LGTM

Pushed an earlier commit by mistake

Picked this issue up in some testing - seems a logical fix.
I have a reproducer, but even the cut-down isn't that clean - do you think this requires a test?

Spelling mistake (and rebase for test changes)

Updating test in light of review suggestions

Updating with link to change for which this is a pre-commit

Pre-commit test

Commited as cf65271e462f0626157a06f2f3a9615b4987c24e

In D111646#3058586, @dstuttard wrote:

Interesting to see that quite a few of the test changes are clustering the loads, but not ending up with a clause (correctly).
Nice result otherwise though.

Interesting to see that quite a few of the test changes are clustering the loads, but not ending up with a clause (correctly).
Nice result otherwise though.

Address review comments

@critson @foad @Flakebi - are you happy with the changes? Can someone approve?

In D103014#3018706, @krisb wrote:

@dstuttard, could you please give an example that reproduces the issue?

ping?

This change has caused some problems for us - resolve doesn't respect the use of subst on windows which we've been using to avoid long file path problems. This is a particular issue for the llvm-lit tests.

Rebase and address some of the review comments

We ran across this issue a while back for graphics, but I'd forgotten to upstream it.

Thanks Stas - LGTM

I'm revisiting this - (sorry for the lengthy hiatus)

Do you think that I should add any other reviewers?

Pre-committed test and rebased on top

In D103323#2856828, @RKSimon wrote:

@dstuttard Please can you rebase? rGcc38f8939da4aec85e7d0ef4de412e30d4de5a14 should give you vector coverage

Updated existing test based on feedback

Adding mul_uint24-amdgcn.ll gfx9 variant

Add test

Updating for review comments.

Some modifications based on review feedback

Thanks for reviews.
Made suggested changes.

Wasn't sure about the new HasSMulHi - any thoughts?

Updates based on review comments

Moved part of this change into D102231 (required for that change).

Moved change from D102066 required for this change (which will be committed first)

I've got a case where we see this issue as well.

Made suggested changes

Fixed sense of NoG16 argument passed in from asmparser and disassembler

See new patch with globalisel implementation

Addressed review comments
Added support for global-isel. Note: global-isel produces less optimal code
presently. Will require some additional work to add extra combines

This is a combination of the 2 previous patches - both asmparser and disassembler. I've also unified the code from SIInstrInfo.cpp into a single function since the code was duplicated 3 times before.

I've combined this with other changes into D102231 which was easier than trying to update this one

I've combined this with some other changes into D102231 - easier than trying to separate all the various bits

a16 and g16 no longer linked

Yes, thanks. Already reverted.

D101619, D101620 are related.