From 7ff899969fbff7f254c0ffabb4480afe1ccbfa22 Mon Sep 17 00:00:00 2001 From: Alex Butler Date: Tue, 9 Jun 2020 16:06:03 +0000 Subject: aarch64: MTE compatible memchr Add support for MTE to memchr. 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: Gabor Kertesz --- sysdeps/aarch64/memchr.S | 187 ++++++++++++++++++++--------------------------- 1 file changed, 80 insertions(+), 107 deletions(-) (limited to 'sysdeps/aarch64') diff --git a/sysdeps/aarch64/memchr.S b/sysdeps/aarch64/memchr.S index 85c65cb..23f30f6 100644 --- a/sysdeps/aarch64/memchr.S +++ b/sysdeps/aarch64/memchr.S @@ -22,8 +22,8 @@ /* Assumptions: * - * ARMv8-a, AArch64 - * Neon Available. + * ARMv8-a, AArch64, Advanced SIMD. + * MTE compatible. */ #ifndef MEMCHR @@ -34,128 +34,101 @@ #define srcin x0 #define chrin w1 #define cntin x2 - #define result x0 #define src x3 -#define tmp x4 -#define wtmp2 w5 -#define synd x6 -#define soff x9 -#define cntrem x10 +#define cntrem x4 +#define synd x5 +#define shift x6 +#define tmp x7 +#define wtmp w7 #define vrepchr v0 -#define vdata1 v1 -#define vdata2 v2 -#define vhas_chr1 v3 -#define vhas_chr2 v4 -#define vrepmask v5 -#define vend v6 +#define qdata q1 +#define vdata v1 +#define vhas_chr v2 +#define vrepmask v3 +#define vend v4 +#define dend d4 /* - * Core algorithm: - * - * For each 32-byte chunk we calculate a 64-bit syndrome value, with two bits - * per byte. For each tuple, bit 0 is set if the relevant byte matched the - * requested character and bit 1 is not used (faster than using a 32bit - * syndrome). Since the bits in the syndrome reflect exactly the order in which - * things occur in the original string, counting trailing zeros allows to - * identify exactly which byte has matched. - */ + Core algorithm: + For each 16-byte chunk we calculate a 64-bit syndrome value with four bits + per byte. For even bytes, bits 0-3 are set if the relevant byte matched the + requested character or the byte is NUL. Bits 4-7 must be zero. Bits 4-7 are + set likewise for odd bytes 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 (MEMCHR) - /* Do not dereference srcin if no bytes to compare. */ - cbz cntin, L(zero_length) - /* - * Magic constant 0x40100401 allows us to identify which lane matches - * the requested byte. - */ - mov wtmp2, #0x0401 - movk wtmp2, #0x4010, lsl #16 + DELOUSE (0) + DELOUSE (2) + bic src, srcin, 15 + cbz cntin, L(nomatch) + ld1 {vdata.16b}, [src] dup vrepchr.16b, chrin - /* Work with aligned 32-byte chunks */ - bic src, srcin, #31 - dup vrepmask.4s, wtmp2 - ands soff, srcin, #31 - and cntrem, cntin, #31 - 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. - */ - - ld1 {vdata1.16b, vdata2.16b}, [src], #32 - 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 soff*2 lower bits */ - lsl tmp, soff, #1 - lsr synd, synd, tmp - lsl 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): - ld1 {vdata1.16b, vdata2.16b}, [src], #32 - 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) + mov wtmp, 0xf00f + dup vrepmask.8h, wtmp + cmeq vhas_chr.16b, vdata.16b, vrepchr.16b + lsl shift, srcin, 2 + and vhas_chr.16b, vhas_chr.16b, vrepmask.16b + addp vend.16b, vhas_chr.16b, vhas_chr.16b /* 128->64 */ + fmov synd, dend + lsr synd, synd, shift + cbz synd, L(start_loop) + rbit synd, synd + clz synd, synd + add result, srcin, synd, lsr 2 + cmp cntin, synd, lsr 2 + csel result, result, xzr, hi + ret + +L(start_loop): + sub tmp, src, srcin + add tmp, tmp, 16 + 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): - /* 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 + soff) % 32)) * 2 upper bits */ - add tmp, cntrem, soff - and tmp, tmp, #31 - sub tmp, tmp, #32 - neg tmp, tmp, lsl #1 - lsl synd, synd, tmp - lsr synd, synd, tmp - -L(tail): - /* Count the trailing zeros using bit reversing */ + 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, srcin, cntin + sub cntrem, tmp, src +#ifndef __AARCH64EB__ rbit synd, synd - /* Compensate the last post-increment */ - sub src, src, #32 - /* Check that we have found a character */ - cmp synd, #0 - /* And count the leading zeros */ +#endif clz synd, synd - /* Compute the potential result */ - add result, src, synd, lsr #1 - /* Select result or NULL */ - csel result, xzr, result, eq + cmp cntrem, synd, lsr 2 + add result, src, synd, lsr 2 + csel result, result, xzr, hi ret -L(zero_length): - mov result, #0 +L(nomatch): + mov result, 0 ret + END (MEMCHR) weak_alias (MEMCHR, memchr) libc_hidden_builtin_def (memchr) -- cgit v1.1