[CUDA] __constant__ aggregate initializer values return undef in host compilation

[Artem-B]

__constant__ aggregate initializer values return undef in host compilation, but things work when we access individual field values or make a local copy.

Reproducer:

• CUDA: https://godbolt.org/z/7ahK78end
• HIP: https://godbolt.org/z/9EW1dTq7s

struct S { int s; };
const __constant__ S const_s = { 77 };

__host__ S works () {
    S s = const_s;
    return s;  // returns {77}
}

__host__ S broken () {
    return const_s;  // returns undef
}

[Artem-B]

@yxsamliu ^^^ FYI. Looks like it affects HIP, too.

[Artem-B]

Things may be a bit worse. Static initializers of aggregates themselves may be broken on the host side, not just accessing the values of known initializers. Or it could be the same problem, just happening a bit earlier.

Here's an example where a local copy no longer works: https://godbolt.org/z/jf95vYhx9

Things still do work if we access an initialized local static __constant__ struct.

[yxsamliu]

At -O0, the emitted IR uses memcpy to copy from the shadow var https://godbolt.org/z/957d89Ma8

So it seems we need to let the shadow var have the initializer in host compilation if we want the code to work with -O0. This seems to be OK since const means immutable, so we have guarantee that the host and device side have the same immutable content in that constant var.

[Artem-B]

What puzzles me is that memcpy is used in both working and non-working cases, but in working case we copy from a fully initialized instance. Why do they have different sources for the memcpy?

@__const.works().s = private unnamed_addr constant %struct.S { i32 77 }, align 4
@const_s = internal constant %struct.S undef, align 4

Few more cases where we allow seeing the initializer value: https://godbolt.org/z/KbMWoMqe7

I think that the source of the inconsistency is that the front end still accesses the initializer value of the variable in cases where it should've used the shadow's initializer. Whether/when we want to propagate the initializer to the shadow is a separate question.