Primarily uploading to get your opinion on whether this is a MSSA bug, or insanity that shouldn't exist in IR.
After inlining, EarlyCSE may be given code like:
; snip %.fca.0.gep.i = getelementptr inbounds [2 x i64], [2 x i64]* %retval.i, i64 0, i64 0 %.fca.1.gep.i = getelementptr inbounds [2 x i64], [2 x i64]* %retval.i, i64 0, i64 1 %.fca.1.load.i = load i64, i64* %.fca.1.gep.i, align 8 call void @llvm.lifetime.end.p0i8(i64 16, i8* %2) %23 = bitcast %class.SkRecorder* %this to %class.SkCanvas* call void @llvm.lifetime.start.p0i8(i64 16, i8* %2) #8 store i64 0, i64* %.fca.0.gep.i, align 8 store i64 %.fca.1.load.i, i64* %.fca.1.gep.i, align 8 call void @whatever(%class.SkCanvas* %23, %struct.SkIRect* dereferenceable(16) %tmpcast.i) ; snip
...where %2 and %tmpcast.i are bitcasted %retval.i pointers. This is a pretty roundabout way of saying "store 0 into %.fca.0.gep.i", but doesn't venture into illegal IR AFAICT.
The issue is that "%fca.1.load.i MustAlias %2" is false, so MSSA doesn't treat it as a clobber for the store into %.fca.1.gep.i. Hence, EarlyCSE considers the last store to be redundant, and eliminates it. Later passes then eliminate the now-dead %.fca.1.load.i, then the store that paired with %.fca.1.load.i dies, presumably due to being unused before lifetime.end, etc.
I imagine we can get into a similar situation if we had code like:
%foo = select %whatever, %ptr1, %ptr2 ; %ptr1 NoAlias %ptr2 %f = load i8, i8* %foo call void @llvm.lifetime.end.p0i8(i64 8, i8* %ptr1) call void @llvm.lifetime.start.p0i8(i64 8, i8* %ptr1) store i8, i8* %foo call void @use(i8* %foo)
...We'd eliminate the store, an optimization after us turns %whatever into a constant truthy value, ...