================
@@ -298,11 +298,15 @@ namespace {
}
Address AtomicInfo::CreateTempAlloca() const {
- Address TempAlloca = CGF.CreateMemTemp(
- (LVal.isBitField() && ValueSizeInBits > AtomicSizeInBits) ? ValueTy
- : AtomicTy,
- getAtomicAlignment(),
- "atomic-temp");
+ QualType TmpTy;
+ // Remove addrspace info from the atomic pointer element when making the
+ // alloca pointer element.
+ if (LVal.isBitField() && ValueSizeInBits > AtomicSizeInBits)
+ TmpTy = ValueTy;
+ else
+ TmpTy = AtomicTy.getUnqualifiedType();
----------------
vtjnash wrote:
So the complication here is that we're getting an atomic operation like `T
_Atomic volatile addrspace(1)*` for some `T` which isn't an integer for some
memory that probably isn't the stack, so now that atomic load needs to be
bitcast into the actual `T`. But the C spec doesn't quite provide accurate
enough information on whether it is acceptable to store `T` to the stack with
all of those qualifiers. Prior to this PR, the qualifiers of T would be
implicitly ignored when making the pointer and load/store for the on-stack
bitcast here. But after this PR, getting the unqualified type for this purpose
needs to be done explicitly instead.
https://github.com/llvm/llvm-project/pull/181390
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
