[LoongArch] Support lowering FrameIndex

[wangleiat]

Summary

Ensure callee-saved registers are accessed relative to the stackpointer.

Depends on D128429

[wangleiat]

Update test to use opaque pointers

[wangleiat]

rebase

[xen0n]

The other changes look good.

llvm/lib/Target/LoongArch/LoongArchFrameLowering.cpp

68–69

This implementation is a simplified version of RISCV's getFrameIndexReference, but we're missing a stack ID check before this. I know we're not utilizing any non-default stack IDs so far but I don't know if such a check/assertion should be kept anyway.

[wangleiat]

llvm/lib/Target/LoongArch/LoongArchFrameLowering.cpp

68–69

Riscv uses stackID to handle ScalableVector, which we currently don't support, at least not currently.

[xen0n]

LGTM apart from the stack ID assertion bit; @SixWeining do you think an assertion should be present nevertheless even if we're only using one type of stack currently?

[SixWeining]

In D128430#3629155, @xen0n wrote:

LGTM apart from the stack ID assertion bit; @SixWeining do you think an assertion should be present nevertheless even if we're only using one type of stack currently?

I think either is fine. I think the reason RISCV does an assertion is that it sets non-default stack ID in its .cpp code.
Considering LoongArch currently only use default stack ID which is not set by target explicitly, maybe there is no strong reason to do an assertion.