Enabling opaque pointers by default

LLVM Project
9

I did some additional testing with the -normalize-opaque-pointers approach, and unfortunately I don’t think this is going to be viable after all. My assumption here was that -opaque-pointers and -normalize-opaque-pointers would produce the same result in nearly all cases, but this turns out to not be the case.

For example, for a set of 311 OpenMP codegen tests passing -normalize-opaque-pointers, 85 of them fail with -opaque-pointers. There are various reasons for this. One is differences in constant expression folding. Another is differences in value names – and in basic block names in particular, which normally do not use FileCheck placeholders.

I don’t think the normalization approach makes sense if it produces different output in such a significant fraction of tests. So I think we should go with the first option. Based on the feedback, this would go as follows:

  1. Add -no-opaque-pointers to Clang tests that are affected by opaque pointers.
  2. Enable opaque pointers in Clang by default.
  3. Wait a while so people following tip-of-trunk can incorporate the change and report any issues.
  4. Gradually remove -no-opaque-pointers from tests while updating check lines. At this point typed pointer support will likely start to bitrot (but we’ll still accept patches for fixing it).