r.clump: Fail with the clump limit instead of overflowing the index - #7853

Fixes #6412.

r.clump stopped with

ERROR: G_realloc: unable to allocate 18446744065119617024 bytes of memory at raster/r.clump/clump.c:633

which reads as a memory problem. It is not one, and the reporter went looking for a disk-based mode because of it.

Where the number comes from

label is a CELL and nalloc an int, both 32 bit:

label++;
cur_clump[col] = label;
if (label >= nalloc) {
    nalloc += INCR;                                    /* INCR is 1024 */
    index = (CELL *)G_realloc(index, nalloc * sizeof(CELL));
}

nalloc starts at INCR and only ever grows by INCR, so it is always a multiple of 1024 and lands exactly on 2³¹, which as a signed 32 bit value is INT_MIN. nalloc * sizeof(CELL) then converts INT_MIN * 4 to size_t:

(2**64) - 2147483648 * 4 = 18446744065119617024

That is the number in the report, so this is the path taken rather than a guess. label itself overflows one step earlier, which is undefined behaviour and would give a negative clump ID.

The fix

@metzm said on the issue that the real problem is the clump count, and that the limit "can not be raised" — clump IDs are the raster values, so the ceiling is one CELL. So this reports the limit instead of overflowing past it, at both places that start a new clump.

MAX_LABEL stops INCR short of INT_MAX rather than at it, so nalloc — which moves in INCR steps and must exceed label — cannot overflow either. The cost is 1024 of 2.1 billion possible clumps.

The new message says what the limit is and what to do about it, since the previous one sent the reporter after the wrong problem.

Verification

I cannot add a regression test for this: reaching it needs a raster with more than 2³¹ clumps, so the reporter's 220000 × 240000 region is roughly the smallest case. Rather than assert it untested, the mechanism above is arithmetic that anyone can check against the reported number.

What I did check: builds clean, clang-format reports no changes, and the four existing tests in raster/r.clump/tests/test_clump.py still pass.