llvm-project
a63a548b - [Driver] Enable -ftime-trace for CUDA/HIP device compilation (#179701)

Commit
2 days ago
[Driver] Enable -ftime-trace for CUDA/HIP device compilation (#179701) Previously, -ftime-trace only generated trace files for host compilation when compiling CUDA/HIP code. Device compilation was excluded because the OffloadingPrefix was non-empty, causing handleTimeTrace() to be skipped. This patch enables -ftime-trace for offload device compilation by: 1. Passing the offloading prefix to handleTimeTrace() 2. Including the bound architecture in the trace filename 3. Deriving the trace output directory from the -o option for device compilation (since the device output is a temp file) Trace files are now generated for each offload target: - Host: output.json - Device: output-hip-amdgcn-amd-amdhsa-gfx906.json Note: When using --save-temps, multiple compilation phases (preprocess, compile, codegen) write to the same trace file, with each phase overwriting the previous. This is pre-existing behavior that also affects regular C++ compilation and is not addressed by this patch. This addresses a long-standing limitation noted in D150282.
Author
Parents
Loading