* Make the necessary data implicitly shared since it might get
accessed/modified from two different threads with follow-up changes.
This applies for UnsavedFiles/UnsavedFile and ProjectPart::arguments().
* Avoid returning references.
Change-Id: I98842c1cb90ae0d344a15c63b72cbc89568722d3
Reviewed-by: Tim Jenssen <tim.jenssen@qt.io>
libclang 3.8 seems to be sensitive to file paths separators [1]. On Windows,
this led to not updated document annotations and/or crashes after reparsing.
When passing file paths to libclang, convert to native separators.
When getting file paths from libclang, convert back.
This handles:
* main file path
* file paths of the unsaved files
* -I<DIR> arguments, the resource path (for builtins) and the paths to the
wrapped qt headers
* included header files from libclang
* source locations from libclang
Also, minimize the conversion in SourceLocation to a minimum by making
filePath() lazy.
[1] https://llvm.org/bugs/show_bug.cgi?id=28381
Change-Id: If5866f34a6fdc6b34b16c022d3988e8e6eae2a0a
Reviewed-by: Christian Stenger <christian.stenger@qt.io>
The issue was re-producible with e.g.:
void g()
{
// Type 'foo.' as fast as possible in the next line
}
This led to "foo->" with completion results as if there was no "foo." at
all in that line.
We relied on a correct position for
translationUnit.sourceLocationAtWithoutReparsing(), but the just typed
characters were not yet reparsed. And we do not want to reparse at that
point since takes too long. We already determine the utf8 position for
the dot character, so simply use that instead.
This completes commit 17c1325cc4.
Change-Id: I669888b5c17ee63b2aec7b16c9921f9d79e281f9
Reviewed-by: David Schulz <david.schulz@theqtcompany.com>
The position of the dot character was determined on an outdated
translation unit.
We queried the translation unit for the source location of the dot
character, but apparently clang_codeCompleteAt() does not update the
source locations for the translation unit. And we do not want to reparse
since this is expensive. Thus, determine the byte position manually by
scanning over the document until the right line/column is reached.
Change-Id: I25e256bb81a83bb71c7e46a0fb3e927bf4031b16
Reviewed-by: Erik Verbruggen <erik.verbruggen@qt.io>
Reviewed-by: Eike Ziller <eike.ziller@qt.io>
...otherwise we might run into the following assertion:
Assertion failed: (IncludeBriefComments ==
this->IncludeBriefCommentsInCodeCompletion), function CodeComplete, file
ASTUnit.cpp, line 2411.
Change-Id: I4723b600c7ac5aa2b4c1cb1827f51156afb492b8
Reviewed-by: Erik Verbruggen <erik.verbruggen@theqtcompany.com>
...by explicitly checking for the dot in the source.
Task-number: QTCREATORBUG-15654
Change-Id: I4172e88a7fbb3015ef391daf13ded1f0002aab9c
Reviewed-by: Marco Bubke <marco.bubke@theqtcompany.com>
This simplifies UnsavedFiles and makes TemporaryModifiedUnsavedFiles
useless.
Change-Id: I1896f971215ed22ce7aa7bf21b16381862b7469d
Reviewed-by: Marco Bubke <marco.bubke@theqtcompany.com>
1 struct Foo { int member; };
2 void f(Foo *foo)
3 {
4 foo.<REQUEST COMPLETION> // correct '.' to '->' and provide results
5 }
The preferred approach would be to check if "foo" in line 4 is of
pointer type, but there is no suitable cursor (only CompoundStmt) at
that position since the code is usually not yet parsed and thus invalid.
Thus, just run the completion as is. If there are not any results for a
dot completion, re-run the completion with "." exchanged by "->". This
approach is inherently slower than the preferred approach implemented in
the built-in code model.
The following rare cases are not handled:
1) Requesting completion after white space:
Works: foo.<COMPLETE HERE>
Fails: foo. <COMPLETE HERE>
2) Opening a file and requesting completion (ctrl+space) without prior
editing. No editing before triggering completion means that no
unsaved file is generated on the backend side, which is a
requirement for the correction.
Task-number: QTCREATORBUG-11581
Change-Id: I6bc8e8594778774ab342755fdb01a8a3e5c52ba0
Reviewed-by: Marco Bubke <marco.bubke@theqtcompany.com>
Includes are now watched by a file watcher. Unsaved file changes are
watched too. If they are changed the translation units which depend on
them are set to a state which require a reparse. Later the diagnostics
of this units are collected and send back to creator.
Change-Id: I2fb5c7dd6644687f22399edd8d18edd6215c9505
Reviewed-by: Nikolai Kosjar <nikolai.kosjar@theqtcompany.com>
using CodeCompletions = QVector<CodeCompletion>;
using CodeCompletionChunks = QVector<CodeCompletionChunk>;
Change-Id: Id729a43566c23dc000e44d79ccf4a49e4f235f51
Reviewed-by: Marco Bubke <marco.bubke@theqtcompany.com>
There are two ways to hot fix the dependency problem. First version is to
disable the caching(not the preamble) but it is triggering bugs. Second
version is to enable caching and reparse before every complete. I opted
for the second version because it is the commonly used path for most
implementations. Later we have to minimize the reparse calls but for that
we need the cooperation of editor.
Change-Id: I878afb896cbf2338cbeb13d9e596a7776dd49854
Reviewed-by: Nikolai Kosjar <nikolai.kosjar@theqtcompany.com>
