Disable region simplification in --aie-substitute-shim-dma-allocations pass

[andrej]

The pattern rewriter in this pass previously also performed region simplification, which caused an assertion error if the MLIR contained blocks with input arguments, such as generated by:

          %7 = scf.if %6 -> (memref<64x32xf32>) {
            scf.yield %C_L1L2_0_0_buff_0 : memref<64x32xf32>
          } else {
            scf.yield %C_L1L2_0_0_buff_1 : memref<64x32xf32>
          }

This fixes that issue. I don't fully understand why the issue exists with region simplification, but the point of this pass is not to simplify regions anyways, so it is fine to disable.

[fifield]

Seems like red-flag that the IR is left in a bad state somewhere. Do you have a failing (fixed) test you can add to the PR?

[jgmelber]

@andrej is this still needed after the dynamic object FIFO implementation landed? @AndraBisca

[andrej]

Oops, forgot about this one. Yes I believe this is still needed. I can work on a test.

[andrej]

@fifield I finally got around to creating a fairly minimal test. I can confirm without this PR, the added test still fails on current MLIR-AIE. Sorry it took so long.

Merging this should allow #1772 to pass.

[fifield]

@fifield I finally got around to creating a fairly minimal test. I can confirm without this PR, the added test still fails on current MLIR-AIE. Sorry it took so long.

Merging this should allow #1772 to pass.

Thanks, out of curiosity I'd still like to take a look at why it was crashing without the modification.

[fifield]

@fifield I finally got around to creating a fairly minimal test. I can confirm without this PR, the added test still fails on current MLIR-AIE. Sorry it took so long.
Merging this should allow #1772 to pass.

Thanks, out of curiosity I'd still like to take a look at why it was crashing without the modification.

The error has little to do with the pass changed in this CR. It has to do with a missing terminator at the end of the aie.device. Still investigating, but I'll note that the aie.device verifier implementation is currently return success();

[andrej]

Please let me know if/when you find the root cause and if you'd still be OK with merging this PR even if it doesn't directly address the root cause.

[andrej]

Should we still merge this after #1936? While the error doesn't exist any more after #1936, I think it's kind of weird that this pass by default performs region simplification when what it purports to do is something else. Or is that normal in MLIR/LLVM world?