← board

The crtl header root resolves against CWD, so building outside the repo root silently uses glibc's headers

Symptom

~/songformatter/render_backend.py (and convertrawtext.py, and SongFormatter.py) fail to compile — but ONLY when the compiler is invoked from their own directory, which is the only way a real user would ever invoke it:

$ cd ~/songformatter
$ pxx render_backend.py out
pascal26:9253: warning: undeclared identifier 'aq' used as value (treated as 0)
pascal26:9255: warning: undeclared identifier 'ap' used as value (treated as 0)
pascal26:9262: error: IR_UNSUPPORTED: frontend could not lower AST node (kind 1)
  near:   return len   >>> static ssize_t dstr_append_data

The same file, same flags, from ~/frank2: ok. The .py is irrelevant — the failing code is lib/vendor/pdfgen/pdfgen.c, pulled in as a UNIT by lib/pcl/mimic_reportlab_pdfgen.pas (uses '../vendor/pdfgen/pdfgen.c'), which any import reportlab reaches.

ap/aq are va_list. They are undeclared because <stdarg.h> came from /usr/include, where va_list is spelled with gcc builtins pxx does not have.

Root cause

AddDefaultCIncludeDirs (compiler/cpreproc.inc:1883) registers two roots:

if ExeDir <> '' then
begin
  ConcatThree(ExeDir, '../lib/crtl/include/', '', d);   { assumes ExeDir = <root>/compiler/ }
  AddCIncludeDir(d);
end;
AddCIncludeDir('lib/crtl/include/');                    { CWD-relative }

The comment says the ExeDir anchor "resolves regardless of CWD". It does — for a compiler built in place at <root>/compiler/. It does not for the binary everyone actually runs: the stable/pinned one lives at <root>/stable_linux_amd64/default/pinned, so ExeDir + '../lib/crtl/include/' is <root>/stable_linux_amd64/lib/crtl/include/, which does not exist. That leaves only the CWD-relative root — so pxx finds its own C headers iff you run it from the repo root, and falls through to /usr/include otherwise.

Confirmed by construction: copy the same stable binary into a layout that DOES have ../lib next to it and the compile gets 5 lines into pxx's own stdarg.h; without it, the same invocation gets 708 lines into glibc's.

Note builtin/ does not have this problem — parser.inc:26881 looks for ExeDir + 'builtin/' and the stable layout ships exactly that. The Pascal unit roots (../lib/rtl, ../lib/pcl) miss for the same reason as crtl, but harmlessly: every caller passes -Fulib/rtl -Fulib/pcl explicitly and a missing unit is a hard error. The C include root is the dangerous one because missing it is silent — /usr/include answers instead.

Why this is a bug and not a packaging nit

The failure mode is the one this repo treats as worst: a wrong VALUE, far from the cause. Alongside the va_list break the same compile reports

warning: undeclared identifier 'M_SQRT2' used as value (treated as 0)

— a math constant silently becoming 0 in a PDF geometry routine. Nothing crashes; the output is just wrong. Any C corpus built from its own source directory is exposed.

Shape of a fix (Track C's call)

Three candidates, not mutually exclusive:

  1. Anchor for the shipped layout too — also register ExeDir + '../../lib/crtl/include/', so <root>/stable_linux_amd64/<x>/pxx resolves. Cheapest; still guesses at a directory depth.
  2. Ship the headers next to the binary, as builtin/ already is, and look for ExeDir + 'crtl/include/' first. Makes the stable binary self-contained, which is what a distributed compiler should be.
  3. Diagnose the fallback. Whatever the search does, taking <stdarg.h> from /usr/include should WARN. The wrong-ABI risk is called out in that procedure's own comment; a program that hits it deserves to be told.

(3) is worth doing regardless of which of (1)/(2) lands.

Reproduce

cd /tmp && pxx -Fu<root>/lib/pcl -Fu<root>/lib/rtl -Fu<root>/lib/rtl/platform/posix \
    <any .npy that does `from reportlab.pdfgen import canvas`> out    # fails
cd <root> && pxx ... same ...                                        # ok
cd /tmp && pxx -I<root>/lib/crtl/include -I<root>/lib/crtl/include/sys \
              -I<root>/lib/crtl/src ... same ...                     # ok

Gate

C tests green + self-host byte-identical; plus a regression that compiles a .c-as-unit from a CWD that is NOT the repo root and asserts pxx's own <stdarg.h> was the one used.

Log