SynthDebugInfo

Abstract

None of the three OSS cuda-tile transform previews ships in tileiras in its original cuda-tile-dialect form. FuseFMA.cpp is superseded by lower compiler layers. LoopSplit.cpp is absent without a TileIR-equivalent replacement. SynthesizeDebugInfoScopes.cpp is replaced by upstream MLIR's LLVM debug-scope pass, with only shared location helper behavior surviving.

That means a reimplementation should not blindly copy the public Transforms/ directory into the tileiras pipeline. The public files are useful for understanding the historical cuda_tile tool, but the released compiler routes FMA, loop shaping, and debug-scope synthesis elsewhere.

FuseFMA

The OSS pass rewrites mulf/addf and mulf/subf patterns into cuda_tile.fma when rounding modes and modifiers agree. Tileiras keeps the cuda_tile.fma operation, but not the pass that searches for these patterns at the cuda_tile layer.

FMA formation is delegated to lower layers:

tileas-legalize-fma-dot handles TileAS-level MMA accumulator contraction.
-nvptx-fma-level controls scalar FMA formation after lowering to LLVM/NVPTX IR.
-enable-fma-to-ffma2 covers the backend's F2 fused variant.

This is a semantic decision. Fusing (a * b) + c changes double-rounding into a single-rounded FMA, so tileiras places the scalar decision under the same backend policy that nvcc --fmad controls.

LoopSplit

The OSS LoopSplit.cpp pass walks cuda_tile.for loops and splits a loop when an inner cuda_tile.if predicate flips at a loop-invariant boundary. Tileiras does not ship that pass and does not provide an equivalent TileIR or TileAS pass.

The nearest named relative is loop unrolling, not loop splitting. Schedule materialization can decompose some guarded loop structure earlier in the pipeline, but it is not the same predicate-based loop-split transform. A compatible clone should not add OSS LoopSplit unless it is intentionally adding functionality beyond tileiras.

Debug Scope Synthesis

The OSS SynthesizeDebugInfoScopes.cpp pass is replaced by upstream MLIR's LLVM function-scope debug pass. The replacement pass is anchored on builtin.module, requires the LLVM dialect, emits compile units with producer "MLIR", and supports the standard emission-kind enum:

Value	Emission kind
`0`	`None`
`1`	`Full`
`2`	`LineTablesOnly`
`3`	`DebugDirectivesOnly`

The important behavioral difference is where locations are attached. The OSS pass rewrites per-op locations to DILocAttr. Tileiras leaves that work for the later ConvertDebugInfoToLLVM path, which consumes debuginfo.value operations after LLVM-dialect lowering. The scope pass itself walks LLVM functions and attaches function-level DISubprogramAttr information.

Delta Summary

OSS transform	Tileiras behavior	Compatibility decision
`FuseFMA.cpp`	Not present as a cuda-tile pass; superseded by TileAS and NVPTX backend policy.	Do not register OSS `fuse-fma` in the tileiras-compatible pipeline.
`LoopSplit.cpp`	Not present; no equivalent TileIR/TileAS split pass.	Do not add a loop-split substitute for compatibility.
`SynthesizeDebugInfoScopes.cpp`	Replaced by upstream LLVM function debug-scope pass.	Use `DIScopeForLLVMFuncOp` and leave per-op location lowering downstream.

Reimplementation Notes

void configure_tileiras_transform_pipeline(Pipeline *pipeline, OptLevel opt_level) {
    add_tileas_legalize_fma_dot(pipeline);
    set_nvptx_fma_level(pipeline, 2);

    if (opt_level == OPT_O3) {
        add_di_scope_for_llvm_func_op(pipeline, DEBUG_DIRECTIVES_ONLY);
    } else {
        add_di_scope_for_llvm_func_op(pipeline, LINE_TABLES_ONLY);
    }

    add_convert_debug_info_to_llvm(pipeline);
}

The key omission is intentional: do not add cuda-tile FuseFMA, cuda-tile LoopSplit, or the cuda-tile-specific debug-scope pass when targeting tileiras behavior.

Keyboard shortcuts