diff options
Diffstat (limited to 'sysdeps/powerpc/powerpc64/power8/strncpy.S')
| -rw-r--r-- | sysdeps/powerpc/powerpc64/power8/strncpy.S | 176 |
1 files changed, 95 insertions, 81 deletions
diff --git a/sysdeps/powerpc/powerpc64/power8/strncpy.S b/sysdeps/powerpc/powerpc64/power8/strncpy.S index 17c3afb5fe..e8c5c71f87 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) 2015-2016 Free Software Foundation, Inc. + Copyright (C) 2015-2018 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 @@ -19,11 +19,32 @@ #include <sysdep.h> #ifdef USE_AS_STPNCPY -# define FUNC_NAME __stpncpy +# ifndef STPNCPY +# define FUNC_NAME __stpncpy +# else +# define FUNC_NAME STPNCPY +# endif #else -# define FUNC_NAME strncpy +# ifndef STRNCPY +# define FUNC_NAME strncpy +# else +# define FUNC_NAME STRNCPY +# endif +#endif /* !USE_AS_STPNCPY */ + +#ifndef MEMSET +/* For builds without IFUNC support, local calls should be made to internal + GLIBC symbol (created by libc_hidden_builtin_def). */ +# ifdef SHARED +# define MEMSET_is_local +# define MEMSET __GI_memset +# else +# define MEMSET memset +# endif #endif +#define FRAMESIZE (FRAME_MIN_SIZE+48) + /* Implements the function char * [r3] strncpy (char *dest [r3], const char *src [r4], size_t n [r5]) @@ -40,7 +61,12 @@ 4k. */ .machine power7 -EALIGN (FUNC_NAME, 4, 0) +#ifdef MEMSET_is_local +ENTRY_TOCLESS (FUNC_NAME, 4) +#else +ENTRY (FUNC_NAME, 4) +#endif + CALL_MCOUNT 3 /* Check if the [src]+15 will cross a 4K page by checking if the bit indicating the page size changes. Basically: @@ -54,8 +80,7 @@ EALIGN (FUNC_NAME, 4, 0) addi r10,r4,16 rlwinm r9,r4,0,19,19 - /* Since it is a leaf function, save some non-volatile registers on the - protected/red zone. */ + /* Save some non-volatile registers on the stack. */ std r26,-48(r1) std r27,-40(r1) @@ -69,6 +94,14 @@ EALIGN (FUNC_NAME, 4, 0) std r30,-16(r1) std r31,-8(r1) + /* Update CFI. */ + cfi_offset(r26, -48) + cfi_offset(r27, -40) + cfi_offset(r28, -32) + cfi_offset(r29, -24) + cfi_offset(r30, -16) + cfi_offset(r31, -8) + beq cr7,L(unaligned_lt_16) rldicl r9,r4,0,61 subfic r8,r9,8 @@ -180,79 +213,66 @@ L(short_path_loop_end): ld r31,-8(r1) blr - /* This code pads the remainder dest with NULL bytes. The algorithm - calculate the remanining size and issues a doubleword unrolled - loops followed by a byte a byte set. */ + /* This code pads the remainder of dest with NULL bytes. The algorithm + calculates the remaining size and calls memset. */ .align 4 L(zero_pad_start): mr r5,r10 mr r9,r6 L(zero_pad_start_1): - srdi. r8,r5,r3 - mr r10,r9 -#ifdef USE_AS_STPNCPY - mr r3,r9 + /* At this point: + - r5 holds the number of bytes that still have to be written to + dest. + - r9 points to the position, in dest, where the first null byte + will be written. + The above statements are true both when control reaches this label + from a branch or when falling through the previous lines. */ +#ifndef USE_AS_STPNCPY + mr r30,r3 /* Save the return value of strncpy. */ +#endif + /* Prepare the call to memset. */ + mr r3,r9 /* Pointer to the area to be zero-filled. */ + li r4,0 /* Byte to be written (zero). */ + + /* We delayed the creation of the stack frame, as well as the saving of + the link register, because only at this point, we are sure that + doing so is actually needed. */ + + /* Save the link register. */ + mflr r0 + std r0,16(r1) + + /* Create the stack frame. */ + stdu r1,-FRAMESIZE(r1) + cfi_adjust_cfa_offset(FRAMESIZE) + cfi_offset(lr, 16) + + bl MEMSET +#ifndef MEMSET_is_local + nop #endif - beq- cr0,L(zero_pad_loop_b_start) - cmpldi cr7,r8,1 - li cr7,0 - std r7,0(r9) - beq cr7,L(zero_pad_loop_b_prepare) - addic. r8,r8,-2 - addi r10,r9,r16 - std r7,8(r9) - beq cr0,L(zero_pad_loop_dw_2) - std r7,16(r9) - li r9,0 - b L(zero_pad_loop_dw_1) - - .align 4 -L(zero_pad_loop_dw): - addi r10,r10,16 - std r9,-8(r10) - beq cr0,L(zero_pad_loop_dw_2) - std r9,0(r10) -L(zero_pad_loop_dw_1): - cmpldi cr7,r8,1 - std r9,0(r10) - addic. r8,r8,-2 - bne cr7,L(zero_pad_loop_dw) - addi r10,r10,8 -L(zero_pad_loop_dw_2): - rldicl r5,r5,0,61 -L(zero_pad_loop_b_start): - cmpdi cr7,r5,0 - addi r5,r5,-1 - addi r9,r10,-1 - add r10,r10,5 - subf r10,r9,r10 - li r8,0 - beq- cr7,L(short_path_loop_end) - - /* Write remaining 1-8 bytes. */ - .align 4 - addi r9,r9,1 - mtocrf 0x1,r10 - bf 29,4f - stw r8,0(r9) - addi r9,r9,4 - .align 4 -4: bf 30,2f - sth r8,0(r9) - addi r9,r9,2 + ld r0,FRAMESIZE+16(r1) - .align 4 -2: bf 31,1f - stb r8,0(r9) +#ifndef USE_AS_STPNCPY + mr r3,r30 /* Restore the return value of strncpy, i.e.: + dest. For stpncpy, the return value is the + same as return value of memset. */ +#endif - /* Restore non-volatile registers. */ -1: ld r26,-48(r1) - ld r27,-40(r1) - ld r28,-32(r1) - ld r29,-24(r1) - ld r30,-16(r1) - ld r31,-8(r1) + /* Restore non-volatile registers and return. */ + ld r26,FRAMESIZE-48(r1) + ld r27,FRAMESIZE-40(r1) + ld r28,FRAMESIZE-32(r1) + ld r29,FRAMESIZE-24(r1) + ld r30,FRAMESIZE-16(r1) + ld r31,FRAMESIZE-8(r1) + /* Restore the stack frame. */ + addi r1,r1,FRAMESIZE + cfi_adjust_cfa_offset(-FRAMESIZE) + /* Restore the link register. */ + mtlr r0 + cfi_restore(lr) blr /* The common case where [src]+16 will not cross a 4K page boundary. @@ -301,7 +321,7 @@ L(pagecross): #endif orc r9,r7,r9 /* Mask bits that are not part of the string. */ - li cr7,0 + li r7,0 cmpb r9,r9,r7 /* Check for null bytes in DWORD1. */ cmpdi cr7,r9,0 bne cr7,L(short_path_prepare_2) @@ -312,14 +332,14 @@ L(pagecross): /* For next checks we have aligned address, so we check for more three doublewords to make sure we can read 16 unaligned bytes to start the bulk copy with 16 aligned addresses. */ - ld cr7,8(r11) + ld r7,8(r11) cmpb r9,r7,r9 cmpdi cr7,r9,0 bne cr7,L(short_path_prepare_2) - addi cr7,r8,-8 + addi r7,r8,-8 cmpldi cr7,r7,8 ble cr7,L(short_path_prepare_2) - ld cr7,16(r11) + ld r7,16(r11) cmpb r9,r7,r9 cmpdi cr7,r9,0 bne cr7,L(short_path_prepare_2) @@ -443,18 +463,12 @@ L(short_path_prepare_2_3): mr r4,r28 mr r9,r29 b L(short_path_2) -L(zero_pad_loop_b_prepare): - addi r10,r9,8 - rldicl r5,r5,0,61 - b L(zero_pad_loop_b_start) L(zero_pad_start_prepare_1): mr r5,r6 mr r9,r8 b L(zero_pad_start_1) END (FUNC_NAME) -#ifdef USE_AS_STPNCPY -libc_hidden_def (__stpncpy) -#else +#ifndef USE_AS_STPNCPY libc_hidden_builtin_def (strncpy) #endif |