08.03.2012 22:08, Manu пишет:
I find myself really wishing for proper multiple return values almost
every day, particularly when I work with maths heavy code, and
especially for efficiently returning error codes in functions I'd
rather not throw from.
Many maths-y functions return some sort of pair; intersections return
(ray, t) or something of that type.
I'm finding HEAPS of SIMD functions want to return pairs (unpacks in
particular): int4 (low, hight) = unpack(someShort8);
Currently I have to duplicate everyting: int4 low =
unpackLow(someShort8); int4 high = unpackHigh(someShort8);
I'm getting really sick of that, it feels so... last millennium.
The point of 'proper' multiple return values is to return each value
in registers, in its own register type, using exactly the same
register assignment pattern as when passing args TO functions.
I don't think this causes any side effects to the ABI, since the arg
registers are already volatile across function calls in the first place.
It just means that the returned-to function can find its return
values already conveniently in an appropriate register, avoiding
memory access.
People argue I should return a tuple, but this isn't really the same,
it has hidden implications that complicate the optimisation potential.
For instance, tuples have an implicit structure/memory layout which
can't be ignored, whereas return values don't have memory allocated,
ie, you can't take the address of a return value without first
assigning it to some local syntactically.
The implementation of efficient tuple return values would be much more
complicated I would imagine too, and the rules are less clear; I can't
intuitively presume what behaviour returning a tuple of different
things should actually have in terms of register assignment. I also
know precisely how multiple return values should work, because it is
exactly the same as passing arguments to the function, but in reverse.
... just saying :)
I'd like to see this as a part of tuple improvement, since tuples may
hold additional compile-time information. This would make it possible to
write efficient code, and at the same time have the ability to reference
tuple components by name rather than position. For example:
{
auto t = getSomeTuple(...); // returns Tuple!(float, "x", float,
"y"), no actual assignment is made
someVar = t.x^^2 + t.y^^2; // t.x and t.y are just aliases for st1
and st0
}
With a tuple unpacking syntax suggested by bearophile, this would be
nicer than any distinct multiple return values syntax, IMO.
