Optimize RDAIO

[alyst]

Another round of RDA performance polishing:

1. Make long R vectors support optional. Disable by default on 32-bit systems (it's unlikely many people work with big dataframes on 32-bit systems, and in all other cases iterating over Int64 variable just slows everything down), R does the same.
2. Get >20% improvement in time and memory by improving type inference of IO variable. Before:

[~]$ julia -e "using DataFrames; @time read_rda(\"test.RData\")"
elapsed time: 15.414509979 seconds (2433718088 bytes allocated, 5.44% gc time)

after

[~]$ julia -e "using DataFrames; @time read_rda(\"test.RData\")"
elapsed time: 12.278712969 seconds (1834821860 bytes allocated, 3.38% gc time)

[coveralls]

Coverage increased (+0.01%) when pulling e6fc839 on alyst:optimize_RDAIO into 0f21922 on JuliaStats:master.

[garborg]

Thanks, @alyst.

What did you find out on the NaN/NA swap we're seeing for ASCII data?

[alyst]

@garborg For ASCII/ASCIIhex output they just output NA if ISNAN(d) is true and that discards NA/NaN distinction, whereas for XDR/native binary formats they write the real as is. I guess it's a rather long standing R bug, but I didn't found any discussion/bug reports.

[garborg]

Ah, interesting, I understand the comment in the tests now. Thanks for expanding.

[alyst]

Filed R-project bug 16137.

[garborg]

Nice!

[StefanKarpinski]

It's kind of amazing that our issue counter is rapidly catching up to the R project.