Rebase again.

Rebase and merge precommit.

Add test with fneg as src modifier.
Something weird happening with patch application, maybe the parent review https://reviews.llvm.org/D155171 should be merged first?

Rebase.

Removed unnecessary flags from fmul instruction in tests.
Add test with abs as a src0 modifier.

Rebase.

Use selection patterns for selecting V_FMA/MAD_MIX* instead of combiners.
With combiners there is a possibility that fptrunc (fmul a, b) would be selected into fptrunc (fma a, b, 0), and some other combiner could transform it back to its original state.

Added tests without denormal flushing as a function attribute.
In all cases except GFX900 it should have the same result as with denormal flushing.

Since it doesn't affect any tests, should I add this line in D153544?

I could add a function attribute containing denormal flushing or add more run lines to these tests.

@foad do you think I should do something similar to Diff 2?
Or should I write AMDGPU-only combiners?
Or something else?

Remove add instruction from fptrunc + fmul -> v_fma/mad_mix*.

Instead of writing patterns that select fma/mad_mix*, write combiners for sdag and global isel that will transform
fptrunc (fmul a, b) into fptrunc (fma a, b, 0), which will later be selected into v_fma/mad_mix*.

Transform fptrunc (mul a, b) -> fma_mix/mad_mix a, b, 0, implicit_def
or
build_vector el0, (fptrunc (mul a, b)) -> fma_mix/mad_mix a, b, 0, el0.

@arsenm Would it be correct if I wrote a pattern in MadFmaMixPats multiclass, that finds (f16 (fptrunc (f32 (fmul %src1, %src2)))) and turns it into v_fma_mixlo_f16 %src1, %src2, 0.
The add would stay the same.

Should it also be done for GlobalISel in the same patch?

What other cases should I cover?
Should I cover the case when we have to pick the higher 16 bits of mul instruction and select v_fma_mixhi_f16?

Remove unnecessary if, use isNullValue instead of isZeroValue.
Thanks @foad.

Rebase.

Instead of call to computeKnownFPClass, use findScalarElement.

Remove *_buffer_store instructions from being optimized.

Add test case with all zero components.

This patch was reverted on upstream, because of failed cts tests.

Change condition in for loop, instead of i >= 0, put i > 0. We don't want to optimize out the 0th element.

In D146737#4353046, @foad wrote:

matejam added a reverting change: rG9c8c31eea439: Revert "[AMDGPU] Trim zero components from buffer and image stores".

Have you got any more details about what was wrong with it?

Move the default case out of the switch.

Refactor.

Do the optimizations for image instructions that were done prior to this patch.

Use Intrinsic opcode to know if the instructions has DMask instead of testing if the instruction has a ConstantInt as the second operand.
Add more run-lines to the test.

Change the name from findDemandedElts to trimTrailingZerosInVector.
Remove some unnecessary dyn_casts.
Refactor and rebase.

Thanks for the review.

Please review this. @foad @arsenm

Rebase and change in comments.

In D148096#4299323, @arsenm wrote:

Needs new tests that now legalize? The test updates only show regressions?

Thank you @foad.
findDemandedElts with correct usage of computeKnownFPClass.

Changes in findDemandedElts, use computeKnownFPClass.

@foad @arsenm would you please review this?
Thank you.

Rebased.

Moved call of moreElementsIf to the bottom. Made changes to customIf call for G_EXTRACT_VECTOR_ELT and G_INSERT_VECTOR_ELT.

Added named LegalityPredicate and LegalizeMutation for checking if the type doesn't have a corresponding AMDGPU RegClass and
if not, widenening the vector to the first next legal RegClass size.

This is an extension of https://reviews.llvm.org/D144198.
It includes checking the exact width in get*ClassForBitWidth and also widening of the vector operand of
G_BUILD_VECTOR, G_INSERT_VECTOR_ELT and G_EXTRACT_VECTOR_ELT instructions.

Rebase and minor changes.

Closed, commit hash: f6e70ed1c73a2f3ac15eb6650423c1c10d278f50

Use of SmallVector instead of VectorMap for tracking which components were already added.
Remove some unnecessary dyn_casts.