https://github.com/llvm/llvm-project/pull/65715

Rebase

Rebase + gentle ping

ping

Update docs

Docs

Fix test
(Note that this depends on some other patches so it can't land yet)

fix small nit

LGTM, don't forget to rebase & rerun tests in case there's more updates needed following MIR patterns landing :)

LGTM but rebase & retest since MIR patterns landed, it should solve a regression we saw during review

In D158598#4611871, @arsenm wrote:

In D158598#4609625, @Pierre-vh wrote:

Additional context: While refactoring this area for some upcoming diffs (removing the deprecated GICombiner backend), I changed some data members from T& to T and got a lot of errors due to this.
While I usually favor const-correctness as much as possible, it's really never respected here. We always use references or pointers, which means we can always call non-const functions on those even in a const function.

Example:

struct Foo {
    int foo() {
        return 0;
    }
};

The foo here would just have to be const? don't see the issue

Additional context: While refactoring this area for some upcoming diffs (removing the deprecated GICombiner backend), I changed some data members from T& to T and got a lot of errors due to this.
While I usually favor const-correctness as much as possible, it's really never respected here. We always use references or pointers, which means we can always call non-const functions on those even in a const function.

ping

LGTM but maybe wait for @arsenm or @b-sumner to comment just in case

Use ArrayRef

Long term I assume we should just have MIR patterns working for ISel, but as long as we rely on DAG patterns I believe this is a good change (we inevitably need to make some tradeoffs in this case anyway), can we land it or are there still some objections?

Actual fix is D157711

Rebase on regenerated checks

Document the new limitation

Rebase

Use implicit op idx

@aemerson are you also fine with landing it?
(Note: I'll be away next week for about 10 days so earliest I can pick up bugs is after the 23rd/24th - I don't mind waiting until I'm back to land this if preferred)

Address comments

Remove useless new testcase

Comments

Thanks for the fix! It's good to commit as-is if you want a quick fix, but the version I propose is the best possible user experience.
Note it's all going away in D156315 anyway.

In D157273#4567283, @arsenm wrote:

In D157273#4565233, @Pierre-vh wrote:

A couple of things to note:

• This is a bit naive, it only works on the Record* and doesn't test for subregs or RC, should it?

Probably can ignore this

• One of the tradeoffs of doing this at the TableGen level is that we don't know about reserved registers. Do they matter in this case? How could we handle them?

It doesn't matter, dead flags shouldn't be set on reserved registers but also shouldn't matter

Comments

Do you think it's worth running a pass like InstSimplify after the tests to check the instruction does get eliminated?
Maybe with a different CHECK line?

In D157273#4565239, @foad wrote:

Are there any cases where this improves codegen?

A couple of things to note:

• This is a bit naive, it only works on the Record* and doesn't test for subregs or RC, should it?
• One of the tradeoffs of doing this at the TableGen level is that we don't know about reserved registers. Do they matter in this case? How could we handle them?

Use a singular, simple OperandTable datastructure for both PatFrags & Rules.
Works a lot better and eliminated about 70 lines of code.

Simplify ParseDirectiveHSAMetadata as it's now HSA only