[DAGCombiner] Improve FMA support for interpolation patterns

[RKSimon]

Summary

This patch adds support for combining patterns such as (FMUL(FADD(1.0, x), y)) and (FMUL(FSUB(x, 1.0), y)) to their FMA equivalents.

This is useful in particular for linear interpolation cases such as (FADD(FMUL(x, t), FMUL(y, FSUB(1.0, t))))

[arsenm]

This mostly LGTM.

There aren't any tests stressing the FMAD path. AMDGPU seems to be only target using it still, and the one test change is in the expansion of an intrinsic which should be removed. If you can add some of those that would be good, otherwise I can try to do it after you commit

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

96

A better name would be AllowFusion or something like that

103–105

I think the AllowFusion/UnsafeFPMath check should be first

112

Usually the int is omitted

113

It seems wrong to use this in the FMAD case, although AMDGPU happens to not care because enableAggressiveFMAFusion always reports true and it seems to be what is used already.

[RKSimon]

There aren't any tests stressing the FMAD path. AMDGPU seems to be only target using it still, and the one test change is in the expansion of an intrinsic which should be removed. If you can add some of those that would be good, otherwise I can try to do it after you commit

Thanks Matt, I can add some FMAD tests for v_mad_f32 - is that the only instruction I should be testing for?

Most of your comments about the preamble are just as relevant for the other FMA pattern combines (visitFADDForFMACombine, visitFSUBForFMACombine); given that I copied+pasted most of it from them should they be updated as well?

[arsenm]

In D13003#249540, @RKSimon wrote:

There aren't any tests stressing the FMAD path. AMDGPU seems to be only target using it still, and the one test change is in the expansion of an intrinsic which should be removed. If you can add some of those that would be good, otherwise I can try to do it after you commit

Thanks Matt, I can add some FMAD tests for v_mad_f32 - is that the only instruction I should be testing for?

Yes. The fneg should be folded in as a source modifier that looks something like v_mad_f32 v0, v1, v2, -v3. Sometimes v_mac_f32 is used, although in these cases that shouldn't happen

Most of your comments about the preamble are just as relevant for the other FMA pattern combines (visitFADDForFMACombine, visitFSUBForFMACombine); given that I copied+pasted most of it from them should they be updated as well?

Yes, probably

[RKSimon]

Updated all FMA combine helpers based on Matt's feedback.

Added AMDGPU FMA/FMAD tests

[arsenm]

LGTM, although I think you should split the renames in the other parts into a separate patch