diff mbox series

[v2] AArch64: Improve A64FX memset

Message ID VE1PR08MB5599AC76DCB2ABEE7A4CC2CE83189@VE1PR08MB5599.eurprd08.prod.outlook.com
State New
Headers show
Series [v2] AArch64: Improve A64FX memset | expand

Checks

Context Check Description
dj/TryBot-apply_patch success Patch applied to master at the time it was sent
dj/TryBot-32bit success Build for i686

Commit Message

Wilco Dijkstra July 9, 2021, 12:23 p.m. UTC
Hi Naohiro,

Here is version 2 which should improve things a lot:

v2: Improve handling of last 512 bytes which improves medium sized memsets.
    Further reduce codesize by removing unnecessary unrolling of dc zva.
    Speed up huge memsets of zero and non-zero.

Reduce the codesize of the A64FX memset by simplifying the small memset code,
better handling of alignment and last 8 vectors as well as removing redundant
instructions and branches. The size for memset goes down from 1032 to 376 bytes.
For large zeroing memsets use DC ZVA, which almost doubles performance. Large
non-zero memsets use the unroll8 loop which is about 10% faster.

Passes GLIBC regress, OK for commit?

---
diff mbox series

Patch

diff --git a/sysdeps/aarch64/multiarch/memset_a64fx.S b/sysdeps/aarch64/multiarch/memset_a64fx.S
index ce54e5418b08c8bc0ecc7affff68a59272ba6397..2737f0cba3e1a9ac887cd8072f6122f4852a9f94 100644
--- a/sysdeps/aarch64/multiarch/memset_a64fx.S
+++ b/sysdeps/aarch64/multiarch/memset_a64fx.S
@@ -30,11 +30,7 @@ 
 #define L2_SIZE         (8*1024*1024)	// L2 8MB - 1MB
 #define CACHE_LINE_SIZE	256
 #define PF_DIST_L1	(CACHE_LINE_SIZE * 16)	// Prefetch distance L1
-#define ZF_DIST		(CACHE_LINE_SIZE * 21)	// Zerofill distance
-#define rest		x8
 #define vector_length	x9
-#define vl_remainder	x10	// vector_length remainder
-#define cl_remainder	x11	// CACHE_LINE_SIZE remainder
 
 #if HAVE_AARCH64_SVE_ASM
 # if IS_IN (libc)
@@ -42,224 +38,126 @@ 
 
 	.arch armv8.2-a+sve
 
-	.macro dc_zva times
-	dc	zva, tmp1
-	add	tmp1, tmp1, CACHE_LINE_SIZE
-	.if \times-1
-	dc_zva "(\times-1)"
-	.endif
-	.endm
-
 	.macro st1b_unroll first=0, last=7
-	st1b	z0.b, p0, [dst, #\first, mul vl]
+	st1b	z0.b, p0, [dst, \first, mul vl]
 	.if \last-\first
 	st1b_unroll "(\first+1)", \last
 	.endif
 	.endm
 
-	.macro shortcut_for_small_size exit
-	// if rest <= vector_length * 2
+
+#undef BTI_C
+#define BTI_C
+
+ENTRY (MEMSET)
+	PTR_ARG (0)
+	SIZE_ARG (2)
+
+	dup	z0.b, valw
 	whilelo	p0.b, xzr, count
+	cntb	vector_length
 	whilelo	p1.b, vector_length, count
+	st1b	z0.b, p0, [dstin, 0, mul vl]
+	st1b	z0.b, p1, [dstin, 1, mul vl]
 	b.last	1f
-	st1b	z0.b, p0, [dstin, #0, mul vl]
-	st1b	z0.b, p1, [dstin, #1, mul vl]
 	ret
-1:	// if rest > vector_length * 8
-	cmp	count, vector_length, lsl 3	// vector_length * 8
-	b.hi	\exit
-	// if rest <= vector_length * 4
-	lsl	tmp1, vector_length, 1	// vector_length * 2
-	whilelo	p2.b, tmp1, count
-	incb	tmp1
-	whilelo	p3.b, tmp1, count
-	b.last	1f
-	st1b	z0.b, p0, [dstin, #0, mul vl]
-	st1b	z0.b, p1, [dstin, #1, mul vl]
-	st1b	z0.b, p2, [dstin, #2, mul vl]
-	st1b	z0.b, p3, [dstin, #3, mul vl]
-	ret
-1:	// if rest <= vector_length * 8
-	lsl	tmp1, vector_length, 2	// vector_length * 4
-	whilelo	p4.b, tmp1, count
-	incb	tmp1
-	whilelo	p5.b, tmp1, count
-	b.last	1f
-	st1b	z0.b, p0, [dstin, #0, mul vl]
-	st1b	z0.b, p1, [dstin, #1, mul vl]
-	st1b	z0.b, p2, [dstin, #2, mul vl]
-	st1b	z0.b, p3, [dstin, #3, mul vl]
-	st1b	z0.b, p4, [dstin, #4, mul vl]
-	st1b	z0.b, p5, [dstin, #5, mul vl]
-	ret
-1:	lsl	tmp1, vector_length, 2	// vector_length * 4
-	incb	tmp1			// vector_length * 5
-	incb	tmp1			// vector_length * 6
-	whilelo	p6.b, tmp1, count
-	incb	tmp1
-	whilelo	p7.b, tmp1, count
-	st1b	z0.b, p0, [dstin, #0, mul vl]
-	st1b	z0.b, p1, [dstin, #1, mul vl]
-	st1b	z0.b, p2, [dstin, #2, mul vl]
-	st1b	z0.b, p3, [dstin, #3, mul vl]
-	st1b	z0.b, p4, [dstin, #4, mul vl]
-	st1b	z0.b, p5, [dstin, #5, mul vl]
-	st1b	z0.b, p6, [dstin, #6, mul vl]
-	st1b	z0.b, p7, [dstin, #7, mul vl]
-	ret
-	.endm
 
-ENTRY (MEMSET)
-
-	PTR_ARG (0)
-	SIZE_ARG (2)
+	// count >= vector_length * 2
+	.p2align 4
+1:	add	dst, dstin, count
+	cmp	count, vector_length, lsl 2
+	b.hi	1f
+	st1b	z0.b, p0, [dst, -2, mul vl]
+	st1b	z0.b, p0, [dst, -1, mul vl]
+	ret
 
-	cbnz	count, 1f
+	// count > vector_length * 4
+1:	cmp	count, vector_length, lsl 3
+	b.hi	L(vl_agnostic)
+	st1b	z0.b, p0, [dstin, 2, mul vl]
+	st1b	z0.b, p0, [dstin, 3, mul vl]
+	st1b	z0.b, p0, [dst, -4, mul vl]
+	st1b	z0.b, p0, [dst, -3, mul vl]
+	st1b	z0.b, p0, [dst, -2, mul vl]
+	st1b	z0.b, p0, [dst, -1, mul vl]
 	ret
-1:	dup	z0.b, valw
-	cntb	vector_length
-	// shortcut for less than vector_length * 8
-	// gives a free ptrue to p0.b for n >= vector_length
-	shortcut_for_small_size L(vl_agnostic)
-	// end of shortcut
 
-L(vl_agnostic): // VL Agnostic
-	mov	rest, count
+	// count >= vector_length * 8
+	.p2align 4
+L(vl_agnostic):
 	mov	dst, dstin
-	add	dstend, dstin, count
-	// if rest >= L2_SIZE && vector_length == 64 then L(L2)
 	mov	tmp1, 64
-	cmp	rest, L2_SIZE
-	ccmp	vector_length, tmp1, 0, cs
-	b.eq	L(L2)
-	// if rest >= L1_SIZE && vector_length == 64 then L(L1_prefetch)
-	cmp	rest, L1_SIZE
+	// if count >= L1_SIZE && vector_length == 64 then L(L1_prefetch)
+	cmp	count, L1_SIZE
 	ccmp	vector_length, tmp1, 0, cs
 	b.eq	L(L1_prefetch)
 
-L(unroll32):
-	lsl	tmp1, vector_length, 3	// vector_length * 8
-	lsl	tmp2, vector_length, 5	// vector_length * 32
-	.p2align 3
-1:	cmp	rest, tmp2
-	b.cc	L(unroll8)
-	st1b_unroll
-	add	dst, dst, tmp1
-	st1b_unroll
-	add	dst, dst, tmp1
-	st1b_unroll
-	add	dst, dst, tmp1
-	st1b_unroll
-	add	dst, dst, tmp1
-	sub	rest, rest, tmp2
-	b	1b
-
+	// count >= 8 * vector_length
 L(unroll8):
 	lsl	tmp1, vector_length, 3
-	.p2align 3
-1:	cmp	rest, tmp1
-	b.cc	L(last)
-	st1b_unroll
+	sub	count, count, tmp1
+	lsl	tmp2, vector_length, 1
+	.p2align 4
+1:	subs	count, count, tmp1
+	st1b_unroll 0, 7
 	add	dst, dst, tmp1
-	sub	rest, rest, tmp1
-	b	1b
-
-L(last):
-	whilelo	p0.b, xzr, rest
-	whilelo	p1.b, vector_length, rest
-	b.last	1f
-	st1b	z0.b, p0, [dst, #0, mul vl]
-	st1b	z0.b, p1, [dst, #1, mul vl]
-	ret
-1:	lsl	tmp1, vector_length, 1	// vector_length * 2
-	whilelo	p2.b, tmp1, rest
-	incb	tmp1
-	whilelo	p3.b, tmp1, rest
-	b.last	1f
-	st1b	z0.b, p0, [dst, #0, mul vl]
-	st1b	z0.b, p1, [dst, #1, mul vl]
-	st1b	z0.b, p2, [dst, #2, mul vl]
-	st1b	z0.b, p3, [dst, #3, mul vl]
-	ret
-1:	lsl	tmp1, vector_length, 2	// vector_length * 4
-	whilelo	p4.b, tmp1, rest
-	incb	tmp1
-	whilelo	p5.b, tmp1, rest
-	incb	tmp1
-	whilelo	p6.b, tmp1, rest
-	incb	tmp1
-	whilelo	p7.b, tmp1, rest
-	st1b	z0.b, p0, [dst, #0, mul vl]
-	st1b	z0.b, p1, [dst, #1, mul vl]
-	st1b	z0.b, p2, [dst, #2, mul vl]
-	st1b	z0.b, p3, [dst, #3, mul vl]
-	st1b	z0.b, p4, [dst, #4, mul vl]
-	st1b	z0.b, p5, [dst, #5, mul vl]
-	st1b	z0.b, p6, [dst, #6, mul vl]
-	st1b	z0.b, p7, [dst, #7, mul vl]
+	b.hi	1b
+
+	add	dst, dst, count
+	add	count, count, tmp1
+	cmp	count, tmp2
+	b.ls	2f
+	add	tmp2, vector_length, vector_length, lsl 2
+	cmp	count, tmp2
+	b.ls	5f
+	st1b	z0.b, p0, [dst, 0, mul vl]
+	st1b	z0.b, p0, [dst, 1, mul vl]
+	st1b	z0.b, p0, [dst, 2, mul vl]
+5:	st1b	z0.b, p0, [dst, 3, mul vl]
+	st1b	z0.b, p0, [dst, 4, mul vl]
+	st1b	z0.b, p0, [dst, 5, mul vl]
+2:	st1b	z0.b, p0, [dst, 6, mul vl]
+	st1b	z0.b, p0, [dst, 7, mul vl]
 	ret
 
-L(L1_prefetch): // if rest >= L1_SIZE
+	// count >= L1_SIZE
 	.p2align 3
+L(L1_prefetch):
+	cmp	count, L2_SIZE
+	b.hs	L(L2)
 1:	st1b_unroll 0, 3
 	prfm	pstl1keep, [dst, PF_DIST_L1]
 	st1b_unroll 4, 7
 	prfm	pstl1keep, [dst, PF_DIST_L1 + CACHE_LINE_SIZE]
 	add	dst, dst, CACHE_LINE_SIZE * 2
-	sub	rest, rest, CACHE_LINE_SIZE * 2
-	cmp	rest, L1_SIZE
-	b.ge	1b
-	cbnz	rest, L(unroll32)
-	ret
+	sub	count, count, CACHE_LINE_SIZE * 2
+	cmp	count, PF_DIST_L1
+	b.hs	1b
+	b	L(unroll8)
 
+	// count >= L2_SIZE
 L(L2):
-	// align dst address at vector_length byte boundary
-	sub	tmp1, vector_length, 1
-	ands	tmp2, dst, tmp1
-	// if vl_remainder == 0
-	b.eq	1f
-	sub	vl_remainder, vector_length, tmp2
-	// process remainder until the first vector_length boundary
-	whilelt	p2.b, xzr, vl_remainder
-	st1b	z0.b, p2, [dst]
-	add	dst, dst, vl_remainder
-	sub	rest, rest, vl_remainder
-	// align dstin address at CACHE_LINE_SIZE byte boundary
-1:	mov	tmp1, CACHE_LINE_SIZE
-	ands	tmp2, dst, CACHE_LINE_SIZE - 1
-	// if cl_remainder == 0
-	b.eq	L(L2_dc_zva)
-	sub	cl_remainder, tmp1, tmp2
-	// process remainder until the first CACHE_LINE_SIZE boundary
-	mov	tmp1, xzr       // index
-2:	whilelt	p2.b, tmp1, cl_remainder
-	st1b	z0.b, p2, [dst, tmp1]
-	incb	tmp1
-	cmp	tmp1, cl_remainder
-	b.lo	2b
-	add	dst, dst, cl_remainder
-	sub	rest, rest, cl_remainder
-
-L(L2_dc_zva):
-	// zero fill
-	mov	tmp1, dst
-	dc_zva	(ZF_DIST / CACHE_LINE_SIZE) - 1
-	mov	zva_len, ZF_DIST
-	add	tmp1, zva_len, CACHE_LINE_SIZE * 2
-	// unroll
-	.p2align 3
-1:	st1b_unroll 0, 3
-	add	tmp2, dst, zva_len
-	dc	 zva, tmp2
-	st1b_unroll 4, 7
-	add	tmp2, tmp2, CACHE_LINE_SIZE
-	dc	zva, tmp2
-	add	dst, dst, CACHE_LINE_SIZE * 2
-	sub	rest, rest, CACHE_LINE_SIZE * 2
-	cmp	rest, tmp1	// ZF_DIST + CACHE_LINE_SIZE * 2
-	b.ge	1b
-	cbnz	rest, L(unroll8)
-	ret
+	tst	valw, 255
+	b.ne	L(unroll8)
+	// align dst to CACHE_LINE_SIZE byte boundary
+	and	tmp1, dst, CACHE_LINE_SIZE - 1
+	sub	tmp1, tmp1, CACHE_LINE_SIZE
+	st1b	z0.b, p0, [dst, 0, mul vl]
+	st1b	z0.b, p0, [dst, 1, mul vl]
+	st1b	z0.b, p0, [dst, 2, mul vl]
+	st1b	z0.b, p0, [dst, 3, mul vl]
+	sub	dst, dst, tmp1
+	add	count, count, tmp1
+
+	// clear cachelines using DC ZVA
+	sub	count, count, CACHE_LINE_SIZE * 4
+	.p2align 4
+1:	dc	zva, dst
+	add	dst, dst, CACHE_LINE_SIZE
+	subs	count, count, CACHE_LINE_SIZE
+	b.hs	1b
+	add	count, count, CACHE_LINE_SIZE * 4
+	b	L(unroll8)
 
 END (MEMSET)
 libc_hidden_builtin_def (MEMSET)