About lowering of tptr dialect

[Realtyxxx]

I have found that in some cases, unprocessed tptr dialects will be generated, such as some cases in @nhat-nguyen ‘s repo triton-scratchpad.
Also I found the discussion you talk in llvm discussions, but the mlir ptr dialect development seems to take some time.
So I did something with reference to other libraries to eliminate tptr dialect:

1. Modify the compile stage to go through the pass manager
2. Modify the tptr dialect, remove the tptr::D efaultMemorySpaceAttr::get(rewriter.getContext()) attribute and related code, remove typeoffset op, and remove the relevant code in the code
3. Implement tptr->llvm pass to convert ptradd，from_memref, to_memref ops;
   I would like to know if this route is in line with the current direction, are you interested in merging the code?

[Realtyxxx]

And the _ttsharedir_to_llir would be like :

pm = ir.pass_manager(mod.context)
            pm.enable_debug()
            triton_shared.to_llir.add_convert_linalg_to_affine_loops(pm)
            triton_shared.to_llir.add_empty_tensor_to_alloc_tensor(pm)
            triton_shared.to_llir.add_one_shot_bufferize_with_options(
                pm, allow_return_allocs_from_loops=True)
            triton_shared.to_llir.add_lower_affine(pm)
            triton_shared.to_llir.add_convert_linalg_to_loops(pm)
            triton_shared.to_llir.add_expand_strided_metadata(pm)
            triton_shared.to_llir.add_convert_scf_to_cf(pm)
            triton_shared.to_llir.add_convert_arith_to_llvm(pm)
            triton_shared.to_llir.add_convert_math_to_llvm(pm)
            triton_shared.to_llir.add_convert_complex_to_llvm(pm)
            triton_shared.to_llir.add_convert_vector_to_llvm(pm)
            triton_shared.to_llir.add_convert_index_to_llvm(pm)
            triton_shared.to_llir.add_memref_expand(pm)
            triton_shared.to_llir.add_finalize_memref_to_llvm(pm)
            triton_shared.to_llir.add_convert_func_to_llvm(pm)
            # ! convert here
            # triton_shared.debug.enable_mlir_debug("tptr-to-llvm")
            triton_shared.to_llir.add_convert_tptr_to_llvm(pm)

            triton_shared.to_llir.add_convert_cf_to_llvm(pm)
            triton_shared.to_llir.add_lower_affine(pm)
            triton_shared.to_llir.add_convert_arith_to_llvm(pm)
            triton_shared.to_llir.add_reconcile_unrealized_casts(pm)
            pm.run(mod)

            Path(llmlir_path).write_text(str(mod))

[Realtyxxx]

Sincerely hope to have suggestions from you @nhat-nguyen

[red1bluelost]

Completely getting rid of tptr will not be possible until ptr has all the necessary features. Some of the ptr features are now available so we could start trimming some of tptr but not all. The PR you have to add tptr-to-llvm is a good stop gap for the CPU backend until ptr dialect has everything.