symm/rijndael-x86ish-aesni.S: Have `endswap_block' copy NKW to ECX.

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

diff --git a/symm/salsa20-x86ish-sse2.S b/symm/salsa20-x86ish-sse2.S
index b3ebf90..47401b7 100644 (file)
--- a/symm/salsa20-x86ish-sse2.S
+++ b/symm/salsa20-x86ish-sse2.S
@@ -42,7 +42,7 @@
 /// Main code.
 
        .arch pentium4
-       .section .text
+       .text
 
 FUNC(salsa20_core_x86ish_sse2)
 
@@ -152,7 +152,7 @@ FUNC(salsa20_core_x86ish_sse2)
        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