[SLP] Add cost model for `llvm.powi.*` intrinsics

[n-omer]

Summary

This patch adds handling for the llvm.powi.* intrinsics in BasicTTIImplBase::getIntrinsicInstrCost() and improves vectorization.

Closes #53887.

[RKSimon]

diff with context?

llvm/include/llvm/CodeGen/TargetLowering.h

2204

Maybe drop the virtual - its easy enough to readd if a target wants to override this.

Would this be better as int Exponent?

[n-omer]

Add context to diff.

[RKSimon]

llvm/test/Transforms/SLPVectorizer/X86/powi-regression.ll

4

This file should probably be pre-committed with current codegen and the patch rebased to show the codegen change.

20

Drop the "local_unnamed_addr #0"

[n-omer]

@RKSimon I've created https://reviews.llvm.org/D128178 for the test.

[n-omer]

Rebase patch and address review comments.

[n-omer]

Add context to diff.

[n-omer]

Remove unused included file.

[RKSimon]

LGTM - cheers

[n-omer]

Reverted due to test failures, investigating further.

[n-omer]

Update failing cost model tests.

[n-omer]

Add context to diff.

[RKSimon]

@craig.topper @dmgreen Are the aarch64/riscv cost changes OK? They are in 'unsupported' sections - so not sure if they have extra meaning?

[dmgreen]

I think it is fine so long as the backend can handle it:
https://godbolt.org/z/5bvEn5563
But needs to be an invalid cost if it cannot:
https://godbolt.org/z/6jK7Kvc4s

Can you change the test lines to this, which should still return an invalid cost:

%powi = call <vscale x 4 x float> @llvm.powi.nxv4f32.i32(<vscale x 4 x float> %vec, i32 %extraarg)

[RKSimon]

Maybe copy the immediate tests below so we have coverage for both constant / variable exponent values?

[n-omer]

Modify AArch64 and RISCV tests.

[dmgreen]

Thanks, LGTM. Although I was wondering why the cost came out as 12, not 14 from the constant 42. It may be possible to get a slightly better costs that more accurately match the code from ExpandPowI.

llvm/include/llvm/CodeGen/BasicTTIImpl.h

1424

I think a more accurate count might be ActiveBits + PopCount - 2, at least from playing around with a few examples.

[n-omer]

Improve the cost computation and update tests.

Thanks @dmgreen.

[dmgreen]

Thanks. I guess that changes all the negative constant values too..

If you want a cost for negative numbers I think it would use the abs of the value in the ActiveBits+PopCount-2 computation, plus the div like you already have it.

llvm/include/llvm/CodeGen/BasicTTIImpl.h

1426

I think it can go via an APInt: RHSC->getValue().countPopulation()

1430

For negative numbers if looks like it should use ((-RHSC).ActiveBits + (-RHSC).PopCount - 2)) * Mul + Div, using the negative -RHSC in the calculations above.

[n-omer]

Fix cost model.

My bad, thanks.

[dmgreen]

Thanks. The ticket is still "closed", so I cant properly accept it again. But it LGTM.