René's URL Explorer Experiment

Title: gh-150424: Widen specialized int fast paths to full int64 range by KRRT7 · Pull Request #150425 · python/cpython · GitHub

Open Graph Title: gh-150424: Widen specialized int fast paths to full int64 range by KRRT7 · Pull Request #150425 · python/cpython

X Title: gh-150424: Widen specialized int fast paths to full int64 range by KRRT7 · Pull Request #150425 · python/cpython

Description: Fixes gh-150424. This PR builds on the initial widening work by splitting the specialised `int` opcodes into two tiers: Compact (`BINARY_OP_{ADD,SUBTRACT,MULTIPLY}_INT`): operands are already in CPython's 30-bit compact range — semantically identical to `main`, no additional overhead. Wide (`BINARY_OP_{ADD,SUBTRACT,MULTIPLY}_INT_WIDE`): new fast path for exact ints in the full `int64_t` range using checked 64-bit arithmetic. Changes: Split specialisation: `specialize.c` prefers the compact variant when both operands are compact, falling back to the wide variant for larger int64-range values `PyCompactLong{Add,Subtract,Multiply}`: return `PyStackRef_NULL` for overflow and OOM (matching `medium_from_stwodigits` in `main`), eliminating a dead `ERROR_IF` branch from every JIT-compiled compact trace New `_wide_op_result(int64_t)` static inline inlines the small-int cache, single-digit freelist alloc, and `_PyLong_FromLarge` call, cutting the `PyLong_FromInt64` → `_PyLong_FromMedium` call chain for medium-range results Tier-2 optimizer: `_GUARD_TOS/NOS_INT` substitutes `_GUARD_TOS/NOS_COMPACT` when the type is already known to be `PyLong_Type`; new `_GUARD_TOS/NOS_INT_WIDE` substitutes `_GUARD_TOS/NOS_OVERFLOWED` under the same condition Six new Tier-2 inplace ops (`BINARY_OP{ADD,SUBTRACT,MULTIPLY}_INT_WIDE_INPLACE` and `_RIGHT` variants) Add regression coverage for compact/wide split, guard substitution, and overflow fallback Comparison: main (`bdf14dc`) vs this branch (`5e01840`). Build: PGO + thin LTO, macOS arm64, Homebrew Clang 22. Microbenchmarks target int arithmetic across compact and non-compact (int64) ranges. pyperformance benchmarks are general workloads not specific to this change — included to confirm no regression. Microbenchmarks (PYTHON_JIT=0) Note: pyperf worker subprocesses exit before the JIT compilation threshold is reached; JIT=1 numbers are within noise of JIT=0 and not shown separately. Benchmark main branch Change int_small 269 ns — not significant int_compact 134 ns — not significant int_intermediate_overflow 1.11 µs 706 ns 1.57x faster int_double_add 1.21 µs 903 ns 1.34x faster int_accumulate 667 ns 517 ns 1.29x faster int_always_large 435 ns 328 ns 1.33x faster int_mixed 155 ns 118 ns 1.32x faster Geometric mean 1.25x faster pyperformance — interpreter only (PYTHON_JIT=0) 12 numeric/float-heavy benchmarks. Build: PGO + thin LTO, macOS arm64, Homebrew Clang 22. Benchmark main branch Change Significant chaos 27.2 ms 27.5 ms 1.01x slower No float 35.6 ms 35.5 ms 1.00x faster No nbody 51.1 ms 50.4 ms 1.01x faster No pidigits 228 ms 228 ms 1.00x No pyflate 212 ms 218 ms 1.03x slower Yes raytrace 129 ms 130 ms 1.00x slower No scimark_fft 134 ms 136 ms 1.01x slower No scimark_lu 54.2 ms 54.0 ms 1.00x faster No scimark_monte_carlo 31.7 ms 31.8 ms 1.00x slower No scimark_sor 57.7 ms 58.1 ms 1.01x slower No scimark_sparse_mat_mult 2.17 ms 2.13 ms 1.02x faster No spectral_norm 47.4 ms 46.8 ms 1.01x faster No pyperformance — JIT (PYTHON_JIT=1) Benchmark main branch Change Significant chaos 27.2 ms 27.6 ms 1.01x slower No float 35.6 ms 35.5 ms 1.00x faster No nbody 51.2 ms 50.5 ms 1.01x faster No pidigits 228 ms 228 ms 1.00x No pyflate 212 ms 218 ms 1.03x slower Yes raytrace 130 ms 130 ms 1.00x slower No scimark_fft 135 ms 136 ms 1.01x slower No scimark_lu 54.1 ms 54.0 ms 1.00x faster No scimark_monte_carlo 31.7 ms 31.8 ms 1.00x slower No scimark_sor 57.5 ms 58.1 ms 1.01x slower No scimark_sparse_mat_mult 2.17 ms 2.25 ms 1.04x slower No spectral_norm 47.6 ms 46.7 ms 1.02x faster Yes pyflate regression (1.03x slower, significant in both modes, t≈−16): likely caused by the `DEOPT_IF(!_PyLong_IsNonNegativeCompact(...))` guards added to `_BINARY_OP_SUBSCR_LIST_INT` / `_STORE_SUBSCR_LIST_INT` in the correctness fix — pyflate does heavy bit-pattern indexing through those bytecodes on every iteration. The guard is necessary for correctness but adds a branch on the hot path that wasn't there before.

Open Graph Description: Fixes gh-150424. This PR builds on the initial widening work by splitting the specialised `int` opcodes into two tiers: Compact (`BINARY_OP_{ADD,SUBTRACT,MULTIPLY}_INT`): operands are already in C...

X Description: Fixes gh-150424. This PR builds on the initial widening work by splitting the specialised `int` opcodes into two tiers: Compact (`BINARY_OP_{ADD,SUBTRACT,MULTIPLY}_INT`): operands are already in C...

Opengraph URL: https://github.com/python/cpython/pull/150425

X: @github

direct link

Domain: github.com

Links:

Viewport: width=device-width