[clang] [windows] [address-sanitizer] stack overflow when test is failing

[hia3]

Description

Catch enters infinite recursion trying to print error message when compiled with clang -faddress-sanitize under Windows

Steps to reproduce

0. Install VS 2015 and clang
1. Put catch.hpp and test.bat from test.zip in one directory, start test.bat
2. Resulting executable, instead of reporting an error, will output:

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

and then it will crash.

[refi64]

FWIW this works normally under Linux.

What if you use normal Clang:

clang++ -fsanitize=address -o test.exe test.cpp

Also, you seem to be building a normal, shared executable. What if you change the library name to clang_rt.asan_dynamic-x86_64.lib? (The one you're using is for applications using the static CRT.)

[hia3]

Compiling with

clang++ -fsanitize=address -o test.exe test.cpp

gives me

==748==AddressSanitizer CHECK failed: D:\src\llvm_package_300231\llvm\projects\compiler-rt\lib\asan\asan_rtl.cc:505 "((!asan_init_is_running && "ASan init calls itself!")) != (0)" (0x0, 0x0)
==748==AddressSanitizer CHECK failed: D:\src\llvm_package_300231\llvm\projects\compiler-rt\lib\asan\asan_poisoning.cc:37 "((AddrIsInMem(addr))) != (0)" (0x0, 0x0)

even if 'test.cpp' is just an empty main. This is why I've added /DEBUG when linking. This has nothing to do with Catch, must be clang bug.

With clang_rt.asan_dynamic-x86_64.lib linking fails:

test.obj : error LNK2019: unresolved external symbol __asan_shadow_memory_dynamic_address referenced in function "struct Catch::IResultCapture & __cdecl Catch::getResultCapture(void)" (?getResultCapture@Catch@@YAAEAUIResultCapture@1@XZ)
test.obj : error LNK2019: unresolved external symbol __asan_option_detect_stack_use_after_return referenced in function "struct Catch::IResultCapture & __cdecl Catch::getResultCapture(void)" (?getResultCapture@Catch@@YAAEAUIResultCapture@1@XZ)
test.exe : fatal error LNK1120: 2 unresolved externals

[horenmar]

Is ASan supported under Windows? Last time I checked, it was "this will very likely work for C, C++ is very much beta".

[ekilmer]

Hello. Now that Address Sanitizer is Generally Available for MSVC, I've noticed that I'm unable to run my tests due to this issue--they get stuck and won't complete.

After some digging, I found a relevant open issue on Microsoft's tracker with a potential workaround that would consist of the following (naive) change to Catch here

Catch2/include/internal/catch_fatal_condition.cpp

Line 64 in b025a00

exceptionHandlerHandle = AddVectoredExceptionHandler(1, handleVectoredException);

and here for the single include header:

Catch2/single_include/catch2/catch.hpp

Line 10787 in b025a00

exceptionHandlerHandle = AddVectoredExceptionHandler(1, handleVectoredException);

- exceptionHandlerHandle = AddVectoredExceptionHandler(1, handleVectoredException);
+ exceptionHandlerHandle = AddVectoredExceptionHandler(0, handleVectoredException);

I've tested this change locally with a modified (read: updated) version of the batch script in the OP with a fully updated Visual Studio (v16.9.1) on latest Catch (v2.13.4), and it seems to run to completion!

echo #define CATCH_CONFIG_MAIN > test.cpp
echo #include "catch.hpp" >> test.cpp
echo TEST_CASE( "t", "[t]" ) { REQUIRE( 2 == 1 ); } >> test.cpp

call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Auxiliary\Build\vcvarsall.bat" amd64
echo Please wait...
cl /EHsc test.cpp -fsanitize=address -c
link /DEBUG test.obj

start test.exe

However, I don't understand the consequences of this change and how it could affect the rest of Catch's logic and exception handling, so hopefully this helps one of the maintainers or someone else more familiar with the code base and C++ in general 🙂

[reder2000]

This is related to the way asan is implemented on x64 msvc. See here doc.
Catch2 is catching tons of 0xc0000005 (FCE - first chance exceptions) and gets confused.
Your fix works but gives up on FCEs.

catch2+asan+msvc is working on x86 unmodified (but nobody builds for x86 anymore)

Maybe catch2 could offer an option to give up on first chance exception or do it automatically on x64 with

... ```

Edit: with catch2 moving as a library, this way to fix depending on asan would not work. 

[unixwitch]

as a workaround, setting CATCH_CONFIG_NO_WINDOWS_SEH fixed this issue for me (tested with 2.13.6).