~mdw
/
catacomb
/ blobdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
|
commitdiff
|
tree
raw
|
inline
| side by side
base/asm-common.h, *.S: Use consistent little-endian notation for SIMD regs.
[catacomb]
/
symm
/
salsa20-x86ish-sse2.S
diff --git
a/symm/salsa20-x86ish-sse2.S
b/symm/salsa20-x86ish-sse2.S
index
9cbaeff
..
ccf912b
100644
(file)
--- a/
symm/salsa20-x86ish-sse2.S
+++ b/
symm/salsa20-x86ish-sse2.S
@@
-25,16
+25,24
@@
/// MA 02111-1307, USA.
///--------------------------------------------------------------------------
/// MA 02111-1307, USA.
///--------------------------------------------------------------------------
-///
External definition
s.
+///
Preliminarie
s.
#include "config.h"
#include "asm-common.h"
#include "config.h"
#include "asm-common.h"
+ .text
+
///--------------------------------------------------------------------------
/// Main code.
///--------------------------------------------------------------------------
/// Main code.
- .arch pentium4
- .text
+FUNC(salsa20_core_x86ish_avx)
+ .arch .avx
+ vzeroupper
+ endprologue
+ // drop through...
+ENDFUNC
+
+ .arch pentium4
FUNC(salsa20_core_x86ish_sse2)
FUNC(salsa20_core_x86ish_sse2)
@@
-57,16
+65,16
@@
FUNC(salsa20_core_x86ish_sse2)
# define OUT edx
# define SAVE0 xmm6
# define SAVE1 xmm7
# define OUT edx
# define SAVE0 xmm6
# define SAVE1 xmm7
-# define SAVE2 [
esp
+ 0]
-# define SAVE3 [
esp
+ 16]
-
- pushreg
ebp
- setfp
ebp
- s
ub esp,
32
- mov IN, [
ebp
+ 12]
- mov OUT, [
ebp
+ 16]
- and
esp
, ~15
- mov NR, [
ebp
+ 8]
+# define SAVE2 [
SP
+ 0]
+# define SAVE3 [
SP
+ 16]
+
+ pushreg
BP
+ setfp
+ s
talloc
32
+ mov IN, [
BP
+ 12]
+ mov OUT, [
BP
+ 16]
+ and
SP
, ~15
+ mov NR, [
BP
+ 8]
#endif
#if CPUFAM_AMD64 && ABI_SYSV
#endif
#if CPUFAM_AMD64 && ABI_SYSV
@@
-99,8
+107,8
@@
FUNC(salsa20_core_x86ish_sse2)
# define OUT r8
# define SAVE0 xmm6
# define SAVE1 xmm7
# define OUT r8
# define SAVE0 xmm6
# define SAVE1 xmm7
-# define SAVE2 [
rsp
+ 32]
-# define SAVE3 [
rsp
+ 48]
+# define SAVE2 [
SP
+ 32]
+# define SAVE3 [
SP
+ 48]
stalloc 64 + 8
savexmm xmm6, 0
stalloc 64 + 8
savexmm xmm6, 0
@@
-293,11
+301,11
@@
FUNC(salsa20_core_x86ish_sse2)
// Tidy things up.
#if CPUFAM_X86
dropfp
// Tidy things up.
#if CPUFAM_X86
dropfp
- popreg
ebp
+ popreg
BP
#endif
#if CPUFAM_AMD64 && ABI_WIN
rstrxmm xmm6, 0
#endif
#if CPUFAM_AMD64 && ABI_WIN
rstrxmm xmm6, 0
- rs
r
rxmm xmm7, 16
+ rs
t
rxmm xmm7, 16
stfree 64 + 8
#endif
stfree 64 + 8
#endif