[release-branch.go1.23] cmd/compile: for arm64 epilog, do SP increment with a single instruction
That way, the frame is atomically popped. Previously, for big frames
the SP was unwound in two steps (because arm64 can only add constants
up to 1<<12 in a single instruction).
Fixes #74693
Change-Id: I382c249194ad7bc9fc19607c27487c58d90d49e5
Reviewed-on: https://go-review.googlesource.com/c/go/+/689235
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
Reviewed-by: Keith Randall <khr@google.com>
(cherry picked from commit f7cc61e7d7f77521e073137c6045ba73f66ef902)
Reviewed-on: https://go-review.googlesource.com/c/go/+/689595
diff --git a/src/cmd/internal/obj/arm64/obj7.go b/src/cmd/internal/obj/arm64/obj7.go
index 20498bc..f15b6a9 100644
--- a/src/cmd/internal/obj/arm64/obj7.go
+++ b/src/cmd/internal/obj/arm64/obj7.go
@@ -882,18 +882,49 @@
p.To.Reg = REGFP
p.To.Offset = REGLINK
- // ADD $aoffset, RSP, RSP
- q = newprog()
- q.As = AADD
- q.From.Type = obj.TYPE_CONST
- q.From.Offset = int64(aoffset)
- q.To.Type = obj.TYPE_REG
- q.To.Reg = REGSP
- q.Spadj = -aoffset
- q.Pos = p.Pos
- q.Link = p.Link
- p.Link = q
- p = q
+ if aoffset < 1<<12 {
+ // ADD $aoffset, RSP, RSP
+ q = newprog()
+ q.As = AADD
+ q.From.Type = obj.TYPE_CONST
+ q.From.Offset = int64(aoffset)
+ q.To.Type = obj.TYPE_REG
+ q.To.Reg = REGSP
+ q.Spadj = -aoffset
+ q.Pos = p.Pos
+ q.Link = p.Link
+ p.Link = q
+ p = q
+ } else {
+ // Put frame size in a separate register and
+ // add it in with a single instruction,
+ // so we never have a partial frame during
+ // the epilog. See issue 73259.
+
+ // MOVD $aoffset, REGTMP
+ q = newprog()
+ q.As = AMOVD
+ q.From.Type = obj.TYPE_CONST
+ q.From.Offset = int64(aoffset)
+ q.To.Type = obj.TYPE_REG
+ q.To.Reg = REGTMP
+ q.Pos = p.Pos
+ q.Link = p.Link
+ p.Link = q
+ p = q
+ // ADD REGTMP, RSP, RSP
+ q = newprog()
+ q.As = AADD
+ q.From.Type = obj.TYPE_REG
+ q.From.Reg = REGTMP
+ q.To.Type = obj.TYPE_REG
+ q.To.Reg = REGSP
+ q.Spadj = -aoffset
+ q.Pos = p.Pos
+ q.Link = p.Link
+ p.Link = q
+ p = q
+ }
}
// If enabled, this code emits 'MOV PC, R27' before every 'MOV LR, PC',
