22 sysutils/strutils symbols FPC has and pxx does not

• Track B (lib/rtl). Found 2026-08-20 by an FPC differential probe of the string/file/date surface.
• Companion finding to the good news below — this is a coverage ticket, not a correctness one.

First, what is already right

A 35-row probe of the implemented string surface (su1.pas in the session scratch) came back byte-identical to FPC, including the cases that usually differ: Copy past the end / with count 0 / with start 0, StringReplace with rfIgnoreCase, CompareStr sign, IntToHex(-1,8), Format with a negative width, QuotedStr doubling. So nothing below is a behaviour difference — each one is a symbol that is simply absent.

Measured

Method: for each candidate, one program that calls it in its FPC signature, compiled by fpc -O- -Mobjfpc and by pxx at 35b69f6e3. Listed only where FPC compiled clean and pxx said undefined variable.

Note on method: an earlier pass probed with @Name and produced six false positives — Sqr, Odd, Trunc, Round, Frac, Int are intrinsics, and FPC refuses @ on them too. They are fine; they are not in the list.

Also missing, and each one stops a program at its var line rather than at a call, so they read as worse than they are:

• TStringArray — and with it SplitString, whose result type it is. These two go together; SplitString('a,b,,c', ',') is the idiom this blocks.
• TProcedure — the parameterless procedure type. A workaround is one line (type TProc0 = procedure;), but every FPC example that takes a callback uses the RTL name.

StrLen and StrPas ARE present, which is what makes the PChar row worth doing as a set: the family is half there, so code reaches for it and then hits a wall on the second call.

Suggested split

The rows are independent; take them a group at a time rather than as one lump. The PChar family and TStringArray+SplitString are the two that most often stop real code. Related, already filed: feature-lib-strutils-ansi-predicate-family, feature-lib-sysutils-strtodate-and-strtodatetime, feature-sysutils-decodedate-missing.

Gate

Track B: build with $(PXX_STABLE), make lib-test / demos. Each group lands with an FPC-differential test — the probe above is the shape: call it, print it, diff against fpc -O-.

Log

• 2026-08-26 — resolved, commit f41c0c42f (the RTL work; write-up in 92f68eea3).

Resolution (2026-08-26, frank1-B-rtl)

All 22 routines and all 3 type names landed. The inventory was re-verified against the pinned stable before implementing rather than trusted: it was accurate except for one row — StrCopy/StrComp/StrScan did exist, in lib/rtl/strings.pas, and were missing only from SysUtils, which is the unit real FPC code names. That is now fixed too, so the row's premise (the family is half there) held even though three of its six entries did not.

Everything below was MEASURED by compiling the same program under fpc -O- -Mobjfpc -Sh and under pxx and diffing the output. A 103-line differential probe covering the whole set came back byte-identical.

The contracts a plausible implementation gets wrong

These are why the ticket asked for measurement, and each one was measured:

• StrNew('') and StrNew(nil) are nil — not an empty allocated buffer.
• StrECopy returns the cursor at the terminating NUL, not Dest.
• StrBufSize(StrNew('abc')) is 4 (len+1), not a rounded capacity.
• CreateDir on an existing directory is False; only ForceDirectories is idempotent.
• GetCurrentDir has no trailing slash; ExpandFileName('') is the cwd with one.
• ExpandFileName collapses ./.. lexically on paths that do not exist, but a leading // survives (POSIX reserves it) where every other slash run collapses.
• DateToStr is 7-3-20, not ISO: / in a format string means DateSeparator (default -) and a lone y is a two-digit year.
• DateTimeToStr drops the time half at exact midnight — once a day a log line loses its clock. Milliseconds do not bring it back.
• TextToFloat trims surrounding blanks but rejects trailing junk: '1.5x' is False, not 1.5.
• Three different split models disagree on a,b,,c: ExtractWord gives a/b/c, ExtractDelimited gives a/b/''/c, ExtractSubstr gives a/b/c/''.
• SplitString's second argument is ONE multi-character separator in FPC 3.2.2, not a character set: 'a1b2c' split on '12' comes back whole.

One deliberate divergence, escalated

FPC ships two incompatible StrAllocs — strings allocates prefix-free, sysutils with a 4-byte size prefix — so uses SysUtils, Strings pairs the first with the second's StrBufSize/StrDispose. Measured: that reads 4294967292 and frees p-4. We ship one implementation; single-unit programs (i.e. all of them) see exactly FPC, and only the programs FPC corrupts can tell the difference. Filed for the owner as decide-stralloc-one-implementation-or-fpcs-two.

Compiler bugs found, none worked around in lib/rtl

• bug-single-char-literal-as-pchar-argument-segfaults — pre-existing and the worst of the three: StrCat(buf, '-') faults where '--' works, because a one-char literal passes its ordinal as the pointer. Reachable through StrCopy/StrCat, which have been here all along; nothing had ever passed a single character to them.
• bug-pchar-difference-in-writeln-arg-segfaults
• bug-dynarray-function-result-passed-directly-types-as-pointer — blocks the Dump(SplitString(line, ',')) idiom.

Only the tests were adjusted around these (documented in place, with the slug to revert at); no library code was reshaped.

Parked

bug-b-f-fixed-point-rounding-of-a-tie-goes-down-where-fpc-goes-up — the one row of twenty where FloatToStrF differs from FPC. Pre-existing in FmtFixed and already reachable through Format('%.2f'); last-digit-only, so Track F by definition (devdocs/progress/float/).

Tests

test/lib_strings_pchar.pas, test/lib_strutils_words.pas, test/lib_sysutils_dirs_dates.pas — wired into lib-test. All three compile and pass under FPC as well as pxx, which is what makes them oracles rather than recordings of our own behaviour.

Gate: make lib-test green end to end. The one failure in that target, MIMIC-XML-ETREE, is pre-existing and unrelated — verified by reverting these three RTL files and watching it fail identically; it is already held in working/bug-n-a-callable-value-reaches-a-str-parameter-and-renders-as-bound-method.