Blame - arch/x86/lib/copy_page_64.S - kernel/msm-5.4

blob: fd2d09afa0974b59a0bf54063682e6502fb44573 [file] [log] [blame]

Greg Kroah-Hartman	b244131	2017-11-01 15:07:57 +0100	[diff] [blame]	1	/* SPDX-License-Identifier: GPL-2.0 */
Linus Torvalds	1da177e	2005-04-16 15:20:36 -0700	[diff] [blame]	2	/* Written 2003 by Andi Kleen, based on a kernel by Evandro Menezes */
Dave Jones	038b0a6	2006-10-04 03:38:54 -0400	[diff] [blame]	3
Jan Beulich	8d379da	2006-09-26 10:52:32 +0200	[diff] [blame]	4	#include <linux/linkage.h>
Borislav Petkov	cd4d09e	2016-01-26 22:12:04 +0100	[diff] [blame]	5	#include <asm/cpufeatures.h>
Andy Lutomirski	59e97e4	2011-07-13 09:24:10 -0400	[diff] [blame]	6	#include <asm/alternative-asm.h>
Al Viro	784d569	2016-01-11 11:04:34 -0500	[diff] [blame]	7	#include <asm/export.h>
Jan Beulich	8d379da	2006-09-26 10:52:32 +0200	[diff] [blame]	8
Borislav Petkov	090a3f6	2015-01-12 18:19:40 +0100	[diff] [blame]	9	/*
				10	* Some CPUs run faster using the string copy instructions (sane microcode).
				11	* It is also a lot simpler. Use this when possible. But, don't use streaming
				12	* copy unless the CPU indicates X86_FEATURE_REP_GOOD. Could vary the
				13	* prefetch distance based on SMP/UP.
				14	*/
Jan Beulich	8d379da	2006-09-26 10:52:32 +0200	[diff] [blame]	15	ALIGN
Borislav Petkov	090a3f6	2015-01-12 18:19:40 +0100	[diff] [blame]	16	ENTRY(copy_page)
Borislav Petkov	090a3f6	2015-01-12 18:19:40 +0100	[diff] [blame]	17	ALTERNATIVE "jmp copy_page_regs", "", X86_FEATURE_REP_GOOD
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	18	movl $4096/8, %ecx
				19	rep movsq
Jan Beulich	8d379da	2006-09-26 10:52:32 +0200	[diff] [blame]	20	ret
Borislav Petkov	090a3f6	2015-01-12 18:19:40 +0100	[diff] [blame]	21	ENDPROC(copy_page)
Al Viro	784d569	2016-01-11 11:04:34 -0500	[diff] [blame]	22	EXPORT_SYMBOL(copy_page)
Jan Beulich	8d379da	2006-09-26 10:52:32 +0200	[diff] [blame]	23
Borislav Petkov	090a3f6	2015-01-12 18:19:40 +0100	[diff] [blame]	24	ENTRY(copy_page_regs)
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	25	subq $2*8, %rsp
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	26	movq %rbx, (%rsp)
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	27	movq %r12, 1*8(%rsp)
Andi Kleen	7bcd3f3	2006-02-03 21:51:02 +0100	[diff] [blame]	28
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	29	movl $(4096/64)-5, %ecx
Andi Kleen	7bcd3f3	2006-02-03 21:51:02 +0100	[diff] [blame]	30	.p2align 4
				31	.Loop64:
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	32	dec %rcx
				33	movq 0x8*0(%rsi), %rax
				34	movq 0x8*1(%rsi), %rbx
				35	movq 0x8*2(%rsi), %rdx
				36	movq 0x8*3(%rsi), %r8
				37	movq 0x8*4(%rsi), %r9
				38	movq 0x8*5(%rsi), %r10
				39	movq 0x8*6(%rsi), %r11
				40	movq 0x8*7(%rsi), %r12
Andi Kleen	7bcd3f3	2006-02-03 21:51:02 +0100	[diff] [blame]	41
				42	prefetcht0 5*64(%rsi)
				43
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	44	movq %rax, 0x8*0(%rdi)
				45	movq %rbx, 0x8*1(%rdi)
				46	movq %rdx, 0x8*2(%rdi)
				47	movq %r8, 0x8*3(%rdi)
				48	movq %r9, 0x8*4(%rdi)
				49	movq %r10, 0x8*5(%rdi)
				50	movq %r11, 0x8*6(%rdi)
				51	movq %r12, 0x8*7(%rdi)
Andi Kleen	7bcd3f3	2006-02-03 21:51:02 +0100	[diff] [blame]	52
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	53	leaq 64 (%rsi), %rsi
				54	leaq 64 (%rdi), %rdi
Andi Kleen	7bcd3f3	2006-02-03 21:51:02 +0100	[diff] [blame]	55
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	56	jnz .Loop64
Andi Kleen	7bcd3f3	2006-02-03 21:51:02 +0100	[diff] [blame]	57
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	58	movl $5, %ecx
Andi Kleen	7bcd3f3	2006-02-03 21:51:02 +0100	[diff] [blame]	59	.p2align 4
				60	.Loop2:
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	61	decl %ecx
Andi Kleen	7bcd3f3	2006-02-03 21:51:02 +0100	[diff] [blame]	62
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	63	movq 0x8*0(%rsi), %rax
				64	movq 0x8*1(%rsi), %rbx
				65	movq 0x8*2(%rsi), %rdx
				66	movq 0x8*3(%rsi), %r8
				67	movq 0x8*4(%rsi), %r9
				68	movq 0x8*5(%rsi), %r10
				69	movq 0x8*6(%rsi), %r11
				70	movq 0x8*7(%rsi), %r12
Andi Kleen	7bcd3f3	2006-02-03 21:51:02 +0100	[diff] [blame]	71
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	72	movq %rax, 0x8*0(%rdi)
				73	movq %rbx, 0x8*1(%rdi)
				74	movq %rdx, 0x8*2(%rdi)
				75	movq %r8, 0x8*3(%rdi)
				76	movq %r9, 0x8*4(%rdi)
				77	movq %r10, 0x8*5(%rdi)
				78	movq %r11, 0x8*6(%rdi)
				79	movq %r12, 0x8*7(%rdi)
Andi Kleen	7bcd3f3	2006-02-03 21:51:02 +0100	[diff] [blame]	80
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	81	leaq 64(%rdi), %rdi
				82	leaq 64(%rsi), %rsi
Andi Kleen	7bcd3f3	2006-02-03 21:51:02 +0100	[diff] [blame]	83	jnz .Loop2
				84
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	85	movq (%rsp), %rbx
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	86	movq 1*8(%rsp), %r12
Ma Ling	269833b	2012-10-18 03:52:45 +0800	[diff] [blame]	87	addq $2*8, %rsp
Andi Kleen	7bcd3f3	2006-02-03 21:51:02 +0100	[diff] [blame]	88	ret
Borislav Petkov	090a3f6	2015-01-12 18:19:40 +0100	[diff] [blame]	89	ENDPROC(copy_page_regs)