cmd/compile: don't elide zero extension on top of signed values

v = ... compute some value, which zeros top 32 bits ...
w = zero-extend v

We want to remove the zero-extension operation, as it doesn't do anything.
But if v is typed as a signed value, and it gets spilled/restored, it
might be re-sign-extended upon restore. So the zero-extend isn't actually
a NOP when there might be calls or other reasons to spill in between v and w.

Fixes #68227

Change-Id: I3b30b8e56c7d70deac1fb09d2becc7395acbadf8
Reviewed-on: https://go-review.googlesource.com/c/go/+/595675
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Joedian Reid <joedian@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
This commit is contained in:
Keith Randall 2024-06-27 20:45:22 -07:00
parent ea537cca31
commit 7f90b960a9
2 changed files with 57 additions and 3 deletions

View file

@ -1287,6 +1287,11 @@ func areAdjacentOffsets(off1, off2, size int64) bool {
// depth limits recursion depth. In AMD64.rules 3 is used as limit,
// because it catches same amount of cases as 4.
func zeroUpper32Bits(x *Value, depth int) bool {
if x.Type.IsSigned() && x.Type.Size() < 8 {
// If the value is signed, it might get re-sign-extended
// during spill and restore. See issue 68227.
return false
}
switch x.Op {
case OpAMD64MOVLconst, OpAMD64MOVLload, OpAMD64MOVLQZX, OpAMD64MOVLloadidx1,
OpAMD64MOVWload, OpAMD64MOVWloadidx1, OpAMD64MOVBload, OpAMD64MOVBloadidx1,
@ -1305,7 +1310,7 @@ func zeroUpper32Bits(x *Value, depth int) bool {
case OpArg: // note: but not ArgIntReg
// amd64 always loads args from the stack unsigned.
// most other architectures load them sign/zero extended based on the type.
return x.Type.Size() == 4 && (x.Type.IsUnsigned() || x.Block.Func.Config.arch == "amd64")
return x.Type.Size() == 4 && x.Block.Func.Config.arch == "amd64"
case OpPhi, OpSelect0, OpSelect1:
// Phis can use each-other as an arguments, instead of tracking visited values,
// just limit recursion depth.
@ -1325,11 +1330,14 @@ func zeroUpper32Bits(x *Value, depth int) bool {
// zeroUpper48Bits is similar to zeroUpper32Bits, but for upper 48 bits.
func zeroUpper48Bits(x *Value, depth int) bool {
if x.Type.IsSigned() && x.Type.Size() < 8 {
return false
}
switch x.Op {
case OpAMD64MOVWQZX, OpAMD64MOVWload, OpAMD64MOVWloadidx1, OpAMD64MOVWloadidx2:
return true
case OpArg: // note: but not ArgIntReg
return x.Type.Size() == 2 && (x.Type.IsUnsigned() || x.Block.Func.Config.arch == "amd64")
return x.Type.Size() == 2 && x.Block.Func.Config.arch == "amd64"
case OpPhi, OpSelect0, OpSelect1:
// Phis can use each-other as an arguments, instead of tracking visited values,
// just limit recursion depth.
@ -1349,11 +1357,14 @@ func zeroUpper48Bits(x *Value, depth int) bool {
// zeroUpper56Bits is similar to zeroUpper32Bits, but for upper 56 bits.
func zeroUpper56Bits(x *Value, depth int) bool {
if x.Type.IsSigned() && x.Type.Size() < 8 {
return false
}
switch x.Op {
case OpAMD64MOVBQZX, OpAMD64MOVBload, OpAMD64MOVBloadidx1:
return true
case OpArg: // note: but not ArgIntReg
return x.Type.Size() == 1 && (x.Type.IsUnsigned() || x.Block.Func.Config.arch == "amd64")
return x.Type.Size() == 1 && x.Block.Func.Config.arch == "amd64"
case OpPhi, OpSelect0, OpSelect1:
// Phis can use each-other as an arguments, instead of tracking visited values,
// just limit recursion depth.