tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
crypto: serpent-sse2 - change transpose_4x4 to only use integer instructions
Matrix transpose macro in serpent-sse2 uses mix of SSE2 integer and SSE floating
point instructions, which might cause performance penality on some CPUs.
This patch replaces transpose_4x4 macro with version that uses only SSE2
integer instructions.
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@mbnet.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
authored by
Jussi Kivilinna
and committed by
Herbert Xu
847cb7ef
4c58464b
+26
-32
+13
-16
arch/x86/crypto/serpent-sse2-i586-asm_32.S
+13
-16
arch/x86/crypto/serpent-sse2-i586-asm_32.S
···
463
463
pand x0, x4; \
464
464
pxor x2, x4;
465
465
466
-
#define transpose_4x4(x0, x1, x2, x3, t1, t2, t3) \
467
-
movdqa x2, t3; \
468
-
movdqa x0, t1; \
469
-
unpcklps x3, t3; \
466
+
#define transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \
470
467
movdqa x0, t2; \
471
-
unpcklps x1, t1; \
472
-
unpckhps x1, t2; \
473
-
movdqa t3, x1; \
474
-
unpckhps x3, x2; \
475
-
movdqa t1, x0; \
476
-
movhlps t1, x1; \
477
-
movdqa t2, t1; \
478
-
movlhps t3, x0; \
479
-
movlhps x2, t1; \
480
-
movhlps t2, x2; \
481
-
movdqa x2, x3; \
482
-
movdqa t1, x2;
468
+
punpckldq x1, x0; \
469
+
punpckhdq x1, t2; \
470
+
movdqa x2, t1; \
471
+
punpckhdq x3, x2; \
472
+
punpckldq x3, t1; \
473
+
movdqa x0, x1; \
474
+
punpcklqdq t1, x0; \
475
+
punpckhqdq t1, x1; \
476
+
movdqa t2, x3; \
477
+
punpcklqdq x2, t2; \
478
+
punpckhqdq x2, x3; \
479
+
movdqa t2, x2;
483
480
484
481
#define read_blocks(in, x0, x1, x2, x3, t0, t1, t2) \
485
482
movdqu (0*4*4)(in), x0; \
+13
-16
arch/x86/crypto/serpent-sse2-x86_64-asm_64.S
+13
-16
arch/x86/crypto/serpent-sse2-x86_64-asm_64.S
···
585
585
get_key(i, 1, RK1); \
586
586
SBOX ## _2(x0 ## 2, x1 ## 2, x2 ## 2, x3 ## 2, x4 ## 2); \
587
587
588
-
#define transpose_4x4(x0, x1, x2, x3, t1, t2, t3) \
589
-
movdqa x2, t3; \
590
-
movdqa x0, t1; \
591
-
unpcklps x3, t3; \
588
+
#define transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \
592
589
movdqa x0, t2; \
593
-
unpcklps x1, t1; \
594
-
unpckhps x1, t2; \
595
-
movdqa t3, x1; \
596
-
unpckhps x3, x2; \
597
-
movdqa t1, x0; \
598
-
movhlps t1, x1; \
599
-
movdqa t2, t1; \
600
-
movlhps t3, x0; \
601
-
movlhps x2, t1; \
602
-
movhlps t2, x2; \
603
-
movdqa x2, x3; \
604
-
movdqa t1, x2;
590
+
punpckldq x1, x0; \
591
+
punpckhdq x1, t2; \
592
+
movdqa x2, t1; \
593
+
punpckhdq x3, x2; \
594
+
punpckldq x3, t1; \
595
+
movdqa x0, x1; \
596
+
punpcklqdq t1, x0; \
597
+
punpckhqdq t1, x1; \
598
+
movdqa t2, x3; \
599
+
punpcklqdq x2, t2; \
600
+
punpckhqdq x2, x3; \
601
+
movdqa t2, x2;
605
602
606
603
#define read_blocks(in, x0, x1, x2, x3, t0, t1, t2) \
607
604
movdqu (0*4*4)(in), x0; \