← board

C pointer-to-array declarator char (*p)[4] hits IR "Unsupported linear node"

Failing test

Gate

Drop 00130.c from test/c-conformance/pxx.skip; runner green.

Triage note (2026-07-06)

Confirmed not a quick fix: needs a real "pointer whose element is an array" type (element = array-of-N, stride N*elem) plus 2-D stride through it for p[i][j]. The (*p)[4] declarator, the decay p = arr, and the double index all depend on that type existing. Codegen ICE "Unsupported linear node Kind=10" is the AST for the unmodelled type reaching codegen. A dedicated modeling change, not a patch.

Confirmed 2026-07-07

Even the bare declarator char (*p)[4]; fails to PARSE ("expected C expression") — the (*name)[N] pointer-to-array declarator isn't recognized at all, and the pointer-to-array TYPE (element = array-of-N, stride N*elem) isn't modelled. Needs both a declarator-parser addition and the type representation. Deep; focused session.

Assessment 2026-07-08 (a-agent) — confirmed a modeling session, scoped

Not started (released without holding). Concrete shape after the array-typedef work: char (*p)[4] = a POINTER whose element is a fixed array-of-4. Needs, in order:

  1. Declarator parse (*name)[N] — mirror the existing fn-ptr (*name)(...) branch in ParseCDeclType (CTypeFnPtrName path); capture N.
  2. Symbol shape — a new field e.g. SymPtrElemArrLen (and dim span for the multi-[N][M] case) on the pointer symbol: "the pointee is an array of N". No existing field carries this (PtrElemTk is scalar-only).
  3. Index lowering, two sites — p[i] must (a) LOAD p's value (it is a pointer variable, unlike an array whose address is used directly), then (b) stride by N*elemSize and yield an ARRAY-typed lvalue so the following [j] indexes within. Reuse the 2-D flat-stride math but with the first "dimension" coming from a pointer deref, not a fixed dim.
  4. Decay assignment p = arr — a 2-D array (SymArrNDims=2,[2,4]) decays to char(*)[4]: assign arr's base address, carry the row length 4 onto p. q = &arr[1][3] and *v already work (scalar element addr / plain deref). The ICE "Unsupported linear node Kind=10" today is the unmodelled AST reaching codegen. Peer of the fn-ptr declarator + the multidim array stride code; ~one focused session. Corpus payoff is modest (rare idiom) — prio stays 55.

FIXED 2026-07-09 (cfront-agent) — ticket closed

Implemented the full pointer-to-array elem (*p)[N] model. Five parts:

  1. Declarator parse (ParseCDeclType): detect the ( * ident ) [ shape (distinct from fn-ptr (*name)(...) and array-of-fnptr (*name[N])(...)), collapse to tyPointer, record row length in new global CTypePtrElemArrLen + the name in CTypeFnPtrName. Also a sibling handler in the ParseCLocalDeclAST multi-declarator loop (00130 is char arr[2][4], (*p)[4], *q; — the loop used to Break on the leading (); it mirrors the loop-tail comma/star setup so a following *q sibling is not dropped.
  2. Symbol field SymPtrElemArrLen (defs.inc) = pointee row length; reset in all four symtab alloc sites (parallel-array landmine).
  3. Index lowering (ParseCPostfix): p[i][j] → AN_INDEX(p, iN + j); the AN_INDEX over a POINTER base loads p and strides by PtrElemTk size, so the address is load(p) + (iN+j)*sizeof(elem). Full two-subscript form only.
  4. Decay p = arr: reuses the existing 2-D-array→pointer decay (base address); the row length is baked onto p at declaration, so no runtime carry.
  5. q = &arr[1][3] / *q / *v already worked.

00130 green (exit 0). Repro test/cptr_to_array_declarator_b206.c (exit 42; char

Log