Fix R CRAN failures.

[trivialfis]

Related: #7260 (comment)

• Remove hist builder dtor.
  I did not reproduce the warning using clang-14, might be missing some flags in my config or it's just a bug in clang-12 used by CRAN test. Anyway, this is a cleanup that removes unnecessary code.
• Initialize some variables.

[hetong007]

Thanks. May I suggest to hold this PR until we see all the checks for 1.5.0.1 on this page: https://cran.r-project.org/web/checks/check_results_xgboost.html. It might take a few days, and we have until Nov 22nd to patch.

[hetong007]

There are additional issues with valgrind: https://www.stats.ox.ac.uk/pub/bdr/memtests/valgrind/xgboost/00check.log:

• An R test failure: Failure (test_helpers.R:231:11): xgb-attribute numeric precision
• Some leaks in the end of the check log.

I didn't remember CRAN has complained about valgrind issues (let's see if they do this time), but if it is fixable on our side maybe we can also try?

[trivialfis]

An R test failure: Failure (test_helpers.R:231:11): xgb-attribute numeric precision

• That's just testing the as.numeric in R. XGBoost is not involved. Will relax the test. A longer-term solution might be using JSON instead.

I did not reproduce the error using valgrind. The patch simply adds tolerance.

at 0x1E1906EF: xgboost::data::SimpleDMatrix::GetGradientIndex(xgboost::BatchParam const&) (packages/tests-vg/xgboost/src/amalgamation/../src/data/simple_dmatrix.cc:94)

• Probably the gpu_id, which is not used in that context. Will fix.

0x1E21680E: xgboost::DMatrix* xgboost::DMatrix::Createxgboost::data::DenseAdapter(xgboost::data::DenseAdapter*, float, int, std::__cxx11::basic_string<char, std::char_traits, std::allocator > const&) (packages/tests-vg/xgboost/src/amalgamation/../src/data/data.cc:828)

Need to find out what's not initialized.

==2937178== by 0x1E148116: XGDMatrixCreateFromCSC_R (packages/tests-vg/xgboost/src/xgboost_R.cc:137)

Leak in libgomp? Not sure if there's anything we can do.

XGDMatrixSetInfo_R (packages/tests-vg/xgboost/src/xgboost_R.cc:203)

std::vector should manage the memory as RAII. Not sure if there's anything we can do.

[trivialfis]

Ran valgrind with Rscript testthat.R. No error or warning from valgrind.

valgrind: 3.15.0
Rscript: 3.6.3
gcc: 9.3.0

[trivialfis]

@hcho3 Do you think we can free some budget for running valgrind with gtest on CI once we remove CUDA10?

[hcho3]

@trivialfis We can just use GitHub Actions for this?

[trivialfis]

But then we will skip all GPU code path.

[hcho3]

@trivialfis I don't think CRAN tests are run for any GPU code path

[trivialfis]

running valgrind with gtest

I want to run valgrind with c++ tests. But there are some warnings whenever GPU code is executed, so we might just limit it to CPU only for now:

==304308== Warning: noted but unhandled ioctl 0x19 with no size/direction hints.
==304308==    This could cause spurious value errors to appear.
==304308==    See README_MISSING_SYSCALL_OR_IOCTL for guidance on writing a proper wrapper.
==304308== Warning: set address range perms: large range [0x10006000000, 0x10106000000) (noaccess)
==304308== Warning: noted but unhandled ioctl 0x49 with no size/direction hints.
==304308==    This could cause spurious value errors to appear.
==304308==    See README_MISSING_SYSCALL_OR_IOCTL for guidance on writing a proper wrapper.
==304308== Warning: noted but unhandled ioctl 0x21 with no size/direction hints.
==304308==    This could cause spurious value errors to appear.
==304308==    See README_MISSING_SYSCALL_OR_IOCTL for guidance on writing a proper wrapper.
==304308== Warning: noted but unhandled ioctl 0x1b with no size/direction hints.
==304308==    This could cause spurious value errors to appear.
==304308==    See README_MISSING_SYSCALL_OR_IOCTL for guidance on writing a proper wrapper.
==304308== Warning: noted but unhandled ioctl 0x44 with no size/direction hints.
==304308==    This could cause spurious value errors to appear.
==304308==    See README_MISSING_SYSCALL_OR_IOCTL for guidance on writing a proper wrapper.

[trivialfis]

I think this PR should contain all the fixes in XGBoost.

[trivialfis]

@hetong007 I looked into the 2 memory leak reports, first one is from libgomp, the second one is from the std::vector implementation in GCC. I don't know how to fix them and so far haven't been able to reproduce them using my own setup. Any chance we can continue the submission?

[hetong007]

How about this: let's wait for all results from different platforms in the check page, and fix whatever we can, then submit a patch (1.5.0.2). After this patch, let's see how CRAN may complaint further.

[trivialfis]

Got it. Thanks for guiding the release.

[hetong007]

I see most CRAN results are available with no new issue. I'm merging this PR.

[trivialfis]

Thanks, I will backport it and bump the version

[hetong007]

Seems CRAN checks are all finished. Please instruct when to submit (before 22nd).

[trivialfis]

I will run through all the tests again today

[trivialfis]

@hetong007 Let me know if there's anything else I can do to help.

[hetong007]

It is undergoing checks by the CRAN team. Will update here.

[hetong007]

On CRAN now. From the status page I see the row with r-devel-linux-x86_64-fedora-clang has its previous WARN fixed. Appreciated!

[trivialfis]

Excellent!