← board

origin/master has advanced 4 commit(s) since this sha. Re-verify at current HEAD before acting — the callback is tagged to the sha that was tested, which may no longer be the state of the tree.

regression: test-core#src:test/test_local_typed_const.pas red at 88b863e7c731 (auto-filed by twatch)

Repro

tools/testmgr.py --tier native --job 'test-core#src:test/test_local_typed_const.pas' at 88b863e7c7311d5f1da708ba3ce7bcfe172e09b3

Range

bad 88b863e7c731, last good 6891a4d56494, 2 commit(s) in range — the watcher narrows this by idle bisect; check tstate/TSTATE.md for the current range.

Log tail

pascal26:41: error: invalid IR symbol reference in load_sym
(tail)
pascal26:41: error: invalid IR symbol reference in load_sym
  near:  WriteLn  SumTable   >>> end  unit

Stub ticket: signal only. Track T agent (face 2) enriches or a dev track takes it from the repro line.

RESOLVED 2026-08-16 (Track P) — mine, fixed forward in 467a4e5da

Caused by 3ed3e2653 (routine-local typed consts made static). Two faults, the second only reachable once the first was fixed:

  1. SymRollbackTo reclaimed the const's symbol index when the routine ended, and the -O2 inliner then copied the body into the caller, where the copy still loads it — verified after the rollback, one past the end. That is the invalid IR symbol reference in load_sym this ticket recorded. -O0/-O1 were clean, which named the inliner.
  2. Fixing the index exposed a silent wrong value at -O2 (Scaled(6) gave 0, not 42): the run-once prologue guard is skipped by an inlined copy, so the initialisation never ran.

SymRollbackTo now keeps its high-water mark above a SymStaticLocal symbol (still unhashing it), and the initialiser is an ordinary PendingInit row emitted once before main begin — no prologue for the inliner to skip.

Verified: this test and test_local_typed_const_is_static.pas both pass at -O0/-O1/-O2/-O3 and match fpc -O- -Mobjfpc. Full write-up in [[bug-p-a-routine-local-typed-const-is-reinitialised-on-every-call]].

Triaged and RESOLVED 2026-08-16 by Track T (face 2)

A true regression, not a flake, and bisected rather than guessed.

3ed3e2653 fix(P): a routine-local typed const is static, initialised once is the commit. Built its parent and ran the test:

3ed3e2653~1   compiles, prints  100 a b c 42 100     <- correct
3ed3e2653..   pascal26:41: error: invalid IR symbol reference in load_sym

What the fix got right, and what it broke

Worth separating, because the fix was not simply wrong — it achieved its target and regressed a neighbour:

shape after 3ed3e2653
const calls: Integer = 0; Inc(calls) — the counter idiom it was FOR fixed: 1 2 3, was 1 1 1
const T: array[0..3] of Integer = (...) read by index still fine
const W: array[1..3] of Char read by index still fine
const Factor: Integer = 7; read in an expression compile error

So the scalar-read arm was the casualty, and the array arms were not — which is the opposite of what the error text ("invalid IR symbol reference") suggests to someone assuming an array-indexing problem.

Root cause (Track P, reported after the bisect)

The symbol index did not die with the routine: SymRollbackTo handed the const's index back, and the -O2 inliner then copied the body into the caller, where the copy still loaded it — one past the end after the rollback. -O0/-O1 clean, -O2/-O3 red, which named the inliner without guessing. Fixing that exposed a second, worse defect underneath: a wrong value at -O2 (7 became 0), because the run-once prologue guard is invisible to a caller that pasted the body, so the initialisation never ran and the BSS slot stayed zero. A compile error became a silent wrong answer.

Fixed in 467a4e5da, carried by pin v342 (315a0f77e).

Verified at HEAD

tools/testmgr.py --tier native --job 'test-core#src:test/test_local_typed_const.pas'
  PASS  test-core#756  unit  0.5s
  testmgr: GREEN

and the three minimised shapes above all compile and print correctly (42, b, 8 9).

Filing note — third auto-filed stub today needing a hand-set track:

Same as [[regression-lib-test-crtl-reachability]] and [[regression-test-core-test-strict-overload-width]]: twatch emits only prio:, so a stub about a Pascal test ranks in Track T's queue until a human moves it. Three in one day is a pattern, not an accident — twatch should guess a track: from the test path when it files.