yxsamliu wrote:

Thanks, both suggestions make sense and I'll take them.

On the split: I'll break this into a runtime PR and a compiler PR. The runtime
part (the compiler-rt profile library changes) is self-contained and can land
first; the compiler part (the LLVM instrumentation and Clang codegen that emit
calls into the runtime) depends on the runtime ABI but builds independently, so
it can follow. That keeps each PR smaller and lets the disruptive part land on
its own.

On the separate library: I like it, and I think it's the cleanest of the
options. I'll add an opt-in `clang_rt.profile_rocm` that carries the ROCm
support (the collection runtime plus the interceptor dependency) and is built
the way that target needs (e.g. /MD on Windows), and leave the existing
`clang_rt.profile` unchanged so current users are not affected. The driver will
link the ROCm library on the host side only when building HIP with
`-fprofile-generate`.

I'll close this PR and open the runtime PR in that shape. Thanks for the review.


https://github.com/llvm/llvm-project/pull/201499
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to