Wrapping subtraction/negation operator oddities

[jmc-88]

The table of operators in the Zig documentation contains the following:

Syntax	Relevant Types	Description	Example
a -% b a -%= b	Integers	Wrapping Subtraction. Guaranteed to have twos-complement wrapping behavior. Invokes Peer Type Resolution for the operands. See also @subWithOverflow.	@as(u32, 0) -% 1 == std.math.maxInt(u32)
-%a	Integers	Wrapping Negation. Guaranteed to have twos-complement wrapping behavior.	-%@as(i32, std.math.minInt(i32)) == std.math.minInt(i32)

I would expect the unary -% to behave identically to @as(some_unsigned_type, 0) -% 1, but instead here's what I'm seeing:

Operation	Expected Result	Actual Result
@as(u32, 0) -% 1	std.math.maxInt(u32)	std.math.maxInt(u32)
-%@as(u32, 1)	std.math.maxInt(u32)	invalid wrapping negation type: 'u32'
-%1	(not sure... it's a comptime_int, maybe it just shouldn't wrap?)	0

The compilation error on -%@as(u32, 1) is particularly interesting. I can't link the source on GitHub since the file is 1.5 MiB big (!), but the failure happens in src/stage1/ir.cpp, inside ir_eval_negation_scalar():

    bool ok_type = ((scalar_type->id == ZigTypeIdInt && scalar_type->data.integral.is_signed) ||
        scalar_type->id == ZigTypeIdComptimeInt || (is_float && !is_wrap_op));

    if (!ok_type) {
        const char *fmt = is_wrap_op ? "invalid wrapping negation type: '%s'" : "invalid negation type: '%s'";
        return ir_add_error(ira, source_instr, buf_sprintf(fmt, buf_ptr(&scalar_type->name)));
    }

I don't know why the requirement for scalar_type->data.integral.is_signed: the wrapping behavior is especially useful on unsigned types. A quick git blame indicates this check has existed since all the way back to 0ad580f.

[Vexu]

Wrapping negation of unsigned integers was accepted in #1770.

[LemonBoy]

See also #7951.