[FEAT-DSL] Update FlyToROCDL Conversion to support full dsl types

[sjfeng1999]

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[Copilot]

Pull request overview

This PR updates the Fly→ROCDL lowering pipeline to better support the full Fly DSL type system (notably swizzles, composed layouts, and BufferDesc “fat pointer” handling), and aligns MLIR tests + Python DSL wrappers with the new printing/lowering behavior.

Changes:

• Add BufferFatPtr helper and refactor BufferDesc pointer lowering/offsetting/copy to use it.
• Extend layout lowering/type inference to handle SwizzleType in crd2idx and introduce fly.decomposition plumbing for copy expansion.
• Update MLIR conversion/transform tests for formatting and i32-based GEP indices; reorganize/extend Python primitive DSL helpers.

Reviewed changes

Copilot reviewed 12 out of 15 changed files in this pull request and generated 9 comments.

Show a summary per file

File	Description
tests/mlir/Transforms/rewrite_func_signature.mlir	Adjusts expected type printing spacing in signatures.
tests/mlir/Transforms/layout_lowering.mlir	Adjusts expected ptr type printing spacing.
tests/mlir/Conversion/pointer_ops.mlir	Updates expected GEP index types/casts (i64 → i32) to match new lowering.
tests/mlir/Conversion/memref_ops.mlir	Updates expected GEP index types/casts (i64 → i32) to match new lowering.
tests/mlir/Conversion/dyn_shared.mlir	Updates expected dynamic shared GEP index type (i64 → i32).
python/flydsl/expr/primitive.py	Reorganizes and expands DSL op wrappers; adds/relocates ptr/memref helpers and “Deprecated” wrappers.
lib/Dialect/FlyROCDL/CDNA3/MmaAtom.cpp	Adjusts CDNA3 MFMA C thread/value layout construction.
lib/Dialect/Fly/Transforms/LayoutLowering.cpp	Adds SwizzleType handling in crd2idx lowering; inserts fly.decomposition before certain copy atom calls; improves GEMM shape handling.
lib/Dialect/Fly/IR/FlyTypeDefs.cpp	Improves printing spacing for types/attrs (commas).
lib/Dialect/Fly/IR/FlyOps.cpp	Extends crd2idx type inference for swizzles; adds DecompositionOp type inference; refines MemRefLoadOp inference for CoordTensorType.
lib/Dialect/Fly/IR/FlyAttrDefs.cpp	Adjusts parsing of leaf/basis attrs to more selectively parse E<mode> suffixes.
lib/Conversion/FlyToROCDL/FlyToROCDL.cpp	Refactors BufferDesc lowering through BufferFatPtr; adds ptr swizzle application at load/store/memcpy sites; simplifies iterator ops lowerings.
lib/Conversion/FlyToROCDL/BufferFatPtr.h	New helper implementing BufferDesc fat-pointer packing/unpacking and swizzled byte-offset computation.
include/flydsl/Dialect/Fly/Transforms/MemrefLowering.td	Adds pattern rewrites for fly.decomposition on memref/coord_tensor.
include/flydsl/Dialect/Fly/IR/FlyOps.td	Broadens crd2idx layout operand type (allows swizzle), adds fly.decomposition op, and tweaks assembly formats.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[coderfeli]

Review Comments

Critical

1. i32 GEP overflow: AddOffsetOpLowering uses i32 offsets for all address spaces including Global. Offsets >2GB will overflow. Should keep i64 for non-BufferDesc pointers.

2. Null deref in AddOffsetOpLowering: offset.getDefiningOp() returns nullptr for block arguments, then defOp->getOperand(0) crashes. Need a null check.

3. Buffer-to-buffer copy dropped: lowerCDNA3BufferCopy now rejects both-sides-BufferDesc and both-sides-non-BufferDesc. Old code handled these cases. Potential regression if any kernel uses buffer-to-buffer copy.

4. CDNA3 MmaAtom C layout: Removed ValM1 from accumulator layout. For mfma with M>16, this changes how accumulator values distribute. Needs correctness verification across all MFMA variants.

Medium

5. Swizzle logic duplicated between BufferFatPtr::swizzleByteOffset and applySwizzleOnPtr — should extract a shared helper.

6. lowerUniversalCopy memcpy length changed from i64 to i32. Confirm LLVM::MemcpyOp accepts i32 length.

7. No MLIR tests for fly.decomposition or swizzle in crd2idx.

8. MakeViewOpLowering relies on use_empty() for CoordTensor — fragile pattern, add a comment explaining why.

[coderfeli]

a little conflict @sjfeng1999 after I changed the crd2idx to use in gemm kernel.

[sjfeng1999]

Review Comments

Critical

1. i32 GEP overflow: AddOffsetOpLowering uses i32 offsets for all address spaces including Global. Offsets >2GB will overflow. Should keep i64 for non-BufferDesc pointers.
2. Null deref in AddOffsetOpLowering: offset.getDefiningOp() returns nullptr for block arguments, then defOp->getOperand(0) crashes. Need a null check.
3. Buffer-to-buffer copy dropped: lowerCDNA3BufferCopy now rejects both-sides-BufferDesc and both-sides-non-BufferDesc. Old code handled these cases. Potential regression if any kernel uses buffer-to-buffer copy.
4. CDNA3 MmaAtom C layout: Removed ValM1 from accumulator layout. For mfma with M>16, this changes how accumulator values distribute. Needs correctness verification across all MFMA variants.

Medium

5. Swizzle logic duplicated between BufferFatPtr::swizzleByteOffset and applySwizzleOnPtr — should extract a shared helper.
6. lowerUniversalCopy memcpy length changed from i64 to i32. Confirm LLVM::MemcpyOp accepts i32 length.
7. No MLIR tests for fly.decomposition or swizzle in crd2idx.
8. MakeViewOpLowering relies on use_empty() for CoordTensor — fragile pattern, add a comment explaining why.

1. The type of offset in the GEP is coherent with the integer width in the int_tuple offset. It's not always i32.
2. After rewrite-func-signature pass, int_tuple value can't appear at the block_arguemnts.
3. It's unproper case anyway. CI will check if there is a regression.
4. Done, it's a true problem.

[coderfeli]

@sjfeng1999 rebased on main and fixed a minor run error in the moe and gemm. take a look.