C: single-subscript row decay of a multi-dim array crashed (NULL pointer)

• Type: bug (Track C — C frontend multi-dim indexing, cparser.inc). Was a SIGSEGV / silent wrong value on valid C.
• Found: 2026-07-18, gcc-differential sweep while fixing [[bug-c-partial-multidim-array-index]].

Repro

static int g[6][4];
int main(void){ g[3][2] = 55; int *r = g[3]; return r[2]; }   /* want 55 */

• gcc: 55. pxx (before): SIGSEGV — g[3] (a single subscript on a >=2-D array = the C row decay to int*) evaluated to a scalar element load (0 = NULL), so r[2] dereferenced NULL. Same for 3-D (int(*r)[4] = g[1]) and for passing a row to a pointer parameter (f(g[i])).

Root

The C multi-dim flatten branch (cparser.inc, NodeArrNDInfo path) required a SECOND [ to fire, so a single subscript fell through to the generic per-node AN_INDEX, which strides by the element type and mis-decays the whole row to a scalar. NodeArrNDInfo only fires for rank >= 2, and the partial-index branch (from [[bug-c-partial-multidim-array-index]]) already builds the correct row pointer for nIdx < rank.

Fix (DONE)

Dropped the (CurTok.Kind = tkLBrack) (second-bracket) requirement from the ND flatten entry, so a single subscript enters with nIdx=1 and, since rank >= 2, takes the partial branch -> row pointer &base + (i-lo0)*rowStride. Full index (all subscripts) and 2-of-N partials are unchanged (the while-loop still consumes every present subscript). Rank-1 arrays never reach here (NodeArrNDInfo needs rank >= 2) and keep the plain AN_INDEX path.

Acceptance — MET

• Repro returns 55; 3-D row decay, deep partial (g[1][2][3] on [2][3][4][2]), and passing a row to an int* parameter all match the gcc oracle.
• C-conformance 220/220, self-host byte-identical. Regression test test/cmultidim_row_decay.c wired into make test-c.

Log

• 2026-07-18 — resolved, commit f8a2cdde.