[TEST] Enable test-alignment.mlir checks & remove failing checks

[davidberard98]

Unless I'm doing something wrong when running lit test, I don't think test-alignment.mlir was actually checking any of the remarks.

There were two issues:

1. -verify-diagnostics is not added to the triton-opt run command, so the expected-remark comments were not executing.
2. expected-remarks was misspelled everywhere in this file, instead written as expeted-remarks.

This PR fixes the typo and adds -verify-diagnostics=only-expected (so that it checks for expected remarks, doesn't require an expected remark for all the remarks that are actually emitted), and it also removes all the failing expected-remarks (replacing them with TODOs). For ease of review, the PR is split into two commits - one which renames expeted-remark -> expected-remark, and one which replaces expected-remark with a TODO for the remarks that were previously failing.

[davidberard98]

cc @makslevental who added -verify-diagnostics=only-expected in MLIR

[makslevental]

Unless I'm doing something wrong when running lit test, I don't think test-alignment.mlir was actually checking any of the remarks.

Definitely wasn't checking before... Currently you have -o /dev/null which I suggest you remove because (testing locally) I see that there are "expected remarks" that are not in fact emitted:

within split at /home/mlevental/dev_projects/triton/test/Analysis/test-alignment.mlir:810 offset :20:6: error: expected remark "contiguity = [1, 1], divisibility = [16, 16], constancy = [1, 1], constant_value = <none>" was not produced
  // expected-remark @below {{contiguity = [1, 1], divisibility = [16, 16], constancy = [1, 1], constant_value = <none>}}
     ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// -----
module {
  tt.func @int_min_does_not_underflow_in_analysis() -> i64 {
    %c-9223372036854775808_i64 = arith.constant -9223372036854775808 : i64
    tt.return %c-9223372036854775808_i64 : i64
  }
}

[davidberard98]

@makslevental These failures don't appear for me when I run locally - e.g. the one at test-alignment.mlir:810 offset :20:6 should be commented out / replaced with the TODO messages by this PR.

regarding -o /dev/null: I'm guessing this was initially added to hide the output, which is the same as the input IR, aside from renamed variables. In my local testing, when things fail, they still fail lit test, and return errors to stderr, even if we have -o /dev/null.

[davidberard98]

also cc @Mogball who introduced the verify-diagnostics testing (since I don't have permissions to request reviewers)

[ThomasRaoux]

so what's going on with the failing cases? Are those cases we regressed? Those use to all pass before moving to remarks?

[davidberard98]

@ThomasRaoux I believe that:

• Some failing cases are regressions
• Some failing cases are formatted in ways that don't work with verify-diagnostics, e.g. comments that previously used CHECK-NEXT were directly replaced with expected-remark @below, when they don't actually apply to the next following op

[ThomasRaoux]

@ThomasRaoux I believe that:

• Some failing cases are regressions
• Some failing cases are formatted in ways that don't work with verify-diagnostics, e.g. comments that previously used CHECK-NEXT were directly replaced with expected-remark @below, when they don't actually apply to the next following op

interesting, thanks for the update