ESP32 managed strings (tyAnsiString runtime on xtensa + riscv32)

• Type: feature
• Status: done 2026-06-18 (all core paths; also exercised on bare-metal in feature-esp32-bare-boot)
• Owner: —
• Opened: 2026-06-17 (from feature-esp32-managed-features)
• Depends-on: feature-esp32-managed-features

Progress

• 2026-06-17 — core done (commit eb849b3): builtinheap restructured so the target-independent ARC runtime (PXXStrFromLit/Concat/IncRef/DecRef/Unique/Eq/ SetLen + MemMove/MemZero + PXXDynSetLen) is no longer behind the blanket {$ifndef PXX_ESP} — only file I/O, managed-element dynarray/record retain, variant and float formatting stay ESP-excluded. Codegen on both ESP backends: IR_STORE_SYM tyAnsiString (literal/char/frozen/handle build + DecRef-old publish), IR_LEA scalar tyAnsiString (read=load handle, write=slot addr), Length [handle-8], index handle+(i-1). test_esp_string → 3/PXX/PXX both ISAs == oracle.
• 2026-06-17 — concat done (commit dd91333): a + b → PXXStrConcat on both backends (handle/tyString/char operands, fresh-temp DecRef, nested concat). test_esp_strcat → "PXX rocks"/"PXX rocks!"/9 both ISAs == oracle.
• 2026-06-18 — riscv32 compare + harness fix (commit c7309b0): PXXStrEq on riscv32 (Y/N/Y/Y/N == oracle); shared EmitStrOperandRISCV32 decomposition; esp_run.sh force-relinks (was running stale images — see Known bugs). xtensa compare + the concat-of-two-string-literals bug remain (below).
• 2026-06-18 — xtensa compare done: PXXStrEq path in the xtensa IR_BINOP (operand decompose mirrors concat: a4=lenA,a5=srcA,a6=lenB,a7=srcB; xor for tkNeq). test_esp_strcmp Y/N/Y/Y/N both ISAs == oracle.
• 2026-06-18 — SetLength + s[i] COW + scope-exit DecRef done (both ISAs). SetLength(s,n)->PXXStrSetLen (tyAnsiString IR_LOAD_SYM lvalue, mirrors aarch64). s[i] write COW: IR_INDEX clones-if-shared via PXXStrUnique when base is an IR_LEA scalar tyAnsiString in write position; IR_STORE_MEM now sets InLValueWrite around the address eval so the IR_LEA keeps the slot addr. Scope-exit DecRef of tyAnsiString locals in EmitProcEpilog (xtensa+riscv32). ROOT-CAUSE FIX: the ESP prologue managed-local nil-init (parser.inc) was a no-op placeholder -> uninitialised local strings held stack garbage and scope-exit DecRef wild-freed them (Guru Meditation / Load access fault). Now zeroes the pointer-sized handle slot. Tests: test_esp_strmut (QXX/QYX/QXX/QY), test_esp_strscope (4000-churn arena survival -> OK4/kept). Full ESP suite byte-equal to x86-64 oracle both ISAs; make test green.

Known bugs (found 2026-06-18)

• concat of two string literals crashes FIXED (commit e1c9198). Root cause was not the EmitStrOperand decomposition: tyString + tyString has no frozen-concat codegen on ESP, so it fell through to the integer binop (adding two const_str addresses), then the sum was read as a string -> garbage len -> arena overflow -> crash. In managed mode the only tyString operands are literals, so 'a' + 'b' (string/char literals) now folds to a single interned literal at IR-lowering (ESP-gated; self-host byte-identical). Residual unfolded tyString concat (nested pure-literal 'a'+'b'+'c') raises a clear ESP-backend error instead of miscompiling.
• xtensa string compare not wired FIXED 2026-06-18. PXXStrEq path added to the xtensa IR_BINOP (decomposition inlined like the xtensa concat, no shared helper). test_esp_strcmp Y/N/Y/Y/N both ISAs.

Remaining

All core paths done (compare, SetLength, s[i] COW, scope-exit DecRef). Ticket ready to close pending board move. Open follow-ups (minor):

• Managed-aggregate locals on ESP (array of AnsiString / record-with-managed fields) still raise "managed aggregate locals not yet supported" in the prologue zero-init — only the scalar pointer-sized handle is nil-inited.
• xtensa char-concat operand uses addi a6, sp, off (±128 imm) — deep nesting with a char operand could overflow; revisit if hit.
• Scope-exit DecRef on ESP only releases scalar tyAnsiString locals (not local dynarrays — those leak on the static arena, harmless; mirror later if needed).

Motivation

PXX defaults to managed strings: PasInitDefines (lexer.inc) seeds PXX_MANAGED_STRING, so AnsiString/String are tyAnsiString (tk=23) — a refcounted handle to a [refcount:8][length:8][data] block (handle = data pointer, length at [handle-8], chars at [handle+0], s[i] = handle+(i-1) 1-based). -uPXX_MANAGED_STRING opts into the frozen tyString (tk=4) inline buffer.

On ESP the managed-string runtime (PXXStr*) is currently guarded out (feature-esp32-managed-features keeps only the allocator + lean dynarray). So s := 'lit' silently stores the frozen-literal pointer instead of building a managed handle, and Length(s) reads [ptr-8] = garbage. Real string support on ESP needs the runtime ported.

Scope

1. Port PXXStr* off the {$ifndef PXX_ESP} guard in builtinheap.pas: PXXStrFromLit, PXXStrConcat, PXXStrIncRef, PXXStrDecRef, PXXStrUnique, PXXStrEq, PXXStrSetLen. They cross-reference each other + PXXAlloc (not external strings), so they should compile on ESP once un-guarded; fix any stage-1 codegen gaps they surface. Register them for ESP in parser.inc (currently allocator-only).
2. Wire managed-string codegen on both ESP backends (mirror i386/aarch64):
   • s := <literal> → PXXStrFromLit(len, src) then publish handle (release old).
   • s := t (lvalue) → IncRef new + DecRef old (ARC).
   • a + b → PXXStrConcat.
   • Length(s) → [handle-8] (nil → 0) — already wired; verify it takes the managed path, not the dynarray one.
   • s[i] (read/write) → handle + (i-1); write position clones via PXXStrUnique (COW).
   • comparisons → PXXStrEq; SetLength(s, n) → PXXStrSetLen.
   • scope-exit DecRef of locals (ARC bookkeeping).
3. IR_LEA for a scalar tyAnsiString read must load the handle (mirror i386 tyAnsiString branch); write position keeps the slot address for PXXStrUnique to publish.

Validation

Print a built/concatenated string char-by-char via PutC(Integer(s[i])) (passthrough already landed) and compare to the x86-64 oracle on both esp32s3 + esp32c3 under tools/esp_run.sh. Cover: literal assign, var copy (refcount), concat, Length, index, SetLength. make test + self-host byte-identical after each step.

Notes

• A frozen-tyString Length/store path was prototyped + reverted during feature-esp32-managed-features (wrong layer — frozen is off-by-default).
• Use --dump-ir to confirm types: tk=23 managed, tk=4 frozen; tkLength builtin id = -44.
• ESP self-host is NOT a goal; validate by output-equality vs x86-64.