lib/glibc
1
0
Fork 0
mirror of https://sourceware.org/git/glibc.git synced 2025-07-30 22:43:12 +03:00
Code Activity

Rewritten v9/64-bit sparc strcmp.

Browse Source
This commit is contained in:
David S. Miller
2011-08-24 01:32:24 -07:00
parent 39dd69dfb9
commit ad69cc2652
2 changed files with 136 additions and 210 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ChangeLog
View File

@ -1,3 +1,7 @@
2011-08-24 David S. Miller <davem@davemloft.net>
* sysdeps/sparc/sparc64/strcmp.S: Rewrite.
2011-08-24 Andreas Schwab <schwab@redhat.com> 2011-08-24 Andreas Schwab <schwab@redhat.com>
* elf/Makefile: Add rules to build and run unload8 test. * elf/Makefile: Add rules to build and run unload8 test.

342
sysdeps/sparc/sparc64/strcmp.S
View File

@ -1,9 +1,8 @@
/* Compare two strings for differences. /* Compare two strings for differences.
For SPARC v9. For SPARC v9.
Copyright (C) 1997, 1999, 2003 Free Software Foundation, Inc. Copyright (C) 2011 Free Software Foundation, Inc.
This file is part of the GNU C Library. This file is part of the GNU C Library.
Contributed by Jan Vondrak <jvon4518@ss1000.ms.mff.cuni.cz> and Contributed by David S. Miller <davem@davemloft.net>
Jakub Jelinek <jj@ultra.linux.cz>.
The GNU C Library is free software; you can redistribute it and/or The GNU C Library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public modify it under the terms of the GNU Lesser General Public
@ -22,259 +21,182 @@
#include <sysdep.h> #include <sysdep.h>
#include <asm/asi.h> #include <asm/asi.h>
#ifndef XCC #ifndef XCC
.register %g2, #scratch .register %g2, #scratch
.register %g3, #scratch .register %g3, #scratch
.register %g6, #scratch .register %g6, #scratch
#endif #endif
/* Normally, this uses #define rSTR1 %o0
((xword - 0x0101010101010101) & 0x8080808080808080) test #define rSTR2 %o1
to find out if any byte in xword could be zero. This is fast, but #define r0101 %o2 /* 0x0101010101010101 */
also gives false alarm for any byte in range 0x81-0xff. It does #define r8080 %o3 /* 0x8080808080808080 */
not matter for correctness, as if this test tells us there could #define rSTRXOR %o4
be some zero byte, we check it byte by byte, but if bytes with #define rWORD1 %o5
high bits set are common in the strings, then this will give poor #define rTMP1 %g1
performance. You can #define EIGHTBIT_NOT_RARE and the algorithm #define rTMP2 %g2
will use one tick slower, but more precise test #define rWORD2 %g3
((xword - 0x0101010101010101) & (~xword) & 0x8080808080808080), #define rSLL %g4
which does not give any false alarms (but if some bits are set, #define rSRL %g5
one cannot assume from it which bytes are zero and which are not). #define rBARREL %g6
It is yet to be measured, what is the correct default for glibc
in these days for an average user. /* There are two cases, either the two pointers are aligned
* identically or they are not. If they have the same
* alignment we can use the normal full speed loop. Otherwise
* we have to use the barrel-shifter version.
*/ */
.text .text
.align 32 .align 32
ENTRY(strcmp) ENTRY(strcmp)
sethi %hi(0x01010101), %g1 /* IEU0 Group */ or rSTR2, rSTR1, rTMP1
andcc %o0, 7, %g0 /* IEU1 */ sethi %hi(0x80808080), r8080
bne,pn %icc, 7f /* CTI */
or %g1, %lo(0x01010101), %g1 /* IEU0 Group */
andcc %o1, 7, %g3 /* IEU1 */ andcc rTMP1, 0x7, %g0
bne,pn %icc, 9f /* CTI */ bne,pn %icc, .Lmaybe_barrel_shift
sllx %g1, 32, %g2 /* IEU0 Group */ or r8080, %lo(0x80808080), r8080
ldx [%o0], %o2 /* Load */ ldx [rSTR1], rWORD1
or %g1, %g2, %g1 /* IEU0 Group */ sub rSTR2, rSTR1, rSTR2
1: ldx [%o1], %o3 /* Load */ sllx r8080, 32, rTMP1
sub %o1, %o0, %o1 /* IEU1 */
sllx %g1, 7, %g2 /* IEU0 Group */
2: add %o0, 8, %o0 /* IEU1 */ ldx [rSTR1 + rSTR2], rWORD2
sub %o2, %g1, %g3 /* IEU0 Group */ or r8080, rTMP1, r8080
subcc %o2, %o3, %g0 /* IEU1 */
bne,pn %xcc, 13f /* CTI */
#ifdef EIGHTBIT_NOT_RARE ba,pt %xcc, .Laligned_loop_entry
andn %g3, %o2, %g4 /* IEU0 Group */ srlx r8080, 7, r0101
ldxa [%o0] ASI_PNF, %o2 /* Load */
andcc %g4, %g2, %g0 /* IEU1 Group */
#else
ldxa [%o0] ASI_PNF, %o2 /* Load Group */
andcc %g3, %g2, %g0 /* IEU1 */
#endif
be,a,pt %xcc, 2b /* CTI */
ldxa [%o1 + %o0] ASI_PNF, %o3 /* Load Group */
addcc %g3, %g1, %o4 /* IEU1 */ .align 32
srlx %g3, 32, %g3 /* IEU0 */ .Laligned_loop_entry:
andcc %g3, %g2, %g0 /* IEU1 Group */ .Laligned_loop:
be,pt %xcc, 3f /* CTI */ add rSTR1, 8, rSTR1
srlx %o4, 56, %o5 /* IEU0 */ sub rWORD1, r0101, rTMP2
andcc %o5, 0xff, %g0 /* IEU1 Group */ xorcc rWORD1, rWORD2, rSTRXOR
be,pn %icc, 4f /* CTI */ bne,pn %xcc, .Lcommon_endstring
srlx %o4, 48, %o5 /* IEU0 */
andcc %o5, 0xff, %g0 /* IEU1 Group */ andn r8080, rWORD1, rTMP1
be,pn %icc, 4f /* CTI */
srlx %o4, 40, %o5 /* IEU0 */
andcc %o5, 0xff, %g0 /* IEU1 Group */
be,pn %icc, 4f /* CTI */ ldxa [rSTR1] ASI_PNF, rWORD1
srlx %o4, 32, %o5 /* IEU0 */ andcc rTMP1, rTMP2, %g0
andcc %o5, 0xff, %g0 /* IEU1 Group */ be,a,pt %xcc, .Laligned_loop
be,pn %icc, 4f /* CTI */
3: srlx %o4, 24, %o5 /* IEU0 */ ldxa [rSTR1 + rSTR2] ASI_PNF, rWORD2
andcc %o5, 0xff, %g0 /* IEU1 Group */
be,pn %icc, 4f /* CTI */
srlx %o4, 16, %o5 /* IEU0 */
andcc %o5, 0xff, %g0 /* IEU1 Group */ .Lcommon_equal:
be,pn %icc, 4f /* CTI */ retl
srlx %o4, 8, %o5 /* IEU0 */ mov 0, %o0
andcc %o5, 0xff, %g0 /* IEU1 Group */
be,pn %icc, 4f /* CTI */ /* All loops terminate here once they find an unequal word.
andcc %o4, 0xff, %g0 /* IEU1 Group */ * If a zero byte appears in the word before the first unequal
bne,a,pn %icc, 2b /* CTI */ * byte, we must report zero. Otherwise we report '1' or '-1'
ldxa [%o1 + %o0] ASI_PNF, %o3 /* Load */ * depending upon whether the first mis-matching byte is larger
* in the first string or the second, respectively.
*
* First we compute a 64-bit mask value that has "0x01" in
* each byte where a zero exists in rWORD1. rSTRXOR holds the
* value (rWORD1 ^ rWORD2). Therefore, if considered as an
* unsigned quantity, our "0x01" mask value is "greater than"
* rSTRXOR then a zero terminating byte comes first and
* therefore we report '0'.
*
* The formula for this mask is:
*
* mask_tmp1 = ~rWORD1 & 0x8080808080808080;
* mask_tmp2 = ((rWORD1 & 0x7f7f7f7f7f7f7f7f) +
* 0x7f7f7f7f7f7f7f7f);
*
* mask = ((mask_tmp1 & ~mask_tmp2) >> 7);
*/
.Lcommon_endstring:
andn rWORD1, r8080, rTMP2
or r8080, 1, %o1
4: retl /* CTI+IEU1 Group */ mov 1, %o0
clr %o0 /* IEU0 */ sub rTMP2, %o1, rTMP2
.align 32 cmp rWORD1, rWORD2
13: mov 0xff, %g6 /* IEU0 Group */ andn rTMP1, rTMP2, rTMP1
#ifdef EIGHTBIT_NOT_RARE
andcc %g4, %g2, %g0 /* IEU1 */
#else
andcc %g3, %g2, %g0 /* IEU1 */
#endif
be,pt %xcc, 25f /* CTI */
addcc %g3, %g1, %o4 /* IEU1 Group */
srlx %g3, 32, %g3 /* IEU0 */ movleu %xcc, -1, %o0
andcc %g3, %g2, %g0 /* IEU1 Group */ srlx rTMP1, 7, rTMP1
be,pt %xcc, 23f /* CTI */
sllx %g6, 56, %o5 /* IEU0 */
andcc %o4, %o5, %g0 /* IEU1 Group */ cmp rTMP1, rSTRXOR
be,pn %xcc, 24f /* CTI */ retl
sllx %g6, 48, %o5 /* IEU0 */ movgu %xcc, 0, %o0
andcc %o4, %o5, %g0 /* IEU1 Group */
be,pn %xcc, 24f /* CTI */ .Lmaybe_barrel_shift:
sllx %g6, 40, %o5 /* IEU0 */ sub rSTR2, rSTR1, rSTR2
andcc %o4, %o5, %g0 /* IEU1 Group */ sllx r8080, 32, rTMP1
be,pn %xcc, 24f /* CTI */
sllx %g6, 32, %o5 /* IEU0 */ or r8080, rTMP1, r8080
andcc %o4, %o5, %g0 /* IEU1 Group */ and rSTR1, 0x7, rTMP2
be,pn %xcc, 24f /* CTI */
23: sllx %g6, 24, %o5 /* IEU0 */
andcc %o4, %o5, %g0 /* IEU1 Group */ srlx r8080, 7, r0101
be,pn %icc, 24f /* CTI */ andn rSTR1, 0x7, rSTR1
sllx %g6, 16, %o5 /* IEU0 */
andcc %o4, %o5, %g0 /* IEU1 Group */
be,pn %icc, 24f /* CTI */ ldxa [rSTR1] ASI_PNF, rWORD1
sllx %g6, 8, %o5 /* IEU0 */ andcc rSTR2, 0x7, rSLL
andcc %o4, %o5, %g0 /* IEU1 Group */ sll rTMP2, 3, rSTRXOR
be,pn %icc, 24f /* CTI */
mov %g6, %o5 /* IEU0 */ bne,pn %icc, .Lneed_barrel_shift
25: cmp %o4, %o3 /* IEU1 Group */ mov -1, rTMP1
5: mov -1, %o0 /* IEU0 */ ldxa [rSTR1 + rSTR2] ASI_PNF, rBARREL
retl /* CTI+IEU1 Group */
movgu %xcc, 1, %o0 /* Single Group */ srlx rTMP1, rSTRXOR, rTMP2
.align 16 orn rWORD1, rTMP2, rWORD1
24: sub %o5, 1, %g6 /* IEU0 Group */ ba,pt %xcc, .Laligned_loop_entry
clr %o0 /* IEU1 */ orn rBARREL, rTMP2, rWORD2
or %o5, %g6, %o5 /* IEU0 Group */
andn %o4, %o5, %o4 /* IEU0 Group */
andn %o3, %o5, %o3 /* IEU1 */ .Lneed_barrel_shift:
cmp %o4, %o3 /* IEU1 Group */ sllx rSLL, 3, rSLL
movgu %xcc, 1, %o0 /* Single Group */ andn rSTR2, 0x7, rSTR2
retl /* CTI+IEU1 Group */
movlu %xcc, -1, %o0 /* Single Group */ ldxa [rSTR1 + rSTR2] ASI_PNF, rBARREL
6: retl /* CTI+IEU1 Group */ mov 64, rTMP2
mov %o4, %o0 /* IEU0 */ sub rTMP2, rSLL, rSRL
.align 16 srlx rTMP1, rSTRXOR, rTMP1
7: ldub [%o0], %o2 /* Load */ add rSTR2, 8, rSTR2
add %o0, 1, %o0 /* IEU1 */
ldub [%o1], %o3 /* Load Group */
sllx %g1, 32, %g2 /* IEU0 */
8: add %o1, 1, %o1 /* IEU1 */ orn rWORD1, rTMP1, rWORD1
subcc %o2, %o3, %o4 /* IEU1 Group */ sllx rBARREL, rSLL, rWORD2
bne,pn %xcc, 6b /* CTI */ ldxa [rSTR1 + rSTR2] ASI_PNF, rBARREL
lduba [%o0] ASI_PNF, %o2 /* Load */
brz,pn %o3, 4b /* CTI+IEU1 Group */ add rSTR1, 8, rSTR1
lduba [%o1] ASI_PNF, %o3 /* Load */ sub rWORD1, r0101, rTMP2
andcc %o0, 7, %g0 /* IEU1 Group */
bne,a,pn %icc, 8b /* CTI */
add %o0, 1, %o0 /* IEU0 */ srlx rBARREL, rSRL, rSTRXOR
or %g1, %g2, %g1 /* IEU0 Group */
andcc %o1, 7, %g3 /* IEU1 */
be,a,pn %icc, 1b /* CTI */
ldxa [%o0] ASI_PNF, %o2 /* Load Group */ or rWORD2, rSTRXOR, rWORD2
9: sllx %g3, 3, %g5 /* IEU0 */
mov 64, %o5 /* IEU1 */
sub %o1, %g3, %o1 /* IEU0 Group */
sub %o5, %g5, %o5 /* IEU1 */ orn rWORD2, rTMP1, rWORD2
ldxa [%o1] ASI_PNF, %g6 /* Load Group */ ba,pt %xcc, .Lbarrel_shift_loop_entry
or %g1, %g2, %g1 /* IEU0 */ andn r8080, rWORD1, rTMP1
sub %o1, %o0, %o1 /* IEU1 */
sllx %g1, 7, %g2 /* IEU0 Group */ .Lbarrel_shift_loop:
add %o1, 8, %o1 /* IEU1 */ sllx rBARREL, rSLL, rWORD2
/* %g1 = 0101010101010101 ldxa [rSTR1 + rSTR2] ASI_PNF, rBARREL
* %g2 = 8080808080800880
* %g5 = number of bits to shift left
* %o5 = number of bits to shift right */
10: sllx %g6, %g5, %o3 /* IEU0 Group */
ldxa [%o1 + %o0] ASI_PNF, %g6 /* Load */
11: srlx %g6, %o5, %o4 /* IEU0 Group */ add rSTR1, 8, rSTR1
ldxa [%o0] ASI_PNF, %o2 /* Load */ sub rWORD1, r0101, rTMP2
or %o3, %o4, %o3 /* IEU1 */
add %o0, 8, %o0 /* IEU0 Group */
subcc %o2, %o3, %g0 /* IEU1 */ srlx rBARREL, rSRL, rSTRXOR
#ifdef EIGHTBIT_NOT_RARE andn r8080, rWORD1, rTMP1
sub %o2, %g1, %g3 /* IEU0 Group */
bne,pn %xcc, 13b /* CTI */
andn %g3, %o2, %g4 /* IEU0 Group */
andcc %g4, %g2, %g0 /* IEU1 Group */ or rWORD2, rSTRXOR, rWORD2
be,pt %xcc, 10b /* CTI */
srlx %g4, 32, %g4 /* IEU0 */
andcc %g4, %g2, %g0 /* IEU1 Group */
#else
bne,pn %xcc, 13b /* CTI */
sub %o2, %g1, %g3 /* IEU0 Group */
andcc %g3, %g2, %g0 /* IEU1 Group */
be,pt %xcc, 10b /* CTI */ .Lbarrel_shift_loop_entry:
srlx %g3, 32, %g3 /* IEU0 */ xorcc rWORD1, rWORD2, rSTRXOR
andcc %g3, %g2, %g0 /* IEU1 Group */ bne,pn %xcc, .Lcommon_endstring
#endif
be,pt %xcc, 12f /* CTI */
srlx %o2, 56, %g3 /* IEU0 */ andcc rTMP1, rTMP2, %g0
andcc %g3, 0xff, %g0 /* IEU1 Group */ be,a,pt %xcc, .Lbarrel_shift_loop
be,pn %icc, 4b /* CTI */ ldxa [rSTR1] ASI_PNF, rWORD1
srlx %o2, 48, %g3 /* IEU0 */
andcc %g3, 0xff, %g0 /* IEU1 Group */ retl
be,pn %icc, 4b /* CTI */ mov 0, %o0
srlx %o2, 40, %g3 /* IEU0 */
andcc %g3, 0xff, %g0 /* IEU1 Group */
be,pn %icc, 4b /* CTI */
srlx %o2, 32, %g3 /* IEU0 */
andcc %g3, 0xff, %g0 /* IEU1 Group */
be,pn %icc, 4b /* CTI */
12: srlx %o2, 24, %g3 /* IEU0 */
andcc %g3, 0xff, %g0 /* IEU1 Group */
be,pn %icc, 4b /* CTI */
srlx %o2, 16, %g3 /* IEU0 */
andcc %g3, 0xff, %g0 /* IEU1 Group */
be,pn %icc, 4b /* CTI */
srlx %o2, 8, %g3 /* IEU0 */
andcc %g3, 0xff, %g0 /* IEU1 Group */
be,pn %icc, 4b /* CTI */
andcc %o2, 0xff, %g0 /* IEU1 Group */
be,pn %icc, 4b /* CTI */
sllx %g6, %g5, %o3 /* IEU0 */
ba,pt %xcc, 11b /* CTI Group */
ldxa [%o1 + %o0] ASI_PNF, %g6 /* Load */
END(strcmp) END(strcmp)
libc_hidden_builtin_def (strcmp) libc_hidden_builtin_def (strcmp)