qiurui
/
glibc
mirror of git://sourceware.org/git/glibc.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

aarch64: MTE compatible memrchr 

Add support for MTE to memrchr. Regression tested with xcheck and benchmarked
with glibc's benchtests on the Cortex-A53, Cortex-A72, and Neoverse N1.

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.

Co-authored-by: Wilco Dijkstra <wilco.dijkstra@arm.com>
This commit is contained in:
Alex Butler 2020-06-09 16:08:07 +00:00 committed by Szabolcs Nagy
parent 7ff899969f
commit df06b0d90f
1 changed files with 84 additions and 117 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

201
sysdeps/aarch64/memrchr.S
View File

@ -22,144 +22,111 @@
/* Assumptions: /* Assumptions:
* *
* ARMv8-a, AArch64 * ARMv8-a, AArch64, Advanced SIMD.
* Neon Available. * MTE compatible.
*/ */
/* Arguments and results. */ /* Arguments and results. */
#define srcin x0 #define srcin x0
#define chrin w1 #define chrin w1
#define cntin x2 #define cntin x2
#define seek_dstin x11
#define seek_dst x12
#define result x0 #define result x0
#define src x3 #define src x3
#define tmp x4 #define cntrem x4
#define wtmp2 w5 #define synd x5
#define synd x6 #define shift x6
#define soff x9 #define tmp x7
#define cntrem x10 #define wtmp w7
#define end x8
#define endm1 x9
#define vrepchr v0 #define vrepchr v0
#define vdata1 v1 #define qdata q1
#define vdata2 v2 #define vdata v1
#define vhas_chr v2
#define vhas_chr1 v3 #define vrepmask v3
#define vhas_chr2 v4 #define vend v4
#define vrepmask v5 #define dend d4
#define vend v6
/* /*
* Core algorithm: Core algorithm:
* For each 16-byte chunk we calculate a 64-bit syndrome value with four bits
* For each 32-byte chunk we calculate a 64-bit syndrome value, with two bits per byte. For even bytes, bits 0-3 are set if the relevant byte matched the
* per byte. For each tuple, bit 0 is set if the relevant byte matched the requested character or the byte is NUL. Bits 4-7 must be zero. Bits 4-7 are
* requested character and bit 1 is not used (faster than using a 32bit set likewise for odd bytes so that adjacent bytes can be merged. Since the
* syndrome). Since the bits in the syndrome reflect exactly the order in which bits in the syndrome reflect the order in which things occur in the original
* things occur in the original string, counting trailing zeros allows to string, counting trailing zeros identifies exactly which byte matched. */
* identify exactly which byte has matched.
*/
ENTRY (__memrchr) ENTRY (__memrchr)
/* Do not dereference srcin if no bytes to compare. */ DELOUSE (0)
cbz cntin, L(zero_length) DELOUSE (2)
/* add end, srcin, cntin
* Magic constant 0x40100401 allows us to identify which lane matches sub endm1, end, 1
* the requested byte. bic src, endm1, 15
*/ cbz cntin, L(nomatch)
mov wtmp2, #0x0401 ld1 {vdata.16b}, [src]
movk wtmp2, #0x4010, lsl #16
dup vrepchr.16b, chrin dup vrepchr.16b, chrin
/* Work with aligned 32-byte chunks */ mov wtmp, 0xf00f
add seek_dstin, cntin, srcin dup vrepmask.8h, wtmp
add tmp, seek_dstin, #31 cmeq vhas_chr.16b, vdata.16b, vrepchr.16b
bic seek_dst, tmp, #31 neg shift, end, lsl 2
dup vrepmask.4s, wtmp2 and vhas_chr.16b, vhas_chr.16b, vrepmask.16b
ands soff, seek_dstin, #31 addp vend.16b, vhas_chr.16b, vhas_chr.16b /* 128->64 */
mov tmp, #32 fmov synd, dend
sub soff,tmp,soff lsl synd, synd, shift
and cntrem, cntin, #31 cbz synd, L(start_loop)
b.eq L(loop)
/*
* Input string is not 32-byte aligned. We calculate the syndrome
* value for the aligned 32 bytes block containing the first bytes
* and mask the irrelevant part.
*/
sub seek_dst, seek_dst, #32
ld1 {vdata1.16b, vdata2.16b}, [seek_dst]
sub tmp, soff, #32
adds cntin, cntin, tmp
cmeq vhas_chr1.16b, vdata1.16b, vrepchr.16b
cmeq vhas_chr2.16b, vdata2.16b, vrepchr.16b
and vhas_chr1.16b, vhas_chr1.16b, vrepmask.16b
and vhas_chr2.16b, vhas_chr2.16b, vrepmask.16b
addp vend.16b, vhas_chr1.16b, vhas_chr2.16b /* 256->128 */
addp vend.16b, vend.16b, vend.16b /* 128->64 */
mov synd, vend.2d[0]
/* Clear the (32-soff)*2 upper bits */
lsl tmp, soff, #1
lsl synd, synd, tmp
lsr synd, synd, tmp
/* The first block can also be the last */
b.ls L(masklast)
/* Have we found something already? */
cbnz synd, L(tail)
L(loop):
sub seek_dst, seek_dst, #32
ld1 {vdata1.16b, vdata2.16b}, [seek_dst]
subs cntin, cntin, #32
cmeq vhas_chr1.16b, vdata1.16b, vrepchr.16b
cmeq vhas_chr2.16b, vdata2.16b, vrepchr.16b
/* If we're out of data we finish regardless of the result */
b.ls L(end)
/* Use a fast check for the termination condition */
orr vend.16b, vhas_chr1.16b, vhas_chr2.16b
addp vend.2d, vend.2d, vend.2d
mov synd, vend.2d[0]
/* We're not out of data, loop if we haven't found the character */
cbz synd, L(loop)
L(end):
/* Termination condition found, let's calculate the syndrome value */
and vhas_chr1.16b, vhas_chr1.16b, vrepmask.16b
and vhas_chr2.16b, vhas_chr2.16b, vrepmask.16b
addp vend.16b, vhas_chr1.16b, vhas_chr2.16b /* 256->128 */
addp vend.16b, vend.16b, vend.16b /* 128->64 */
mov synd, vend.2d[0]
/* Only do the clear for the last possible block */
b.hi L(tail)
L(masklast):
/* Clear the (32 - ((cntrem + (32-soff)) % 32)) * 2 lower bits */
add tmp, cntrem, soff
and tmp, tmp, #31
sub tmp, tmp, #32
neg tmp, tmp, lsl #1
lsr synd, synd, tmp
lsl synd, synd, tmp
L(tail):
/* Compensate the last post-increment*/
add seek_dst, seek_dst, #32
/* Check that we have found a character */
cmp synd, #0
/* And count the leading zeros */
clz synd, synd clz synd, synd
/* Compute the potential result */ sub result, endm1, synd, lsr 2
sub result, seek_dst, synd, lsr #1 cmp cntin, synd, lsr 2
sub result, result, #1 csel result, result, xzr, hi
/* Select result or NULL */
csel result, xzr, result, eq
ret ret
L(zero_length): L(start_loop):
mov result, #0 sub tmp, end, src
subs cntrem, cntin, tmp
b.ls L(nomatch)
/* Make sure that it won't overread by a 16-byte chunk */
add tmp, cntrem, 15
tbnz tmp, 4, L(loop32_2)
.p2align 4
L(loop32):
ldr qdata, [src, -16]!
cmeq vhas_chr.16b, vdata.16b, vrepchr.16b
umaxp vend.16b, vhas_chr.16b, vhas_chr.16b /* 128->64 */
fmov synd, dend
cbnz synd, L(end)
L(loop32_2):
ldr qdata, [src, -16]!
subs cntrem, cntrem, 32
cmeq vhas_chr.16b, vdata.16b, vrepchr.16b
b.ls L(end)
umaxp vend.16b, vhas_chr.16b, vhas_chr.16b /* 128->64 */
fmov synd, dend
cbz synd, L(loop32)
L(end):
and vhas_chr.16b, vhas_chr.16b, vrepmask.16b
addp vend.16b, vhas_chr.16b, vhas_chr.16b /* 128->64 */
fmov synd, dend
add tmp, src, 15
#ifdef __AARCH64EB__
rbit synd, synd
#endif
clz synd, synd
sub tmp, tmp, synd, lsr 2
cmp tmp, srcin
csel result, tmp, xzr, hs
ret ret
L(nomatch):
mov result, 0
ret
END (__memrchr) END (__memrchr)
weak_alias (__memrchr, memrchr) weak_alias (__memrchr, memrchr)
libc_hidden_builtin_def (memrchr) libc_hidden_builtin_def (memrchr)