perf: speed up subnormal float parsing - #1237

Summary

Speed up JSON and lax string-to-float decoding of subnormal numbers without changing their IEEE-754 results.

Problem

#1232 reports extremely slow decoding of [5e-324] * 65_536. Exponents below -307 currently go straight to the exact high-precision decimal parser. On this Linux x86-64 / CPython 3.11.15 / GCC 15 host, the reported array takes about 1.075 seconds on current main.

Solution

• Extend the generated powers-of-ten table to -342 (35 additional entries), retaining the existing generator and attribution.
• Round tiny values directly in units of 2**-1074. Bound the approximate product by an interval less than two high-word units wide; if that interval touches a halfway boundary, keep the exact fallback. This avoids double rounding, handles underflow and carry into the normal range, and guards all 64-bit shifts.
• Retain the existing truncated-mantissa endpoint check, exact HPD fallback, overflow behavior, Decimal/float_hook paths and integer behavior.
• Add 530 parameterized correctness cases covering both signs, signed zero, table boundaries, exact halfway neighbours, JSON numeric keys and lax conversion. Add five CodSpeed benchmark inputs, including ordinary-float controls; no wall-clock threshold is added to CI.

Local medians from seven baseline/fixed rounds, reversing their order every round; each array has 65,536 values and outputs are checked bit-for-bit outside the timed region:

The ordinary controls changed by +0.35%, +0.14%, and -0.88% respectively in these medians. These are local measurements, not cross-platform guarantees. Ambiguous and very long decimal strings may still use HPD; this is not a general parser rewrite.

Testing

All builds/tests ran offline in a credential-free Bubblewrap namespace with environment, filesystem, network and source-import canaries. Equivalent underlying commands from the just recipes were invoked directly using the prepared dependency layer:

• python setup.py build_ext --inplace --force — passed; the same pre-existing _POSIX_C_SOURCE redefinition warning appears on main.
• python scripts/generate_atof_consts.py — generated the checked-in table.
• pytest -o testpaths=tests/unit -p no:cacheprovider --basetemp=/tmp/pytest -q -rs — 6,987 passed, 156 skipped. Pristine main with the same final dependency layer: 6,457 passed, 156 skipped. Skips are existing Python-version/optional-feature guards.
• pytest --doctest-modules --pyargs msgspec -p no:cacheprovider --basetemp=/tmp/pytest -q — 7 passed.
• ruff check, ruff format --check, python -m codespell_lib — passed (Ruff 0.16.9, Codespell 2.4.3). Unchanged main also passes these hooks.
• mypy tests/typing, pyright --pythonpath <resolved CPython executable> tests/typing, pyrefly check tests/typing, python scripts/stubtest.py — passed. Pyrefly emits the expected PYTHONPATH-environment notice; no type diagnostics. Explicit Pyright runtime selection avoids a sandbox interpreter-probe error.
• pytest tests/prof/perf/test_encode_decode.py -k 'decode_small_floats and amd' --codspeed --codspeed-mode walltime --codspeed-warmup-time 0.1 --codspeed-max-time 0.2 -p no:cacheprovider --basetemp=/tmp/pytest -q — 5 passed. This uses current pytest-codspeed 5.0.3 flags; the legacy --benchmark-name / --benchmark-disable-gc recipe flags were not accepted by this installation.
• External numerical harness: 6,317,572 mantissa/exponent pairs against strtod, compiled with -fsanitize=undefined -fno-sanitize-recover=all; repeated with -U__SIZEOF_INT128__ to exercise the portable multiplication path. Both passed: 6,312,610 accepted approximations, 4,962 exact-fallback requests, no mismatches or sanitizer findings.
• External Python differential probe: 60,000 signed decimal strings, including long/truncated mantissas, checked against float(text).hex() through JSON and lax conversion — passed.
• Performance RED/GREEN/sabotage: a local diagnostic asserting subnormal parsing is below 20x the ordinary-float control failed on original code (289x), passed with this change (1.10x), failed again against the preserved original build (277x), and passed after restoring the fixed build (1.10x). Correctness tests intentionally also pass on main: this fixes performance, not rounding semantics. The diagnostic timing threshold is not in CI.
• Independent read-only AI diff review passed with no security or logic findings. Its suggestion for direct fast-path coverage was exercised by the external native harness.

Not run locally: non-Linux platforms, other Python versions, full Sphinx docs build (no docs changed), and hosted CI. Maintainer review and CI remain required; no merge is requested outside normal controls.

Related Issue

Fixes #1232.

Automation disclosure

Implemented and validated by Hermes DEV, an autonomous coding agent working for arindamsikder, with a separate AI reviewer. No human pre-review is claimed.