[PATCH] D90851: [clang-tidy] Extending bugprone-signal-handler with POSIX functions.

Wed Feb 3 16:25:15 PST 2021

alexfh added inline comments.

================
Comment at: clang-tools-extra/test/clang-tidy/checkers/Inputs/Headers/system-header-posix-api.h:1
-//===--- signal.h - Stub header for tests -----------------------*- C++ -*-===//
+//===--- system-header-posix-api.h - Stub header for tests ------*- C++ -*-===//
 //
----------------
aaron.ballman wrote:
> balazske wrote:
> > aaron.ballman wrote:
> > > balazske wrote:
> > > > aaron.ballman wrote:
> > > > > I think we should strive to replicate the system headers rather than fake up a system header (these headers can be used by more than one check). So I think we want to keep signal.h and stdlib.h, and should include the POSIX-specific functionality with a macro. This also helps us test the behavior on systems like Windows which are not POSIX systems.
> > > > My concern was to add these many small header files with just some functions in them. For `accept` more than one header is needed if we want to exactly replicate the system files. And data types like `size_t` should have a common header too. So I decided to have one header that contains all system functions and data types. This can be used by multiple tests and extended as needed.
> > > I don't think we're too worried about having a bunch of small test headers around -- I think it's more important the headers used to check system header behavior be understandable as to what you're getting from them. For instance, there are clang-tidy checks for llvm-libc that may have very different needs from what you're doing here.
> > > 
> > > What's more, these changes break existing tests -- I don't see any companion changes to fix those up.
> > On my system (Ubuntu) I do not see failing tests, and these changes do not touch files that were created before D87449, no new problems should happen. I am not against "mirroring" the POSIX API header structure into the test stub header structure, only do not like the overhead of adding these many files with just 1-2 lines (that is needed for this test) in them. It is not done this way in the clang tests either (there are multiple "system-header-simulator" files that contain every declaration usable for specific purposes at one or more tests). Also I want to have the opinion of another reviewer for this question.
> > On my system (Ubuntu) I do not see failing tests, and these changes do not touch files that were created before D87449, no new problems should happen.
> 
> You renamed stdlib.h and `lllvmlibc-restrict-system-libc-headers.cpp` includes `stdlib.h`, but after closer inspection, I see now that it's including one from a different directory. So this doesn't break the things I thought it was breaking, good!
> 
> > Also I want to have the opinion of another reviewer for this question.
> 
> I'm happy to go with whatever @alexfh thinks.
> 
Based on the experience with our internal set of checks and tests for them, I tend to think that keeping mock API definitions arranged in a similar way to the corresponding real library is a more sustainable way to manage mock headers. When arranged this way, it's easier to find the origins of, verify, and update definitions, when, for example, a higher fidelity replication of a system/STL/... entity becomes necessary. Dependencies between these mock headers (if necessary) can repeat those in the real library, and #includes of these mock headers can be composed in a similar way. Though the structure of the mock headers can start be approximate and the level of detail can be added when necessary. For example, one can start with a mock header for <string> that would contain initializer_list implementation. But when another test would need <initializer_list>, the corresponding entities can be moved to a separate initializer_list header.

As for different header contents depending on target platform, preprocessor macros seem to be a better way to handle this than using separate mock headers.

The set of mock headers can be shared by all tests that examine the corresponding API. I don't see good reasons to keep different mocks of the same API for different tests.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D90851/new/

https://reviews.llvm.org/D90851