Should prefer IV when both IV and NV are OK #39
Description
Hello,
After a lot of digging, I finally managed to understand why this module in my test script was always assuming that integers were huge float64's. Here's my discovery:
use Devel::Peek;
use Data::MessagePack;
my $mp = new Data::MessagePack;
$mp->prefer_integer(1);
my $a = [ 30000, 123, 300, 1000, 1000, 500, 500 ];
print 'Clean: ', unpack('H*', $mp->pack($a)), "\n";
#print Dump($a);
my $b = pack('wwwwwww', @{ $a });
#print Dump($a);
print 'After pack(): ', unpack('H*', $mp->pack($a)), "\n";This outputs:
Clean: 97cd75307bcd012ccd03e8cd03e8cd01f4cd01f4
After pack(): 97cb40dd4c0000000000cb405ec00000000000cb4072c00000000000cb408f400000000000cb408f400000000000cb407f400000000000cb407f400000000000
Using Devel::Peek's Dump() I finally figured out that pack() does its reading in a way which triggers Perl to add an NV in addition to the original IV. Before pack() the first value is:
SV = IV(0x17d32b0) at 0x17d32c0
REFCNT = 1
FLAGS = (PADMY,IOK,pIOK)
IV = 30000
But after pack() reads it, it becomes:
SV = PVNV(0x1934710) at 0x17d32c0
REFCNT = 1
FLAGS = (PADMY,IOK,NOK,pIOK,pNOK)
IV = 30000
NV = 30000
PV = 0
My understanding is that the integer version of the value is still usable, but MessagePack prioritizes the float version. My first attempt at fixing that (PP.pm#274 and pack.c#193) broke test 21 "dirty float" so there seems to be more to Perl's IOK/NOK than meets my eye.
I'll experiment further and hopefully come up with a pull request. I figure there must be an efficient way to determine whether the IV loses precision vs the NV. If it doesn't then IV wins, if it does then NV wins.