[AI] Bunch of bugs found by autonomous scanner

[qarmin]

I’m developing a project locally that verifies whether a given CPU correctly handles AVX/SSE and other SIMD operations, and it contains a table of reference results generated on my own CPU (hopefully without any errors), which can then be used to validate the results obtained on other hardware.

I thought that testing against these reference results might be useful for this project, as it could help increase the test coverage.

With the help of AI, I automatically generated a list of potential issues from such tool, along with test code, bug explanations and possible solutions, and I hope that this can be useful and help with the project - rustc_codegen_gcc_bughunt_20260902_142949.zip

#	Class	Candidate	Fix lines	Root cause	Introduced by
001	crash	back/write.rs fat-LTO codegen path SIGSEGVs instead of erroring when liblto_plugin.so is missing	n/a	-fuse-linker-plugin added unconditionally for fat LTO with no check the plugin is available; GCC's driver fails cleanly but rustc then SIGSEGVs inside libgccjit	c92f054
004	logic	builder.rs atomic_extremum CAS-loop comparison has no signed/unsigned variant	+18/-6	atomic_rmw routes both signed and unsigned fetch_max/fetch_min through one comparison typed by atomic_load's fixed builtin return type, so unsigned compares silently run as signed across the sign bit	afae271
005	memory-safety	builder.rs shuffle_vector mask-padding branch reads v2 past its length	+18/-9	broadcasting a short vector to more than double its lanes reads v2 out of bounds in the padding loop; harmless since the mask never selects the overrun slots, but a genuine UBSan-flagged OOB read	faebf73
006	crash	type_of.rs type_uint_from_ty returns signed GCC types for unsigned Rust types	+6/-6	every UintTy variant mapped to the signed GCC type instead of unsigned, so simd_scatter's pointer-cast target mismatches its stored value's sign and libgccjit hard-rejects the assignment at compile time	fac57d9
008	memory-safety	intrinsic/simd.rs gather() touches every SIMD lane's pointer even when masked off	+51/-36	gather() dereferenced/wrote through every lane's pointer unconditionally before masking the result via shuffle_vector, so a disabled lane's null/dangling pointer got touched anyway in simd_gather and simd_scatter	not determined
010	logic	int.rs u128-to-f32 cast via __floatuntisf gets miscompiled by GCC's optimizer under a black_box barrier	n/a	u128 as f32 computes the correct runtime result via the extern __floatuntisf call, but GCC's optimizer (-O, opt-level>=1) statically folds away the assert_eq check against f32::INFINITY - an upstream GCC optimizer bug, not a backend cast-lowering bug	not determined
018	logic	builder.rs lshr/ashr share gcc_lshr's native >> with no sign-forcing cast	+26/-4	a value whose GCC-declared type ends up wrongly signed (e.g. via ptrtoint bridging a pointer-to-usize transmute through a signed isize) gets an arithmetic shift where a logical one (or vice versa) was required	afae271
025	logic	intrinsic/simd.rs simd_saturating_sub negates rhs to reuse the saturating-add algorithm	+14/-11	negating the element type's minimum value (e.g. i8::MIN) overflows and wraps back to itself instead of its true negation, corrupting the sum the borrowed saturating-add overflow check is based on, so a MIN-valued rhs lane returns a wrong, non-saturated result	d725cfb
026	logic	int.rs int_to_float_cast native branch casts using the operand's own declared sign instead of the caller-requested one	+12/-0	uitofp/sitofp share a native-type cast with no sign-forcing cast, so a value whose GCC type ends up wrongly signed (e.g. via a pointer-to-usize transmute) converts to float using the wrong sign interpretation	41f20fa
030	logic	std::arch::x86_64 _mm_subs_epi8/_mm_subs_epi16 return wrong values for i8::MIN/i16::MIN operands	+14/-11 (candidate 025's patch)	the pinned toolchain's stdarch implements these intrinsics via simd_saturating_sub, so this is candidate 025's negate-of-MIN saturating-sub bug reached through a different Rust API surface, not a separate libgccjit-builtin dispatch bug as originally hypothesized	d725cfb
031	crash	intrinsic/llvm.rs intrinsic() missing explicit mappings for 5 of 6 rdrand/rdseed builtins	+25/-2	only llvm.x86.rdrand.64 had an explicit override + out-parameter repacking; its five siblings fell through to archs.rs's generated unimplemented!() catch-all and panicked instead of compiling	68ac3a4
032	crash	builder.rs atomic lowering always emits a _atomic*_16 libcall with no -latomic link flag	+395/-1	every atomic op (including 16-byte cmpxchg/RMW) unconditionally calls the sized _atomic*_{size} GCC builtin instead of inlining cmpxchg16b when the feature is enabled, and the backend never adds -latomic at link time, leaving the libcall reference dangling	afae271

[bjorn3]

builder.rs atomic lowering always emits a __atomic_*_16 libcall with no -latomic link flag

Atomics are not exposed on targets where libatomic is necessary because Rust guarantees that atomic operations are lock free. libatomic is not guaranteed to be lock free and thus never used. I presume the __atomic_*_16 libcalls are emitted by cg_gcc because that is the only way to tell GCC to emit atomic operations. On all targets where Rust supports atomics in the first place, this should always lower to atomic instructions and never result in calls to libatomic.

[antoyo]

On all targets where Rust supports atomics in the first place, this should always lower to atomic instructions and never result in calls to libatomic.

I don't think this is always the case for GCC: I've seen some cases on Aarch64 where LSE/LSE2 is not enabled lower to a libatomic call and enabling LSE does not seem to allow GCC to drop that call.

[bjorn3]

The outlined atomics for when LSE is not known to be supported should be the __sync_* functions which are guaranteed to be lock free, not the __atomic_* functions which may use a mutex internally. The __sync_* functions are provided by libgcc, compiler-rt and compiler_builtins rather than by libatomic.

[antoyo]

Can we specify the ordering with __sync_* functions?

[bjorn3]

Looks like I misremembered what functions -moutline-atomics produces on arm64. It uses __aarch64_* functions: https://rust.godbolt.org/z/fnvhv4sn1