Added support for C++17 or later in tests
Pointers to member functions are now handled correctly in C++20
Details
- Reviewers
njames93 carlosgalvezp isuckatcs
Diff Detail
- Repository
- rG LLVM Github Monorepo
Event Timeline
clang-tools-extra/clang-tidy/utils/ExceptionAnalyzer.cpp | ||
---|---|---|
164–166 | Are you sure noexcept is stored in the type? The test case you modified throw_noexcept_catch_regular() tests this scenario and in that case the types seem to be the same even though one of the is noexcept an the other is not. If the FIXME is valid the proper way would be to implement it in this patch. | |
clang-tools-extra/test/clang-tidy/checkers/bugprone/exception-escape.cpp | ||
414 | The name of the function suggests that we throw a noexcept function pointer and catch a non-noexcept one. Please don't change it. |
clang-tools-extra/clang-tidy/utils/ExceptionAnalyzer.cpp | ||
---|---|---|
164–166 | Yes I'm sure. | |
clang-tools-extra/test/clang-tidy/checkers/bugprone/exception-escape.cpp | ||
414 | Ack. |
Update, added support for comparing function pointers without
exception specification.
Please cover the changes in as much test cases as possible.
clang-tools-extra/clang-tidy/utils/ExceptionAnalyzer.cpp | ||
---|---|---|
152 | The naming suggests that we check for equality, however in reality we check for convertability here. For example if one of the prototypes has exception spec and the other doesn't have one, this can still return true. Also instead of manually matching everything, you could take a look at ODRHash or StructuralEquivalenceContext. ODRHash for example doesn't take the FunctionProtoType into account when it generates the hashes AFAIK, so maybe you could use it to avoid these manual checks. I don't know if StructuralEquivalenceContext uses it or not, but you might consider taking a look at that too. Maybe you can also consider pasting the appropriate section of the standard here, so that whenever someone reads the code they'll understand why are we doing what we're doing. And please cover this function in positive and negative test cases too. | |
164–166 | I did some investigation regarding this and if I dump the type of void no_throw() noexcept both with -std=c++11 and -std=c++17 I get the same result. FunctionProtoType 'void (void) noexcept' cdecl `-BuiltinType 'void' I agree however that function pointer conversion was not properly implemented. | |
171–172 | I think we want to call this function for member function pointers too. The[[ https://github.com/cplusplus/draft/releases | C++ N4917 draft ]] says the following: 7.3.14 Function pointer conversions [conv.fctptr] A prvalue of type “pointer to noexcept function” can be converted to a prvalue of type “pointer to function”. The result is a pointer to the function. A prvalue of type “pointer to member of type noexcept function” can be converted to a prvalue of type “pointer to member of type function”. The result designates the member function. | |
281 | Please cover this line with both positive and negative test cases. Also upon looking up both N4849 (C++ 20 draft) and N4917 (C++ 23 draft), they both say for qualification conversion that
Why are we not allowing them if the standard is at least C++ 20? |
clang-tools-extra/clang-tidy/utils/ExceptionAnalyzer.cpp | ||
---|---|---|
281 | Please resolve this thread. |
Add one test.
clang-tools-extra/clang-tidy/utils/ExceptionAnalyzer.cpp | ||
---|---|---|
281 | Positive case is tested by throw_basefn_catch_derivedfn test and throw_derivedfn_catch_basefn. For members there are tests throw_basem_catch_basem, throw_basem_catch_derivedm, throw_derivedm_catch_basem that also covers this correctly. Tested this with GCC, and behavior is proper and independent to C++ standard. |
clang-tools-extra/clang-tidy/utils/ExceptionAnalyzer.cpp | ||
---|---|---|
281 |
I don't know how to deal with this tbh. In a static analyzer we usually want to consider what the standard says and not what a specific compiler implementation does, as the compiler can still be buggy, lack the proper support for the standard, etc. Maybe we can add a FIXME here that explains, why this check is here and that it's the opposite of what the standard says. Then later if it causes issues, it will be easy to see why that happens. @xazax.hun do you have a prefered method to deal with these situations in the analyzer? If there is one, we could also try it here. |
clang-tools-extra/clang-tidy/utils/ExceptionAnalyzer.cpp | ||
---|---|---|
281 | If you talk about about '7.3.5 Qualification conversions' then this is only about cv conversion, not about overall conversion from base to derived type, and this is verify by SatisfiesCVRules lambda. Without this line, check fails to detect issue in throw_basefn_catch_derivedfn function under C++20. Other interesting thing is that even this code: try { throw &baseMember::foo; } catch(const void(baseMember::*)()) { } Causes exception not to be catch, but if you remove const then it's catch. Right now both GCC and Clang in C++20 and C++17 work in same way, and this `if` should be here so check would be align with how compiler works. |
clang-tools-extra/clang-tidy/utils/ExceptionAnalyzer.cpp | ||
---|---|---|
281 |
Then how would you reason about this snippet? MSVC detects that the exception won't be caught and reports a warning, while gcc and clang fails to do it. In this case different compilers work differently. To be fair, the exception should be caught, since the thrown object is convertible to the handler (which is probably why gcc and clang don't report a warning).
Then do a proper investigation why that happens instead of adding random lines, which contradict what should happen, so that the test can pass. Probably the condition is wrong here and we should return false if !isSameP_i even if the standard is C++20 except for some cases with arrays. Look up the C++17 and the C++20 standards and compare them. |
The naming suggests that we check for equality, however in reality we check for convertability here. For example if one of the prototypes has exception spec and the other doesn't have one, this can still return true.
Also instead of manually matching everything, you could take a look at ODRHash or StructuralEquivalenceContext. ODRHash for example doesn't take the FunctionProtoType into account when it generates the hashes AFAIK, so maybe you could use it to avoid these manual checks. I don't know if StructuralEquivalenceContext uses it or not, but you might consider taking a look at that too.
Maybe you can also consider pasting the appropriate section of the standard here, so that whenever someone reads the code they'll understand why are we doing what we're doing.
And please cover this function in positive and negative test cases too.