[PATCH] D140558: AMDGPU: Fix broken and permissive handling of printf format strings

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added a comment.

In D140558#4025024 <https://reviews.llvm.org/D140558#4025024>, @sameerds wrote:

> This pass is used only by OpenCL on AMDGPU. We don't really need to handle errors at the LLVM IR level, since Clang will have already validated the printf call under the OpenCL spec. It's sufficient to just return -1 on anything invalid according to OpenCL printf spec.

This is operating on IR, it needs to handle the whole IR. This isn't diagnosing end user errors, but everything that could have happened to the IR in and after the emission in clang. It needs to validate it's handling only exactly what's expected from clang.

> If we really should keep this check, can you please make sure it passes internal OpenCL builds on AMDGPU?

Yes, but the printf testing is clearly not worth anything since it didn't catch any of these opaque pointer issues


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D140558/new/

https://reviews.llvm.org/D140558