Fix Theta(N^2) memory usage of EXPECT_EQ(string) when the strings don't match.
The underlying CalculateOptimalEdits() implementation used a simple
dynamic-programming approach that always used N^2 memory and time. This meant
that tests for equality of large strings were ticking time bombs: They'd work
fine as long as the test passed, but as soon as the strings differed the test
would OOM, which is very hard to debug.
I switched it out for a Dijkstra search, which is still worst-case O(N^2), but
in the usual case of mostly-matching strings, it is much closer to linear.
PiperOrigin-RevId: 210405025
GTEST_TEST_FILTER_ENV_VAR_ was used to specify an environment variable to obtain
the default test filter from. By default it was unset which broke
"--test_filter" for bazel. This CL eliminates GTEST_TEST_FILTER_ENV_VAR_ and
explicitly obtains the default test filter from the environment variable
TESTBRIDGE_TEST_ONLY if it exists.
This canonicalizes demangled names by omitting a nested inline namespace within namespace std if the name of the nested namespace begins with a double underscore. This improves compatibility with libc++.
This allows doing things like TEST_P(TestFixture, MAYBE(TestName))
for nicer conditional test disabling.
Upstream of cr/188748737.
Tested:
Added unit tests MacroNamingTest and MacroNamingTestNonParametrized.
Google Test has recently (02/09/2018) switched the default death test
style from "fast" to "threadsafe" in
ec7faa943d
Threadsafe death tests have been used internally for a while, and are
proven to be a better default.
However, adopting this better default can be challenging for large
projects with a significant investment in custom infrastructure built on
top of Google Test. The same custom infrastructure can make it difficult
for large projects to switch back to the old default by passing in
--gtest_death_test_style=fast.
For the reasons above, the default switch is considered too disruptive,
and this CL reverts it. This CL also introduces the
GTEST_DEFAULT_DEATH_TEST_STYLE preprocesor macro, which replaces the
hard-coded default. The macro can be defined in
gtest/internal/custom/gtest-port.h by projects that are ready to migrate
to thread-safe death tests.
_HAS_EXCEPTIONS is specific to the MSVC STL and defining it to 0 causes
problems with libc++, so libc++ users may leave it undefined. This can
cause GTEST_HAS_EXCEPTIONS to be defined incorrectly if the user has
disabled exceptions via the compiler, which can lead to build errors.
_CPPUNWIND is a builtin macro provided by the compiler so it should
work with both STLs.
clang-cl is clang for Windows running in MSVC mode. Chromium uses it for
Windows builds. clang-cl is weird in that it defines __clang__ and
_MSC_VER, but *NOT* __GNUC__. This is vaguely analogous to how normal
clang defines __clang__ (what it is) and __GNUC__ (what it is compatible
with).
However, clang-cl still implements most GCC extensions, being clang.
Notably, the way to control -Wformat-literal is still with
__attribute__((__format__)). For better error-checking and strict
-Wformatl-literal compatibility (see
53c478d639), define
GTEST_ATTRIBUTE_PRINTF_ in clang-cl too.
On platforms with std::tuple and not std::tr1::tuple, GTEST_HAS_COMBINE
gets turned off when it works fine (due to GTEST_TUPLE_NAMESPACE_).
Elsewhere in the project, several GTEST_HAS_TR1_TUPLE checks
additionally check GTEST_HAS_STD_TUPLE_, so use that formulation.
(The ones that don't are specific to std::tr1::tuple and are followed by
an identical GTEST_HAS_STD_TUPLE_ version underneath it.)
In particular, this fixes testing::Combine on MSVC 2017, which regressed
here:
https://github.com/google/googletest/pull/1348#issuecomment-353879010
ThreadLocal class needs to be have default visibility.
Root cause is gtest uses typeinfo for the ThreadLocal class.
The problem manifests When gtest/gmock are built as a shared library
with libc++. When a class is used in typeinfo, it must have default
visibility.
There is an explanation about typeinfo and visibility here:
https://developer.apple.com/library/content/documentation/DeveloperTools/Conceptual/CppRuntimeEnv/Articles/SymbolVisibility.html
When libc++ is used with gtest in shared library mode, any tests
that are compiled with -fvisibility=hidden and exercise the
macro EXPECT_CALL, it results in an abort like:
[ FATAL ] /usr/include/gtest/internal/gtest-port.h:1394::
Condition typeid(*base) == typeid(Derived) failed.
This is because the typeinfo for ThreadLocal class is not visible.
Therefore, linker failed to match it to the shared library symbol, creating a
new symbol instead.
This fixes https://github.com/google/googletest/issues/1207.
