The wall, and how to stand in front of it

$ pascal26 --mimic-fpc-compiler \
    -Fu/usr/share/fpcsrc/3.2.2/compiler \
    -Fu/usr/share/fpcsrc/3.2.2/compiler/x86_64 cf.pas
pascal26:136: error: unknown type: TExecuteFlags
  in: /usr/share/fpcsrc/3.2.2/compiler/cfileutl.pas
  near: ComLine : AnsiString ; Flags : >>> TExecuteFlags = [
pascal26:137: error: unknown type: TExecuteFlags

cf.pas is program cf; uses comphook, finput, cfileutl; begin WriteLn('cf ok') end. The second -Fu is not optional — without it cpuinfo is not found first and you never reach this. Writing the configuration down is the point of this ticket: the march is driven by hand, and frankB's standing note is that "a substitute for a flag is a configuration nobody else runs, so every number it produces is unshared."

What it is NOT

The parse is fine. Two overloads with a defaulted set after an open-array parameter compile and run:

type TEF = set of (efInheritsHandles);
function A(const P: AnsiString; const C: AnsiString;       F: TEF = []): LongInt;
function B(const P: AnsiString; const C: array of AnsiString; F: TEF = []): LongInt;

prints 1 2. The type is simply absent from our RTL.

What FPC declares

rtl/objpas/sysutils/sysutilh.inc:

type
  TExecuteFlags = set of (ExecInheritsHandles);
function ExecuteProcess(const Path, ComLine: AnsiString; Flags: TExecuteFlags = []): Integer;
function ExecuteProcess(const Path: AnsiString; const ComLine: array of AnsiString;
                        Flags: TExecuteFlags = []): Integer;

plus RawByteString/UnicodeString spellings and EOSError.

The machinery is already here

lib/rtl/subprocess.pas:102 builds Popen on PalVforkAndExec(prog, argvp, envp, ...) and waits with Popen.wait. ExecuteProcess is the same three steps with a different argument shape — split ComLine on whitespace for the string overload, take the array directly for the other. A type-only stub is worse than nothing: cfileutl's implementation section calls ExecuteProcess, so parsing further only moves the failure to link time.

Where it came from

The march's previous wall, cclasses.pas:2909, is gone — see [[bug-p-a-double-deref-in-fpcs-cclasses-is-refused-and-the-obvious-reduction-compiles]], attributed to a4cbaa1de. Four walls have now fallen in this unit in sequence (TFPCHeapStatus, the forward pointer-to-array, Prefetch, the double deref); each was invisible until the one in front of it cleared, which is why a march wall is worth a ticket only once it is the frontier.

2026-09-11, frankS — it is the frontier again, and now it is the WHOLE frontier

Attempt 7 of [[umbrella-pxx-compiles-fpc-itself]] (probe #11, whole corpus, e013c4344): 119 of 207 units stop here, up from a handful when this was filed, because everything that used to stop earlier in cclasses.pas now walks past it. cclasses.pas itself compiles.

DO NOT RANK THIS ON 119. The same attempt measured what clearing a wall of that size is actually worth: the Finalize(x, n) wall had 150 units stacked on it and clearing it moved BOTH-OK by three — the unit itself plus its two direct dependents. The walls are stacked in a few shared units, so a blocker's unit count is a queue position. Rank this on the fact that it is a small surface over machinery that already exists, which is what the section above says.

Nothing in this ticket's diagnosis has changed and none of it needed re-deriving — including the part that matters most, and it is easy to skip on a reread: a type-only stub is worse than nothing, because cfileutl's implementation calls ExecuteProcess and adding the type alone only moves the failure to link time. A second, smaller thing sits immediately behind it and is unmeasured because nothing reaches it today: cfileutl then writes const ExecuteProcess = 'Do not use' deprecated '<msg>' to hide the sysutils name — a const carrying a deprecated directive. Whoever clears the type should expect it and not read it as a regression.

Resolved 2026-09-11 (frankH, Track B)

lib/rtl/sysutils.pas gains EOSError (with ErrorCode), TExecuteFlags = set of (ExecInheritsHandles), and both ExecuteProcess overloads — the string form (whitespace-split, quote-naive, as fpc's is) and the array of AnsiString form. Built on PalVforkAndExec + PalWait4 + EnvironmentBlock, which were all already here; no new capability, a surface.

Exit-code convention is (status shr 8) and 255. A failed spawn and a child that could not exec both raise EOSError carrying the code, matching fpc.

The corpus moved, and the delta was A/B'd on ONE binary

Same compiler/pascal26, the change stashed and restored between the two halves, so this is not a pull's delta read as mine:

Both columns are one wall counted three times, and then one wall counted twice. All three "before" rows are the SAME LINE of the same file, and the two IsATTY rows are both comptty.pas:66 — termio.IsATTY(t) — reached through a shared dependency. Units-compiling moved by ZERO, which is the umbrella's own finding for the fifth time: a wall's population is a queue position, not a size. The two new walls (TDoubleRec, termio.IsATTY) are unowned and unfiled as of this resolution.

Verification

• Oracle: every expected value in test/lib_sysutils_executeprocess.pas was read off fpc 3.2.2 on that same source, not predicted. fpc passes the fixture whole (EXECPROC OK); so does pxx under the pin.
• Positive control: with lib/rtl/sysutils.pas stashed the fixture cannot compile — pascal26:60: error: undefined variable (ExecuteProcess), rc=1.
• Row control: tools/expect_same.sh rejects a wrong expectation for this row (rc=1), so the row can fail.
• An earlier draft's array-argc-0 row was a guard that could not fail and was replaced. It put the space inside the COMMAND STRING and expected 0: [sh,-c,'exit $#'] and its re-split [sh,-c,exit,$#] both leave $# at 0, so the row certified the property it was written to test in either direction. The row now puts the space in a POSITIONAL — $0=zero, $1='a b', $2=c, so $# is 2 and a re-split would answer 3.
• make lib-test does not reach this row: it dies ~180 lines earlier at Makefile:33417, lib-units: FAIL mimic_queue — unknown type: TPyDeque, the standing owner-only pin cliff. The row was run by executing the recipe's two lines verbatim against the same $(TESTTMP).

Filed on the way

[[bug-p-an-array-constructor-in-argument-position-is-typed-as-a-set]] — ['x'] in argument position is typed as a set, so an array of T overload is unreachable for it. Against these declarations it does not refuse, it selects the WRONG overload silently. That is why no row here uses a [...] literal.

Log

• 2026-09-11 — resolved; this names the commit that carried the resolve, which is not always the one that carried the change — commit 70220c6f4.