xref: /openbmc/linux/tools/arch/x86/lib/memset_64.S (revision 7f02ce62)
1/* SPDX-License-Identifier: GPL-2.0 */
2/* Copyright 2002 Andi Kleen, SuSE Labs */
3
4#include <linux/linkage.h>
5#include <asm/cpufeatures.h>
6#include <asm/alternative.h>
7#include <asm/export.h>
8
9.section .noinstr.text, "ax"
10
11/*
12 * ISO C memset - set a memory block to a byte value. This function uses fast
13 * string to get better performance than the original function. The code is
14 * simpler and shorter than the original function as well.
15 *
16 * rdi   destination
17 * rsi   value (char)
18 * rdx   count (bytes)
19 *
20 * rax   original destination
21 *
22 * The FSRS alternative should be done inline (avoiding the call and
23 * the disgusting return handling), but that would require some help
24 * from the compiler for better calling conventions.
25 *
26 * The 'rep stosb' itself is small enough to replace the call, but all
27 * the register moves blow up the code. And two of them are "needed"
28 * only for the return value that is the same as the source input,
29 * which the compiler could/should do much better anyway.
30 */
31SYM_FUNC_START(__memset)
32	ALTERNATIVE "jmp memset_orig", "", X86_FEATURE_FSRS
33
34	movq %rdi,%r9
35	movb %sil,%al
36	movq %rdx,%rcx
37	rep stosb
38	movq %r9,%rax
39	RET
40SYM_FUNC_END(__memset)
41EXPORT_SYMBOL(__memset)
42
43SYM_FUNC_ALIAS(memset, __memset)
44EXPORT_SYMBOL(memset)
45
46SYM_FUNC_START_LOCAL(memset_orig)
47	movq %rdi,%r10
48
49	/* expand byte value  */
50	movzbl %sil,%ecx
51	movabs $0x0101010101010101,%rax
52	imulq  %rcx,%rax
53
54	/* align dst */
55	movl  %edi,%r9d
56	andl  $7,%r9d
57	jnz  .Lbad_alignment
58.Lafter_bad_alignment:
59
60	movq  %rdx,%rcx
61	shrq  $6,%rcx
62	jz	 .Lhandle_tail
63
64	.p2align 4
65.Lloop_64:
66	decq  %rcx
67	movq  %rax,(%rdi)
68	movq  %rax,8(%rdi)
69	movq  %rax,16(%rdi)
70	movq  %rax,24(%rdi)
71	movq  %rax,32(%rdi)
72	movq  %rax,40(%rdi)
73	movq  %rax,48(%rdi)
74	movq  %rax,56(%rdi)
75	leaq  64(%rdi),%rdi
76	jnz    .Lloop_64
77
78	/* Handle tail in loops. The loops should be faster than hard
79	   to predict jump tables. */
80	.p2align 4
81.Lhandle_tail:
82	movl	%edx,%ecx
83	andl    $63&(~7),%ecx
84	jz 		.Lhandle_7
85	shrl	$3,%ecx
86	.p2align 4
87.Lloop_8:
88	decl   %ecx
89	movq  %rax,(%rdi)
90	leaq  8(%rdi),%rdi
91	jnz    .Lloop_8
92
93.Lhandle_7:
94	andl	$7,%edx
95	jz      .Lende
96	.p2align 4
97.Lloop_1:
98	decl    %edx
99	movb 	%al,(%rdi)
100	leaq	1(%rdi),%rdi
101	jnz     .Lloop_1
102
103.Lende:
104	movq	%r10,%rax
105	RET
106
107.Lbad_alignment:
108	cmpq $7,%rdx
109	jbe	.Lhandle_7
110	movq %rax,(%rdi)	/* unaligned store */
111	movq $8,%r8
112	subq %r9,%r8
113	addq %r8,%rdi
114	subq %r8,%rdx
115	jmp .Lafter_bad_alignment
116.Lfinal:
117SYM_FUNC_END(memset_orig)
118