From 52628ee3f51f011b463aaedb1a28aa0524b43cb3 Mon Sep 17 00:00:00 2001
From: miod <>
Date: Sun, 13 Apr 2014 15:16:40 +0000
Subject: Import OpenSSL 1.0.1g

---
 src/lib/libcrypto/bn/asm/mips-mont.pl        |  2 +-
 src/lib/libcrypto/bn/asm/mips.pl             | 46 +++++++++++++---------------
 src/lib/libcrypto/bn/asm/modexp512-x86_64.pl |  3 +-
 src/lib/libcrypto/bn/asm/parisc-mont.pl      |  4 ++-
 src/lib/libcrypto/bn/asm/x86_64-gf2m.pl      |  3 +-
 src/lib/libcrypto/bn/asm/x86_64-mont.pl      |  3 +-
 src/lib/libcrypto/bn/asm/x86_64-mont5.pl     |  7 +++--
 7 files changed, 36 insertions(+), 32 deletions(-)

(limited to 'src/lib/libcrypto/bn/asm')

diff --git a/src/lib/libcrypto/bn/asm/mips-mont.pl b/src/lib/libcrypto/bn/asm/mips-mont.pl
index b944a12b8e..caae04ed3a 100644
--- a/src/lib/libcrypto/bn/asm/mips-mont.pl
+++ b/src/lib/libcrypto/bn/asm/mips-mont.pl
@@ -133,7 +133,7 @@ $code.=<<___;
 	bnez	$at,1f
 	li	$t0,0
 	slt	$at,$num,17	# on in-order CPU
-	bnezl	$at,bn_mul_mont_internal
+	bnez	$at,bn_mul_mont_internal
 	nop
 1:	jr	$ra
 	li	$a0,0
diff --git a/src/lib/libcrypto/bn/asm/mips.pl b/src/lib/libcrypto/bn/asm/mips.pl
index c162a3ec23..d2f3ef7bbf 100644
--- a/src/lib/libcrypto/bn/asm/mips.pl
+++ b/src/lib/libcrypto/bn/asm/mips.pl
@@ -140,10 +140,10 @@ $code.=<<___;
 	.set	reorder
 	li	$minus4,-4
 	and	$ta0,$a2,$minus4
-	$LD	$t0,0($a1)
 	beqz	$ta0,.L_bn_mul_add_words_tail
 
 .L_bn_mul_add_words_loop:
+	$LD	$t0,0($a1)
 	$MULTU	$t0,$a3
 	$LD	$t1,0($a0)
 	$LD	$t2,$BNSZ($a1)
@@ -200,10 +200,9 @@ $code.=<<___;
 	$ADDU	$v0,$ta2
 	sltu	$at,$ta3,$at
 	$ST	$ta3,-$BNSZ($a0)
-	$ADDU	$v0,$at
 	.set	noreorder
-	bgtzl	$ta0,.L_bn_mul_add_words_loop
-	$LD	$t0,0($a1)
+	bgtz	$ta0,.L_bn_mul_add_words_loop
+	$ADDU	$v0,$at
 
 	beqz	$a2,.L_bn_mul_add_words_return
 	nop
@@ -300,10 +299,10 @@ $code.=<<___;
 	.set	reorder
 	li	$minus4,-4
 	and	$ta0,$a2,$minus4
-	$LD	$t0,0($a1)
 	beqz	$ta0,.L_bn_mul_words_tail
 
 .L_bn_mul_words_loop:
+	$LD	$t0,0($a1)
 	$MULTU	$t0,$a3
 	$LD	$t2,$BNSZ($a1)
 	$LD	$ta0,2*$BNSZ($a1)
@@ -341,10 +340,9 @@ $code.=<<___;
 	$ADDU	$v0,$at
 	sltu	$ta3,$v0,$at
 	$ST	$v0,-$BNSZ($a0)
-	$ADDU	$v0,$ta3,$ta2
 	.set	noreorder
-	bgtzl	$ta0,.L_bn_mul_words_loop
-	$LD	$t0,0($a1)
+	bgtz	$ta0,.L_bn_mul_words_loop
+	$ADDU	$v0,$ta3,$ta2
 
 	beqz	$a2,.L_bn_mul_words_return
 	nop
@@ -429,10 +427,10 @@ $code.=<<___;
 	.set	reorder
 	li	$minus4,-4
 	and	$ta0,$a2,$minus4
-	$LD	$t0,0($a1)
 	beqz	$ta0,.L_bn_sqr_words_tail
 
 .L_bn_sqr_words_loop:
+	$LD	$t0,0($a1)
 	$MULTU	$t0,$t0
 	$LD	$t2,$BNSZ($a1)
 	$LD	$ta0,2*$BNSZ($a1)
@@ -463,11 +461,10 @@ $code.=<<___;
 	mflo	$ta3
 	mfhi	$ta2
 	$ST	$ta3,-2*$BNSZ($a0)
-	$ST	$ta2,-$BNSZ($a0)
 
 	.set	noreorder
-	bgtzl	$ta0,.L_bn_sqr_words_loop
-	$LD	$t0,0($a1)
+	bgtz	$ta0,.L_bn_sqr_words_loop
+	$ST	$ta2,-$BNSZ($a0)
 
 	beqz	$a2,.L_bn_sqr_words_return
 	nop
@@ -547,10 +544,10 @@ $code.=<<___;
 	.set	reorder
 	li	$minus4,-4
 	and	$at,$a3,$minus4
-	$LD	$t0,0($a1)
 	beqz	$at,.L_bn_add_words_tail
 
 .L_bn_add_words_loop:
+	$LD	$t0,0($a1)
 	$LD	$ta0,0($a2)
 	subu	$a3,4
 	$LD	$t1,$BNSZ($a1)
@@ -589,11 +586,10 @@ $code.=<<___;
 	$ADDU	$t3,$ta3,$v0
 	sltu	$v0,$t3,$ta3
 	$ST	$t3,-$BNSZ($a0)
-	$ADDU	$v0,$t9
 	
 	.set	noreorder
-	bgtzl	$at,.L_bn_add_words_loop
-	$LD	$t0,0($a1)
+	bgtz	$at,.L_bn_add_words_loop
+	$ADDU	$v0,$t9
 
 	beqz	$a3,.L_bn_add_words_return
 	nop
@@ -679,10 +675,10 @@ $code.=<<___;
 	.set	reorder
 	li	$minus4,-4
 	and	$at,$a3,$minus4
-	$LD	$t0,0($a1)
 	beqz	$at,.L_bn_sub_words_tail
 
 .L_bn_sub_words_loop:
+	$LD	$t0,0($a1)
 	$LD	$ta0,0($a2)
 	subu	$a3,4
 	$LD	$t1,$BNSZ($a1)
@@ -722,11 +718,10 @@ $code.=<<___;
 	$SUBU	$t3,$ta3,$v0
 	sgtu	$v0,$t3,$ta3
 	$ST	$t3,-$BNSZ($a0)
-	$ADDU	$v0,$t9
 
 	.set	noreorder
-	bgtzl	$at,.L_bn_sub_words_loop
-	$LD	$t0,0($a1)
+	bgtz	$at,.L_bn_sub_words_loop
+	$ADDU	$v0,$t9
 
 	beqz	$a3,.L_bn_sub_words_return
 	nop
@@ -819,7 +814,7 @@ ___
 $code.=<<___;
 	.set	reorder
 	move	$ta3,$ra
-	bal	bn_div_words
+	bal	bn_div_words_internal
 	move	$ra,$ta3
 	$MULTU	$ta2,$v0
 	$LD	$t2,-2*$BNSZ($a3)
@@ -840,8 +835,9 @@ $code.=<<___;
 	sltu	$ta0,$a1,$a2
 	or	$t8,$ta0
 	.set	noreorder
-	beqzl	$at,.L_bn_div_3_words_inner_loop
+	beqz	$at,.L_bn_div_3_words_inner_loop
 	$SUBU	$v0,1
+	$ADDU	$v0,1
 	.set	reorder
 .L_bn_div_3_words_inner_loop_done:
 	.set	noreorder
@@ -902,7 +898,8 @@ $code.=<<___;
 	and	$t2,$a0
 	$SRL	$at,$a1,$t1
 	.set	noreorder
-	bnezl	$t2,.+8
+	beqz	$t2,.+12
+	nop
 	break	6		# signal overflow
 	.set	reorder
 	$SLL	$a0,$t9
@@ -917,7 +914,8 @@ $code.=<<___;
 	$SRL	$DH,$a2,4*$BNSZ	# bits
 	sgeu	$at,$a0,$a2
 	.set	noreorder
-	bnezl	$at,.+8
+	beqz	$at,.+12
+	nop
 	$SUBU	$a0,$a2
 	.set	reorder
 
diff --git a/src/lib/libcrypto/bn/asm/modexp512-x86_64.pl b/src/lib/libcrypto/bn/asm/modexp512-x86_64.pl
index 54aeb01921..bfd6e97541 100644
--- a/src/lib/libcrypto/bn/asm/modexp512-x86_64.pl
+++ b/src/lib/libcrypto/bn/asm/modexp512-x86_64.pl
@@ -68,7 +68,8 @@ $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
 die "can't locate x86_64-xlate.pl";
 
-open STDOUT,"| $^X $xlate $flavour $output";
+open OUT,"| \"$^X\" $xlate $flavour $output";
+*STDOUT=*OUT;
 
 use strict;
 my $code=".text\n\n";
diff --git a/src/lib/libcrypto/bn/asm/parisc-mont.pl b/src/lib/libcrypto/bn/asm/parisc-mont.pl
index 4a766a87fb..c02ef6f014 100644
--- a/src/lib/libcrypto/bn/asm/parisc-mont.pl
+++ b/src/lib/libcrypto/bn/asm/parisc-mont.pl
@@ -40,7 +40,7 @@
 # of arithmetic operations, most notably multiplications. It requires
 # more memory references, most notably to tp[num], but this doesn't
 # seem to exhaust memory port capacity. And indeed, dedicated PA-RISC
-# 2.0 code path, provides virtually same performance as pa-risc2[W].s:
+# 2.0 code path provides virtually same performance as pa-risc2[W].s:
 # it's ~10% better for shortest key length and ~10% worse for longest
 # one.
 #
@@ -988,6 +988,8 @@ foreach (split("\n",$code)) {
 	# assemble 2.0 instructions in 32-bit mode...
 	s/^\s+([a-z]+)([\S]*)\s+([\S]*)/&assemble($1,$2,$3)/e if ($BN_SZ==4);
 
+	s/\bbv\b/bve/gm	if ($SIZE_T==8);
+
 	print $_,"\n";
 }
 close STDOUT;
diff --git a/src/lib/libcrypto/bn/asm/x86_64-gf2m.pl b/src/lib/libcrypto/bn/asm/x86_64-gf2m.pl
index 1658acbbdd..226c66c35e 100644
--- a/src/lib/libcrypto/bn/asm/x86_64-gf2m.pl
+++ b/src/lib/libcrypto/bn/asm/x86_64-gf2m.pl
@@ -31,7 +31,8 @@ $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
 die "can't locate x86_64-xlate.pl";
 
-open STDOUT,"| $^X $xlate $flavour $output";
+open OUT,"| \"$^X\" $xlate $flavour $output";
+*STDOUT=*OUT;
 
 ($lo,$hi)=("%rax","%rdx");	$a=$lo;
 ($i0,$i1)=("%rsi","%rdi");
diff --git a/src/lib/libcrypto/bn/asm/x86_64-mont.pl b/src/lib/libcrypto/bn/asm/x86_64-mont.pl
index 5d79b35e1c..17fb94c84c 100755
--- a/src/lib/libcrypto/bn/asm/x86_64-mont.pl
+++ b/src/lib/libcrypto/bn/asm/x86_64-mont.pl
@@ -40,7 +40,8 @@ $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
 die "can't locate x86_64-xlate.pl";
 
-open STDOUT,"| $^X $xlate $flavour $output";
+open OUT,"| \"$^X\" $xlate $flavour $output";
+*STDOUT=*OUT;
 
 # int bn_mul_mont(
 $rp="%rdi";	# BN_ULONG *rp,
diff --git a/src/lib/libcrypto/bn/asm/x86_64-mont5.pl b/src/lib/libcrypto/bn/asm/x86_64-mont5.pl
index 057cda28aa..dae0fe2453 100755
--- a/src/lib/libcrypto/bn/asm/x86_64-mont5.pl
+++ b/src/lib/libcrypto/bn/asm/x86_64-mont5.pl
@@ -28,7 +28,8 @@ $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
 die "can't locate x86_64-xlate.pl";
 
-open STDOUT,"| $^X $xlate $flavour $output";
+open OUT,"| \"$^X\" $xlate $flavour $output";
+*STDOUT=*OUT;
 
 # int bn_mul_mont_gather5(
 $rp="%rdi";	# BN_ULONG *rp,
@@ -900,8 +901,8 @@ $code.=<<___;
 	jnz	.Lgather
 ___
 $code.=<<___ if ($win64);
-	movaps	%xmm6,(%rsp)
-	movaps	%xmm7,0x10(%rsp)
+	movaps	(%rsp),%xmm6
+	movaps	0x10(%rsp),%xmm7
 	lea	0x28(%rsp),%rsp
 ___
 $code.=<<___;
-- 
cgit v1.2.3-55-g6feb