1 /* Optimized memcpy implementation for PowerPC64/POWER7.
2 Copyright (C) 2010-2013 Free Software Foundation, Inc.
3 Contributed by Luis Machado <luisgpm@br.ibm.com>.
4 This file is part of the GNU C Library.
6 The GNU C Library is free software; you can redistribute it and/or
7 modify it under the terms of the GNU Lesser General Public
8 License as published by the Free Software Foundation; either
9 version 2.1 of the License, or (at your option) any later version.
11 The GNU C Library is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 Lesser General Public License for more details.
16 You should have received a copy of the GNU Lesser General Public
17 License along with the GNU C Library; if not, see
18 <http://www.gnu.org/licenses/>. */
23 /* __ptr_t [r3] memcpy (__ptr_t dst [r3], __ptr_t src [r4], size_t len [r5]);
35 ble cr1, L(copy_LT_32) /* If move < 32 bytes use short move
38 andi. 11,3,7 /* Check alignment of DST. */
41 clrldi 10,4,61 /* Check alignment of SRC. */
42 cmpld cr6,10,11 /* SRC and DST alignments match? */
45 bne cr6,L(copy_GE_32_unaligned)
47 srdi 9,5,3 /* Number of full quadwords remaining. */
49 beq L(copy_GE_32_aligned_cont)
55 /* Get the SRC aligned to 8 bytes. */
73 clrldi 10,12,61 /* Check alignment of SRC again. */
74 srdi 9,31,3 /* Number of full doublewords remaining. */
76 L(copy_GE_32_aligned_cont):
86 /* Copy 1~3 doublewords so the main loop starts
87 at a multiple of 32 bytes. */
106 1: /* Copy 1 doubleword and set the counter. */
116 /* Main aligned copy loop. Copies up to 128-bytes at a time. */
119 /* check for any 32-byte or 64-byte lumps that are outside of a
120 nice 128-byte range. R8 contains the number of 32-byte
121 lumps, so drop this into the CR, and use the SO/EQ bits to help
122 handle the 32- or 64- byte lumps. Then handle the rest with an
123 unrolled 128-bytes-at-a-time copy loop. */
130 /* if the SO bit (indicating a 32-byte lump) is not set, move along. */
131 bns cr7,L(aligned_64byte)
140 /* if the EQ bit (indicating a 64-byte lump) is not set, move along. */
141 bne cr7,L(aligned_128setup)
154 /* Set up for the 128-byte at a time copy loop. */
156 cmpdi 8,0 # Any 4x lumps left?
157 beq 3f # if not, move along.
160 mtctr 8 # otherwise, load the ctr and begin.
165 /* for the 2nd + iteration of this loop. */
187 bdnz L(aligned_128head)
190 /* Check for tail bytes. */
199 /* At this point we have a tail of 0-7 bytes and we know that the
200 destination is doubleword-aligned. */
201 4: /* Copy 4 bytes. */
208 2: /* Copy 2 bytes. */
215 1: /* Copy 1 byte. */
220 0: /* Return original DST pointer. */
225 /* Handle copies of 0~31 bytes. */
233 /* At least 9 bytes to go. */
239 beq L(copy_LT_32_aligned)
241 /* Force 4-bytes alignment for SRC. */
250 1: bf 31,L(end_4bytes_alignment)
258 L(end_4bytes_alignment):
262 L(copy_LT_32_aligned):
263 /* At least 6 bytes to go, and SRC is word-aligned. */
277 8: /* Copy 8 bytes. */
286 4: /* Copy 4 bytes. */
293 2: /* Copy 2-3 bytes. */
305 1: /* Copy 1 byte. */
310 0: /* Return original DST pointer. */
314 /* Handles copies of 0~8 bytes. */
319 /* Though we could've used ld/std here, they are still
320 slow for unaligned cases. */
326 ld 3,-16(1) /* Return original DST pointers. */
330 4: /* Copies 4~7 bytes. */
345 5: /* Copy 1 byte. */
351 0: /* Return original DST pointer. */
355 /* Handle copies of 32+ bytes where DST is aligned (to quadword) but
356 SRC is not. Use aligned quadword loads from SRC, shifted to realign
357 the data, allowing for aligned DST stores. */
359 L(copy_GE_32_unaligned):
360 clrldi 0,0,60 /* Number of bytes until the 1st
362 andi. 11,3,15 /* Check alignment of DST (against
364 srdi 9,5,4 /* Number of full quadwords remaining. */
366 beq L(copy_GE_32_unaligned_cont)
368 /* SRC is not quadword aligned, get it aligned. */
373 /* Vector instructions work best when proper alignment (16-bytes)
374 is present. Move 0~15 bytes as needed to get DST quadword-aligned. */
375 1: /* Copy 1 byte. */
382 2: /* Copy 2 bytes. */
389 4: /* Copy 4 bytes. */
396 8: /* Copy 8 bytes. */
404 clrldi 10,12,60 /* Check alignment of SRC. */
405 srdi 9,31,4 /* Number of full quadwords remaining. */
407 /* The proper alignment is present, it is OK to copy the bytes now. */
408 L(copy_GE_32_unaligned_cont):
410 /* Setup two indexes to speed up the indexed vector operations. */
412 li 6,16 /* Index for 16-bytes offsets. */
413 li 7,32 /* Index for 32-bytes offsets. */
415 srdi 8,31,5 /* Setup the loop counter. */
422 bf 31,L(setup_unaligned_loop)
424 /* Copy another 16 bytes to align to 32-bytes due to the loop . */
432 L(setup_unaligned_loop):
434 ble cr6,L(end_unaligned_loop)
436 /* Copy 32 bytes at a time using vector instructions. */
440 /* Note: vr6/vr10 may contain data that was already copied,
441 but in order to get proper alignment, we may have to copy
442 some portions again. This is faster than having unaligned
443 vector instructions though. */
445 lvx 4,11,6 /* vr4 = r11+16. */
446 vperm 6,3,4,5 /* Merge the correctly-aligned portions
447 of vr3/vr4 into vr6. */
448 lvx 3,11,7 /* vr3 = r11+32. */
449 vperm 10,4,3,5 /* Merge the correctly-aligned portions
450 of vr3/vr4 into vr10. */
456 bdnz L(unaligned_loop)
459 L(end_unaligned_loop):
461 /* Check for tail bytes. */
469 /* We have 1~15 tail bytes to copy, and DST is quadword aligned. */
470 8: /* Copy 8 bytes. */
479 4: /* Copy 4 bytes. */
486 2: /* Copy 2~3 bytes. */
493 1: /* Copy 1 byte. */
498 0: /* Return original DST pointer. */
503 END_GEN_TB (memcpy,TB_TOCLESS)
504 libc_hidden_builtin_def (memcpy)