Tracking: full device debug metadata for variables and types #226

Activity

1. nihalpasham

   added

   enhancementNew feature or requestNew feature or request

   debug infoDebug information in generated PTX/IR (line/source maps)Debug information in generated PTX/IR (line/source maps)

   llvm-exportTextual LLVM/NVVM IR exporter (llvm-export crate)Textual LLVM/NVVM IR exporter (llvm-export crate)

   codegenDevice code-generation pipeline (Rust MIR to IR to PTX)Device code-generation pipeline (Rust MIR to IR to PTX)

   trackingUmbrella tracking issueUmbrella tracking issue

   IR-loweringLowering between dialects (dialect-mir to LLVM dialect)Lowering between dialects (dialect-mir to LLVM dialect)

   on Jun 15, 2026
2. nihalpasham commented on Jun 16, 2026

   

   nihalpasham

   on Jun 16, 2026

   CollaboratorAuthor

   More actions

   Opened #233 as the first CUDA_OXIDE_DEBUG=full implementation slice.

   What #233 covers:

   rustc var_debug_info
     -> tagged MIR local alloca
     -> LLVM alloca + DILocalVariable + DIType + dbg.declare
     -> LLVM opt may salvage into #dbg_value
     -> cuda-gdb can inspect supported locals/args where locations survive
   

   A few choices from the implementation:

   • full debug skips Pliron mem2reg for now, because Pliron does not yet rewrite dbg.declare into dbg.value
   • LLVM opt -O2 stays enabled, since LLVM can salvage many debug locations itself
   • line-table mode strips only PTX .target ..., debug after llc so source breakpoints stay fast
   • full mode keeps PTX .target ..., debug because cuda-gdb needs it for local/argument inspection

   What remains here:

   • richer Rust type trees: ADTs, tuples, arrays, slices, closures
   • fragments/projections like fields, derefs, destructuring
   • precise lexical scopes
   • constants from var_debug_info
   • Pliron-side debug salvage so optimized passes can produce dbg.value directly
3. nihalpasham

   changed the title [-]Track full device debug metadata for variables and types[/-] [+]Tracking: full device debug metadata for variables and types[/+] on Jun 24, 2026
4. nihalpasham

   added sub-issues on Aug 21, 2026

   • Full debug: preserve pointee types for supported thin pointers and references #1119
   • Full debug: describe ordinary Rust device globals (AS1) #1120
   • Full debug: describe static shared-memory objects (AS3) #1121
   • Full debug: emit source-facing device function names and linkage names #1122
5. nihalpasham commented on Aug 21, 2026

   

   nihalpasham

   on Aug 21, 2026

   CollaboratorAuthor

   More actions

   cuda-devtools debugger-assessment follow-ups (validated against main 6c5458fe991bbde32c5bee74d87822aef1b5a691):

   Full variable/type debug children:

   • Full debug: preserve pointee types for supported thin pointers and references #1119 — preserve supported thin-pointer/reference pointee types
   • Full debug: describe ordinary Rust device globals (AS1) #1120 — describe ordinary Rust device globals (AS1)
   • Full debug: describe static shared-memory objects (AS3) #1121 — describe static shared-memory objects (AS3)
   • Full debug: emit source-facing device function names and linkage names #1122 — emit source-facing function names plus linkage names

   Related line-table follow-ups (outside #226's stated scope; see #224/#225):

   • Debug info: keep #[kernel] prologue instructions off the attribute line #1117 — keep generated #[kernel] prologue off attribute lines
   • Debug info: preserve MIR assert locations on generated traps #1118 — preserve MIR assert locations on generated traps

   External cuda-devtools dependency:

   • PTX trap is misclassified as stack overflow after ptxas reuses R1 on a terminal local-memory path #1123 — track the sm_120 terminal-trap R1-reuse recurrence against NVBugs 710864/711266 and relate the open work in 5158747

   The trap() versus breakpoint() API question is expected behavior, not an eighth issue: they intentionally match CUDA C++ __trap() and __brkpt().
6. nihalpasham

   added this to the cuda oxide v0.5.0 milestone on Oct 25, 2026