid.c -I./utils
+	if [ $? -ne 0 ]
+	then
+		echo "ERROR:"
+		echo "Failed to build SSE detection utility."
+		echo ""
+		exit 1
+	fi
+	sse_ver=`./sse_level`
+	if [ $? -ne 0 ]
+	then
+		rm -f sse_level
+		echo "ERROR:"
+		echo "SSE version detection utility. Try configuring with --no-sse-check option."
+		echo ""
+		exit 1
+	fi
+	rm -f sse_level
+	extra_opt_flags="${extra_opt_flags} -msse${sse_ver}"
+else
+	extra_opt_flags="${extra_opt_flags} ${default_sse}"
+fi
+
+
 echo $plat | egrep 'x86_64|amd64' > /dev/null
 if [ $? -eq 0 ]
 then
@@ -299,7 +331,7 @@ main(void)
 }
 __EOF
 
-gcc ${m64_flag} -I${openssl_incdir} -L${openssl_libdir} tst.c -o tst
+gcc ${extra_opt_flags} -I${openssl_incdir} -L${openssl_libdir} tst.c -o tst
 if [ $? -ne 0 ]
 then
 	echo "Unable to compile OpenSSL test program please check OpenSSL installation."
@@ -335,7 +367,7 @@ main(void)
 }
 __EOF
 
-gcc ${m64_flag} -I${openssl_incdir} -L${openssl_libdir} -O0 -g tst.c -o tst -lcrypto >/dev/null 2>&1
+gcc ${extra_opt_flags} -I${openssl_incdir} -L${openssl_libdir} -O0 -g tst.c -o tst -lcrypto >/dev/null 2>&1
 if [ $? -ne 0 ]
 then
 	openssl_incdir="${openssl_incdir} -D__OSSL_OLD__"
@@ -465,7 +497,7 @@ sha256asmobjsvar="SHA256ASM_OBJS"
 sha256objsvar="SHA256_OBJS"
 yasmvar="YASM"
 fptr_flag_var="FPTR_FLAG"
-m64_flag_var="M64_FLAG"
+extra_opt_flags_var="EXTRA_OPT_FLAGS"
 
 openssllibdirvar="OPENSSL_LIBDIR"
 opensslincdirvar="OPENSSL_INCDIR"
@@ -517,6 +549,6 @@ s#@${keccak_srcs_var}@#${keccak_srcs}#g
 s#@${keccak_hdrs_var}@#${keccak_hdrs}#g
 s#@${keccak_srcs_var}@#${keccak_srcs}#g
 s#@${keccak_srcs_asm_var}@#${keccak_srcs_asm}#g
-s#@${m64_flag_var}@#${m64_flag}#g
+s#@${extra_opt_flags_var}@#${extra_opt_flags}#g
 " > Makefile
 

utils/cpuid.c

@@ -29,6 +29,10 @@
 #include "cpuid.h"
 
 #ifdef	__x86_64__
+
+#define SSE4_1_FLAG     0x080000
+#define SSE4_2_FLAG     0x100000
+
 void
 exec_cpuid(uint32_t *regs)
 {
@@ -107,6 +111,7 @@ cpuid_basic_identify(processor_info_t *pc)
 	raw.vendor_str[12] = 0;
 	pc->avx_level = 0;
 	pc->sse_level = 0;
+	pc->sse_sub_level = 0;
 
 	if (strcmp(raw.vendor_str, "GenuineIntel") == 0) {
 		pc->proc_type = PROC_X64_INTEL;
@@ -119,8 +124,12 @@ cpuid_basic_identify(processor_info_t *pc)
 	if (raw.basic_cpuid[0][0] >= 1) {
 		// ECX has SSE 4.2 and AVX flags
 		// Bit 20 is SSE 4.2 and bit 28 indicates AVX
-		if (raw.basic_cpuid[1][2] & (1 << 20)) {
+		if (raw.basic_cpuid[1][2] & SSE4_1_FLAG) {
 			pc->sse_level = 4;
+			pc->sse_sub_level = 1;
+			if (raw.basic_cpuid[1][2] & SSE4_2_FLAG) {
+				pc->sse_sub_level = 2;
+			}
 		} else {
 			pc->sse_level = 3;
 		}

utils/cpuid.h

@@ -26,6 +26,8 @@
 #ifndef __CPUID_H__
 #define __CPUID_H__
 
+#include "utils.h"
+
 #ifdef	__x86_64__
 #define VENDOR_STR_MAX          16
 #define BRAND_STR_MAX           64

utils/sse_level.c

@@ -0,0 +1,17 @@
+#include <stdio.h>
+#include <utils.h>
+#include <cpuid.h>
+
+int
+main(void)
+{
+	processor_info_t pc;
+	cpuid_basic_identify(&pc);
+	printf("%d", pc.sse_level);
+	if (pc.sse_sub_level > 0)
+		printf(".%d\n", pc.sse_sub_level);
+	else
+		printf("\n");
+	return (0);
+}
+

utils/utils.h

@@ -146,6 +146,7 @@ typedef enum {
 
 typedef struct {
 	int sse_level;
+	int sse_sub_level;
 	int avx_level;
 	proc_type_t proc_type;
 } processor_info_t;