[PATCH] D117778: [OpaquePtrs] Add -normalize-opaque-pointers option

[Nikita Popov via Phabricator via llvm-commits]

nikic added a comment.

In D117778#3261936 <https://reviews.llvm.org/D117778#3261936>, @aeubanks wrote:

> The issue with the opaque pointer migration is that passes sometimes don't handle them correctly, not that we can't print the final IR. This patch still doesn't test that passes properly handle both typed pointers and opaque pointers, unless the idea is that we'd have RUN lines with both `-opaque-pointers` and `-normalize-opaque-pointers` which this patch doesn't do for the modified tests.

The idea is that the tests can be run with both `-opaque-pointers=0` and `-opaque-pointers=1`. They will use `-opaque-pointers=0` by default, but they can be run with `-opaque-pointers=1` and will pass -- for now, that would just be manual runs by interested parties, but if we can have a significant fraction of tests passing, that may be a buildbot.

Of course, another possibility would be to add both a `-normalize-opaque-pointers` and an `-opaque-pointers' RUN line, so that everyone tests both modes. That would double the number of RUN lines (for tests using pointers) though, but maybe it's preferred?

> As for a flag flip day, it would be per-frontend right? We can already incrementally update tests (including frontend e.g. clang) by adding `-opaque-pointers` to RUN lines.

Adding `-opaque-pointers` to tests means we lose typed pointer test coverage though. I don't think that's a good idea while typed pointers are still the default mode.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D117778/new/

https://reviews.llvm.org/D117778