diff options
author | Al Viro <viro@zeniv.linux.org.uk> | 2008-08-17 21:05:42 -0400 |
---|---|---|
committer | H. Peter Anvin <hpa@zytor.com> | 2008-10-22 22:55:20 -0700 |
commit | bb8985586b7a906e116db835c64773b7a7d51663 (patch) | |
tree | de93ae58e88cc563d95cc124a73f3930594c6100 /arch/x86/include/asm/xor_64.h | |
parent | 8ede0bdb63305d3353efd97e9af6210afb05734e (diff) | |
download | linux-3.10-bb8985586b7a906e116db835c64773b7a7d51663.tar.gz linux-3.10-bb8985586b7a906e116db835c64773b7a7d51663.tar.bz2 linux-3.10-bb8985586b7a906e116db835c64773b7a7d51663.zip |
x86, um: ... and asm-x86 move
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Diffstat (limited to 'arch/x86/include/asm/xor_64.h')
-rw-r--r-- | arch/x86/include/asm/xor_64.h | 361 |
1 files changed, 361 insertions, 0 deletions
diff --git a/arch/x86/include/asm/xor_64.h b/arch/x86/include/asm/xor_64.h new file mode 100644 index 00000000000..2d3a18de295 --- /dev/null +++ b/arch/x86/include/asm/xor_64.h @@ -0,0 +1,361 @@ +#ifndef ASM_X86__XOR_64_H +#define ASM_X86__XOR_64_H + +/* + * Optimized RAID-5 checksumming functions for MMX and SSE. + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation; either version 2, or (at your option) + * any later version. + * + * You should have received a copy of the GNU General Public License + * (for example /usr/src/linux/COPYING); if not, write to the Free + * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + */ + + +/* + * Cache avoiding checksumming functions utilizing KNI instructions + * Copyright (C) 1999 Zach Brown (with obvious credit due Ingo) + */ + +/* + * Based on + * High-speed RAID5 checksumming functions utilizing SSE instructions. + * Copyright (C) 1998 Ingo Molnar. + */ + +/* + * x86-64 changes / gcc fixes from Andi Kleen. + * Copyright 2002 Andi Kleen, SuSE Labs. + * + * This hasn't been optimized for the hammer yet, but there are likely + * no advantages to be gotten from x86-64 here anyways. + */ + +typedef struct { + unsigned long a, b; +} __attribute__((aligned(16))) xmm_store_t; + +/* Doesn't use gcc to save the XMM registers, because there is no easy way to + tell it to do a clts before the register saving. */ +#define XMMS_SAVE \ +do { \ + preempt_disable(); \ + asm volatile( \ + "movq %%cr0,%0 ;\n\t" \ + "clts ;\n\t" \ + "movups %%xmm0,(%1) ;\n\t" \ + "movups %%xmm1,0x10(%1) ;\n\t" \ + "movups %%xmm2,0x20(%1) ;\n\t" \ + "movups %%xmm3,0x30(%1) ;\n\t" \ + : "=&r" (cr0) \ + : "r" (xmm_save) \ + : "memory"); \ +} while (0) + +#define XMMS_RESTORE \ +do { \ + asm volatile( \ + "sfence ;\n\t" \ + "movups (%1),%%xmm0 ;\n\t" \ + "movups 0x10(%1),%%xmm1 ;\n\t" \ + "movups 0x20(%1),%%xmm2 ;\n\t" \ + "movups 0x30(%1),%%xmm3 ;\n\t" \ + "movq %0,%%cr0 ;\n\t" \ + : \ + : "r" (cr0), "r" (xmm_save) \ + : "memory"); \ + preempt_enable(); \ +} while (0) + +#define OFFS(x) "16*("#x")" +#define PF_OFFS(x) "256+16*("#x")" +#define PF0(x) " prefetchnta "PF_OFFS(x)"(%[p1]) ;\n" +#define LD(x, y) " movaps "OFFS(x)"(%[p1]), %%xmm"#y" ;\n" +#define ST(x, y) " movaps %%xmm"#y", "OFFS(x)"(%[p1]) ;\n" +#define PF1(x) " prefetchnta "PF_OFFS(x)"(%[p2]) ;\n" +#define PF2(x) " prefetchnta "PF_OFFS(x)"(%[p3]) ;\n" +#define PF3(x) " prefetchnta "PF_OFFS(x)"(%[p4]) ;\n" +#define PF4(x) " prefetchnta "PF_OFFS(x)"(%[p5]) ;\n" +#define PF5(x) " prefetchnta "PF_OFFS(x)"(%[p6]) ;\n" +#define XO1(x, y) " xorps "OFFS(x)"(%[p2]), %%xmm"#y" ;\n" +#define XO2(x, y) " xorps "OFFS(x)"(%[p3]), %%xmm"#y" ;\n" +#define XO3(x, y) " xorps "OFFS(x)"(%[p4]), %%xmm"#y" ;\n" +#define XO4(x, y) " xorps "OFFS(x)"(%[p5]), %%xmm"#y" ;\n" +#define XO5(x, y) " xorps "OFFS(x)"(%[p6]), %%xmm"#y" ;\n" + + +static void +xor_sse_2(unsigned long bytes, unsigned long *p1, unsigned long *p2) +{ + unsigned int lines = bytes >> 8; + unsigned long cr0; + xmm_store_t xmm_save[4]; + + XMMS_SAVE; + + asm volatile( +#undef BLOCK +#define BLOCK(i) \ + LD(i, 0) \ + LD(i + 1, 1) \ + PF1(i) \ + PF1(i + 2) \ + LD(i + 2, 2) \ + LD(i + 3, 3) \ + PF0(i + 4) \ + PF0(i + 6) \ + XO1(i, 0) \ + XO1(i + 1, 1) \ + XO1(i + 2, 2) \ + XO1(i + 3, 3) \ + ST(i, 0) \ + ST(i + 1, 1) \ + ST(i + 2, 2) \ + ST(i + 3, 3) \ + + + PF0(0) + PF0(2) + + " .align 32 ;\n" + " 1: ;\n" + + BLOCK(0) + BLOCK(4) + BLOCK(8) + BLOCK(12) + + " addq %[inc], %[p1] ;\n" + " addq %[inc], %[p2] ;\n" + " decl %[cnt] ; jnz 1b" + : [p1] "+r" (p1), [p2] "+r" (p2), [cnt] "+r" (lines) + : [inc] "r" (256UL) + : "memory"); + + XMMS_RESTORE; +} + +static void +xor_sse_3(unsigned long bytes, unsigned long *p1, unsigned long *p2, + unsigned long *p3) +{ + unsigned int lines = bytes >> 8; + xmm_store_t xmm_save[4]; + unsigned long cr0; + + XMMS_SAVE; + + asm volatile( +#undef BLOCK +#define BLOCK(i) \ + PF1(i) \ + PF1(i + 2) \ + LD(i, 0) \ + LD(i + 1, 1) \ + LD(i + 2, 2) \ + LD(i + 3, 3) \ + PF2(i) \ + PF2(i + 2) \ + PF0(i + 4) \ + PF0(i + 6) \ + XO1(i, 0) \ + XO1(i + 1, 1) \ + XO1(i + 2, 2) \ + XO1(i + 3, 3) \ + XO2(i, 0) \ + XO2(i + 1, 1) \ + XO2(i + 2, 2) \ + XO2(i + 3, 3) \ + ST(i, 0) \ + ST(i + 1, 1) \ + ST(i + 2, 2) \ + ST(i + 3, 3) \ + + + PF0(0) + PF0(2) + + " .align 32 ;\n" + " 1: ;\n" + + BLOCK(0) + BLOCK(4) + BLOCK(8) + BLOCK(12) + + " addq %[inc], %[p1] ;\n" + " addq %[inc], %[p2] ;\n" + " addq %[inc], %[p3] ;\n" + " decl %[cnt] ; jnz 1b" + : [cnt] "+r" (lines), + [p1] "+r" (p1), [p2] "+r" (p2), [p3] "+r" (p3) + : [inc] "r" (256UL) + : "memory"); + XMMS_RESTORE; +} + +static void +xor_sse_4(unsigned long bytes, unsigned long *p1, unsigned long *p2, + unsigned long *p3, unsigned long *p4) +{ + unsigned int lines = bytes >> 8; + xmm_store_t xmm_save[4]; + unsigned long cr0; + + XMMS_SAVE; + + asm volatile( +#undef BLOCK +#define BLOCK(i) \ + PF1(i) \ + PF1(i + 2) \ + LD(i, 0) \ + LD(i + 1, 1) \ + LD(i + 2, 2) \ + LD(i + 3, 3) \ + PF2(i) \ + PF2(i + 2) \ + XO1(i, 0) \ + XO1(i + 1, 1) \ + XO1(i + 2, 2) \ + XO1(i + 3, 3) \ + PF3(i) \ + PF3(i + 2) \ + PF0(i + 4) \ + PF0(i + 6) \ + XO2(i, 0) \ + XO2(i + 1, 1) \ + XO2(i + 2, 2) \ + XO2(i + 3, 3) \ + XO3(i, 0) \ + XO3(i + 1, 1) \ + XO3(i + 2, 2) \ + XO3(i + 3, 3) \ + ST(i, 0) \ + ST(i + 1, 1) \ + ST(i + 2, 2) \ + ST(i + 3, 3) \ + + + PF0(0) + PF0(2) + + " .align 32 ;\n" + " 1: ;\n" + + BLOCK(0) + BLOCK(4) + BLOCK(8) + BLOCK(12) + + " addq %[inc], %[p1] ;\n" + " addq %[inc], %[p2] ;\n" + " addq %[inc], %[p3] ;\n" + " addq %[inc], %[p4] ;\n" + " decl %[cnt] ; jnz 1b" + : [cnt] "+c" (lines), + [p1] "+r" (p1), [p2] "+r" (p2), [p3] "+r" (p3), [p4] "+r" (p4) + : [inc] "r" (256UL) + : "memory" ); + + XMMS_RESTORE; +} + +static void +xor_sse_5(unsigned long bytes, unsigned long *p1, unsigned long *p2, + unsigned long *p3, unsigned long *p4, unsigned long *p5) +{ + unsigned int lines = bytes >> 8; + xmm_store_t xmm_save[4]; + unsigned long cr0; + + XMMS_SAVE; + + asm volatile( +#undef BLOCK +#define BLOCK(i) \ + PF1(i) \ + PF1(i + 2) \ + LD(i, 0) \ + LD(i + 1, 1) \ + LD(i + 2, 2) \ + LD(i + 3, 3) \ + PF2(i) \ + PF2(i + 2) \ + XO1(i, 0) \ + XO1(i + 1, 1) \ + XO1(i + 2, 2) \ + XO1(i + 3, 3) \ + PF3(i) \ + PF3(i + 2) \ + XO2(i, 0) \ + XO2(i + 1, 1) \ + XO2(i + 2, 2) \ + XO2(i + 3, 3) \ + PF4(i) \ + PF4(i + 2) \ + PF0(i + 4) \ + PF0(i + 6) \ + XO3(i, 0) \ + XO3(i + 1, 1) \ + XO3(i + 2, 2) \ + XO3(i + 3, 3) \ + XO4(i, 0) \ + XO4(i + 1, 1) \ + XO4(i + 2, 2) \ + XO4(i + 3, 3) \ + ST(i, 0) \ + ST(i + 1, 1) \ + ST(i + 2, 2) \ + ST(i + 3, 3) \ + + + PF0(0) + PF0(2) + + " .align 32 ;\n" + " 1: ;\n" + + BLOCK(0) + BLOCK(4) + BLOCK(8) + BLOCK(12) + + " addq %[inc], %[p1] ;\n" + " addq %[inc], %[p2] ;\n" + " addq %[inc], %[p3] ;\n" + " addq %[inc], %[p4] ;\n" + " addq %[inc], %[p5] ;\n" + " decl %[cnt] ; jnz 1b" + : [cnt] "+c" (lines), + [p1] "+r" (p1), [p2] "+r" (p2), [p3] "+r" (p3), [p4] "+r" (p4), + [p5] "+r" (p5) + : [inc] "r" (256UL) + : "memory"); + + XMMS_RESTORE; +} + +static struct xor_block_template xor_block_sse = { + .name = "generic_sse", + .do_2 = xor_sse_2, + .do_3 = xor_sse_3, + .do_4 = xor_sse_4, + .do_5 = xor_sse_5, +}; + +#undef XOR_TRY_TEMPLATES +#define XOR_TRY_TEMPLATES \ +do { \ + xor_speed(&xor_block_sse); \ +} while (0) + +/* We force the use of the SSE xor block because it can write around L2. + We may also be able to load into the L1 only depending on how the cpu + deals with a load to a line that is being prefetched. */ +#define XOR_SELECT_TEMPLATE(FASTEST) (&xor_block_sse) + +#endif /* ASM_X86__XOR_64_H */ |