[DAG] Constant Folding for U/SMUL_LOHI

[Pierre-vh]

No description provided.

[Pierre-vh]

I did this after looking at #69324, where we emitted a couple of muls that were obviously foldable.
Now we also have a s_add that could be folded but it's emitted from a (bitcast (build_vector lo, hi)) and we can't look through bitcast/splat in the DAG apparently.

Again no sure how effective this is, i just thought it's a nice to have. We have one more instruction in the testcase but it's because we scalarized a bit more and needed to copy to a VGPR, so overall I think it improved it.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[jayfoad]

Do you have a plan to fold this add?

[Pierre-vh]

Not yet, it looks like this:

      t420: v2i32 = BUILD_VECTOR Constant:i32<16951>, Constant:i32<0>
    t421: i64 = bitcast t420
      t414: v2i32 = BUILD_VECTOR Constant:i32<54974720>, Constant:i32<0>
    t415: i64 = bitcast t414
  t190: i64 = add t421, t415

We can't do a generic combine for the bitcast I think. I tried it and it caused loops. I think the easiest would be to make a DAG equivalent of the GISel helper that looks through splats and just returns the value directly.

[RKSimon]

We might be able to replace dyn_cast<ConstantSDNode> checks in SelectionDAG::FoldConstantArithmetic with ComputeKnownBits calls that looks for isConstant scalar values?

[jayfoad]

I don't understand what the issue is here. Why can't we fold (bitcast (build_vector constant:i32, constant:i32)) -> constant:i64 ?

[Pierre-vh]

We already have a ConstantFoldBITCASTofBUILD_VECTOR fold but it only runs before LegalizeTypes/Ops. Some targets need the bitcasts apparently. I tried doing it more often but only in specific cases (<=64 bit ints) but I still get infinite loops everywhere.

I think we need to add some helper to get splat values instead of checking ConstantSDNode when constant folding add. What do you think?

[arsenm]

Shouldn't it just be a matter of legality checks to avoid the loops?