diff options
Diffstat (limited to 'libc/string/xtensa/memcpy.S')
| -rw-r--r-- | libc/string/xtensa/memcpy.S | 297 |
1 files changed, 297 insertions, 0 deletions
diff --git a/libc/string/xtensa/memcpy.S b/libc/string/xtensa/memcpy.S new file mode 100644 index 000000000..19f3a6818 --- /dev/null +++ b/libc/string/xtensa/memcpy.S @@ -0,0 +1,297 @@ +/* Optimized memcpy for Xtensa. + Copyright (C) 2001, 2007 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, write to the Free + Software Foundation, Inc., 51 Franklin Street - Fifth Floor, + Boston, MA 02110-1301, USA. */ + +#include "../../sysdeps/linux/xtensa/sysdep.h" +#include <bits/xtensa-config.h> + + .macro src_b r, w0, w1 +#ifdef __XTENSA_EB__ + src \r, \w0, \w1 +#else + src \r, \w1, \w0 +#endif + .endm + + .macro ssa8 r +#ifdef __XTENSA_EB__ + ssa8b \r +#else + ssa8l \r +#endif + .endm + +/* If the Xtensa Unaligned Load Exception option is not used, this + code can run a few cycles faster by relying on the low address bits + being ignored. However, if the code is then run with an Xtensa ISS + client that checks for unaligned accesses, it will produce a lot of + warning messages. Set this flag to disable the use of unaligned + accesses and keep the ISS happy. */ + +#define UNALIGNED_ADDRESSES_CHECKED 1 + +/* Do not use .literal_position in the ENTRY macro. */ +#undef LITERAL_POSITION +#define LITERAL_POSITION + + +/* void *memcpy (void *dst, const void *src, size_t len) + + The algorithm is as follows: + + If the destination is unaligned, align it by conditionally + copying 1- and/or 2-byte pieces. + + If the source is aligned, copy 16 bytes with a loop, and then finish up + with 8, 4, 2, and 1-byte copies conditional on the length. + + Else (if source is unaligned), do the same, but use SRC to align the + source data. + + This code tries to use fall-through branches for the common + case of aligned source and destination and multiple of 4 (or 8) length. */ + + +/* Byte by byte copy. */ + + .text + .align 4 + .literal_position +__memcpy_aux: + + /* Skip a byte to get 1 mod 4 alignment for LOOPNEZ + (0 mod 4 alignment for LBEG). */ + .byte 0 + +.Lbytecopy: +#if XCHAL_HAVE_LOOPS + loopnez a4, 2f +#else + beqz a4, 2f + add a7, a3, a4 // a7 = end address for source +#endif +1: l8ui a6, a3, 0 + addi a3, a3, 1 + s8i a6, a5, 0 + addi a5, a5, 1 +#if !XCHAL_HAVE_LOOPS + blt a3, a7, 1b +#endif +2: retw + + +/* Destination is unaligned. */ + + .align 4 +.Ldst1mod2: // dst is only byte aligned + + /* Do short copies byte-by-byte. */ + _bltui a4, 7, .Lbytecopy + + /* Copy 1 byte. */ + l8ui a6, a3, 0 + addi a3, a3, 1 + addi a4, a4, -1 + s8i a6, a5, 0 + addi a5, a5, 1 + + /* Return to main algorithm if dst is now aligned. */ + _bbci.l a5, 1, .Ldstaligned + +.Ldst2mod4: // dst has 16-bit alignment + + /* Do short copies byte-by-byte. */ + _bltui a4, 6, .Lbytecopy + + /* Copy 2 bytes. */ + l8ui a6, a3, 0 + l8ui a7, a3, 1 + addi a3, a3, 2 + addi a4, a4, -2 + s8i a6, a5, 0 + s8i a7, a5, 1 + addi a5, a5, 2 + + /* dst is now aligned; return to main algorithm. */ + j .Ldstaligned + + +ENTRY (memcpy) + /* a2 = dst, a3 = src, a4 = len */ + + mov a5, a2 // copy dst so that a2 is return value + _bbsi.l a2, 0, .Ldst1mod2 + _bbsi.l a2, 1, .Ldst2mod4 +.Ldstaligned: + + /* Get number of loop iterations with 16B per iteration. */ + srli a7, a4, 4 + + /* Check if source is aligned. */ + movi a8, 3 + _bany a3, a8, .Lsrcunaligned + + /* Destination and source are word-aligned, use word copy. */ +#if XCHAL_HAVE_LOOPS + loopnez a7, 2f +#else + beqz a7, 2f + slli a8, a7, 4 + add a8, a8, a3 // a8 = end of last 16B source chunk +#endif +1: l32i a6, a3, 0 + l32i a7, a3, 4 + s32i a6, a5, 0 + l32i a6, a3, 8 + s32i a7, a5, 4 + l32i a7, a3, 12 + s32i a6, a5, 8 + addi a3, a3, 16 + s32i a7, a5, 12 + addi a5, a5, 16 +#if !XCHAL_HAVE_LOOPS + blt a3, a8, 1b +#endif + + /* Copy any leftover pieces smaller than 16B. */ +2: bbci.l a4, 3, 3f + + /* Copy 8 bytes. */ + l32i a6, a3, 0 + l32i a7, a3, 4 + addi a3, a3, 8 + s32i a6, a5, 0 + s32i a7, a5, 4 + addi a5, a5, 8 + +3: bbsi.l a4, 2, 4f + bbsi.l a4, 1, 5f + bbsi.l a4, 0, 6f + retw + + /* Copy 4 bytes. */ +4: l32i a6, a3, 0 + addi a3, a3, 4 + s32i a6, a5, 0 + addi a5, a5, 4 + bbsi.l a4, 1, 5f + bbsi.l a4, 0, 6f + retw + + /* Copy 2 bytes. */ +5: l16ui a6, a3, 0 + addi a3, a3, 2 + s16i a6, a5, 0 + addi a5, a5, 2 + bbsi.l a4, 0, 6f + retw + + /* Copy 1 byte. */ +6: l8ui a6, a3, 0 + s8i a6, a5, 0 + +.Ldone: + retw + + +/* Destination is aligned; source is unaligned. */ + + .align 4 +.Lsrcunaligned: + /* Avoid loading anything for zero-length copies. */ + _beqz a4, .Ldone + + /* Copy 16 bytes per iteration for word-aligned dst and + unaligned src. */ + ssa8 a3 // set shift amount from byte offset +#if UNALIGNED_ADDRESSES_CHECKED + and a11, a3, a8 // save unalignment offset for below + sub a3, a3, a11 // align a3 +#endif + l32i a6, a3, 0 // load first word +#if XCHAL_HAVE_LOOPS + loopnez a7, 2f +#else + beqz a7, 2f + slli a10, a7, 4 + add a10, a10, a3 // a10 = end of last 16B source chunk +#endif +1: l32i a7, a3, 4 + l32i a8, a3, 8 + src_b a6, a6, a7 + s32i a6, a5, 0 + l32i a9, a3, 12 + src_b a7, a7, a8 + s32i a7, a5, 4 + l32i a6, a3, 16 + src_b a8, a8, a9 + s32i a8, a5, 8 + addi a3, a3, 16 + src_b a9, a9, a6 + s32i a9, a5, 12 + addi a5, a5, 16 +#if !XCHAL_HAVE_LOOPS + blt a3, a10, 1b +#endif + +2: bbci.l a4, 3, 3f + + /* Copy 8 bytes. */ + l32i a7, a3, 4 + l32i a8, a3, 8 + src_b a6, a6, a7 + s32i a6, a5, 0 + addi a3, a3, 8 + src_b a7, a7, a8 + s32i a7, a5, 4 + addi a5, a5, 8 + mov a6, a8 + +3: bbci.l a4, 2, 4f + + /* Copy 4 bytes. */ + l32i a7, a3, 4 + addi a3, a3, 4 + src_b a6, a6, a7 + s32i a6, a5, 0 + addi a5, a5, 4 + mov a6, a7 +4: +#if UNALIGNED_ADDRESSES_CHECKED + add a3, a3, a11 // readjust a3 with correct misalignment +#endif + bbsi.l a4, 1, 5f + bbsi.l a4, 0, 6f + retw + + /* Copy 2 bytes. */ +5: l8ui a6, a3, 0 + l8ui a7, a3, 1 + addi a3, a3, 2 + s8i a6, a5, 0 + s8i a7, a5, 1 + addi a5, a5, 2 + bbsi.l a4, 0, 6f + retw + + /* Copy 1 byte. */ +6: l8ui a6, a3, 0 + s8i a6, a5, 0 + retw + +libc_hidden_def (memcpy) |