diff options
Diffstat (limited to 'REORG.TODO/sysdeps/powerpc/powerpc64/power7/memchr.S')
| -rw-r--r-- | REORG.TODO/sysdeps/powerpc/powerpc64/power7/memchr.S | 199 |
1 files changed, 199 insertions, 0 deletions
diff --git a/REORG.TODO/sysdeps/powerpc/powerpc64/power7/memchr.S b/REORG.TODO/sysdeps/powerpc/powerpc64/power7/memchr.S new file mode 100644 index 0000000000..5e9707aa02 --- /dev/null +++ b/REORG.TODO/sysdeps/powerpc/powerpc64/power7/memchr.S @@ -0,0 +1,199 @@ +/* Optimized memchr 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] memchr (char *s [r3], int byte [r4], int size [r5]) */ + +#ifndef MEMCHR +# define MEMCHR __memchr +#endif + .machine power7 +ENTRY (MEMCHR) + CALL_MCOUNT 3 + dcbt 0,r3 + clrrdi r8,r3,3 + insrdi r4,r4,8,48 + + /* Calculate the last acceptable address and check for possible + addition overflow by using satured math: + r7 = r3 + r5 + r7 |= -(r7 < x) */ + add r7,r3,r5 + subfc r6,r3,r7 + subfe r9,r9,r9 + extsw r6,r9 + or r7,r7,r6 + + insrdi r4,r4,16,32 + cmpldi r5,32 + li r9, -1 + rlwinm r6,r3,3,26,28 /* Calculate padding. */ + insrdi r4,r4,32,0 + addi r7,r7,-1 +#ifdef __LITTLE_ENDIAN__ + sld r9,r9,r6 +#else + srd r9,r9,r6 +#endif + ble L(small_range) + + ld r12,0(r8) /* Load doubleword from memory. */ + cmpb r3,r12,r4 /* Check for BYTEs in DWORD1. */ + and r3,r3,r9 + clrldi r5,r7,61 /* Byte count - 1 in last dword. */ + clrrdi r7,r7,3 /* Address of last doubleword. */ + cmpldi cr7,r3,0 /* Does r3 indicate we got a hit? */ + 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. */ + bt 28,L(loop_setup) + + /* Handle DWORD2 of pair. */ + ldu r12,8(r8) + 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 last byte of the string, ie. the dword at + (s + size - 1) & ~7, or r7. 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 = (r7 - r8) / 16 */ + sub r6,r7,r8 + srdi r6,r6,4 /* Number of loop iterations. */ + mtctr r6 /* Setup the counter. */ + + /* Main loop to look for BYTE in the string. Since + it's a small loop (8 instructions), align it to 32-bytes. */ + .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. */ + ld r12,8(r8) + ldu r11,16(r8) + cmpb r3,r12,r4 + cmpb r9,r11,r4 + or r6,r9,r3 /* Merge everything in one doubleword. */ + cmpldi cr7,r6,0 + bne cr7,L(found) + bdnz L(loop) + + /* We may have one more dword to read. */ + cmpld r8,r7 + beqlr + + ldu r12,8(r8) + cmpb r3,r12,r4 + cmpldi cr6,r3,0 + bne cr6,L(done) + blr + + .align 4 +L(found): + /* OK, one (or both) of the doublewords contains BYTE. Check + the first doubleword and decrement the address in case the first + doubleword really contains BYTE. */ + cmpldi cr6,r3,0 + addi r8,r8,-8 + bne cr6,L(done) + + /* BYTE must be in the second doubleword. 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 + doubleword from the string. Use that to calculate the pointer. + We need to make sure BYTE is *before* the end of the range. */ +L(done): +#ifdef __LITTLE_ENDIAN__ + addi r0,r3,-1 + andc r0,r0,r3 + popcntd r0,r0 /* Count trailing zeros. */ +#else + cntlzd r0,r3 /* Count leading zeros before the match. */ +#endif + cmpld r8,r7 /* Are we on the last dword? */ + srdi r0,r0,3 /* Convert leading/trailing zeros to bytes. */ + add r3,r8,r0 + cmpld cr7,r0,r5 /* If on the last dword, check byte offset. */ + bnelr + blelr 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) + ld r12,0(r8) /* Load word from memory. */ + cmpb r3,r12,r4 /* Check for BYTE in DWORD1. */ + and r3,r3,r9 + cmpldi cr7,r3,0 + clrldi r5,r7,61 /* Byte count - 1 in last dword. */ + clrrdi r7,r7,3 /* Address of last doubleword. */ + cmpld r8,r7 /* Are we done already? */ + bne cr7,L(done) + beqlr + + ldu r12,8(r8) + cmpb r3,r12,r4 + cmpldi cr6,r3,0 + cmpld r8,r7 + bne cr6,L(done) /* Found something. */ + beqlr /* Hit end of string (length). */ + + ldu r12,8(r8) + cmpb r3,r12,r4 + cmpldi cr6,r3,0 + cmpld r8,r7 + bne cr6,L(done) + beqlr + + ldu r12,8(r8) + cmpb r3,r12,r4 + cmpldi cr6,r3,0 + cmpld r8,r7 + bne cr6,L(done) + beqlr + + ldu r12,8(r8) + cmpb r3,r12,r4 + cmpldi cr6,r3,0 + bne cr6,L(done) + blr + +END (MEMCHR) +weak_alias (__memchr, memchr) +libc_hidden_builtin_def (memchr) |