aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorAndi Kleen <ak@suse.de>2005-11-05 11:25:54 -0500
committerLinus Torvalds <torvalds@g5.osdl.org>2005-11-14 22:55:17 -0500
commita5b250a428aabc619ace872f8220a7d0b8f7d557 (patch)
tree11cabf07982ae37f94bc929f9a605cbbd20e35ab
parenta6f5deb2be4c82f24fefadcbf7e448f540c05ae6 (diff)
[PATCH] x86_64: Remove optimization for B stepping AMD K8
B stepping were the first shipping Opterons. memcpy/memset/copy_page/ clear_page had special optimized version for them. These are really old and in the minority now and the difference to the generic versions (using rep microcode) is not that big anyways. So just remove them. TODO: figure out optimized versions for Intel Netburst based EM64T Signed-off-by: Andi Kleen <ak@suse.de> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Diffstat
-rw-r--r--arch/x86_64/lib/clear_page.S38
-rw-r--r--arch/x86_64/lib/copy_page.S87
-rw-r--r--arch/x86_64/lib/memcpy.S93
-rw-r--r--arch/x86_64/lib/memset.S94
4 files changed, 2 insertions, 310 deletions
diff --git a/arch/x86_64/lib/clear_page.S b/arch/x86_64/lib/clear_page.S
index 30a9da458c15..43d9fa136180 100644
--- a/arch/x86_64/lib/clear_page.S
+++ b/arch/x86_64/lib/clear_page.S
@@ -5,46 +5,8 @@
5 .globl clear_page 5 .globl clear_page
6 .p2align 4 6 .p2align 4
7clear_page: 7clear_page:
8 xorl %eax,%eax
9 movl $4096/64,%ecx
10 .p2align 4
11.Lloop:
12 decl %ecx
13#define PUT(x) movq %rax,x*8(%rdi)
14 movq %rax,(%rdi)
15 PUT(1)
16 PUT(2)
17 PUT(3)
18 PUT(4)
19 PUT(5)
20 PUT(6)
21 PUT(7)
22 leaq 64(%rdi),%rdi
23 jnz .Lloop
24 nop
25 ret
26clear_page_end:
27
28 /* C stepping K8 run faster using the string instructions.
29 It is also a lot simpler. Use this when possible */
30
31#include <asm/cpufeature.h>
32
33 .section .altinstructions,"a"
34 .align 8
35 .quad clear_page
36 .quad clear_page_c
37 .byte X86_FEATURE_K8_C
38 .byte clear_page_end-clear_page
39 .byte clear_page_c_end-clear_page_c
40 .previous
41
42 .section .altinstr_replacement,"ax"
43clear_page_c:
44 movl $4096/8,%ecx 8 movl $4096/8,%ecx
45 xorl %eax,%eax 9 xorl %eax,%eax
46 rep 10 rep
47 stosq 11 stosq
48 ret 12 ret
49clear_page_c_end:
50 .previous
diff --git a/arch/x86_64/lib/copy_page.S b/arch/x86_64/lib/copy_page.S
index dd3aa47b6bf5..621a19769406 100644
--- a/arch/x86_64/lib/copy_page.S
+++ b/arch/x86_64/lib/copy_page.S
@@ -8,94 +8,7 @@
8 .globl copy_page 8 .globl copy_page
9 .p2align 4 9 .p2align 4
10copy_page: 10copy_page:
11 subq $3*8,%rsp
12 movq %rbx,(%rsp)
13 movq %r12,1*8(%rsp)
14 movq %r13,2*8(%rsp)
15
16 movl $(4096/64)-5,%ecx
17 .p2align 4
18.Loop64:
19 dec %rcx
20
21 movq (%rsi), %rax
22 movq 8 (%rsi), %rbx
23 movq 16 (%rsi), %rdx
24 movq 24 (%rsi), %r8
25 movq 32 (%rsi), %r9
26 movq 40 (%rsi), %r10
27 movq 48 (%rsi), %r11
28 movq 56 (%rsi), %r12
29
30 prefetcht0 5*64(%rsi)
31
32 movq %rax, (%rdi)
33 movq %rbx, 8 (%rdi)
34 movq %rdx, 16 (%rdi)
35 movq %r8, 24 (%rdi)
36 movq %r9, 32 (%rdi)
37 movq %r10, 40 (%rdi)
38 movq %r11, 48 (%rdi)
39 movq %r12, 56 (%rdi)
40
41 leaq 64 (%rsi), %rsi
42 leaq 64 (%rdi), %rdi
43
44 jnz .Loop64
45
46 movl $5,%ecx
47 .p2align 4
48.Loop2:
49 decl %ecx
50
51 movq (%rsi), %rax
52 movq 8 (%rsi), %rbx
53 movq 16 (%rsi), %rdx
54 movq 24 (%rsi), %r8
55 movq 32 (%rsi), %r9
56 movq 40 (%rsi), %r10
57 movq 48 (%rsi), %r11
58 movq 56 (%rsi), %r12
59
60 movq %rax, (%rdi)
61 movq %rbx, 8 (%rdi)
62 movq %rdx, 16 (%rdi)
63 movq %r8, 24 (%rdi)
64 movq %r9, 32 (%rdi)
65 movq %r10, 40 (%rdi)
66 movq %r11, 48 (%rdi)
67 movq %r12, 56 (%rdi)
68
69 leaq 64(%rdi),%rdi
70 leaq 64(%rsi),%rsi
71
72 jnz .Loop2
73
74 movq (%rsp),%rbx
75 movq 1*8(%rsp),%r12
76 movq 2*8(%rsp),%r13
77 addq $3*8,%rsp
78 ret
79
80 /* C stepping K8 run faster using the string copy instructions.
81 It is also a lot simpler. Use this when possible */
82
83#include <asm/cpufeature.h>
84
85 .section .altinstructions,"a"
86 .align 8
87 .quad copy_page
88 .quad copy_page_c
89 .byte X86_FEATURE_K8_C
90 .byte copy_page_c_end-copy_page_c
91 .byte copy_page_c_end-copy_page_c
92 .previous
93
94 .section .altinstr_replacement,"ax"
95copy_page_c:
96 movl $4096/8,%ecx 11 movl $4096/8,%ecx
97 rep 12 rep
98 movsq 13 movsq
99 ret 14 ret
100copy_page_c_end:
101 .previous
diff --git a/arch/x86_64/lib/memcpy.S b/arch/x86_64/lib/memcpy.S
index c6c46494fef5..92dd80544602 100644
--- a/arch/x86_64/lib/memcpy.S
+++ b/arch/x86_64/lib/memcpy.S
@@ -11,6 +11,8 @@
11 * 11 *
12 * Output: 12 * Output:
13 * rax original destination 13 * rax original destination
14 *
15 * TODO: check best memcpy for PSC
14 */ 16 */
15 17
16 .globl __memcpy 18 .globl __memcpy
@@ -18,95 +20,6 @@
18 .p2align 4 20 .p2align 4
19__memcpy: 21__memcpy:
20memcpy: 22memcpy:
21 pushq %rbx
22 movq %rdi,%rax
23
24 movl %edx,%ecx
25 shrl $6,%ecx
26 jz .Lhandle_tail
27
28 .p2align 4
29.Lloop_64:
30 decl %ecx
31
32 movq (%rsi),%r11
33 movq 8(%rsi),%r8
34
35 movq %r11,(%rdi)
36 movq %r8,1*8(%rdi)
37
38 movq 2*8(%rsi),%r9
39 movq 3*8(%rsi),%r10
40
41 movq %r9,2*8(%rdi)
42 movq %r10,3*8(%rdi)
43
44 movq 4*8(%rsi),%r11
45 movq 5*8(%rsi),%r8
46
47 movq %r11,4*8(%rdi)
48 movq %r8,5*8(%rdi)
49
50 movq 6*8(%rsi),%r9
51 movq 7*8(%rsi),%r10
52
53 movq %r9,6*8(%rdi)
54 movq %r10,7*8(%rdi)
55
56 leaq 64(%rsi),%rsi
57 leaq 64(%rdi),%rdi
58 jnz .Lloop_64
59
60.Lhandle_tail:
61 movl %edx,%ecx
62 andl $63,%ecx
63 shrl $3,%ecx
64 jz .Lhandle_7
65 .p2align 4
66.Lloop_8:
67 decl %ecx
68 movq (%rsi),%r8
69 movq %r8,(%rdi)
70 leaq 8(%rdi),%rdi
71 leaq 8(%rsi),%rsi
72 jnz .Lloop_8
73
74.Lhandle_7:
75 movl %edx,%ecx
76 andl $7,%ecx
77 jz .Lende
78 .p2align 4
79.Lloop_1:
80 movb (%rsi),%r8b
81 movb %r8b,(%rdi)
82 incq %rdi
83 incq %rsi
84 decl %ecx
85 jnz .Lloop_1
86
87.Lende:
88 popq %rbx
89 ret
90.Lfinal:
91
92 /* C stepping K8 run faster using the string copy instructions.
93 It is also a lot simpler. Use this when possible */
94
95 .section .altinstructions,"a"
96 .align 8
97 .quad memcpy
98 .quad memcpy_c
99 .byte X86_FEATURE_K8_C
100 .byte .Lfinal-memcpy
101 .byte memcpy_c_end-memcpy_c
102 .previous
103
104 .section .altinstr_replacement,"ax"
105 /* rdi destination
106 * rsi source
107 * rdx count
108 */
109memcpy_c:
110 movq %rdi,%rax 23 movq %rdi,%rax
111 movl %edx,%ecx 24 movl %edx,%ecx
112 shrl $3,%ecx 25 shrl $3,%ecx
@@ -117,5 +30,3 @@ memcpy_c:
117 rep 30 rep
118 movsb 31 movsb
119 ret 32 ret
120memcpy_c_end:
121 .previous
diff --git a/arch/x86_64/lib/memset.S b/arch/x86_64/lib/memset.S
index 4b4c40638640..2aa48f24ed1e 100644
--- a/arch/x86_64/lib/memset.S
+++ b/arch/x86_64/lib/memset.S
@@ -13,98 +13,6 @@
13 .p2align 4 13 .p2align 4
14memset: 14memset:
15__memset: 15__memset:
16 movq %rdi,%r10
17 movq %rdx,%r11
18
19 /* expand byte value */
20 movzbl %sil,%ecx
21 movabs $0x0101010101010101,%rax
22 mul %rcx /* with rax, clobbers rdx */
23
24 /* align dst */
25 movl %edi,%r9d
26 andl $7,%r9d
27 jnz .Lbad_alignment
28.Lafter_bad_alignment:
29
30 movl %r11d,%ecx
31 shrl $6,%ecx
32 jz .Lhandle_tail
33
34 .p2align 4
35.Lloop_64:
36 decl %ecx
37 movq %rax,(%rdi)
38 movq %rax,8(%rdi)
39 movq %rax,16(%rdi)
40 movq %rax,24(%rdi)
41 movq %rax,32(%rdi)
42 movq %rax,40(%rdi)
43 movq %rax,48(%rdi)
44 movq %rax,56(%rdi)
45 leaq 64(%rdi),%rdi
46 jnz .Lloop_64
47
48 /* Handle tail in loops. The loops should be faster than hard
49 to predict jump tables. */
50 .p2align 4
51.Lhandle_tail:
52 movl %r11d,%ecx
53 andl $63&(~7),%ecx
54 jz .Lhandle_7
55 shrl $3,%ecx
56 .p2align 4
57.Lloop_8:
58 decl %ecx
59 movq %rax,(%rdi)
60 leaq 8(%rdi),%rdi
61 jnz .Lloop_8
62
63.Lhandle_7:
64 movl %r11d,%ecx
65 andl $7,%ecx
66 jz .Lende
67 .p2align 4
68.Lloop_1:
69 decl %ecx
70 movb %al,(%rdi)
71 leaq 1(%rdi),%rdi
72 jnz .Lloop_1
73
74.Lende:
75 movq %r10,%rax
76 ret
77
78.Lbad_alignment:
79 cmpq $7,%r11
80 jbe .Lhandle_7
81 movq %rax,(%rdi) /* unaligned store */
82 movq $8,%r8
83 subq %r9,%r8
84 addq %r8,%rdi
85 subq %r8,%r11
86 jmp .Lafter_bad_alignment
87
88 /* C stepping K8 run faster using the string instructions.
89 It is also a lot simpler. Use this when possible */
90
91#include <asm/cpufeature.h>
92
93 .section .altinstructions,"a"
94 .align 8
95 .quad memset
96 .quad memset_c
97 .byte X86_FEATURE_K8_C
98 .byte memset_c_end-memset_c
99 .byte memset_c_end-memset_c
100 .previous
101
102 .section .altinstr_replacement,"ax"
103 /* rdi destination
104 * rsi value
105 * rdx count
106 */
107memset_c:
108 movq %rdi,%r9 16 movq %rdi,%r9
109 movl %edx,%r8d 17 movl %edx,%r8d
110 andl $7,%r8d 18 andl $7,%r8d
@@ -121,5 +29,3 @@ memset_c:
121 stosb 29 stosb
122 movq %r9,%rax 30 movq %r9,%rax
123 ret 31 ret
124memset_c_end:
125 .previous
generated by cgit v1.2.2 (git 2.25.0) at 2024-12-12 09:41:30 -0500