[PATCH] D33589: clang-format: consider not splitting tokens in optimization

[Francois Ferrand via Phabricator via cfe-commits]

Typz added a comment.

In https://reviews.llvm.org/D33589#875039, @djasper wrote:

> I think doing the computation twice is fine. Or at least, I'd need a test case where it actually shows substantial overhead before doing what you are doing here. Understand that creating more States and making the State object itself larger also has cost and that cost occurs in the combinatorial exploration of the solution space. Doing an additional computation at the end should be comparatively cheap. Look at it this way: During the exploration of the solution space, we might enter breakProtrudingToken many times for the same comment. One more time during reconstruction of the solution is not that harmful.


I 've just tried to implement this (e.g. make the 2 calls also in DryRun), and I am running into an assertion in WhitespaceManager :

  [ RUN      ] FormatTest.ConfigurableUseOfTab
  FormatTests: /workspace/llvm/tools/clang/lib/Format/WhitespaceManager.cpp:112: void clang::format::WhitespaceManager::calculateLineBreakInformation(): Assertion `PreviousOriginalWhitespaceEndOffset <= OriginalWhitespaceStartOffset' failed.

This remind there was another "reason" for limiting this to DryRun (not sure it is a good or bad reason): it happens only in the optimizer, all other cases where the indenter is used are not affected.

I am still trying to get to the bottom of this assertion, any hint where to look for?


https://reviews.llvm.org/D33589