Remove non-API calls in C

[MichaelChirico]

We are getting these NOTEs on CRAN, e.g.

Found non-API calls to R: ‘SETLENGTH’, ‘SET_S4_OBJECT’,
‘SET_TRUELENGTH’, ‘UNSET_S4_OBJECT’

I'm really skeptical about the SETLENGTH and SET_TRUELENGTH parts, which have long been a key underlying feature of data.table internals. I propose we try and keep using those unless CRAN really forces our hands CRAN is really forcing our hands :). It be really great if someone wanted to do a performance comparison of the package with/without those calls.

Anyway, it should be easier to remove the non-API S4 calls.

The list does keep changing as this is an actively evolving piece of r-devel. Here's the current list of things we should (eventually) address:

• LEVELS
• OBJECT
• Rf_GetOption
• NAMED
• SET_S4_OBJECT
• SET_TYPEOF
• STRING_PTR
• UNSET_S4_OBJECT
• isFrame
• SETLENGTH
• SET_TRUELENGTH
• SET_GROWABLE_BIT
• TRUELENGTH

[tdhock]

there is a web page which lists all of the API calls, grouped into 3 categories
https://yutannihilation.github.io/R-fun-API/
see also the R-devel thread which is linked from that page

[ecoRoland2]

You should probably be proactive and contact R-core regarding your need of an API to SETLENGTH and SET_TRUELENGTH. I would hate seeing slower performance in data.table. It could cause some difficult decisions at my day job.

[MichaelChirico]

if this is important for you (or anyone else reading) please take the time to construct a performance comparison.

[ecoRoland2]

I would but unfortunately I think you need at least basic knowledge of C to do that.

[ben-schwen]

AFAIU we can replace SETLENGTH calls with Rf_lengthgets calls and reassignment. But there is no safe way to handle SET_TRUELENGTH right?

[tdhock]

Where is SET_TRUELENGTH used? https://github.com/search?q=repo%3ARdatatable%2Fdata.table%20SET_TRUELENGTH&type=code
one place is in rbindlist which is a pretty important function for performance.
@Anirban166 can you please construct a PR with a single change SET_TRUELENGTH -> Rf_lengthgets in rbindlist along with an atime performance test?

[ben-schwen]

Where is SET_TRUELENGTH used?

We use it as part of memrecycle and subsetting so its basically in most operations :D

But we are not alone with this problem r-lib/vctrs#1933

[Anirban166]

@Anirban166 can you please construct a PR with a single change SET_TRUELENGTH -> Rf_lengthgets in rbindlist along with an atime performance test?

Sure, but I'm wondering what the tests will be focused on (since this affects multiple functions) and then what about SETLENGTH?

[MichaelChirico]

AFAIU we can replace SETLENGTH calls with Rf_lengthgets calls and reassignment.

I don't see lengthgets mentioned in WRE.

By the same definition, touching S4 objects from C is also non-API! 🤔

[MichaelChirico]

I see, lengthgets is just what length<- calls under the hood. But that induces a copy:

a = 1:10
address(a)
# [1] "0x5651b8564b30"
length(a) <- 20
address(a)
# [1] "0x5651b8ab5c68"

[MichaelChirico]

It looks like more of our usage has been marked as non-API in the meantime:

Result: NOTE 
  File ‘data.table/libs/data_table.so’:
    Found non-API calls to R: ‘LEVELS’, ‘NAMED’, ‘SETLENGTH’,
      ‘SET_GROWABLE_BIT’, ‘SET_S4_OBJECT’, ‘SET_TRUELENGTH’,
      ‘UNSET_S4_OBJECT’

new:

LEVELS():

data.table/src/data.table.h

Lines 30 to 32 in fab93a9

	#define IS_UTF8(x) (LEVELS(x) & 8)
	#define IS_ASCII(x) (LEVELS(x) & 64)
	#define IS_LATIN(x) (LEVELS(x) & 4)

data.table/src/bmerge.c

Line 20 in fab93a9

#define ENC_KNOWN(x) (LEVELS(x) & 12)

NAMED():

data.table/src/data.table.h

Lines 58 to 63 in fab93a9

	#ifndef MAYBE_SHARED
	# define MAYBE_SHARED(x) (NAMED(x) > 1)
	#endif
	#ifndef MAYBE_REFERENCED
	# define MAYBE_REFERENCED(x) ( NAMED(x) > 0 )
	#endif

data.table/src/assign.c

Lines 548 to 552 in fab93a9

	i+1, NAMED(thisvalue), MAYBE_SHARED(thisvalue), length(values), length(cols));
	}
	thisvalue = copyAsPlain(thisvalue); // PROTECT not needed as assigned as element to protected list below.
	} else {
	if (verbose) Rprintf(_("Direct plonk of unnamed RHS, no copy. NAMED==%d, MAYBE_SHARED==%d\n"), NAMED(thisvalue), MAYBE_SHARED(thisvalue)); // e.g. DT[,a:=as.character(a)] as tested by 754.5

SET_GROWABLE_BIT():

data.table/src/data.table.h

Lines 10 to 12 in fab93a9

	#if !defined(R_VERSION) || R_VERSION < R_Version(3, 4, 0)
	# define SET_GROWABLE_BIT(x) // #3292
	#endif

data.table/src/freadR.c

Line 535 in fab93a9

SET_GROWABLE_BIT(VECTOR_ELT(DT,i)); // #3292