mirror of
https://sourceware.org/git/glibc.git
synced 2024-11-21 20:40:05 +00:00
aarch64: MTE compatible strchr
Introduce an Arm MTE compatible strchr implementation. The existing implementation assumes that any access to the pages in which the string resides is safe. This assumption is not true when MTE is enabled. This patch updates the algorithm to ensure that accesses remain within the bounds of an MTE tag (16-byte chunks) and improves overall performance. Benchmarked on Cortex-A72, Cortex-A53, Neoverse N1. Co-authored-by: Wilco Dijkstra <wilco.dijkstra@arm.com>
This commit is contained in:
Andrea Corallo
Szabolcs Nagy
parent
f7de454f20
commit
49beaaec1b
@ -22,118 +22,98 @@
|
||||
|
||||
/* Assumptions:
|
||||
*
|
||||
* ARMv8-a, AArch64
|
||||
* ARMv8-a, AArch64, Advanced SIMD.
|
||||
* MTE compatible.
|
||||
*/
|
||||
|
||||
/* Arguments and results. */
|
||||
#define srcin x0
|
||||
#define chrin w1
|
||||
|
||||
#define result x0
|
||||
|
||||
#define src x2
|
||||
#define tmp1 x3
|
||||
#define wtmp2 w4
|
||||
#define tmp3 x5
|
||||
#define tmp1 x1
|
||||
#define wtmp2 w3
|
||||
#define tmp3 x3
|
||||
|
||||
#define vrepchr v0
|
||||
#define vdata1 v1
|
||||
#define vdata2 v2
|
||||
#define vhas_nul1 v3
|
||||
#define vhas_nul2 v4
|
||||
#define vhas_chr1 v5
|
||||
#define vhas_chr2 v6
|
||||
#define vrepmask_0 v7
|
||||
#define vrepmask_c v16
|
||||
#define vend1 v17
|
||||
#define vend2 v18
|
||||
#define vdata v1
|
||||
#define qdata q1
|
||||
#define vhas_nul v2
|
||||
#define vhas_chr v3
|
||||
#define vrepmask v4
|
||||
#define vrepmask2 v5
|
||||
#define vend v6
|
||||
#define dend d6
|
||||
|
||||
/* Core algorithm.
|
||||
For each 32-byte hunk we calculate a 64-bit syndrome value, with
|
||||
two bits per byte (LSB is always in bits 0 and 1, for both big
|
||||
and little-endian systems). Bit 0 is set iff the relevant byte
|
||||
matched the requested character. Bit 1 is set iff the
|
||||
relevant byte matched the NUL end of string (we trigger off bit0
|
||||
for the special case of looking for NUL). Since the bits
|
||||
in the syndrome reflect exactly the order in which things occur
|
||||
in the original string a count_trailing_zeros() operation will
|
||||
identify exactly which byte is causing the termination, and why. */
|
||||
/* Core algorithm.
|
||||
|
||||
/* Locals and temporaries. */
|
||||
For each 16-byte chunk we calculate a 64-bit syndrome value with four bits
|
||||
per byte. For even bytes, bits 0-1 are set if the relevant byte matched the
|
||||
requested character, bits 2-3 are set if the byte is NUL (or matched), and
|
||||
bits 4-7 are not used and must be zero if none of bits 0-3 are set). Odd
|
||||
bytes set bits 4-7 so that adjacent bytes can be merged. Since the bits
|
||||
in the syndrome reflect the order in which things occur in the original
|
||||
string, counting trailing zeros identifies exactly which byte matched. */
|
||||
|
||||
ENTRY (strchr)
|
||||
DELOUSE (0)
|
||||
mov wtmp2, #0x0401
|
||||
movk wtmp2, #0x4010, lsl #16
|
||||
bic src, srcin, 15
|
||||
dup vrepchr.16b, chrin
|
||||
bic src, srcin, #31
|
||||
dup vrepmask_c.4s, wtmp2
|
||||
ands tmp1, srcin, #31
|
||||
add vrepmask_0.4s, vrepmask_c.4s, vrepmask_c.4s // lsl #1
|
||||
b.eq L(loop)
|
||||
ld1 {vdata.16b}, [src]
|
||||
mov wtmp2, 0x3003
|
||||
dup vrepmask.8h, wtmp2
|
||||
cmeq vhas_nul.16b, vdata.16b, 0
|
||||
cmeq vhas_chr.16b, vdata.16b, vrepchr.16b
|
||||
mov wtmp2, 0xf00f
|
||||
dup vrepmask2.8h, wtmp2
|
||||
|
||||
/* Input string is not 32-byte aligned. Rather than forcing
|
||||
the padding bytes to a safe value, we calculate the syndrome
|
||||
for all the bytes, but then mask off those bits of the
|
||||
syndrome that are related to the padding. */
|
||||
ld1 {vdata1.16b, vdata2.16b}, [src], #32
|
||||
neg tmp1, tmp1
|
||||
cmeq vhas_nul1.16b, vdata1.16b, #0
|
||||
cmeq vhas_chr1.16b, vdata1.16b, vrepchr.16b
|
||||
cmeq vhas_nul2.16b, vdata2.16b, #0
|
||||
cmeq vhas_chr2.16b, vdata2.16b, vrepchr.16b
|
||||
and vhas_nul1.16b, vhas_nul1.16b, vrepmask_0.16b
|
||||
and vhas_nul2.16b, vhas_nul2.16b, vrepmask_0.16b
|
||||
and vhas_chr1.16b, vhas_chr1.16b, vrepmask_c.16b
|
||||
and vhas_chr2.16b, vhas_chr2.16b, vrepmask_c.16b
|
||||
orr vend1.16b, vhas_nul1.16b, vhas_chr1.16b
|
||||
orr vend2.16b, vhas_nul2.16b, vhas_chr2.16b
|
||||
lsl tmp1, tmp1, #1
|
||||
addp vend1.16b, vend1.16b, vend2.16b // 256->128
|
||||
mov tmp3, #~0
|
||||
addp vend1.16b, vend1.16b, vend2.16b // 128->64
|
||||
lsr tmp1, tmp3, tmp1
|
||||
bit vhas_nul.16b, vhas_chr.16b, vrepmask.16b
|
||||
and vhas_nul.16b, vhas_nul.16b, vrepmask2.16b
|
||||
lsl tmp3, srcin, 2
|
||||
addp vend.16b, vhas_nul.16b, vhas_nul.16b /* 128->64 */
|
||||
|
||||
mov tmp3, vend1.2d[0]
|
||||
bic tmp1, tmp3, tmp1 // Mask padding bits.
|
||||
cbnz tmp1, L(tail)
|
||||
|
||||
L(loop):
|
||||
ld1 {vdata1.16b, vdata2.16b}, [src], #32
|
||||
cmeq vhas_nul1.16b, vdata1.16b, #0
|
||||
cmeq vhas_chr1.16b, vdata1.16b, vrepchr.16b
|
||||
cmeq vhas_nul2.16b, vdata2.16b, #0
|
||||
cmeq vhas_chr2.16b, vdata2.16b, vrepchr.16b
|
||||
/* Use a fast check for the termination condition. */
|
||||
orr vend1.16b, vhas_nul1.16b, vhas_chr1.16b
|
||||
orr vend2.16b, vhas_nul2.16b, vhas_chr2.16b
|
||||
orr vend1.16b, vend1.16b, vend2.16b
|
||||
addp vend1.2d, vend1.2d, vend1.2d
|
||||
mov tmp1, vend1.2d[0]
|
||||
fmov tmp1, dend
|
||||
lsr tmp1, tmp1, tmp3
|
||||
cbz tmp1, L(loop)
|
||||
|
||||
/* Termination condition found. Now need to establish exactly why
|
||||
we terminated. */
|
||||
and vhas_nul1.16b, vhas_nul1.16b, vrepmask_0.16b
|
||||
and vhas_nul2.16b, vhas_nul2.16b, vrepmask_0.16b
|
||||
and vhas_chr1.16b, vhas_chr1.16b, vrepmask_c.16b
|
||||
and vhas_chr2.16b, vhas_chr2.16b, vrepmask_c.16b
|
||||
orr vend1.16b, vhas_nul1.16b, vhas_chr1.16b
|
||||
orr vend2.16b, vhas_nul2.16b, vhas_chr2.16b
|
||||
addp vend1.16b, vend1.16b, vend2.16b // 256->128
|
||||
addp vend1.16b, vend1.16b, vend2.16b // 128->64
|
||||
|
||||
mov tmp1, vend1.2d[0]
|
||||
L(tail):
|
||||
sub src, src, #32
|
||||
rbit tmp1, tmp1
|
||||
clz tmp1, tmp1
|
||||
/* Tmp1 is even if the target charager was found first. Otherwise
|
||||
we've found the end of string and we weren't looking for NUL. */
|
||||
tst tmp1, #1
|
||||
add result, src, tmp1, lsr #1
|
||||
/* Tmp1 is an even multiple of 2 if the target character was
|
||||
found first. Otherwise we've found the end of string. */
|
||||
tst tmp1, 2
|
||||
add result, srcin, tmp1, lsr 2
|
||||
csel result, result, xzr, eq
|
||||
ret
|
||||
|
||||
.p2align 4
|
||||
L(loop):
|
||||
ldr qdata, [src, 16]!
|
||||
cmeq vhas_chr.16b, vdata.16b, vrepchr.16b
|
||||
cmhs vhas_nul.16b, vhas_chr.16b, vdata.16b
|
||||
umaxp vend.16b, vhas_nul.16b, vhas_nul.16b
|
||||
fmov tmp1, dend
|
||||
cbz tmp1, L(loop)
|
||||
|
||||
#ifdef __AARCH64EB__
|
||||
bif vhas_nul.16b, vhas_chr.16b, vrepmask.16b
|
||||
and vhas_nul.16b, vhas_nul.16b, vrepmask2.16b
|
||||
addp vend.16b, vhas_nul.16b, vhas_nul.16b /* 128->64 */
|
||||
fmov tmp1, dend
|
||||
#else
|
||||
bit vhas_nul.16b, vhas_chr.16b, vrepmask.16b
|
||||
and vhas_nul.16b, vhas_nul.16b, vrepmask2.16b
|
||||
addp vend.16b, vhas_nul.16b, vhas_nul.16b /* 128->64 */
|
||||
fmov tmp1, dend
|
||||
rbit tmp1, tmp1
|
||||
#endif
|
||||
clz tmp1, tmp1
|
||||
/* Tmp1 is an even multiple of 2 if the target character was
|
||||
found first. Otherwise we've found the end of string. */
|
||||
tst tmp1, 2
|
||||
add result, src, tmp1, lsr 2
|
||||
csel result, result, xzr, eq
|
||||
ret
|
||||
|
||||
END (strchr)
|
||||
libc_hidden_builtin_def (strchr)
|
||||
weak_alias (strchr, index)
|
||||
|
||||
Loading…
Reference in New Issue
Block a user