[Triton-MLIR] Fix a few minor issues

[shintaro-iwasaki]

1. In Triton-MLIR's combine pass, I didn't check if getDefiningOp() returns nullptr or not. Unlike dynamic_cast, llvm::dyn_cast fails if nullptr is passed. The first two commits fix this issue and add tests to check the behavior.

2. Pipeline transformation in TritonGPU-MLIR uses a wrong index to access the mask argument of insert_slice_async. The last commit fixes this issue.

(Note that the these issues are found to try to compile the tutorial matmul code.)

[shintaro-iwasaki]

This PR also updates the version of clang-format on ubuntu-latest based on an update in #662.

With this change, it passes tests on ubuntu-latest: shintaro-iwasaki#6

[Jokeren]

Unlike dynamic_cast, llvm::dyn_cast fails if nullptr is passed

I don't quite get this one. Doesn't the following code also fail when nullptr is passed?

    auto *constantMaskCandidate = mask.getDefiningOp();
    if (!constantMaskCandidate)
      return mlir::failure();

[shintaro-iwasaki]

Thanks for taking a look at this PR @Jokeren! Sorry for confusion. The first commit handles the last "NG" case: llvm::dyn_cast does not seem to accept nullptr.

// Assume constantMaskCandidate is nullptr
auto *constantMaskCandidate = mask.getDefiningOp();

// OK
if (!constantMaskCandidate)
  return mlir::failure();

// OK: dynamic_cast<> can take nullptr 
#if 0
// The following code does not work because LLVM does not enable RTTI.
if (!dynamic_cast<arith::ConstantOp *>(constantMaskCandidate))
  return mlir::failure();
#endif

// NG: llvm::dyn_cast<> cannot take nullptr (it causes assert() or SEGV) for some reasons
if (!llvm::dyn_cast<arith::ConstantOp>(constantMaskCandidate))
  return mlir::failure();    

[ptillet]

What about llvm::dyn_cast_or_null ?

[shintaro-iwasaki]

Thanks, @ptillet. You are completely right. I updated the first commit to use dyn_cast_or_null.