/* SPDX-License-Identifier: GPL-2.0 */ /* * "memcpy" implementation of SuperH * * Copyright (C) 1999 Niibe Yutaka * Copyright (c) 2002 STMicroelectronics Ltd * Modified from memcpy.S and micro-optimised for SH4 * Stuart Menefy (stuart.menefy@st.com) *
*/
#include <linux/linkage.h>
/* * void *memcpy(void *dst, const void *src, size_t n); * * It is assumed that there is no overlap between src and dst. * If there is an overlap, then the results are undefined.
*/
! ! GHIJ KLMN OPQR --> ...G HIJK LMNO PQR. !
! Size is 16 or greater, and may have trailing bytes
.balign 32
.Lcase1: ! Read a long word and write a long word at once ! At the start of each iteration, r7 contains last long load
add #-1,r5 ! 79 EX
mov r4,r2 ! 5 MT (0 cycles latency)
mov.l @(r0,r5),r7 ! 21 LS (2 cycles latency)
add #-4,r5 ! 50 EX
! Size is 16 or greater, and may have trailing bytes
.balign 32
.Lcase3: ! Read a long word and write a long word at once ! At the start of each iteration, r7 contains last long load
add #-3,r5 ! 79 EX
mov r4,r2 ! 5 MT (0 cycles latency)
mov.l @(r0,r5),r7 ! 21 LS (2 cycles latency)
add #-4,r5 ! 50 EX
! Arguments are not nicely long word aligned or zero len. ! Check for small copies, and if so do a simple byte at a time copy. ! ! Deciding on an exact value of 'small' is not easy, as the point at which ! using the optimised routines become worthwhile varies (these are the ! cycle counts for differnet sizes using byte-at-a-time vs. optimised): ! size byte-at-time long word byte ! 16 42 39-40 46-50 50-55 ! 24 58 43-44 54-58 62-67 ! 36 82 49-50 66-70 80-85 ! However the penalty for getting it 'wrong' is much higher for long word ! aligned data (and this is more common), so use a value of 16.
! Copy the cache line aligned blocks ! ! In use: r0, r2, r4, r5 ! Scratch: r1, r3, r6, r7 ! ! We could do this with the four scratch registers, but if src ! and dest hit the same cache line, this will thrash, so make ! use of additional registers. ! ! We also need r0 as a temporary (for movca), so 'undo' the invariant: ! r5: src (was r0+r5) ! r1: dest (was r0) ! this can be reversed at the end, so we don't need to save any extra ! state. ! 1: mov.l r8, @-r15 ! 30 LS
add r0, r5 ! 49 EX
add #-2, r5 ! 50 EX
bt/s 1f ! 110 BR
and r1, r2 ! 78 EX
! Copy a short word one at a time until we are cache line aligned ! Normal values: r0, r2, r3, r4 ! Unused: r1, r6, r7 ! Mod: r5 (=r5-2) !
add #2, r3 ! 50 EX
! Copy the cache line aligned blocks ! ! In use: r0, r2, r4, r5 (=r5-2) ! Scratch: r1, r3, r6, r7 ! ! We could do this with the four scratch registers, but if src ! and dest hit the same cache line, this will thrash, so make ! use of additional registers. ! ! We also need r0 as a temporary (for movca), so 'undo' the invariant: ! r5: src (was r0+r5) ! r1: dest (was r0) ! this can be reversed at the end, so we don't need to save any extra ! state. ! 1: mov.l r8, @-r15 ! 30 LS
add r0, r5 ! 49 EX
! 17 cycles, 32 bytes per iteration
#ifdef CONFIG_CPU_LITTLE_ENDIAN 2: mov.w @r5+, r0 ! 14 LS (latency=2) ..JI
add #-0x20, r1 ! 50 EX
mov.l @r5+, r3 ! 15 LS (latency=2) NMLK
mov.l @r5+, r6 ! 15 LS (latency=2) RQPO
shll16 r0 ! 103 EX JI..
mov.l @r5+, r7 ! 15 LS (latency=2)
xtrct r3, r0 ! 48 EX LKJI
mov.l @r5+, r8 ! 15 LS (latency=2)
xtrct r6, r3 ! 48 EX PONM
mov.l @r5+, r9 ! 15 LS (latency=2)
xtrct r7, r6 ! 48 EX
mov.l @r5+, r10 ! 15 LS (latency=2)
xtrct r8, r7 ! 48 EX
mov.l @r5+, r11 ! 15 LS (latency=2)
xtrct r9, r8 ! 48 EX
mov.w @r5+, r12 ! 15 LS (latency=2)
xtrct r10, r9 ! 48 EX
movca.l r0,@r1 ! 40 LS (latency=3-7)
xtrct r11, r10 ! 48 EX
mov.l r3, @(0x04,r1) ! 33 LS
xtrct r12, r11 ! 48 EX
mov.l r6, @(0x08,r1) ! 33 LS
mov.l r7, @(0x0c,r1) ! 33 LS
mov.l r8, @(0x10,r1) ! 33 LS
add #-0x40, r5 ! 50 EX
mov.l r9, @(0x14,r1) ! 33 LS
cmp/eq r2,r1 ! 54 MT
mov.l r10, @(0x18,r1) ! 33 LS
bf/s 2b ! 109 BR
mov.l r11, @(0x1c,r1) ! 33 LS
#else 2: mov.w @(0x1e,r5), r0 ! 17 LS (latency=2)
add #-2, r5 ! 50 EX
mov.l @(0x1c,r5), r3 ! 18 LS (latency=2)
add #-4, r1 ! 50 EX
mov.l @(0x18,r5), r6 ! 18 LS (latency=2)
shll16 r0 ! 103 EX
mov.l @(0x14,r5), r7 ! 18 LS (latency=2)
xtrct r3, r0 ! 48 EX
mov.l @(0x10,r5), r8 ! 18 LS (latency=2)
xtrct r6, r3 ! 48 EX
mov.l @(0x0c,r5), r9 ! 18 LS (latency=2)
xtrct r7, r6 ! 48 EX
mov.l @(0x08,r5), r10 ! 18 LS (latency=2)
xtrct r8, r7 ! 48 EX
mov.l @(0x04,r5), r11 ! 18 LS (latency=2)
xtrct r9, r8 ! 48 EX
mov.l @(0x00,r5), r12 ! 18 LS (latency=2)
xtrct r10, r9 ! 48 EX
movca.l r0,@r1 ! 40 LS (latency=3-7)
add #-0x1c, r1 ! 50 EX
mov.l r3, @(0x18,r1) ! 33 LS
xtrct r11, r10 ! 48 EX
mov.l r6, @(0x14,r1) ! 33 LS
xtrct r12, r11 ! 48 EX
mov.l r7, @(0x10,r1) ! 33 LS
mov.l r8, @(0x0c,r1) ! 33 LS
add #-0x1e, r5 ! 50 EX
mov.l r9, @(0x08,r1) ! 33 LS
cmp/eq r2,r1 ! 54 MT
mov.l r10, @(0x04,r1) ! 33 LS
bf/s 2b ! 109 BR
mov.l r11, @(0x00,r1) ! 33 LS
#endif
mov.l @r15+, r12
mov r1, r0 ! 5 MT (latency=0)
mov.l @r15+, r11 ! 15 LS sub r1, r5 ! 75 EX
mov.l @r15+, r10 ! 15 LS
cmp/eq r4, r0 ! 54 MT
bf/s 1f ! 109 BR
mov.l @r15+, r9 ! 15 LS
rts 1: mov.l @r15+, r8 ! 15 LS
add #0x1e, r5 ! 50 EX
! Finish off a short word at a time ! r5 must be invariant - 2 10: mov r4,r2 ! 5 MT (latency=0)
add #1,r2 ! 50 EX
cmp/hi r2, r0 ! 57 MT
bf/s 1f ! 109 BR
add #2, r2 ! 50 EX
3: mov.w @(r0,r5),r1 ! 20 LS
cmp/hi r2,r0 ! 57 MT
bt/s 3b ! 109 BR
mov.w r1,@-r0 ! 29 LS 1:
! ! Finally, copy the last byte if necessary
cmp/eq r4,r0 ! 54 MT
bt/s 9b
add #1,r5
mov.b @(r0,r5),r1
rts
mov.b r1,@-r0
Messung V0.5 in Prozent
¤ Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.0.12Bemerkung:
(vorverarbeitet am 2026-06-08)
¤
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.