[mlir][arith] Canonicalization patterns for arith.select

[peterbell10]

This adds the following canonicalization patterns:

• Inverting condition:
  • select(not(pred), a, b) => select(pred, b, a)
• Merging consecutive selects with the same predicate
  • select(pred, select(pred, a, b), c) => select(pred, a, c)
  • select(pred, a, select(pred, b, c)) => select(pred, a, c)
• Merging consecutive selects with a common value:
  • select(predA, select(predB, a, b), b) => select(and(predA, predB), a, b)
  • select(predA, select(predB, b, a), b) => select(and(predA, not(predB)), a, b)
  • select(predA, a, select(predB, a, b)) => select(or(predA, predB), a, b)
  • select(predA, a, select(predB, b, a)) => select(or(predA, not(predB)), a, b)

[peterbell10]

@matthias-springer I see you've reviewed a few mlir PRs recently. Would you be able to help find a reviewer for this? I can see that @llvmbot has pinged teams on other PRs but not here and am concerned it might be missed.

[peterbell10]

@matthias-springer PTAL

[matthias-springer]

Thanks! Looks good to me! Do you have commit access or do you need help landing this?

[peterbell10]

No I don't have commit rights, this is actually my first LLVM contribution. Any help would be appreciated, thanks!

[hanhanW]

FYI, this is wrong because it could be vector of i1 types, e.g., vector<4xi1>.

This patch introduces failure in downstream project (e.g., IREE). I'm working on a fix now.

[hanhanW]

Here is a failure example:

  func.func @foo(%arg0: vector<4xf32>, %arg1: vector<4xf32>) -> vector<4xf32> {
    %0 = arith.cmpf ult, %arg0, %arg1 : vector<4xf32>
    %1 = arith.select %0, %arg0, %arg1 : vector<4xi1>, vector<4xf32>
    %2 = arith.cmpf uno, %arg1, %arg1 : vector<4xf32>
    %3 = arith.select %2, %arg1, %1 : vector<4xi1>, vector<4xf32>
    return %3 : vector<4xf32>
  }

[hanhanW]

As mentioned above, it broke the downstream project. So I reverted the commit. I have a WIP fix and will send it out for review.