From 563d3e5948a17328819f90b2ab20477256ada1f4 Mon Sep 17 00:00:00 2001
From: Andy Polyakov <appro@openssl.org>
Date: Fri, 14 Sep 2007 21:06:14 +0000
Subject: [PATCH] Engage new x86 assembler modules.

---
 Configure                    |  6 ++--
 TABLE                        | 56 ++++++++++++++++++------------------
 crypto/sha/Makefile          | 18 ++++++++----
 crypto/sha/asm/sha512-586.pl | 14 ++++++---
 4 files changed, 53 insertions(+), 41 deletions(-)

diff --git a/Configure b/Configure
index e26e6e4e6e..b35cc61459 100755
--- a/Configure
+++ b/Configure
@@ -116,9 +116,9 @@ my $tlib="-lnsl -lsocket";
 my $bits1="THIRTY_TWO_BIT ";
 my $bits2="SIXTY_FOUR_BIT ";
 
-my $x86_elf_asm="x86cpuid-elf.o:bn86-elf.o co86-elf.o mo86-elf.o:dx86-elf.o yx86-elf.o:ax86-elf.o:bx86-elf.o:mx86-elf.o:sx86-elf.o s512sse2-elf.o:cx86-elf.o:rx86-elf.o:rm86-elf.o:r586-elf.o:wp_block.o w86mmx-elf.o:";
-my $x86_coff_asm="x86cpuid-cof.o:bn86-cof.o co86-cof.o mo86-cof.o:dx86-cof.o yx86-cof.o:ax86-cof.o:bx86-cof.o:mx86-cof.o:sx86-cof.o s512sse2-cof.o:cx86-cof.o:rx86-cof.o:rm86-cof.o:r586-cof.o:wp_block.o w86mmx-cof.o:";
-my $x86_out_asm="x86cpuid-out.o:bn86-out.o co86-out.o mo86-out.o:dx86-out.o yx86-out.o:ax86-out.o:bx86-out.o:mx86-out.o:sx86-out.o s512sse2-out.o:cx86-out.o:rx86-out.o:rm86-out.o:r586-out.o:wp_block.o w86mmx-out.o:";
+my $x86_elf_asm="x86cpuid-elf.o:bn86-elf.o co86-elf.o mo86-elf.o:dx86-elf.o yx86-elf.o:ax86-elf.o:bx86-elf.o:mx86-elf.o:sx86-elf.o sha256x86-elf.o sha512x86-elf.o:cx86-elf.o:rx86-elf.o:rm86-elf.o:r586-elf.o:wp_block.o w86mmx-elf.o:";
+my $x86_coff_asm="x86cpuid-cof.o:bn86-cof.o co86-cof.o mo86-cof.o:dx86-cof.o yx86-cof.o:ax86-cof.o:bx86-cof.o:mx86-cof.o:sx86-cof.o sha256x86-cof.o sha512x86-cof.o:cx86-cof.o:rx86-cof.o:rm86-cof.o:r586-cof.o:wp_block.o w86mmx-cof.o:";
+my $x86_out_asm="x86cpuid-out.o:bn86-out.o co86-out.o mo86-out.o:dx86-out.o yx86-out.o:ax86-out.o:bx86-out.o:mx86-out.o:sx86-out.o sha256x86-out.o sha512x86-out.o:cx86-out.o:rx86-out.o:rm86-out.o:r586-out.o:wp_block.o w86mmx-out.o:";
 
 my $x86_64_asm="x86_64cpuid.o:x86_64-gcc.o x86_64-mont.o::aes-x86_64.o::md5-x86_64.o:sha1-x86_64.o sha256-x86_64.o sha512-x86_64.o::rc4-x86_64.o:::wp-x86_64.o:";
 my $ia64_asm="ia64cpuid.o:bn-ia64.o::aes_core.o aes_cbc.o aes-ia64.o::md5-ia64.o:sha1-ia64.o sha256-ia64.o sha512-ia64.o::rc4-ia64.o rc4_skey.o::::";
diff --git a/TABLE b/TABLE
index 195cc64860..3219b8bc10 100644
--- a/TABLE
+++ b/TABLE
@@ -217,7 +217,7 @@ $des_obj      = dx86-out.o yx86-out.o
 $aes_obj      = ax86-out.o
 $bf_obj       = bx86-out.o
 $md5_obj      = mx86-out.o
-$sha1_obj     = sx86-out.o s512sse2-out.o
+$sha1_obj     = sx86-out.o sha256x86-out.o sha512x86-out.o
 $cast_obj     = cx86-out.o
 $rc4_obj      = rx86-out.o
 $rmd160_obj   = rm86-out.o
@@ -246,7 +246,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -304,7 +304,7 @@ $des_obj      = dx86-cof.o yx86-cof.o
 $aes_obj      = ax86-cof.o
 $bf_obj       = bx86-cof.o
 $md5_obj      = mx86-cof.o
-$sha1_obj     = sx86-cof.o s512sse2-cof.o
+$sha1_obj     = sx86-cof.o sha256x86-cof.o sha512x86-cof.o
 $cast_obj     = cx86-cof.o
 $rc4_obj      = rx86-cof.o
 $rmd160_obj   = rm86-cof.o
@@ -362,7 +362,7 @@ $des_obj      = dx86-out.o yx86-out.o
 $aes_obj      = ax86-out.o
 $bf_obj       = bx86-out.o
 $md5_obj      = mx86-out.o
-$sha1_obj     = sx86-out.o s512sse2-out.o
+$sha1_obj     = sx86-out.o sha256x86-out.o sha512x86-out.o
 $cast_obj     = cx86-out.o
 $rc4_obj      = rx86-out.o
 $rmd160_obj   = rm86-out.o
@@ -913,7 +913,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -942,7 +942,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -971,7 +971,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -1232,7 +1232,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -1435,7 +1435,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -1522,7 +1522,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -1551,7 +1551,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -1638,7 +1638,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -1667,7 +1667,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -1725,7 +1725,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -1754,7 +1754,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -1783,7 +1783,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -1986,7 +1986,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -2131,7 +2131,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -2624,7 +2624,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -2943,7 +2943,7 @@ $des_obj      = dx86-out.o yx86-out.o
 $aes_obj      = ax86-out.o
 $bf_obj       = bx86-out.o
 $md5_obj      = mx86-out.o
-$sha1_obj     = sx86-out.o s512sse2-out.o
+$sha1_obj     = sx86-out.o sha256x86-out.o sha512x86-out.o
 $cast_obj     = cx86-out.o
 $rc4_obj      = rx86-out.o
 $rmd160_obj   = rm86-out.o
@@ -2972,7 +2972,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -3059,7 +3059,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -3378,7 +3378,7 @@ $des_obj      = dx86-cof.o yx86-cof.o
 $aes_obj      = ax86-cof.o
 $bf_obj       = bx86-cof.o
 $md5_obj      = mx86-cof.o
-$sha1_obj     = sx86-cof.o s512sse2-cof.o
+$sha1_obj     = sx86-cof.o sha256x86-cof.o sha512x86-cof.o
 $cast_obj     = cx86-cof.o
 $rc4_obj      = rx86-cof.o
 $rmd160_obj   = rm86-cof.o
@@ -3813,7 +3813,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -3842,7 +3842,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -4074,7 +4074,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -4422,7 +4422,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
@@ -4451,7 +4451,7 @@ $des_obj      = dx86-elf.o yx86-elf.o
 $aes_obj      = ax86-elf.o
 $bf_obj       = bx86-elf.o
 $md5_obj      = mx86-elf.o
-$sha1_obj     = sx86-elf.o s512sse2-elf.o
+$sha1_obj     = sx86-elf.o sha256x86-elf.o sha512x86-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
diff --git a/crypto/sha/Makefile b/crypto/sha/Makefile
index 1231f1da38..e02d9f1dc3 100644
--- a/crypto/sha/Makefile
+++ b/crypto/sha/Makefile
@@ -45,18 +45,24 @@ lib:    $(LIBOBJ)
 # ELF
 sx86-elf.s: asm/sha1-586.pl ../perlasm/x86asm.pl
 	(cd asm; $(PERL) sha1-586.pl elf $(CFLAGS) $(PROCESSOR) > ../$@)
-s512sse2-elf.s:	asm/sha512-sse2.pl ../perlasm/x86asm.pl
-	(cd asm; $(PERL) sha512-sse2.pl elf $(CFLAGS) $(PROCESSOR) > ../$@)
+sha256x86-elf.s:	asm/sha256-586.pl ../perlasm/x86asm.pl
+	(cd asm; $(PERL) sha256-586.pl elf $(CFLAGS) $(PROCESSOR) > ../$@)
+sha512x86-elf.s:	asm/sha512-586.pl ../perlasm/x86asm.pl
+	(cd asm; $(PERL) sha512-586.pl elf $(CFLAGS) $(PROCESSOR) > ../$@)
 # COFF
 sx86-cof.s: asm/sha1-586.pl ../perlasm/x86asm.pl
 	(cd asm; $(PERL) sha1-586.pl coff $(CFLAGS) $(PROCESSOR) > ../$@)
-s512sse2-cof.s:     asm/sha512-sse2.pl ../perlasm/x86asm.pl
-	(cd asm; $(PERL) sha512-sse2.pl coff $(CFLAGS) $(PROCESSOR) > ../$@)
+sha256x86-cof.s:	asm/sha256-586.pl ../perlasm/x86asm.pl
+	(cd asm; $(PERL) sha256-586.pl coff $(CFLAGS) $(PROCESSOR) > ../$@)
+sha512x86-cof.s:	asm/sha512-586.pl ../perlasm/x86asm.pl
+	(cd asm; $(PERL) sha512-586.pl coff $(CFLAGS) $(PROCESSOR) > ../$@)
 # a.out
 sx86-out.s: asm/sha1-586.pl ../perlasm/x86asm.pl
 	(cd asm; $(PERL) sha1-586.pl a.out $(CFLAGS) $(PROCESSOR) > ../$@)
-s512sse2-out.s:     asm/sha512-sse2.pl ../perlasm/x86asm.pl
-	(cd asm; $(PERL) sha512-sse2.pl a.out $(CFLAGS) $(PROCESSOR) > ../$@)
+sha256x86-out.s:	asm/sha256-586.pl ../perlasm/x86asm.pl
+	(cd asm; $(PERL) sha256-586.pl a.out $(CFLAGS) $(PROCESSOR) > ../$@)
+sha512x86-out.s:	asm/sha512-586.pl ../perlasm/x86asm.pl
+	(cd asm; $(PERL) sha512-586.pl a.out $(CFLAGS) $(PROCESSOR) > ../$@)
 
 sha1-ia64.s:   asm/sha1-ia64.pl
 	(cd asm; $(PERL) sha1-ia64.pl $(CFLAGS) ) > $@
diff --git a/crypto/sha/asm/sha512-586.pl b/crypto/sha/asm/sha512-586.pl
index 8f215ac247..49a9814544 100644
--- a/crypto/sha/asm/sha512-586.pl
+++ b/crypto/sha/asm/sha512-586.pl
@@ -38,6 +38,11 @@ require "x86asm.pl";
 
 &asm_init($ARGV[0],"sha512-586.pl",$ARGV[$#ARGV] eq "386");
 
+$sse2=0;
+for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
+
+&external_label("OPENSSL_ia32cap_P") if ($sse2);
+
 $Tlo=&DWP(0,"esp");	$Thi=&DWP(4,"esp");
 $Alo=&DWP(8,"esp");	$Ahi=&DWP(8+4,"esp");
 $Blo=&DWP(16,"esp");	$Bhi=&DWP(16+4,"esp");
@@ -70,7 +75,7 @@ sub BODY_00_15_sse2 {
 	&movq	("mm1",$E);			# %mm1 is sliding right
 	&movq	("mm2",$E);			# %mm2 is sliding left
 	&psrlq	("mm1",14);
-	&movq	($Esse2,$E);			# module-scheduled save e
+	&movq	($Esse2,$E);			# modulo-scheduled save e
 	&psllq	("mm2",23);
 	&movq	("mm3","mm1");			# %mm3 is T1
 	&psrlq	("mm1",4);
@@ -109,7 +114,7 @@ sub BODY_00_15_sse2 {
 	&pxor	("mm7","mm6");
 	&psllq	("mm6",6);
 	&pxor	("mm7","mm5");
-	&movq	(&QWP(0,"esp"),$A);		# module-scheduled save a
+	&movq	(&QWP(0,"esp"),$A);		# modulo-scheduled save a
 	&pxor	("mm7","mm6");			# T2=Sigma0_512(a)
 
 	&movq	("mm5",$A);			# %mm5=a
@@ -274,7 +279,8 @@ sub BODY_00_15_x86 {
 	&mov	(&DWP(8,"esp"),"eax");	# inp+num*128
 	&mov	(&DWP(12,"esp"),"ebx");	# saved sp
 
-	&picmeup("edx","OPENSSL_ia32cap_P",$K512,&label("pic_point"));
+if ($sse2) {
+	&picmeup("edx","OPENSSL_ia32cap_P",$K512,&label("K512"));
 	&bt	(&DWP(0,"edx"),26);
 	&jnc	(&label("loop_x86"));
 
@@ -403,7 +409,7 @@ sub BODY_00_15_x86 {
 	&emms	();
 	&mov	("esp",&DWP(8*10+12,"esp"));	# restore sp
 &function_end_A();
-
+}
 &set_label("loop_x86",16);
     # copy input block to stack reversing byte and qword order
     for ($i=0;$i<8;$i++) {
-- 
GitLab