aboutsummaryrefslogtreecommitdiff
path: root/sysdeps/powerpc/powerpc64/power7
diff options
context:
space:
mode:
authorAdhemerval Zanella <azanella@linux.vnet.ibm.com>2014-12-23 13:39:23 -0500
committerAdhemerval Zanella <azanella@linux.vnet.ibm.com>2015-01-13 11:28:40 -0500
commit9f2f36e5a91c2ce6edba5415e176155eb1008ae1 (patch)
treef772f8f5005de041205a3f1ed0317ce5bf992fc6 /sysdeps/powerpc/powerpc64/power7
parent94c9680945369d63ef9ed266a29f28ebaaaeb5ce (diff)
downloadglibc-9f2f36e5a91c2ce6edba5415e176155eb1008ae1.zip
glibc-9f2f36e5a91c2ce6edba5415e176155eb1008ae1.tar.gz
glibc-9f2f36e5a91c2ce6edba5415e176155eb1008ae1.tar.bz2
powerpc: Optimized strncat for POWER7/PPC64
With 3eb38795dbbbd816 (Simplify strncat) the generic algorithms uses strlen, strnlen, and memcpy. This is faster than POWER7 current implementation, especially for unaligned strings (where POWER7 code uses byte-byte operations). This patch removes the assembly implementation and uses a multiarch specialization based on default algorithm calling optimized POWER7 symbols.
Diffstat (limited to 'sysdeps/powerpc/powerpc64/power7')
-rw-r--r--sysdeps/powerpc/powerpc64/power7/strncat.S228
1 files changed, 0 insertions, 228 deletions
diff --git a/sysdeps/powerpc/powerpc64/power7/strncat.S b/sysdeps/powerpc/powerpc64/power7/strncat.S
deleted file mode 100644
index 05502ac..0000000
--- a/sysdeps/powerpc/powerpc64/power7/strncat.S
+++ /dev/null
@@ -1,228 +0,0 @@
-/* Optimized strncat implementation for PowerPC64/POWER7.
-
- Copyright (C) 2014-2015 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
- modify it under the terms of the GNU Lesser General Public
- License as published by the Free Software Foundation; either
- version 2.1 of the License, or (at your option) any later version.
-
- The GNU C Library is distributed in the hope that it will be useful,
- but WITHOUT ANY WARRANTY; without even the implied warranty of
- MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
- Lesser General Public License for more details.
-
- You should have received a copy of the GNU Lesser General Public
- License along with the GNU C Library; if not, see
- <http://www.gnu.org/licenses/>. */
-
-/* The algorithm is as follows for aligned memory access :
-
- if address of s2 is divisible by 0x7UL,
- perform aligned doubleword catenation
- else
- perform unaligned catenation
-
- The aligned comparison are made using cmpb instructions. */
-
-/* char* [r3] strncat (const char *s1 [r3],
- const char *s2 [r4],
- size_t size [r5]) */
-
-#include <sysdep.h>
-
-#ifndef STRNCAT
-# undef strncat
-# define STRNCAT strncat
-#endif
-
-#ifndef STRLEN
-/* For builds with no IFUNC support, local calls should be made to internal
- GLIBC symbol (created by libc_hidden_builtin_def). */
-# ifdef SHARED
-# define STRLEN __GI_strlen
-# else
-# define STRLEN strlen
-# endif
-#endif
-
-#define FRAMESIZE (FRAME_MIN_SIZE+32)
-
- .machine power7
-EALIGN(STRNCAT, 4, 0)
- CALL_MCOUNT 3
-
- mflr r0 /* Load link register LR to r0. */
-
-/* We shall use r29, r30 and r31 non volatile register for retention.
- Save all the callee registers in the GPR save area. */
- std r29, -24(r1) /* Save callers register r29. */
- std r30, -16(r1) /* Save callers register r30. */
- std r31, -8(r1) /* Save callers register r31. */
-
- std r0, 16(r1) /* Store the link register. */
- stdu r1, -FRAMESIZE(r1) /* Create the stack frame. */
-
-/* Improve performance with CPU pre-fetch. */
- dcbt 0, r3 /* Pre-fetch str to avoid cache
- miss. */
- dcbt 0, r4 /* Pre-fetch accept to avoid cache
- miss. */
-
- mr. r29, r5 /* Save "n" in r29. */
- mr r30, r3 /* Save "s1" in r30 from r3. */
- beq cr0,L(done)
-
- mr r31, r4 /* Save "s2" in r31 from r4. */
- bl STRLEN /* Call optimized strlen on s1; goto
- end of s1. */
- nop
- cmpldi cr7, r29, 7 /* If s2 is <=7 process
- byte-by-byte. */
- add r3, r30, r3 /* Grab the last character of s1. */
- bgt cr7,L(alignment) /* Process by aligned strings. */
-
- cmpldi cr7, r29, 3 /* If n is >= 4, we can
- byte-unroll. */
- addi r9, r3, -1 /* Make "s1" point before next
- character, increment when read. */
- bgt cr7, L(bytes_unroll) /* Process each byte. */
-
-L(byte_by_byte):
- lbz r10, 0(r31)
- addi r8, r9, 1
- cmpdi cr7, r10, 0 /* Check for NULL in "s2". */
- stb r10, 1(r9)
- beq cr7, L(done)
- add r9, r9, r29
- subf r9, r8, r9
- addi r9, r9, 1
- mtctr r9
- b L(branch2)
- .p2align 4
-L(branch1):
- lbzu r10, 1(r31)
- cmpdi cr7, r10, 0
- stbu r10, 1(r8)
- beq cr7,L(done)
-L(branch2):
- mr r9, r8
- bdnz L(branch1)
- beq cr7,L(done)
-L(nullTerminate):
- li r10, 0 /* Load NULL for termination. */
- stb r10, 1(r9) /* Append or terminate s1 with
- NULL. */
- .p2align 4 /* A small section here. */
-L(done): /* We return now. */
- addi r1, r1, FRAMESIZE /* Restore stack pointer. */
- mr r3, r30 /* Set the return value length of
- string. */
- ld r0, 16(r1) /* Read the saved link register. */
- ld r29, -24(r1) /* Restore save register r29. */
- ld r30, -16(r1) /* Restore save register r30. */
- ld r31, -8(r1) /* Restore save register r31. */
- mtlr r0 /* Restore link register. */
- blr /* Branch to link register. */
-
- .p2align 4
-L(alignment):
- rldicl. r9, r31, 0, 61 /* Check if s2 is 8byte aligned */
- beq cr0,L(dwordAligned)
-
- .p2align 4
-/* Unaligned bytes in string, so process byte by byte.
- POWER7 has performance gains over loop unroll. */
-L(bytes_unroll):
- addi r9, r3, -1
- srdi r10, r29, 2
- mtctr r10
- b L(L10)
- .p2align 4
-L(L44):
- lbz r10, 1(r31) /* Load byte. */
- cmpdi cr7, r10, 0 /* Compare ; if byte not zero,
- continue. */
- stb r10, 2(r9) /* Store byte */
- beq cr7, L(done)
- addi r31, r31, 4
-
- lbz r10, -2(r31) /* Perform loop unroll here on byte
- load and store. */
- cmpdi cr7, r10, 0
- stb r10, 3(r9)
- beq cr7, L(done)
-
- lbz r10, -1(r31) /* Loop unroll here. */
- cmpdi cr7, r10, 0
- stbu r10, 4(r9)
- beq cr7, L(done)
-
- bdz L(leftNbytes)
-
-L(L10):
- lbz r10, 0(r31) /* Loop unroll here. */
- cmpdi cr7, r10, 0
- stb r10, 1(r9)
- bne cr7,L(L44)
- b L(done)
- .p2align 4
-/* If s2 is double word aligned, we load and store double word. */
-L(dwordAligned):
-/* read, write 8 bytes at a time */
- srdi r8, r29, 3 /* Compute count for CTR to loop;
- count = n/8. */
- li r7, 0 /* Load r7 with NULL. */
- li r10, 0 /* Load r10 with MASK '0'. */
-
- mtctr r8 /* Move count to CTR. */
-L(loop8):
- ld r9, 0(r31) /* Read double word from s2. */
- cmpb r6, r9, r10 /* Compare bytes in s2 we read
- just now. */
- cmpdi r6, 0 /* If cmpb returned NULL,
- we continue. */
- bne+ L(a8)
- std r9, 0(r3) /* Append double word from s2
- with s1. */
- addi r3, r3, 8 /* Increment s1. */
- addi r31, r31, 8 /* Increment s2. */
- subi r29, r29, 8 /* Decrement count by 8. */
- bdnz L(loop8) /* Continue until "count" is
- non zero. */
-
-L(a8):
- cmpdi r29, 0 /* If "n" is already zero, we skip. */
- beq+ L(align8align)
-
- mtctr r29 /* Process left over bytes in "n". */
-L(unaligned0):
- lbz r9, 0(r31) /* Read a byte from s2. */
- cmpw r9, r7 /* If byte is NULL, we stop here . */
- beq+ L(align8align) /* Skip processing further if NULL. */
- stb r9, 0(r3) /* If not NULL, store byte into s1. */
- addi r3, r3, 1 /* Increment s1 by 1. */
- addi r31, r31, 1 /* Increment s2 by 1. */
- bdnz L(unaligned0) /* Decrement counter "n" and loop
- until non zero. */
-L(align8align):
- stb r7, 0(r3) /* Terminate s1 with NULL. */
-
- addi r1, r1, FRAMESIZE /* Restore stack pointer. */
- mr r3, r30 /* Set the return value, length of
- string. */
- ld r0, 16(r1) /* Read the saved link register. */
- ld r29, -24(r1) /* Restore save register r29. */
- ld r30, -16(r1) /* Restore save register r30. */
- ld r31, -8(r1) /* Restore save register r31. */
- mtlr r0 /* Restore link register. */
- blr /* Branch to link register */
-
- .p2align 4
-L(leftNbytes):
- rldicl. r29, r29, 0, 62 /* Check if n>0 and n < 4 bytes. */
- bne cr0,L(byte_by_byte) /* Process bytes one by one. */
- b L(nullTerminate) /* Now, finish catenation with
- NULL termination. */
-END(STRNCAT)