String literal operand is not GC-rooted across an allocating call in the same expression (stale handle after evacuation)

[proggeramlug]

Summary

A string-literal handle is loaded before an allocating call in the same expression and reused after it. The handle global is a GC root that evacuation rewrites; the register holding the pre-call value is stale. Under enough allocation the concatenation silently produces a truncated or empty string.

This is exactly the hazard crates/perry-codegen/src/expr/temp_root.rs (#6951) was built to close — "an already-evaluated operand waiting for its sibling ... the slot is a mutable root, so an evacuating cycle rewrites it and the register pushed beforehand is stale" — with the string-literal-handle load not covered by it.

Silent wrong output, exit code 0. No crash, no diagnostic.

Repro

class Rec {
  id: number; score: number;
  constructor(id: number, score: number) { this.id = id; this.score = score; }
}
function make(i: number): Rec { const r = new Rec(i, 0); r.score = r.id * 1.5; return r; }
function run(n: number): number {
  let acc = 0;
  for (let i = 0; i < n; i++) { const r = make(i); acc = acc + r.score; }
  return acc;
}

console.log("acc:" + run(10000000));   // literal evaluated BEFORE the allocating call
const v = run(10000000);
console.log("hoisted:" + v);           // same concat, call hoisted out

$ node --experimental-strip-types repro.ts     # v26.5.1, the .node-version pin
acc:74999992500000
hoisted:74999992500000

$ perry compile repro.ts -o repro && ./repro
                                                <- EMPTY LINE
hoisted:74999992500000
$ echo $?
0

Only the first statement is wrong. Hoisting the call out of the concatenation fixes it, which localises the bug to operand evaluation order rather than to run, to js_string_concat_value, or to the arithmetic (the number is correct in both).

The corruption is allocation-count dependent and its manifestation varies, which is the signature of reading through a stale/reused heap slot rather than of a logic error:

iterations	observed
1,000	acc:1249000 (correct)
100,000	acc:12499900000 (correct)
1,000,000	1249999000000 — prefix gone
10,000,000	S124999990000000 / r1249999000000 / empty — prefix replaced by garbage

Root cause, from the emitted IR

--trace llvm, @main, the two statements side by side:

; BROKEN — literal handle read first, used after the allocating call
%r1 = load double, ptr @minrepro2_ts_.str.2.handle            ; <-- read BEFORE
%r2 = call double @perry_fn_minrepro2_ts__run__spec_i32(i32 10000000)   ; 10M allocs -> evacuating GC
%r3 = bitcast double %r1 to i64                               ; <-- STALE
%r4 = and i64 %r3, 281474976710655
%r5 = call i64 @js_string_concat_value(i64 %r4, double %r2)

; CORRECT — call is a separate statement, so the handle is read after it
%r15 = call double @perry_fn_minrepro2_ts__run__spec_i32(i32 10000000)
store double %r15, ptr %r14
%r16 = load double, ptr @minrepro2_ts_.str.3.handle           ; <-- read AFTER

There is no js_gc_temp_root_push protecting %r1 across the call. The temp_root triple is emitted in this same function — but only later, around the console.log spread array (%r8/%r10), not around the literal operand.

Two possible fixes, both already-established patterns in the codebase:

• push the loaded literal with temp_root_push_i64 before the sibling operand is evaluated and re-read it with temp_root_get_i64 afterwards (the documented contract in temp_root.rs), or
• sink the load of the .handle global below the call so it is re-read after any collection point.

Scope

• Reproduces on main (3f1a0853a = e2557c1a9 + test(repsel): promotion census with a ratcheted, falsifiable floor #7104) and on the perf(repsel): return-shape facts — Ptr<Shape> survives the return escape (#7034 §4) #7107 branch.
• Independent of representation selection: identical with PERRY_PTR_SHAPE_LOCALS=0 and on the pre-perf(repsel): return-shape facts — Ptr<Shape> survives the return escape (#7034 §4) #7107 compiler. It is not a repsel defect.
• Default GC settings — no PERRY_GC_FORCE_EVACUATE, no PERRY_CONSERVATIVE_STACK_SCAN=off. temp_root.rs's header notes conservative stack scanning hid this class of bug; here it does not, presumably because the stale value is a rewritten root rather than a freed object.
• Trigger is a binary + whose left operand is a string literal and whose right operand is a call that allocates enough to drive an evacuating cycle. "a" + f() is an extremely common shape, so the exposure is likely wider than this one expression form.

How it was found

Fell out of building a hot-path workload for the representation-selection A/B on #7107 (console.log("acc:" + run(N)) at N = 10⁷). Filing separately so it does not ride on a perf PR.