This is an archive of the discontinued LLVM Phabricator instance.

Fix a few issues while skipping function bodies
ClosedPublic

Authored by ogoffart on May 31 2016, 10:03 AM.

Details

Summary

Fix a few issues while skipping function bodies

  • In functions with try { } catch { }, only the try block would be skipped, not the catch blocks
  • The template functions would still be parsed.
  • The initializers within a constructor would still be parsed.
  • The inline functions within class would still be stored, only to be discared later.
  • Invalid code with try would assert (as in "int foo() try assert_here")

This attempt to do even less while skipping function bodies.

Diff Detail

Event Timeline

ogoffart updated this revision to Diff 59084.May 31 2016, 10:03 AM
ogoffart retitled this revision from to Fix a few issues while skipping function bodies.
ogoffart updated this object.
ogoffart added reviewers: cfe-commits, rsmith, akyrtzi.
rsmith added inline comments.May 31 2016, 10:41 AM
lib/Parse/ParseStmt.cpp
1987

Comments should start with a capital letter and end with a period.

1988

This is not correct. Skipping constructor initializer lists correctly is hard (which is probably why we used to parse the initializer list rather than skipping it). You can use Parser::ConsumeAndStoreFunctionPrologue to get this (approximately) correct.

ogoffart updated this revision to Diff 59197.Jun 1 2016, 4:29 AM

Right, i forgot about the C++11 initializer list syntax. I hope I got it right now.

The idea is that when we see a ") {" or "} {" in the ctor-initializers, (optionally with "..."), it is necessarily the start of the body.

Unless there might be lambda expressions within a template aregument, as in:

  
A::A() : Base<[](){return 42; }()>(){}

But this is not valid because lambda expression cannot appear in this context as far as i know since lambda expression cannot appear in template argument

rsmith edited edge metadata.Jun 9 2016, 1:46 PM

Please call Parser::ConsumeAndStoreFunctionPrologue rather than trying to reinvent it. You're still getting a number of cases wrong that it handles properly.

You also need to handle the case where the code completion token appears within the constructor *mem-initializer*s.

ogoffart updated this revision to Diff 60340.Jun 10 2016, 4:48 AM
ogoffart edited edge metadata.

Using Parser::ConsumeAndStoreFunctionPrologue this time

rsmith accepted this revision.Jun 16 2016, 12:22 PM
rsmith edited edge metadata.

LGTM

This revision is now accepted and ready to land.Jun 16 2016, 12:22 PM
This revision was automatically updated to reflect the committed changes.
chapuni added inline comments.
cfe/trunk/unittests/Tooling/ToolingTest.cpp
315 ↗(On Diff #61035)

It didn't pass for targeting *-msvc. Tweaked in r272985.