[AArch64] Reverting recent optimized memset().
This commit is contained in:
Marcus Shawcroft
parent
c028685518
commit
c7806ef76a
|
|
@ -1,3 +1,8 @@
|
||||||
|
2015-07-15 Wilco Dijkstra <wdijkstr@arm.com>
|
||||||
|
|
||||||
|
* libc/machine/aarch64/memset.S (memset):
|
||||||
|
Revert: Rewrite of optimized memset.
|
||||||
|
|
||||||
2015-07-13 Wilco Dijkstra <wdijkstr@arm.com>
|
2015-07-13 Wilco Dijkstra <wdijkstr@arm.com>
|
||||||
|
|
||||||
* libc/machine/aarch64/memset.S (memset):
|
* libc/machine/aarch64/memset.S (memset):
|
||||||
|
|
|
||||||
|
|
@ -24,37 +24,10 @@
|
||||||
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||||
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */
|
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */
|
||||||
|
|
||||||
/*
|
|
||||||
* Copyright (c) 2015 ARM Ltd
|
|
||||||
* All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* 1. Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* 2. Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* 3. The name of the company may not be used to endorse or promote
|
|
||||||
* products derived from this software without specific prior written
|
|
||||||
* permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY ARM LTD ``AS IS'' AND ANY EXPRESS OR IMPLIED
|
|
||||||
* WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
|
|
||||||
* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
|
|
||||||
* IN NO EVENT SHALL ARM LTD BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
|
|
||||||
* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
|
|
||||||
* TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
|
|
||||||
* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
|
|
||||||
* NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
|
|
||||||
* SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
/* Assumptions:
|
/* Assumptions:
|
||||||
*
|
*
|
||||||
* ARMv8-a, AArch64, unaligned accesses
|
* ARMv8-a, AArch64
|
||||||
|
* Unaligned accesses
|
||||||
*
|
*
|
||||||
*/
|
*/
|
||||||
|
|
||||||
|
|
@ -62,20 +35,32 @@
|
||||||
/* See memset-stub.c */
|
/* See memset-stub.c */
|
||||||
#else
|
#else
|
||||||
|
|
||||||
#define dstin x0
|
/* By default we assume that the DC instruction can be used to zero
|
||||||
#define val x1
|
data blocks more efficiently. In some circumstances this might be
|
||||||
#define valw w1
|
unsafe, for example in an asymmetric multiprocessor environment with
|
||||||
#define count x2
|
different DC clear lengths (neither the upper nor lower lengths are
|
||||||
#define dst x3
|
safe to use). The feature can be disabled by defining DONT_USE_DC.
|
||||||
#define dstend x4
|
|
||||||
#define tmp1 x5
|
If code may be run in a virtualized environment, then define
|
||||||
#define tmp1w w5
|
MAYBE_VIRT. This will cause the code to cache the system register
|
||||||
#define tmp2 x6
|
values rather than re-reading them each call. */
|
||||||
#define tmp2w w6
|
|
||||||
#define zva_len x7
|
#define dstin x0
|
||||||
#define zva_lenw w7
|
#define val w1
|
||||||
|
#define count x2
|
||||||
|
#define tmp1 x3
|
||||||
|
#define tmp1w w3
|
||||||
|
#define tmp2 x4
|
||||||
|
#define tmp2w w4
|
||||||
|
#define zva_len_x x5
|
||||||
|
#define zva_len w5
|
||||||
|
#define zva_bits_x x6
|
||||||
|
|
||||||
|
#define A_l x7
|
||||||
|
#define A_lw w7
|
||||||
|
#define dst x8
|
||||||
|
#define tmp3w w9
|
||||||
|
|
||||||
#define L(l) .L ## l
|
|
||||||
|
|
||||||
.macro def_fn f p2align=0
|
.macro def_fn f p2align=0
|
||||||
.text
|
.text
|
||||||
|
|
@ -87,153 +72,175 @@
|
||||||
|
|
||||||
def_fn memset p2align=6
|
def_fn memset p2align=6
|
||||||
|
|
||||||
dup v0.16B, valw
|
mov dst, dstin /* Preserve return value. */
|
||||||
add dstend, dstin, count
|
ands A_lw, val, #255
|
||||||
|
#ifndef DONT_USE_DC
|
||||||
cmp count, 96
|
b.eq .Lzero_mem
|
||||||
b.hi L(set_long)
|
#endif
|
||||||
cmp count, 16
|
orr A_lw, A_lw, A_lw, lsl #8
|
||||||
b.hs L(set_medium)
|
orr A_lw, A_lw, A_lw, lsl #16
|
||||||
mov val, v0.D[0]
|
orr A_l, A_l, A_l, lsl #32
|
||||||
|
.Ltail_maybe_long:
|
||||||
/* Set 0..15 bytes. */
|
cmp count, #64
|
||||||
tbz count, 3, 1f
|
b.ge .Lnot_short
|
||||||
str val, [dstin]
|
.Ltail_maybe_tiny:
|
||||||
str val, [dstend, -8]
|
cmp count, #15
|
||||||
ret
|
b.le .Ltail15tiny
|
||||||
nop
|
.Ltail63:
|
||||||
1: tbz count, 2, 2f
|
ands tmp1, count, #0x30
|
||||||
str valw, [dstin]
|
b.eq .Ltail15
|
||||||
str valw, [dstend, -4]
|
add dst, dst, tmp1
|
||||||
ret
|
cmp tmp1w, #0x20
|
||||||
2: cbz count, 3f
|
b.eq 1f
|
||||||
strb valw, [dstin]
|
b.lt 2f
|
||||||
tbz count, 1, 3f
|
stp A_l, A_l, [dst, #-48]
|
||||||
strh valw, [dstend, -2]
|
1:
|
||||||
3: ret
|
stp A_l, A_l, [dst, #-32]
|
||||||
|
2:
|
||||||
/* Set 17..96 bytes. */
|
stp A_l, A_l, [dst, #-16]
|
||||||
L(set_medium):
|
|
||||||
str q0, [dstin]
|
.Ltail15:
|
||||||
tbnz count, 6, L(set96)
|
and count, count, #15
|
||||||
str q0, [dstend, -16]
|
add dst, dst, count
|
||||||
tbz count, 5, 1f
|
stp A_l, A_l, [dst, #-16] /* Repeat some/all of last store. */
|
||||||
str q0, [dstin, 16]
|
ret
|
||||||
str q0, [dstend, -32]
|
|
||||||
1: ret
|
.Ltail15tiny:
|
||||||
|
/* Set up to 15 bytes. Does not assume earlier memory
|
||||||
.p2align 4
|
being set. */
|
||||||
/* Set 64..96 bytes. Write 64 bytes from the start and
|
tbz count, #3, 1f
|
||||||
32 bytes from the end. */
|
str A_l, [dst], #8
|
||||||
L(set96):
|
1:
|
||||||
str q0, [dstin, 16]
|
tbz count, #2, 1f
|
||||||
stp q0, q0, [dstin, 32]
|
str A_lw, [dst], #4
|
||||||
stp q0, q0, [dstend, -32]
|
1:
|
||||||
ret
|
tbz count, #1, 1f
|
||||||
|
strh A_lw, [dst], #2
|
||||||
.p2align 3
|
1:
|
||||||
nop
|
tbz count, #0, 1f
|
||||||
L(set_long):
|
strb A_lw, [dst]
|
||||||
and valw, valw, 255
|
1:
|
||||||
bic dst, dstin, 15
|
ret
|
||||||
str q0, [dstin]
|
|
||||||
cmp count, 256
|
/* Critical loop. Start at a new cache line boundary. Assuming
|
||||||
ccmp valw, 0, 0, cs
|
* 64 bytes per line, this ensures the entire loop is in one line. */
|
||||||
b.eq L(try_zva)
|
.p2align 6
|
||||||
L(no_zva):
|
.Lnot_short:
|
||||||
sub count, dstend, dst /* Count is 16 too large. */
|
neg tmp2, dst
|
||||||
add dst, dst, 16
|
ands tmp2, tmp2, #15
|
||||||
sub count, count, 64 + 16 /* Adjust count and bias for loop. */
|
b.eq 2f
|
||||||
1: stp q0, q0, [dst], 64
|
/* Bring DST to 128-bit (16-byte) alignment. We know that there's
|
||||||
stp q0, q0, [dst, -32]
|
* more than that to set, so we simply store 16 bytes and advance by
|
||||||
L(tail64):
|
* the amount required to reach alignment. */
|
||||||
subs count, count, 64
|
sub count, count, tmp2
|
||||||
b.hi 1b
|
stp A_l, A_l, [dst]
|
||||||
2: stp q0, q0, [dstend, -64]
|
add dst, dst, tmp2
|
||||||
stp q0, q0, [dstend, -32]
|
/* There may be less than 63 bytes to go now. */
|
||||||
ret
|
cmp count, #63
|
||||||
|
b.le .Ltail63
|
||||||
.p2align 3
|
2:
|
||||||
L(try_zva):
|
sub dst, dst, #16 /* Pre-bias. */
|
||||||
mrs tmp1, dczid_el0
|
sub count, count, #64
|
||||||
tbnz tmp1w, 4, L(no_zva)
|
1:
|
||||||
and tmp1w, tmp1w, 4
|
stp A_l, A_l, [dst, #16]
|
||||||
cmp tmp1w, 4 /* ZVA size is 64 bytes. */
|
stp A_l, A_l, [dst, #32]
|
||||||
b.ne L(zva_128)
|
stp A_l, A_l, [dst, #48]
|
||||||
|
stp A_l, A_l, [dst, #64]!
|
||||||
/* Write the first and last 64 byte aligned block using stp rather
|
subs count, count, #64
|
||||||
than using DC ZVA. This is faster on some cores.
|
b.ge 1b
|
||||||
*/
|
tst count, #0x3f
|
||||||
L(zva_64):
|
add dst, dst, #16
|
||||||
str q0, [dst, 16]
|
b.ne .Ltail63
|
||||||
stp q0, q0, [dst, 32]
|
ret
|
||||||
bic dst, dst, 63
|
|
||||||
stp q0, q0, [dst, 64]
|
#ifndef DONT_USE_DC
|
||||||
stp q0, q0, [dst, 96]
|
/* For zeroing memory, check to see if we can use the ZVA feature to
|
||||||
sub count, dstend, dst /* Count is now 128 too large. */
|
* zero entire 'cache' lines. */
|
||||||
sub count, count, 128+64+64 /* Adjust count and bias for loop. */
|
.Lzero_mem:
|
||||||
add dst, dst, 128
|
mov A_l, #0
|
||||||
nop
|
cmp count, #63
|
||||||
1: dc zva, dst
|
b.le .Ltail_maybe_tiny
|
||||||
add dst, dst, 64
|
neg tmp2, dst
|
||||||
subs count, count, 64
|
ands tmp2, tmp2, #15
|
||||||
b.hi 1b
|
b.eq 1f
|
||||||
stp q0, q0, [dst, 0]
|
sub count, count, tmp2
|
||||||
stp q0, q0, [dst, 32]
|
stp A_l, A_l, [dst]
|
||||||
stp q0, q0, [dstend, -64]
|
add dst, dst, tmp2
|
||||||
stp q0, q0, [dstend, -32]
|
cmp count, #63
|
||||||
ret
|
b.le .Ltail63
|
||||||
|
1:
|
||||||
.p2align 3
|
/* For zeroing small amounts of memory, it's not worth setting up
|
||||||
L(zva_128):
|
* the line-clear code. */
|
||||||
cmp tmp1w, 5 /* ZVA size is 128 bytes. */
|
cmp count, #128
|
||||||
b.ne L(zva_other)
|
b.lt .Lnot_short
|
||||||
|
#ifdef MAYBE_VIRT
|
||||||
str q0, [dst, 16]
|
/* For efficiency when virtualized, we cache the ZVA capability. */
|
||||||
stp q0, q0, [dst, 32]
|
adrp tmp2, .Lcache_clear
|
||||||
stp q0, q0, [dst, 64]
|
ldr zva_len, [tmp2, #:lo12:.Lcache_clear]
|
||||||
stp q0, q0, [dst, 96]
|
tbnz zva_len, #31, .Lnot_short
|
||||||
bic dst, dst, 127
|
cbnz zva_len, .Lzero_by_line
|
||||||
sub count, dstend, dst /* Count is now 128 too large. */
|
mrs tmp1, dczid_el0
|
||||||
sub count, count, 128+128 /* Adjust count and bias for loop. */
|
tbz tmp1, #4, 1f
|
||||||
add dst, dst, 128
|
/* ZVA not available. Remember this for next time. */
|
||||||
1: dc zva, dst
|
mov zva_len, #~0
|
||||||
add dst, dst, 128
|
str zva_len, [tmp2, #:lo12:.Lcache_clear]
|
||||||
subs count, count, 128
|
b .Lnot_short
|
||||||
b.hi 1b
|
1:
|
||||||
stp q0, q0, [dstend, -128]
|
mov tmp3w, #4
|
||||||
stp q0, q0, [dstend, -96]
|
and zva_len, tmp1w, #15 /* Safety: other bits reserved. */
|
||||||
stp q0, q0, [dstend, -64]
|
lsl zva_len, tmp3w, zva_len
|
||||||
stp q0, q0, [dstend, -32]
|
str zva_len, [tmp2, #:lo12:.Lcache_clear]
|
||||||
ret
|
#else
|
||||||
|
mrs tmp1, dczid_el0
|
||||||
L(zva_other):
|
tbnz tmp1, #4, .Lnot_short
|
||||||
mov tmp2w, 4
|
mov tmp3w, #4
|
||||||
lsl zva_lenw, tmp2w, tmp1w
|
and zva_len, tmp1w, #15 /* Safety: other bits reserved. */
|
||||||
add tmp1, zva_len, 64 /* Max alignment bytes written. */
|
lsl zva_len, tmp3w, zva_len
|
||||||
cmp count, tmp1
|
#endif
|
||||||
blo L(no_zva)
|
|
||||||
|
.Lzero_by_line:
|
||||||
sub tmp2, zva_len, 1
|
/* Compute how far we need to go to become suitably aligned. We're
|
||||||
add tmp1, dst, zva_len
|
* already at quad-word alignment. */
|
||||||
add dst, dst, 16
|
cmp count, zva_len_x
|
||||||
subs count, tmp1, dst /* Actual alignment bytes to write. */
|
b.lt .Lnot_short /* Not enough to reach alignment. */
|
||||||
bic tmp1, tmp1, tmp2 /* Aligned dc zva start address. */
|
sub zva_bits_x, zva_len_x, #1
|
||||||
beq 2f
|
neg tmp2, dst
|
||||||
1: stp q0, q0, [dst], 64
|
ands tmp2, tmp2, zva_bits_x
|
||||||
stp q0, q0, [dst, -32]
|
b.eq 1f /* Already aligned. */
|
||||||
subs count, count, 64
|
/* Not aligned, check that there's enough to copy after alignment. */
|
||||||
b.hi 1b
|
sub tmp1, count, tmp2
|
||||||
2: mov dst, tmp1
|
cmp tmp1, #64
|
||||||
sub count, dstend, tmp1 /* Remaining bytes to write. */
|
ccmp tmp1, zva_len_x, #8, ge /* NZCV=0b1000 */
|
||||||
subs count, count, zva_len
|
b.lt .Lnot_short
|
||||||
b.lo 4f
|
/* We know that there's at least 64 bytes to zero and that it's safe
|
||||||
3: dc zva, dst
|
* to overrun by 64 bytes. */
|
||||||
add dst, dst, zva_len
|
mov count, tmp1
|
||||||
subs count, count, zva_len
|
2:
|
||||||
b.hs 3b
|
stp A_l, A_l, [dst]
|
||||||
4: add count, count, zva_len
|
stp A_l, A_l, [dst, #16]
|
||||||
b L(tail64)
|
stp A_l, A_l, [dst, #32]
|
||||||
|
subs tmp2, tmp2, #64
|
||||||
.size memset, . - memset
|
stp A_l, A_l, [dst, #48]
|
||||||
|
add dst, dst, #64
|
||||||
|
b.ge 2b
|
||||||
|
/* We've overrun a bit, so adjust dst downwards. */
|
||||||
|
add dst, dst, tmp2
|
||||||
|
1:
|
||||||
|
sub count, count, zva_len_x
|
||||||
|
3:
|
||||||
|
dc zva, dst
|
||||||
|
add dst, dst, zva_len_x
|
||||||
|
subs count, count, zva_len_x
|
||||||
|
b.ge 3b
|
||||||
|
ands count, count, zva_bits_x
|
||||||
|
b.ne .Ltail_maybe_long
|
||||||
|
ret
|
||||||
|
.size memset, .-memset
|
||||||
|
#ifdef MAYBE_VIRT
|
||||||
|
.bss
|
||||||
|
.p2align 2
|
||||||
|
.Lcache_clear:
|
||||||
|
.space 4
|
||||||
|
#endif
|
||||||
|
#endif /* DONT_USE_DC */
|
||||||
#endif
|
#endif
|
||||||
|
|
|
||||||
Loading…
Reference in New Issue