CodeQL C++ Extractor — GCC 12.2.0 + Recent LLVM libc++ Extraction Failures #22514

Summary

The CodeQL C++ extractor (v2.26.0, --mimic mode, --gnu_version 120200) produces parse errors in nearly all translation units when processing a codebase compiled with GCC 12.2.0 paired with a recent LLVM libc++ standard library. The C++ compiler is q++ (QNX SDP 8.0.0), which combines a GCC 12.2.0 frontend with LLVM libc++ headers. All errors are in libc++ system headers (std::__2:: namespace) with no errors in user code. The real q++ compiler compiles every translation unit without issue.

Switching the same codebase to libstdc++ (with the same GCC 12.2.0 version) eliminates all errors.

Environment

Symptoms

Errors span multiple libc++ headers and all follow the same pattern: the extractor's internal C++ frontend fails template deduction or SFINAE substitution where GCC 12.2.0 succeeds. Examples:

error: class "std::__2::iterator_traits<std::__2::__wrap_iter<char *>>"
  has no member "iterator_category"

error: no instance of overloaded function "std::__2::__search_impl"
  matches the argument list
candidate function template failed deduction

error: no instance of constructor
  "std::__2::pair<_T1, _T2>::pair [...]" matches the argument list

Reproduction

The real compiler succeeds; the extractor fails:

EXTRACTOR="$CODEQL_BUNDLE/cpp/tools/linux64/extractor"
COMPILER="/path/to/gcc-12.2.0-based-compiler"

cat > test.cpp << 'EOF'
#include <algorithm>
#include <vector>
#include <string>
void f() {
    std::vector<char> v(10);
    std::fill(v.begin(), v.end(), 0);
    std::string s(10, ' ');
    std::fill(s.begin(), s.end(), 'x');
}
EOF

"$COMPILER" -std=c++17 -c test.cpp -o /dev/null       # exit 0
"$EXTRACTOR" --mimic "$COMPILER" --object_filename /dev/null \
  --error_limit 1000 -std=c++17 -c test.cpp -o /dev/null 2>&1 | grep "error:"

Root Cause

The extractor's internal C++ frontend does not fully replicate GCC 12.2.0's template deduction and SFINAE substitution behavior when processing recent LLVM libc++ headers. This manifests across multiple libc++ internals — iterator traits (__wrap_iter), algorithm SFINAE guards (__enable_if_t), and conditional constructors (pair, optional) — but the underlying cause is the same: the extractor frontend reaches different SFINAE outcomes than GCC 12.2.0.

What works

• GCC 12.2.0 + libstdc++: 0 extraction errors
• GCC 11.4 + LLVM 14 libc++: 0 extraction errors (older libc++, different SFINAE patterns)

codeql-extraction-logs.zip