Workaround for OSS-Fuzz std::format detection

[kmilos]

No description provided.

[neheb]

Why not test for std::format directly?

[kmilos]

From the CheckCXXSymbolExists docs:

If the symbol is a type, enum value, or C++ template it will not be recognized: consider using the CheckTypeSize or CheckSourceCompiles module instead.

It looks like it might not always work w/ non-templated std:vformat either:

This command is unreliable when <symbol> is (potentially) an overloaded function.

Switching to CheckSourceCompiles would be just as verbose, if not more effort...

This way the check/workaround is the same in CMake and source, and IMHO easily identifiable and removable in the future when OSS-Fuzz image is updated.

In any case, let's first confirm this approach works w/ -DCMAKE_CXX_STANDARD=20? @kevinbackhouse

[neheb]

@@ -56,7 +56,7 @@ deps += cpp.find_library('ws2_32', required: host_machine.system() == 'windows')
 deps += cpp.find_library('psapi', required: host_machine.system() == 'windows')
 deps += cpp.find_library('procstat', required: host_machine.system() == 'freebsd')
 
-if not cpp.has_header_symbol('format', '__cpp_lib_format')
+if not cpp.has_header_symbol('format', 'std::format')
   deps += dependency('fmt')
 endif
 

on meson works. It's actually shorter...

[kevinbackhouse]

Wow! Is this the answer to the OSS-Fuzz build failure?

[neheb]

Yeah even

env CXX=clang++ CXXFLAGS=-stdlib=libc++ meson -Dcpp_std=c++20 a

succeeds.

edit: oh but it fails to compile as the code ends up using fmt.

[kevinbackhouse]

I'm testing it now.

[neheb]

@kmilos what do you think of this alternate solution?

neheb@3f7fbda

[kevinbackhouse]

It works! Awesome work @kmilos!

[kmilos]

Sure, the other proposal looks good. Just wasn't sure it'd work going by CMake docs...

[kevinbackhouse]

@neheb's: is your solution a permanent fix? I.e. we won't need to remember to revert it at some stage in the future when OSS-Fuzz gets fixed? If so, that sounds even better.

[neheb]

oh this is hilarious. So it works under meson but not cmake:

 error: address of overloaded function 'forma    t' does not match required type 'int'
 35             7 |   return ((int*)(&std::format))[argc];
 34               |                   ^~~~~~~~~~~
 33         /usr/bin/../include/c++/v1/__format/format_functions.h:476:1: note: candidate template ignored: could not match 'string (format_string<_Args...>,     _Args &&...)' (aka 'basic_string<char> (basic_format_string<char, type_identity_t<type-parameter-0-0>...>, _Args &&...)') against 'int'
 32           476 | format(format_string<_Args...> __fmt, _Args&&... __args) {
 31               | ^
 30         /usr/bin/../include/c++/v1/__format/format_functions.h:483:1: note: candidate template ignored: could not match 'wstring (wformat_string<_Args...>    , _Args &&...)' (aka 'basic_string<wchar_t> (basic_format_string<wchar_t, type_identity_t<type-parameter-0-0>...>, _Args &&...)') against 'int'
 29           483 | format(wformat_string<_Args...> __fmt, _Args&&... __args) {
 28               | ^
 27         /usr/bin/../include/c++/v1/__format/format_functions.h:609:1: note: candidate template ignored: could not match 'string (locale, format_string<_Ar    gs...>, _Args &&...)' (aka 'basic_string<char> (std::locale, basic_format_string<char, type_identity_t<type-parameter-0-0>...>, _Args &&...)') against 'in    t'
 26           609 | format(locale __loc, format_string<_Args...> __fmt, _Args&&... __args) {
 25               | ^
 24         /usr/bin/../include/c++/v1/__format/format_functions.h:616:1: note: candidate template ignored: could not match 'wstring (locale, wformat_string<_    Args...>, _Args &&...)' (aka 'basic_string<wchar_t> (std::locale, basic_format_string<wchar_t, type_identity_t<type-parameter-0-0>...>, _Args &&...)') aga    inst 'int'
 23           616 | format(locale __loc, wformat_string<_Args...> __fmt, _Args&&... __args) {
 22               | ^
 21         1 error generated.
 20         ninja: build stopped: subcommand failed.

That has to be a joke.

[kevinbackhouse]

I've squashed the commits on my testing branch so that it's almost ready to submit: kevinbackhouse/oss-fuzz#12

[neheb]

OK. Finally fixed CMake: https://github.com/neheb/exiv2/commit/d8867845918cc2e7e75e9ee9bbfd57994db73c00.patch

My solution is more generic. I can't say I like matching against compilers.

[kevinbackhouse]

For future reference this is the link to the discussions that we had about the OSS-Fuzz build failure (which started after #2769 was merged): #3108

I also asked for help on the OSS-Fuzz repo but nobody has replied so far.

[kmilos]

I'll close this PR then in favour of the other, more generic solution.

[kevinbackhouse]

I'm testing @neheb's version now

[neheb]

Looks like my version was missing

--- a/src/image_int.cpp
+++ b/src/image_int.cpp
@@ -1,5 +1,6 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
+#include "config.h"
 #include "image_int.hpp"
 
 #include <cstdarg>

Added and force pushed.

[neheb]

For reference, I'll note that you cannot mix and match libc or libc++. OSS-Fuzz builds with -stdlib=libc++ (LLVM) when the OS uses libstdc++ (GNU). That's the whole reason for the errors with fmt.

This also happens with system INIReader, which doesn't seem to be used.

[kmilos]

you cannot mix and match libc or libc++. OSS-Fuzz builds with -stdlib=libc++ (LLVM) when the OS uses libstdc++ (GNU)

... when trying to link static libs.