|
 |
ce426f |
# commit 33ee81de05e83ce12f32a491270bb4c1611399c7
|
|
|
ce426f |
# Author: Alan Modra <amodra@gmail.com>
|
|
|
ce426f |
# Date: Sat Aug 17 18:40:48 2013 +0930
|
|
|
ce426f |
#
|
|
|
ce426f |
# PowerPC LE strnlen
|
|
|
ce426f |
# http://sourceware.org/ml/libc-alpha/2013-08/msg00098.html
|
|
|
ce426f |
#
|
|
|
ce426f |
# The existing strnlen code has a number of defects, so this patch is more
|
|
|
ce426f |
# than just adding little-endian support. The changes here are similar to
|
|
|
ce426f |
# those for memchr.
|
|
|
ce426f |
#
|
|
|
ce426f |
# * sysdeps/powerpc/powerpc64/power7/strnlen.S (strnlen): Add
|
|
|
ce426f |
# little-endian support. Remove unnecessary "are we done" tests.
|
|
|
ce426f |
# Handle "s" wrapping around zero and extremely large "size".
|
|
|
ce426f |
# Correct main loop count. Handle single left-over word from main
|
|
|
ce426f |
# loop inline rather than by using small_loop. Correct comments.
|
|
|
ce426f |
# Delete "zero" tail, use "end_max" instead.
|
|
|
ce426f |
# * sysdeps/powerpc/powerpc32/power7/strnlen.S: Likewise.
|
|
|
ce426f |
#
|
|
|
ce426f |
diff -urN glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power7/strnlen.S glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power7/strnlen.S
|
|
|
ce426f |
--- glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power7/strnlen.S 2014-05-28 12:40:17.000000000 -0500
|
|
|
ce426f |
+++ glibc-2.17-c758a686/sysdeps/powerpc/powerpc32/power7/strnlen.S 2014-05-28 12:44:52.000000000 -0500
|
|
|
ce426f |
@@ -30,51 +30,47 @@
|
|
|
ce426f |
add r7,r3,r4 /* Calculate the last acceptable address. */
|
|
|
ce426f |
cmplwi r4,16
|
|
|
ce426f |
li r0,0 /* Word with null chars. */
|
|
|
ce426f |
+ addi r7,r7,-1
|
|
|
ce426f |
ble L(small_range)
|
|
|
ce426f |
|
|
|
ce426f |
- cmplw cr7,r3,r7 /* Is the address equal or less than r3? If
|
|
|
ce426f |
- it's equal or less, it means size is either 0
|
|
|
ce426f |
- or a negative number. */
|
|
|
ce426f |
- ble cr7,L(proceed)
|
|
|
ce426f |
-
|
|
|
ce426f |
- li r7,-1 /* Make r11 the biggest if r4 <= 0. */
|
|
|
ce426f |
-L(proceed):
|
|
|
ce426f |
rlwinm r6,r3,3,27,28 /* Calculate padding. */
|
|
|
ce426f |
lwz r12,0(r8) /* Load word from memory. */
|
|
|
ce426f |
cmpb r10,r12,r0 /* Check for null bytes in DWORD1. */
|
|
|
ce426f |
+#ifdef __LITTLE_ENDIAN__
|
|
|
ce426f |
+ srw r10,r10,r6
|
|
|
ce426f |
+ slw r10,r10,r6
|
|
|
ce426f |
+#else
|
|
|
ce426f |
slw r10,r10,r6
|
|
|
ce426f |
srw r10,r10,r6
|
|
|
ce426f |
+#endif
|
|
|
ce426f |
cmplwi cr7,r10,0 /* If r10 == 0, no null's have been found. */
|
|
|
ce426f |
bne cr7,L(done)
|
|
|
ce426f |
|
|
|
ce426f |
- /* Are we done already? */
|
|
|
ce426f |
- addi r9,r8,4
|
|
|
ce426f |
- cmplw cr6,r9,r7
|
|
|
ce426f |
- bge cr6,L(end_max)
|
|
|
ce426f |
-
|
|
|
ce426f |
+ clrrwi r7,r7,2 /* Address of last word. */
|
|
|
ce426f |
mtcrf 0x01,r8
|
|
|
ce426f |
/* Are we now aligned to a doubleword boundary? If so, skip to
|
|
|
ce426f |
the main loop. Otherwise, go through the alignment code. */
|
|
|
ce426f |
|
|
|
ce426f |
bt 29,L(loop_setup)
|
|
|
ce426f |
|
|
|
ce426f |
- /* Handle DWORD2 of pair. */
|
|
|
ce426f |
+ /* Handle WORD2 of pair. */
|
|
|
ce426f |
lwzu r12,4(r8)
|
|
|
ce426f |
cmpb r10,r12,r0
|
|
|
ce426f |
cmplwi cr7,r10,0
|
|
|
ce426f |
bne cr7,L(done)
|
|
|
ce426f |
|
|
|
ce426f |
- /* Are we done already? */
|
|
|
ce426f |
- addi r9,r8,4
|
|
|
ce426f |
- cmplw cr6,r9,r7
|
|
|
ce426f |
- bge cr6,L(end_max)
|
|
|
ce426f |
-
|
|
|
ce426f |
L(loop_setup):
|
|
|
ce426f |
- sub r5,r7,r9
|
|
|
ce426f |
+ /* The last word we want to read in the loop below is the one
|
|
|
ce426f |
+ containing the last byte of the string, ie. the word at
|
|
|
ce426f |
+ (s + size - 1) & ~3, or r7. The first word read is at
|
|
|
ce426f |
+ r8 + 4, we read 2 * cnt words, so the last word read will
|
|
|
ce426f |
+ be at r8 + 4 + 8 * cnt - 4. Solving for cnt gives
|
|
|
ce426f |
+ cnt = (r7 - r8) / 8 */
|
|
|
ce426f |
+ sub r5,r7,r8
|
|
|
ce426f |
srwi r6,r5,3 /* Number of loop iterations. */
|
|
|
ce426f |
mtctr r6 /* Setup the counter. */
|
|
|
ce426f |
- b L(loop)
|
|
|
ce426f |
- /* Main loop to look for the null byte backwards in the string. Since
|
|
|
ce426f |
+
|
|
|
ce426f |
+ /* Main loop to look for the null byte in the string. Since
|
|
|
ce426f |
it's a small loop (< 8 instructions), align it to 32-bytes. */
|
|
|
ce426f |
.p2align 5
|
|
|
ce426f |
L(loop):
|
|
|
ce426f |
@@ -90,15 +86,18 @@
|
|
|
ce426f |
cmplwi cr7,r5,0
|
|
|
ce426f |
bne cr7,L(found)
|
|
|
ce426f |
bdnz L(loop)
|
|
|
ce426f |
- /* We're here because the counter reached 0, and that means we
|
|
|
ce426f |
- didn't have any matches for null in the whole range. Just return
|
|
|
ce426f |
- the original size. */
|
|
|
ce426f |
- addi r9,r8,4
|
|
|
ce426f |
- cmplw cr6,r9,r7
|
|
|
ce426f |
- blt cr6,L(loop_small)
|
|
|
ce426f |
+
|
|
|
ce426f |
+ /* We may have one more word to read. */
|
|
|
ce426f |
+ cmplw cr6,r8,r7
|
|
|
ce426f |
+ beq cr6,L(end_max)
|
|
|
ce426f |
+
|
|
|
ce426f |
+ lwzu r12,4(r8)
|
|
|
ce426f |
+ cmpb r10,r12,r0
|
|
|
ce426f |
+ cmplwi cr6,r10,0
|
|
|
ce426f |
+ bne cr6,L(done)
|
|
|
ce426f |
|
|
|
ce426f |
L(end_max):
|
|
|
ce426f |
- sub r3,r7,r3
|
|
|
ce426f |
+ mr r3,r4
|
|
|
ce426f |
blr
|
|
|
ce426f |
|
|
|
ce426f |
/* OK, one (or both) of the words contains a null byte. Check
|
|
|
ce426f |
@@ -123,49 +122,56 @@
|
|
|
ce426f |
We need to make sure the null char is *before* the end of the
|
|
|
ce426f |
range. */
|
|
|
ce426f |
L(done):
|
|
|
ce426f |
- cntlzw r0,r10 /* Count leading zeroes before the match. */
|
|
|
ce426f |
- srwi r0,r0,3 /* Convert leading zeroes to bytes. */
|
|
|
ce426f |
- add r9,r8,r0
|
|
|
ce426f |
- sub r6,r9,r3 /* Length until the match. */
|
|
|
ce426f |
- cmplw r9,r7
|
|
|
ce426f |
- bgt L(end_max)
|
|
|
ce426f |
- mr r3,r6
|
|
|
ce426f |
- blr
|
|
|
ce426f |
-
|
|
|
ce426f |
- .align 4
|
|
|
ce426f |
-L(zero):
|
|
|
ce426f |
- li r3,0
|
|
|
ce426f |
+#ifdef __LITTLE_ENDIAN__
|
|
|
ce426f |
+ addi r0,r10,-1
|
|
|
ce426f |
+ andc r0,r0,r10
|
|
|
ce426f |
+ popcntw r0,r0
|
|
|
ce426f |
+#else
|
|
|
ce426f |
+ cntlzw r0,r10 /* Count leading zeros before the match. */
|
|
|
ce426f |
+#endif
|
|
|
ce426f |
+ sub r3,r8,r3
|
|
|
ce426f |
+ srwi r0,r0,3 /* Convert leading/trailing zeros to bytes. */
|
|
|
ce426f |
+ add r3,r3,r0 /* Length until the match. */
|
|
|
ce426f |
+ cmplw r3,r4
|
|
|
ce426f |
+ blelr
|
|
|
ce426f |
+ mr r3,r4
|
|
|
ce426f |
blr
|
|
|
ce426f |
|
|
|
ce426f |
-/* Deals with size <= 32. */
|
|
|
ce426f |
+/* Deals with size <= 16. */
|
|
|
ce426f |
.align 4
|
|
|
ce426f |
L(small_range):
|
|
|
ce426f |
cmplwi r4,0
|
|
|
ce426f |
- beq L(zero)
|
|
|
ce426f |
+ beq L(end_max)
|
|
|
ce426f |
+
|
|
|
ce426f |
+ clrrwi r7,r7,2 /* Address of last word. */
|
|
|
ce426f |
|
|
|
ce426f |
rlwinm r6,r3,3,27,28 /* Calculate padding. */
|
|
|
ce426f |
lwz r12,0(r8) /* Load word from memory. */
|
|
|
ce426f |
cmpb r10,r12,r0 /* Check for null bytes in WORD1. */
|
|
|
ce426f |
+#ifdef __LITTLE_ENDIAN__
|
|
|
ce426f |
+ srw r10,r10,r6
|
|
|
ce426f |
+ slw r10,r10,r6
|
|
|
ce426f |
+#else
|
|
|
ce426f |
slw r10,r10,r6
|
|
|
ce426f |
srw r10,r10,r6
|
|
|
ce426f |
+#endif
|
|
|
ce426f |
cmplwi cr7,r10,0
|
|
|
ce426f |
bne cr7,L(done)
|
|
|
ce426f |
|
|
|
ce426f |
- addi r9,r8,4
|
|
|
ce426f |
- cmplw r9,r7
|
|
|
ce426f |
- bge L(end_max)
|
|
|
ce426f |
- b L(loop_small)
|
|
|
ce426f |
+ cmplw r8,r7
|
|
|
ce426f |
+ beq L(end_max)
|
|
|
ce426f |
|
|
|
ce426f |
.p2align 5
|
|
|
ce426f |
L(loop_small):
|
|
|
ce426f |
lwzu r12,4(r8)
|
|
|
ce426f |
cmpb r10,r12,r0
|
|
|
ce426f |
- addi r9,r8,4
|
|
|
ce426f |
cmplwi cr6,r10,0
|
|
|
ce426f |
bne cr6,L(done)
|
|
|
ce426f |
- cmplw r9,r7
|
|
|
ce426f |
- bge L(end_max)
|
|
|
ce426f |
- b L(loop_small)
|
|
|
ce426f |
+ cmplw r8,r7
|
|
|
ce426f |
+ bne L(loop_small)
|
|
|
ce426f |
+ mr r3,r4
|
|
|
ce426f |
+ blr
|
|
|
ce426f |
+
|
|
|
ce426f |
END (BP_SYM (__strnlen))
|
|
|
ce426f |
weak_alias (BP_SYM (__strnlen), BP_SYM(strnlen))
|
|
|
ce426f |
libc_hidden_builtin_def (strnlen)
|
|
|
ce426f |
diff -urN glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power7/strnlen.S glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power7/strnlen.S
|
|
|
ce426f |
--- glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power7/strnlen.S 2014-05-28 12:40:17.000000000 -0500
|
|
|
ce426f |
+++ glibc-2.17-c758a686/sysdeps/powerpc/powerpc64/power7/strnlen.S 2014-05-28 13:24:41.000000000 -0500
|
|
|
ce426f |
@@ -26,33 +26,29 @@
|
|
|
ce426f |
ENTRY (BP_SYM (__strnlen))
|
|
|
ce426f |
CALL_MCOUNT 2
|
|
|
ce426f |
dcbt 0,r3
|
|
|
ce426f |
- clrrdi r8,r3,3
|
|
|
ce426f |
+ clrrdi r8,r3,3
|
|
|
ce426f |
add r7,r3,r4 /* Calculate the last acceptable address. */
|
|
|
ce426f |
cmpldi r4,32
|
|
|
ce426f |
li r0,0 /* Doubleword with null chars. */
|
|
|
ce426f |
+ addi r7,r7,-1
|
|
|
ce426f |
+
|
|
|
ce426f |
/* If we have less than 33 bytes to search, skip to a faster code. */
|
|
|
ce426f |
ble L(small_range)
|
|
|
ce426f |
|
|
|
ce426f |
- cmpld cr7,r3,r7 /* Is the address equal or less than r3? If
|
|
|
ce426f |
- it's equal or less, it means size is either 0
|
|
|
ce426f |
- or a negative number. */
|
|
|
ce426f |
- ble cr7,L(proceed)
|
|
|
ce426f |
-
|
|
|
ce426f |
- li r7,-1 /* Make r11 the biggest if r4 <= 0. */
|
|
|
ce426f |
-L(proceed):
|
|
|
ce426f |
rlwinm r6,r3,3,26,28 /* Calculate padding. */
|
|
|
ce426f |
ld r12,0(r8) /* Load doubleword from memory. */
|
|
|
ce426f |
cmpb r10,r12,r0 /* Check for null bytes in DWORD1. */
|
|
|
ce426f |
+#ifdef __LITTLE_ENDIAN__
|
|
|
ce426f |
+ srd r10,r10,r6
|
|
|
ce426f |
+ sld r10,r10,r6
|
|
|
ce426f |
+#else
|
|
|
ce426f |
sld r10,r10,r6
|
|
|
ce426f |
srd r10,r10,r6
|
|
|
ce426f |
+#endif
|
|
|
ce426f |
cmpldi cr7,r10,0 /* If r10 == 0, no null's have been found. */
|
|
|
ce426f |
bne cr7,L(done)
|
|
|
ce426f |
|
|
|
ce426f |
- /* Are we done already? */
|
|
|
ce426f |
- addi r9,r8,8
|
|
|
ce426f |
- cmpld cr6,r9,r7
|
|
|
ce426f |
- bge cr6,L(end_max)
|
|
|
ce426f |
-
|
|
|
ce426f |
+ clrrdi r7,r7,3 /* Address of last doubleword. */
|
|
|
ce426f |
mtcrf 0x01,r8
|
|
|
ce426f |
/* Are we now aligned to a quadword boundary? If so, skip to
|
|
|
ce426f |
the main loop. Otherwise, go through the alignment code. */
|
|
|
ce426f |
@@ -65,17 +61,18 @@
|
|
|
ce426f |
cmpldi cr7,r10,0
|
|
|
ce426f |
bne cr7,L(done)
|
|
|
ce426f |
|
|
|
ce426f |
- /* Are we done already? */
|
|
|
ce426f |
- addi r9,r8,8
|
|
|
ce426f |
- cmpld cr6,r9,r7
|
|
|
ce426f |
- bge cr6,L(end_max)
|
|
|
ce426f |
-
|
|
|
ce426f |
L(loop_setup):
|
|
|
ce426f |
- sub r5,r7,r9
|
|
|
ce426f |
+ /* The last dword we want to read in the loop below is the one
|
|
|
ce426f |
+ containing the last byte of the string, ie. the dword at
|
|
|
ce426f |
+ (s + size - 1) & ~7, or r7. The first dword read is at
|
|
|
ce426f |
+ r8 + 8, we read 2 * cnt dwords, so the last dword read will
|
|
|
ce426f |
+ be at r8 + 8 + 16 * cnt - 8. Solving for cnt gives
|
|
|
ce426f |
+ cnt = (r7 - r8) / 16 */
|
|
|
ce426f |
+ sub r5,r7,r8
|
|
|
ce426f |
srdi r6,r5,4 /* Number of loop iterations. */
|
|
|
ce426f |
mtctr r6 /* Setup the counter. */
|
|
|
ce426f |
- b L(loop)
|
|
|
ce426f |
- /* Main loop to look for the null byte backwards in the string. Since
|
|
|
ce426f |
+
|
|
|
ce426f |
+ /* Main loop to look for the null byte in the string. Since
|
|
|
ce426f |
it's a small loop (< 8 instructions), align it to 32-bytes. */
|
|
|
ce426f |
.p2align 5
|
|
|
ce426f |
L(loop):
|
|
|
ce426f |
@@ -91,15 +88,18 @@
|
|
|
ce426f |
cmpldi cr7,r5,0
|
|
|
ce426f |
bne cr7,L(found)
|
|
|
ce426f |
bdnz L(loop)
|
|
|
ce426f |
- /* We're here because the counter reached 0, and that means we
|
|
|
ce426f |
- didn't have any matches for null in the whole range. Just return
|
|
|
ce426f |
- the original size. */
|
|
|
ce426f |
- addi r9,r8,8
|
|
|
ce426f |
- cmpld cr6,r9,r7
|
|
|
ce426f |
- blt cr6,L(loop_small)
|
|
|
ce426f |
+
|
|
|
ce426f |
+ /* We may have one more dword to read. */
|
|
|
ce426f |
+ cmpld cr6,r8,r7
|
|
|
ce426f |
+ beq cr6,L(end_max)
|
|
|
ce426f |
+
|
|
|
ce426f |
+ ldu r12,8(r8)
|
|
|
ce426f |
+ cmpb r10,r12,r0
|
|
|
ce426f |
+ cmpldi cr6,r10,0
|
|
|
ce426f |
+ bne cr6,L(done)
|
|
|
ce426f |
|
|
|
ce426f |
L(end_max):
|
|
|
ce426f |
- sub r3,r7,r3
|
|
|
ce426f |
+ mr r3,r4
|
|
|
ce426f |
blr
|
|
|
ce426f |
|
|
|
ce426f |
/* OK, one (or both) of the doublewords contains a null byte. Check
|
|
|
ce426f |
@@ -121,52 +121,59 @@
|
|
|
ce426f |
/* r10 has the output of the cmpb instruction, that is, it contains
|
|
|
ce426f |
0xff in the same position as the null byte in the original
|
|
|
ce426f |
doubleword from the string. Use that to calculate the length.
|
|
|
ce426f |
- We need to make sure the null char is *before* the start of the
|
|
|
ce426f |
- range (since we're going backwards). */
|
|
|
ce426f |
+ We need to make sure the null char is *before* the end of the
|
|
|
ce426f |
+ range. */
|
|
|
ce426f |
L(done):
|
|
|
ce426f |
- cntlzd r0,r10 /* Count leading zeroes before the match. */
|
|
|
ce426f |
- srdi r0,r0,3 /* Convert leading zeroes to bytes. */
|
|
|
ce426f |
- add r9,r8,r0
|
|
|
ce426f |
- sub r6,r9,r3 /* Length until the match. */
|
|
|
ce426f |
- cmpld r9,r7
|
|
|
ce426f |
- bgt L(end_max)
|
|
|
ce426f |
- mr r3,r6
|
|
|
ce426f |
- blr
|
|
|
ce426f |
-
|
|
|
ce426f |
- .align 4
|
|
|
ce426f |
-L(zero):
|
|
|
ce426f |
- li r3,0
|
|
|
ce426f |
+#ifdef __LITTLE_ENDIAN__
|
|
|
ce426f |
+ addi r0,r10,-1
|
|
|
ce426f |
+ andc r0,r0,r10
|
|
|
ce426f |
+ popcntd r0,r0
|
|
|
ce426f |
+#else
|
|
|
ce426f |
+ cntlzd r0,r10 /* Count leading zeros before the match. */
|
|
|
ce426f |
+#endif
|
|
|
ce426f |
+ sub r3,r8,r3
|
|
|
ce426f |
+ srdi r0,r0,3 /* Convert leading/trailing zeros to bytes. */
|
|
|
ce426f |
+ add r3,r3,r0 /* Length until the match. */
|
|
|
ce426f |
+ cmpld r3,r4
|
|
|
ce426f |
+ blelr
|
|
|
ce426f |
+ mr r3,r4
|
|
|
ce426f |
blr
|
|
|
ce426f |
|
|
|
ce426f |
/* Deals with size <= 32. */
|
|
|
ce426f |
.align 4
|
|
|
ce426f |
L(small_range):
|
|
|
ce426f |
cmpldi r4,0
|
|
|
ce426f |
- beq L(zero)
|
|
|
ce426f |
+ beq L(end_max)
|
|
|
ce426f |
+
|
|
|
ce426f |
+ clrrdi r7,r7,3 /* Address of last doubleword. */
|
|
|
ce426f |
|
|
|
ce426f |
rlwinm r6,r3,3,26,28 /* Calculate padding. */
|
|
|
ce426f |
- ld r12,0(r8) /* Load word from memory. */
|
|
|
ce426f |
+ ld r12,0(r8) /* Load doubleword from memory. */
|
|
|
ce426f |
cmpb r10,r12,r0 /* Check for null bytes in DWORD1. */
|
|
|
ce426f |
+#ifdef __LITTLE_ENDIAN__
|
|
|
ce426f |
+ srd r10,r10,r6
|
|
|
ce426f |
+ sld r10,r10,r6
|
|
|
ce426f |
+#else
|
|
|
ce426f |
sld r10,r10,r6
|
|
|
ce426f |
srd r10,r10,r6
|
|
|
ce426f |
+#endif
|
|
|
ce426f |
cmpldi cr7,r10,0
|
|
|
ce426f |
bne cr7,L(done)
|
|
|
ce426f |
|
|
|
ce426f |
- addi r9,r8,8
|
|
|
ce426f |
- cmpld r9,r7
|
|
|
ce426f |
- bge L(end_max)
|
|
|
ce426f |
- b L(loop_small)
|
|
|
ce426f |
+ cmpld r8,r7
|
|
|
ce426f |
+ beq L(end_max)
|
|
|
ce426f |
|
|
|
ce426f |
.p2align 5
|
|
|
ce426f |
L(loop_small):
|
|
|
ce426f |
ldu r12,8(r8)
|
|
|
ce426f |
cmpb r10,r12,r0
|
|
|
ce426f |
- addi r9,r8,8
|
|
|
ce426f |
cmpldi cr6,r10,0
|
|
|
ce426f |
bne cr6,L(done)
|
|
|
ce426f |
- cmpld r9,r7
|
|
|
ce426f |
- bge L(end_max)
|
|
|
ce426f |
- b L(loop_small)
|
|
|
ce426f |
+ cmpld r8,r7
|
|
|
ce426f |
+ bne L(loop_small)
|
|
|
ce426f |
+ mr r3,r4
|
|
|
ce426f |
+ blr
|
|
|
ce426f |
+
|
|
|
ce426f |
END (BP_SYM (__strnlen))
|
|
|
ce426f |
weak_alias (BP_SYM (__strnlen), BP_SYM(strnlen))
|
|
|
ce426f |
libc_hidden_builtin_def (strnlen)
|