[GlobalISel] [AArch64] Fold G_PTRTOINT(G_CONSTANT)

[andcarminati]

Summary

Fold the following pattern:

    %0:_(p0) = G_CONSTANT i64 0
    %1:_(s64) = G_PTRTOINT %0:_(p0)
    %2 = COPY %1(s64)

Into:
    %0:_(s64) = G_CONSTANT i64 0
    %1 = COPY %0(s64)

[arsenm]

Where are these coming from? I thought we directly emitted pointer constants now?

[andcarminati]

In D140302#4004906, @arsenm wrote:

Where are these coming from? I thought we directly emitted pointer constants now?

Well, we have this folding implemented in our development branch (custom target) and we think it is interesting for AArch64 too.

[arsenm]

In D140302#4004906, @arsenm wrote:

Where are these coming from? I thought we directly emitted pointer constants now?

I think this is special cased for ConstantPointerNull

llvm/include/llvm/Target/GlobalISel/Combine.td

514

Should be able to just use replaceInstWithConstant

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

2095–2098

We already have a matchConstantOp, but that checks for a specific constant. Should have another generic one that checks for any constant

[arsenm]

Maybe should have the constant folding MIRBuilder do this too

[paquette]

In D140302#4004963, @andcarminati wrote:

In D140302#4004906, @arsenm wrote:

Where are these coming from? I thought we directly emitted pointer constants now?

Well, we have this folding implemented in our development branch (custom target) and we think it is interesting for AArch64 too.

How much of a code size improvement did you see with this change?

[andcarminati]

The patch was updated to the main branch.

Some background: we are upstreaming this pattern through the AArch64 because the target backend that we are developing in our company was not upstreamed yet. About usability, we observed that this pattern is being matched and can impact the final code.