Investigation: conda-forge scikit-learn slower than PyPI (ExtraTrees case) #34869

sklearn/tree/_partitioner.pyx (DensePartitioner.find_min_max, .partition_samples) and sklearn/tree/_utils.pxd call isnan(value) on float32_t data, via from libc.math cimport isnan. This runs once per sample per feature scan during fit; never during predict.

The feedstock's CI pin file (.ci_support/linux_64_python3.12.____cpython.yaml at the 1.9.0 release commit) sets c_stdlib: sysroot, c_stdlib_version: '2.17'. That sysroot's math.h has:

#  define isnan(x) \
     (sizeof (x) == sizeof (float) ? __isnanf (x) \
      : sizeof (x) == sizeof (double) ? __isnan (x) : __isnanl (x))

Once the preprocessor expands this, the identifier isnan is gone from
the translation unit — GCC has nothing left to recognize as a builtin, so
__isnanf (an ordinary external glibc symbol, not a compiler intrinsic)
gets a real out-of-line call. Modern glibc instead defines isnan(x) as
__builtin_isnan(x) directly, which GCC folds to 2 instructions
(ucomiss/setp). PyPI/manylinux wheels don't hit this because they're
built with devtoolset-style toolchains: modern compiler and modern
headers, achieving old-system compatibility via glibc symbol versioning
at link time rather than by literally compiling against old headers.

Confirmed end to end:

• perf-profiled real installed PyPI vs conda-forge packages
  (ExtraTreesRegressor.fit, n_features=30): conda-forge spends 12.6% of
  samples in __isnanf alone; PyPI doesn't show it in the top 15 at all.
• Disassembled both real .so files directly: conda-forge's
  find_min_max has call *0x1791f(%rip) # <__isnanf@GLIBC_2.2.5>;
  PyPI's has no call at all.
• Reproduced from source: rebuilding the actual v1.9.0 tag with the exact
  matching GCC (14.3.0-19, confirmed via the real binary's own
  .comment section) reproduced nothing until sysroot_linux-64=2.17
  was pinned explicitly (every environment had otherwise resolved the
  latest default, 2.39) — at which point the non-inlined call and the
  larger function size reappeared byte-for-byte.
• Same isnan() call pattern, same functions, confirmed still present on
  current main (the categorical-feature/Breiman-shortcut work added
  since 1.9.0 didn't touch this code path — find_min_max compiles to the
  identical byte size on both).

• -O2 vs -O3 in isolation (~1.7% effect — this is the Meson bug above)
• -march=nocona -mtune=haswell vs no override
• GCC version: 16.2.0, 13.4.0, and the exact 14.3.0-19 patch build
  (pinning gcc_impl_linux-64/gxx_impl_linux-64/binutils_impl_linux-64
  precisely, since gcc_impl_linux-64 has its own build-number sequence
  independent of the gcc_linux-64 wrapper)
• Cython version (real binary uses 3.2.5; early rebuilds had picked up
  3.3.0 by default — controlled for, no effect)
• Exact v1.9.0 source tag vs current main
• Editable install vs literal python -m build -w wheel build (matching
  build.sh exactly)
• -fno-merge-constants (the extra flag conda-forge adds under
  CONDA_BUILD=1)
• joblib version, core-count detection, BLAS threading config — all
  identical between real PyPI and conda-forge installs