* Added scopeinfo member to token class
Moved ScopeInfo2 declaration here as well because that's where it needs to be now.
* Added scopeinfo accessors and declaration to class
* Add new method for calculating scopes
This replaces the methods in the TemplateSimplifier which calculate the current scope as the token list is iterated. The old method required checking if the scope had changed for every token multiple times (for multiple iterations), which was surprisingly costly. Calculating scopes in advance like this decreases runtime on a worst-case file by around thirty percent.
ScopeInfo objects are disposed of when the TemplateSimplification is done as they are not used later.
* Add calculateScopes method to header
* Removed code that calculated current scope
This has been replaced by code that calculates the scopes up front and stores them with each token, which is much faster.
* Fixed compile errors from extra parentheses
* Added missing code to fix memory leak
* Added code to actually clean up ScopeInfo structs
* Tidy up a dodgy for loop
* Convert argument to const ref
* Calculate missing scopes
As the templatesimplificator expands templates and does multiple passes it needs to make sure all scopes are calculated.
* Remove copying the scope to the next token
This is now done properly when scopes are calculated.
* Remove call to calculateScopes
This is now done by the TemplateSimplifier.
* Recalculate scopes for every pass of simplifyTemplates
* Add code to calculate extra scopes as they are added
I thought that this might be useful for calculating scopes when Tokens are created, but as there are several ways of creating Tokens that don't guarantee that they are placed in a list it is easier to just calculate scopes when you know you have a list and when you know you're adding to a list.
* Fix several bugs and poorly designed code
Remove the global scopes collection, and clean them up instead by iterating through the tokenlist to find them. This means scopes can be calculated by functions in the Token class as well as in the Tokenizer class without leaking the scope object.
Fix a couple of bugs in the calculateScopes method and make it more efficient.
* Remove unnecessary calls to calculateScopes
* Move brace to correct position
Calculating scopes during insertToken only needs to happen if we created a new Token.
* Handle 'using namespace' declarations separately
This fixes a bug caused by a statement matching 'struct B < 0 > ;'
* Fix argument name mismatch
* Actually use newScopeInfo when inserting Token
* Switch to using shared_ptr to hold scopeInfos
This means ScopeInfo2 objects get properly cleaned up when they are no longer needed.
* Change ScopeInfo member to be a shared_ptr
* Update code to use shared_ptr
* Add missing include for shared_ptr
* Remove unnecessary cleanup code
This has been replaced by shared_ptr for ScopeInfo2 objects
This will warn for cases where searching in an associative container happens before insertion, like this:
```cpp
void f1(std::set<unsigned>& s, unsigned x) {
if (s.find(x) == s.end()) {
s.insert(x);
}
}
void f2(std::map<unsigned, unsigned>& m, unsigned x) {
if (m.find(x) == m.end()) {
m.emplace(x, 1);
} else {
m[x] = 1;
}
}
```
In the case of the map it could be written as `m[x] = 1` as it will create the key if it doesnt exist, so the extra search is not necessary.
I have this marked as `performance` as it is mostly concerning performance, but there could be a copy-paste error possibly, although I dont think thats common.
Change the astStringVerbose() recursion to extend a string instead of
returning one. This has the benefit that for tokens where the recursion
runs deep (typically large arrays), the time savings can be substantial
(see comments on benchmarks further down).
The reason is that previously, for each token, the astString of its
operands was constructed, and then appended to this tokens astString.
This led to a lot of unnecessary string copying (and with that
allocations). Instead, by passing the string by reference, the number
of temporary strings is greatly reduced.
Another way of seeing it is that previously, the string was constructed
from end to beginning, but now it is constructed from the beginning to
end. There was no notable speedup by preallocating the entire string
using string::reserve() (at least not on Linux).
To benchmark, the changes and master were tested on Linux using the
commands:
make
time cppcheck --debug --verbose $file >/dev/null
i.e., the cppcheck binary was compiled with the settings in the
Makefile. Printing the output to screen or file will of course take
longer time.
In Trac ticket #8355 which triggered this change, an example file from the
Wine repository was attached. Running the above cppcheck on master took
24 minutes and with the changes in this commmit, took 22 seconds.
Another test made was on lib/tokenlist.cpp in the cppcheck repo, which is
more "normal" file. On that file there was no measurable time difference.
A synthetic benchmark was generated to illustrate the effects on dumping
the ast for arrays of different sizes. The generate code looked as
follows:
const int array[] = {...};
with different number of elements. The results are as follows (times are
in seconds):
N master optimized
10 0.1 0.1
100 0.1 0.1
1000 2.8 0.7
2000 19 1.8
3000 53 3.8
5000 350 10
10000 3215 38
As we can see, for small arrays, there is no time difference, but for
large arrays the time savings are substantial.
This handles concatenated strings and characters from simplecpp.
Previously, L'c' would be preprocessed to the tokens "L" and "'c'".
cppcheck would then remove the "L" token and set "'c'" to be a wide
character literal. Now, it needs to remove the prefix instead.
When doing this, add handling of utf32 encoded literals (U) and UTF-8
encoded literals (u8).
* Code changes for Token::mImpl optimisation
* Added new TokenImpl optimisation
Moving members to the TokenImpl struct reduces the size of the Token class, which is a fairly significant optimisation. In my testing on Windows with 32-bit Release-PCRE, this change reduced the size of the Token class from 108 bits to 52 bits and reduced run-time of my test case by around 20%.
* Several optimisations
Deleted some code that ran very slowly and did nothing, as there is no need to change a Token's string to null if you are about to delete it.
Added a frontToken to simplifyCalculations to reduce the amount of work it has to do on already-simplified calculations.
Moved template removal to the end of the list as this reduces redundant iteration and saves time.
* Added tok argument to simplifyCalculations
This means callers can avoid unnecessary work if they know which tokens have already been simplified. Passing nullptr indicates the original behaviour (starting from the front of the list).
* Removed mention of member from another change
* Re-added and optimised some code deleted in error
Changing mTemplateInstantiations to a vector avoids the high cost of doing repeated linear searches. Changing how the code iterates through the array was necessary because the vector can be resized at several points during the loop, which breaks existing references and iterators.
* Changed mTemplateInstantiations to a vector
This is an optimisation that makes repeated linear searches of this collection significantly faster.
Also added a copy constructor to TokenAndName so code can make copies of these objects to keep a reference if a vector gets resized.
* A cleaner optimisation to removing template tokens
This reverts the previous change to made mInstantiatedTemplates a vector and the iterator changes to support this, and makes mTypesUsedInTemplateInstantiation so the eraseTokens logic can be unified.
* Reverted vector to list
Also made mTypesUsedInTemplateInstantiation a vector of TokenAndName objects so it can share the same logic as the other members.
* Added member for template simplifier pointer
This can be used more efficiently than marking Tokens with a flag and then searching through all templates to find the one that matches.
* Turned loop inside out
This means we only have to iterate through the std::list once. std::list is very expensive to iterate through.
* Latest code from danmar and fixed optimisations
In particular I have optimised simplifying template instantiation names as this was incredibly slow because of the number of times it had to iterate through the template instantiation list. Previous optimisations to this weren't very effective and broke some edge cases.
* Added changes from danmar
Made mExplicitInstantiationsToDelete a vector of TokenAndName to be consistent with the rest of the members, which are cleaned up very efficiently.
* Tokens can have many templateSimplifierPointers
* templateSimplifierPointers must be kept in sync
* Fixed#8693 (Template specialization: Constructor detected as normal function (functionStatic error))
Refactor template simplifier to remove the existing full specialization
function expandSpecialized and allow full specializations to use the
existing function expandTemplate. The function expandTemplate was
modified to either expand the template like it originally did by copying
it or to modify the template in place. Both instantiated and
uninstantiated full specializations are modified in place. This also
fixes#8692 and probably other related tickets as well.
The function simplifyTemplates now tries twice to simplify templates so
more templates can be simplified. We should try as many times as
necessary to find all possible templates. We can't do that now because
uninstantiated templates are left unchanged. It is relatively straight
forward to have the new code also expand in place uninstantiated
templates with their symbolic types but namespaces are not handled
properly (ticket #8671) and it would introduce regressions.
* Fix travis warnings.
* Add tests for invalid ranges
* Refactor loadLibErrors
This reduces the amount of code slightly and will simplify adding
more tests.
* Handle empty valid field
Before this change, the sequence <valid></valid> in a config file would
result in a segmentation fault. Now an empty field results in the error
message:
cppcheck: Failed to load library configuration file 'mycfg.cfg'. Bad attribute value '""'
* Add support for valid for floating point arguments
Previously, it was not possible to add valid ranges to floating point
arguments since it only handled integers. This made ranges not work well
for floating point arguments since arguments were cast to integers
before the ranges were handled.
Fix this by using doubles instead of integers if the argument is a float.
Add some tests for this and make sure errors are printed with enough
precision (somewhat arbitrarily chosen).
Note that it is still only possible to add integer ranges (i.e. -1:1).
* Add support for floats in configuration valid range
Now that it is possible to handle decimal arguments, there is no reason
to not allow non-integer ranges. Take care to not allow broken
configurations.
* Move check to within if-clause
* Move asin{,f,l} and acos{,f,l} input checks to config file
Daniel Marjamäki
versat
Paul Fultz II
rebnridgway
orbitcowboy
adamjrichter
Rikard Falkeborn
IOBYTE