diff options
Diffstat (limited to 'sysdeps/powerpc/powerpc64/power8/strncpy.S')
| -rw-r--r-- | sysdeps/powerpc/powerpc64/power8/strncpy.S | 54 |
1 files changed, 45 insertions, 9 deletions
diff --git a/sysdeps/powerpc/powerpc64/power8/strncpy.S b/sysdeps/powerpc/powerpc64/power8/strncpy.S index 5fda953526..17c3afb5fe 100644 --- a/sysdeps/powerpc/powerpc64/power8/strncpy.S +++ b/sysdeps/powerpc/powerpc64/power8/strncpy.S @@ -1,5 +1,5 @@ /* Optimized strncpy/stpncpy implementation for PowerPC64/POWER8. - Copyright (C) 2015 Free Software Foundation, Inc. + Copyright (C) 2015-2016 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 @@ -64,7 +64,7 @@ EALIGN (FUNC_NAME, 4, 0) std r28,-32(r1) std r29,-24(r1) - cmpld r7,r9,r8 + cmpld cr7,r9,r8 std r30,-16(r1) std r31,-8(r1) @@ -82,44 +82,80 @@ EALIGN (FUNC_NAME, 4, 0) L(short_path): mr r9,r3 L(short_path_1): + /* Return if there are no more bytes to be written. */ cmpdi cr7,r5,0 beq cr7,L(short_path_loop_end_1) L(short_path_2): + /* Copy one char from src (r4) and write it to dest (r9). If it is the + end-of-string, start the null padding. Continue, otherwise. */ lbz r10,0(r4) cmpdi cr7,r10,0 stb r10,0(r9) beq cr7,L(zero_pad_start_1) + /* If there are no more bytes to be written, return. */ cmpdi cr0,r5,1 addi r8,r9,1 addi r6,r5,-1 beq cr0,L(short_path_loop_end_0) + /* Copy another char from src (r4) to dest (r9). Check again if it is + the end-of-string. If so, start the null padding. */ lbz r10,1(r4) cmpdi cr7,r10,0 stb r10,1(r9) beq cr7,L(zero_pad_start_prepare_1) + /* Eagerly decrement r5 by 3, which is the number of bytes already + written, plus one write that will be performed later on. */ addi r10,r5,-3 b L(short_path_loop_1) .align 4 L(short_path_loop): + /* At this point, the induction variable, r5, as well as the pointers + to dest and src (r9 and r4, respectivelly) have been updated. + + Note: The registers r7 and r10 are induction variables derived from + r5. They are used to determine if the total number of writes has + been reached at every other write. + + Copy one char from src (r4) and write it to dest (r9). If it is the + end-of-string, start the null padding. Continue, otherwise. */ lbz r8,0(r4) addi r7,r10,-2 cmpdi cr5,r8,0 stb r8,0(r9) beq cr5,L(zero_pad_start_1) - beq r7,L(short_path_loop_end_0) + beq cr7,L(short_path_loop_end_0) + /* Copy another char from src (r4) to dest (r9). Check again if it is + the end-of-string. If so, start the null padding. */ lbz r8,1(r4) cmpdi cr7,r8,0 stb r8,1(r9) beq cr7,L(zero_pad_start) mr r10,r7 L(short_path_loop_1): + /* This block is reached after two chars have been already written to + dest. Nevertheless, r5 (the induction variable), r9 (the pointer to + dest), and r4 (the pointer to src) have not yet been updated. + + At this point: + r5 holds the count of bytes yet to be written plus 2. + r9 points to the last two chars that were already written to dest. + r4 points to the last two chars that were already copied from src. + + The algorithm continues by decrementing r5, the induction variable, + so that it reflects the last two writes. The pointers to dest (r9) + and to src (r4) are increment by two, for the same reason. + + Note: Register r10 is another induction variable, derived from r5, + which determines if the total number of writes has been reached. */ addic. r5,r5,-2 addi r9,r9,2 - cmpdi cr7,r10,0 + cmpdi cr7,r10,0 /* Eagerly check if the next write is the last. */ addi r4,r4,2 addi r6,r9,1 - bne cr0,L(short_path_loop) + bne cr0,L(short_path_loop) /* Check if the total number of writes + has been reached at every other + write. */ #ifdef USE_AS_STPNCPY mr r3,r9 b L(short_path_loop_end) @@ -234,7 +270,7 @@ L(unaligned_lt_16): bne cr7,L(short_path_prepare_2) addi r6,r5,-8 std r7,0(r3) - addi r9,r3,r8 + addi r9,r3,8 cmpldi cr7,r6,7 addi r7,r4,8 ble cr7,L(short_path_prepare_1_1) @@ -288,11 +324,11 @@ L(pagecross): cmpdi cr7,r9,0 bne cr7,L(short_path_prepare_2) addi r8,r8,-16 - cmpldi r7,r8,8 + cmpldi cr7,r8,8 ble cr7,L(short_path_prepare_2) ld r8,24(r11) cmpb r9,r8,r9 - cmpdi r7,r9,0 + cmpdi cr7,r9,0 bne cr7,L(short_path_prepare_2) /* No null byte found in the 32 bytes readed and length not reached, @@ -367,7 +403,7 @@ L(loop_16b): cmpb r7,r0,r30 or. r7,r8,r7 addi r12,r12,-32 - cmpldi r7,r12,15 + cmpldi cr7,r12,15 addi r11,r11,32 bne cr0,L(short_path_2) std r10,16(r6) |