|
 |
a41fe4 |
From 186092c6ba8825598ffdbf15dbf0823c771f560d Mon Sep 17 00:00:00 2001
|
|
|
a41fe4 |
From: Wilco Dijkstra <wdijkstr@arm.com>
|
|
|
a41fe4 |
Date: Tue, 10 Aug 2021 13:42:07 +0100
|
|
|
a41fe4 |
Subject: [PATCH] [3/5] AArch64: Improve A64FX memset for remaining bytes
|
|
|
a41fe4 |
|
|
|
a41fe4 |
Simplify handling of remaining bytes. Avoid lots of taken branches and complex
|
|
|
a41fe4 |
whilelo computations, instead unconditionally write vectors from the end.
|
|
|
a41fe4 |
|
|
|
a41fe4 |
Reviewed-by: Naohiro Tamura <naohirot@fujitsu.com>
|
|
|
a41fe4 |
---
|
|
|
a41fe4 |
sysdeps/aarch64/multiarch/memset_a64fx.S | 46 +++++++-----------------
|
|
|
a41fe4 |
1 file changed, 13 insertions(+), 33 deletions(-)
|
|
|
a41fe4 |
|
|
|
a41fe4 |
diff --git a/sysdeps/aarch64/multiarch/memset_a64fx.S b/sysdeps/aarch64/multiarch/memset_a64fx.S
|
|
|
a41fe4 |
index 75cf43ae79..337c86be6f 100644
|
|
|
a41fe4 |
--- a/sysdeps/aarch64/multiarch/memset_a64fx.S
|
|
|
a41fe4 |
+++ b/sysdeps/aarch64/multiarch/memset_a64fx.S
|
|
|
a41fe4 |
@@ -130,38 +130,19 @@ L(unroll8):
|
|
|
a41fe4 |
b 1b
|
|
|
a41fe4 |
|
|
|
a41fe4 |
L(last):
|
|
|
a41fe4 |
- whilelo p0.b, xzr, rest
|
|
|
a41fe4 |
- whilelo p1.b, vector_length, rest
|
|
|
a41fe4 |
- b.last 1f
|
|
|
a41fe4 |
- st1b z0.b, p0, [dst, #0, mul vl]
|
|
|
a41fe4 |
- st1b z0.b, p1, [dst, #1, mul vl]
|
|
|
a41fe4 |
- ret
|
|
|
a41fe4 |
-1: lsl tmp1, vector_length, 1 // vector_length * 2
|
|
|
a41fe4 |
- whilelo p2.b, tmp1, rest
|
|
|
a41fe4 |
- incb tmp1
|
|
|
a41fe4 |
- whilelo p3.b, tmp1, rest
|
|
|
a41fe4 |
- b.last 1f
|
|
|
a41fe4 |
- st1b z0.b, p0, [dst, #0, mul vl]
|
|
|
a41fe4 |
- st1b z0.b, p1, [dst, #1, mul vl]
|
|
|
a41fe4 |
- st1b z0.b, p2, [dst, #2, mul vl]
|
|
|
a41fe4 |
- st1b z0.b, p3, [dst, #3, mul vl]
|
|
|
a41fe4 |
- ret
|
|
|
a41fe4 |
-1: lsl tmp1, vector_length, 2 // vector_length * 4
|
|
|
a41fe4 |
- whilelo p4.b, tmp1, rest
|
|
|
a41fe4 |
- incb tmp1
|
|
|
a41fe4 |
- whilelo p5.b, tmp1, rest
|
|
|
a41fe4 |
- incb tmp1
|
|
|
a41fe4 |
- whilelo p6.b, tmp1, rest
|
|
|
a41fe4 |
- incb tmp1
|
|
|
a41fe4 |
- whilelo p7.b, tmp1, rest
|
|
|
a41fe4 |
- st1b z0.b, p0, [dst, #0, mul vl]
|
|
|
a41fe4 |
- st1b z0.b, p1, [dst, #1, mul vl]
|
|
|
a41fe4 |
- st1b z0.b, p2, [dst, #2, mul vl]
|
|
|
a41fe4 |
- st1b z0.b, p3, [dst, #3, mul vl]
|
|
|
a41fe4 |
- st1b z0.b, p4, [dst, #4, mul vl]
|
|
|
a41fe4 |
- st1b z0.b, p5, [dst, #5, mul vl]
|
|
|
a41fe4 |
- st1b z0.b, p6, [dst, #6, mul vl]
|
|
|
a41fe4 |
- st1b z0.b, p7, [dst, #7, mul vl]
|
|
|
a41fe4 |
+ cmp count, vector_length, lsl 1
|
|
|
a41fe4 |
+ b.ls 2f
|
|
|
a41fe4 |
+ add tmp2, vector_length, vector_length, lsl 2
|
|
|
a41fe4 |
+ cmp count, tmp2
|
|
|
a41fe4 |
+ b.ls 5f
|
|
|
a41fe4 |
+ st1b z0.b, p0, [dstend, -8, mul vl]
|
|
|
a41fe4 |
+ st1b z0.b, p0, [dstend, -7, mul vl]
|
|
|
a41fe4 |
+ st1b z0.b, p0, [dstend, -6, mul vl]
|
|
|
a41fe4 |
+5: st1b z0.b, p0, [dstend, -5, mul vl]
|
|
|
a41fe4 |
+ st1b z0.b, p0, [dstend, -4, mul vl]
|
|
|
a41fe4 |
+ st1b z0.b, p0, [dstend, -3, mul vl]
|
|
|
a41fe4 |
+2: st1b z0.b, p0, [dstend, -2, mul vl]
|
|
|
a41fe4 |
+ st1b z0.b, p0, [dstend, -1, mul vl]
|
|
|
a41fe4 |
ret
|
|
|
a41fe4 |
|
|
|
a41fe4 |
L(L1_prefetch): // if rest >= L1_SIZE
|
|
|
a41fe4 |
@@ -199,7 +180,6 @@ L(L2):
|
|
|
a41fe4 |
subs count, count, CACHE_LINE_SIZE
|
|
|
a41fe4 |
b.hi 1b
|
|
|
a41fe4 |
add count, count, CACHE_LINE_SIZE
|
|
|
a41fe4 |
- add dst, dst, CACHE_LINE_SIZE
|
|
|
a41fe4 |
b L(last)
|
|
|
a41fe4 |
|
|
|
a41fe4 |
END (MEMSET)
|
|
|
a41fe4 |
--
|
|
|
a41fe4 |
2.31.1
|
|
|
a41fe4 |
|