Some versions of MinGW do not support enough of Win32 API to let us work
with SEH, so SEH is now MSVC only (+ configurable toggle).
Also made use of gmtime_s MSVC only (as oposed to Windows only).
Fixes#805
Because the signal changes were in a different branch from the windows.h
related changes, the SEH handling code included the header directly.
Fixes#803
This prevents Clang from complaining about unused value in expressions
containing explicit casts used in the THROW assertion macro family.
Example:
`REQUIRE_THROWS_AS(static_cast<bool>(object), std::bad_cast);` would
trigger `-Wunused-value` warning. Now it does not.
Credits to Arto Bendiken, who submitted a PR almost 3 years ago, but his
branch has since died and I was unable to merge it.
The integrated assembler segment was missing an underscore:
"_asm__". Also we remove the "DEBUG" macro check, so we are consistent
with the linux and windows variant.
Now breaking into gdb on failure should work via:
gdb --args test_executable --break
Using sizeof(expr) can trigger a compile-time error,
"lambda expressions are not allowed in an unevaluated context", when passing
expression containing lambda, like a std algorithm. This error is considered
a standard defect, as it is meant to prevent lambdas in decltype
or templates, but not in sizeof.
This reverts commit 227598af47.
In reality, this is a relatively small performance improvement,
especially with the previous improvements removing lots of superfluous
string handling, but still was measurable.
This is sane, because those `const char*`s are given to us by compiler,
from the text area and thus we do not have to manage their lifetimes. We
also never want to change them.
Also moved copy constructor to compiler-generated methods, not sure why
it wasn't -- even before it was the same as a compiler would generate.
This means that all tabs used in indentation are now 4 spaces and that
there should be no more trailing whitespace.
Ill also look into creating a pre-commit hook that will prevent this
from happening in the future.
Fixes#105
Rewrote main wrapping loop. Now uses iterators instead of indices and intermediate strings.
Differentiates between chars to wrap before, after or instead of.
Doesn’t preserve trailing newlines.
Wraps or more characters.
Dropped support for using tab character as an indent setting control char.
Hopefully avoids all the undefined behaviour and other bugs of the previous implementation.
Don't duplicate Catch::isDebuggerActive() check many times, do it just once
in CATCH_BREAK_INTO_DEBUGGER() definition and use a separate CATCH_TRAP()
macro for the really platform-dependent part.
* Empty strings are now direct constructed as `std::string()`, not as empty string literals.
* `startsWith` and `endsWith` no longer construct new a string. This should be an improvement
for libstdc++ when using older standards, as it doesn't use SSO but COW and thus even short
strings are expensive to first create.
* Various places now use char literal instead of string literals containing single char.
** `startsWith` and `endsWith` now also have overload that takes single char.
Generally the performance improvements under VS2015 are small, as going from short string
to char is mostly meaningless because of SSO (Catch doesn't push string handling that hard)
and previous commit removed most string handling if tests pass, which is the expect case.
This fixes the case when we pass signal to previously registered
handler, and it needs to transform the signal into different one.
Still problematic: What if the signal handler we replaced does not
terminate the application? We can end up in a weird state and loop
forever.
Possible solution: Deregister our signal handlers, CALL the previous
signal handler explicitly and if control returns, abort. This would
however complicate our code quite a bit, as we would have to parse the
sigaction we delegate to, decide whether to use signal handler or signal
action, etc...
Only some "signals" are handled under Windows, because Windows does not
use signals per-se and the mechanics are different. For now, we handle
sigsegv, stack overflow, div-by-zero and sigill. We can also
meaningfully
add various floating point errors, but not sigterm and family, because
sigterm is not a structured exception under Windows.
There is also no catch-all, because that would also catch various
debugger-related exceptions, like EXCEPTION_BREAKPOINT.
Also stops Catch from assuming its the only signal user in the binary,
and makes it restore the signal handlers it has replaced. Same goes for
the signal stack.
The signal stack itself probably shouldn't be always reallocated for
fragmentation reasons, but that can be fixed later on.
Now if we detect C++11 compiler, or MSVC in version corresponding to VS2015,
we switch from using `std::random_shuffle` to `std::shuffle`.
`std::random_shuffle` was officially deprecated in C++14, and removed in C++17.
Also removed guarded inclusion of `<random>` header, as there was nothing
in the header that used it.
Catch passes ::tolower into std::transform with string iterators.
::tolower has the signature int(int), which triggers a stealth narrowing
warning inside std::transform, because transform calls
*_Dest = _Fn(*_First), which implicitly narrows an int to a char.
For this particular application the narrowing is fine, so explicitly
narrow in a lambda.
Catch passes an RNG which accepts int to random_shuffle. Inside
random_shuffle, the STL tries to call that RNG with the difference_type
of the user provided iterators. For std::vector, this is ptrdiff_t,
which on amd64 builds is wider than int. This triggers a narrowing
warning because the 64 bit difference is being truncated to 32 bits.
Note that this RNG implementation still does not produce a correctly
uniformly shuffled result -- it's currently asserting that std::rand
can produce 1000000 which is false -- but I don't know enough about
how much repeatable shuffles are necessary here, so I'm leaving that
alone for now.
SCENARIO does not add leading spaces to the test name (only BDD-style section
names are modified), so the trimming is not necessary. But if the name is
trimmed, it makes it harder to correlate the output of XML reporter with tests
that have leading spaces in their name: e.g. these tests will have the same name
attribute:
TEST_CASE("Test") {}
TEST_CASE(" Test") {}
