Merge branch 'master' of git.distorted.org.uk:~mdw/publish/public-git/catacomb

[catacomb] / symm / salsa20-x86ish-sse2.S

diff --git a/symm/salsa20-x86ish-sse2.S b/symm/salsa20-x86ish-sse2.S
index a168d79..d74836b 100644 (file)
--- a/symm/salsa20-x86ish-sse2.S
+++ b/symm/salsa20-x86ish-sse2.S
@@ -146,13 +146,13 @@ FUNC(salsa20_core_x86ish_sse2)
        movdqu  xmm2, [IN + 32]
        movdqu  xmm3, [IN + 48]
 
-       ## Take a copy for later.
+       // Take a copy for later.
        movdqa  SAVE0, xmm0
        movdqa  SAVE1, xmm1
        movdqa  SAVE2, xmm2
        movdqa  SAVE3, xmm3
 
-loop:
+0:
        // Apply a column quarterround to each of the columns simultaneously.
        // Alas, there doesn't seem to be a packed doubleword rotate, so we
        // have to synthesize it.
@@ -256,7 +256,7 @@ loop:
        // Decrement the loop counter and see if we should go round again.
        // Later processors fuse this pair into a single uop.
        sub     NR, 2
-       ja      loop
+       ja      0b
 
        // Almost there.  Firstly, the feedforward addition, and then we have
        // to write out the result.  Here we have to undo the permutation