iree: Transform preprocessing spec is unusably slow on real-world inputs

Simplest transform spec to illustrate the issue:

// spec.mlir
module attributes { transform.with_named_sequence } {
  transform.named_sequence @match_generic(%entry: !transform.any_op {transform.readonly}) -> !transform.any_op {
    transform.match.operation_name %entry ["linalg.generic"] : !transform.any_op
    transform.yield %entry : !transform.any_op
  }

  transform.named_sequence @__transform_main(%variant_op: !transform.any_op {transform.consumed}) {
    %generic = transform.collect_matching @match_generic in %variant_op
      : (!transform.any_op) -> !transform.any_op

    transform.print %generic {name = "Generic"} : !transform.any_op
    // Transform the matched generic.
    transform.yield
  }

  transform.named_sequence @__preprocessing_main(%variant_op: !transform.any_op {transform.readonly}) {
    transform.yield
  }
} //  module

Repro:

wget https://sharkpublic.blob.core.windows.net/sharkpublic/ian/stable_diffusion_xl_base_1_0_64_1024x1024_fp16_unet_linalg_nithin.mlir

tools/iree-compile --iree-hal-target-backends=rocm --iree-rocm-target-chip=gfx942 stable_diffusion_xl_base_1_0_64_1024x1024_fp16_unet_linalg_nithin.mlir --iree-preprocessing-transform-spec-filename=spec.mlir

This input mlir has <50 kLOC and contains 8583 linalg.generic ops. With the transform script above, I can only match ~560 linag.generics in 30 seconds. This doesn’t make it a viable solution for real-world inputs.

Are there some knobs I could tweak to make it go faster? Alternatively, maybe we could do some pre-filtering at the level of C++, e.g., wrap the transform dialect interpreter in a pass pipeline (that could be specified as a CLI pipeline string).

cc: @Groverkss @harsh-nod @ftynse @matthias-springer @maerhart