summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorZhang Xianyi <traits.zhang@gmail.com>2013-07-01 16:09:05 +0800
committerZhang Xianyi <traits.zhang@gmail.com>2013-07-01 16:09:05 +0800
commit57944538b64b11a644c11a2fe3dabb0503a4a815 (patch)
tree446aa64e45baa5f223a2360bec2c69c216ba9c4a
parent3ce2c62b0bb5ccff69d89117fd9a894814b89b1a (diff)
downloadopenblas-57944538b64b11a644c11a2fe3dabb0503a4a815.tar.gz
openblas-57944538b64b11a644c11a2fe3dabb0503a4a815.tar.bz2
openblas-57944538b64b11a644c11a2fe3dabb0503a4a815.zip
Use ALIGN_5 instead of .algin 32 in assembly kernel. Added ALIGN_5 for 32-bit OSX.
-rw-r--r--common_x86.h4
-rw-r--r--kernel/x86_64/dgemv_n_bulldozer.S2
2 files changed, 3 insertions, 3 deletions
diff --git a/common_x86.h b/common_x86.h
index 4316318ec..12b348b8e 100644
--- a/common_x86.h
+++ b/common_x86.h
@@ -335,6 +335,7 @@ REALNAME:
#define ALIGN_2 .align 2
#define ALIGN_3 .align 3
#define ALIGN_4 .align 4
+#define ALIGN_5 .align 5
#define ffreep fstp
#endif
@@ -356,11 +357,10 @@ REALNAME:
#ifndef ALIGN_6
#define ALIGN_6 .align 64
-
+#endif
// ffreep %st(0).
// Because Clang didn't support ffreep, we directly use the opcode.
// Please check out http://www.sandpile.org/x86/opc_fpu.htm
#ifndef ffreep
#define ffreep .byte 0xdf, 0xc0 #
#endif
-#endif
diff --git a/kernel/x86_64/dgemv_n_bulldozer.S b/kernel/x86_64/dgemv_n_bulldozer.S
index c954f1929..ef2c4e23f 100644
--- a/kernel/x86_64/dgemv_n_bulldozer.S
+++ b/kernel/x86_64/dgemv_n_bulldozer.S
@@ -313,7 +313,7 @@
decq I
jle .L14
- .align 32
+ ALIGN_5
.L13: