Add optimized salsa20 amd64 assembly implementation

.gitignore

@@ -5,7 +5,6 @@
 *.log
 *.o
 *.plist
-*.s
 *.scan
 *.status
 *.tar.*

configure.ac

@@ -34,6 +34,7 @@ AC_SUBST(DLL_VERSION)
 
 LX_CFLAGS=${CFLAGS-NONE}
 AC_PROG_CC_C99
+AM_PROG_AS
 AC_USE_SYSTEM_EXTENSIONS
 CPPFLAGS="$CPPFLAGS -D_FORTIFY_SOURCE=2"
 
@@ -184,6 +185,24 @@ AC_C_BIGENDIAN(
   AC_MSG_WARN([universal endianess])
 )
 
+AC_MSG_CHECKING(if we can assemble basic amd64 code)
+HAVE_AMD64_ASM_V=0
+AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[
+]], [[
+#if defined(__amd64) || defined(__amd64__) || defined(__x86_64__)
+/* neat */
+#else
+# error !amd64
+#endif
+__asm__("pxor %xmm12,%xmm6");
+]])],
+[AC_MSG_RESULT(yes)
+ AC_DEFINE([HAVE_AMD64_ASM], [1], [basic amd64 code can be assembled])
+ HAVE_AMD64_ASM_V=1],
+[AC_MSG_RESULT(no)])
+AM_CONDITIONAL([HAVE_AMD64_ASM], [test $HAVE_AMD64_ASM_V = 1])
+AC_SUBST(HAVE_AMD64_ASM_V)
+
 AC_MSG_CHECKING(for 128-bit arithmetic)
 HAVE_TI_MODE_V=0
 AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[

src/libsodium/Makefile.am

@@ -137,9 +137,6 @@ libsodium_la_SOURCES = \
 	crypto_stream/aes256estream/hongjun/aes256.h \
 	crypto_stream/aes256estream/hongjun/api.h \
 	crypto_stream/aes256estream/hongjun/ecrypt-sync.h \
-	crypto_stream/salsa20/ref/api.h \
-	crypto_stream/salsa20/ref/stream_salsa20.c \
-	crypto_stream/salsa20/ref/xor_salsa20.c \
 	crypto_stream/salsa2012/ref/api.h \
 	crypto_stream/salsa2012/ref/stream_salsa2012.c \
 	crypto_stream/salsa2012/ref/xor_salsa2012.c \
@@ -173,6 +170,17 @@ libsodium_la_SOURCES += \
 	crypto_scalarmult/curve25519/ref/smult_curve25519_ref.c
 endif
 
+if HAVE_AMD64_ASM
+libsodium_la_SOURCES += \
+	crypto_stream/salsa20/amd64_xmm6/api.h \
+	crypto_stream/salsa20/amd64_xmm6/stream_salsa20_amd64_xmm6.s
+else
+libsodium_la_SOURCES += \
+	crypto_stream/salsa20/ref/api.h \
+	crypto_stream/salsa20/ref/stream_salsa20_ref.c \
+	crypto_stream/salsa20/ref/xor_salsa20_ref.c
+endif
+
 libsodium_la_LDFLAGS = \
 	$(AM_LDFLAGS) \
 	-export-dynamic \

src/libsodium/crypto_stream/salsa20/amd64_xmm6/api.h

@@ -0,0 +1 @@
+#include "crypto_stream_salsa20.h"

src/libsodium/crypto_stream/salsa20/amd64_xmm6/stream_salsa20_amd64_xmm6.s

@@ -0,0 +1,947 @@
+
+#if defined(__amd64) || defined(__amd64__) || defined(__x86_64__)
+
+.text
+.p2align 5
+
+.globl _crypto_stream_salsa20
+.globl crypto_stream_salsa20
+_crypto_stream_salsa20:
+crypto_stream_salsa20:
+mov %rsp,%r11
+and $31,%r11
+add $480,%r11
+sub %r11,%rsp
+movq %r11,352(%rsp)
+movq %r12,360(%rsp)
+movq %r13,368(%rsp)
+movq %r14,376(%rsp)
+movq %r15,384(%rsp)
+movq %rbx,392(%rsp)
+movq %rbp,400(%rsp)
+mov  %rsi,%r9
+mov  %rdi,%rdi
+mov  %rdi,%rsi
+mov  %rdx,%rdx
+mov  %rcx,%r10
+cmp  $0,%r9
+jbe ._done
+
+mov  $0,%rax
+mov  %r9,%rcx
+rep stosb
+sub  %r9,%rdi
+jmp ._start
+
+.text
+.p2align 5
+
+.globl _crypto_stream_salsa20_xor
+.globl crypto_stream_salsa20_xor
+_crypto_stream_salsa20_xor:
+crypto_stream_salsa20_xor:
+mov %rsp,%r11
+and $31,%r11
+add $480,%r11
+sub %r11,%rsp
+movq %r11,352(%rsp)
+movq %r12,360(%rsp)
+movq %r13,368(%rsp)
+movq %r14,376(%rsp)
+movq %r15,384(%rsp)
+movq %rbx,392(%rsp)
+movq %rbp,400(%rsp)
+mov  %rdi,%rdi
+mov  %rsi,%rsi
+mov  %rdx,%r9
+mov  %rcx,%rdx
+mov  %r8,%r10
+cmp  $0,%r9
+jbe ._done
+
+._start:
+movl   20(%r10),%ecx
+movl   0(%r10),%r8d
+movl   0(%rdx),%eax
+movl   16(%r10),%r11d
+movl %ecx,0(%rsp)
+movl %r8d,4+0(%rsp)
+movl %eax,8+0(%rsp)
+movl %r11d,12+0(%rsp)
+mov  $0,%rcx
+movl   24(%r10),%r8d
+movl   4(%r10),%eax
+movl   4(%rdx),%edx
+movl %ecx,16(%rsp)
+movl %r8d,4+16(%rsp)
+movl %eax,8+16(%rsp)
+movl %edx,12+16(%rsp)
+movl   12(%r10),%edx
+mov  $0,%rcx
+movl   28(%r10),%r8d
+movl   8(%r10),%eax
+movl %edx,32(%rsp)
+movl %ecx,4+32(%rsp)
+movl %r8d,8+32(%rsp)
+movl %eax,12+32(%rsp)
+mov  $1634760805,%rdx
+mov  $857760878,%rcx
+mov  $2036477234,%r8
+mov  $1797285236,%rax
+movl %edx,48(%rsp)
+movl %ecx,4+48(%rsp)
+movl %r8d,8+48(%rsp)
+movl %eax,12+48(%rsp)
+cmp  $256,%r9
+jb ._bytesbetween1and255
+
+movdqa 48(%rsp),%xmm0
+pshufd $0x55,%xmm0,%xmm1
+pshufd $0xaa,%xmm0,%xmm2
+pshufd $0xff,%xmm0,%xmm3
+pshufd $0x00,%xmm0,%xmm0
+movdqa %xmm1,64(%rsp)
+movdqa %xmm2,80(%rsp)
+movdqa %xmm3,96(%rsp)
+movdqa %xmm0,112(%rsp)
+movdqa 0(%rsp),%xmm0
+pshufd $0xaa,%xmm0,%xmm1
+pshufd $0xff,%xmm0,%xmm2
+pshufd $0x00,%xmm0,%xmm3
+pshufd $0x55,%xmm0,%xmm0
+movdqa %xmm1,128(%rsp)
+movdqa %xmm2,144(%rsp)
+movdqa %xmm3,160(%rsp)
+movdqa %xmm0,176(%rsp)
+movdqa 16(%rsp),%xmm0
+pshufd $0xff,%xmm0,%xmm1
+pshufd $0x55,%xmm0,%xmm2
+pshufd $0xaa,%xmm0,%xmm0
+movdqa %xmm1,192(%rsp)
+movdqa %xmm2,208(%rsp)
+movdqa %xmm0,224(%rsp)
+movdqa 32(%rsp),%xmm0
+pshufd $0x00,%xmm0,%xmm1
+pshufd $0xaa,%xmm0,%xmm2
+pshufd $0xff,%xmm0,%xmm0
+movdqa %xmm1,240(%rsp)
+movdqa %xmm2,256(%rsp)
+movdqa %xmm0,272(%rsp)
+
+._bytesatleast256:
+movl 16(%rsp),%edx
+movl 4+32(%rsp),%ecx
+movl %edx,288(%rsp)
+movl %ecx,304(%rsp)
+add  $1,%rdx
+shl  $32,%rcx
+add  %rcx,%rdx
+mov  %rdx,%rcx
+shr  $32,%rcx
+movl %edx,4+288(%rsp)
+movl %ecx,4+304(%rsp)
+add  $1,%rdx
+shl  $32,%rcx
+add  %rcx,%rdx
+mov  %rdx,%rcx
+shr  $32,%rcx
+movl %edx,8+288(%rsp)
+movl %ecx,8+304(%rsp)
+add  $1,%rdx
+shl  $32,%rcx
+add  %rcx,%rdx
+mov  %rdx,%rcx
+shr  $32,%rcx
+movl %edx,12+288(%rsp)
+movl %ecx,12+304(%rsp)
+add  $1,%rdx
+shl  $32,%rcx
+add  %rcx,%rdx
+mov  %rdx,%rcx
+shr  $32,%rcx
+movl %edx,16(%rsp)
+movl %ecx,4+32(%rsp)
+movq %r9,408(%rsp)
+mov  $20,%rdx
+movdqa 64(%rsp),%xmm0
+movdqa 80(%rsp),%xmm1
+movdqa 96(%rsp),%xmm2
+movdqa 256(%rsp),%xmm3
+movdqa 272(%rsp),%xmm4
+movdqa 128(%rsp),%xmm5
+movdqa 144(%rsp),%xmm6
+movdqa 176(%rsp),%xmm7
+movdqa 192(%rsp),%xmm8
+movdqa 208(%rsp),%xmm9
+movdqa 224(%rsp),%xmm10
+movdqa 304(%rsp),%xmm11
+movdqa 112(%rsp),%xmm12
+movdqa 160(%rsp),%xmm13
+movdqa 240(%rsp),%xmm14
+movdqa 288(%rsp),%xmm15
+
+._mainloop1:
+movdqa %xmm1,320(%rsp)
+movdqa %xmm2,336(%rsp)
+movdqa %xmm13,%xmm1
+paddd %xmm12,%xmm1
+movdqa %xmm1,%xmm2
+pslld $7,%xmm1
+pxor  %xmm1,%xmm14
+psrld $25,%xmm2
+pxor  %xmm2,%xmm14
+movdqa %xmm7,%xmm1
+paddd %xmm0,%xmm1
+movdqa %xmm1,%xmm2
+pslld $7,%xmm1
+pxor  %xmm1,%xmm11
+psrld $25,%xmm2
+pxor  %xmm2,%xmm11
+movdqa %xmm12,%xmm1
+paddd %xmm14,%xmm1
+movdqa %xmm1,%xmm2
+pslld $9,%xmm1
+pxor  %xmm1,%xmm15
+psrld $23,%xmm2
+pxor  %xmm2,%xmm15
+movdqa %xmm0,%xmm1
+paddd %xmm11,%xmm1
+movdqa %xmm1,%xmm2
+pslld $9,%xmm1
+pxor  %xmm1,%xmm9
+psrld $23,%xmm2
+pxor  %xmm2,%xmm9
+movdqa %xmm14,%xmm1
+paddd %xmm15,%xmm1
+movdqa %xmm1,%xmm2
+pslld $13,%xmm1
+pxor  %xmm1,%xmm13
+psrld $19,%xmm2
+pxor  %xmm2,%xmm13
+movdqa %xmm11,%xmm1
+paddd %xmm9,%xmm1
+movdqa %xmm1,%xmm2
+pslld $13,%xmm1
+pxor  %xmm1,%xmm7
+psrld $19,%xmm2
+pxor  %xmm2,%xmm7
+movdqa %xmm15,%xmm1
+paddd %xmm13,%xmm1
+movdqa %xmm1,%xmm2
+pslld $18,%xmm1
+pxor  %xmm1,%xmm12
+psrld $14,%xmm2
+pxor  %xmm2,%xmm12
+movdqa 320(%rsp),%xmm1
+movdqa %xmm12,320(%rsp)
+movdqa %xmm9,%xmm2
+paddd %xmm7,%xmm2
+movdqa %xmm2,%xmm12
+pslld $18,%xmm2
+pxor  %xmm2,%xmm0
+psrld $14,%xmm12
+pxor  %xmm12,%xmm0
+movdqa %xmm5,%xmm2
+paddd %xmm1,%xmm2
+movdqa %xmm2,%xmm12
+pslld $7,%xmm2
+pxor  %xmm2,%xmm3
+psrld $25,%xmm12
+pxor  %xmm12,%xmm3
+movdqa 336(%rsp),%xmm2
+movdqa %xmm0,336(%rsp)
+movdqa %xmm6,%xmm0
+paddd %xmm2,%xmm0
+movdqa %xmm0,%xmm12
+pslld $7,%xmm0
+pxor  %xmm0,%xmm4
+psrld $25,%xmm12
+pxor  %xmm12,%xmm4
+movdqa %xmm1,%xmm0
+paddd %xmm3,%xmm0
+movdqa %xmm0,%xmm12
+pslld $9,%xmm0
+pxor  %xmm0,%xmm10
+psrld $23,%xmm12
+pxor  %xmm12,%xmm10
+movdqa %xmm2,%xmm0
+paddd %xmm4,%xmm0
+movdqa %xmm0,%xmm12
+pslld $9,%xmm0
+pxor  %xmm0,%xmm8
+psrld $23,%xmm12
+pxor  %xmm12,%xmm8
+movdqa %xmm3,%xmm0
+paddd %xmm10,%xmm0
+movdqa %xmm0,%xmm12
+pslld $13,%xmm0
+pxor  %xmm0,%xmm5
+psrld $19,%xmm12
+pxor  %xmm12,%xmm5
+movdqa %xmm4,%xmm0
+paddd %xmm8,%xmm0
+movdqa %xmm0,%xmm12
+pslld $13,%xmm0
+pxor  %xmm0,%xmm6
+psrld $19,%xmm12
+pxor  %xmm12,%xmm6
+movdqa %xmm10,%xmm0
+paddd %xmm5,%xmm0
+movdqa %xmm0,%xmm12
+pslld $18,%xmm0
+pxor  %xmm0,%xmm1
+psrld $14,%xmm12
+pxor  %xmm12,%xmm1
+movdqa 320(%rsp),%xmm0
+movdqa %xmm1,320(%rsp)
+movdqa %xmm4,%xmm1
+paddd %xmm0,%xmm1
+movdqa %xmm1,%xmm12
+pslld $7,%xmm1
+pxor  %xmm1,%xmm7
+psrld $25,%xmm12
+pxor  %xmm12,%xmm7
+movdqa %xmm8,%xmm1
+paddd %xmm6,%xmm1
+movdqa %xmm1,%xmm12
+pslld $18,%xmm1
+pxor  %xmm1,%xmm2
+psrld $14,%xmm12
+pxor  %xmm12,%xmm2
+movdqa 336(%rsp),%xmm12
+movdqa %xmm2,336(%rsp)
+movdqa %xmm14,%xmm1
+paddd %xmm12,%xmm1
+movdqa %xmm1,%xmm2
+pslld $7,%xmm1
+pxor  %xmm1,%xmm5
+psrld $25,%xmm2
+pxor  %xmm2,%xmm5
+movdqa %xmm0,%xmm1
+paddd %xmm7,%xmm1
+movdqa %xmm1,%xmm2
+pslld $9,%xmm1
+pxor  %xmm1,%xmm10
+psrld $23,%xmm2
+pxor  %xmm2,%xmm10
+movdqa %xmm12,%xmm1
+paddd %xmm5,%xmm1
+movdqa %xmm1,%xmm2
+pslld $9,%xmm1
+pxor  %xmm1,%xmm8
+psrld $23,%xmm2
+pxor  %xmm2,%xmm8
+movdqa %xmm7,%xmm1
+paddd %xmm10,%xmm1
+movdqa %xmm1,%xmm2
+pslld $13,%xmm1
+pxor  %xmm1,%xmm4
+psrld $19,%xmm2
+pxor  %xmm2,%xmm4
+movdqa %xmm5,%xmm1
+paddd %xmm8,%xmm1
+movdqa %xmm1,%xmm2
+pslld $13,%xmm1
+pxor  %xmm1,%xmm14
+psrld $19,%xmm2
+pxor  %xmm2,%xmm14
+movdqa %xmm10,%xmm1
+paddd %xmm4,%xmm1
+movdqa %xmm1,%xmm2
+pslld $18,%xmm1
+pxor  %xmm1,%xmm0
+psrld $14,%xmm2
+pxor  %xmm2,%xmm0
+movdqa 320(%rsp),%xmm1
+movdqa %xmm0,320(%rsp)
+movdqa %xmm8,%xmm0
+paddd %xmm14,%xmm0
+movdqa %xmm0,%xmm2
+pslld $18,%xmm0
+pxor  %xmm0,%xmm12
+psrld $14,%xmm2
+pxor  %xmm2,%xmm12
+movdqa %xmm11,%xmm0
+paddd %xmm1,%xmm0
+movdqa %xmm0,%xmm2
+pslld $7,%xmm0
+pxor  %xmm0,%xmm6
+psrld $25,%xmm2
+pxor  %xmm2,%xmm6
+movdqa 336(%rsp),%xmm2
+movdqa %xmm12,336(%rsp)
+movdqa %xmm3,%xmm0
+paddd %xmm2,%xmm0
+movdqa %xmm0,%xmm12
+pslld $7,%xmm0
+pxor  %xmm0,%xmm13
+psrld $25,%xmm12
+pxor  %xmm12,%xmm13
+movdqa %xmm1,%xmm0
+paddd %xmm6,%xmm0
+movdqa %xmm0,%xmm12
+pslld $9,%xmm0
+pxor  %xmm0,%xmm15
+psrld $23,%xmm12
+pxor  %xmm12,%xmm15
+movdqa %xmm2,%xmm0
+paddd %xmm13,%xmm0
+movdqa %xmm0,%xmm12
+pslld $9,%xmm0
+pxor  %xmm0,%xmm9
+psrld $23,%xmm12
+pxor  %xmm12,%xmm9
+movdqa %xmm6,%xmm0
+paddd %xmm15,%xmm0
+movdqa %xmm0,%xmm12
+pslld $13,%xmm0
+pxor  %xmm0,%xmm11
+psrld $19,%xmm12
+pxor  %xmm12,%xmm11
+movdqa %xmm13,%xmm0
+paddd %xmm9,%xmm0
+movdqa %xmm0,%xmm12
+pslld $13,%xmm0
+pxor  %xmm0,%xmm3
+psrld $19,%xmm12
+pxor  %xmm12,%xmm3
+movdqa %xmm15,%xmm0
+paddd %xmm11,%xmm0
+movdqa %xmm0,%xmm12
+pslld $18,%xmm0
+pxor  %xmm0,%xmm1
+psrld $14,%xmm12
+pxor  %xmm12,%xmm1
+movdqa %xmm9,%xmm0
+paddd %xmm3,%xmm0
+movdqa %xmm0,%xmm12
+pslld $18,%xmm0
+pxor  %xmm0,%xmm2
+psrld $14,%xmm12
+pxor  %xmm12,%xmm2
+movdqa 320(%rsp),%xmm12
+movdqa 336(%rsp),%xmm0
+sub  $2,%rdx
+ja ._mainloop1
+
+paddd 112(%rsp),%xmm12
+paddd 176(%rsp),%xmm7
+paddd 224(%rsp),%xmm10
+paddd 272(%rsp),%xmm4
+movd   %xmm12,%rdx
+movd   %xmm7,%rcx
+movd   %xmm10,%r8
+movd   %xmm4,%r9
+pshufd $0x39,%xmm12,%xmm12
+pshufd $0x39,%xmm7,%xmm7
+pshufd $0x39,%xmm10,%xmm10
+pshufd $0x39,%xmm4,%xmm4
+xorl 0(%rsi),%edx
+xorl 4(%rsi),%ecx
+xorl 8(%rsi),%r8d
+xorl 12(%rsi),%r9d
+movl   %edx,0(%rdi)
+movl   %ecx,4(%rdi)
+movl   %r8d,8(%rdi)
+movl   %r9d,12(%rdi)
+movd   %xmm12,%rdx
+movd   %xmm7,%rcx
+movd   %xmm10,%r8
+movd   %xmm4,%r9
+pshufd $0x39,%xmm12,%xmm12
+pshufd $0x39,%xmm7,%xmm7
+pshufd $0x39,%xmm10,%xmm10
+pshufd $0x39,%xmm4,%xmm4
+xorl 64(%rsi),%edx
+xorl 68(%rsi),%ecx
+xorl 72(%rsi),%r8d
+xorl 76(%rsi),%r9d
+movl   %edx,64(%rdi)
+movl   %ecx,68(%rdi)
+movl   %r8d,72(%rdi)
+movl   %r9d,76(%rdi)
+movd   %xmm12,%rdx
+movd   %xmm7,%rcx
+movd   %xmm10,%r8
+movd   %xmm4,%r9
+pshufd $0x39,%xmm12,%xmm12
+pshufd $0x39,%xmm7,%xmm7
+pshufd $0x39,%xmm10,%xmm10
+pshufd $0x39,%xmm4,%xmm4
+xorl 128(%rsi),%edx
+xorl 132(%rsi),%ecx
+xorl 136(%rsi),%r8d
+xorl 140(%rsi),%r9d
+movl   %edx,128(%rdi)
+movl   %ecx,132(%rdi)
+movl   %r8d,136(%rdi)
+movl   %r9d,140(%rdi)
+movd   %xmm12,%rdx
+movd   %xmm7,%rcx
+movd   %xmm10,%r8
+movd   %xmm4,%r9
+xorl 192(%rsi),%edx
+xorl 196(%rsi),%ecx
+xorl 200(%rsi),%r8d
+xorl 204(%rsi),%r9d
+movl   %edx,192(%rdi)
+movl   %ecx,196(%rdi)
+movl   %r8d,200(%rdi)
+movl   %r9d,204(%rdi)
+paddd 240(%rsp),%xmm14
+paddd 64(%rsp),%xmm0
+paddd 128(%rsp),%xmm5
+paddd 192(%rsp),%xmm8
+movd   %xmm14,%rdx
+movd   %xmm0,%rcx
+movd   %xmm5,%r8
+movd   %xmm8,%r9
+pshufd $0x39,%xmm14,%xmm14
+pshufd $0x39,%xmm0,%xmm0
+pshufd $0x39,%xmm5,%xmm5
+pshufd $0x39,%xmm8,%xmm8
+xorl 16(%rsi),%edx
+xorl 20(%rsi),%ecx
+xorl 24(%rsi),%r8d
+xorl 28(%rsi),%r9d
+movl   %edx,16(%rdi)
+movl   %ecx,20(%rdi)
+movl   %r8d,24(%rdi)
+movl   %r9d,28(%rdi)
+movd   %xmm14,%rdx
+movd   %xmm0,%rcx
+movd   %xmm5,%r8
+movd   %xmm8,%r9
+pshufd $0x39,%xmm14,%xmm14
+pshufd $0x39,%xmm0,%xmm0
+pshufd $0x39,%xmm5,%xmm5
+pshufd $0x39,%xmm8,%xmm8
+xorl 80(%rsi),%edx
+xorl 84(%rsi),%ecx
+xorl 88(%rsi),%r8d
+xorl 92(%rsi),%r9d
+movl   %edx,80(%rdi)
+movl   %ecx,84(%rdi)
+movl   %r8d,88(%rdi)
+movl   %r9d,92(%rdi)
+movd   %xmm14,%rdx
+movd   %xmm0,%rcx
+movd   %xmm5,%r8
+movd   %xmm8,%r9
+pshufd $0x39,%xmm14,%xmm14
+pshufd $0x39,%xmm0,%xmm0
+pshufd $0x39,%xmm5,%xmm5
+pshufd $0x39,%xmm8,%xmm8
+xorl 144(%rsi),%edx
+xorl 148(%rsi),%ecx
+xorl 152(%rsi),%r8d
+xorl 156(%rsi),%r9d
+movl   %edx,144(%rdi)
+movl   %ecx,148(%rdi)
+movl   %r8d,152(%rdi)
+movl   %r9d,156(%rdi)
+movd   %xmm14,%rdx
+movd   %xmm0,%rcx
+movd   %xmm5,%r8
+movd   %xmm8,%r9
+xorl 208(%rsi),%edx
+xorl 212(%rsi),%ecx
+xorl 216(%rsi),%r8d
+xorl 220(%rsi),%r9d
+movl   %edx,208(%rdi)
+movl   %ecx,212(%rdi)
+movl   %r8d,216(%rdi)
+movl   %r9d,220(%rdi)
+paddd 288(%rsp),%xmm15
+paddd 304(%rsp),%xmm11
+paddd 80(%rsp),%xmm1
+paddd 144(%rsp),%xmm6
+movd   %xmm15,%rdx
+movd   %xmm11,%rcx
+movd   %xmm1,%r8
+movd   %xmm6,%r9
+pshufd $0x39,%xmm15,%xmm15
+pshufd $0x39,%xmm11,%xmm11
+pshufd $0x39,%xmm1,%xmm1
+pshufd $0x39,%xmm6,%xmm6
+xorl 32(%rsi),%edx
+xorl 36(%rsi),%ecx
+xorl 40(%rsi),%r8d
+xorl 44(%rsi),%r9d
+movl   %edx,32(%rdi)
+movl   %ecx,36(%rdi)
+movl   %r8d,40(%rdi)
+movl   %r9d,44(%rdi)
+movd   %xmm15,%rdx
+movd   %xmm11,%rcx
+movd   %xmm1,%r8
+movd   %xmm6,%r9
+pshufd $0x39,%xmm15,%xmm15
+pshufd $0x39,%xmm11,%xmm11
+pshufd $0x39,%xmm1,%xmm1
+pshufd $0x39,%xmm6,%xmm6
+xorl 96(%rsi),%edx
+xorl 100(%rsi),%ecx
+xorl 104(%rsi),%r8d
+xorl 108(%rsi),%r9d
+movl   %edx,96(%rdi)
+movl   %ecx,100(%rdi)
+movl   %r8d,104(%rdi)
+movl   %r9d,108(%rdi)
+movd   %xmm15,%rdx
+movd   %xmm11,%rcx
+movd   %xmm1,%r8
+movd   %xmm6,%r9
+pshufd $0x39,%xmm15,%xmm15
+pshufd $0x39,%xmm11,%xmm11
+pshufd $0x39,%xmm1,%xmm1
+pshufd $0x39,%xmm6,%xmm6
+xorl 160(%rsi),%edx
+xorl 164(%rsi),%ecx
+xorl 168(%rsi),%r8d
+xorl 172(%rsi),%r9d
+movl   %edx,160(%rdi)
+movl   %ecx,164(%rdi)
+movl   %r8d,168(%rdi)
+movl   %r9d,172(%rdi)
+movd   %xmm15,%rdx
+movd   %xmm11,%rcx
+movd   %xmm1,%r8
+movd   %xmm6,%r9
+xorl 224(%rsi),%edx
+xorl 228(%rsi),%ecx
+xorl 232(%rsi),%r8d
+xorl 236(%rsi),%r9d
+movl   %edx,224(%rdi)
+movl   %ecx,228(%rdi)
+movl   %r8d,232(%rdi)
+movl   %r9d,236(%rdi)
+paddd 160(%rsp),%xmm13
+paddd 208(%rsp),%xmm9
+paddd 256(%rsp),%xmm3
+paddd 96(%rsp),%xmm2
+movd   %xmm13,%rdx
+movd   %xmm9,%rcx
+movd   %xmm3,%r8
+movd   %xmm2,%r9
+pshufd $0x39,%xmm13,%xmm13
+pshufd $0x39,%xmm9,%xmm9
+pshufd $0x39,%xmm3,%xmm3
+pshufd $0x39,%xmm2,%xmm2
+xorl 48(%rsi),%edx
+xorl 52(%rsi),%ecx
+xorl 56(%rsi),%r8d
+xorl 60(%rsi),%r9d
+movl   %edx,48(%rdi)
+movl   %ecx,52(%rdi)
+movl   %r8d,56(%rdi)
+movl   %r9d,60(%rdi)
+movd   %xmm13,%rdx
+movd   %xmm9,%rcx
+movd   %xmm3,%r8
+movd   %xmm2,%r9
+pshufd $0x39,%xmm13,%xmm13
+pshufd $0x39,%xmm9,%xmm9
+pshufd $0x39,%xmm3,%xmm3
+pshufd $0x39,%xmm2,%xmm2
+xorl 112(%rsi),%edx
+xorl 116(%rsi),%ecx
+xorl 120(%rsi),%r8d
+xorl 124(%rsi),%r9d
+movl   %edx,112(%rdi)
+movl   %ecx,116(%rdi)
+movl   %r8d,120(%rdi)
+movl   %r9d,124(%rdi)
+movd   %xmm13,%rdx
+movd   %xmm9,%rcx
+movd   %xmm3,%r8
+movd   %xmm2,%r9
+pshufd $0x39,%xmm13,%xmm13
+pshufd $0x39,%xmm9,%xmm9
+pshufd $0x39,%xmm3,%xmm3
+pshufd $0x39,%xmm2,%xmm2
+xorl 176(%rsi),%edx
+xorl 180(%rsi),%ecx
+xorl 184(%rsi),%r8d
+xorl 188(%rsi),%r9d
+movl   %edx,176(%rdi)
+movl   %ecx,180(%rdi)
+movl   %r8d,184(%rdi)
+movl   %r9d,188(%rdi)
+movd   %xmm13,%rdx
+movd   %xmm9,%rcx
+movd   %xmm3,%r8
+movd   %xmm2,%r9
+xorl 240(%rsi),%edx
+xorl 244(%rsi),%ecx
+xorl 248(%rsi),%r8d
+xorl 252(%rsi),%r9d
+movl   %edx,240(%rdi)
+movl   %ecx,244(%rdi)
+movl   %r8d,248(%rdi)
+movl   %r9d,252(%rdi)
+movq 408(%rsp),%r9
+sub  $256,%r9
+add  $256,%rsi
+add  $256,%rdi
+cmp  $256,%r9
+jae ._bytesatleast256
+
+cmp  $0,%r9
+jbe ._done
+
+._bytesbetween1and255:
+cmp  $64,%r9
+jae ._nocopy
+
+mov  %rdi,%rdx
+leaq 416(%rsp),%rdi
+mov  %r9,%rcx
+rep movsb
+leaq 416(%rsp),%rdi
+leaq 416(%rsp),%rsi
+
+._nocopy:
+movq %r9,408(%rsp)
+movdqa 48(%rsp),%xmm0
+movdqa 0(%rsp),%xmm1
+movdqa 16(%rsp),%xmm2
+movdqa 32(%rsp),%xmm3
+movdqa %xmm1,%xmm4
+mov  $20,%rcx
+
+._mainloop2:
+paddd %xmm0,%xmm4
+movdqa %xmm0,%xmm5
+movdqa %xmm4,%xmm6
+pslld $7,%xmm4
+psrld $25,%xmm6
+pxor  %xmm4,%xmm3
+pxor  %xmm6,%xmm3
+paddd %xmm3,%xmm5
+movdqa %xmm3,%xmm4
+movdqa %xmm5,%xmm6
+pslld $9,%xmm5
+psrld $23,%xmm6
+pxor  %xmm5,%xmm2
+pshufd $0x93,%xmm3,%xmm3
+pxor  %xmm6,%xmm2
+paddd %xmm2,%xmm4
+movdqa %xmm2,%xmm5
+movdqa %xmm4,%xmm6
+pslld $13,%xmm4
+psrld $19,%xmm6
+pxor  %xmm4,%xmm1
+pshufd $0x4e,%xmm2,%xmm2
+pxor  %xmm6,%xmm1
+paddd %xmm1,%xmm5
+movdqa %xmm3,%xmm4
+movdqa %xmm5,%xmm6
+pslld $18,%xmm5
+psrld $14,%xmm6
+pxor  %xmm5,%xmm0
+pshufd $0x39,%xmm1,%xmm1
+pxor  %xmm6,%xmm0
+paddd %xmm0,%xmm4
+movdqa %xmm0,%xmm5
+movdqa %xmm4,%xmm6
+pslld $7,%xmm4
+psrld $25,%xmm6
+pxor  %xmm4,%xmm1
+pxor  %xmm6,%xmm1
+paddd %xmm1,%xmm5
+movdqa %xmm1,%xmm4
+movdqa %xmm5,%xmm6
+pslld $9,%xmm5
+psrld $23,%xmm6
+pxor  %xmm5,%xmm2
+pshufd $0x93,%xmm1,%xmm1
+pxor  %xmm6,%xmm2
+paddd %xmm2,%xmm4
+movdqa %xmm2,%xmm5
+movdqa %xmm4,%xmm6
+pslld $13,%xmm4
+psrld $19,%xmm6
+pxor  %xmm4,%xmm3
+pshufd $0x4e,%xmm2,%xmm2
+pxor  %xmm6,%xmm3
+paddd %xmm3,%xmm5
+movdqa %xmm1,%xmm4
+movdqa %xmm5,%xmm6
+pslld $18,%xmm5
+psrld $14,%xmm6
+pxor  %xmm5,%xmm0
+pshufd $0x39,%xmm3,%xmm3
+pxor  %xmm6,%xmm0
+paddd %xmm0,%xmm4
+movdqa %xmm0,%xmm5
+movdqa %xmm4,%xmm6
+pslld $7,%xmm4
+psrld $25,%xmm6
+pxor  %xmm4,%xmm3
+pxor  %xmm6,%xmm3
+paddd %xmm3,%xmm5
+movdqa %xmm3,%xmm4
+movdqa %xmm5,%xmm6
+pslld $9,%xmm5
+psrld $23,%xmm6
+pxor  %xmm5,%xmm2
+pshufd $0x93,%xmm3,%xmm3
+pxor  %xmm6,%xmm2
+paddd %xmm2,%xmm4
+movdqa %xmm2,%xmm5
+movdqa %xmm4,%xmm6
+pslld $13,%xmm4
+psrld $19,%xmm6
+pxor  %xmm4,%xmm1
+pshufd $0x4e,%xmm2,%xmm2
+pxor  %xmm6,%xmm1
+paddd %xmm1,%xmm5
+movdqa %xmm3,%xmm4
+movdqa %xmm5,%xmm6
+pslld $18,%xmm5
+psrld $14,%xmm6
+pxor  %xmm5,%xmm0
+pshufd $0x39,%xmm1,%xmm1
+pxor  %xmm6,%xmm0
+paddd %xmm0,%xmm4
+movdqa %xmm0,%xmm5
+movdqa %xmm4,%xmm6
+pslld $7,%xmm4
+psrld $25,%xmm6
+pxor  %xmm4,%xmm1
+pxor  %xmm6,%xmm1
+paddd %xmm1,%xmm5
+movdqa %xmm1,%xmm4
+movdqa %xmm5,%xmm6
+pslld $9,%xmm5
+psrld $23,%xmm6
+pxor  %xmm5,%xmm2
+pshufd $0x93,%xmm1,%xmm1
+pxor  %xmm6,%xmm2
+paddd %xmm2,%xmm4
+movdqa %xmm2,%xmm5
+movdqa %xmm4,%xmm6
+pslld $13,%xmm4
+psrld $19,%xmm6
+pxor  %xmm4,%xmm3
+pshufd $0x4e,%xmm2,%xmm2
+pxor  %xmm6,%xmm3
+sub  $4,%rcx
+paddd %xmm3,%xmm5
+movdqa %xmm1,%xmm4
+movdqa %xmm5,%xmm6
+pslld $18,%xmm5
+pxor   %xmm7,%xmm7
+psrld $14,%xmm6
+pxor  %xmm5,%xmm0
+pshufd $0x39,%xmm3,%xmm3
+pxor  %xmm6,%xmm0
+ja ._mainloop2
+
+paddd 48(%rsp),%xmm0
+paddd 0(%rsp),%xmm1
+paddd 16(%rsp),%xmm2
+paddd 32(%rsp),%xmm3
+movd   %xmm0,%rcx
+movd   %xmm1,%r8
+movd   %xmm2,%r9
+movd   %xmm3,%rax
+pshufd $0x39,%xmm0,%xmm0
+pshufd $0x39,%xmm1,%xmm1
+pshufd $0x39,%xmm2,%xmm2
+pshufd $0x39,%xmm3,%xmm3
+xorl 0(%rsi),%ecx
+xorl 48(%rsi),%r8d
+xorl 32(%rsi),%r9d
+xorl 16(%rsi),%eax
+movl   %ecx,0(%rdi)
+movl   %r8d,48(%rdi)
+movl   %r9d,32(%rdi)
+movl   %eax,16(%rdi)
+movd   %xmm0,%rcx
+movd   %xmm1,%r8
+movd   %xmm2,%r9
+movd   %xmm3,%rax
+pshufd $0x39,%xmm0,%xmm0
+pshufd $0x39,%xmm1,%xmm1
+pshufd $0x39,%xmm2,%xmm2
+pshufd $0x39,%xmm3,%xmm3
+xorl 20(%rsi),%ecx
+xorl 4(%rsi),%r8d
+xorl 52(%rsi),%r9d
+xorl 36(%rsi),%eax
+movl   %ecx,20(%rdi)
+movl   %r8d,4(%rdi)
+movl   %r9d,52(%rdi)
+movl   %eax,36(%rdi)
+movd   %xmm0,%rcx
+movd   %xmm1,%r8
+movd   %xmm2,%r9
+movd   %xmm3,%rax
+pshufd $0x39,%xmm0,%xmm0
+pshufd $0x39,%xmm1,%xmm1
+pshufd $0x39,%xmm2,%xmm2
+pshufd $0x39,%xmm3,%xmm3
+xorl 40(%rsi),%ecx
+xorl 24(%rsi),%r8d
+xorl 8(%rsi),%r9d
+xorl 56(%rsi),%eax
+movl   %ecx,40(%rdi)
+movl   %r8d,24(%rdi)
+movl   %r9d,8(%rdi)
+movl   %eax,56(%rdi)
+movd   %xmm0,%rcx
+movd   %xmm1,%r8
+movd   %xmm2,%r9
+movd   %xmm3,%rax
+xorl 60(%rsi),%ecx
+xorl 44(%rsi),%r8d
+xorl 28(%rsi),%r9d
+xorl 12(%rsi),%eax
+movl   %ecx,60(%rdi)
+movl   %r8d,44(%rdi)
+movl   %r9d,28(%rdi)
+movl   %eax,12(%rdi)
+movq 408(%rsp),%r9
+movl 16(%rsp),%ecx
+movl 4+32(%rsp),%r8d
+add  $1,%rcx
+shl  $32,%r8
+add  %r8,%rcx
+mov  %rcx,%r8
+shr  $32,%r8
+movl %ecx,16(%rsp)
+movl %r8d,4+32(%rsp)
+cmp  $64,%r9
+
+ja ._bytesatleast65
+
+jae ._bytesatleast64
+
+mov  %rdi,%rsi
+mov  %rdx,%rdi
+mov  %r9,%rcx
+rep movsb
+
+._bytesatleast64:
+._done:
+movq 352(%rsp),%r11
+movq 360(%rsp),%r12
+movq 368(%rsp),%r13
+movq 376(%rsp),%r14
+movq 384(%rsp),%r15
+movq 392(%rsp),%rbx
+movq 400(%rsp),%rbp
+add %r11,%rsp
+xor %rax,%rax
+xor %rdx,%rdx
+ret
+
+._bytesatleast65:
+sub  $64,%r9
+add  $64,%rdi
+add  $64,%rsi
+jmp ._bytesbetween1and255
+
+#endif

src/libsodium/crypto_stream/salsa20/ref/stream_salsa20_ref.c

@@ -7,6 +7,8 @@ Public domain.
 #include "api.h"
 #include "crypto_core_salsa20.h"
 
+#ifndef HAVE_AMD64_ASM
+
 typedef unsigned int uint32;
 
 static const unsigned char sigma[16] = {
@@ -49,3 +51,5 @@ int crypto_stream(
   }
   return 0;
 }
+
+#endif

src/libsodium/crypto_stream/salsa20/ref/xor_salsa20_ref.c

@@ -7,6 +7,8 @@ Public domain.
 #include "api.h"
 #include "crypto_core_salsa20.h"
 
+#ifndef HAVE_AMD64_ASM
+
 typedef unsigned int uint32;
 
 static const unsigned char sigma[16] = {
@@ -52,3 +54,5 @@ int crypto_stream_xor(
   }
   return 0;
 }
+
+#endif

src/libsodium/include/sodium/crypto_scalarmult_curve25519.h.in

@@ -2,7 +2,9 @@
 #define crypto_scalarmult_curve25519_H
 
 #if @HAVE_TI_MODE_V@
-# define SODIUM_HAVE_TI_MODE
+# ifndef SODIUM_HAVE_TI_MODE
+#  define SODIUM_HAVE_TI_MODE
+# endif
 #endif
 
 #include "export.h"

src/libsodium/include/sodium/crypto_stream_salsa20.h

@@ -9,6 +9,12 @@
  *  the crypto_box functions.
  */
 
+#if 1
+# ifndef SODIUM_HAVE_AMD64_ASM
+#  define SODIUM_HAVE_AMD64_ASM
+# endif
+#endif
+
 #include "export.h"
 
 #define crypto_stream_salsa20_KEYBYTES 32U
@@ -24,7 +30,10 @@ int crypto_stream_salsa20(unsigned char *,unsigned long long,const unsigned char
 SODIUM_EXPORT
 int crypto_stream_salsa20_xor(unsigned char *,const unsigned char *,unsigned long long,const unsigned char *,const unsigned char *);
 
-#if 1
+#ifdef SODIUM_HAVE_AMD64_ASM
+# define crypto_stream_salsa20_amd64_xmm6 crypto_stream_salsa20
+# define crypto_stream_salsa20_amd64_xmm6_xor crypto_stream_salsa20_xor
+#else
 # define crypto_stream_salsa20_ref crypto_stream_salsa20
 # define crypto_stream_salsa20_ref_xor crypto_stream_salsa20_xor
 #endif
@@ -33,7 +42,4 @@ int crypto_stream_salsa20_xor(unsigned char *,const unsigned char *,unsigned lon
 }
 #endif
 
-#define crypto_stream_salsa20_ref crypto_stream_salsa20
-#define crypto_stream_salsa20_ref_xor crypto_stream_salsa20_xor
-
 #endif

src/libsodium/include/sodium/crypto_stream_salsa20.h.in

@@ -9,6 +9,12 @@
  *  the crypto_box functions.
  */
 
+#if @HAVE_AMD64_ASM_V@
+# ifndef SODIUM_HAVE_AMD64_ASM
+#  define SODIUM_HAVE_AMD64_ASM
+# endif
+#endif
+
 #include "export.h"
 
 #define crypto_stream_salsa20_KEYBYTES 32U
@@ -24,16 +30,16 @@ int crypto_stream_salsa20(unsigned char *,unsigned long long,const unsigned char
 SODIUM_EXPORT
 int crypto_stream_salsa20_xor(unsigned char *,const unsigned char *,unsigned long long,const unsigned char *,const unsigned char *);
 
-#if 1    
+#ifdef SODIUM_HAVE_AMD64_ASM
+# define crypto_stream_salsa20_amd64_xmm6 crypto_stream_salsa20
+# define crypto_stream_salsa20_amd64_xmm6_xor crypto_stream_salsa20_xor
+#else
 # define crypto_stream_salsa20_ref crypto_stream_salsa20
 # define crypto_stream_salsa20_ref_xor crypto_stream_salsa20_xor
-#endif    
+#endif
 
 #ifdef __cplusplus
 }
 #endif
 
-#define crypto_stream_salsa20_ref crypto_stream_salsa20
-#define crypto_stream_salsa20_ref_xor crypto_stream_salsa20_xor
-
 #endif