Since the last status update the opaque pointer migration has reached the point where both LLVM and Clang fully support opaque pointers (modulo some minor pending patches). I think it is time to talk about enabling opaque pointers by default.
The core problem with enabling opaque pointers is that they have major impact on tests: At least any tests with pointer types in their CHECK lines will have different output when opaque pointers are enabled (even if the difference is only cosmetic). As a rough estimate, there are 2.3k affected Clang tests and 4.5k affected LLVM tests. Updating so many tests in an atomic enablement commit is unrealistic, and would cause massive churn on reverts (and I do expect this to go through some revert cycles).
Something worth noting is that code that works with typed pointers will usually work with opaque pointers, but the converse is not true (this is because typed pointers require correct bitcasts, while opaque pointers don’t). As such, I think our general approach should be to enable opaque pointers on the Clang side first, so that the primary consumer no longer cares about typed pointers, and then mass-migrate tests to use opaque pointers.
There’s a few ways I could see this playing out. One of them would be:
- Annotate all affected Clang tests with
-opaque-pointers=0. - Enable opaque pointers in Clang by default.
- Gradually remove
-opaque-pointers=0from Clang tests while updating test expectations. - Gradually switch LLVM tests to use opaque pointers.
- Globally enable opaque pointers by default in LLVM.
Another one would be based on âš™ D122488 [OpaquePtrs][Clang] Add -normalize-opaque-pointers cc1 option, which adds an option for performing codegen with typed pointers and then rendering the final result with opaque pointers, with the intention that we still test typed pointer bitcast consistency, but make the final result the same as if opaque pointers were used.
- Gradually migrate all affected Clang tests to
-normalize-opaque-pointers. - Enable opaque pointers in Clang by default.
- Gradually switch LLVM tests to use opaque pointers.
- Globally enable opaque pointers by default in LLVM.
This one has the advantage that ability to test Clang with typed pointers would be retained, at least to some degree. However, I’m not really certain how well the “typed codegen rendered as opaque” to “opaque codegen” equivalence would work in practice, it’s possible that there would still be differences in results.
I’d appreciate some input on the preferred way to handle this. As discussed previously, the general plan here is that once opaque pointers are the default, typed pointer support will be maintained on a best-effort basis until the LLVM 15 release branch – that is, we’ll not go out of the way to break it, but also aren’t responsible for keeping it working and (most importantly) tested.
