Rebase

remove erroneous autogen note.

rebase

One more rebase and upgradign tests:

• Several tests where converted to opaque pointer type.
• rG37bc62ed0a24303aa572155009358b8937ab8b4c

This is looking good. I'm going to accept as it, but it'd be better to wait a little to give the opportunity for someone more familiar with this to shime in.

Rebase on top of the fix for v8i1-masks.ll

Rebase on top of @RKSimon 's element shuffle work.

Rebase, bitreverse.ll is definitively fixed.

If you are not sure, I'm not going to land this right away to give anyone who has concern the opportunity to comment.

rebase on top of D141884

One less regression, two more regressions :P

Rebase on top of D140993 and a recent version of main,
and regenerate tests accordingly.

Specialize the transform, which avoids problems with the transform creating an infinite loop.

Yes, it shouldn't in itself, but throw another transform in the loop, and it may well be.

Ok Let's try to make a list of what we want to see fixed before we proceed.

• AMDGPU: widen-smrd-loads.ll ? It doesn't look so bad, but having someone familiar with that target would help. I'm also not sure about the changes related to stores.
• ARM: Seems like wins accross the board by now.
• Hexagon: autohvx/mulh.ll needs someone to look into it.
• PowerPC: is the change in aix32-cc-abi-vaarg.ll a problem? The regression in store-forward-be64.ll is addressed by D140993
• RISCV: Seems like wins accross the board.
• SystemZ: No meaningful changes.
• Thumb2: mve-vst3.ll is worth looking into.
• X86: bitreverse.ll, dagcombine-select.ll, insert-into-constant-vector.ll, insertelement-var-index.ll, pr53419.ll, v8i1-masks.ll, vector-shuffle-combining.ll looks bad.

rebase

Ok, it turns out I'm an idiot and the correct transformation is already done. Sorry everybody.

Lesson of the day, don't submit patches at 3am. It'll get this sorted out.

@RKSimon Yes, I can do that in January. There is one thing I'm not sure of, where do we put the bar for this to be merged? If we want no regression at all, this is going to be very difficult to get it past the line.

The change that are relatively small all come either better or equivalent. While it's hard to evaluate the giant multiplications ones, this definitively looks like a win.

Rebase and cleanup

rebase

Long awaited rebase

Add a test case for this spercifically in select_const and modify cmov-promotion so that this optimization doesn't kick in.

Precommit trivial changes and rebase

Fix the latest broken test

Update the condition in which we transform select to math as to preserve opportunities to optimize select_cc

@RKSimon I'll try what you suggest tomorrow. I've been sick so things are moving slowly on my end. Apologies.

When select_cc takes precedence

What happens if select -> arithmetic takes priority

Dear PowerPC folks, your help would be greatly appreciated here.

I restored the select_cc creation, but now it generates some select_cc that the powerpc backend doesn't quite know what to do with.

In D118037#3731784, @RKSimon wrote:

Please can you rebase to see where we currently are? Also D127115 affects these tests - so maybe see whether this SETULT change is still needed after that patch or whether something else can further improve the codegen?

Fix CodeGen/PowerPC/aix32-cc-abi-vaarg.ll

rebase

rebase

Alive is happy with the correctness of the transform, should we land this or are there extra concerns to be addressed?

• Limit the transform to 32bits and 64bits.
• Add 64 bits test cases.

Update comment and use auto

Rebase on top of D131358

Reabse and variosu tweaks.

Could you add the new tests in main, on their own, then rebase that patch in top of this, so we can see what the effect is?

This didn't look like it was getting much traction. Processing things in topological order should provide a good chunk of the value, so we should revisit later.

Fix CodeGen/PowerPC/nofpexcept.ll

Avoid complex combines when we expect to fold into select_cc

Should I proceed with this? This isn't vital, but it is fairly trivial, and seems better to me.

Rebase