[jmgilbert2]

Cleans up the warnings that GCC 10 to 15 report with -Wall -Wextra on EASTL's library and test sources. Several of them were real bugs. Addresses part of upstream electronicarts#260 and fixes upstream electronicarts#421 and PR electronicarts#595.

How it was checked

• GCC: compile-only sweeps on GitHub Actions (the gcc-probe* scratch branches), GCC 10 to 15. Each covered library sources, test sources and repros, in six builds: -O0 Debug, -O1, -O2, -O3, -O2 with the April 2025 deprecations disabled, and -O2 -std=c++17.
  • EASTL's own warning sites went from 73 to the few listed under "Not fixed here".
  • Every test-code warning is gone.
  • Nothing new appeared in any build.
• Windows: EASTL's test suite on MSVC and clang-cl × C++17/latest × Debug/Release. 8/8 passed, and the C++17 and latest outputs are identical.

Real bugs fixed

• tuple_vector: variadicAnd(bool cond, bool conds...) was a C varargs function, so the iterator validation ignored every element type after the second. It's now a variadic template.
• DecodePart (DecodePart pointer type punning breaks strict aliasing electronicarts/EASTL#421): the forwarding overloads accessed the callers' pointer variables through references to other pointer types, which breaks strict aliasing. They now convert the pointer values and write them back.
• segmented_vector: begin() const and end() const returned an iterator reinterpreted as a const_iterator, and end() left two members uninitialized.
• SetAssertionFailureFunction (PR Forward pContext to SetAssertionFailureFunction electronicarts/EASTL#595): the legacy overload dropped pContext.
• Tests:
  • TestIntrusiveSDList printed c_str() of a destroyed temporary.
  • A precedence bug in TestSegmentedVector meant one check didn't test the values.

electronicarts#260: fixed_vector without overflow

EASTL_CRASH() now traps after its invalid write on GCC and clang, so the compiler knows the exhausted allocator never returns null. That clears the -Wnonnull warnings from fixed_string and vector growth. The assign()-past-capacity path stays (see below).

Warnings on correct code

• Code changes:
  • a void* cast for -Wclass-memaccess;
  • braces for -Wempty-body;
  • EA_MAYBE_UNUSED parameters;
  • a named extract_key instead of a temporary, which GCC 13 took for a dangling reference;
  • an always-false trait instead of static_assert(I != I);
  • literal operators without the space that C++23 deprecates.
• Suppressions: narrow, version-guarded suppressions for GCC false positives in inlined paths. Each covers one function or statement and says why the code is correct:
  • Compare;
  • CharStringUninitializedCopy;
  • shared_array::unique;
  • radix_sort's copy-back;
  • basic_string::erase;
  • fixed_vector_allocator::deallocate;
  • BitsetBase::set;
  • the two memmove helpers that vector::insert reaches.

Not fixed here

• Left on purpose, because the only fixes cost a branch or change semantics:
  • GCC 10 to 15 report -Wnonnull at memory.h:319 when a fixed_vector without overflow is assigned past its capacity. Avoiding it would add a branch to every reallocation.
  • GCC 10 to 15 report -Wuninitialized/-Wmaybe-uninitialized in eastl::swap (utility.h:56) from string's Layout swap. Layout's move operations must keep swapping.
• expected's constructors set mHasValue before constructing the value or error in the constructor body. If that throws, the base destructor destroys an object that was never constructed. GCC reports -Wmaybe-uninitialized on those exception paths in TestExpected (vector.h), plus a GCC 10-only one in the value-to-error assignment (smart_ptr.h:167). optional's copy/move path has the same order. Left alone: our builds disable exceptions.
• smart_ptr_deleter<void> frees with delete[] (char*)p, which mismatches TestSmartPtr's scoped_ptr<void>(new int) (GCC: -Wmismatched-new-delete). Left alone by decision.
• With the April 2025 deprecations disabled (legacy optional), GCC 13 to 15 report -Wmaybe-uninitialized in sort.h and optional.h for optional<int>.
• EABase's typedef char char8_t; gives -Wc++20-compat in C++17 builds on GCC 13+; that's for EABase.
• GCC 11 ignores the suppression pragmas for some inlined code, even at the warning's own location; GCC 12+ honors them. Two of the false positives above stay on GCC 11 only:
  • Compare (-Wstringop-overread);
  • the fixed_vector deallocation, at -O3 (-Wfree-nonheap-object).
• Only EASTL's sources, and the headers of the other packages that its tests include, were compiled with GCC; the other EA repos weren't.

🤖 Generated with Claude Code