← board

The measurement

program dm;
uses erroru, sysutils;
var mem: SizeUInt; i: Integer; s: string;
begin
  mem := 0;
  DoMem(mem);
  for i := 1 to 100 do begin s := IntToStr(i); s := s + 'x'; end;
  writeln('control lost = ', DoMem(mem));
end.



pxx: [HEAP] Size: 262144 Kb,  Used: 240 bytes,  Lost: 208 bytes    -> 208
fpc: [HEAP] Size: 352 Kb,     Used: 1344 bytes, Lost: 64 bytes     ->  64

erroru.pp is FPC's own testsuite helper; DoMem is GetFPCHeapStatus.CurrHeapUsed before and after.

Why it is filed rather than fixed, and what it exculpates

texception3.pp was skipped for "RTL ExitCode variable missing; try/finally with exit/break/continue". Both clauses are stale. Re-measured, the row compiles, runs, and passes all 119 exception sub-tests; its only failure is the last statement:

if DoMem(mem)<>0 then
  begin
    writeln('exception generates memory holes');
    do_error(99999);
  end;

exception generates memory holes is the message that comes out, and it is not what is wrong — the control above shows the same non-zero answer for a program containing no exception at all, under both compilers. The pxx allocation census agrees: a 1000-iteration raise/handle loop gives allocs=1871 frees=1868 live=3.

That is the whole reason this ticket exists. The reading "exception handling leaks 3408 bytes" is coherent, comes from FPC's own testsuite, and is false — and I had already written it into pxx.skip before running the control. A message printed by a test is a claim about the test's own model, not a measurement.

The residual, which is what someone owns here

Neither compiler returns CurrHeapUsed to its starting value, so a test asserting DoMem = 0 is asserting something about an allocator rather than about the code under it. Whether pxx's 208 is defensible is not established. Two sub-questions:

  1. Should CurrHeapUsed fall when a block is freed? pxx bump-allocates from one arena; if the number is a high-water mark by construction, say so in the function's own contract, because the name promises otherwise.
  2. Size: 262144 Kb against fpc's 352 Kb is the arena reservation. Nobody has said that is wrong, and nobody has said it is right either.

Low prio: no compiling program's ANSWERS depend on it — only a program that introspects the allocator, and this repo does not chase parity on those (CLAUDE.md, "compatible with FPC means the value, not the intermediate's type"). It is filed because "not exception handling" is half a finding.