[libc-commits] [PATCH] D152575: Added modf for NVPTX and AMDGPU targets to implement 'libmgpu.a' for math on the GPU

[Siva Chandra via Phabricator via libc-commits]

sivachandra added a comment.

In D152575#4421826 <https://reviews.llvm.org/D152575#4421826>, @jhuber6 wrote:

> That's a good question, we'd probably like to have both if we'd like to perform some performance tests, but I'm not sure if that's a good reason to keep it in-tree if we have a suitable alternative. @sivachandra What do you think?

Per my understanding from the earlier discussion, we will add a vendor-wrapper only if the in-tree implementation is not sufficiently proven to be a good enough replacement for the vendor implementation. If that is correct, I would like to stick to that. You shouldn't need a wrapper in the libc for differential testing.

That said, will it ever be that a builtin for a floating point primitive will not be available? If yes, then the ideal approach we should take is to "fix the compiler". If that is not practical, we can take up adding vendor wrappers at that time with reduced scope (as in, that option is taken only if the builtin is not available.)


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D152575/new/

https://reviews.llvm.org/D152575