perf(stable-api): inline num2long/num2ulong Bignum fast path for MRI

[ianks]

Summary

For Bignums that fit in c_long/c_ulong, read BDIGIT digits directly from RBignum instead of calling rb_num2long/rb_num2ulong. Falls back to the dylib for overflow (which raises RangeError), Float coercion, and to_int dispatch.

Only active on 64-bit targets (#[cfg(target_pointer_width = "64")]). 32-bit always falls through to the dylib.

Layout

RBignum { RBasic basic; union { { usize len; u32 *digits; } heap; u32 ary[2]; } as_; }

• BIGNUM_EMBED_FLAG = RUBY_FL_USER2 — digits are in as_.ary (embedded)
• BIGNUM_EMBED_LEN_MASK = USER3|USER4|USER5, shift = RUBY_FL_USHIFT + 3 = 15
• BIGNUM_POSITIVE_P = flags & RUBY_FL_USER1
• BIGNUM_EMBED_LEN_MAX = 2 on 64-bit (= sizeof(u64)/sizeof(u32))

Fits-in-i64 check (64-bit, BDIGIT=u32)

Digit count	Signed	Unsigned
0	→ 0	→ 0
1	always fits	always fits
2	range-check hi+lo	range-check
3+	fall back (RangeError)	fall back

Asm: fully inlined; fixnum path unchanged, Bignum fast path adds ~10 instructions before the dylib fallback.

Parity tests added (10 new)

• 2**32 (1 BDIGIT, positive)
• 2**62 (2 BDIGITs, fits)
• -(2**32) (negative, 1 BDIGIT)
• -(2**62) (negative, 2 BDIGITs)
• Boundary values at i64::MAX, i64::MIN
• Unsigned variants

189/189 tests pass.

Stack

1. perf(stable-api): inline num2dbl T_FLOAT fast path for MRI #733 — num2dbl T_FLOAT fast path
2. perf(stable-api): inline dbl2num flonum encode for MRI #734 — dbl2num flonum encode fast path
3. perf(stable-api): inline rhash_size for MRI #735 — rhash_size AR/ST inline ⬅ base
4. This PR — num2long/num2ulong Bignum fast path

🤖 Generated with Claude Code