Add missing pshufd.

diff --git a/x86_64/sha_ni/sha1-compress.asm b/x86_64/sha_ni/sha1-compress.asm
index 3eb73366abaee7df73ae8e93050c4d7eb3ac6712..f0097eef06f8415240e6a0710326abc8be9c4a74 100644 (file)
--- a/x86_64/sha_ni/sha1-compress.asm
+++ b/x86_64/sha_ni/sha1-compress.asm
@@ -72,6 +72,8 @@ PROLOGUE(_nettle_sha1_compress)
        movd    16(STATE), E0
        movups  (INPUT), MSG0
        movdqa  .Lswap_mask(%rip), SWAP_MASK
+       pshufd  $0x1b, ABCD, ABCD
+       pshufd  $0x1b, E0, E0
        movdqa  ABCD, ABCD_ORIG
        movdqa  E0, E_ORIG
        pshufb  SWAP_MASK, MSG0
@@ -136,7 +138,9 @@ PROLOGUE(_nettle_sha1_compress)
        sha1nexte E_ORIG, E0
        paddd   ABCD_ORIG, ABCD
 
+       pshufd  $0x1b, ABCD, ABCD
        movups  ABCD, (STATE)
+       pshufd  $0x1b, E0, E0
        movd    E0, 16(STATE)
 
        W64_EXIT(2, 10)