Scalar functions fail or return wrong results on BIGINT argument overflow

[SEPURI-SAI-KRISHNA]

Four scalar functions perform unchecked 64-bit arithmetic on their user-supplied BIGINT arguments. When a value near the BIGINT bounds is passed, the arithmetic overflows and the function either returns a wrong result or fails with a raw ArithmeticException instead of the documented NULL / 0.

Query	Actual	Expected
SELECT slice(ARRAY[1, 2, 3, 4], 1, 9223372036854775807)	[]	[1, 2, 3, 4]
SELECT element_at(ARRAY[1, 2, 3], -9223372036854775808)	ArithmeticException: integer overflow	NULL
SELECT split_part('abc', '', 3000000000)	ArithmeticException: integer overflow	NULL
SELECT strpos('abc/xyz', '/', -9223372036854775808)	'instance' must be a positive or negative number.	0

The slice case is the most serious: it returns a wrong result silently rather than failing, so a query asking for "everything from position N onwards" via a large sentinel length quietly loses all its rows.

Details

slice computes its end index as fromIndex + length without an overflow check. A large length wraps the sum negative, which then loses the min(..., size + 1) comparison, so the fromIndex >= toIndex guard treats a fully in-range slice as empty.

element_at guards its bounds with Math.abs(index) > arrayLength. Math.abs(Long.MIN_VALUE) is negative, so the guard does not fire and execution falls through to toIntExact(arrayLength + index), which throws.

split_part converts its index with toIntExact when the delimiter is empty, so any index above Integer.MAX_VALUE throws — even though such an index is simply past the end of the string, a case the function already handles by returning NULL (split_part('abc', '', 99) → NULL).

strpos negates a negative instance with Math.abs before searching from the end. Math.abs(Long.MIN_VALUE) is negative, so stringPositionFromEnd rejects it as non-negative and reports "'instance' must be a positive or negative number" for a value that is in fact negative. The @ScalarFunction(value = "strpos", neverFails = false /* for instance==0 */) annotation documents instance == 0 as the only failure mode, so this is an unintended second one.

Related existing handling

This class of input is already guarded elsewhere in the same code:

• substring uses Ints.saturatedCast
• overlay uses LongMath.saturatedAdd
• json_array_get special-cases Long.MIN_VALUE and returns NULL

These four functions appear to have been missed.

Environment

Reproduced on master.