Location information with inlined frames

[maleadt]

PTX ISA 7.2 now supports .loc with inlined_at: https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#debugging-directives-loc. This is huge, as we currently attribute so much source code to e.g. LLVM.jl's call_function.

We should look into making LLVM emit this.

[maleadt]

Had a brief look into this, and it's not immediately obvious where to best add this.

The .loc statements are generated by the generic MCAsmStreamer's emitDwarfLocDirective, https://github.com/llvm/llvm-project/blob/e63455d5e0e5d148215a61c0e4d10f00aaf6eb56/llvm/lib/MC/MCAsmStreamer.cpp#L1570-L1573, where there's no access to the full debug location (to access the inlinedAt info). I guess this needs to be moved upstream, e.g., a more specific DwarfDebug::emitInlinedAtLocDirective/MCAsmStreamer::emitDwarfInlinedAtLocDirective that could be added to DwarfDebug::beginInstruction directly (i.e. not using recordSouceLine where the full debug location is lost), https://github.com/llvm/llvm-project/blob/829616c24119f000132913616a7cd4ad930ca84a/llvm/lib/CodeGen/AsmPrinter/DwarfDebug.cpp#L1964, similar to emitInitialLocDirective as called by beginFunction, https://github.com/llvm/llvm-project/blob/829616c24119f000132913616a7cd4ad930ca84a/llvm/lib/CodeGen/AsmPrinter/DwarfDebug.cpp#L2154-L2155.

The other problem is that these inlined_at .loc statements are PTX-specific, so I'm not sure whether that functionality belongs in DwarfDebug/MCAsmStreamer, or what would be the best way to customize this. Maybe a class PTXDwarfDebug : public DwarfDebug?

[maleadt]

LLVM 23 will support this: llvm/llvm-project#170239

[maleadt]

Another step towards this: JuliaGPU/GPUCompiler.jl#801