yxsamliu wrote:

Updated this to only skip the early cc1 optimizer for the default non-RDC HIP 
path.

The motivation is that default non-RDC now uses the linker wrapper `--no-lto` 
path, where bitcode is compiled to ISA through the normal non-LTO optimization 
pipeline. Running the optimizer in the earlier device cc1 job gives us two 
optimization pipelines, and the first one is not parallelized.

RDC and explicit offload LTO are now preserved: those cc1 jobs still produce 
optimized bitcode (`-flto=full`) and do not get `-disable-llvm-passes`, because 
the LTO flow expects optimized input bitcode. Final driver LLVM output 
(`-emit-llvm`) also keeps the requested optimization level.

https://github.com/llvm/llvm-project/pull/206350
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to