diff options
Diffstat (limited to 'REORG.TODO/sysdeps/powerpc/powerpc64/power7/memrchr.S')
| -rw-r--r-- | REORG.TODO/sysdeps/powerpc/powerpc64/power7/memrchr.S | 201 |
1 files changed, 201 insertions, 0 deletions
diff --git a/REORG.TODO/sysdeps/powerpc/powerpc64/power7/memrchr.S b/REORG.TODO/sysdeps/powerpc/powerpc64/power7/memrchr.S new file mode 100644 index 0000000000..4276768915 --- /dev/null +++ b/REORG.TODO/sysdeps/powerpc/powerpc64/power7/memrchr.S @@ -0,0 +1,201 @@ +/* Optimized memrchr implementation for PowerPC64/POWER7 using cmpb insn. + Copyright (C) 2010-2017 Free Software Foundation, Inc. + Contributed by Luis Machado <luisgpm@br.ibm.com>. + 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/>. */ + +#include <sysdep.h> + +/* int [r3] memrchr (char *s [r3], int byte [r4], int size [r5]) */ + +#ifndef MEMRCHR +# define MEMRCHR __memrchr +#endif + .machine power7 +ENTRY (MEMRCHR) + CALL_MCOUNT 3 + add r7,r3,r5 /* Calculate the last acceptable address. */ + neg r0,r7 + addi r7,r7,-1 + mr r10,r3 + clrrdi r6,r7,7 + li r9,3<<5 + dcbt r9,r6,8 /* Stream hint, decreasing addresses. */ + + /* Replicate BYTE to doubleword. */ + insrdi r4,r4,8,48 + insrdi r4,r4,16,32 + insrdi r4,r4,32,0 + li r6,-8 + li r9,-1 + rlwinm r0,r0,3,26,28 /* Calculate padding. */ + clrrdi r8,r7,3 + srd r9,r9,r0 + cmpldi r5,32 + clrrdi r0,r10,3 + ble L(small_range) + +#ifdef __LITTLE_ENDIAN__ + ldx r12,0,r8 +#else + ldbrx r12,0,r8 /* Load reversed doubleword from memory. */ +#endif + cmpb r3,r12,r4 /* Check for BYTE in DWORD1. */ + and r3,r3,r9 + cmpldi cr7,r3,0 /* If r3 == 0, no BYTEs have been found. */ + bne cr7,L(done) + + mtcrf 0x01,r8 + /* Are we now aligned to a quadword boundary? If so, skip to + the main loop. Otherwise, go through the alignment code. */ + bf 28,L(loop_setup) + + /* Handle DWORD2 of pair. */ +#ifdef __LITTLE_ENDIAN__ + ldx r12,r8,r6 +#else + ldbrx r12,r8,r6 +#endif + addi r8,r8,-8 + cmpb r3,r12,r4 + cmpldi cr7,r3,0 + bne cr7,L(done) + +L(loop_setup): + /* The last dword we want to read in the loop below is the one + containing the first byte of the string, ie. the dword at + s & ~7, or r0. The first dword read is at r8 - 8, we + read 2 * cnt dwords, so the last dword read will be at + r8 - 8 - 16 * cnt + 8. Solving for cnt gives + cnt = (r8 - r0) / 16 */ + sub r5,r8,r0 + addi r8,r8,-8 + srdi r9,r5,4 /* Number of loop iterations. */ + mtctr r9 /* Setup the counter. */ + + /* Main loop to look for BYTE backwards in the string. + FIXME: Investigate whether 32 byte align helps with this + 9 instruction loop. */ + .align 5 +L(loop): + /* Load two doublewords, compare and merge in a + single register for speed. This is an attempt + to speed up the byte-checking process for bigger strings. */ + +#ifdef __LITTLE_ENDIAN__ + ldx r12,0,r8 + ldx r11,r8,r6 +#else + ldbrx r12,0,r8 + ldbrx r11,r8,r6 +#endif + cmpb r3,r12,r4 + cmpb r9,r11,r4 + or r5,r9,r3 /* Merge everything in one doubleword. */ + cmpldi cr7,r5,0 + bne cr7,L(found) + addi r8,r8,-16 + bdnz L(loop) + + /* We may have one more word to read. */ + cmpld r8,r0 + bnelr + +#ifdef __LITTLE_ENDIAN__ + ldx r12,0,r8 +#else + ldbrx r12,0,r8 +#endif + cmpb r3,r12,r4 + cmpldi cr7,r3,0 + bne cr7,L(done) + blr + + .align 4 +L(found): + /* OK, one (or both) of the dwords contains BYTE. Check + the first dword. */ + cmpldi cr6,r3,0 + bne cr6,L(done) + + /* BYTE must be in the second word. Adjust the address + again and move the result of cmpb to r3 so we can calculate the + pointer. */ + + mr r3,r9 + addi r8,r8,-8 + + /* r3 has the output of the cmpb instruction, that is, it contains + 0xff in the same position as BYTE in the original + word from the string. Use that to calculate the pointer. + We need to make sure BYTE is *before* the end of the + range. */ +L(done): + cntlzd r9,r3 /* Count leading zeros before the match. */ + cmpld r8,r0 /* Are we on the last word? */ + srdi r6,r9,3 /* Convert leading zeros to bytes. */ + addi r0,r6,-7 + sub r3,r8,r0 + cmpld cr7,r3,r10 + bnelr + bgelr cr7 + li r3,0 + blr + + .align 4 +L(null): + li r3,0 + blr + +/* Deals with size <= 32. */ + .align 4 +L(small_range): + cmpldi r5,0 + beq L(null) + +#ifdef __LITTLE_ENDIAN__ + ldx r12,0,r8 +#else + ldbrx r12,0,r8 /* Load reversed doubleword from memory. */ +#endif + cmpb r3,r12,r4 /* Check for BYTE in DWORD1. */ + and r3,r3,r9 + cmpldi cr7,r3,0 + bne cr7,L(done) + + /* Are we done already? */ + cmpld r8,r0 + addi r8,r8,-8 + beqlr + + .align 5 +L(loop_small): +#ifdef __LITTLE_ENDIAN__ + ldx r12,0,r8 +#else + ldbrx r12,0,r8 +#endif + cmpb r3,r12,r4 + cmpld r8,r0 + cmpldi cr7,r3,0 + bne cr7,L(done) + addi r8,r8,-8 + bne L(loop_small) + blr + +END (MEMRCHR) +weak_alias (__memrchr, memrchr) +libc_hidden_builtin_def (memrchr) |