You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 
 
 

2943 lines
66 KiB

# commit 759cfef3ac4c07dba1ece0bbc1207e099348816d
# Author: Alan Modra <amodra@gmail.com>
# Date: Sat Aug 17 18:47:22 2013 +0930
#
# PowerPC LE memcpy
# http://sourceware.org/ml/libc-alpha/2013-08/msg00103.html
#
# LIttle-endian support for memcpy. I spent some time cleaning up the
# 64-bit power7 memcpy, in order to avoid the extra alignment traps
# power7 takes for little-endian. It probably would have been better
# to copy the linux kernel version of memcpy.
#
# * sysdeps/powerpc/powerpc32/power4/memcpy.S: Add little endian support.
# * sysdeps/powerpc/powerpc32/power6/memcpy.S: Likewise.
# * sysdeps/powerpc/powerpc32/power7/memcpy.S: Likewise.
# * sysdeps/powerpc/powerpc32/power7/mempcpy.S: Likewise.
# * sysdeps/powerpc/powerpc64/memcpy.S: Likewise.
# * sysdeps/powerpc/powerpc64/power4/memcpy.S: Likewise.
# * sysdeps/powerpc/powerpc64/power6/memcpy.S: Likewise.
# * sysdeps/powerpc/powerpc64/power7/memcpy.S: Likewise.
# * sysdeps/powerpc/powerpc64/power7/mempcpy.S: Likewise. Make better
# use of regs. Use power7 mtocrf. Tidy function tails.
#
diff -urN glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power4/memcpy.S glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power4/memcpy.S
--- glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power4/memcpy.S 2014-05-29 13:04:56.000000000 -0500
+++ glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power4/memcpy.S 2014-05-29 13:04:56.000000000 -0500
@@ -205,15 +205,28 @@
blt cr6,5f
srwi 7,6,16
bgt cr6,3f
+#ifdef __LITTLE_ENDIAN__
+ sth 7,0(3)
+#else
sth 6,0(3)
+#endif
b 7f
.align 4
3:
+#ifdef __LITTLE_ENDIAN__
+ rotlwi 6,6,24
+ stb 6,0(3)
+ sth 7,1(3)
+#else
stb 7,0(3)
sth 6,1(3)
+#endif
b 7f
.align 4
5:
+#ifdef __LITTLE_ENDIAN__
+ rotlwi 6,6,8
+#endif
stb 6,0(3)
7:
cmplwi cr1,10,16
@@ -341,13 +354,23 @@
bf 30,1f
/* there are at least two words to copy, so copy them */
+#ifdef __LITTLE_ENDIAN__
+ srw 0,6,10
+ slw 8,7,9
+#else
slw 0,6,10 /* shift 1st src word to left align it in R0 */
srw 8,7,9 /* shift 2nd src word to right align it in R8 */
+#endif
or 0,0,8 /* or them to get word to store */
lwz 6,8(5) /* load the 3rd src word */
stw 0,0(4) /* store the 1st dst word */
+#ifdef __LITTLE_ENDIAN__
+ srw 0,7,10
+ slw 8,6,9
+#else
slw 0,7,10 /* now left align 2nd src word into R0 */
srw 8,6,9 /* shift 3rd src word to right align it in R8 */
+#endif
or 0,0,8 /* or them to get word to store */
lwz 7,12(5)
stw 0,4(4) /* store the 2nd dst word */
@@ -355,8 +378,13 @@
addi 5,5,16
bf 31,4f
/* there is a third word to copy, so copy it */
+#ifdef __LITTLE_ENDIAN__
+ srw 0,6,10
+ slw 8,7,9
+#else
slw 0,6,10 /* shift 3rd src word to left align it in R0 */
srw 8,7,9 /* shift 4th src word to right align it in R8 */
+#endif
or 0,0,8 /* or them to get word to store */
stw 0,0(4) /* store 3rd dst word */
mr 6,7
@@ -366,8 +394,13 @@
b 4f
.align 4
1:
+#ifdef __LITTLE_ENDIAN__
+ srw 0,6,10
+ slw 8,7,9
+#else
slw 0,6,10 /* shift 1st src word to left align it in R0 */
srw 8,7,9 /* shift 2nd src word to right align it in R8 */
+#endif
addi 5,5,8
or 0,0,8 /* or them to get word to store */
bf 31,4f
@@ -380,23 +413,43 @@
.align 4
4:
/* copy 16 bytes at a time */
+#ifdef __LITTLE_ENDIAN__
+ srw 0,6,10
+ slw 8,7,9
+#else
slw 0,6,10
srw 8,7,9
+#endif
or 0,0,8
lwz 6,0(5)
stw 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srw 0,7,10
+ slw 8,6,9
+#else
slw 0,7,10
srw 8,6,9
+#endif
or 0,0,8
lwz 7,4(5)
stw 0,4(4)
+#ifdef __LITTLE_ENDIAN__
+ srw 0,6,10
+ slw 8,7,9
+#else
slw 0,6,10
srw 8,7,9
+#endif
or 0,0,8
lwz 6,8(5)
stw 0,8(4)
+#ifdef __LITTLE_ENDIAN__
+ srw 0,7,10
+ slw 8,6,9
+#else
slw 0,7,10
srw 8,6,9
+#endif
or 0,0,8
lwz 7,12(5)
stw 0,12(4)
@@ -405,8 +458,13 @@
bdnz+ 4b
8:
/* calculate and store the final word */
+#ifdef __LITTLE_ENDIAN__
+ srw 0,6,10
+ slw 8,7,9
+#else
slw 0,6,10
srw 8,7,9
+#endif
or 0,0,8
stw 0,0(4)
3:
diff -urN glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power6/memcpy.S glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power6/memcpy.S
--- glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power6/memcpy.S 2014-05-29 13:04:56.000000000 -0500
+++ glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power6/memcpy.S 2014-05-29 13:04:56.000000000 -0500
@@ -221,15 +221,28 @@
blt cr6,5f
srwi 7,6,16
bgt cr6,3f
+#ifdef __LITTLE_ENDIAN__
+ sth 7,0(3)
+#else
sth 6,0(3)
+#endif
b 7f
.align 4
3:
+#ifdef __LITTLE_ENDIAN__
+ rotlwi 6,6,24
+ stb 6,0(3)
+ sth 7,1(3)
+#else
stb 7,0(3)
sth 6,1(3)
+#endif
b 7f
.align 4
5:
+#ifdef __LITTLE_ENDIAN__
+ rotlwi 6,6,8
+#endif
stb 6,0(3)
7:
cmplwi cr1,10,16
@@ -579,7 +592,11 @@
lwz 6,-1(4)
cmplwi cr6,31,4
srwi 8,31,5 /* calculate the 32 byte loop count */
+#ifdef __LITTLE_ENDIAN__
+ srwi 6,6,8
+#else
slwi 6,6,8
+#endif
clrlwi 31,31,27 /* The remaining bytes, < 32. */
blt cr5,L(wdu1_32tail)
mtctr 8
@@ -587,8 +604,12 @@
lwz 8,3(4)
lwz 7,4(4)
+#ifdef __LITTLE_ENDIAN__
+ rldimi 6,8,24,32
+#else
/* Equivalent to: srwi 8,8,32-8; or 6,6,8 */
rlwimi 6,8,8,(32-8),31
+#endif
b L(wdu1_loop32x)
.align 4
L(wdu1_loop32):
@@ -597,8 +618,12 @@
lwz 7,4(4)
stw 10,-8(3)
stw 11,-4(3)
+#ifdef __LITTLE_ENDIAN__
+ rldimi 6,8,24,32
+#else
/* Equivalent to srwi 8,8,32-8; or 6,6,8 */
rlwimi 6,8,8,(32-8),31
+#endif
L(wdu1_loop32x):
lwz 10,8(4)
lwz 11,12(4)
@@ -615,7 +640,11 @@
stw 6,16(3)
stw 7,20(3)
addi 3,3,32
+#ifdef __LITTLE_ENDIAN__
+ srwi 6,8,8
+#else
slwi 6,8,8
+#endif
bdnz+ L(wdu1_loop32)
stw 10,-8(3)
stw 11,-4(3)
@@ -626,8 +655,12 @@
blt cr6,L(wdu_4tail)
/* calculate and store the final word */
lwz 8,3(4)
-/* Equivalent to: srwi 8,8,32-9; or 6,6,8 */
+#ifdef __LITTLE_ENDIAN__
+ rldimi 6,8,24,32
+#else
+/* Equivalent to: srwi 8,8,32-8; or 6,6,8 */
rlwimi 6,8,8,(32-8),31
+#endif
b L(wdu_32tailx)
L(wdu2_32):
@@ -635,7 +668,11 @@
lwz 6,-2(4)
cmplwi cr6,31,4
srwi 8,31,5 /* calculate the 32 byte loop count */
+#ifdef __LITTLE_ENDIAN__
+ srwi 6,6,16
+#else
slwi 6,6,16
+#endif
clrlwi 31,31,27 /* The remaining bytes, < 32. */
blt cr5,L(wdu2_32tail)
mtctr 8
@@ -643,8 +680,11 @@
lwz 8,2(4)
lwz 7,4(4)
-/* Equivalent to: srwi 8,8,32-8; or 6,6,8 */
+#ifdef __LITTLE_ENDIAN__
+ rldimi 6,8,16,32
+#else
rlwimi 6,8,16,(32-16),31
+#endif
b L(wdu2_loop32x)
.align 4
L(wdu2_loop32):
@@ -653,8 +693,11 @@
lwz 7,4(4)
stw 10,-8(3)
stw 11,-4(3)
-/* Equivalent to srwi 8,8,32-8; or 6,6,8 */
+#ifdef __LITTLE_ENDIAN__
+ rldimi 6,8,16,32
+#else
rlwimi 6,8,16,(32-16),31
+#endif
L(wdu2_loop32x):
lwz 10,8(4)
lwz 11,12(4)
@@ -672,7 +715,11 @@
stw 6,16(3)
stw 7,20(3)
addi 3,3,32
+#ifdef __LITTLE_ENDIAN__
+ srwi 6,8,16
+#else
slwi 6,8,16
+#endif
bdnz+ L(wdu2_loop32)
stw 10,-8(3)
stw 11,-4(3)
@@ -683,8 +730,11 @@
blt cr6,L(wdu_4tail)
/* calculate and store the final word */
lwz 8,2(4)
-/* Equivalent to: srwi 8,8,32-9; or 6,6,8 */
+#ifdef __LITTLE_ENDIAN__
+ rldimi 6,8,16,32
+#else
rlwimi 6,8,16,(32-16),31
+#endif
b L(wdu_32tailx)
L(wdu3_32):
@@ -692,7 +742,11 @@
lwz 6,-3(4)
cmplwi cr6,31,4
srwi 8,31,5 /* calculate the 32 byte loop count */
+#ifdef __LITTLE_ENDIAN__
+ srwi 6,6,24
+#else
slwi 6,6,24
+#endif
clrlwi 31,31,27 /* The remaining bytes, < 32. */
blt cr5,L(wdu3_32tail)
mtctr 8
@@ -700,8 +754,11 @@
lwz 8,1(4)
lwz 7,4(4)
-/* Equivalent to: srwi 8,8,32-8; or 6,6,8 */
+#ifdef __LITTLE_ENDIAN__
+ rldimi 6,8,8,32
+#else
rlwimi 6,8,24,(32-24),31
+#endif
b L(wdu3_loop32x)
.align 4
L(wdu3_loop32):
@@ -710,8 +767,11 @@
lwz 7,4(4)
stw 10,-8(3)
stw 11,-4(3)
-/* Equivalent to srwi 8,8,32-8; or 6,6,8 */
+#ifdef __LITTLE_ENDIAN__
+ rldimi 6,8,8,32
+#else
rlwimi 6,8,24,(32-24),31
+#endif
L(wdu3_loop32x):
lwz 10,8(4)
lwz 11,12(4)
@@ -728,7 +788,11 @@
stw 6,16(3)
stw 7,20(3)
addi 3,3,32
+#ifdef __LITTLE_ENDIAN__
+ srwi 6,8,24
+#else
slwi 6,8,24
+#endif
bdnz+ L(wdu3_loop32)
stw 10,-8(3)
stw 11,-4(3)
@@ -739,8 +803,11 @@
blt cr6,L(wdu_4tail)
/* calculate and store the final word */
lwz 8,1(4)
-/* Equivalent to: srwi 8,8,32-9; or 6,6,8 */
+#ifdef __LITTLE_ENDIAN__
+ rldimi 6,8,8,32
+#else
rlwimi 6,8,24,(32-24),31
+#endif
b L(wdu_32tailx)
.align 4
L(wdu_32tailx):
diff -urN glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power7/memcpy.S glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power7/memcpy.S
--- glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power7/memcpy.S 2014-05-29 13:04:56.000000000 -0500
+++ glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power7/memcpy.S 2014-05-29 13:04:56.000000000 -0500
@@ -385,7 +385,7 @@
beq L(copy_GE_32_unaligned_cont)
- /* SRC is not quadword aligned, get it aligned. */
+ /* DST is not quadword aligned, get it aligned. */
mtcrf 0x01,0
subf 31,0,5
@@ -437,13 +437,21 @@
mr 11,12
mtcrf 0x01,9
cmplwi cr6,9,1
+#ifdef __LITTLE_ENDIAN__
+ lvsr 5,0,12
+#else
lvsl 5,0,12
+#endif
lvx 3,0,12
bf 31,L(setup_unaligned_loop)
/* Copy another 16 bytes to align to 32-bytes due to the loop . */
lvx 4,12,6
+#ifdef __LITTLE_ENDIAN__
+ vperm 6,4,3,5
+#else
vperm 6,3,4,5
+#endif
addi 11,12,16
addi 10,3,16
stvx 6,0,3
@@ -463,11 +471,17 @@
vector instructions though. */
lvx 4,11,6 /* vr4 = r11+16. */
- vperm 6,3,4,5 /* Merge the correctly-aligned portions
- of vr3/vr4 into vr6. */
+#ifdef __LITTLE_ENDIAN__
+ vperm 6,4,3,5
+#else
+ vperm 6,3,4,5
+#endif
lvx 3,11,7 /* vr3 = r11+32. */
- vperm 10,4,3,5 /* Merge the correctly-aligned portions
- of vr3/vr4 into vr10. */
+#ifdef __LITTLE_ENDIAN__
+ vperm 10,3,4,5
+#else
+ vperm 10,4,3,5
+#endif
addi 11,11,32
stvx 6,0,10
stvx 10,10,6
diff -urN glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power7/mempcpy.S glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power7/mempcpy.S
--- glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power7/mempcpy.S 2014-05-29 13:04:56.000000000 -0500
+++ glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power7/mempcpy.S 2014-05-29 13:04:56.000000000 -0500
@@ -327,7 +327,7 @@
beq L(copy_GE_32_unaligned_cont)
- /* SRC is not quadword aligned, get it aligned. */
+ /* DST is not quadword aligned, get it aligned. */
mtcrf 0x01,0
subf 31,0,5
@@ -379,13 +379,21 @@
mr 11,12
mtcrf 0x01,9
cmplwi cr6,9,1
- lvsl 5,0,12
+#ifdef __LITTLE_ENDIAN__
+ lvsr 5,0,12
+#else
+ lvsl 5,0,12
+#endif
lvx 3,0,12
bf 31,L(setup_unaligned_loop)
/* Copy another 16 bytes to align to 32-bytes due to the loop . */
lvx 4,12,6
- vperm 6,3,4,5
+#ifdef __LITTLE_ENDIAN__
+ vperm 6,4,3,5
+#else
+ vperm 6,3,4,5
+#endif
addi 11,12,16
addi 10,3,16
stvx 6,0,3
@@ -405,11 +413,17 @@
vector instructions though. */
lvx 4,11,6 /* vr4 = r11+16. */
- vperm 6,3,4,5 /* Merge the correctly-aligned portions
- of vr3/vr4 into vr6. */
+#ifdef __LITTLE_ENDIAN__
+ vperm 6,4,3,5
+#else
+ vperm 6,3,4,5
+#endif
lvx 3,11,7 /* vr3 = r11+32. */
- vperm 10,4,3,5 /* Merge the correctly-aligned portions
- of vr3/vr4 into vr10. */
+#ifdef __LITTLE_ENDIAN__
+ vperm 10,3,4,5
+#else
+ vperm 10,4,3,5
+#endif
addi 11,11,32
stvx 6,0,10
stvx 10,10,6
diff -urN glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/memcpy.S glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/memcpy.S
--- glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/memcpy.S 2014-05-29 13:04:56.000000000 -0500
+++ glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/memcpy.S 2014-05-29 13:04:56.000000000 -0500
@@ -214,15 +214,28 @@
blt cr6,5f
srdi 7,6,16
bgt cr6,3f
+#ifdef __LITTLE_ENDIAN__
+ sth 7,0(3)
+#else
sth 6,0(3)
+#endif
b 7f
.align 4
3:
+#ifdef __LITTLE_ENDIAN__
+ rotlwi 6,6,24
+ stb 6,0(3)
+ sth 7,1(3)
+#else
stb 7,0(3)
sth 6,1(3)
+#endif
b 7f
.align 4
5:
+#ifdef __LITTLE_ENDIAN__
+ rotlwi 6,6,8
+#endif
stb 6,0(3)
7:
cmpldi cr1,10,16
@@ -330,7 +343,11 @@
ld 7,8(5)
subfic 9,10,64
beq 2f
+#ifdef __LITTLE_ENDIAN__
+ srd 0,6,10
+#else
sld 0,6,10
+#endif
cmpldi 11,1
mr 6,7
addi 4,4,-8
@@ -338,15 +355,25 @@
b 1f
2: addi 5,5,8
.align 4
+#ifdef __LITTLE_ENDIAN__
+0: srd 0,6,10
+ sld 8,7,9
+#else
0: sld 0,6,10
srd 8,7,9
+#endif
cmpldi 11,2
ld 6,8(5)
or 0,0,8
addi 11,11,-2
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srd 0,7,10
+1: sld 8,6,9
+#else
sld 0,7,10
1: srd 8,6,9
+#endif
or 0,0,8
beq 8f
ld 7,16(5)
diff -urN glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power4/memcpy.S glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power4/memcpy.S
--- glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power4/memcpy.S 2014-05-29 13:04:56.000000000 -0500
+++ glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power4/memcpy.S 2014-05-29 13:05:51.000000000 -0500
@@ -1,5 +1,5 @@
/* Optimized memcpy implementation for PowerPC64.
- Copyright (C) 2003, 2006, 2011 Free Software Foundation, Inc.
+ Copyright (C) 2003-2014 Free Software Foundation, Inc.
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
@@ -17,26 +17,24 @@
<http://www.gnu.org/licenses/>. */
#include <sysdep.h>
-#include <bp-sym.h>
-#include <bp-asm.h>
/* __ptr_t [r3] memcpy (__ptr_t dst [r3], __ptr_t src [r4], size_t len [r5]);
Returns 'dst'.
- Memcpy handles short copies (< 32-bytes) using a binary move blocks
- (no loops) of lwz/stw. The tail (remaining 1-3) bytes is handled
- with the appropriate combination of byte and halfword load/stores.
- There is minimal effort to optimize the alignment of short moves.
+ Memcpy handles short copies (< 32-bytes) using a binary move blocks
+ (no loops) of lwz/stw. The tail (remaining 1-3) bytes is handled
+ with the appropriate combination of byte and halfword load/stores.
+ There is minimal effort to optimize the alignment of short moves.
The 64-bit implementations of POWER3 and POWER4 do a reasonable job
- of handling unligned load/stores that do not cross 32-byte boundries.
+ of handling unaligned load/stores that do not cross 32-byte boundaries.
Longer moves (>= 32-bytes) justify the effort to get at least the
destination doubleword (8-byte) aligned. Further optimization is
- posible when both source and destination are doubleword aligned.
+ possible when both source and destination are doubleword aligned.
Each case has a optimized unrolled loop. */
.machine power4
-EALIGN (BP_SYM (memcpy), 5, 0)
+EALIGN (memcpy, 5, 0)
CALL_MCOUNT 3
cmpldi cr1,5,31
@@ -44,20 +42,20 @@
std 3,-16(1)
std 31,-8(1)
cfi_offset(31,-8)
- andi. 11,3,7 /* check alignement of dst. */
+ andi. 11,3,7 /* check alignment of dst. */
clrldi 0,0,61 /* Number of bytes until the 1st doubleword of dst. */
- clrldi 10,4,61 /* check alignement of src. */
+ clrldi 10,4,61 /* check alignment of src. */
cmpldi cr6,5,8
ble- cr1,.L2 /* If move < 32 bytes use short move code. */
- cmpld cr6,10,11
+ cmpld cr6,10,11
mr 12,4
srdi 9,5,3 /* Number of full double words remaining. */
mtcrf 0x01,0
mr 31,5
beq .L0
-
+
subf 31,0,5
- /* Move 0-7 bytes as needed to get the destination doubleword alligned. */
+ /* Move 0-7 bytes as needed to get the destination doubleword aligned. */
1: bf 31,2f
lbz 6,0(12)
addi 12,12,1
@@ -74,17 +72,17 @@
stw 6,0(3)
addi 3,3,4
0:
- clrldi 10,12,61 /* check alignement of src again. */
+ clrldi 10,12,61 /* check alignment of src again. */
srdi 9,31,3 /* Number of full double words remaining. */
-
- /* Copy doublewords from source to destination, assumpting the
+
+ /* Copy doublewords from source to destination, assuming the
destination is aligned on a doubleword boundary.
At this point we know there are at least 25 bytes left (32-7) to copy.
- The next step is to determine if the source is also doubleword aligned.
+ The next step is to determine if the source is also doubleword aligned.
If not branch to the unaligned move code at .L6. which uses
a load, shift, store strategy.
-
+
Otherwise source and destination are doubleword aligned, and we can
the optimized doubleword copy loop. */
.L0:
@@ -97,14 +95,14 @@
Use a unrolled loop to copy 4 doubleword (32-bytes) per iteration.
If the copy is not an exact multiple of 32 bytes, 1-3
doublewords are copied as needed to set up the main loop. After
- the main loop exits there may be a tail of 1-7 bytes. These byte are
+ the main loop exits there may be a tail of 1-7 bytes. These byte are
copied a word/halfword/byte at a time as needed to preserve alignment. */
srdi 8,31,5
cmpldi cr1,9,4
cmpldi cr6,11,0
mr 11,12
-
+
bf 30,1f
ld 6,0(12)
ld 7,8(12)
@@ -115,7 +113,7 @@
addi 10,3,16
bf 31,4f
ld 0,16(12)
- std 0,16(3)
+ std 0,16(3)
blt cr1,3f
addi 11,12,24
addi 10,3,24
@@ -129,7 +127,7 @@
addi 11,12,8
std 6,0(3)
addi 10,3,8
-
+
.align 4
4:
ld 6,0(11)
@@ -144,7 +142,7 @@
std 0,24(10)
addi 10,10,32
bdnz 4b
-3:
+3:
rldicr 0,31,0,60
mtcrf 0x01,31
@@ -152,9 +150,9 @@
.L9:
add 3,3,0
add 12,12,0
-
+
/* At this point we have a tail of 0-7 bytes and we know that the
- destiniation is double word aligned. */
+ destination is double word aligned. */
4: bf 29,2f
lwz 6,0(12)
addi 12,12,4
@@ -173,29 +171,29 @@
ld 31,-8(1)
ld 3,-16(1)
blr
-
-/* Copy up to 31 bytes. This divided into two cases 0-8 bytes and 9-31
- bytes. Each case is handled without loops, using binary (1,2,4,8)
- tests.
-
+
+/* Copy up to 31 bytes. This divided into two cases 0-8 bytes and 9-31
+ bytes. Each case is handled without loops, using binary (1,2,4,8)
+ tests.
+
In the short (0-8 byte) case no attempt is made to force alignment
- of either source or destination. The hardware will handle the
- unaligned load/stores with small delays for crossing 32- 64-byte, and
+ of either source or destination. The hardware will handle the
+ unaligned load/stores with small delays for crossing 32- 64-byte, and
4096-byte boundaries. Since these short moves are unlikely to be
- unaligned or cross these boundaries, the overhead to force
+ unaligned or cross these boundaries, the overhead to force
alignment is not justified.
-
+
The longer (9-31 byte) move is more likely to cross 32- or 64-byte
boundaries. Since only loads are sensitive to the 32-/64-byte
- boundaries it is more important to align the source then the
+ boundaries it is more important to align the source then the
destination. If the source is not already word aligned, we first
- move 1-3 bytes as needed. Since we are only word aligned we don't
- use double word load/stores to insure that all loads are aligned.
+ move 1-3 bytes as needed. Since we are only word aligned we don't
+ use double word load/stores to insure that all loads are aligned.
While the destination and stores may still be unaligned, this
is only an issue for page (4096 byte boundary) crossing, which
should be rare for these short moves. The hardware handles this
- case automatically with a small delay. */
-
+ case automatically with a small delay. */
+
.align 4
.L2:
mtcrf 0x01,5
@@ -216,15 +214,28 @@
blt cr6,5f
srdi 7,6,16
bgt cr6,3f
+#ifdef __LITTLE_ENDIAN__
+ sth 7,0(3)
+#else
sth 6,0(3)
+#endif
b 7f
.align 4
3:
+#ifdef __LITTLE_ENDIAN__
+ rotlwi 6,6,24
+ stb 6,0(3)
+ sth 7,1(3)
+#else
stb 7,0(3)
sth 6,1(3)
+#endif
b 7f
.align 4
5:
+#ifdef __LITTLE_ENDIAN__
+ rotlwi 6,6,8
+#endif
stb 6,0(3)
7:
cmpldi cr1,10,16
@@ -258,11 +269,11 @@
lwz 6,0(12)
addi 12,12,4
stw 6,0(3)
- addi 3,3,4
+ addi 3,3,4
2: /* Move 2-3 bytes. */
bf 30,1f
lhz 6,0(12)
- sth 6,0(3)
+ sth 6,0(3)
bf 31,0f
lbz 7,2(12)
stb 7,2(3)
@@ -283,8 +294,8 @@
mr 12,4
bne cr6,4f
/* Would have liked to use use ld/std here but the 630 processors are
- slow for load/store doubles that are not at least word aligned.
- Unaligned Load/Store word execute with only a 1 cycle penaltity. */
+ slow for load/store doubles that are not at least word aligned.
+ Unaligned Load/Store word execute with only a 1 cycle penalty. */
lwz 6,0(4)
lwz 7,4(4)
stw 6,0(3)
@@ -299,14 +310,14 @@
6:
bf 30,5f
lhz 7,4(4)
- sth 7,4(3)
+ sth 7,4(3)
bf 31,0f
lbz 8,6(4)
stb 8,6(3)
ld 3,-16(1)
blr
.align 4
-5:
+5:
bf 31,0f
lbz 6,4(4)
stb 6,4(3)
@@ -336,13 +347,23 @@
bf 30,1f
/* there are at least two DWs to copy */
+#ifdef __LITTLE_ENDIAN__
+ srd 0,6,10
+ sld 8,7,9
+#else
sld 0,6,10
srd 8,7,9
+#endif
or 0,0,8
ld 6,16(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srd 0,7,10
+ sld 8,6,9
+#else
sld 0,7,10
srd 8,6,9
+#endif
or 0,0,8
ld 7,24(5)
std 0,8(4)
@@ -351,8 +372,13 @@
blt cr6,8f /* if total DWs = 3, then bypass loop */
bf 31,4f
/* there is a third DW to copy */
+#ifdef __LITTLE_ENDIAN__
+ srd 0,6,10
+ sld 8,7,9
+#else
sld 0,6,10
srd 8,7,9
+#endif
or 0,0,8
std 0,0(4)
mr 6,7
@@ -363,8 +389,13 @@
b 4f
.align 4
1:
+#ifdef __LITTLE_ENDIAN__
+ srd 0,6,10
+ sld 8,7,9
+#else
sld 0,6,10
srd 8,7,9
+#endif
addi 5,5,16
or 0,0,8
bf 31,4f
@@ -375,23 +406,44 @@
addi 4,4,8
.align 4
/* copy 32 bytes at a time */
-4: sld 0,6,10
+4:
+#ifdef __LITTLE_ENDIAN__
+ srd 0,6,10
+ sld 8,7,9
+#else
+ sld 0,6,10
srd 8,7,9
+#endif
or 0,0,8
ld 6,0(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srd 0,7,10
+ sld 8,6,9
+#else
sld 0,7,10
srd 8,6,9
+#endif
or 0,0,8
ld 7,8(5)
std 0,8(4)
+#ifdef __LITTLE_ENDIAN__
+ srd 0,6,10
+ sld 8,7,9
+#else
sld 0,6,10
srd 8,7,9
+#endif
or 0,0,8
ld 6,16(5)
std 0,16(4)
+#ifdef __LITTLE_ENDIAN__
+ srd 0,7,10
+ sld 8,6,9
+#else
sld 0,7,10
srd 8,6,9
+#endif
or 0,0,8
ld 7,24(5)
std 0,24(4)
@@ -401,9 +453,14 @@
.align 4
8:
/* calculate and store the final DW */
+#ifdef __LITTLE_ENDIAN__
+ srd 0,6,10
+ sld 8,7,9
+#else
sld 0,6,10
srd 8,7,9
- or 0,0,8
+#endif
+ or 0,0,8
std 0,0(4)
3:
rldicr 0,31,0,60
@@ -413,5 +470,5 @@
ld 31,-8(1)
ld 3,-16(1)
blr
-END_GEN_TB (BP_SYM (memcpy),TB_TOCLESS)
+END_GEN_TB (memcpy,TB_TOCLESS)
libc_hidden_builtin_def (memcpy)
diff -urN glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power6/memcpy.S glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power6/memcpy.S
--- glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power6/memcpy.S 2014-05-29 13:04:56.000000000 -0500
+++ glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power6/memcpy.S 2014-05-29 13:05:27.000000000 -0500
@@ -1,5 +1,5 @@
/* Optimized memcpy implementation for PowerPC64.
- Copyright (C) 2003, 2006, 2007, 2011 Free Software Foundation, Inc.
+ Copyright (C) 2003-2014 Free Software Foundation, Inc.
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
@@ -17,52 +17,50 @@
<http://www.gnu.org/licenses/>. */
#include <sysdep.h>
-#include <bp-sym.h>
-#include <bp-asm.h>
/* __ptr_t [r3] memcpy (__ptr_t dst [r3], __ptr_t src [r4], size_t len [r5]);
Returns 'dst'.
- Memcpy handles short copies (< 32-bytes) using a binary move blocks
- (no loops) of lwz/stw. The tail (remaining 1-3) bytes is handled
- with the appropriate combination of byte and halfword load/stores.
- There is minimal effort to optimize the alignment of short moves.
+ Memcpy handles short copies (< 32-bytes) using a binary move blocks
+ (no loops) of lwz/stw. The tail (remaining 1-3) bytes is handled
+ with the appropriate combination of byte and halfword load/stores.
+ There is minimal effort to optimize the alignment of short moves.
The 64-bit implementations of POWER3 and POWER4 do a reasonable job
- of handling unligned load/stores that do not cross 32-byte boundries.
+ of handling unaligned load/stores that do not cross 32-byte boundaries.
Longer moves (>= 32-bytes) justify the effort to get at least the
destination doubleword (8-byte) aligned. Further optimization is
- posible when both source and destination are doubleword aligned.
- Each case has a optimized unrolled loop.
-
- For POWER6 unaligned loads will take a 20+ cycle hicup for any
+ possible when both source and destination are doubleword aligned.
+ Each case has a optimized unrolled loop.
+
+ For POWER6 unaligned loads will take a 20+ cycle hiccup for any
L1 cache miss that crosses a 32- or 128-byte boundary. Store
- is more forgiving and does not take a hicup until page or
- segment boundaries. So we require doubleword alignment for
+ is more forgiving and does not take a hiccup until page or
+ segment boundaries. So we require doubleword alignment for
the source but may take a risk and only require word alignment
for the destination. */
.machine "power6"
-EALIGN (BP_SYM (memcpy), 7, 0)
+EALIGN (memcpy, 7, 0)
CALL_MCOUNT 3
cmpldi cr1,5,31
neg 0,3
std 3,-16(1)
std 31,-8(1)
- andi. 11,3,7 /* check alignement of dst. */
+ andi. 11,3,7 /* check alignment of dst. */
clrldi 0,0,61 /* Number of bytes until the 1st doubleword of dst. */
- clrldi 10,4,61 /* check alignement of src. */
+ clrldi 10,4,61 /* check alignment of src. */
cmpldi cr6,5,8
ble- cr1,.L2 /* If move < 32 bytes use short move code. */
mtcrf 0x01,0
- cmpld cr6,10,11
+ cmpld cr6,10,11
srdi 9,5,3 /* Number of full double words remaining. */
beq .L0
-
+
subf 5,0,5
- /* Move 0-7 bytes as needed to get the destination doubleword alligned.
- Duplicate some code to maximize fall-throught and minimize agen delays. */
+ /* Move 0-7 bytes as needed to get the destination doubleword aligned.
+ Duplicate some code to maximize fall-through and minimize agen delays. */
1: bf 31,2f
lbz 6,0(4)
stb 6,0(3)
@@ -78,7 +76,7 @@
lwz 6,1(4)
stw 6,1(3)
b 0f
-
+
2: bf 30,4f
lhz 6,0(4)
sth 6,0(3)
@@ -86,26 +84,26 @@
lwz 6,2(4)
stw 6,2(3)
b 0f
-
+
4: bf 29,0f
lwz 6,0(4)
stw 6,0(3)
-0:
+0:
/* Add the number of bytes until the 1st doubleword of dst to src and dst. */
add 4,4,0
add 3,3,0
-
- clrldi 10,4,61 /* check alignement of src again. */
+
+ clrldi 10,4,61 /* check alignment of src again. */
srdi 9,5,3 /* Number of full double words remaining. */
-
- /* Copy doublewords from source to destination, assumpting the
+
+ /* Copy doublewords from source to destination, assuming the
destination is aligned on a doubleword boundary.
At this point we know there are at least 25 bytes left (32-7) to copy.
- The next step is to determine if the source is also doubleword aligned.
+ The next step is to determine if the source is also doubleword aligned.
If not branch to the unaligned move code at .L6. which uses
a load, shift, store strategy.
-
+
Otherwise source and destination are doubleword aligned, and we can
the optimized doubleword copy loop. */
.align 4
@@ -123,14 +121,14 @@
the main loop exits there may be a tail of 1-7 bytes. These byte
are copied a word/halfword/byte at a time as needed to preserve
alignment.
-
+
For POWER6 the L1 is store-through and the L2 is store-in. The
L2 is clocked at half CPU clock so we can store 16 bytes every
other cycle. POWER6 also has a load/store bypass so we can do
- load, load, store, store every 2 cycles.
-
+ load, load, store, store every 2 cycles.
+
The following code is sensitive to cache line alignment. Do not
- make any change with out first making sure thay don't result in
+ make any change with out first making sure they don't result in
splitting ld/std pairs across a cache line. */
mtcrf 0x02,5
@@ -273,7 +271,7 @@
std 8,16+96(10)
std 0,24+96(10)
ble cr5,L(das_loop_e)
-
+
mtctr 12
.align 4
L(das_loop2):
@@ -326,10 +324,10 @@
.align 4
L(das_tail):
beq cr1,0f
-
+
L(das_tail2):
/* At this point we have a tail of 0-7 bytes and we know that the
- destiniation is double word aligned. */
+ destination is double word aligned. */
4: bf 29,2f
lwz 6,0(4)
stw 6,0(3)
@@ -344,7 +342,7 @@
lbz 6,4(4)
stb 6,4(3)
b 0f
-
+
2: bf 30,1f
lhz 6,0(4)
sth 6,0(3)
@@ -352,7 +350,7 @@
lbz 6,2(4)
stb 6,2(3)
b 0f
-
+
1: bf 31,0f
lbz 6,0(4)
stb 6,0(3)
@@ -361,7 +359,7 @@
ld 3,-16(1)
blr
-/* Copy up to 31 bytes. This divided into two cases 0-8 bytes and 9-31
+/* Copy up to 31 bytes. This divided into two cases 0-8 bytes and 9-31
bytes. Each case is handled without loops, using binary (1,2,4,8)
tests.
@@ -402,15 +400,28 @@
blt cr6,5f
srdi 7,6,16
bgt cr6,3f
+#ifdef __LITTLE_ENDIAN__
+ sth 7,0(3)
+#else
sth 6,0(3)
+#endif
b 7f
.align 4
3:
+#ifdef __LITTLE_ENDIAN__
+ rotlwi 6,6,24
+ stb 6,0(3)
+ sth 7,1(3)
+#else
stb 7,0(3)
sth 6,1(3)
+#endif
b 7f
.align 4
5:
+#ifdef __LITTLE_ENDIAN__
+ rotlwi 6,6,8
+#endif
stb 6,0(3)
7:
cmpldi cr1,10,16
@@ -421,7 +432,7 @@
/* At least 6 bytes left and the source is word aligned. This allows
some speculative loads up front. */
/* We need to special case the fall-through because the biggest delays
- are due to address computation not being ready in time for the
+ are due to address computation not being ready in time for the
AGEN. */
lwz 6,0(12)
lwz 7,4(12)
@@ -452,7 +463,7 @@
ld 3,-16(1)
blr
.align 4
-L(dus_tail16p8): /* less then 8 bytes left. */
+L(dus_tail16p8): /* less than 8 bytes left. */
beq cr1,L(dus_tailX) /* exactly 16 bytes, early exit. */
cmpldi cr1,10,20
bf 29,L(dus_tail16p2)
@@ -466,7 +477,7 @@
ld 3,-16(1)
blr
.align 4
-L(dus_tail16p4): /* less then 4 bytes left. */
+L(dus_tail16p4): /* less than 4 bytes left. */
addi 12,12,24
addi 3,3,24
bgt cr0,L(dus_tail2)
@@ -474,7 +485,7 @@
ld 3,-16(1)
blr
.align 4
-L(dus_tail16p2): /* 16 bytes moved, less then 4 bytes left. */
+L(dus_tail16p2): /* 16 bytes moved, less than 4 bytes left. */
addi 12,12,16
addi 3,3,16
b L(dus_tail2)
@@ -499,7 +510,7 @@
ld 3,-16(1)
blr
.align 4
-L(dus_tail8p4): /* less then 4 bytes left. */
+L(dus_tail8p4): /* less than 4 bytes left. */
addi 12,12,8
addi 3,3,8
bgt cr1,L(dus_tail2)
@@ -510,14 +521,14 @@
.align 4
L(dus_tail4): /* Move 4 bytes. */
/* r6 already loaded speculatively. If we are here we know there is
- more then 4 bytes left. So there is no need to test. */
+ more than 4 bytes left. So there is no need to test. */
addi 12,12,4
stw 6,0(3)
addi 3,3,4
L(dus_tail2): /* Move 2-3 bytes. */
bf 30,L(dus_tail1)
lhz 6,0(12)
- sth 6,0(3)
+ sth 6,0(3)
bf 31,L(dus_tailX)
lbz 7,2(12)
stb 7,2(3)
@@ -537,7 +548,7 @@
.LE8:
mr 12,4
bne cr6,L(dus_4)
-/* Exactly 8 bytes. We may cross a 32-/128-byte boundry and take a ~20
+/* Exactly 8 bytes. We may cross a 32-/128-byte boundary and take a ~20
cycle delay. This case should be rare and any attempt to avoid this
would take most of 20 cycles any way. */
ld 6,0(4)
@@ -552,7 +563,7 @@
stw 6,0(3)
bf 30,L(dus_5)
lhz 7,4(4)
- sth 7,4(3)
+ sth 7,4(3)
bf 31,L(dus_0)
lbz 8,6(4)
stb 8,6(3)
@@ -590,20 +601,31 @@
bge cr0, L(du4_do)
blt cr5, L(du1_do)
beq cr5, L(du2_do)
- b L(du3_do)
-
+ b L(du3_do)
+
.align 4
L(du1_do):
bf 30,L(du1_1dw)
/* there are at least two DWs to copy */
+ /* FIXME: can combine last shift and "or" into "rldimi" */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 8
+ sldi 8,7, 64-8
+#else
sldi 0,6, 8
srdi 8,7, 64-8
+#endif
or 0,0,8
ld 6,16(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 8
+ sldi 8,6, 64-8
+#else
sldi 0,7, 8
srdi 8,6, 64-8
+#endif
or 0,0,8
ld 7,24(5)
std 0,8(4)
@@ -612,8 +634,13 @@
blt cr6,L(du1_fini) /* if total DWs = 3, then bypass loop */
bf 31,L(du1_loop)
/* there is a third DW to copy */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 8
+ sldi 8,7, 64-8
+#else
sldi 0,6, 8
srdi 8,7, 64-8
+#endif
or 0,0,8
std 0,0(4)
mr 6,7
@@ -624,8 +651,13 @@
b L(du1_loop)
.align 4
L(du1_1dw):
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 8
+ sldi 8,7, 64-8
+#else
sldi 0,6, 8
srdi 8,7, 64-8
+#endif
addi 5,5,16
or 0,0,8
bf 31,L(du1_loop)
@@ -637,23 +669,43 @@
.align 4
/* copy 32 bytes at a time */
L(du1_loop):
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 8
+ sldi 8,7, 64-8
+#else
sldi 0,6, 8
srdi 8,7, 64-8
+#endif
or 0,0,8
ld 6,0(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 8
+ sldi 8,6, 64-8
+#else
sldi 0,7, 8
srdi 8,6, 64-8
+#endif
or 0,0,8
ld 7,8(5)
std 0,8(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 8
+ sldi 8,7, 64-8
+#else
sldi 0,6, 8
srdi 8,7, 64-8
+#endif
or 0,0,8
ld 6,16(5)
std 0,16(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 8
+ sldi 8,6, 64-8
+#else
sldi 0,7, 8
srdi 8,6, 64-8
+#endif
or 0,0,8
ld 7,24(5)
std 0,24(4)
@@ -663,9 +715,14 @@
.align 4
L(du1_fini):
/* calculate and store the final DW */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 8
+ sldi 8,7, 64-8
+#else
sldi 0,6, 8
srdi 8,7, 64-8
- or 0,0,8
+#endif
+ or 0,0,8
std 0,0(4)
b L(du_done)
@@ -674,13 +731,23 @@
bf 30,L(du2_1dw)
/* there are at least two DWs to copy */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 16
+ sldi 8,7, 64-16
+#else
sldi 0,6, 16
srdi 8,7, 64-16
+#endif
or 0,0,8
ld 6,16(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 16
+ sldi 8,6, 64-16
+#else
sldi 0,7, 16
srdi 8,6, 64-16
+#endif
or 0,0,8
ld 7,24(5)
std 0,8(4)
@@ -689,8 +756,13 @@
blt cr6,L(du2_fini) /* if total DWs = 3, then bypass loop */
bf 31,L(du2_loop)
/* there is a third DW to copy */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 16
+ sldi 8,7, 64-16
+#else
sldi 0,6, 16
srdi 8,7, 64-16
+#endif
or 0,0,8
std 0,0(4)
mr 6,7
@@ -701,8 +773,13 @@
b L(du2_loop)
.align 4
L(du2_1dw):
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 16
+ sldi 8,7, 64-16
+#else
sldi 0,6, 16
srdi 8,7, 64-16
+#endif
addi 5,5,16
or 0,0,8
bf 31,L(du2_loop)
@@ -714,23 +791,43 @@
.align 4
/* copy 32 bytes at a time */
L(du2_loop):
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 16
+ sldi 8,7, 64-16
+#else
sldi 0,6, 16
srdi 8,7, 64-16
+#endif
or 0,0,8
ld 6,0(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 16
+ sldi 8,6, 64-16
+#else
sldi 0,7, 16
srdi 8,6, 64-16
+#endif
or 0,0,8
ld 7,8(5)
std 0,8(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 16
+ sldi 8,7, 64-16
+#else
sldi 0,6, 16
srdi 8,7, 64-16
+#endif
or 0,0,8
ld 6,16(5)
std 0,16(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 16
+ sldi 8,6, 64-16
+#else
sldi 0,7, 16
srdi 8,6, 64-16
+#endif
or 0,0,8
ld 7,24(5)
std 0,24(4)
@@ -740,9 +837,14 @@
.align 4
L(du2_fini):
/* calculate and store the final DW */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 16
+ sldi 8,7, 64-16
+#else
sldi 0,6, 16
srdi 8,7, 64-16
- or 0,0,8
+#endif
+ or 0,0,8
std 0,0(4)
b L(du_done)
@@ -751,13 +853,23 @@
bf 30,L(du3_1dw)
/* there are at least two DWs to copy */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 24
+ sldi 8,7, 64-24
+#else
sldi 0,6, 24
srdi 8,7, 64-24
+#endif
or 0,0,8
ld 6,16(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 24
+ sldi 8,6, 64-24
+#else
sldi 0,7, 24
srdi 8,6, 64-24
+#endif
or 0,0,8
ld 7,24(5)
std 0,8(4)
@@ -766,8 +878,13 @@
blt cr6,L(du3_fini) /* if total DWs = 3, then bypass loop */
bf 31,L(du3_loop)
/* there is a third DW to copy */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 24
+ sldi 8,7, 64-24
+#else
sldi 0,6, 24
srdi 8,7, 64-24
+#endif
or 0,0,8
std 0,0(4)
mr 6,7
@@ -778,8 +895,13 @@
b L(du3_loop)
.align 4
L(du3_1dw):
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 24
+ sldi 8,7, 64-24
+#else
sldi 0,6, 24
srdi 8,7, 64-24
+#endif
addi 5,5,16
or 0,0,8
bf 31,L(du3_loop)
@@ -791,23 +913,43 @@
.align 4
/* copy 32 bytes at a time */
L(du3_loop):
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 24
+ sldi 8,7, 64-24
+#else
sldi 0,6, 24
srdi 8,7, 64-24
+#endif
or 0,0,8
ld 6,0(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 24
+ sldi 8,6, 64-24
+#else
sldi 0,7, 24
srdi 8,6, 64-24
+#endif
or 0,0,8
ld 7,8(5)
std 0,8(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 24
+ sldi 8,7, 64-24
+#else
sldi 0,6, 24
srdi 8,7, 64-24
+#endif
or 0,0,8
ld 6,16(5)
std 0,16(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 24
+ sldi 8,6, 64-24
+#else
sldi 0,7, 24
srdi 8,6, 64-24
+#endif
or 0,0,8
ld 7,24(5)
std 0,24(4)
@@ -817,9 +959,14 @@
.align 4
L(du3_fini):
/* calculate and store the final DW */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 24
+ sldi 8,7, 64-24
+#else
sldi 0,6, 24
srdi 8,7, 64-24
- or 0,0,8
+#endif
+ or 0,0,8
std 0,0(4)
b L(du_done)
@@ -834,13 +981,23 @@
bf 30,L(du4_1dw)
/* there are at least two DWs to copy */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 32
+ sldi 8,7, 64-32
+#else
sldi 0,6, 32
srdi 8,7, 64-32
+#endif
or 0,0,8
ld 6,16(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 32
+ sldi 8,6, 64-32
+#else
sldi 0,7, 32
srdi 8,6, 64-32
+#endif
or 0,0,8
ld 7,24(5)
std 0,8(4)
@@ -849,8 +1006,13 @@
blt cr6,L(du4_fini) /* if total DWs = 3, then bypass loop */
bf 31,L(du4_loop)
/* there is a third DW to copy */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 32
+ sldi 8,7, 64-32
+#else
sldi 0,6, 32
srdi 8,7, 64-32
+#endif
or 0,0,8
std 0,0(4)
mr 6,7
@@ -861,8 +1023,13 @@
b L(du4_loop)
.align 4
L(du4_1dw):
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 32
+ sldi 8,7, 64-32
+#else
sldi 0,6, 32
srdi 8,7, 64-32
+#endif
addi 5,5,16
or 0,0,8
bf 31,L(du4_loop)
@@ -874,23 +1041,43 @@
.align 4
/* copy 32 bytes at a time */
L(du4_loop):
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 32
+ sldi 8,7, 64-32
+#else
sldi 0,6, 32
srdi 8,7, 64-32
+#endif
or 0,0,8
ld 6,0(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 32
+ sldi 8,6, 64-32
+#else
sldi 0,7, 32
srdi 8,6, 64-32
+#endif
or 0,0,8
ld 7,8(5)
std 0,8(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 32
+ sldi 8,7, 64-32
+#else
sldi 0,6, 32
srdi 8,7, 64-32
+#endif
or 0,0,8
ld 6,16(5)
std 0,16(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 32
+ sldi 8,6, 64-32
+#else
sldi 0,7, 32
srdi 8,6, 64-32
+#endif
or 0,0,8
ld 7,24(5)
std 0,24(4)
@@ -900,9 +1087,14 @@
.align 4
L(du4_fini):
/* calculate and store the final DW */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 32
+ sldi 8,7, 64-32
+#else
sldi 0,6, 32
srdi 8,7, 64-32
- or 0,0,8
+#endif
+ or 0,0,8
std 0,0(4)
b L(du_done)
@@ -911,13 +1103,23 @@
bf 30,L(du5_1dw)
/* there are at least two DWs to copy */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 40
+ sldi 8,7, 64-40
+#else
sldi 0,6, 40
srdi 8,7, 64-40
+#endif
or 0,0,8
ld 6,16(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 40
+ sldi 8,6, 64-40
+#else
sldi 0,7, 40
srdi 8,6, 64-40
+#endif
or 0,0,8
ld 7,24(5)
std 0,8(4)
@@ -926,8 +1128,13 @@
blt cr6,L(du5_fini) /* if total DWs = 3, then bypass loop */
bf 31,L(du5_loop)
/* there is a third DW to copy */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 40
+ sldi 8,7, 64-40
+#else
sldi 0,6, 40
srdi 8,7, 64-40
+#endif
or 0,0,8
std 0,0(4)
mr 6,7
@@ -938,8 +1145,13 @@
b L(du5_loop)
.align 4
L(du5_1dw):
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 40
+ sldi 8,7, 64-40
+#else
sldi 0,6, 40
srdi 8,7, 64-40
+#endif
addi 5,5,16
or 0,0,8
bf 31,L(du5_loop)
@@ -951,23 +1163,43 @@
.align 4
/* copy 32 bytes at a time */
L(du5_loop):
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 40
+ sldi 8,7, 64-40
+#else
sldi 0,6, 40
srdi 8,7, 64-40
+#endif
or 0,0,8
ld 6,0(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 40
+ sldi 8,6, 64-40
+#else
sldi 0,7, 40
srdi 8,6, 64-40
+#endif
or 0,0,8
ld 7,8(5)
std 0,8(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 40
+ sldi 8,7, 64-40
+#else
sldi 0,6, 40
srdi 8,7, 64-40
+#endif
or 0,0,8
ld 6,16(5)
std 0,16(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 40
+ sldi 8,6, 64-40
+#else
sldi 0,7, 40
srdi 8,6, 64-40
+#endif
or 0,0,8
ld 7,24(5)
std 0,24(4)
@@ -977,9 +1209,14 @@
.align 4
L(du5_fini):
/* calculate and store the final DW */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 40
+ sldi 8,7, 64-40
+#else
sldi 0,6, 40
srdi 8,7, 64-40
- or 0,0,8
+#endif
+ or 0,0,8
std 0,0(4)
b L(du_done)
@@ -988,13 +1225,23 @@
bf 30,L(du6_1dw)
/* there are at least two DWs to copy */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 48
+ sldi 8,7, 64-48
+#else
sldi 0,6, 48
srdi 8,7, 64-48
+#endif
or 0,0,8
ld 6,16(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 48
+ sldi 8,6, 64-48
+#else
sldi 0,7, 48
srdi 8,6, 64-48
+#endif
or 0,0,8
ld 7,24(5)
std 0,8(4)
@@ -1003,8 +1250,13 @@
blt cr6,L(du6_fini) /* if total DWs = 3, then bypass loop */
bf 31,L(du6_loop)
/* there is a third DW to copy */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 48
+ sldi 8,7, 64-48
+#else
sldi 0,6, 48
srdi 8,7, 64-48
+#endif
or 0,0,8
std 0,0(4)
mr 6,7
@@ -1015,8 +1267,13 @@
b L(du6_loop)
.align 4
L(du6_1dw):
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 48
+ sldi 8,7, 64-48
+#else
sldi 0,6, 48
srdi 8,7, 64-48
+#endif
addi 5,5,16
or 0,0,8
bf 31,L(du6_loop)
@@ -1028,23 +1285,43 @@
.align 4
/* copy 32 bytes at a time */
L(du6_loop):
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 48
+ sldi 8,7, 64-48
+#else
sldi 0,6, 48
srdi 8,7, 64-48
+#endif
or 0,0,8
ld 6,0(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 48
+ sldi 8,6, 64-48
+#else
sldi 0,7, 48
srdi 8,6, 64-48
+#endif
or 0,0,8
ld 7,8(5)
std 0,8(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 48
+ sldi 8,7, 64-48
+#else
sldi 0,6, 48
srdi 8,7, 64-48
+#endif
or 0,0,8
ld 6,16(5)
std 0,16(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 48
+ sldi 8,6, 64-48
+#else
sldi 0,7, 48
srdi 8,6, 64-48
+#endif
or 0,0,8
ld 7,24(5)
std 0,24(4)
@@ -1054,9 +1331,14 @@
.align 4
L(du6_fini):
/* calculate and store the final DW */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 48
+ sldi 8,7, 64-48
+#else
sldi 0,6, 48
srdi 8,7, 64-48
- or 0,0,8
+#endif
+ or 0,0,8
std 0,0(4)
b L(du_done)
@@ -1065,13 +1347,23 @@
bf 30,L(du7_1dw)
/* there are at least two DWs to copy */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 56
+ sldi 8,7, 64-56
+#else
sldi 0,6, 56
srdi 8,7, 64-56
+#endif
or 0,0,8
ld 6,16(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 56
+ sldi 8,6, 64-56
+#else
sldi 0,7, 56
srdi 8,6, 64-56
+#endif
or 0,0,8
ld 7,24(5)
std 0,8(4)
@@ -1080,8 +1372,13 @@
blt cr6,L(du7_fini) /* if total DWs = 3, then bypass loop */
bf 31,L(du7_loop)
/* there is a third DW to copy */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 56
+ sldi 8,7, 64-56
+#else
sldi 0,6, 56
srdi 8,7, 64-56
+#endif
or 0,0,8
std 0,0(4)
mr 6,7
@@ -1092,8 +1389,13 @@
b L(du7_loop)
.align 4
L(du7_1dw):
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 56
+ sldi 8,7, 64-56
+#else
sldi 0,6, 56
srdi 8,7, 64-56
+#endif
addi 5,5,16
or 0,0,8
bf 31,L(du7_loop)
@@ -1105,23 +1407,43 @@
.align 4
/* copy 32 bytes at a time */
L(du7_loop):
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 56
+ sldi 8,7, 64-56
+#else
sldi 0,6, 56
srdi 8,7, 64-56
+#endif
or 0,0,8
ld 6,0(5)
std 0,0(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 56
+ sldi 8,6, 64-56
+#else
sldi 0,7, 56
srdi 8,6, 64-56
+#endif
or 0,0,8
ld 7,8(5)
std 0,8(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 56
+ sldi 8,7, 64-56
+#else
sldi 0,6, 56
srdi 8,7, 64-56
+#endif
or 0,0,8
ld 6,16(5)
std 0,16(4)
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,7, 56
+ sldi 8,6, 64-56
+#else
sldi 0,7, 56
srdi 8,6, 64-56
+#endif
or 0,0,8
ld 7,24(5)
std 0,24(4)
@@ -1131,12 +1453,17 @@
.align 4
L(du7_fini):
/* calculate and store the final DW */
+#ifdef __LITTLE_ENDIAN__
+ srdi 0,6, 56
+ sldi 8,7, 64-56
+#else
sldi 0,6, 56
srdi 8,7, 64-56
- or 0,0,8
+#endif
+ or 0,0,8
std 0,0(4)
b L(du_done)
-
+
.align 4
L(du_done):
rldicr 0,31,0,60
@@ -1144,9 +1471,9 @@
beq cr1,0f /* If the tail is 0 bytes we are done! */
add 3,3,0
- add 12,12,0
+ add 12,12,0
/* At this point we have a tail of 0-7 bytes and we know that the
- destiniation is double word aligned. */
+ destination is double word aligned. */
4: bf 29,2f
lwz 6,0(12)
addi 12,12,4
@@ -1165,5 +1492,5 @@
ld 31,-8(1)
ld 3,-16(1)
blr
-END_GEN_TB (BP_SYM (memcpy),TB_TOCLESS)
+END_GEN_TB (memcpy,TB_TOCLESS)
libc_hidden_builtin_def (memcpy)
diff -urN glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power7/memcpy.S glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power7/memcpy.S
--- glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power7/memcpy.S 2014-05-29 13:04:56.000000000 -0500
+++ glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power7/memcpy.S 2014-05-29 13:05:40.000000000 -0500
@@ -1,5 +1,5 @@
/* Optimized memcpy implementation for PowerPC64/POWER7.
- Copyright (C) 2010, 2011 Free Software Foundation, Inc.
+ Copyright (C) 2010-2014 Free Software Foundation, Inc.
Contributed by Luis Machado <luisgpm@br.ibm.com>.
This file is part of the GNU C Library.
@@ -18,425 +18,366 @@
<http://www.gnu.org/licenses/>. */
#include <sysdep.h>
-#include <bp-sym.h>
-#include <bp-asm.h>
/* __ptr_t [r3] memcpy (__ptr_t dst [r3], __ptr_t src [r4], size_t len [r5]);
Returns 'dst'. */
+#define dst 11 /* Use r11 so r3 kept unchanged. */
+#define src 4
+#define cnt 5
+
.machine power7
-EALIGN (BP_SYM (memcpy), 5, 0)
+EALIGN (memcpy, 5, 0)
CALL_MCOUNT 3
- cmpldi cr1,5,31
+ cmpldi cr1,cnt,31
neg 0,3
- std 3,-16(1)
- std 31,-8(1)
- cfi_offset(31,-8)
ble cr1, L(copy_LT_32) /* If move < 32 bytes use short move
code. */
- andi. 11,3,7 /* Check alignment of DST. */
-
+#ifdef __LITTLE_ENDIAN__
+/* In little-endian mode, power7 takes an alignment trap on any lxvd2x
+ or stxvd2x crossing a 32-byte boundary, so ensure the aligned_copy
+ loop is only used for quadword aligned copies. */
+ andi. 10,3,15
+ clrldi 11,4,60
+#else
+ andi. 10,3,7 /* Check alignment of DST. */
+ clrldi 11,4,61 /* Check alignment of SRC. */
+#endif
+ cmpld cr6,10,11 /* SRC and DST alignments match? */
- clrldi 10,4,61 /* Check alignment of SRC. */
- cmpld cr6,10,11 /* SRC and DST alignments match? */
- mr 12,4
- mr 31,5
+ mr dst,3
bne cr6,L(copy_GE_32_unaligned)
+ beq L(aligned_copy)
- srdi 9,5,3 /* Number of full quadwords remaining. */
-
- beq L(copy_GE_32_aligned_cont)
-
- clrldi 0,0,61
- mtcrf 0x01,0
- subf 31,0,5
-
- /* Get the SRC aligned to 8 bytes. */
-
-1: bf 31,2f
- lbz 6,0(12)
- addi 12,12,1
- stb 6,0(3)
- addi 3,3,1
-2: bf 30,4f
- lhz 6,0(12)
- addi 12,12,2
- sth 6,0(3)
- addi 3,3,2
-4: bf 29,0f
- lwz 6,0(12)
- addi 12,12,4
- stw 6,0(3)
- addi 3,3,4
-0:
- clrldi 10,12,61 /* Check alignment of SRC again. */
- srdi 9,31,3 /* Number of full doublewords remaining. */
-
-L(copy_GE_32_aligned_cont):
-
- clrldi 11,31,61
- mtcrf 0x01,9
-
- srdi 8,31,5
- cmpldi cr1,9,4
- cmpldi cr6,11,0
- mr 11,12
+ mtocrf 0x01,0
+#ifdef __LITTLE_ENDIAN__
+ clrldi 0,0,60
+#else
+ clrldi 0,0,61
+#endif
- /* Copy 1~3 doublewords so the main loop starts
- at a multiple of 32 bytes. */
-
- bf 30,1f
- ld 6,0(12)
- ld 7,8(12)
- addi 11,12,16
- mtctr 8
- std 6,0(3)
- std 7,8(3)
- addi 10,3,16
- bf 31,4f
- ld 0,16(12)
- std 0,16(3)
- blt cr1,3f
- addi 11,12,24
- addi 10,3,24
- b 4f
-
- .align 4
-1: /* Copy 1 doubleword and set the counter. */
- mr 10,3
- mtctr 8
- bf 31,4f
- ld 6,0(12)
- addi 11,12,8
- std 6,0(3)
- addi 10,3,8
+/* Get the DST and SRC aligned to 8 bytes (16 for little-endian). */
+1:
+ bf 31,2f
+ lbz 6,0(src)
+ addi src,src,1
+ stb 6,0(dst)
+ addi dst,dst,1
+2:
+ bf 30,4f
+ lhz 6,0(src)
+ addi src,src,2
+ sth 6,0(dst)
+ addi dst,dst,2
+4:
+ bf 29,8f
+ lwz 6,0(src)
+ addi src,src,4
+ stw 6,0(dst)
+ addi dst,dst,4
+8:
+#ifdef __LITTLE_ENDIAN__
+ bf 28,16f
+ ld 6,0(src)
+ addi src,src,8
+ std 6,0(dst)
+ addi dst,dst,8
+16:
+#endif
+ subf cnt,0,cnt
+/* Main aligned copy loop. Copies 128 bytes at a time. */
L(aligned_copy):
- /* Main aligned copy loop. Copies up to 128-bytes at a time. */
- .align 4
-4:
- /* check for any 32-byte or 64-byte lumps that are outside of a
- nice 128-byte range. R8 contains the number of 32-byte
- lumps, so drop this into the CR, and use the SO/EQ bits to help
- handle the 32- or 64- byte lumps. Then handle the rest with an
- unrolled 128-bytes-at-a-time copy loop. */
- mtocrf 1,8
- li 6,16 # 16() index
- li 7,32 # 32() index
- li 8,48 # 48() index
-
-L(aligned_32byte):
- /* if the SO bit (indicating a 32-byte lump) is not set, move along. */
- bns cr7,L(aligned_64byte)
- lxvd2x 6,0,11
- lxvd2x 7,11,6
- addi 11,11,32
- stxvd2x 6,0,10
- stxvd2x 7,10,6
- addi 10,10,32
-
-L(aligned_64byte):
- /* if the EQ bit (indicating a 64-byte lump) is not set, move along. */
- bne cr7,L(aligned_128setup)
- lxvd2x 6,0,11
- lxvd2x 7,11,6
- lxvd2x 8,11,7
- lxvd2x 9,11,8
- addi 11,11,64
- stxvd2x 6,0,10
- stxvd2x 7,10,6
- stxvd2x 8,10,7
- stxvd2x 9,10,8
- addi 10,10,64
-
-L(aligned_128setup):
- /* Set up for the 128-byte at a time copy loop. */
- srdi 8,31,7
- cmpdi 8,0 # Any 4x lumps left?
- beq 3f # if not, move along.
- lxvd2x 6,0,11
- lxvd2x 7,11,6
- mtctr 8 # otherwise, load the ctr and begin.
- li 8,48 # 48() index
+ li 6,16
+ li 7,32
+ li 8,48
+ mtocrf 0x02,cnt
+ srdi 12,cnt,7
+ cmpdi 12,0
+ beq L(aligned_tail)
+ lxvd2x 6,0,src
+ lxvd2x 7,src,6
+ mtctr 12
b L(aligned_128loop)
+ .align 4
L(aligned_128head):
/* for the 2nd + iteration of this loop. */
- lxvd2x 6,0,11
- lxvd2x 7,11,6
+ lxvd2x 6,0,src
+ lxvd2x 7,src,6
L(aligned_128loop):
- lxvd2x 8,11,7
- lxvd2x 9,11,8
- stxvd2x 6,0,10
- addi 11,11,64
- stxvd2x 7,10,6
- stxvd2x 8,10,7
- stxvd2x 9,10,8
- lxvd2x 6,0,11
- lxvd2x 7,11,6
- addi 10,10,64
- lxvd2x 8,11,7
- lxvd2x 9,11,8
- addi 11,11,64
- stxvd2x 6,0,10
- stxvd2x 7,10,6
- stxvd2x 8,10,7
- stxvd2x 9,10,8
- addi 10,10,64
+ lxvd2x 8,src,7
+ lxvd2x 9,src,8
+ stxvd2x 6,0,dst
+ addi src,src,64
+ stxvd2x 7,dst,6
+ stxvd2x 8,dst,7
+ stxvd2x 9,dst,8
+ lxvd2x 6,0,src
+ lxvd2x 7,src,6
+ addi dst,dst,64
+ lxvd2x 8,src,7
+ lxvd2x 9,src,8
+ addi src,src,64
+ stxvd2x 6,0,dst
+ stxvd2x 7,dst,6
+ stxvd2x 8,dst,7
+ stxvd2x 9,dst,8
+ addi dst,dst,64
bdnz L(aligned_128head)
-3:
- /* Check for tail bytes. */
- rldicr 0,31,0,60
- mtcrf 0x01,31
- beq cr6,0f
-
-.L9:
- add 3,3,0
- add 12,12,0
-
- /* At this point we have a tail of 0-7 bytes and we know that the
- destination is doubleword-aligned. */
-4: /* Copy 4 bytes. */
- bf 29,2f
-
- lwz 6,0(12)
- addi 12,12,4
- stw 6,0(3)
- addi 3,3,4
-2: /* Copy 2 bytes. */
- bf 30,1f
-
- lhz 6,0(12)
- addi 12,12,2
- sth 6,0(3)
- addi 3,3,2
-1: /* Copy 1 byte. */
- bf 31,0f
-
- lbz 6,0(12)
- stb 6,0(3)
-0: /* Return original DST pointer. */
- ld 31,-8(1)
- ld 3,-16(1)
+L(aligned_tail):
+ mtocrf 0x01,cnt
+ bf 25,32f
+ lxvd2x 6,0,src
+ lxvd2x 7,src,6
+ lxvd2x 8,src,7
+ lxvd2x 9,src,8
+ addi src,src,64
+ stxvd2x 6,0,dst
+ stxvd2x 7,dst,6
+ stxvd2x 8,dst,7
+ stxvd2x 9,dst,8
+ addi dst,dst,64
+32:
+ bf 26,16f
+ lxvd2x 6,0,src
+ lxvd2x 7,src,6
+ addi src,src,32
+ stxvd2x 6,0,dst
+ stxvd2x 7,dst,6
+ addi dst,dst,32
+16:
+ bf 27,8f
+ lxvd2x 6,0,src
+ addi src,src,16
+ stxvd2x 6,0,dst
+ addi dst,dst,16
+8:
+ bf 28,4f
+ ld 6,0(src)
+ addi src,src,8
+ std 6,0(dst)
+ addi dst,dst,8
+4: /* Copies 4~7 bytes. */
+ bf 29,L(tail2)
+ lwz 6,0(src)
+ stw 6,0(dst)
+ bf 30,L(tail5)
+ lhz 7,4(src)
+ sth 7,4(dst)
+ bflr 31
+ lbz 8,6(src)
+ stb 8,6(dst)
+ /* Return original DST pointer. */
blr
- /* Handle copies of 0~31 bytes. */
- .align 4
+
+/* Handle copies of 0~31 bytes. */
+ .align 4
L(copy_LT_32):
- cmpldi cr6,5,8
- mr 12,4
- mtcrf 0x01,5
+ mr dst,3
+ cmpldi cr6,cnt,8
+ mtocrf 0x01,cnt
ble cr6,L(copy_LE_8)
/* At least 9 bytes to go. */
neg 8,4
- clrrdi 11,4,2
- andi. 0,8,3
- cmpldi cr1,5,16
- mr 10,5
+ andi. 0,8,3
+ cmpldi cr1,cnt,16
beq L(copy_LT_32_aligned)
- /* Force 4-bytes alignment for SRC. */
- mtocrf 0x01,0
- subf 10,0,5
-2: bf 30,1f
-
- lhz 6,0(12)
- addi 12,12,2
- sth 6,0(3)
- addi 3,3,2
-1: bf 31,L(end_4bytes_alignment)
-
- lbz 6,0(12)
- addi 12,12,1
- stb 6,0(3)
- addi 3,3,1
+ /* Force 4-byte alignment for SRC. */
+ mtocrf 0x01,0
+ subf cnt,0,cnt
+2:
+ bf 30,1f
+ lhz 6,0(src)
+ addi src,src,2
+ sth 6,0(dst)
+ addi dst,dst,2
+1:
+ bf 31,L(end_4bytes_alignment)
+ lbz 6,0(src)
+ addi src,src,1
+ stb 6,0(dst)
+ addi dst,dst,1
- .align 4
+ .align 4
L(end_4bytes_alignment):
- cmpldi cr1,10,16
- mtcrf 0x01,10
+ cmpldi cr1,cnt,16
+ mtocrf 0x01,cnt
L(copy_LT_32_aligned):
/* At least 6 bytes to go, and SRC is word-aligned. */
blt cr1,8f
/* Copy 16 bytes. */
- lwz 6,0(12)
- lwz 7,4(12)
- stw 6,0(3)
- lwz 8,8(12)
- stw 7,4(3)
- lwz 6,12(12)
- addi 12,12,16
- stw 8,8(3)
- stw 6,12(3)
- addi 3,3,16
+ lwz 6,0(src)
+ lwz 7,4(src)
+ stw 6,0(dst)
+ lwz 8,8(src)
+ stw 7,4(dst)
+ lwz 6,12(src)
+ addi src,src,16
+ stw 8,8(dst)
+ stw 6,12(dst)
+ addi dst,dst,16
8: /* Copy 8 bytes. */
- bf 28,4f
+ bf 28,L(tail4)
+ lwz 6,0(src)
+ lwz 7,4(src)
+ addi src,src,8
+ stw 6,0(dst)
+ stw 7,4(dst)
+ addi dst,dst,8
+
+ .align 4
+/* Copies 4~7 bytes. */
+L(tail4):
+ bf 29,L(tail2)
+ lwz 6,0(src)
+ stw 6,0(dst)
+ bf 30,L(tail5)
+ lhz 7,4(src)
+ sth 7,4(dst)
+ bflr 31
+ lbz 8,6(src)
+ stb 8,6(dst)
+ /* Return original DST pointer. */
+ blr
- lwz 6,0(12)
- lwz 7,4(12)
- addi 12,12,8
- stw 6,0(3)
- stw 7,4(3)
- addi 3,3,8
-4: /* Copy 4 bytes. */
- bf 29,2f
-
- lwz 6,0(12)
- addi 12,12,4
- stw 6,0(3)
- addi 3,3,4
-2: /* Copy 2-3 bytes. */
+ .align 4
+/* Copies 2~3 bytes. */
+L(tail2):
bf 30,1f
-
- lhz 6,0(12)
- sth 6,0(3)
- bf 31,0f
- lbz 7,2(12)
- stb 7,2(3)
- ld 3,-16(1)
+ lhz 6,0(src)
+ sth 6,0(dst)
+ bflr 31
+ lbz 7,2(src)
+ stb 7,2(dst)
blr
- .align 4
-1: /* Copy 1 byte. */
- bf 31,0f
+ .align 4
+L(tail5):
+ bflr 31
+ lbz 6,4(src)
+ stb 6,4(dst)
+ blr
- lbz 6,0(12)
- stb 6,0(3)
-0: /* Return original DST pointer. */
- ld 3,-16(1)
+ .align 4
+1:
+ bflr 31
+ lbz 6,0(src)
+ stb 6,0(dst)
+ /* Return original DST pointer. */
blr
- /* Handles copies of 0~8 bytes. */
- .align 4
+
+/* Handles copies of 0~8 bytes. */
+ .align 4
L(copy_LE_8):
- bne cr6,4f
+ bne cr6,L(tail4)
/* Though we could've used ld/std here, they are still
slow for unaligned cases. */
- lwz 6,0(4)
- lwz 7,4(4)
- stw 6,0(3)
- stw 7,4(3)
- ld 3,-16(1) /* Return original DST pointers. */
+ lwz 6,0(src)
+ lwz 7,4(src)
+ stw 6,0(dst)
+ stw 7,4(dst)
blr
- .align 4
-4: /* Copies 4~7 bytes. */
- bf 29,2b
- lwz 6,0(4)
- stw 6,0(3)
- bf 30,5f
- lhz 7,4(4)
- sth 7,4(3)
- bf 31,0f
- lbz 8,6(4)
- stb 8,6(3)
- ld 3,-16(1)
- blr
-
- .align 4
-5: /* Copy 1 byte. */
- bf 31,0f
-
- lbz 6,4(4)
- stb 6,4(3)
-
-0: /* Return original DST pointer. */
- ld 3,-16(1)
- blr
-
- /* Handle copies of 32+ bytes where DST is aligned (to quadword) but
- SRC is not. Use aligned quadword loads from SRC, shifted to realign
- the data, allowing for aligned DST stores. */
- .align 4
+/* Handle copies of 32+ bytes where DST is aligned (to quadword) but
+ SRC is not. Use aligned quadword loads from SRC, shifted to realign
+ the data, allowing for aligned DST stores. */
+ .align 4
L(copy_GE_32_unaligned):
- clrldi 0,0,60 /* Number of bytes until the 1st
- quadword. */
- andi. 11,3,15 /* Check alignment of DST (against
- quadwords). */
- srdi 9,5,4 /* Number of full quadwords remaining. */
+ clrldi 0,0,60 /* Number of bytes until the 1st dst quadword. */
+#ifndef __LITTLE_ENDIAN__
+ andi. 10,3,15 /* Check alignment of DST (against quadwords). */
+#endif
+ srdi 9,cnt,4 /* Number of full quadwords remaining. */
beq L(copy_GE_32_unaligned_cont)
- /* SRC is not quadword aligned, get it aligned. */
+ /* DST is not quadword aligned, get it aligned. */
- mtcrf 0x01,0
- subf 31,0,5
+ mtocrf 0x01,0
+ subf cnt,0,cnt
/* Vector instructions work best when proper alignment (16-bytes)
is present. Move 0~15 bytes as needed to get DST quadword-aligned. */
-1: /* Copy 1 byte. */
+1:
bf 31,2f
-
- lbz 6,0(12)
- addi 12,12,1
- stb 6,0(3)
- addi 3,3,1
-2: /* Copy 2 bytes. */
+ lbz 6,0(src)
+ addi src,src,1
+ stb 6,0(dst)
+ addi dst,dst,1
+2:
bf 30,4f
-
- lhz 6,0(12)
- addi 12,12,2
- sth 6,0(3)
- addi 3,3,2
-4: /* Copy 4 bytes. */
+ lhz 6,0(src)
+ addi src,src,2
+ sth 6,0(dst)
+ addi dst,dst,2
+4:
bf 29,8f
-
- lwz 6,0(12)
- addi 12,12,4
- stw 6,0(3)
- addi 3,3,4
-8: /* Copy 8 bytes. */
+ lwz 6,0(src)
+ addi src,src,4
+ stw 6,0(dst)
+ addi dst,dst,4
+8:
bf 28,0f
-
- ld 6,0(12)
- addi 12,12,8
- std 6,0(3)
- addi 3,3,8
+ ld 6,0(src)
+ addi src,src,8
+ std 6,0(dst)
+ addi dst,dst,8
0:
- clrldi 10,12,60 /* Check alignment of SRC. */
- srdi 9,31,4 /* Number of full quadwords remaining. */
+ srdi 9,cnt,4 /* Number of full quadwords remaining. */
/* The proper alignment is present, it is OK to copy the bytes now. */
L(copy_GE_32_unaligned_cont):
/* Setup two indexes to speed up the indexed vector operations. */
- clrldi 11,31,60
- li 6,16 /* Index for 16-bytes offsets. */
+ clrldi 10,cnt,60
+ li 6,16 /* Index for 16-bytes offsets. */
li 7,32 /* Index for 32-bytes offsets. */
- cmpldi cr1,11,0
- srdi 8,31,5 /* Setup the loop counter. */
- mr 10,3
- mr 11,12
- mtcrf 0x01,9
- cmpldi cr6,9,1
- lvsl 5,0,12
- lvx 3,0,12
- bf 31,L(setup_unaligned_loop)
-
- /* Copy another 16 bytes to align to 32-bytes due to the loop . */
- lvx 4,12,6
- vperm 6,3,4,5
- addi 11,12,16
- addi 10,3,16
- stvx 6,0,3
+ cmpldi cr1,10,0
+ srdi 8,cnt,5 /* Setup the loop counter. */
+ mtocrf 0x01,9
+ cmpldi cr6,9,1
+#ifdef __LITTLE_ENDIAN__
+ lvsr 5,0,src
+#else
+ lvsl 5,0,src
+#endif
+ lvx 3,0,src
+ li 0,0
+ bf 31,L(setup_unaligned_loop)
+
+ /* Copy another 16 bytes to align to 32-bytes due to the loop. */
+ lvx 4,src,6
+#ifdef __LITTLE_ENDIAN__
+ vperm 6,4,3,5
+#else
+ vperm 6,3,4,5
+#endif
+ addi src,src,16
+ stvx 6,0,dst
+ addi dst,dst,16
vor 3,4,4
+ clrrdi 0,src,60
L(setup_unaligned_loop):
- mtctr 8
- ble cr6,L(end_unaligned_loop)
+ mtctr 8
+ ble cr6,L(end_unaligned_loop)
/* Copy 32 bytes at a time using vector instructions. */
- .align 4
+ .align 4
L(unaligned_loop):
/* Note: vr6/vr10 may contain data that was already copied,
@@ -444,63 +385,56 @@
some portions again. This is faster than having unaligned
vector instructions though. */
- lvx 4,11,6 /* vr4 = r11+16. */
- vperm 6,3,4,5 /* Merge the correctly-aligned portions
- of vr3/vr4 into vr6. */
- lvx 3,11,7 /* vr3 = r11+32. */
- vperm 10,4,3,5 /* Merge the correctly-aligned portions
- of vr3/vr4 into vr10. */
- addi 11,11,32
- stvx 6,0,10
- stvx 10,10,6
- addi 10,10,32
-
+ lvx 4,src,6
+#ifdef __LITTLE_ENDIAN__
+ vperm 6,4,3,5
+#else
+ vperm 6,3,4,5
+#endif
+ lvx 3,src,7
+#ifdef __LITTLE_ENDIAN__
+ vperm 10,3,4,5
+#else
+ vperm 10,4,3,5
+#endif
+ addi src,src,32
+ stvx 6,0,dst
+ stvx 10,dst,6
+ addi dst,dst,32
bdnz L(unaligned_loop)
- .align 4
+ clrrdi 0,src,60
+
+ .align 4
L(end_unaligned_loop):
/* Check for tail bytes. */
- rldicr 0,31,0,59
- mtcrf 0x01,31
- beq cr1,0f
+ mtocrf 0x01,cnt
+ beqlr cr1
- add 3,3,0
- add 12,12,0
+ add src,src,0
/* We have 1~15 tail bytes to copy, and DST is quadword aligned. */
-8: /* Copy 8 bytes. */
+ /* Copy 8 bytes. */
bf 28,4f
-
- lwz 6,0(12)
- lwz 7,4(12)
- addi 12,12,8
- stw 6,0(3)
- stw 7,4(3)
- addi 3,3,8
-4: /* Copy 4 bytes. */
- bf 29,2f
-
- lwz 6,0(12)
- addi 12,12,4
- stw 6,0(3)
- addi 3,3,4
-2: /* Copy 2~3 bytes. */
- bf 30,1f
-
- lhz 6,0(12)
- addi 12,12,2
- sth 6,0(3)
- addi 3,3,2
-1: /* Copy 1 byte. */
- bf 31,0f
-
- lbz 6,0(12)
- stb 6,0(3)
-0: /* Return original DST pointer. */
- ld 31,-8(1)
- ld 3,-16(1)
+ lwz 6,0(src)
+ lwz 7,4(src)
+ addi src,src,8
+ stw 6,0(dst)
+ stw 7,4(dst)
+ addi dst,dst,8
+4: /* Copy 4~7 bytes. */
+ bf 29,L(tail2)
+ lwz 6,0(src)
+ stw 6,0(dst)
+ bf 30,L(tail5)
+ lhz 7,4(src)
+ sth 7,4(dst)
+ bflr 31
+ lbz 8,6(src)
+ stb 8,6(dst)
+ /* Return original DST pointer. */
blr
-END_GEN_TB (BP_SYM (memcpy),TB_TOCLESS)
+END_GEN_TB (memcpy,TB_TOCLESS)
libc_hidden_builtin_def (memcpy)
diff -urN glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power7/mempcpy.S glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power7/mempcpy.S
--- glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power7/mempcpy.S 2014-05-29 13:04:56.000000000 -0500
+++ glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power7/mempcpy.S 2014-05-29 13:04:56.000000000 -0500
@@ -367,13 +367,21 @@
mr 11,12
mtcrf 0x01,9
cmpldi cr6,9,1
- lvsl 5,0,12
+#ifdef __LITTLE_ENDIAN__
+ lvsr 5,0,12
+#else
+ lvsl 5,0,12
+#endif
lvx 3,0,12
bf 31,L(setup_unaligned_loop)
/* Copy another 16 bytes to align to 32-bytes due to the loop . */
lvx 4,12,6
- vperm 6,3,4,5
+#ifdef __LITTLE_ENDIAN__
+ vperm 6,4,3,5
+#else
+ vperm 6,3,4,5
+#endif
addi 11,12,16
addi 10,3,16
stvx 6,0,3
@@ -393,11 +401,17 @@
vector instructions though. */
lvx 4,11,6 /* vr4 = r11+16. */
- vperm 6,3,4,5 /* Merge the correctly-aligned portions
- of vr3/vr4 into vr6. */
+#ifdef __LITTLE_ENDIAN__
+ vperm 6,4,3,5
+#else
+ vperm 6,3,4,5
+#endif
lvx 3,11,7 /* vr3 = r11+32. */
- vperm 10,4,3,5 /* Merge the correctly-aligned portions
- of vr3/vr4 into vr10. */
+#ifdef __LITTLE_ENDIAN__
+ vperm 10,3,4,5
+#else
+ vperm 10,4,3,5
+#endif
addi 11,11,32
stvx 6,0,10
stvx 10,10,6