xref: /linux/tools/arch/x86/lib/memset_64.S (revision 7c7f9dd1ea3fc6b175b7227bb473dc883b925548)
1/* SPDX-License-Identifier: GPL-2.0 */
2/* Copyright 2002 Andi Kleen, SuSE Labs */
3
4#include <linux/export.h>
5#include <linux/linkage.h>
6#include <linux/cfi_types.h>
7#include <asm/cpufeatures.h>
8#include <asm/alternative.h>
9
10.section .noinstr.text, "ax"
11
12/*
13 * ISO C memset - set a memory block to a byte value. This function uses fast
14 * string to get better performance than the original function. The code is
15 * simpler and shorter than the original function as well.
16 *
17 * rdi   destination
18 * rsi   value (char)
19 * rdx   count (bytes)
20 *
21 * rax   original destination
22 *
23 * The FSRS alternative should be done inline (avoiding the call and
24 * the disgusting return handling), but that would require some help
25 * from the compiler for better calling conventions.
26 *
27 * The 'rep stosb' itself is small enough to replace the call, but all
28 * the register moves blow up the code. And two of them are "needed"
29 * only for the return value that is the same as the source input,
30 * which the compiler could/should do much better anyway.
31 */
32SYM_TYPED_FUNC_START(__memset)
33	ALTERNATIVE "jmp memset_orig", "", X86_FEATURE_FSRS
34
35	movq %rdi,%r9
36	movb %sil,%al
37	movq %rdx,%rcx
38	rep stosb
39	movq %r9,%rax
40	RET
41SYM_FUNC_END(__memset)
42EXPORT_SYMBOL(__memset)
43
44SYM_FUNC_ALIAS_MEMFUNC(memset, __memset)
45SYM_PIC_ALIAS(memset)
46EXPORT_SYMBOL(memset)
47
48SYM_FUNC_START_LOCAL(memset_orig)
49	movq %rdi,%r10
50
51	/* expand byte value  */
52	movzbl %sil,%ecx
53	movabs $0x0101010101010101,%rax
54	imulq  %rcx,%rax
55
56	/* align dst */
57	movl  %edi,%r9d
58	andl  $7,%r9d
59	jnz  .Lbad_alignment
60.Lafter_bad_alignment:
61
62	movq  %rdx,%rcx
63	shrq  $6,%rcx
64	jz	 .Lhandle_tail
65
66	.p2align 4
67.Lloop_64:
68	decq  %rcx
69	movq  %rax,(%rdi)
70	movq  %rax,8(%rdi)
71	movq  %rax,16(%rdi)
72	movq  %rax,24(%rdi)
73	movq  %rax,32(%rdi)
74	movq  %rax,40(%rdi)
75	movq  %rax,48(%rdi)
76	movq  %rax,56(%rdi)
77	leaq  64(%rdi),%rdi
78	jnz    .Lloop_64
79
80	/* Handle tail in loops. The loops should be faster than hard
81	   to predict jump tables. */
82	.p2align 4
83.Lhandle_tail:
84	movl	%edx,%ecx
85	andl    $63&(~7),%ecx
86	jz 		.Lhandle_7
87	shrl	$3,%ecx
88	.p2align 4
89.Lloop_8:
90	decl   %ecx
91	movq  %rax,(%rdi)
92	leaq  8(%rdi),%rdi
93	jnz    .Lloop_8
94
95.Lhandle_7:
96	andl	$7,%edx
97	jz      .Lende
98	.p2align 4
99.Lloop_1:
100	decl    %edx
101	movb 	%al,(%rdi)
102	leaq	1(%rdi),%rdi
103	jnz     .Lloop_1
104
105.Lende:
106	movq	%r10,%rax
107	RET
108
109.Lbad_alignment:
110	cmpq $7,%rdx
111	jbe	.Lhandle_7
112	movq %rax,(%rdi)	/* unaligned store */
113	movq $8,%r8
114	subq %r9,%r8
115	addq %r8,%rdi
116	subq %r8,%rdx
117	jmp .Lafter_bad_alignment
118.Lfinal:
119SYM_FUNC_END(memset_orig)
120