aarch64: MTE compatible memchr

Message ID VI1PR08MB366340C245F23F517B712F5CED820@VI1PR08MB3663.eurprd08.prod.outlook.com
State Committed
Headers
Series aarch64: MTE compatible memchr |

Commit Message

Alex Butler June 9, 2020, 4:06 p.m. UTC
  This patch adds an MTE compatible implementation of memchr.

Please see the benchmark results for the performance uplift.

| length | position | alignment | uplift A72  | uplift A53  | uplift N1   |
|--------+----------+-----------+-------------|-------------|-------------|
|   2048 |       32 |         0 |       1.53x |       1.02x |       1.21x |
|    256 |       64 |         1 |       1.57x |       1.03x |       1.02x |
|   2048 |       32 |         0 |       1.53x |       1.02x |       1.21x |
|    256 |       64 |         1 |       1.57x |       1.03x |       1.00x |
|   2048 |       64 |         0 |       1.58x |       0.95x |       0.95x |
|    256 |       64 |         2 |       1.57x |       1.03x |       1.01x |
|   2048 |       64 |         0 |       1.58x |       0.95x |       0.95x |
|    256 |       64 |         2 |       1.57x |       1.03x |       1.00x |
|   2048 |      128 |         0 |       1.35x |       0.90x |       0.96x |
|    256 |       64 |         3 |       1.57x |       1.03x |       1.00x |
|   2048 |      128 |         0 |       1.35x |       0.90x |       0.96x |
|    256 |       64 |         3 |       1.57x |       1.03x |       1.00x |
|   2048 |      256 |         0 |       1.10x |       0.84x |       1.01x |
|    256 |       64 |         4 |       1.57x |       1.03x |       1.00x |
|   2048 |      256 |         0 |       1.10x |       0.83x |       1.01x |
|    256 |       64 |         4 |       1.57x |       1.03x |       1.00x |
|   2048 |      512 |         0 |       1.03x |       0.80x |       0.98x |
|    256 |       64 |         5 |       1.57x |       1.03x |       1.00x |
|   2048 |      512 |         0 |       1.01x |       0.80x |       0.98x |
|    256 |       64 |         5 |       1.57x |       1.03x |       1.00x |
|   2048 |     1024 |         0 |       0.92x |       0.78x |       0.96x |
|    256 |       64 |         6 |       1.57x |       1.03x |       1.00x |
|   2048 |     1024 |         0 |       0.92x |       0.78x |       0.97x |
|    256 |       64 |         6 |       1.57x |       1.03x |       1.00x |
|   2048 |     2048 |         0 |       0.88x |       0.76x |       0.95x |
|    256 |       64 |         7 |       1.57x |       1.03x |       1.00x |
|   2048 |     2048 |         0 |       0.86x |       0.76x |       0.96x |
|    256 |       64 |         7 |       1.05x |       1.02x |       1.00x |
|      2 |        1 |         0 |       1.38x |       1.56x |       1.38x |
|      2 |        1 |         0 |       1.37x |       1.56x |       1.38x |
|      2 |        1 |         1 |       1.37x |       1.72x |       1.55x |
|      2 |        1 |         1 |       1.37x |       1.72x |       1.55x |
|      3 |        2 |         0 |       1.38x |       1.56x |       1.38x |
|      3 |        2 |         0 |       1.37x |       1.56x |       1.37x |
|      3 |        2 |         2 |       1.38x |       1.71x |       1.55x |
|      3 |        2 |         2 |       1.37x |       1.72x |       1.51x |
|      4 |        3 |         0 |       1.37x |       1.56x |       1.38x |
|      4 |        3 |         0 |       1.37x |       1.56x |       1.37x |
|      4 |        3 |         3 |       1.37x |       1.72x |       1.50x |
|      4 |        3 |         3 |       1.37x |       1.72x |       1.55x |
|      5 |        4 |         0 |       1.37x |       1.56x |       1.38x |
|      5 |        4 |         0 |       1.37x |       1.56x |       1.37x |
|      5 |        4 |         4 |       1.37x |       1.72x |       1.50x |
|      5 |        4 |         4 |       1.37x |       1.72x |       1.55x |
|      6 |        5 |         0 |       1.37x |       1.56x |       1.38x |
|      6 |        5 |         0 |       1.37x |       1.56x |       1.38x |
|      6 |        5 |         5 |       1.38x |       1.72x |       1.54x |
|      6 |        5 |         5 |       1.38x |       1.72x |       1.56x |
|      7 |        6 |         0 |       1.38x |       1.56x |       1.38x |
|      7 |        6 |         0 |       1.37x |       1.56x |       1.38x |
|      7 |        6 |         6 |       1.37x |       1.72x |       1.50x |
|      7 |        6 |         6 |       1.37x |       1.72x |       1.54x |
|      8 |        7 |         0 |       1.38x |       1.56x |       1.38x |
|      8 |        7 |         0 |       1.38x |       1.56x |       1.37x |
|      8 |        7 |         7 |       1.37x |       1.72x |       1.53x |
|      8 |        7 |         7 |       1.37x |       1.72x |       1.50x |
|      9 |        8 |         0 |       1.37x |       1.56x |       1.38x |
|      9 |        8 |         0 |       1.38x |       1.56x |       1.38x |
|      9 |        8 |         0 |       1.37x |       1.56x |       1.37x |
|      9 |        8 |         0 |       1.38x |       1.56x |       1.37x |
|     10 |        9 |         0 |       1.37x |       1.56x |       1.38x |
|     10 |        9 |         0 |       1.37x |       1.56x |       1.37x |
|     10 |        9 |         1 |       1.38x |       1.72x |       1.55x |
|     10 |        9 |         1 |       1.37x |       1.72x |       1.50x |
|     11 |       10 |         0 |       1.38x |       1.56x |       1.38x |
|     11 |       10 |         0 |       1.37x |       1.56x |       1.37x |
|     11 |       10 |         2 |       1.38x |       1.72x |       1.55x |
|     11 |       10 |         2 |       1.37x |       1.72x |       1.55x |
|     12 |       11 |         0 |       1.38x |       1.56x |       1.38x |
|     12 |       11 |         0 |       1.38x |       1.56x |       1.37x |
|     12 |       11 |         3 |       1.38x |       1.72x |       1.50x |
|     12 |       11 |         3 |       1.38x |       1.72x |       1.50x |
|     13 |       12 |         0 |       1.37x |       1.56x |       1.38x |
|     13 |       12 |         0 |       1.37x |       1.56x |       1.37x |
|     13 |       12 |         4 |       0.88x |       1.05x |       1.03x |
|     13 |       12 |         4 |       0.88x |       1.05x |       1.03x |
|     14 |       13 |         0 |       1.37x |       1.56x |       1.38x |
|     14 |       13 |         0 |       1.37x |       1.56x |       1.37x |
|     14 |       13 |         5 |       0.88x |       1.05x |       1.03x |
|     14 |       13 |         5 |       0.88x |       1.05x |       1.04x |
|     15 |       14 |         0 |       1.37x |       1.56x |       1.38x |
|     15 |       14 |         0 |       1.37x |       1.56x |       1.38x |
|     15 |       14 |         6 |       0.88x |       1.05x |       1.02x |
|     15 |       14 |         6 |       0.88x |       1.05x |       1.03x |
|     16 |       15 |         0 |       1.38x |       1.56x |       1.38x |
|     16 |       15 |         0 |       1.38x |       1.56x |       1.37x |
|     16 |       15 |         7 |       0.88x |       1.05x |       1.03x |
|     16 |       15 |         7 |       0.88x |       1.05x |       1.00x |
|     17 |       16 |         0 |       0.88x |       0.95x |       0.92x |
|     17 |       16 |         0 |       0.88x |       0.95x |       0.92x |
|     17 |       16 |         0 |       0.88x |       0.95x |       0.92x |
|     17 |       16 |         0 |       0.88x |       0.95x |       0.92x |
|     18 |       17 |         0 |       0.88x |       0.95x |       0.92x |
|     18 |       17 |         0 |       0.88x |       0.95x |       0.92x |
|     18 |       17 |         1 |       0.88x |       1.05x |       1.01x |
|     18 |       17 |         1 |       0.88x |       1.05x |       1.03x |
|     19 |       18 |         0 |       0.88x |       0.95x |       0.92x |
|     19 |       18 |         0 |       0.88x |       0.95x |       0.92x |
|     19 |       18 |         2 |       0.88x |       1.05x |       1.11x |
|     19 |       18 |         2 |       0.88x |       1.05x |       1.09x |
|     20 |       19 |         0 |       0.88x |       0.95x |       1.00x |
|     20 |       19 |         0 |       0.88x |       0.95x |       1.00x |
|     20 |       19 |         3 |       0.88x |       1.05x |       1.12x |
|     20 |       19 |         3 |       0.88x |       1.05x |       1.13x |
|     21 |       20 |         0 |       0.88x |       0.95x |       0.92x |
|     21 |       20 |         0 |       0.88x |       0.95x |       0.92x |
|     21 |       20 |         4 |       0.88x |       1.05x |       1.03x |
|     21 |       20 |         4 |       0.88x |       1.05x |       1.00x |
|     22 |       21 |         0 |       0.88x |       0.95x |       1.00x |
|     22 |       21 |         0 |       0.88x |       0.95x |       1.00x |
|     22 |       21 |         5 |       0.88x |       1.05x |       1.09x |
|     22 |       21 |         5 |       0.88x |       1.05x |       1.09x |
|     23 |       22 |         0 |       0.88x |       0.95x |       1.00x |
|     23 |       22 |         0 |       0.88x |       0.95x |       1.00x |
|     23 |       22 |         6 |       0.88x |       1.05x |       1.13x |
|     23 |       22 |         6 |       0.88x |       1.05x |       1.14x |
|     24 |       23 |         0 |       0.88x |       0.95x |       1.00x |
|     24 |       23 |         0 |       0.88x |       0.95x |       1.00x |
|     24 |       23 |         7 |       0.88x |       1.05x |       1.13x |
|     24 |       23 |         7 |       0.88x |       1.05x |       1.11x |
|     25 |       24 |         0 |       0.88x |       0.95x |       1.00x |
|     25 |       24 |         0 |       0.88x |       0.95x |       1.00x |
|     25 |       24 |         0 |       0.88x |       0.95x |       1.00x |
|     25 |       24 |         0 |       0.88x |       0.95x |       1.00x |
|     26 |       25 |         0 |       0.88x |       0.95x |       1.00x |
|     26 |       25 |         0 |       0.88x |       0.95x |       0.92x |
|     26 |       25 |         1 |       0.88x |       1.05x |       1.03x |
|     26 |       25 |         1 |       0.88x |       1.05x |       1.03x |
|     27 |       26 |         0 |       0.88x |       0.95x |       0.92x |
|     27 |       26 |         0 |       0.88x |       0.95x |       0.92x |
|     27 |       26 |         2 |       0.88x |       1.05x |       1.03x |
|     27 |       26 |         2 |       0.88x |       1.05x |       1.00x |
|     28 |       27 |         0 |       0.88x |       0.95x |       0.92x |
|     28 |       27 |         0 |       0.88x |       0.95x |       0.92x |
|     28 |       27 |         3 |       0.88x |       1.05x |       1.00x |
|     28 |       27 |         3 |       0.88x |       1.05x |       1.03x |
|     29 |       28 |         0 |       0.88x |       0.95x |       0.92x |
|     29 |       28 |         0 |       0.88x |       0.95x |       0.92x |
|     29 |       28 |         4 |       1.24x |       1.17x |       1.15x |
|     29 |       28 |         4 |       1.21x |       1.17x |       1.19x |
|     30 |       29 |         0 |       0.88x |       0.95x |       0.92x |
|     30 |       29 |         0 |       0.88x |       0.95x |       0.92x |
|     30 |       29 |         5 |       1.22x |       1.17x |       1.15x |
|     30 |       29 |         5 |       1.22x |       1.17x |       1.19x |
|     31 |       30 |         0 |       0.88x |       0.95x |       0.92x |
|     31 |       30 |         0 |       0.88x |       0.95x |       0.92x |
|     31 |       30 |         6 |       1.22x |       1.17x |       1.17x |
|     31 |       30 |         6 |       1.22x |       1.17x |       1.16x |
|     32 |       31 |         0 |       0.88x |       0.95x |       0.92x |
|     32 |       31 |         0 |       0.88x |       0.95x |       0.92x |
|     32 |       31 |         7 |       1.22x |       1.17x |       1.19x |
|     32 |       31 |         7 |       1.22x |       1.17x |       1.16x |

This patch passes the tests with no regressions.

8< --- 8< --- 8<
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 <gabor.kertesz@arm.com>
  

Comments

Szabolcs Nagy June 23, 2020, 4:57 p.m. UTC | #1
The 06/09/2020 16:06, Alex Butler wrote:
> 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 <gabor.kertesz@arm.com>

thanks, committed.
  

Patch

From a1c8d7d647174e8d029d06bdf48ed75ffbc19b21 Mon Sep 17 00:00:00 2001
From: Alex Butler <alex.butler@arm.com>
Date: Wed, 13 May 2020 22:46:33 +0100
Subject: [PATCH] aarch64: add MTE compatible memchr

---
 sysdeps/aarch64/memchr.S | 187 ++++++++++++++++++++---------------------------
 1 file changed, 80 insertions(+), 107 deletions(-)

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)
-- 
2.7.4