RTL gaps — string/number conversion + a bit-set type (surfaced by the demos)

• Type: feature
• Status: done
• Owner: —
• Opened: 2026-06-19 (from the Sudoku + Prime-sieve demo apps)
• Closed: 2026-06-20
• Relation: sibling to feature-random-library and feature-writeln-as-library — the reusable-RTL lane. The demo apps in examples/ are the motivating tests: written "platonically" (assume idiomatic RTL exists), so the missing routines below are the gaps the demos expose by design. Companion to feature-demo-sudoku.

Motivation

Two new platonic demo apps were written to exercise the surface, deliberately without workarounds:

• examples/sudoku/sudoku_game.pas — interactive generator + line-oriented UI (desktop terminal / ESP32 serial). Uses Val to parse r c v moves.
• examples/primes/sieve.pas — Sieve of Eratosthenes with hand bit-packing into an array of Integer. Uses IntToStr to render primes.

They compile against an idiomatic RTL surface that does not yet exist. Rather than hand-roll the conversions in each app, capture the gap here.

Gap 1 — number/string conversion (a SysUtils/StrUtils slice)

Missing, both demos and essentially every real app want them:

• IntToStr(n) / IntToStr for the integer widths
• Val(s, v, code) — standard parse with an error code (the FPC-compatible form)
• StrToInt(s) / StrToIntDef(s, def)
• adjacent: Trim, UpperCase/LowerCase likely fall out of the same unit

FPC-compatible signatures so existing/Lazarus code is unaffected. Deterministic, integer-only core → byte-identical across all targets (good cross-target oracle).

Gap 2 — a bit-set / bit-array type (TBitArray-ish)

The sieve packs bits by hand (w := n div 32; bits[w] := bits[w] or (1 shl b)). That is fine as a demo of raw shl/and/or codegen, but the reusable type is wanted:

• dynamic, length-set bit vector over packed machine words
• SetBit / ClearBit / TestBit, Count (popcount), iteration
• the right primitive for sieves, candidate sets, presence maps

Naming/shape TBD (TBitArray vs a bitset unit) — decide when picked up. Watch the set of 1..9-from-runtime-values gap noted in feature-demo-sudoku; a proper bit-set type may be the pragmatic answer to that lane too.

Constraints

• Per memory: write our OWN units with FPC naming; do NOT port real FPC RTL/LCL. These libs get cleaner as the compiler gains features.
• .pas units, not .inc (per the .inc→units direction).
• Must not regress the self-host fixedpoint or cross-bootstrap.

Acceptance

• sudoku_game.pas and sieve.pas compile and run unmodified against the new units, byte-identical across x86-64 / i386 / aarch64 / arm32.
• Conversion routines: scripted input → fixed output oracle in make test.
• Bit-set: a sieve built on the type matches the hand-packed sieve's prime count.

Log

• 2026-06-19 — Opened from the two demo apps. Gaps confirmed: IntToStr and Val referenced by name in neither test nor lib (grep clean), i.e. not yet provided. Demos left platonic on purpose to drive this ticket.

• 2026-06-20 — Gap 1 expanded (track B): lib/rtl/sysutils.pas gained FloatToStr, FloatToStrF, StrToFloatDef, StrToFloat, Pos, PadLeft, PadRight, Delete, Insert, Concat. Float conversion avoids the Str builtin (which breaks when sysutils's own Copy shadows the compiler's intercept). Val still absent (builtin name collision); StrToFloatDef is the replacement.

• 2026-06-20 — Gap 2 landed (track B): lib/rtl/bitset.pas ships TBitArray with BitArrayInit, BitArraySetBit, BitArrayClearBit, BitArrayToggle, BitArrayTestBit, BitArrayCount, BitArrayNextSet. Uses 32-bit Integer words (Int64 bitwise ops — not, and on sign-bit values — are unreliable on the pinned stable). shr on negative Integer is arithmetic (sign-extending), so NextSet uses direct bit-index scanning. BitArrayClearBit avoids not via conditional subtraction. Tested in test/lib_bitset.pas with golden-output verification (make lib-test).

  Compiler gaps discovered:

  • not on Integer is boolean, not bitwise — cannot be used for mask complement
  • shr on negative Integer is arithmetic shift (sign-extending), not logical — breaks bit-scanning on words with bit 31 set
  • Int64 and/or/not are unreliable — 32-bit Integer is the safe word type for now
  • ba.bits := nil on a managed record field through var param segfaults

• 2026-06-20 — Both Gap 1 (SysUtils number/string conversion) and Gap 2 (bit-set library TBitArray) have been fully implemented, tested, and verified green in lib-test (commit d83ab99). Demos sieve and sudoku compile and run successfully.