vikramRH wrote:

> If it's not failing in expensive checks with the legacy PM with the verifier 
> enabled, why is it tripping the verifier with the NewPM? That seems to imply 
> an issue with the new pipeline to me.

I think there is a difference in understanding about the motivation of the 
change here,
1. This tries to control the machine verifier being added to the default 
pipelines (O0, O1...) with or without expensive checks enabled. The primary 
motivation was several verifier failures that I saw while doing NPM testing 
downstream. (i.e compiling several ROCm math libraries and applications using 
the AMDGPU NPM backend with expensive checks disabled).
2. Legacy PM does not enable machine verifier as part of default pipelines at 
all (Only IR verifiers are enabled at specific points). machine verifier would 
still trip if I try to compile the before said applications with expensive 
checks enabled compiler.
3. When I said the "expensive build checks currenlty being run for the target" 
in my previous comment, I was referring to the tests run as part of LLVM 
pre/post merge checks which I think are not exhaustive enough. (I think its 
just lit tests ?)



https://github.com/llvm/llvm-project/pull/176693
_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits

Reply via email to