Issues under raising opaque SSA-defined bounds to ISL

[xys-syx]

I have met the issues the raising pass produces MLIR loop bounds from opaque LLVM ssa values that isl cannot map to named symbolic parameters, leading to unexpected unnamed parametersand failure to infer kernel dimensions. And the raising pass segfaults on translation units containing loops whose bounds come from indirect memory accesses.

The output messages are:
Hundreds of src/external/isl/isl_space.c:3351: unexpected unnamed parameters or found loop with positive increment and non-constant bounds.

[Pangoraw]

do you have a reproducer / input ir ?

[xys-syx]

Sure, I have a minimal reproducer at here wsmoses/Enzyme-GPU-Tests#5

[wsmoses]

@xys-syx can you paste here an mlir file and the corresponding pass which when it runs on the IR, breaks?

[xys-syx]

@wsmoses Hello, I have created the milr files, which are for two ISL errors:

• hang on enzymexlamlir-opt isl.mlir --affine-cfg, reproduces the affine-cfg hang
• enzymexlamlir-opt isl_error.mlir --affine-cfg, error messages are Assertion: "flattening unimplemented for semi-affine integer sets"
  But I cannot paste the mlir file here, as github tell me "There was a problem saving your comment. Please try again." So I have commit to this PR: Align Makefile with LBM and add reproducer.cu wsmoses/Enzyme-GPU-Tests#5