diff options
| author | Mark Kettenis <kettenis@cvs.openbsd.org> | 2020-10-09 17:36:48 +0000 |
|---|---|---|
| committer | Mark Kettenis <kettenis@cvs.openbsd.org> | 2020-10-09 17:36:48 +0000 |
| commit | 967365ce097b36c5276c49b77ee3d8b9fb1b8f1c (patch) | |
| tree | 3d0dd9b8eded0436a5f7e46986492665d17b5225 /sys | |
| parent | 7aa42f974b2ff58ece3fbf84e22fb69c5f501293 (diff) | |
Optimize copyin(9), copyout(9) and kcopy(9) by doing 16-byte copies if
possible. No doubt further optimizations are poissible, but this
change results in a nice balance between code size and speed and is
still easy to understand.
ok patrick@
Diffstat (limited to 'sys')
| -rw-r--r-- | sys/arch/arm64/arm64/copy.S | 54 |
1 files changed, 42 insertions, 12 deletions
diff --git a/sys/arch/arm64/arm64/copy.S b/sys/arch/arm64/arm64/copy.S index db2a6c393a2..1a4ed25b080 100644 --- a/sys/arch/arm64/arm64/copy.S +++ b/sys/arch/arm64/arm64/copy.S @@ -1,4 +1,4 @@ -/* $OpenBSD: copy.S,v 1.8 2020/10/04 20:03:57 kettenis Exp $ */ +/* $OpenBSD: copy.S,v 1.9 2020/10/09 17:36:47 kettenis Exp $ */ /* * Copyright (c) 2015 Dale Rahn <drahn@dalerahn.com> * Copyright (c) 2014 Patrick Wildt <patrick@blueri.se> @@ -52,13 +52,25 @@ ENTRY(copyin) adr x5, .Lcopyfault str x5, [x3, #(PCB_ONFAULT)] // set handler -// This probably should be optimized -2: ldtrb w6, [x0] + cmp x2, #16 + b.lo .Lcopyin1 +2: ldtr x6, [x0] + ldtr x7, [x0, #8] + stp x6, x7, [x1], #16 + add x0, x0, #16 + sub x2, x2, #16 + cmp x2, #16 + b.hs 2b + +.Lcopyin1: + cbz x2, .Lcopyin0 +3: ldtrb w6, [x0] strb w6, [x1], #1 add x0, x0, #1 sub x2, x2, #1 - cbnz x2, 2b + cbnz x2, 3b +.Lcopyin0: str x4, [x3, #(PCB_ONFAULT)] // clear handler mov x0, xzr RETGUARD_CHECK(copy, x15) @@ -130,13 +142,25 @@ ENTRY(copyout) adr x5, .Lcopyfault str x5, [x3, #(PCB_ONFAULT)] // set handler -// This probably should be optimized -2: ldrb w6, [x0], #1 + cmp x2, #16 + b.lo .Lcopyout1 +2: ldp x6, x7, [x0], #16 + sttr x6, [x1] + sttr x7, [x1, #8] + add x1, x1, #16 + sub x2, x2, #16 + cmp x2, #16 + b.hs 2b + +.Lcopyout1: + cbz x2, .Lcopyout0 +3: ldrb w6, [x0], #1 sttrb w6, [x1] add x1, x1, #1 sub x2, x2, #1 - cbnz x2, 2b + cbnz x2, 3b +.Lcopyout0: str x4, [x3, #(PCB_ONFAULT)] // clear handler mov x0, xzr RETGUARD_CHECK(copy, x15) @@ -163,13 +187,19 @@ ENTRY(kcopy) adr x5, .Lcopyfault str x5, [x3, #(PCB_ONFAULT)] // set handler - cmp x2, #8 - b.lo .Lkcopy4 -2: ldr x6, [x0], #8 + cmp x2, #16 + b.lo .Lkcopy8 +2: ldp x6, x7, [x0], #16 + stp x6, x7, [x1], #16 + sub x2, x2, #16 + cmp x2, #16 + b.hs 2b + +.Lkcopy8: + tbz x2, #3, .Lkcopy4 + ldr x6, [x0], #8 str x6, [x1], #8 sub x2, x2, #8 - cmp x2, #8 - b.hs 2b .Lkcopy4: tbz x2, #2, .Lkcopy1 |