Regen assembly files for i386.

secure/lib/libcrypto/Makefile.asm

@@ -59,6 +59,9 @@ sha256-armv8.S:	sha512-armv8.pl
 	${LCRYPTO_SRC}/crypto/whrlpool/asm \
 	${LCRYPTO_SRC}/engines/asm
 
+# cpuid
+SRCS+=	x86_64cpuid.pl
+
 # aes
 SRCS=	aes-x86_64.pl aesni-mb-x86_64.pl aesni-sha1-x86_64.pl \
 	aesni-sha256-x86_64.pl aesni-x86_64.pl bsaes-x86_64.pl \
@@ -77,9 +80,6 @@ SRCS+=	chacha-x86_64.pl
 # ec
 SRCS+=	ecp_nistz256-x86_64.pl x25519-x86_64.pl
 
-# engines
-SRCS+=	e_padlock-x86_64.pl
-
 # md5
 SRCS+=	md5-x86_64.pl
 
@@ -99,8 +99,8 @@ SRCS+=	keccak1600-x86_64.pl sha1-mb-x86_64.pl sha1-x86_64.pl \
 # whrlpool
 SRCS+=	wp-x86_64.pl
 
-# cpuid
+# engines
-SRCS+=	x86_64cpuid.pl
+SRCS+=	e_padlock-x86_64.pl
 
 SHA_ASM=	sha256-x86_64 sha512-x86_64
 SHA_SRC=	sha512-x86_64.pl
@@ -175,22 +175,30 @@ aes-armv4.S:	aes-armv4.pl
 	${LCRYPTO_SRC}/crypto/bf/asm \
 	${LCRYPTO_SRC}/crypto/bn/asm \
 	${LCRYPTO_SRC}/crypto/camellia/asm \
+	${LCRYPTO_SRC}/crypto/cast/asm \
+	${LCRYPTO_SRC}/crypto/chacha/asm \
 	${LCRYPTO_SRC}/crypto/des/asm \
+	${LCRYPTO_SRC}/crypto/ec/asm \
 	${LCRYPTO_SRC}/crypto/md5/asm \
 	${LCRYPTO_SRC}/crypto/modes/asm \
+	${LCRYPTO_SRC}/crypto/poly1305/asm \
 	${LCRYPTO_SRC}/crypto/rc4/asm \
 	${LCRYPTO_SRC}/crypto/rc5/asm \
 	${LCRYPTO_SRC}/crypto/ripemd/asm \
 	${LCRYPTO_SRC}/crypto/sha/asm \
-	${LCRYPTO_SRC}/crypto/whrlpool/asm
+	${LCRYPTO_SRC}/crypto/whrlpool/asm \
+	${LCRYPTO_SRC}/engines/asm
 
-PERLPATH=	-I${LCRYPTO_SRC}/crypto/des/asm -I${LCRYPTO_SRC}/crypto/perlasm
+#PERLPATH=	-I${LCRYPTO_SRC}/crypto/des/asm -I${LCRYPTO_SRC}/crypto/perlasm
+
+# cpuid
+SRCS=	x86cpuid.pl
 
 # aes
-SRCS=	aes-586.pl aesni-x86.pl vpaes-x86.pl
+SRCS+=	aes-586.pl aesni-x86.pl vpaes-x86.pl
 
 # blowfish
-SRCS+=	bf-586.pl bf-686.pl
+SRCS+=	bf-586.pl
 
 # bn
 SRCS+=	bn-586.pl co-586.pl x86-gf2m.pl x86-mont.pl
@@ -198,15 +206,27 @@ SRCS+=	bn-586.pl co-586.pl x86-gf2m.pl x86-mont.pl
 # camellia
 SRCS+=	cmll-x86.pl
 
+# cast
+SRCS+=	cast-586.pl
+
+# chacha
+SRCS+=	chacha-x86.pl
+
 # des
 SRCS+=	crypt586.pl des-586.pl
 
+# ec
+SRCS+=	ecp_nistz256-x86.pl
+
 # md5
 SRCS+=	md5-586.pl
 
 # modes
 SRCS+=	ghash-x86.pl
 
+# poly1305
+SRCS+=	poly1305-x86.pl
+
 # rc4
 SRCS+=	rc4-586.pl
 
@@ -222,25 +242,26 @@ SRCS+=	sha1-586.pl sha256-586.pl sha512-586.pl
 # whrlpool
 SRCS+=	wp-mmx.pl
 
-# cpuid
+# engines
-SRCS+=	x86cpuid.pl
+SRCS+=	e_padlock-x86.pl
 
 ASM=	${SRCS:R:S/$/.S/}
 
 all:	${ASM}
 
-CLEANFILES=	${ASM}
+CLEANFILES=	${ASM} ${SRCS:R:S/$/.s/}
 .SUFFIXES:	.pl
 
 .pl.S:
 	( echo '/* $$'FreeBSD'$$ */' ;\
 	echo '/* Do not modify. This file is auto-generated from ${.IMPSRC:T}. */' ;\
 	echo '#ifdef PIC' ;\
-	env CC=cc perl ${PERLPATH} ${.IMPSRC} elf ${CFLAGS} -fpic -DPIC ;\
+	env CC=cc perl ${PERLPATH} ${.IMPSRC} elf ${CFLAGS} -fpic -DPIC ${.IMPSRC:R:S/$/.s/} ;\
+	cat ${.IMPSRC:R:S/$/.s/} ;\
 	echo '#else' ;\
-	env CC=cc perl ${PERLPATH} ${.IMPSRC} elf ${CFLAGS} ;\
+	env CC=cc perl ${PERLPATH} ${.IMPSRC} elf ${CFLAGS} ${.IMPSRC:R:S/$/.s/} ;\
-	echo '#endif') |\
+	cat ${.IMPSRC:R:S/$/.s/} ;\
-	sed -E 's|(\.file[[:blank:]]+)".*"|\1"${.TARGET}"|' > ${.TARGET}
+	echo '#endif' ) > ${.TARGET}
 .endif
 
 .include <bsd.prog.mk>

secure/lib/libcrypto/i386/aes-586.S

@@ -1,7 +1,6 @@
 /* $FreeBSD$ */
 /* Do not modify. This file is auto-generated from aes-586.pl. */
 #ifdef PIC
-.file	"aes-586.S"
 .text
 .type	_x86_AES_encrypt_compact,@function
 .align	16
@@ -2999,19 +2998,19 @@ _x86_AES_set_encrypt_key:
 	popl	%ebp
 	ret
 .size	_x86_AES_set_encrypt_key,.-_x86_AES_set_encrypt_key
-.globl	private_AES_set_encrypt_key
+.globl	AES_set_encrypt_key
-.type	private_AES_set_encrypt_key,@function
+.type	AES_set_encrypt_key,@function
 .align	16
-private_AES_set_encrypt_key:
+AES_set_encrypt_key:
-.L_private_AES_set_encrypt_key_begin:
+.L_AES_set_encrypt_key_begin:
 	call	_x86_AES_set_encrypt_key
 	ret
-.size	private_AES_set_encrypt_key,.-.L_private_AES_set_encrypt_key_begin
+.size	AES_set_encrypt_key,.-.L_AES_set_encrypt_key_begin
-.globl	private_AES_set_decrypt_key
+.globl	AES_set_decrypt_key
-.type	private_AES_set_decrypt_key,@function
+.type	AES_set_decrypt_key,@function
 .align	16
-private_AES_set_decrypt_key:
+AES_set_decrypt_key:
-.L_private_AES_set_decrypt_key_begin:
+.L_AES_set_decrypt_key_begin:
 	call	_x86_AES_set_encrypt_key
 	cmpl	$0,%eax
 	je	.L054proceed
@@ -3240,13 +3239,12 @@ private_AES_set_decrypt_key:
 	popl	%ebx
 	popl	%ebp
 	ret
-.size	private_AES_set_decrypt_key,.-.L_private_AES_set_decrypt_key_begin
+.size	AES_set_decrypt_key,.-.L_AES_set_decrypt_key_begin
 .byte	65,69,83,32,102,111,114,32,120,56,54,44,32,67,82,89
 .byte	80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114
 .byte	111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
 .comm	OPENSSL_ia32cap_P,16,4
 #else
-.file	"aes-586.S"
 .text
 .type	_x86_AES_encrypt_compact,@function
 .align	16
@@ -6244,19 +6242,19 @@ _x86_AES_set_encrypt_key:
 	popl	%ebp
 	ret
 .size	_x86_AES_set_encrypt_key,.-_x86_AES_set_encrypt_key
-.globl	private_AES_set_encrypt_key
+.globl	AES_set_encrypt_key
-.type	private_AES_set_encrypt_key,@function
+.type	AES_set_encrypt_key,@function
 .align	16
-private_AES_set_encrypt_key:
+AES_set_encrypt_key:
-.L_private_AES_set_encrypt_key_begin:
+.L_AES_set_encrypt_key_begin:
 	call	_x86_AES_set_encrypt_key
 	ret
-.size	private_AES_set_encrypt_key,.-.L_private_AES_set_encrypt_key_begin
+.size	AES_set_encrypt_key,.-.L_AES_set_encrypt_key_begin
-.globl	private_AES_set_decrypt_key
+.globl	AES_set_decrypt_key
-.type	private_AES_set_decrypt_key,@function
+.type	AES_set_decrypt_key,@function
 .align	16
-private_AES_set_decrypt_key:
+AES_set_decrypt_key:
-.L_private_AES_set_decrypt_key_begin:
+.L_AES_set_decrypt_key_begin:
 	call	_x86_AES_set_encrypt_key
 	cmpl	$0,%eax
 	je	.L054proceed
@@ -6485,7 +6483,7 @@ private_AES_set_decrypt_key:
 	popl	%ebx
 	popl	%ebp
 	ret
-.size	private_AES_set_decrypt_key,.-.L_private_AES_set_decrypt_key_begin
+.size	AES_set_decrypt_key,.-.L_AES_set_decrypt_key_begin
 .byte	65,69,83,32,102,111,114,32,120,56,54,44,32,67,82,89
 .byte	80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114
 .byte	111,64,111,112,101,110,115,115,108,46,111,114,103,62,0

secure/lib/libcrypto/i386/bf-586.S

@@ -1,7 +1,6 @@
 /* $FreeBSD$ */
 /* Do not modify. This file is auto-generated from bf-586.pl. */
 #ifdef PIC
-.file	"bf-586.S"
 .text
 .globl	BF_encrypt
 .type	BF_encrypt,@function
@@ -898,7 +897,6 @@ BF_cbc_encrypt:
 .align	64
 .size	BF_cbc_encrypt,.-.L_BF_cbc_encrypt_begin
 #else
-.file	"bf-586.S"
 .text
 .globl	BF_encrypt
 .type	BF_encrypt,@function

secure/lib/libcrypto/i386/chacha-x86.S

@@ -0,0 +1,741 @@
+/* $FreeBSD$ */
+/* Do not modify. This file is auto-generated from chacha-x86.pl. */
+#ifdef PIC
+.text
+.globl	ChaCha20_ctr32
+.type	ChaCha20_ctr32,@function
+.align	16
+ChaCha20_ctr32:
+.L_ChaCha20_ctr32_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	xorl	%eax,%eax
+	cmpl	28(%esp),%eax
+	je	.L000no_data
+	movl	32(%esp),%esi
+	movl	36(%esp),%edi
+	subl	$132,%esp
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	%eax,80(%esp)
+	movl	%ebx,84(%esp)
+	movl	%ecx,88(%esp)
+	movl	%edx,92(%esp)
+	movl	16(%esi),%eax
+	movl	20(%esi),%ebx
+	movl	24(%esi),%ecx
+	movl	28(%esi),%edx
+	movl	%eax,96(%esp)
+	movl	%ebx,100(%esp)
+	movl	%ecx,104(%esp)
+	movl	%edx,108(%esp)
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	subl	$1,%eax
+	movl	%eax,112(%esp)
+	movl	%ebx,116(%esp)
+	movl	%ecx,120(%esp)
+	movl	%edx,124(%esp)
+	jmp	.L001entry
+.align	16
+.L002outer_loop:
+	movl	%ebx,156(%esp)
+	movl	%eax,152(%esp)
+	movl	%ecx,160(%esp)
+.L001entry:
+	movl	$1634760805,%eax
+	movl	$857760878,4(%esp)
+	movl	$2036477234,8(%esp)
+	movl	$1797285236,12(%esp)
+	movl	84(%esp),%ebx
+	movl	88(%esp),%ebp
+	movl	104(%esp),%ecx
+	movl	108(%esp),%esi
+	movl	116(%esp),%edx
+	movl	120(%esp),%edi
+	movl	%ebx,20(%esp)
+	movl	%ebp,24(%esp)
+	movl	%ecx,40(%esp)
+	movl	%esi,44(%esp)
+	movl	%edx,52(%esp)
+	movl	%edi,56(%esp)
+	movl	92(%esp),%ebx
+	movl	124(%esp),%edi
+	movl	112(%esp),%edx
+	movl	80(%esp),%ebp
+	movl	96(%esp),%ecx
+	movl	100(%esp),%esi
+	addl	$1,%edx
+	movl	%ebx,28(%esp)
+	movl	%edi,60(%esp)
+	movl	%edx,112(%esp)
+	movl	$10,%ebx
+	jmp	.L003loop
+.align	16
+.L003loop:
+	addl	%ebp,%eax
+	movl	%ebx,128(%esp)
+	movl	%ebp,%ebx
+	xorl	%eax,%edx
+	roll	$16,%edx
+	addl	%edx,%ecx
+	xorl	%ecx,%ebx
+	movl	52(%esp),%edi
+	roll	$12,%ebx
+	movl	20(%esp),%ebp
+	addl	%ebx,%eax
+	xorl	%eax,%edx
+	movl	%eax,(%esp)
+	roll	$8,%edx
+	movl	4(%esp),%eax
+	addl	%edx,%ecx
+	movl	%edx,48(%esp)
+	xorl	%ecx,%ebx
+	addl	%ebp,%eax
+	roll	$7,%ebx
+	xorl	%eax,%edi
+	movl	%ecx,32(%esp)
+	roll	$16,%edi
+	movl	%ebx,16(%esp)
+	addl	%edi,%esi
+	movl	40(%esp),%ecx
+	xorl	%esi,%ebp
+	movl	56(%esp),%edx
+	roll	$12,%ebp
+	movl	24(%esp),%ebx
+	addl	%ebp,%eax
+	xorl	%eax,%edi
+	movl	%eax,4(%esp)
+	roll	$8,%edi
+	movl	8(%esp),%eax
+	addl	%edi,%esi
+	movl	%edi,52(%esp)
+	xorl	%esi,%ebp
+	addl	%ebx,%eax
+	roll	$7,%ebp
+	xorl	%eax,%edx
+	movl	%esi,36(%esp)
+	roll	$16,%edx
+	movl	%ebp,20(%esp)
+	addl	%edx,%ecx
+	movl	44(%esp),%esi
+	xorl	%ecx,%ebx
+	movl	60(%esp),%edi
+	roll	$12,%ebx
+	movl	28(%esp),%ebp
+	addl	%ebx,%eax
+	xorl	%eax,%edx
+	movl	%eax,8(%esp)
+	roll	$8,%edx
+	movl	12(%esp),%eax
+	addl	%edx,%ecx
+	movl	%edx,56(%esp)
+	xorl	%ecx,%ebx
+	addl	%ebp,%eax
+	roll	$7,%ebx
+	xorl	%eax,%edi
+	roll	$16,%edi
+	movl	%ebx,24(%esp)
+	addl	%edi,%esi
+	xorl	%esi,%ebp
+	roll	$12,%ebp
+	movl	20(%esp),%ebx
+	addl	%ebp,%eax
+	xorl	%eax,%edi
+	movl	%eax,12(%esp)
+	roll	$8,%edi
+	movl	(%esp),%eax
+	addl	%edi,%esi
+	movl	%edi,%edx
+	xorl	%esi,%ebp
+	addl	%ebx,%eax
+	roll	$7,%ebp
+	xorl	%eax,%edx
+	roll	$16,%edx
+	movl	%ebp,28(%esp)
+	addl	%edx,%ecx
+	xorl	%ecx,%ebx
+	movl	48(%esp),%edi
+	roll	$12,%ebx
+	movl	24(%esp),%ebp
+	addl	%ebx,%eax
+	xorl	%eax,%edx
+	movl	%eax,(%esp)
+	roll	$8,%edx
+	movl	4(%esp),%eax
+	addl	%edx,%ecx
+	movl	%edx,60(%esp)
+	xorl	%ecx,%ebx
+	addl	%ebp,%eax
+	roll	$7,%ebx
+	xorl	%eax,%edi
+	movl	%ecx,40(%esp)
+	roll	$16,%edi
+	movl	%ebx,20(%esp)
+	addl	%edi,%esi
+	movl	32(%esp),%ecx
+	xorl	%esi,%ebp
+	movl	52(%esp),%edx
+	roll	$12,%ebp
+	movl	28(%esp),%ebx
+	addl	%ebp,%eax
+	xorl	%eax,%edi
+	movl	%eax,4(%esp)
+	roll	$8,%edi
+	movl	8(%esp),%eax
+	addl	%edi,%esi
+	movl	%edi,48(%esp)
+	xorl	%esi,%ebp
+	addl	%ebx,%eax
+	roll	$7,%ebp
+	xorl	%eax,%edx
+	movl	%esi,44(%esp)
+	roll	$16,%edx
+	movl	%ebp,24(%esp)
+	addl	%edx,%ecx
+	movl	36(%esp),%esi
+	xorl	%ecx,%ebx
+	movl	56(%esp),%edi
+	roll	$12,%ebx
+	movl	16(%esp),%ebp
+	addl	%ebx,%eax
+	xorl	%eax,%edx
+	movl	%eax,8(%esp)
+	roll	$8,%edx
+	movl	12(%esp),%eax
+	addl	%edx,%ecx
+	movl	%edx,52(%esp)
+	xorl	%ecx,%ebx
+	addl	%ebp,%eax
+	roll	$7,%ebx
+	xorl	%eax,%edi
+	roll	$16,%edi
+	movl	%ebx,28(%esp)
+	addl	%edi,%esi
+	xorl	%esi,%ebp
+	movl	48(%esp),%edx
+	roll	$12,%ebp
+	movl	128(%esp),%ebx
+	addl	%ebp,%eax
+	xorl	%eax,%edi
+	movl	%eax,12(%esp)
+	roll	$8,%edi
+	movl	(%esp),%eax
+	addl	%edi,%esi
+	movl	%edi,56(%esp)
+	xorl	%esi,%ebp
+	roll	$7,%ebp
+	decl	%ebx
+	jnz	.L003loop
+	movl	160(%esp),%ebx
+	addl	$1634760805,%eax
+	addl	80(%esp),%ebp
+	addl	96(%esp),%ecx
+	addl	100(%esp),%esi
+	cmpl	$64,%ebx
+	jb	.L004tail
+	movl	156(%esp),%ebx
+	addl	112(%esp),%edx
+	addl	120(%esp),%edi
+	xorl	(%ebx),%eax
+	xorl	16(%ebx),%ebp
+	movl	%eax,(%esp)
+	movl	152(%esp),%eax
+	xorl	32(%ebx),%ecx
+	xorl	36(%ebx),%esi
+	xorl	48(%ebx),%edx
+	xorl	56(%ebx),%edi
+	movl	%ebp,16(%eax)
+	movl	%ecx,32(%eax)
+	movl	%esi,36(%eax)
+	movl	%edx,48(%eax)
+	movl	%edi,56(%eax)
+	movl	4(%esp),%ebp
+	movl	8(%esp),%ecx
+	movl	12(%esp),%esi
+	movl	20(%esp),%edx
+	movl	24(%esp),%edi
+	addl	$857760878,%ebp
+	addl	$2036477234,%ecx
+	addl	$1797285236,%esi
+	addl	84(%esp),%edx
+	addl	88(%esp),%edi
+	xorl	4(%ebx),%ebp
+	xorl	8(%ebx),%ecx
+	xorl	12(%ebx),%esi
+	xorl	20(%ebx),%edx
+	xorl	24(%ebx),%edi
+	movl	%ebp,4(%eax)
+	movl	%ecx,8(%eax)
+	movl	%esi,12(%eax)
+	movl	%edx,20(%eax)
+	movl	%edi,24(%eax)
+	movl	28(%esp),%ebp
+	movl	40(%esp),%ecx
+	movl	44(%esp),%esi
+	movl	52(%esp),%edx
+	movl	60(%esp),%edi
+	addl	92(%esp),%ebp
+	addl	104(%esp),%ecx
+	addl	108(%esp),%esi
+	addl	116(%esp),%edx
+	addl	124(%esp),%edi
+	xorl	28(%ebx),%ebp
+	xorl	40(%ebx),%ecx
+	xorl	44(%ebx),%esi
+	xorl	52(%ebx),%edx
+	xorl	60(%ebx),%edi
+	leal	64(%ebx),%ebx
+	movl	%ebp,28(%eax)
+	movl	(%esp),%ebp
+	movl	%ecx,40(%eax)
+	movl	160(%esp),%ecx
+	movl	%esi,44(%eax)
+	movl	%edx,52(%eax)
+	movl	%edi,60(%eax)
+	movl	%ebp,(%eax)
+	leal	64(%eax),%eax
+	subl	$64,%ecx
+	jnz	.L002outer_loop
+	jmp	.L005done
+.L004tail:
+	addl	112(%esp),%edx
+	addl	120(%esp),%edi
+	movl	%eax,(%esp)
+	movl	%ebp,16(%esp)
+	movl	%ecx,32(%esp)
+	movl	%esi,36(%esp)
+	movl	%edx,48(%esp)
+	movl	%edi,56(%esp)
+	movl	4(%esp),%ebp
+	movl	8(%esp),%ecx
+	movl	12(%esp),%esi
+	movl	20(%esp),%edx
+	movl	24(%esp),%edi
+	addl	$857760878,%ebp
+	addl	$2036477234,%ecx
+	addl	$1797285236,%esi
+	addl	84(%esp),%edx
+	addl	88(%esp),%edi
+	movl	%ebp,4(%esp)
+	movl	%ecx,8(%esp)
+	movl	%esi,12(%esp)
+	movl	%edx,20(%esp)
+	movl	%edi,24(%esp)
+	movl	28(%esp),%ebp
+	movl	40(%esp),%ecx
+	movl	44(%esp),%esi
+	movl	52(%esp),%edx
+	movl	60(%esp),%edi
+	addl	92(%esp),%ebp
+	addl	104(%esp),%ecx
+	addl	108(%esp),%esi
+	addl	116(%esp),%edx
+	addl	124(%esp),%edi
+	movl	%ebp,28(%esp)
+	movl	156(%esp),%ebp
+	movl	%ecx,40(%esp)
+	movl	152(%esp),%ecx
+	movl	%esi,44(%esp)
+	xorl	%esi,%esi
+	movl	%edx,52(%esp)
+	movl	%edi,60(%esp)
+	xorl	%eax,%eax
+	xorl	%edx,%edx
+.L006tail_loop:
+	movb	(%esi,%ebp,1),%al
+	movb	(%esp,%esi,1),%dl
+	leal	1(%esi),%esi
+	xorb	%dl,%al
+	movb	%al,-1(%ecx,%esi,1)
+	decl	%ebx
+	jnz	.L006tail_loop
+.L005done:
+	addl	$132,%esp
+.L000no_data:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	ChaCha20_ctr32,.-.L_ChaCha20_ctr32_begin
+.byte	67,104,97,67,104,97,50,48,32,102,111,114,32,120,56,54
+.byte	44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32
+.byte	60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111
+.byte	114,103,62,0
+#else
+.text
+.globl	ChaCha20_ctr32
+.type	ChaCha20_ctr32,@function
+.align	16
+ChaCha20_ctr32:
+.L_ChaCha20_ctr32_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	xorl	%eax,%eax
+	cmpl	28(%esp),%eax
+	je	.L000no_data
+	movl	32(%esp),%esi
+	movl	36(%esp),%edi
+	subl	$132,%esp
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	%eax,80(%esp)
+	movl	%ebx,84(%esp)
+	movl	%ecx,88(%esp)
+	movl	%edx,92(%esp)
+	movl	16(%esi),%eax
+	movl	20(%esi),%ebx
+	movl	24(%esi),%ecx
+	movl	28(%esi),%edx
+	movl	%eax,96(%esp)
+	movl	%ebx,100(%esp)
+	movl	%ecx,104(%esp)
+	movl	%edx,108(%esp)
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	subl	$1,%eax
+	movl	%eax,112(%esp)
+	movl	%ebx,116(%esp)
+	movl	%ecx,120(%esp)
+	movl	%edx,124(%esp)
+	jmp	.L001entry
+.align	16
+.L002outer_loop:
+	movl	%ebx,156(%esp)
+	movl	%eax,152(%esp)
+	movl	%ecx,160(%esp)
+.L001entry:
+	movl	$1634760805,%eax
+	movl	$857760878,4(%esp)
+	movl	$2036477234,8(%esp)
+	movl	$1797285236,12(%esp)
+	movl	84(%esp),%ebx
+	movl	88(%esp),%ebp
+	movl	104(%esp),%ecx
+	movl	108(%esp),%esi
+	movl	116(%esp),%edx
+	movl	120(%esp),%edi
+	movl	%ebx,20(%esp)
+	movl	%ebp,24(%esp)
+	movl	%ecx,40(%esp)
+	movl	%esi,44(%esp)
+	movl	%edx,52(%esp)
+	movl	%edi,56(%esp)
+	movl	92(%esp),%ebx
+	movl	124(%esp),%edi
+	movl	112(%esp),%edx
+	movl	80(%esp),%ebp
+	movl	96(%esp),%ecx
+	movl	100(%esp),%esi
+	addl	$1,%edx
+	movl	%ebx,28(%esp)
+	movl	%edi,60(%esp)
+	movl	%edx,112(%esp)
+	movl	$10,%ebx
+	jmp	.L003loop
+.align	16
+.L003loop:
+	addl	%ebp,%eax
+	movl	%ebx,128(%esp)
+	movl	%ebp,%ebx
+	xorl	%eax,%edx
+	roll	$16,%edx
+	addl	%edx,%ecx
+	xorl	%ecx,%ebx
+	movl	52(%esp),%edi
+	roll	$12,%ebx
+	movl	20(%esp),%ebp
+	addl	%ebx,%eax
+	xorl	%eax,%edx
+	movl	%eax,(%esp)
+	roll	$8,%edx
+	movl	4(%esp),%eax
+	addl	%edx,%ecx
+	movl	%edx,48(%esp)
+	xorl	%ecx,%ebx
+	addl	%ebp,%eax
+	roll	$7,%ebx
+	xorl	%eax,%edi
+	movl	%ecx,32(%esp)
+	roll	$16,%edi
+	movl	%ebx,16(%esp)
+	addl	%edi,%esi
+	movl	40(%esp),%ecx
+	xorl	%esi,%ebp
+	movl	56(%esp),%edx
+	roll	$12,%ebp
+	movl	24(%esp),%ebx
+	addl	%ebp,%eax
+	xorl	%eax,%edi
+	movl	%eax,4(%esp)
+	roll	$8,%edi
+	movl	8(%esp),%eax
+	addl	%edi,%esi
+	movl	%edi,52(%esp)
+	xorl	%esi,%ebp
+	addl	%ebx,%eax
+	roll	$7,%ebp
+	xorl	%eax,%edx
+	movl	%esi,36(%esp)
+	roll	$16,%edx
+	movl	%ebp,20(%esp)
+	addl	%edx,%ecx
+	movl	44(%esp),%esi
+	xorl	%ecx,%ebx
+	movl	60(%esp),%edi
+	roll	$12,%ebx
+	movl	28(%esp),%ebp
+	addl	%ebx,%eax
+	xorl	%eax,%edx
+	movl	%eax,8(%esp)
+	roll	$8,%edx
+	movl	12(%esp),%eax
+	addl	%edx,%ecx
+	movl	%edx,56(%esp)
+	xorl	%ecx,%ebx
+	addl	%ebp,%eax
+	roll	$7,%ebx
+	xorl	%eax,%edi
+	roll	$16,%edi
+	movl	%ebx,24(%esp)
+	addl	%edi,%esi
+	xorl	%esi,%ebp
+	roll	$12,%ebp
+	movl	20(%esp),%ebx
+	addl	%ebp,%eax
+	xorl	%eax,%edi
+	movl	%eax,12(%esp)
+	roll	$8,%edi
+	movl	(%esp),%eax
+	addl	%edi,%esi
+	movl	%edi,%edx
+	xorl	%esi,%ebp
+	addl	%ebx,%eax
+	roll	$7,%ebp
+	xorl	%eax,%edx
+	roll	$16,%edx
+	movl	%ebp,28(%esp)
+	addl	%edx,%ecx
+	xorl	%ecx,%ebx
+	movl	48(%esp),%edi
+	roll	$12,%ebx
+	movl	24(%esp),%ebp
+	addl	%ebx,%eax
+	xorl	%eax,%edx
+	movl	%eax,(%esp)
+	roll	$8,%edx
+	movl	4(%esp),%eax
+	addl	%edx,%ecx
+	movl	%edx,60(%esp)
+	xorl	%ecx,%ebx
+	addl	%ebp,%eax
+	roll	$7,%ebx
+	xorl	%eax,%edi
+	movl	%ecx,40(%esp)
+	roll	$16,%edi
+	movl	%ebx,20(%esp)
+	addl	%edi,%esi
+	movl	32(%esp),%ecx
+	xorl	%esi,%ebp
+	movl	52(%esp),%edx
+	roll	$12,%ebp
+	movl	28(%esp),%ebx
+	addl	%ebp,%eax
+	xorl	%eax,%edi
+	movl	%eax,4(%esp)
+	roll	$8,%edi
+	movl	8(%esp),%eax
+	addl	%edi,%esi
+	movl	%edi,48(%esp)
+	xorl	%esi,%ebp
+	addl	%ebx,%eax
+	roll	$7,%ebp
+	xorl	%eax,%edx
+	movl	%esi,44(%esp)
+	roll	$16,%edx
+	movl	%ebp,24(%esp)
+	addl	%edx,%ecx
+	movl	36(%esp),%esi
+	xorl	%ecx,%ebx
+	movl	56(%esp),%edi
+	roll	$12,%ebx
+	movl	16(%esp),%ebp
+	addl	%ebx,%eax
+	xorl	%eax,%edx
+	movl	%eax,8(%esp)
+	roll	$8,%edx
+	movl	12(%esp),%eax
+	addl	%edx,%ecx
+	movl	%edx,52(%esp)
+	xorl	%ecx,%ebx
+	addl	%ebp,%eax
+	roll	$7,%ebx
+	xorl	%eax,%edi
+	roll	$16,%edi
+	movl	%ebx,28(%esp)
+	addl	%edi,%esi
+	xorl	%esi,%ebp
+	movl	48(%esp),%edx
+	roll	$12,%ebp
+	movl	128(%esp),%ebx
+	addl	%ebp,%eax
+	xorl	%eax,%edi
+	movl	%eax,12(%esp)
+	roll	$8,%edi
+	movl	(%esp),%eax
+	addl	%edi,%esi
+	movl	%edi,56(%esp)
+	xorl	%esi,%ebp
+	roll	$7,%ebp
+	decl	%ebx
+	jnz	.L003loop
+	movl	160(%esp),%ebx
+	addl	$1634760805,%eax
+	addl	80(%esp),%ebp
+	addl	96(%esp),%ecx
+	addl	100(%esp),%esi
+	cmpl	$64,%ebx
+	jb	.L004tail
+	movl	156(%esp),%ebx
+	addl	112(%esp),%edx
+	addl	120(%esp),%edi
+	xorl	(%ebx),%eax
+	xorl	16(%ebx),%ebp
+	movl	%eax,(%esp)
+	movl	152(%esp),%eax
+	xorl	32(%ebx),%ecx
+	xorl	36(%ebx),%esi
+	xorl	48(%ebx),%edx
+	xorl	56(%ebx),%edi
+	movl	%ebp,16(%eax)
+	movl	%ecx,32(%eax)
+	movl	%esi,36(%eax)
+	movl	%edx,48(%eax)
+	movl	%edi,56(%eax)
+	movl	4(%esp),%ebp
+	movl	8(%esp),%ecx
+	movl	12(%esp),%esi
+	movl	20(%esp),%edx
+	movl	24(%esp),%edi
+	addl	$857760878,%ebp
+	addl	$2036477234,%ecx
+	addl	$1797285236,%esi
+	addl	84(%esp),%edx
+	addl	88(%esp),%edi
+	xorl	4(%ebx),%ebp
+	xorl	8(%ebx),%ecx
+	xorl	12(%ebx),%esi
+	xorl	20(%ebx),%edx
+	xorl	24(%ebx),%edi
+	movl	%ebp,4(%eax)
+	movl	%ecx,8(%eax)
+	movl	%esi,12(%eax)
+	movl	%edx,20(%eax)
+	movl	%edi,24(%eax)
+	movl	28(%esp),%ebp
+	movl	40(%esp),%ecx
+	movl	44(%esp),%esi
+	movl	52(%esp),%edx
+	movl	60(%esp),%edi
+	addl	92(%esp),%ebp
+	addl	104(%esp),%ecx
+	addl	108(%esp),%esi
+	addl	116(%esp),%edx
+	addl	124(%esp),%edi
+	xorl	28(%ebx),%ebp
+	xorl	40(%ebx),%ecx
+	xorl	44(%ebx),%esi
+	xorl	52(%ebx),%edx
+	xorl	60(%ebx),%edi
+	leal	64(%ebx),%ebx
+	movl	%ebp,28(%eax)
+	movl	(%esp),%ebp
+	movl	%ecx,40(%eax)
+	movl	160(%esp),%ecx
+	movl	%esi,44(%eax)
+	movl	%edx,52(%eax)
+	movl	%edi,60(%eax)
+	movl	%ebp,(%eax)
+	leal	64(%eax),%eax
+	subl	$64,%ecx
+	jnz	.L002outer_loop
+	jmp	.L005done
+.L004tail:
+	addl	112(%esp),%edx
+	addl	120(%esp),%edi
+	movl	%eax,(%esp)
+	movl	%ebp,16(%esp)
+	movl	%ecx,32(%esp)
+	movl	%esi,36(%esp)
+	movl	%edx,48(%esp)
+	movl	%edi,56(%esp)
+	movl	4(%esp),%ebp
+	movl	8(%esp),%ecx
+	movl	12(%esp),%esi
+	movl	20(%esp),%edx
+	movl	24(%esp),%edi
+	addl	$857760878,%ebp
+	addl	$2036477234,%ecx
+	addl	$1797285236,%esi
+	addl	84(%esp),%edx
+	addl	88(%esp),%edi
+	movl	%ebp,4(%esp)
+	movl	%ecx,8(%esp)
+	movl	%esi,12(%esp)
+	movl	%edx,20(%esp)
+	movl	%edi,24(%esp)
+	movl	28(%esp),%ebp
+	movl	40(%esp),%ecx
+	movl	44(%esp),%esi
+	movl	52(%esp),%edx
+	movl	60(%esp),%edi
+	addl	92(%esp),%ebp
+	addl	104(%esp),%ecx
+	addl	108(%esp),%esi
+	addl	116(%esp),%edx
+	addl	124(%esp),%edi
+	movl	%ebp,28(%esp)
+	movl	156(%esp),%ebp
+	movl	%ecx,40(%esp)
+	movl	152(%esp),%ecx
+	movl	%esi,44(%esp)
+	xorl	%esi,%esi
+	movl	%edx,52(%esp)
+	movl	%edi,60(%esp)
+	xorl	%eax,%eax
+	xorl	%edx,%edx
+.L006tail_loop:
+	movb	(%esi,%ebp,1),%al
+	movb	(%esp,%esi,1),%dl
+	leal	1(%esi),%esi
+	xorb	%dl,%al
+	movb	%al,-1(%ecx,%esi,1)
+	decl	%ebx
+	jnz	.L006tail_loop
+.L005done:
+	addl	$132,%esp
+.L000no_data:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	ChaCha20_ctr32,.-.L_ChaCha20_ctr32_begin
+.byte	67,104,97,67,104,97,50,48,32,102,111,114,32,120,56,54
+.byte	44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32
+.byte	60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111
+.byte	114,103,62,0
+#endif

secure/lib/libcrypto/i386/cmll-x86.S

@@ -1,7 +1,6 @@
 /* $FreeBSD$ */
 /* Do not modify. This file is auto-generated from cmll-x86.pl. */
 #ifdef PIC
-.file	"cmll-x86.S"
 .text
 .globl	Camellia_EncryptBlock_Rounds
 .type	Camellia_EncryptBlock_Rounds,@function
@@ -1540,11 +1539,11 @@ Camellia_Ekeygen:
 	popl	%ebp
 	ret
 .size	Camellia_Ekeygen,.-.L_Camellia_Ekeygen_begin
-.globl	private_Camellia_set_key
+.globl	Camellia_set_key
-.type	private_Camellia_set_key,@function
+.type	Camellia_set_key,@function
 .align	16
-private_Camellia_set_key:
+Camellia_set_key:
-.L_private_Camellia_set_key_begin:
+.L_Camellia_set_key_begin:
 	pushl	%ebx
 	movl	8(%esp),%ecx
 	movl	12(%esp),%ebx
@@ -1574,7 +1573,7 @@ private_Camellia_set_key:
 .L014done:
 	popl	%ebx
 	ret
-.size	private_Camellia_set_key,.-.L_private_Camellia_set_key_begin
+.size	Camellia_set_key,.-.L_Camellia_set_key_begin
 .align	64
 .LCamellia_SIGMA:
 .long	2694735487,1003262091,3061508184,1286239154,3337565999,3914302142,1426019237,4057165596,283453434,3731369245,2958461122,3018244605,0,0,0,0
@@ -2377,7 +2376,6 @@ Camellia_cbc_encrypt:
 .byte	32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115
 .byte	115,108,46,111,114,103,62,0
 #else
-.file	"cmll-x86.S"
 .text
 .globl	Camellia_EncryptBlock_Rounds
 .type	Camellia_EncryptBlock_Rounds,@function
@@ -3916,11 +3914,11 @@ Camellia_Ekeygen:
 	popl	%ebp
 	ret
 .size	Camellia_Ekeygen,.-.L_Camellia_Ekeygen_begin
-.globl	private_Camellia_set_key
+.globl	Camellia_set_key
-.type	private_Camellia_set_key,@function
+.type	Camellia_set_key,@function
 .align	16
-private_Camellia_set_key:
+Camellia_set_key:
-.L_private_Camellia_set_key_begin:
+.L_Camellia_set_key_begin:
 	pushl	%ebx
 	movl	8(%esp),%ecx
 	movl	12(%esp),%ebx
@@ -3950,7 +3948,7 @@ private_Camellia_set_key:
 .L014done:
 	popl	%ebx
 	ret
-.size	private_Camellia_set_key,.-.L_private_Camellia_set_key_begin
+.size	Camellia_set_key,.-.L_Camellia_set_key_begin
 .align	64
 .LCamellia_SIGMA:
 .long	2694735487,1003262091,3061508184,1286239154,3337565999,3914302142,1426019237,4057165596,283453434,3731369245,2958461122,3018244605,0,0,0,0

secure/lib/libcrypto/i386/co-586.S

@@ -1,7 +1,6 @@
 /* $FreeBSD$ */
 /* Do not modify. This file is auto-generated from co-586.pl. */
 #ifdef PIC
-.file	"co-586.S"
 .text
 .globl	bn_mul_comba8
 .type	bn_mul_comba8,@function
@@ -1256,7 +1255,6 @@ bn_sqr_comba4:
 	ret
 .size	bn_sqr_comba4,.-.L_bn_sqr_comba4_begin
 #else
-.file	"co-586.S"
 .text
 .globl	bn_mul_comba8
 .type	bn_mul_comba8,@function

secure/lib/libcrypto/i386/crypt586.S

@@ -1,7 +1,6 @@
 /* $FreeBSD$ */
 /* Do not modify. This file is auto-generated from crypt586.pl. */
 #ifdef PIC
-.file	"crypt586.S"
 .text
 .globl	fcrypt_body
 .type	fcrypt_body,@function
@@ -881,7 +880,6 @@ fcrypt_body:
 	ret
 .size	fcrypt_body,.-.L_fcrypt_body_begin
 #else
-.file	"crypt586.S"
 .text
 .globl	fcrypt_body
 .type	fcrypt_body,@function

secure/lib/libcrypto/i386/des-586.S

@@ -1,7 +1,6 @@
 /* $FreeBSD$ */
 /* Do not modify. This file is auto-generated from des-586.pl. */
 #ifdef PIC
-.file	"des-586.S"
 .text
 .globl	DES_SPtrans
 .type	_x86_DES_encrypt,@function
@@ -1840,7 +1839,6 @@ DES_SPtrans:
 .long	128,545390592,8519808,0
 .long	536870912,545259648,131072,8519808
 #else
-.file	"des-586.S"
 .text
 .globl	DES_SPtrans
 .type	_x86_DES_encrypt,@function

secure/lib/libcrypto/i386/poly1305-x86.S

@@ -0,0 +1,557 @@
+/* $FreeBSD$ */
+/* Do not modify. This file is auto-generated from poly1305-x86.pl. */
+#ifdef PIC
+.text
+.align	64
+.globl	poly1305_init
+.type	poly1305_init,@function
+.align	16
+poly1305_init:
+.L_poly1305_init_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%edi
+	movl	24(%esp),%esi
+	movl	28(%esp),%ebp
+	xorl	%eax,%eax
+	movl	%eax,(%edi)
+	movl	%eax,4(%edi)
+	movl	%eax,8(%edi)
+	movl	%eax,12(%edi)
+	movl	%eax,16(%edi)
+	movl	%eax,20(%edi)
+	cmpl	$0,%esi
+	je	.L000nokey
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	andl	$268435455,%eax
+	andl	$268435452,%ebx
+	andl	$268435452,%ecx
+	andl	$268435452,%edx
+	movl	%eax,24(%edi)
+	movl	%ebx,28(%edi)
+	movl	%ecx,32(%edi)
+	movl	%edx,36(%edi)
+	movl	$0,%eax
+.L000nokey:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	poly1305_init,.-.L_poly1305_init_begin
+.globl	poly1305_blocks
+.type	poly1305_blocks,@function
+.align	16
+poly1305_blocks:
+.L_poly1305_blocks_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%edi
+	movl	24(%esp),%esi
+	movl	28(%esp),%ecx
+.L001enter_blocks:
+	andl	$-15,%ecx
+	jz	.L002nodata
+	subl	$64,%esp
+	movl	24(%edi),%eax
+	movl	28(%edi),%ebx
+	leal	(%esi,%ecx,1),%ebp
+	movl	32(%edi),%ecx
+	movl	36(%edi),%edx
+	movl	%ebp,92(%esp)
+	movl	%esi,%ebp
+	movl	%eax,36(%esp)
+	movl	%ebx,%eax
+	shrl	$2,%eax
+	movl	%ebx,40(%esp)
+	addl	%ebx,%eax
+	movl	%ecx,%ebx
+	shrl	$2,%ebx
+	movl	%ecx,44(%esp)
+	addl	%ecx,%ebx
+	movl	%edx,%ecx
+	shrl	$2,%ecx
+	movl	%edx,48(%esp)
+	addl	%edx,%ecx
+	movl	%eax,52(%esp)
+	movl	%ebx,56(%esp)
+	movl	%ecx,60(%esp)
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%esi
+	movl	16(%edi),%edi
+	jmp	.L003loop
+.align	32
+.L003loop:
+	addl	(%ebp),%eax
+	adcl	4(%ebp),%ebx
+	adcl	8(%ebp),%ecx
+	adcl	12(%ebp),%esi
+	leal	16(%ebp),%ebp
+	adcl	96(%esp),%edi
+	movl	%eax,(%esp)
+	movl	%esi,12(%esp)
+	mull	36(%esp)
+	movl	%edi,16(%esp)
+	movl	%eax,%edi
+	movl	%ebx,%eax
+	movl	%edx,%esi
+	mull	60(%esp)
+	addl	%eax,%edi
+	movl	%ecx,%eax
+	adcl	%edx,%esi
+	mull	56(%esp)
+	addl	%eax,%edi
+	movl	12(%esp),%eax
+	adcl	%edx,%esi
+	mull	52(%esp)
+	addl	%eax,%edi
+	movl	(%esp),%eax
+	adcl	%edx,%esi
+	mull	40(%esp)
+	movl	%edi,20(%esp)
+	xorl	%edi,%edi
+	addl	%eax,%esi
+	movl	%ebx,%eax
+	adcl	%edx,%edi
+	mull	36(%esp)
+	addl	%eax,%esi
+	movl	%ecx,%eax
+	adcl	%edx,%edi
+	mull	60(%esp)
+	addl	%eax,%esi
+	movl	12(%esp),%eax
+	adcl	%edx,%edi
+	mull	56(%esp)
+	addl	%eax,%esi
+	movl	16(%esp),%eax
+	adcl	%edx,%edi
+	imull	52(%esp),%eax
+	addl	%eax,%esi
+	movl	(%esp),%eax
+	adcl	$0,%edi
+	mull	44(%esp)
+	movl	%esi,24(%esp)
+	xorl	%esi,%esi
+	addl	%eax,%edi
+	movl	%ebx,%eax
+	adcl	%edx,%esi
+	mull	40(%esp)
+	addl	%eax,%edi
+	movl	%ecx,%eax
+	adcl	%edx,%esi
+	mull	36(%esp)
+	addl	%eax,%edi
+	movl	12(%esp),%eax
+	adcl	%edx,%esi
+	mull	60(%esp)
+	addl	%eax,%edi
+	movl	16(%esp),%eax
+	adcl	%edx,%esi
+	imull	56(%esp),%eax
+	addl	%eax,%edi
+	movl	(%esp),%eax
+	adcl	$0,%esi
+	mull	48(%esp)
+	movl	%edi,28(%esp)
+	xorl	%edi,%edi
+	addl	%eax,%esi
+	movl	%ebx,%eax
+	adcl	%edx,%edi
+	mull	44(%esp)
+	addl	%eax,%esi
+	movl	%ecx,%eax
+	adcl	%edx,%edi
+	mull	40(%esp)
+	addl	%eax,%esi
+	movl	12(%esp),%eax
+	adcl	%edx,%edi
+	mull	36(%esp)
+	addl	%eax,%esi
+	movl	16(%esp),%ecx
+	adcl	%edx,%edi
+	movl	%ecx,%edx
+	imull	60(%esp),%ecx
+	addl	%ecx,%esi
+	movl	20(%esp),%eax
+	adcl	$0,%edi
+	imull	36(%esp),%edx
+	addl	%edi,%edx
+	movl	24(%esp),%ebx
+	movl	28(%esp),%ecx
+	movl	%edx,%edi
+	shrl	$2,%edx
+	andl	$3,%edi
+	leal	(%edx,%edx,4),%edx
+	addl	%edx,%eax
+	adcl	$0,%ebx
+	adcl	$0,%ecx
+	adcl	$0,%esi
+	adcl	$0,%edi
+	cmpl	92(%esp),%ebp
+	jne	.L003loop
+	movl	84(%esp),%edx
+	addl	$64,%esp
+	movl	%eax,(%edx)
+	movl	%ebx,4(%edx)
+	movl	%ecx,8(%edx)
+	movl	%esi,12(%edx)
+	movl	%edi,16(%edx)
+.L002nodata:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	poly1305_blocks,.-.L_poly1305_blocks_begin
+.globl	poly1305_emit
+.type	poly1305_emit,@function
+.align	16
+poly1305_emit:
+.L_poly1305_emit_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%ebp
+.L004enter_emit:
+	movl	24(%esp),%edi
+	movl	(%ebp),%eax
+	movl	4(%ebp),%ebx
+	movl	8(%ebp),%ecx
+	movl	12(%ebp),%edx
+	movl	16(%ebp),%esi
+	addl	$5,%eax
+	adcl	$0,%ebx
+	adcl	$0,%ecx
+	adcl	$0,%edx
+	adcl	$0,%esi
+	shrl	$2,%esi
+	negl	%esi
+	andl	%esi,%eax
+	andl	%esi,%ebx
+	andl	%esi,%ecx
+	andl	%esi,%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	notl	%esi
+	movl	(%ebp),%eax
+	movl	4(%ebp),%ebx
+	movl	8(%ebp),%ecx
+	movl	12(%ebp),%edx
+	movl	28(%esp),%ebp
+	andl	%esi,%eax
+	andl	%esi,%ebx
+	andl	%esi,%ecx
+	andl	%esi,%edx
+	orl	(%edi),%eax
+	orl	4(%edi),%ebx
+	orl	8(%edi),%ecx
+	orl	12(%edi),%edx
+	addl	(%ebp),%eax
+	adcl	4(%ebp),%ebx
+	adcl	8(%ebp),%ecx
+	adcl	12(%ebp),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	poly1305_emit,.-.L_poly1305_emit_begin
+.byte	80,111,108,121,49,51,48,53,32,102,111,114,32,120,56,54
+.byte	44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32
+.byte	60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111
+.byte	114,103,62,0
+.align	4
+#else
+.text
+.align	64
+.globl	poly1305_init
+.type	poly1305_init,@function
+.align	16
+poly1305_init:
+.L_poly1305_init_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%edi
+	movl	24(%esp),%esi
+	movl	28(%esp),%ebp
+	xorl	%eax,%eax
+	movl	%eax,(%edi)
+	movl	%eax,4(%edi)
+	movl	%eax,8(%edi)
+	movl	%eax,12(%edi)
+	movl	%eax,16(%edi)
+	movl	%eax,20(%edi)
+	cmpl	$0,%esi
+	je	.L000nokey
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	andl	$268435455,%eax
+	andl	$268435452,%ebx
+	andl	$268435452,%ecx
+	andl	$268435452,%edx
+	movl	%eax,24(%edi)
+	movl	%ebx,28(%edi)
+	movl	%ecx,32(%edi)
+	movl	%edx,36(%edi)
+	movl	$0,%eax
+.L000nokey:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	poly1305_init,.-.L_poly1305_init_begin
+.globl	poly1305_blocks
+.type	poly1305_blocks,@function
+.align	16
+poly1305_blocks:
+.L_poly1305_blocks_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%edi
+	movl	24(%esp),%esi
+	movl	28(%esp),%ecx
+.L001enter_blocks:
+	andl	$-15,%ecx
+	jz	.L002nodata
+	subl	$64,%esp
+	movl	24(%edi),%eax
+	movl	28(%edi),%ebx
+	leal	(%esi,%ecx,1),%ebp
+	movl	32(%edi),%ecx
+	movl	36(%edi),%edx
+	movl	%ebp,92(%esp)
+	movl	%esi,%ebp
+	movl	%eax,36(%esp)
+	movl	%ebx,%eax
+	shrl	$2,%eax
+	movl	%ebx,40(%esp)
+	addl	%ebx,%eax
+	movl	%ecx,%ebx
+	shrl	$2,%ebx
+	movl	%ecx,44(%esp)
+	addl	%ecx,%ebx
+	movl	%edx,%ecx
+	shrl	$2,%ecx
+	movl	%edx,48(%esp)
+	addl	%edx,%ecx
+	movl	%eax,52(%esp)
+	movl	%ebx,56(%esp)
+	movl	%ecx,60(%esp)
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%esi
+	movl	16(%edi),%edi
+	jmp	.L003loop
+.align	32
+.L003loop:
+	addl	(%ebp),%eax
+	adcl	4(%ebp),%ebx
+	adcl	8(%ebp),%ecx
+	adcl	12(%ebp),%esi
+	leal	16(%ebp),%ebp
+	adcl	96(%esp),%edi
+	movl	%eax,(%esp)
+	movl	%esi,12(%esp)
+	mull	36(%esp)
+	movl	%edi,16(%esp)
+	movl	%eax,%edi
+	movl	%ebx,%eax
+	movl	%edx,%esi
+	mull	60(%esp)
+	addl	%eax,%edi
+	movl	%ecx,%eax
+	adcl	%edx,%esi
+	mull	56(%esp)
+	addl	%eax,%edi
+	movl	12(%esp),%eax
+	adcl	%edx,%esi
+	mull	52(%esp)
+	addl	%eax,%edi
+	movl	(%esp),%eax
+	adcl	%edx,%esi
+	mull	40(%esp)
+	movl	%edi,20(%esp)
+	xorl	%edi,%edi
+	addl	%eax,%esi
+	movl	%ebx,%eax
+	adcl	%edx,%edi
+	mull	36(%esp)
+	addl	%eax,%esi
+	movl	%ecx,%eax
+	adcl	%edx,%edi
+	mull	60(%esp)
+	addl	%eax,%esi
+	movl	12(%esp),%eax
+	adcl	%edx,%edi
+	mull	56(%esp)
+	addl	%eax,%esi
+	movl	16(%esp),%eax
+	adcl	%edx,%edi
+	imull	52(%esp),%eax
+	addl	%eax,%esi
+	movl	(%esp),%eax
+	adcl	$0,%edi
+	mull	44(%esp)
+	movl	%esi,24(%esp)
+	xorl	%esi,%esi
+	addl	%eax,%edi
+	movl	%ebx,%eax
+	adcl	%edx,%esi
+	mull	40(%esp)
+	addl	%eax,%edi
+	movl	%ecx,%eax
+	adcl	%edx,%esi
+	mull	36(%esp)
+	addl	%eax,%edi
+	movl	12(%esp),%eax
+	adcl	%edx,%esi
+	mull	60(%esp)
+	addl	%eax,%edi
+	movl	16(%esp),%eax
+	adcl	%edx,%esi
+	imull	56(%esp),%eax
+	addl	%eax,%edi
+	movl	(%esp),%eax
+	adcl	$0,%esi
+	mull	48(%esp)
+	movl	%edi,28(%esp)
+	xorl	%edi,%edi
+	addl	%eax,%esi
+	movl	%ebx,%eax
+	adcl	%edx,%edi
+	mull	44(%esp)
+	addl	%eax,%esi
+	movl	%ecx,%eax
+	adcl	%edx,%edi
+	mull	40(%esp)
+	addl	%eax,%esi
+	movl	12(%esp),%eax
+	adcl	%edx,%edi
+	mull	36(%esp)
+	addl	%eax,%esi
+	movl	16(%esp),%ecx
+	adcl	%edx,%edi
+	movl	%ecx,%edx
+	imull	60(%esp),%ecx
+	addl	%ecx,%esi
+	movl	20(%esp),%eax
+	adcl	$0,%edi
+	imull	36(%esp),%edx
+	addl	%edi,%edx
+	movl	24(%esp),%ebx
+	movl	28(%esp),%ecx
+	movl	%edx,%edi
+	shrl	$2,%edx
+	andl	$3,%edi
+	leal	(%edx,%edx,4),%edx
+	addl	%edx,%eax
+	adcl	$0,%ebx
+	adcl	$0,%ecx
+	adcl	$0,%esi
+	adcl	$0,%edi
+	cmpl	92(%esp),%ebp
+	jne	.L003loop
+	movl	84(%esp),%edx
+	addl	$64,%esp
+	movl	%eax,(%edx)
+	movl	%ebx,4(%edx)
+	movl	%ecx,8(%edx)
+	movl	%esi,12(%edx)
+	movl	%edi,16(%edx)
+.L002nodata:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	poly1305_blocks,.-.L_poly1305_blocks_begin
+.globl	poly1305_emit
+.type	poly1305_emit,@function
+.align	16
+poly1305_emit:
+.L_poly1305_emit_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%ebp
+.L004enter_emit:
+	movl	24(%esp),%edi
+	movl	(%ebp),%eax
+	movl	4(%ebp),%ebx
+	movl	8(%ebp),%ecx
+	movl	12(%ebp),%edx
+	movl	16(%ebp),%esi
+	addl	$5,%eax
+	adcl	$0,%ebx
+	adcl	$0,%ecx
+	adcl	$0,%edx
+	adcl	$0,%esi
+	shrl	$2,%esi
+	negl	%esi
+	andl	%esi,%eax
+	andl	%esi,%ebx
+	andl	%esi,%ecx
+	andl	%esi,%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	notl	%esi
+	movl	(%ebp),%eax
+	movl	4(%ebp),%ebx
+	movl	8(%ebp),%ecx
+	movl	12(%ebp),%edx
+	movl	28(%esp),%ebp
+	andl	%esi,%eax
+	andl	%esi,%ebx
+	andl	%esi,%ecx
+	andl	%esi,%edx
+	orl	(%edi),%eax
+	orl	4(%edi),%ebx
+	orl	8(%edi),%ecx
+	orl	12(%edi),%edx
+	addl	(%ebp),%eax
+	adcl	4(%ebp),%ebx
+	adcl	8(%ebp),%ecx
+	adcl	12(%ebp),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	poly1305_emit,.-.L_poly1305_emit_begin
+.byte	80,111,108,121,49,51,48,53,32,102,111,114,32,120,56,54
+.byte	44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32
+.byte	60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111
+.byte	114,103,62,0
+.align	4
+#endif

secure/lib/libcrypto/i386/rc4-586.S

@@ -1,7 +1,6 @@
 /* $FreeBSD$ */
 /* Do not modify. This file is auto-generated from rc4-586.pl. */
 #ifdef PIC
-.file	"rc4-586.S"
 .text
 .globl	RC4
 .type	RC4,@function
@@ -269,11 +268,11 @@ RC4:
 	popl	%ebp
 	ret
 .size	RC4,.-.L_RC4_begin
-.globl	private_RC4_set_key
+.globl	RC4_set_key
-.type	private_RC4_set_key,@function
+.type	RC4_set_key,@function
 .align	16
-private_RC4_set_key:
+RC4_set_key:
-.L_private_RC4_set_key_begin:
+.L_RC4_set_key_begin:
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -346,7 +345,7 @@ private_RC4_set_key:
 	popl	%ebx
 	popl	%ebp
 	ret
-.size	private_RC4_set_key,.-.L_private_RC4_set_key_begin
+.size	RC4_set_key,.-.L_RC4_set_key_begin
 .globl	RC4_options
 .type	RC4_options,@function
 .align	16
@@ -383,7 +382,6 @@ RC4_options:
 .size	RC4_options,.-.L_RC4_options_begin
 .comm	OPENSSL_ia32cap_P,16,4
 #else
-.file	"rc4-586.S"
 .text
 .globl	RC4
 .type	RC4,@function
@@ -648,11 +646,11 @@ RC4:
 	popl	%ebp
 	ret
 .size	RC4,.-.L_RC4_begin
-.globl	private_RC4_set_key
+.globl	RC4_set_key
-.type	private_RC4_set_key,@function
+.type	RC4_set_key,@function
 .align	16
-private_RC4_set_key:
+RC4_set_key:
-.L_private_RC4_set_key_begin:
+.L_RC4_set_key_begin:
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -722,7 +720,7 @@ private_RC4_set_key:
 	popl	%ebx
 	popl	%ebp
 	ret
-.size	private_RC4_set_key,.-.L_private_RC4_set_key_begin
+.size	RC4_set_key,.-.L_RC4_set_key_begin
 .globl	RC4_options
 .type	RC4_options,@function
 .align	16

secure/lib/libcrypto/i386/rc5-586.S

@@ -1,7 +1,6 @@
 /* $FreeBSD$ */
 /* Do not modify. This file is auto-generated from rc5-586.pl. */
 #ifdef PIC
-.file	"rc5-586.S"
 .text
 .globl	RC5_32_encrypt
 .type	RC5_32_encrypt,@function
@@ -566,7 +565,6 @@ RC5_32_cbc_encrypt:
 .align	64
 .size	RC5_32_cbc_encrypt,.-.L_RC5_32_cbc_encrypt_begin
 #else
-.file	"rc5-586.S"
 .text
 .globl	RC5_32_encrypt
 .type	RC5_32_encrypt,@function

secure/lib/libcrypto/i386/rmd-586.S

@@ -1,7 +1,6 @@
 /* $FreeBSD$ */
 /* Do not modify. This file is auto-generated from rmd-586.pl. */
 #ifdef PIC
-.file	"rmd-586.S"
 .text
 .globl	ripemd160_block_asm_data_order
 .type	ripemd160_block_asm_data_order,@function
@@ -1967,7 +1966,6 @@ ripemd160_block_asm_data_order:
 	ret
 .size	ripemd160_block_asm_data_order,.-.L_ripemd160_block_asm_data_order_begin
 #else
-.file	"rmd-586.S"
 .text
 .globl	ripemd160_block_asm_data_order
 .type	ripemd160_block_asm_data_order,@function

secure/lib/libcrypto/i386/vpaes-x86.S

@@ -1,7 +1,6 @@
 /* $FreeBSD$ */
 /* Do not modify. This file is auto-generated from vpaes-x86.pl. */
 #ifdef PIC
-.file	"vpaes-x86.S"
 .text
 .align	64
 .L_vpaes_consts:
@@ -663,7 +662,6 @@ vpaes_cbc_encrypt:
 	ret
 .size	vpaes_cbc_encrypt,.-.L_vpaes_cbc_encrypt_begin
 #else
-.file	"vpaes-x86.S"
 .text
 .align	64
 .L_vpaes_consts:

secure/lib/libcrypto/i386/wp-mmx.S

@@ -1,7 +1,6 @@
 /* $FreeBSD$ */
 /* Do not modify. This file is auto-generated from wp-mmx.pl. */
 #ifdef PIC
-.file	"wp-mmx.S"
 .text
 .globl	whirlpool_block_mmx
 .type	whirlpool_block_mmx,@function
@@ -1109,7 +1108,6 @@ whirlpool_block_mmx:
 .byte	202,45,191,7,173,90,131,51
 .size	whirlpool_block_mmx,.-.L_whirlpool_block_mmx_begin
 #else
-.file	"wp-mmx.S"
 .text
 .globl	whirlpool_block_mmx
 .type	whirlpool_block_mmx,@function

secure/lib/libcrypto/i386/x86-gf2m.S

@@ -1,7 +1,6 @@
 /* $FreeBSD$ */
 /* Do not modify. This file is auto-generated from x86-gf2m.pl. */
 #ifdef PIC
-.file	"x86-gf2m.S"
 .text
 .type	_mul_1x1_mmx,@function
 .align	16
@@ -251,18 +250,6 @@ bn_GF2m_mul_2x2:
 	movl	4(%edx),%edx
 	testl	$8388608,%eax
 	jz	.L001ialu
-	testl	$16777216,%eax
-	jz	.L002mmx
-	testl	$2,%edx
-	jz	.L002mmx
-	movups	8(%esp),%xmm0
-	shufps	$177,%xmm0,%xmm0
-.byte	102,15,58,68,192,1
-	movl	4(%esp),%eax
-	movups	%xmm0,(%eax)
-	ret
-.align	16
-.L002mmx:
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -348,7 +335,6 @@ bn_GF2m_mul_2x2:
 .byte	62,0
 .comm	OPENSSL_ia32cap_P,16,4
 #else
-.file	"x86-gf2m.S"
 .text
 .type	_mul_1x1_mmx,@function
 .align	16
@@ -595,18 +581,6 @@ bn_GF2m_mul_2x2:
 	movl	4(%edx),%edx
 	testl	$8388608,%eax
 	jz	.L000ialu
-	testl	$16777216,%eax
-	jz	.L001mmx
-	testl	$2,%edx
-	jz	.L001mmx
-	movups	8(%esp),%xmm0
-	shufps	$177,%xmm0,%xmm0
-.byte	102,15,58,68,192,1
-	movl	4(%esp),%eax
-	movups	%xmm0,(%eax)
-	ret
-.align	16
-.L001mmx:
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi

secure/lib/libcrypto/i386/x86cpuid.S

@@ -1,7 +1,6 @@
 /* $FreeBSD$ */
 /* Do not modify. This file is auto-generated from x86cpuid.pl. */
 #ifdef PIC
-.file	"x86cpuid.S"
 .text
 .globl	OPENSSL_ia32_cpuid
 .type	OPENSSL_ia32_cpuid,@function
@@ -237,18 +236,6 @@ OPENSSL_wipe_cpu:
 	movl	(%ecx),%ecx
 	btl	$1,(%ecx)
 	jnc	.L016no_x87
-	andl	$83886080,%ecx
-	cmpl	$83886080,%ecx
-	jne	.L017no_sse2
-	pxor	%xmm0,%xmm0
-	pxor	%xmm1,%xmm1
-	pxor	%xmm2,%xmm2
-	pxor	%xmm3,%xmm3
-	pxor	%xmm4,%xmm4
-	pxor	%xmm5,%xmm5
-	pxor	%xmm6,%xmm6
-	pxor	%xmm7,%xmm7
-.L017no_sse2:
 .long	4007259865,4007259865,4007259865,4007259865,2430851995
 .L016no_x87:
 	leal	4(%esp),%eax
@@ -264,42 +251,15 @@ OPENSSL_atomic_add:
 	pushl	%ebx
 	nop
 	movl	(%edx),%eax
-.L018spin:
+.L017spin:
 	leal	(%eax,%ecx,1),%ebx
 	nop
 .long	447811568
-	jne	.L018spin
+	jne	.L017spin
 	movl	%ebx,%eax
 	popl	%ebx
 	ret
 .size	OPENSSL_atomic_add,.-.L_OPENSSL_atomic_add_begin
-.globl	OPENSSL_indirect_call
-.type	OPENSSL_indirect_call,@function
-.align	16
-OPENSSL_indirect_call:
-.L_OPENSSL_indirect_call_begin:
-	pushl	%ebp
-	movl	%esp,%ebp
-	subl	$28,%esp
-	movl	12(%ebp),%ecx
-	movl	%ecx,(%esp)
-	movl	16(%ebp),%edx
-	movl	%edx,4(%esp)
-	movl	20(%ebp),%eax
-	movl	%eax,8(%esp)
-	movl	24(%ebp),%eax
-	movl	%eax,12(%esp)
-	movl	28(%ebp),%eax
-	movl	%eax,16(%esp)
-	movl	32(%ebp),%eax
-	movl	%eax,20(%esp)
-	movl	36(%ebp),%eax
-	movl	%eax,24(%esp)
-	call	*8(%ebp)
-	movl	%ebp,%esp
-	popl	%ebp
-	ret
-.size	OPENSSL_indirect_call,.-.L_OPENSSL_indirect_call_begin
 .globl	OPENSSL_cleanse
 .type	OPENSSL_cleanse,@function
 .align	16
@@ -309,71 +269,187 @@ OPENSSL_cleanse:
 	movl	8(%esp),%ecx
 	xorl	%eax,%eax
 	cmpl	$7,%ecx
-	jae	.L019lot
+	jae	.L018lot
 	cmpl	$0,%ecx
-	je	.L020ret
+	je	.L019ret
-.L021little:
+.L020little:
 	movb	%al,(%edx)
 	subl	$1,%ecx
 	leal	1(%edx),%edx
-	jnz	.L021little
+	jnz	.L020little
-.L020ret:
+.L019ret:
 	ret
 .align	16
-.L019lot:
+.L018lot:
 	testl	$3,%edx
-	jz	.L022aligned
+	jz	.L021aligned
 	movb	%al,(%edx)
 	leal	-1(%ecx),%ecx
 	leal	1(%edx),%edx
-	jmp	.L019lot
+	jmp	.L018lot
-.L022aligned:
+.L021aligned:
 	movl	%eax,(%edx)
 	leal	-4(%ecx),%ecx
 	testl	$-4,%ecx
 	leal	4(%edx),%edx
-	jnz	.L022aligned
+	jnz	.L021aligned
 	cmpl	$0,%ecx
-	jne	.L021little
+	jne	.L020little
 	ret
 .size	OPENSSL_cleanse,.-.L_OPENSSL_cleanse_begin
-.globl	OPENSSL_ia32_rdrand
+.globl	CRYPTO_memcmp
-.type	OPENSSL_ia32_rdrand,@function
+.type	CRYPTO_memcmp,@function
 .align	16
-OPENSSL_ia32_rdrand:
+CRYPTO_memcmp:
-.L_OPENSSL_ia32_rdrand_begin:
+.L_CRYPTO_memcmp_begin:
-	movl	$8,%ecx
+	pushl	%esi
+	pushl	%edi
+	movl	12(%esp),%esi
+	movl	16(%esp),%edi
+	movl	20(%esp),%ecx
+	xorl	%eax,%eax
+	xorl	%edx,%edx
+	cmpl	$0,%ecx
+	je	.L022no_data
 .L023loop:
-.byte	15,199,240
+	movb	(%esi),%dl
-	jc	.L024break
+	leal	1(%esi),%esi
-	loop	.L023loop
+	xorb	(%edi),%dl
-.L024break:
+	leal	1(%edi),%edi
-	cmpl	$0,%eax
+	orb	%dl,%al
-	cmovel	%ecx,%eax
+	decl	%ecx
+	jnz	.L023loop
+	negl	%eax
+	shrl	$31,%eax
+.L022no_data:
+	popl	%edi
+	popl	%esi
 	ret
-.size	OPENSSL_ia32_rdrand,.-.L_OPENSSL_ia32_rdrand_begin
+.size	CRYPTO_memcmp,.-.L_CRYPTO_memcmp_begin
-.globl	OPENSSL_ia32_rdseed
+.globl	OPENSSL_instrument_bus
-.type	OPENSSL_ia32_rdseed,@function
+.type	OPENSSL_instrument_bus,@function
 .align	16
-OPENSSL_ia32_rdseed:
+OPENSSL_instrument_bus:
-.L_OPENSSL_ia32_rdseed_begin:
+.L_OPENSSL_instrument_bus_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	$0,%eax
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	OPENSSL_instrument_bus,.-.L_OPENSSL_instrument_bus_begin
+.globl	OPENSSL_instrument_bus2
+.type	OPENSSL_instrument_bus2,@function
+.align	16
+OPENSSL_instrument_bus2:
+.L_OPENSSL_instrument_bus2_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	$0,%eax
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	OPENSSL_instrument_bus2,.-.L_OPENSSL_instrument_bus2_begin
+.globl	OPENSSL_ia32_rdrand_bytes
+.type	OPENSSL_ia32_rdrand_bytes,@function
+.align	16
+OPENSSL_ia32_rdrand_bytes:
+.L_OPENSSL_ia32_rdrand_bytes_begin:
+	pushl	%edi
+	pushl	%ebx
+	xorl	%eax,%eax
+	movl	12(%esp),%edi
+	movl	16(%esp),%ebx
+	cmpl	$0,%ebx
+	je	.L024done
 	movl	$8,%ecx
 .L025loop:
-.byte	15,199,248
+.byte	15,199,242
 	jc	.L026break
 	loop	.L025loop
+	jmp	.L024done
+.align	16
 .L026break:
-	cmpl	$0,%eax
+	cmpl	$4,%ebx
-	cmovel	%ecx,%eax
+	jb	.L027tail
+	movl	%edx,(%edi)
+	leal	4(%edi),%edi
+	addl	$4,%eax
+	subl	$4,%ebx
+	jz	.L024done
+	movl	$8,%ecx
+	jmp	.L025loop
+.align	16
+.L027tail:
+	movb	%dl,(%edi)
+	leal	1(%edi),%edi
+	incl	%eax
+	shrl	$8,%edx
+	decl	%ebx
+	jnz	.L027tail
+.L024done:
+	xorl	%edx,%edx
+	popl	%ebx
+	popl	%edi
 	ret
-.size	OPENSSL_ia32_rdseed,.-.L_OPENSSL_ia32_rdseed_begin
+.size	OPENSSL_ia32_rdrand_bytes,.-.L_OPENSSL_ia32_rdrand_bytes_begin
+.globl	OPENSSL_ia32_rdseed_bytes
+.type	OPENSSL_ia32_rdseed_bytes,@function
+.align	16
+OPENSSL_ia32_rdseed_bytes:
+.L_OPENSSL_ia32_rdseed_bytes_begin:
+	pushl	%edi
+	pushl	%ebx
+	xorl	%eax,%eax
+	movl	12(%esp),%edi
+	movl	16(%esp),%ebx
+	cmpl	$0,%ebx
+	je	.L028done
+	movl	$8,%ecx
+.L029loop:
+.byte	15,199,250
+	jc	.L030break
+	loop	.L029loop
+	jmp	.L028done
+.align	16
+.L030break:
+	cmpl	$4,%ebx
+	jb	.L031tail
+	movl	%edx,(%edi)
+	leal	4(%edi),%edi
+	addl	$4,%eax
+	subl	$4,%ebx
+	jz	.L028done
+	movl	$8,%ecx
+	jmp	.L029loop
+.align	16
+.L031tail:
+	movb	%dl,(%edi)
+	leal	1(%edi),%edi
+	incl	%eax
+	shrl	$8,%edx
+	decl	%ebx
+	jnz	.L031tail
+.L028done:
+	xorl	%edx,%edx
+	popl	%ebx
+	popl	%edi
+	ret
+.size	OPENSSL_ia32_rdseed_bytes,.-.L_OPENSSL_ia32_rdseed_bytes_begin
 .hidden	OPENSSL_cpuid_setup
 .hidden	OPENSSL_ia32cap_P
 .comm	OPENSSL_ia32cap_P,16,4
 .section	.init
 	call	OPENSSL_cpuid_setup
 #else
-.file	"x86cpuid.S"
 .text
 .globl	OPENSSL_ia32_cpuid
 .type	OPENSSL_ia32_cpuid,@function
@@ -600,18 +676,6 @@ OPENSSL_wipe_cpu:
 	movl	(%ecx),%ecx
 	btl	$1,(%ecx)
 	jnc	.L013no_x87
-	andl	$83886080,%ecx
-	cmpl	$83886080,%ecx
-	jne	.L014no_sse2
-	pxor	%xmm0,%xmm0
-	pxor	%xmm1,%xmm1
-	pxor	%xmm2,%xmm2
-	pxor	%xmm3,%xmm3
-	pxor	%xmm4,%xmm4
-	pxor	%xmm5,%xmm5
-	pxor	%xmm6,%xmm6
-	pxor	%xmm7,%xmm7
-.L014no_sse2:
 .long	4007259865,4007259865,4007259865,4007259865,2430851995
 .L013no_x87:
 	leal	4(%esp),%eax
@@ -627,42 +691,15 @@ OPENSSL_atomic_add:
 	pushl	%ebx
 	nop
 	movl	(%edx),%eax
-.L015spin:
+.L014spin:
 	leal	(%eax,%ecx,1),%ebx
 	nop
 .long	447811568
-	jne	.L015spin
+	jne	.L014spin
 	movl	%ebx,%eax
 	popl	%ebx
 	ret
 .size	OPENSSL_atomic_add,.-.L_OPENSSL_atomic_add_begin
-.globl	OPENSSL_indirect_call
-.type	OPENSSL_indirect_call,@function
-.align	16
-OPENSSL_indirect_call:
-.L_OPENSSL_indirect_call_begin:
-	pushl	%ebp
-	movl	%esp,%ebp
-	subl	$28,%esp
-	movl	12(%ebp),%ecx
-	movl	%ecx,(%esp)
-	movl	16(%ebp),%edx
-	movl	%edx,4(%esp)
-	movl	20(%ebp),%eax
-	movl	%eax,8(%esp)
-	movl	24(%ebp),%eax
-	movl	%eax,12(%esp)
-	movl	28(%ebp),%eax
-	movl	%eax,16(%esp)
-	movl	32(%ebp),%eax
-	movl	%eax,20(%esp)
-	movl	36(%ebp),%eax
-	movl	%eax,24(%esp)
-	call	*8(%ebp)
-	movl	%ebp,%esp
-	popl	%ebp
-	ret
-.size	OPENSSL_indirect_call,.-.L_OPENSSL_indirect_call_begin
 .globl	OPENSSL_cleanse
 .type	OPENSSL_cleanse,@function
 .align	16
@@ -672,64 +709,181 @@ OPENSSL_cleanse:
 	movl	8(%esp),%ecx
 	xorl	%eax,%eax
 	cmpl	$7,%ecx
-	jae	.L016lot
+	jae	.L015lot
 	cmpl	$0,%ecx
-	je	.L017ret
+	je	.L016ret
-.L018little:
+.L017little:
 	movb	%al,(%edx)
 	subl	$1,%ecx
 	leal	1(%edx),%edx
-	jnz	.L018little
+	jnz	.L017little
-.L017ret:
+.L016ret:
 	ret
 .align	16
-.L016lot:
+.L015lot:
 	testl	$3,%edx
-	jz	.L019aligned
+	jz	.L018aligned
 	movb	%al,(%edx)
 	leal	-1(%ecx),%ecx
 	leal	1(%edx),%edx
-	jmp	.L016lot
+	jmp	.L015lot
-.L019aligned:
+.L018aligned:
 	movl	%eax,(%edx)
 	leal	-4(%ecx),%ecx
 	testl	$-4,%ecx
 	leal	4(%edx),%edx
-	jnz	.L019aligned
+	jnz	.L018aligned
 	cmpl	$0,%ecx
-	jne	.L018little
+	jne	.L017little
 	ret
 .size	OPENSSL_cleanse,.-.L_OPENSSL_cleanse_begin
-.globl	OPENSSL_ia32_rdrand
+.globl	CRYPTO_memcmp
-.type	OPENSSL_ia32_rdrand,@function
+.type	CRYPTO_memcmp,@function
 .align	16
-OPENSSL_ia32_rdrand:
+CRYPTO_memcmp:
-.L_OPENSSL_ia32_rdrand_begin:
+.L_CRYPTO_memcmp_begin:
-	movl	$8,%ecx
+	pushl	%esi
+	pushl	%edi
+	movl	12(%esp),%esi
+	movl	16(%esp),%edi
+	movl	20(%esp),%ecx
+	xorl	%eax,%eax
+	xorl	%edx,%edx
+	cmpl	$0,%ecx
+	je	.L019no_data
 .L020loop:
-.byte	15,199,240
+	movb	(%esi),%dl
-	jc	.L021break
+	leal	1(%esi),%esi
-	loop	.L020loop
+	xorb	(%edi),%dl
-.L021break:
+	leal	1(%edi),%edi
-	cmpl	$0,%eax
+	orb	%dl,%al
-	cmovel	%ecx,%eax
+	decl	%ecx
+	jnz	.L020loop
+	negl	%eax
+	shrl	$31,%eax
+.L019no_data:
+	popl	%edi
+	popl	%esi
 	ret
-.size	OPENSSL_ia32_rdrand,.-.L_OPENSSL_ia32_rdrand_begin
+.size	CRYPTO_memcmp,.-.L_CRYPTO_memcmp_begin
-.globl	OPENSSL_ia32_rdseed
+.globl	OPENSSL_instrument_bus
-.type	OPENSSL_ia32_rdseed,@function
+.type	OPENSSL_instrument_bus,@function
 .align	16
-OPENSSL_ia32_rdseed:
+OPENSSL_instrument_bus:
-.L_OPENSSL_ia32_rdseed_begin:
+.L_OPENSSL_instrument_bus_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	$0,%eax
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	OPENSSL_instrument_bus,.-.L_OPENSSL_instrument_bus_begin
+.globl	OPENSSL_instrument_bus2
+.type	OPENSSL_instrument_bus2,@function
+.align	16
+OPENSSL_instrument_bus2:
+.L_OPENSSL_instrument_bus2_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	$0,%eax
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	OPENSSL_instrument_bus2,.-.L_OPENSSL_instrument_bus2_begin
+.globl	OPENSSL_ia32_rdrand_bytes
+.type	OPENSSL_ia32_rdrand_bytes,@function
+.align	16
+OPENSSL_ia32_rdrand_bytes:
+.L_OPENSSL_ia32_rdrand_bytes_begin:
+	pushl	%edi
+	pushl	%ebx
+	xorl	%eax,%eax
+	movl	12(%esp),%edi
+	movl	16(%esp),%ebx
+	cmpl	$0,%ebx
+	je	.L021done
 	movl	$8,%ecx
 .L022loop:
-.byte	15,199,248
+.byte	15,199,242
 	jc	.L023break
 	loop	.L022loop
+	jmp	.L021done
+.align	16
 .L023break:
-	cmpl	$0,%eax
+	cmpl	$4,%ebx
-	cmovel	%ecx,%eax
+	jb	.L024tail
+	movl	%edx,(%edi)
+	leal	4(%edi),%edi
+	addl	$4,%eax
+	subl	$4,%ebx
+	jz	.L021done
+	movl	$8,%ecx
+	jmp	.L022loop
+.align	16
+.L024tail:
+	movb	%dl,(%edi)
+	leal	1(%edi),%edi
+	incl	%eax
+	shrl	$8,%edx
+	decl	%ebx
+	jnz	.L024tail
+.L021done:
+	xorl	%edx,%edx
+	popl	%ebx
+	popl	%edi
 	ret
-.size	OPENSSL_ia32_rdseed,.-.L_OPENSSL_ia32_rdseed_begin
+.size	OPENSSL_ia32_rdrand_bytes,.-.L_OPENSSL_ia32_rdrand_bytes_begin
+.globl	OPENSSL_ia32_rdseed_bytes
+.type	OPENSSL_ia32_rdseed_bytes,@function
+.align	16
+OPENSSL_ia32_rdseed_bytes:
+.L_OPENSSL_ia32_rdseed_bytes_begin:
+	pushl	%edi
+	pushl	%ebx
+	xorl	%eax,%eax
+	movl	12(%esp),%edi
+	movl	16(%esp),%ebx
+	cmpl	$0,%ebx
+	je	.L025done
+	movl	$8,%ecx
+.L026loop:
+.byte	15,199,250
+	jc	.L027break
+	loop	.L026loop
+	jmp	.L025done
+.align	16
+.L027break:
+	cmpl	$4,%ebx
+	jb	.L028tail
+	movl	%edx,(%edi)
+	leal	4(%edi),%edi
+	addl	$4,%eax
+	subl	$4,%ebx
+	jz	.L025done
+	movl	$8,%ecx
+	jmp	.L026loop
+.align	16
+.L028tail:
+	movb	%dl,(%edi)
+	leal	1(%edi),%edi
+	incl	%eax
+	shrl	$8,%edx
+	decl	%ebx
+	jnz	.L028tail
+.L025done:
+	xorl	%edx,%edx
+	popl	%ebx
+	popl	%edi
+	ret
+.size	OPENSSL_ia32_rdseed_bytes,.-.L_OPENSSL_ia32_rdseed_bytes_begin
 .hidden	OPENSSL_cpuid_setup
 .hidden	OPENSSL_ia32cap_P
 .comm	OPENSSL_ia32cap_P,16,4