fix use of non-API calls to R's C code

[paciorek]

CRAN is now flagging this issue with a deadline of April 9.

Version: 1.4.1
Check: compiled code
Result: NOTE
  File ‘nimble/libs/nimble.so’:
    Found non-API calls to R: ‘Rf_findVar’, ‘Rf_findVarInFrame’
  
  Compiled code should not call non-API entry points in R.
  
  See ‘Writing portable packages’ in the ‘Writing R Extensions’ manual,
  and section ‘Moving into C API compliance’ for issues with the use of
  non-API entry points.

From quick glance at the docs, it looks like we can replace those non-API calls. The docs say "Use R_getVar or R_getVarEx, added in R 4.5.0. In some cases using eval may suffice. "

Rcpp discussed this, making the comment about using the old code for older R versions. I guess we need to think about whether we are ok with having nimble 1.4.2 require R>=4.5.0. Probably?

[paciorek]

Before we submit to CRAN, I want to (and will) check Rcpp source code to see if the old calls exist there, as a way of trying to determine if it's ok to have them for use with older R versions.

[perrydv]

They old calls do appear, but remember they are giving code used in on-the-fly compilation, not in a package DLL, I think.

Potentially a relevant test would be to build a binary with R > 4.5.0 and then install and use it with R < 4.5.0.

[paciorek]

We submitted to CRAN and they now report additional non-API calls:

"""
Dear maintainer,

package nimble_1.4.2.tar.gz does not pass the incoming checks automatically, please see the following pre-tests (additional issue checks):
Windows: https://win-builder.r-project.org/incoming_pretest/nimble_1.4.2_20260327_175144/Windows/00check.log
Status: 1 NOTE
Debian: https://win-builder.r-project.org/incoming_pretest/nimble_1.4.2_20260327_175144/Debian/00check.log
Status: 1 NOTE

Last released version's CRAN status: OK: 6, NOTE: 4, WARNING: 4
See: https://cran.r-project.org/web/checks/check_results_nimble.html

Last released version's additional issues:
rchk https://raw.githubusercontent.com/kalibera/cran-checks/master/rchk/results/nimble.out

CRAN Web: https://cran.r-project.org/package=nimble

Please fix all problems and resubmit a fixed version via the webform.
If you are not sure how to fix the problems shown, please ask for help on the R-package-devel mailing list:
https://stat.ethz.ch/mailman/listinfo/r-package-devel
If you are fairly certain the rejection is a false positive, please reply-all to this message and explain.

More details are given in the directory:
https://win-builder.r-project.org/incoming_pretest/nimble_1.4.2_20260327_175144/
The files will be removed after roughly 7 days.

*** Strong rev. depends ***: baorista BayesNSGP BayesSIM compareMCMCs DPCD InvStablePrior ivd MultiNMix nimbleAPT nimbleCarbon nimbleEcology nimbleHMC nimbleMacros nimbleNoBounds nimbleQuad nimbleSCR nimbleSMC nimblewomble scPloidy spatialAtomizeR Xcertainty

Best regards,
CRAN teams' auto-check service
Flavor: r-devel-windows-x86_64
Check: compiled code, Result: NA
File 'nimble/libs/x64/nimble.dll':
Found non-API calls to R: 'R_MissingArg', 'R_UnboundValue'

Compiled code should not call non-API entry points in R.

See 'Writing portable packages' in the 'Writing R Extensions' manual,
and section 'Moving into C API compliance' for issues with the use of
non-API entry points.

Flavor: r-devel-windows-x86_64
Check: Overall checktime, Result: NOTE
Overall checktime 12 min > 10 min

Flavor: r-devel-linux-x86_64-debian-gcc
Check: compiled code, Result: NOTE
File 'nimble/libs/nimble.so':
Found non-API calls to R: 'R_MissingArg', 'R_UnboundValue'

Compiled code should not call non-API entry points in R.

See 'Writing portable packages' in the 'Writing R Extensions' manual,
and section 'Moving into C API compliance' for issues with the use of
non-API entry points.
"""

[paciorek]

I tried to run checking under the "rchk" option of rbut but I got a different issue, which seems quite odd. I'm tempted to dismiss it as an anomaly since it's not happening in other contexts, but in any event it is preventing us from seeing if rhub would show us all the non-API calls. (actually see next comment -- this seems like a real issue)

9 warnings generated.
wllvm++ -stdlib=libc++ -std=gnu++20 -I"/opt/R/devel-rchk/lib/R/include" -DNDEBUG -DR_NO_REMAP -I../inst/include    -I/usr/local/include    -fpic  -Wall -g -O0 -fPIC -I/usr/include/libcxxabi  -c nimbleGraph.cpp -o nimbleGraph.o
wllvm++ -stdlib=libc++ -std=gnu++20 -shared -L/usr/local/lib -o nimble.so RcppUtils.o dists.o eigenUsingClasses.o nimble.o nimbleGraph.o -L/opt/R/devel-rchk/lib/R/lib -lRlapack -L/opt/R/devel-rchk/lib/R/lib -lRblas -lgfortran -lm -lquadmath
installing to /opt/R/devel-rchk/packages/lib/nimble/libs
* DONE (nimble)
objdump: Warning: Unrecognized form: 0x23
objdump: Warning: Unrecognized form: 0x23
<snip>
objdump: Warning: Unrecognized form: 0x23
objdump: Warning: Unrecognized form: 0x22
Running bcheck
==== rchk bcheck =========================================
ERROR: too many states (abstraction error?) in function strptime_internal
ERROR: too many states (abstraction error?) in function bcEval_loop
ERROR: too many states (abstraction error?) in function RunGenCollect

Function C_getConditionallyIndependentSets
  [PB] has too high protection stack depth results will be incomplete
  [UP] protect stack is too deep, unprotecting all variables, results will be incomplete
  [PB] has an unsupported form of unprotect (not constant, not variable), results will be incomplete /__w/comfortable-birthstone-zebratailedlizard-nimble/comfortable-birthstone-zebratailedlizard-nimble/src/nimbleGraph.cpp:679
  [UP] unsupported form of unprotect, unprotecting all variables, results will be incomplete /__w/comfortable-birthstone-zebratailedlizard-nimble/comfortable-birthstone-zebratailedlizard-nimble/src/nimbleGraph.cpp:679

Function getSEXPdims(SEXPREC*)
  [UP] calling allocating function SEXP_2_vectorInt(SEXPREC*, int) with argument allocated using Rf_getAttrib(V,S:dim) /__w/comfortable-birthstone-zebratailedlizard-nimble/comfortable-birthstone-zebratailedlizard-nimble/src/RcppUtils.cpp:52
Analyzed 9662 functions, traversed 22266 states.
------------------------------------------------------
Run # Fail for rchk errors.

[paciorek]

A couple other comments:

1. I am now seeing that the nimble checks published on CRAN do mention R_MissingArg, which we apparently hadn't noticed, I guess because the eamil from CRAN didn't mention it. But the checks did not mention R_UnboundValue.
2. The "rchk" link on the nimble checks published on CRAN also mention the issue seen in the rhub running of rchk, so actually that does seem like a legitimate issue:

Package nimble version 1.4.1
Package built using 89420/R 4.6.0; x86_64-pc-linux-gnu; 2026-02-16 13:59:43 UTC; unix   
Checked with rchk version 35618ebbccf3cd0b45a3530e6303970a22a9056b LLVM version 14.0.6
More information at https://github.com/kalibera/cran-checks/blob/master/rchk/PROTECT.md
For rchk in docker image see https://github.com/kalibera/rchk/blob/master/doc/DOCKER.md

Function getSEXPdims(SEXPREC*)
  [UP] calling allocating function SEXP_2_vectorInt(SEXPREC*, int) with argument allocated using Rf_getAttrib(V,S:dim) /unknown:0

[paciorek]

I'm looking in r-devel and r-package-devel about R_MissingArg and R_UnboundValue.

As of 30 March 26 (!) Luke Tierney says that R_MissingArg should no longer generate a check NOTE.

[perrydv]

I asked (paid) claude.ai through VS Code this query and attach its report below. I may add comments as I look through it and compare to WRE. This is done on fix-findVar-API-calls. Prompt:
"Please find all uses of non-API R internals such as Rf_findVar or R_UnboundValue that might be hiding in inst/CppCode, inst/include/nimble, or src. CRAN now has automated checking for these as it becomes strict about R's API for internal C (and fortran) calls."

Report (updated after further input and better format):

Audit: Non-API R Internals in nimble C/C++ Code

Context

CRAN has tightened enforcement of R's official C API, particularly starting with R 4.5. Packages that call internal R functions or rely on symbols not in the official API (documented in Writing R Extensions) now get flagged by automated CRAN checks. The branch fix-findVar-API-calls already addressed Rf_findVar/Rf_findVarInFrame via version-conditional compilation, but several other non-API internals remain.

---

Findings by Severity

TIER 1 — Already addressed (verify, don't break)

Rf_findVar / Rf_findVarInFrame — wrapped in version guards in inst/include/nimble/Utils.h:41–75

• R < 4.5.0: calls Rf_findVarInFrame(Senv, Ssym) and Rf_findVar(Senv, Ssym) — these are the functions CRAN is newly flagging
• R ≥ 4.5.0: uses R_getVarEx(sym__, Senv, FALSE, R_UnboundValue) and R_getVar(sym__, Senv, TRUE) — the new official replacements
• Status: Correct approach; the version guard preserves backward compatibility. No action needed unless CRAN also flags the R_getVarEx path (see Tier 2).

---

TIER 2 — Likely flagged by CRAN (need investigation/fix)

R_UnboundValue + R_getVarEx → replace both with R_getVar

• Location: inst/include/nimble/Utils.h:65
• Current code: R_getVarEx(sym__, Senv, FALSE, R_UnboundValue) — returns R_UnboundValue when the variable is not found in the frame
• Fix: R_getVar(sym__, Senv, FALSE) — R_getVar already returns R_UnboundValue when not found (no inheritance), so R_getVarEx and the R_UnboundValue sentinel are both unnecessary. This makes NIM_FINDVARINFRAME identical in structure to NIM_FINDVAR (which already uses R_getVar).
• Result: Eliminates both R_getVarEx and R_UnboundValue from the R ≥ 4.5 branch.

R_MissingArg

• Location: inst/CppCode/RcppUtils.cpp:865 — used in varAndIndices_2_LANGSXP() to represent blank indices in expressions like A[i,]
• Status: As of 2026-03-30, R-devel list confirmed R_MissingArg will be placed back on the allowed (non-flagged) list. No action needed.

R_BracketSymbol

• Location: inst/CppCode/RcppUtils.cpp:861 — SETCAR(t, R_BracketSymbol)
• Status: In <Rinternals.h>, not in the formal API. Alternative: Rf_install("[").

SET_TAG()

• Location: inst/CppCode/nimOptim.cpp:270–289 — sets named argument tags on LANGSXP for the custom_optim call (9 uses)
• Status: SET_TAG is in <Rinternals.h>. It does not have a stable Rf_-prefixed replacement. This is used to pass named arguments in a language object. This is one of the harder ones to fix.

---

TIER 3 — Widespread internal macros (less urgent, watch for future flagging)

These macros come from <Rinternals.h> and <Rdefines.h> and are used pervasively. CRAN has not (yet) flagged these as errors for typical packages, but they are technically internal:

SETCAR(), SETCADR(), SETCADDR(), CDR()

Used for constructing LANGSXP (language call) objects:

• inst/include/nimble/RcppUtils.h:52–53 — macro SET_NEXT_LANG_ARG
• inst/CppCode/RcppUtils.cpp:845–909 — makeAsNumeric_LANGSXP, varAndIndices_2_LANGSXP, makeParsedVarList, makeNewNimbleList
• inst/CppCode/nimOptim.cpp:268–291
• These are the standard C API for pairlist/LANGSXP construction and are widely used; unlikely to be flagged soon.

GET_SLOT()

Used for S4 .xData slot access (accessing environments inside reference objects):

• inst/include/nimble/accessorClasses.h:805,809,813,817,828,849,870,891 — 8 uses via NIM_FINDVARINFRAME(GET_SLOT(...), ...)
• inst/include/nimble/RcppNimbleUtils.h:35
• Alternative: R_do_slot(obj, Rf_install(".xData")) — this is in the official API.

REAL(), INTEGER(), LOGICAL() (mutable pointer access)

• inst/include/nimble/RcppNimbleUtils.h:251,256,275,280,298,303,314,320,334,341,355,362
• These are still in the official API. For R ≥ 4.2, read-only access should use REAL_RO() etc., but REAL() for write access remains valid.

VECTOR_ELT(), SET_VECTOR_ELT(), STRING_ELT(), SET_STRING_ELT()

• Used throughout accessorClasses.h, RcppNimbleUtils.h, RcppUtils.cpp, predefinedNimbleLists.cpp, nimOptim.cpp
• These are in the official API. No action needed.

TYPEOF(), CHAR(), LENGTH()

• Used throughout. These are in the official API. No action needed.

---

TIER 4 — Headers (low urgency, but worth noting)

#include <Rinternals.h> appears in 20 header files under inst/include/nimble/. This header exposes all internal structures. While CRAN doesn't currently flag the include itself, using only the official API subset from it is best practice. Files:

• Utils.h, RcppUtils.h, accessorClasses.h, NamedObjects.h, dllFinalizer.h, nimbleGraph.h, predefinedNimbleLists.h, nimDerivs_*.h (13 files)

#include "Rdefines.h" in inst/include/nimble/RcppUtils.h:32 — this is a legacy compatibility header; <Rinternals.h> already included separately covers everything it provides.

---

Priority Action Items

Priority	Symbol/Function	File(s)	Fix
High	R_BracketSymbol	RcppUtils.cpp:861	Replace with Rf_install("[")
N/A	R_MissingArg	RcppUtils.cpp:865	Confirmed allowed on R-devel list (2026-03-30); no action needed
Medium	SET_TAG()	nimOptim.cpp:270–289	No direct API replacement; may need Rcpp or Rf_cons-based approach
High	R_UnboundValue + R_getVarEx	Utils.h:65	Replace R_getVarEx(sym, env, FALSE, R_UnboundValue) with R_getVar(sym, env, FALSE)
Low	GET_SLOT()	accessorClasses.h, RcppNimbleUtils.h	Replace with R_do_slot()
Low	#include "Rdefines.h"	RcppUtils.h:32	Remove (redundant with <Rinternals.h>)

---

Verification

After any changes:

1. Run R CMD check --as-cran — check for NOTE: Found non-API calls or similar
2. Run existing test suite: devtools::test() or R CMD check
3. Test on R ≥ 4.5 specifically (the version where new API enforcement began)
4. Check CRAN's check_r_package_url output for the package if available
5. Cross-reference R 4.5 API changes in NEWS and WRE §6.13

[perrydv]

Some relevant background links:

• June 2025 r-devel posts by Luke Tierney
• Writing R Extensions Moving into C API Compliance. (Unfortunately this is not a complete guide.)
• 4.5.0 NEWS See "CHANGES IN R 4.5.0 C-LEVEL API"

[paciorek]

A bit of progress on the Rhub side of things. I ran the gcc16 platform (option number 22) and with that I was able to see the NOTEs about non-API calls:

* checking compiled code ... NOTE
File ‘nimble/libs/nimble.so’:
  Found non-API calls to R: ‘R_MissingArg’, ‘R_UnboundValue’
Compiled code should not call non-API entry points in R.

So we at least have a way to check things.

I also see Rcpp changed to using R_UnboundValue as of 2025-12-16.