Rounding Inf and NaN

[dukeleto]

$ ./perl6

say ceiling(NaN)

-2147483648

say floor(NaN)

-2147483648

say floor(Inf)

-2147483648

say ceiling(Inf)

-2147483648

[dukeleto]

Attempt at testing rounding of NaN, but I get runtime error about 'attempt to access code outside of current code segment'

[dukeleto]

Complete Error message of the failing example test

[dukeleto]

Passing NaN-rounding-related tests

[dukeleto]

After a little more digging, it seems that Rakudo's floor() passes the string 'NaN' down to Parrot's rounding functions, but PASM's floor() and cousins only act on numeric registers, so 'NaN' and 'Inf' turn into some machine-dependent integer. Code written in PIR works just fine though.

[coke]

Per original poster, this is a rakudo issue, not a parrot issue; created a ticket for rakudo at

 http://rt.perl.org/rt3/Ticket/Display.html?id=63440

which just points back here for now. Resolving ticket.

[NotFound]

A shortened version of the patch for the arithmetic test added in r36971

[pmichaud]

(reopening ticket with message I sent to parrot-dev earlier this week)

FWIW, I don't entirely agree with the premise that this is strictly a Rakudo bug. I find that Parrot has some undefined behavior when dealing with NaN, Inf, and integers, and little documentation to indicate what the expected behavior (if any) should be. As an example:

    $ cat x.pir
    .sub 'main'
        $N0 = 3.5
        say $N0
        $I0 = floor $N0
        say $I0

        $N0 = 'Inf'
        say $N0
        $I0 = floor $N0
        say $I0
    .end

    $ ./parrot x.pir
    3.5
    3
    Inf
    -2147483648
    $

On my machine, Parrot seems to always treat an integer conversion of any infinity as -2147483648. Better would seem to be to throw a range exception or the like, instead of silently converting to a non-infinite integer value.

Pm

[dukeleto]

The example that pmichaud++ gave is now a TODO test in t/op/arithmetics.t

[dukeleto]

Tests relating to Inf/NaN now live in t/op/inf_nan.t . I have added many tests recently , and the same bug that pmichaud++ describes happens to the fdiv, mod and cmod functions.