[AMD] DCE/canonicalize true epilogue conditionals

[makslevental]

Waiting on llvm/llvm-project#133151 upstream.

This PR adds a pattern that folds "true" arith.cmpi operations to arith.constant true; e.g.

%c0 = arith.constant 0 : i32
%c1024_i32 = arith.constant 1024 : i32
%cmpsge = arith.cmpi sge, %c1024_i32, %c0 : i32

->

%cmpsge = arith.constant true

(after DCE).

The specific use case is "unguarding" the epilogue in pipelined loops (e.g., as produced by tritonamdgpu-stream-pipeline). So e.g.,

tt.func @assume_matmul(%arg0: index, %arg1: index, %arg2: index, %arg3: !tt.ptr<f16>, %arg4: !tt.ptr<f16>) -> tensor<128x128xf32, #mma> {
  ...
  %20:6 = scf.for ... {
    scf.yield ...
  }
  ...
  %27 = arith.cmpi sge, %26, %c1 : index
  %31 = scf.if %27 -> (tensor<128x128xf32, #mma>) {
    %33 = tt.dot %28, %30, %20#2
    scf.yield %33 : tensor<128x128xf32, #mma>
  } else {
    scf.yield %20#2 : tensor<128x128xf32, #mma>
  }
  %32 = arith.select %27, %31, %20#2 : tensor<128x128xf32, #mma>
  ttg.local_dealloc %10 : !ttg.memdesc<1x128x32xf16, #shared, #smem, mutable>
  ttg.local_dealloc %11 : !ttg.memdesc<1x32x128xf16, #shared1, #smem, mutable>
  tt.return %32 : tensor<128x128xf32, #mma>
}

becomes

tt.func @assume_matmul(%arg0: index, %arg1: index, %arg2: index, %arg3: !tt.ptr<f16>, %arg4: !tt.ptr<f16>) -> tensor<128x128xf32, #mma> {
  ...
  %20:6 = scf.for ... {
    scf.yield ... 
  }
  %21 = ttg.local_load %20#4
  %22 = ttg.local_load %20#5
  %23 = arith.mulf %22, %cst
  %24 = tt.dot %21, %23, %20#2
  ttg.local_dealloc %10 : !ttg.memdesc<1x128x32xf16, #shared, #smem, mutable>
  ttg.local_dealloc %11 : !ttg.memdesc<1x32x128xf16, #shared1, #smem, mutable>
  tt.return %24 : tensor<128x128xf32, #mma>
}

Notice both the scf.if and arith.select are canonicalized away.

Note, this usually requires the use of tl.assume to hint/constrain the operands of the arith.cmpi; specifically wrt the original loop bounds something like %stop // %step >= 1 (or whatever the arithmetic on the loop bounds needs to be...).

Currently this is failing because I need to cherry-pick/PR an LLVM bump.

Waiting on #6334.

[makslevental]

Some kind of bug around here https://github.com/llvm/llvm-project/blob/8726e973459d93d34653946ba1e01ad198cdf11f/mlir/lib/Dialect/Arith/Transforms/IntRangeOptimizations.cpp#L56-L81 related to how the constant is materialized. Will figure it out next week.

[makslevental]

Upstream bug fix: llvm/llvm-project#133556

[makslevental]

Same failure as here #6343 - related to a recent change @Mogball made upstream also to range analysis.

[Mogball]

I put a fix in the branch. IntRangeAnalysis will now return a dummy return for noninteger values, because it has to return something.