Bump tested IREE version to most recent nightly.

[ScottTodd]

We're approaching our 3.1.0 release across all IREE packages (iree-org/iree#19192), so testing closer to HEAD.

Commit range: iree-org/iree@iree-3.1.0rc20241212...iree-3.1.0rc20241220

[ScottTodd]

Some TKW tests seem to be stalling on mi250 and mi300 runners: https://github.com/iree-org/iree-turbine/actions/runs/12438563153?pr=354. Might have to debug locally.

cc @raikonenfnu @harsh-nod @Hardcode84

[ScottTodd]

Yeah jobs are getting stuck consistently. First run wasn't making progress after nearly 2 hours: https://github.com/iree-org/iree-turbine/actions/runs/12438563153/job/34730710267?pr=354#step:7:39.

[Hardcode84]

I will look into it.

[Hardcode84]

https://gist.githubusercontent.com/Hardcode84/dd213d889a1c7a96618bd2bab5c451ca/raw/89681b9add6ab5f0d468aa115c96b83b004142f0/gistfile1.txt

After OptimizeIntArithmeticPass attention loop iteration count is replaced with huge value, causing kernel to hang.

There may be an issue with OptimizeIntArithmeticPass or upstream integer range analysis, I'm looking into it.

CC @krzysz00

[Hardcode84]

Here is the reduced example:

func.func @test3(%6: index) -> index {
  %c64 = arith.constant 64 : index
  %55 = arith.ceildivsi %6, %c64 : index
  return %55 : index
}

mlir-opt --debug-only=int-range-analysis --int-range-optimizations repro.mlir

func.func @test3(%arg0: index) -> index {
  %c144115188075855872 = arith.constant 144115188075855872 : index
  return %c144115188075855872 : index
}

ceildivsi range analysis is very broken

[Hardcode84]

Workaround for the ceildiv bug: #355
Upstream fix: llvm/llvm-project#121062

[ScottTodd]

Workaround for the ceildiv bug: #355 Upstream fix: llvm/llvm-project#121062

Thanks! Looks like the upstream LLVM change hasn't landed in IREE quite yet, so proceeding with the workaround for now SGTM.

[Hardcode84]

Merged the workaround

[ScottTodd]

Got a clean CI run after retrying a few flaky tests/workflows. I'd like to land this then try incrementally updating again.