Add python 3.15, small fixes, update vendored packages

.github/workflows/ci.yml

@@ -18,23 +18,16 @@ jobs:
       fail-fast: false
       matrix:
         os: [ubuntu-latest, macos-14, windows-latest]
-        python: ["3.12", "3.13", "3.14", "3.14t", "3.15", "3.15t"]
-        exclude:
-          # 3.15 is still pre-release (alpha at time of writing) and the
-          # free-threaded Windows binary distributed via setup-python's
-          # python-versions repo has flaked here at least once
-          # (concurrent jobs racing for the same download). Re-add this
-          # cell once 3.15t has shipped a stable Windows release; until
-          # then ubuntu and macos provide free-threaded 3.15 coverage.
-          - os: windows-latest
-            python: "3.15t"
+        python: ["3.13", "3.14", "3.14t", "3.15", "3.15t"]
     steps:
       - uses: actions/checkout@v5
 
       - uses: actions/setup-python@v6
         with:
+          # Every entry is a final release (3.15.0 shipped 2026-10-09;
+          # setup-python's manifest carries 3.15 and 3.15t for all three
+          # OSes), so no allow-prereleases.
           python-version: ${{ matrix.python }}
-          allow-prereleases: true
 
       - name: Install build deps and package
         run: |
@@ -53,3 +46,40 @@ jobs:
         run: |
           python -m pytest tests/cpython_suite --tb=short || true
         shell: bash
+
+  test-all-backends:
+    # Every optional backend (fmt, ryu_opt, zmij, wuffs, ffc, stock) compiled
+    # in, so tests/test_backends.py can check each one against stock.
+    # The wheels only ship the defaults; this job is what keeps the
+    # others honest — and, on all three OSes, what proves the pure-C
+    # zmij.c / zmij_dtoa.c build with MSVC and Apple clang as well.
+    name: test / all backends / py3.14 / ${{ matrix.os }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-14, windows-latest]
+    steps:
+      - uses: actions/checkout@v5
+
+      - uses: actions/setup-python@v6
+        with:
+          python-version: "3.14"
+
+      - name: Install with every backend compiled in
+        # bash on every OS: the default Windows shell is PowerShell, where
+        # the backslash continuations below are not line continuations.
+        shell: bash
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install ".[test]" \
+            -C cmake.define.FLOATIUM_FORMAT_BACKEND=all \
+            -C cmake.define.FLOATIUM_PARSE_BACKEND=all
+
+      - name: Show info
+        run: |
+          python -c "import floatium; print(floatium.info())"
+
+      - name: Run unit + parity + edge + backend tests
+        run: |
+          python -m pytest tests -x --tb=short --ignore=tests/cpython_suite

.github/workflows/cpython-suite.yml

@@ -34,7 +34,6 @@ jobs:
       - uses: actions/setup-python@v6
         with:
           python-version: ${{ matrix.python }}
-          allow-prereleases: true
 
       - run: python -m pip install --upgrade pip && python -m pip install .
 

.github/workflows/wheels.yml

@@ -17,7 +17,7 @@ jobs:
     steps:
       - uses: actions/checkout@v5
       - uses: actions/setup-python@v6
-        with: { python-version: "3.12" }
+        with: { python-version: "3.13" }
       - run: |
           python -m pip install --upgrade pip build
           python -m build --sdist --outdir dist
@@ -40,18 +40,16 @@ jobs:
     steps:
       - uses: actions/checkout@v5
 
-      - uses: pypa/cibuildwheel@v3.4.1
+      - name: Build wheels (cp313, cp314, cp314t, cp315, cp315t)
+        uses: pypa/cibuildwheel@v4.3.1
         env:
           # Build config lives in [tool.cibuildwheel] in pyproject.toml.
-          # cpython-freethreading is still required in cibuildwheel
-          # 3.4.1 for the cp313t / cp314t selectors — without it they
-          # fail with "Invalid build selector ... wasn't enabled". The
-          # token emits a deprecation warning (it is being renamed in a
-          # future cibuildwheel) but still works and is mandatory here.
-          # No cpython-prerelease: the build list is all released
-          # CPython (3.15 is omitted — 3.4.1 has no config for it).
+          # cibuildwheel 4.x: cp315/cp315t are in the default catalog
+          # (3.15.0 final since 4.3.1) and free-threaded 3.14+ builds
+          # need no `enable` token — the old cpython-freethreading enable
+          # was removed in 4.0.0 along with the 3.13t builds (which is
+          # why there are no cp313t wheels from 0.15.0 on).
           CIBW_BUILD_VERBOSITY: 1
-          CIBW_ENABLE: cpython-freethreading
 
       - uses: actions/upload-artifact@v5
         with:

BENCHMARKS.md

@@ -2,18 +2,20 @@
 
 Nanoseconds per operation for `repr(float)`, fixed-precision f-string
 formatting, and `float(str)`, comparing stock CPython against every
-floatium format backend (`fmt_opt`, `fmt`, `ryu_opt`) and every parse
-backend (`fast_float`, `wuffs`, `stock`).
+floatium format backend (`fmt_opt`, `fmt`, `ryu_opt`, `zmij`) and every parse
+backend (`fast_float`, `wuffs`, `ffc`, `stock`).
 
 ## Methodology
 
-- **floatium:** 0.14.2, release build, all backends compiled in.
+- **floatium:** 0.16.0 (zmij a9d6363, ffc.h v26.09.01, fmt 12.2.0,
+  fast_float v8.3.2), release build, all backends compiled in.
 - **CPython:** 3.14.3.
-- **Date:** 2026-05-16.
+- **Date:** 2026-10-19.
+- **Script:** `FLOATIUM_AUTOPATCH=0 python -m bench.bench_backend_matrix`.
 - **Corpora:** the five canonical corpora from `bench/corpora.py` —
   `random_uniform`, `random_bits`, `financial`, `scientific`,
   `integer_valued`; 2000 values each.
-- **Timing:** mirrors `bench/bench_ns_per_op.py` — 9 outer samples per
+- **Timing:** same as `bench/bench_ns_per_op.py` — 9 outer samples per
   cell, each timing 5 inner loops over the whole corpus; the reported
   figure is the median of the fastest third (3 of 9), robust to
   background jitter without picking the cold-cache outlier.
@@ -28,30 +30,38 @@ backend (`fast_float`, `wuffs`, `stock`).
 ## Format backends
 
 `repr(x)` and `f"{x:.4f}"` over each corpus. Speedup is `stock / backend`;
-values above 1.00x are faster than stock.
-
-| Corpus | Operation | Stock (ns) | fmt_opt | fmt | ryu_opt | fmt_opt speedup | fmt speedup | ryu_opt speedup |
-|---|---|---:|---:|---:|---:|---:|---:|---:|
-| random_uniform | `repr(x)` | 285.3 | 97.9 | 98.3 | 96.4 | 2.92x | 2.90x | 2.96x |
-| random_uniform | `f"{x:.4f}"` | 119.2 | 103.4 | 103.5 | 104.8 | 1.15x | 1.15x | 1.14x |
-| random_bits | `repr(x)` | 819.6 | 135.5 | 135.2 | 146.1 | 6.05x | 6.06x | 5.61x |
-| random_bits | `f"{x:.4f}"` | 1,956.3 | 193.9 | 5,549.3 | 212.2 | 10.09x | 0.35x | 9.22x |
-| financial | `repr(x)` | 172.0 | 80.7 | 79.6 | 101.6 | 2.13x | 2.16x | 1.69x |
-| financial | `f"{x:.4f}"` | 144.5 | 100.5 | 100.1 | 115.1 | 1.44x | 1.44x | 1.25x |
-| scientific | `repr(x)` | 639.8 | 133.6 | 132.7 | 141.0 | 4.79x | 4.82x | 4.54x |
-| scientific | `f"{x:.4f}"` | 1,078.6 | 158.1 | 2,972.4 | 170.7 | 6.82x | 0.36x | 6.32x |
-| integer_valued | `repr(x)` | 143.1 | 88.7 | 88.2 | 80.7 | 1.61x | 1.62x | 1.77x |
-| integer_valued | `f"{x:.4f}"` | 168.0 | 104.8 | 104.5 | 107.6 | 1.60x | 1.61x | 1.56x |
+values above 1.00x are faster than stock. `zmij` is the pure-C
+single-library backend proposed for CPython; its `repr` carries a text
+re-read of ~10 ns that a to-decimal entry point in Żmij's C API would
+remove, and its fixed-precision path on huge magnitudes (`random_bits`)
+is bigint-based and ~2x slower than Ryu's `d2fixed`, while still 4x
+faster than stock.
+
+| Corpus | Operation | Stock (ns) | fmt_opt | fmt | ryu_opt | zmij | fmt_opt speedup | fmt speedup | ryu_opt speedup | zmij speedup |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| random_uniform | `repr(x)` | 284.6 | 91.0 | 90.6 | 93.8 | 100.7 | 3.13x | 3.14x | 3.03x | 2.83x |
+| random_uniform | `f"{x:.4f}"` | 119.2 | 97.9 | 97.6 | 102.9 | 109.8 | 1.22x | 1.22x | 1.16x | 1.09x |
+| random_bits | `repr(x)` | 821.9 | 130.9 | 129.0 | 146.1 | 153.3 | 6.28x | 6.37x | 5.62x | 5.36x |
+| random_bits | `f"{x:.4f}"` | 1,936.1 | 193.6 | 5,550.4 | 213.6 | 452.8 | 10.00x | 0.35x | 9.07x | 4.28x |
+| financial | `repr(x)` | 169.5 | 73.5 | 73.5 | 99.2 | 89.3 | 2.30x | 2.31x | 1.71x | 1.90x |
+| financial | `f"{x:.4f}"` | 143.5 | 95.2 | 96.0 | 114.1 | 114.0 | 1.51x | 1.49x | 1.26x | 1.26x |
+| scientific | `repr(x)` | 633.8 | 127.6 | 126.9 | 142.1 | 149.7 | 4.97x | 5.00x | 4.46x | 4.23x |
+| scientific | `f"{x:.4f}"` | 1,079.2 | 159.3 | 2,973.1 | 172.2 | 279.4 | 6.77x | 0.36x | 6.27x | 3.86x |
+| integer_valued | `repr(x)` | 141.8 | 81.8 | 82.0 | 78.0 | 92.4 | 1.73x | 1.73x | 1.82x | 1.54x |
+| integer_valued | `f"{x:.4f}"` | 167.6 | 98.2 | 99.0 | 107.2 | 109.3 | 1.71x | 1.69x | 1.56x | 1.53x |
 
 ## Parse backend
 
 `float(s)` over each corpus, `s` being a stock CPython repr. Speedup is
-`stock / backend`.
-
-| Corpus | Operation | Stock (ns) | fast_float | wuffs | stock | fast_float speedup | wuffs speedup | stock speedup |
-|---|---|---:|---:|---:|---:|---:|---:|---:|
-| random_uniform | `float(s)` | 121.4 | 43.9 | 55.6 | 106.3 | 2.76x | 2.18x | 1.14x |
-| random_bits | `float(s)` | 277.6 | 59.1 | 293.4 | 208.7 | 4.70x | 0.95x | 1.33x |
-| financial | `float(s)` | 36.8 | 37.0 | 39.8 | 78.0 | 0.99x | 0.92x | 0.47x |
-| scientific | `float(s)` | 213.7 | 57.4 | 66.5 | 183.3 | 3.72x | 3.21x | 1.17x |
-| integer_valued | `float(s)` | 42.8 | 42.0 | 46.6 | 64.1 | 1.02x | 0.92x | 0.67x |
+`stock / backend`. `ffc` (ffc.h, the C99 port of fast_float) is the
+pure-C parse backend proposed for CPython; it tracks `fast_float`
+within noise, while `wuffs` loses on 17-digit inputs (`random_bits`)
+where its high-precision-decimal fallback runs.
+
+| Corpus | Operation | Stock (ns) | fast_float | wuffs | ffc | stock | fast_float speedup | wuffs speedup | ffc speedup | stock speedup |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| random_uniform | `float(s)` | 120.8 | 45.6 | 56.2 | 44.2 | 109.3 | 2.65x | 2.15x | 2.73x | 1.11x |
+| random_bits | `float(s)` | 276.5 | 58.8 | 293.4 | 59.5 | 212.1 | 4.70x | 0.94x | 4.65x | 1.30x |
+| financial | `float(s)` | 36.6 | 37.6 | 39.9 | 37.3 | 79.9 | 0.98x | 0.92x | 0.98x | 0.46x |
+| scientific | `float(s)` | 214.3 | 57.1 | 67.3 | 56.9 | 187.7 | 3.75x | 3.18x | 3.77x | 1.14x |
+| integer_valued | `float(s)` | 43.0 | 43.4 | 47.9 | 44.1 | 66.4 | 0.99x | 0.90x | 0.98x | 0.65x |

CMakeLists.txt

@@ -1,18 +1,24 @@
 cmake_minimum_required(VERSION 3.20)
 project(floatium LANGUAGES C CXX)
 
+set(CMAKE_C_STANDARD 11)              # zmij.c: static_assert, stdbool
+set(CMAKE_C_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_EXTENSIONS OFF)
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 
 # --- Backend selection ------------------------------------------------------
-# Format backend: fmt_opt  (default — fmt for repr/exp, Ryu d2fixed for fixed),
+# Format backend: zmij     (default — pure-C Żmij for every mode; one unmodified
+#                           upstream library, the CPython-proposal backend),
+#                 fmt_opt  (fmt for repr/exp, Ryu d2fixed for fixed),
 #                 fmt      (plain fmt for everything, for A/B measurement),
 #                 ryu_opt  (pure-C: Ryu d2s for repr, d2exp for %e/%g, d2fixed
 #                           for %f — zero C++ in the format path),
 #                 all.
-# Parse backend:  fast_float (default), wuffs (pure-C Eisel-Lemire + HPD),
+# Parse backend:  ffc       (default — pure-C ffc.h, the C99 port of
+#                           fast_float; the CPython-proposal backend),
+#                 fast_float (the C++ original), wuffs (pure-C Eisel-Lemire + HPD),
 #                 stock     (libc strtod via floatium's wrapper), all.
 #
 # For a "stock CPython" format baseline, just `floatium.uninstall()` —
@@ -23,10 +29,13 @@ set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 #
 # With "all", every backend is compiled in and the active one is chosen at
 # runtime via env vars FLOATIUM_FORMAT_BACKEND / FLOATIUM_PARSE_BACKEND.
-set(FLOATIUM_FORMAT_BACKEND "fmt_opt" CACHE STRING "Format backend: fmt_opt, fmt, ryu_opt, all")
-set(FLOATIUM_PARSE_BACKEND  "fast_float" CACHE STRING "Parse backend: fast_float, wuffs, stock, all")
-set_property(CACHE FLOATIUM_FORMAT_BACKEND PROPERTY STRINGS fmt_opt fmt ryu_opt all)
-set_property(CACHE FLOATIUM_PARSE_BACKEND  PROPERTY STRINGS fast_float wuffs stock all)
+# The defaults changed from fmt_opt + fast_float to zmij + ffc in 0.16.0:
+# the shipped wheel is now the exact pure-C configuration proposed for
+# CPython, so every user runs the candidate code.
+set(FLOATIUM_FORMAT_BACKEND "zmij" CACHE STRING "Format backend: fmt_opt, fmt, ryu_opt, zmij, all")
+set(FLOATIUM_PARSE_BACKEND  "ffc" CACHE STRING "Parse backend: fast_float, wuffs, ffc, stock, all")
+set_property(CACHE FLOATIUM_FORMAT_BACKEND PROPERTY STRINGS fmt_opt fmt ryu_opt zmij all)
+set_property(CACHE FLOATIUM_PARSE_BACKEND  PROPERTY STRINGS fast_float wuffs ffc stock all)
 
 # --- Python -----------------------------------------------------------------
 # scikit-build-core passes Python3_EXECUTABLE via the FindPython hint file.
@@ -66,6 +75,14 @@ if(FLOATIUM_FORMAT_BACKEND STREQUAL "ryu_opt" OR FLOATIUM_FORMAT_BACKEND STREQUA
     set(FLOATIUM_HAS_RYU_OPT_FORMAT 1)
     set(FLOATIUM_NEEDS_RYU_D2FIXED 1)
 endif()
+if(FLOATIUM_FORMAT_BACKEND STREQUAL "zmij" OR FLOATIUM_FORMAT_BACKEND STREQUAL "all")
+    list(APPEND FLOATIUM_SRCS
+        src/backends/zmij_format.cc
+        src/cpython_adapter/zmij_dtoa.c
+        third_party/zmij/zmij.c
+    )
+    set(FLOATIUM_HAS_ZMIJ_FORMAT 1)
+endif()
 # Ryu d2fixed.c is shared between fmt_opt and ryu_opt — compile it once.
 if(FLOATIUM_NEEDS_RYU_D2FIXED)
     list(APPEND FLOATIUM_SRCS third_party/ryu/d2fixed.c)
@@ -84,6 +101,13 @@ if(FLOATIUM_PARSE_BACKEND STREQUAL "wuffs" OR FLOATIUM_PARSE_BACKEND STREQUAL "a
     )
     set(FLOATIUM_HAS_WUFFS_PARSE 1)
 endif()
+if(FLOATIUM_PARSE_BACKEND STREQUAL "ffc" OR FLOATIUM_PARSE_BACKEND STREQUAL "all")
+    list(APPEND FLOATIUM_SRCS
+        src/backends/ffc_parse.cc
+        src/cpython_adapter/ffc_strtod.c
+    )
+    set(FLOATIUM_HAS_FFC_PARSE 1)
+endif()
 if(FLOATIUM_PARSE_BACKEND STREQUAL "stock" OR FLOATIUM_PARSE_BACKEND STREQUAL "all")
     list(APPEND FLOATIUM_SRCS src/backends/stock_parse.cc)
     set(FLOATIUM_HAS_STOCK_PARSE 1)
@@ -98,6 +122,8 @@ target_include_directories(_ext PRIVATE
     third_party/fast_float
     third_party/ryu
     third_party/wuffs
+    third_party/zmij
+    third_party/ffc
 )
 
 # Match the build flags the CPython fmt-fastfloat branch uses for the
@@ -128,19 +154,21 @@ endif()
 # wheel but still need a single default; prefer the optimized fmt-based one.
 set(FLOATIUM_DEFAULT_FORMAT "${FLOATIUM_FORMAT_BACKEND}")
 if(FLOATIUM_DEFAULT_FORMAT STREQUAL "all")
-    set(FLOATIUM_DEFAULT_FORMAT "fmt_opt")
+    set(FLOATIUM_DEFAULT_FORMAT "zmij")
 endif()
 set(FLOATIUM_DEFAULT_PARSE "${FLOATIUM_PARSE_BACKEND}")
 if(FLOATIUM_DEFAULT_PARSE STREQUAL "all")
-    set(FLOATIUM_DEFAULT_PARSE "fast_float")
+    set(FLOATIUM_DEFAULT_PARSE "ffc")
 endif()
 
 target_compile_definitions(_ext PRIVATE
     $<$<BOOL:${FLOATIUM_HAS_FMT_FORMAT}>:FLOATIUM_HAS_FMT_FORMAT=1>
     $<$<BOOL:${FLOATIUM_HAS_FMT_OPT_FORMAT}>:FLOATIUM_HAS_FMT_OPT_FORMAT=1>
     $<$<BOOL:${FLOATIUM_HAS_RYU_OPT_FORMAT}>:FLOATIUM_HAS_RYU_OPT_FORMAT=1>
+    $<$<BOOL:${FLOATIUM_HAS_ZMIJ_FORMAT}>:FLOATIUM_HAS_ZMIJ_FORMAT=1>
     $<$<BOOL:${FLOATIUM_HAS_FAST_FLOAT_PARSE}>:FLOATIUM_HAS_FAST_FLOAT_PARSE=1>
     $<$<BOOL:${FLOATIUM_HAS_WUFFS_PARSE}>:FLOATIUM_HAS_WUFFS_PARSE=1>
+    $<$<BOOL:${FLOATIUM_HAS_FFC_PARSE}>:FLOATIUM_HAS_FFC_PARSE=1>
     $<$<BOOL:${FLOATIUM_HAS_STOCK_PARSE}>:FLOATIUM_HAS_STOCK_PARSE=1>
     FLOATIUM_DEFAULT_FORMAT_BACKEND="${FLOATIUM_DEFAULT_FORMAT}"
     FLOATIUM_DEFAULT_PARSE_BACKEND="${FLOATIUM_DEFAULT_PARSE}"