* ConcatAnalyzer improvements.
Deduplicate code.
Improve type inference.
Allow literal type inference when only one side has multiple types (fixes#5483).
Fix invalid type inference with negative int as right operand.
* Fix inference to be lowercase-string when concatenating int.
* Fix TNonEmptyLowercaseString to not be subtype of TNonFalsyString.
'0' is a non-empty-lowercase-string that is falsy.
* Fix other issues with non-falsy-string.
* Nest ands and ors
Co-authored-by: Matthew Brown <github@muglug.com>
* #4997 added more precise stub for `count()` returning `0` or `positive-int` on known types
* #4997 updated `count()` to support `\SimpleXmlElement` and `\ResourceBundle` counting, as well as handling hardcoded 2-element-arrays cases
This patch:
* adds support for `count(\SimpleXmlElement)` (https://www.php.net/manual/en/simplexmlelement.count.php)
* adds support for `count(\ResourceBundle)` (https://www.php.net/manual/en/resourcebundle.count.php)
* removes usage of global constants from stub (not supported - see https://www.php.net/manual/en/function.count.php)
* adds support for identifying fixed-element-count arrays, for example `count(callable&array)`, which is always `2`
* #4997 adapted `FunctionCallReturnTypeFetcher` to infer `TPositiveInt` for `count(TNonEmptyArray)` and `count(TNonEmptyList)`
* The `FunctionCallReturnTypeFetcher` is responsible for defining the precise type of a `\count(T)`
expression when given a `T`, so we baked the whole type resolution for `positive-int`, `0` and
`positive-int|0` directly in there.
While this complicates things, it is also true that it is not possible right now (for the stubs)
to provide the level of detail around `count()` that is required by the type inference system
for such a complex function with so many different semantics.
* Run tests in random order
Being able to run tests in any order is a pre-requisite for being able
to run them in parallel.
* Reset type coverage between tests, fix affected tests
* Reset parser and lexer between test runs and on php version change
Previously lexer was reset, but parser kept the reference to the old
one, and reference to the parser was kept by StatementsProvider. This
resulted in order-dependent tests - if the parser was first initialized
with phpVersion set to 7.4 then arrow functions worked fine, but were
failing when the parser was initially constructed with settings for 7.3
This can be demonstrated on current master by upgrading to
nikic/php-parser:4.9 and running:
```
vendor/bin/phpunit --no-coverage --filter="inferredArgArrowFunction" tests/ClosureTest.php
```
Now all tests using PHP 7.4 features must set the PHP version
accordingly.
* Marked more tests using 7.4 syntax
* Reset newline-between-annotation flag between tests
* Resolve real paths before passing them to checkPaths
When checkPaths is called from psalm.php the paths are resolved, so we
just mimicking SUT behaviour here.
* Restore newline-between-annotations in DocCommentTest
* Tweak Appveyor caches
* Tweak TravisCI caches
* Tweak CircleCI caches
* Run tests in parallel
Use `vendor/bin/paratest` instead of `vendor/bin/phpunit`
* Use default paratest runner on Windows
WrapperRunner is not supported on Windows.
* TRAVIS_TAG could be empty
* Restore appveyor conditional caching
