Adds size attribute aliases for DataFrame (closes #630)

[coatless]

No description provided.

[codecov-io]

Current coverage is 92.50% (diff: 100%)

Merging #638 into master will not change coverage

@@             master       #638   diff @@
==========================================
  Files            65         65          
  Lines          3309       3309          
  Methods           0          0          
  Messages          0          0          
  Branches          0          0          
==========================================
  Hits           3061       3061          
  Misses          248        248          
  Partials          0          0          

Powered by Codecov. Last update 271148a...e9db599

[eddelbuettel]

I don't think we have consensus that we what or need that.

I presume I could get my arm twisted to allow nrow and nrows; I see nothing but needed pain in removing the old one.

[coatless]

The goal behind the PR was to align the DataFrame class to match with:

1. the R functions ncol() and nrow();
2. the Matrix class member functions .nrow(), .ncol(), .rows(), .cols().

In an ideal world, the old interface probably should be cleaned up as a result. But, it most certainly can be left in place and the nrows() aspect can be omitted from documentation.

By my count, the following ~20 packages on CRAN would be affected:

• BacArena
• Biocomb
• Cyclops
• deepboost
• dplyr
• FeatureHashing
• FIT
• feather
• gaston
• medfate
• missDeaths
• Rfast
• Rborist
• raptr
• readr
• readstat13
• reshape2
• tweenr
• valr
• wsrf

[eddelbuettel]

I have no interest / time / energy to chase twenty package maintainers over essentially nothing.

I have tried to get five to change an error of theirs for three weeks now and only got half. No, we are not going to break that interface.

Additional accessors are perfectly fine.

[eddelbuettel]

I took the language about deprecation out and simplified a little. Now it is a straight-up extension of the API, plus extra tests, and should not create any trouble.

Second set of eyes anyone? @thirdwing @nathan-russell @kevinushey @jjallaire ?

[kevinushey]

LGTM. As an aside, these APIs should probably be returning R_xlen_t, but that's probably for a separate PR.

[eddelbuettel]

Good catch. The actual worker function here still uses LENGTH(), and that should be XLENGTH() with the commensurate change in returning R_xlen_t.

[coatless]

The following seems to work:

        inline R_xlen_t nrow() const {
            SEXP rn = R_NilValue ;
            SEXP att = ATTRIB( Parent::get__() )  ;
            while( att != R_NilValue ){
                if( TAG(att) == R_RowNamesSymbol ) {
                    rn = CAR(att) ;
                    break ;
                }
                att = CDR(att) ;
            }
            if (Rf_isNull(rn))
                return 0;
            if (TYPEOF(rn) == INTSXP && XLENGTH(rn) == 2 && INTEGER(rn)[0] == NA_INTEGER)
                return abs(INTEGER(rn)[1]);
            return XLENGTH(rn);
        }

        // Offer multiple variants to accomodate both old interface here and signatures in other classes
        inline R_xlen_t nrows() const { return DataFrame_Impl::nrow(); }
        inline R_xlen_t rows()  const { return DataFrame_Impl::nrow(); }

        inline R_xlen_t ncol()  const { return DataFrame_Impl::length(); }
        inline R_xlen_t cols()  const { return DataFrame_Impl::length(); }

Given the above discussion, I'll toss the above in a new issue and then submit a PR with the above change. Or, should I just submit the code change within here and add a new note to NEWS?

[kevinushey]

Sorry, looks like I'm not quite right -- while the number of columns (length) can indeed be a 'long' vector, the row.names attribute is still expected to be a 'short' vector. From

https://github.com/wch/r-source/blob/45576f65c403f46e03b88e4e0301bde330aea3be/src/main/attrib.c#L181-L203

The code for extracting the true length from a potentially 'compact' representation is:

int n = (isInteger(s) && LENGTH(s) == 2 && INTEGER(s)[0] == NA_INTEGER)
    ? INTEGER(s)[1] : (isNull(s) ? 0 : LENGTH(s));
ans = ScalarInteger((type == 1) ? n : abs(n));

So R indeed is still expecting an integer number of rows for data frames, if I understand correctly. I was under the impression that 'long' vectors in data frames had been supported now, but perhaps not (or at least, not for this specific API)

[eddelbuettel]

Thank you both for following up. So I guess I'll merge as is...

[coatless]

Based on @kevinushey's digging, the aliases for columns should have their return type updated to match the definition out of Vector.

        inline R_xlen_t ncol()  const { return DataFrame_Impl::length(); }
        inline R_xlen_t cols()  const { return DataFrame_Impl::length(); }

[eddelbuettel]

Thanks for making the change. This should now be good.

[kevinushey]

LGTM as well; thanks!

[eddelbuettel]

So in it goes -- thanks to @coatless for doing the work and to @kevinushey for being eagle-eyed as usual.