================
@@ -1302,6 +1302,7 @@ ItaniumRecordLayoutBuilder::LayoutBase(const
BaseSubobjectInfo *Base) {
setSize(std::max(getSize(), Offset + Layout.getSize()));
// Remember max struct/class alignment.
+ UnadjustedAlignment = std::max(UnadjustedAlignment, PreferredBaseAlign);
----------------
rjmccall wrote:
Does MSVC actually use the ARM ABI? Windows usually just follows its own rules.
Ah, well, doesn't really matter.
It seems natural that this would use the alignment of the base class rather
than its natural alignment, since that amounts to a consistent rule for all
subobjects. I mean, it doesn't *really* make sense to only ignore non-natural
alignment at the outermost level, but frankly this is a silly rule to begin
with. Every justification for honoring non-natural alignment at the subobject
level doubles as an argument for honoring non-natural alignment at the
aggregate level, which is the most sensible thing to do, and yet here we are.
Normally, whenever we're talking about base class alignment, we have to be very
careful to pick correctly between base-subobject alignment and complete-object
alignment. This should probably be base-subobject alignment, but I don't think
it's possible to detect a difference because any class with virtual bases is
not trivially-copyable and so will always be passed indirectly.
Darwin is just going to have to not follow the ARM spec here. We will not
consider an ABI break for something like this.
https://github.com/llvm/llvm-project/pull/135564
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
