pxx compiler hangs (non-termination) on two valid C programs

• Type: bug (compiler non-termination). Worst class — not a wrong answer, no output at all: the compiler never returns. A hang in a gate wedges the whole run.
• Track: A (front/middle-end — lowering or an optimization pass that doesn't converge). Both inputs are valid C that gcc compiles in well under a second.
• Found by: the one-time gcc c-torture harvest (2026-07-15), [[feature-t-gcc-torture-runner]]. Surfaced because pxx spun 35 min at 100% CPU on the first one and stalled the harvest until a compile timeout was added.

Repro

timeout 5 compiler/pascal26 -Ilib/crtl/include -Ilib/crtl/src \
  library_candidates/gcc-torture/execute/pr44164.c /tmp/x   # exit 124 = HANG (24 lines)
timeout 5 compiler/pascal26 -Ilib/crtl/include -Ilib/crtl/src \
  library_candidates/gcc-torture/execute/pr23324.c /tmp/x   # exit 124 = HANG (133 lines)

Both: pxx exit=124 (killed by timeout, running at 100% CPU); gcc -w -c <f> compiles each in <1s. No diagnostic, no partial output — pure non-termination.

Leads (reduction started, NOT isolated — that's the fix)

• pr44164.c (24 lines, the tractable one): nested structs X{Y{YY{Z{int}}}} + an empty compound literal a.b = (struct Y){} assigned to a nested field, inside a __attribute__((noinline,noclone)) function that reads p->i, clobbers *p via the aggregate store, then reads p->i again (the gcc PR44164 store-forwarding/aliasing shape). Minimal reductions of JUST the empty compound literal on a nested struct do NOT hang (tested struct Y {}, (struct Y){} one/two levels deep — all terminate), so the trigger needs the aliasing/self-referential-store context or an opt pass that fails to converge on it. Start there.
• pr23324.c (133 lines): many odd-width signed bitfields (:2 :3 :5 :6 :7 :9 :10 :12) across nested structs + an empty union union at6 {} + long long. May be a distinct root cause (bitfield layout) or the same non-convergence; also referenced in [[bug-c-bitfield-promotion-and-layout-cluster]] but centralized here as a hang.

Why it matters beyond these two files

A compiler that can non-terminate on valid input is a latent hazard for every corpus and the self-host gate. The watcher now guards against a stall (compile timeout in the harvest; testmgr's per-job timeout catches it in tiers), but the underlying non-termination should be fixed, not just timed out around.

Acceptance

Both files compile in bounded time under pxx (match gcc's sub-second); a minimal reduced repro is added as a test/ regression; if the two have distinct root causes, split into two fixes but keep this ticket until both terminate.

Log

• 2026-07-15 — resolved, commit PENDING.