A fast path for fixed-point float formatting

• Type: optimization — Track O (Track A file ownership: compiler/builtin/pylib.pas)
• Opened: 2026-08-13, by the fix for [[bug-nilpy-float-formatting-manufactures-ties-by-scaling]], which is what made it slower and is not to be undone.

Measured

200,000 "%.2f" % (i / 7.0) in a NilPy loop, same box, same program:

So ~1.85us -> ~8.8us per format. Correctness first: the old path printed 0.2 for "%.1f" % 0.15 where CPython prints 0.1, and got the error in both directions, so this is a cost that was paid for something.

Why it costs what it does

PyExDecDigits expands the mantissa into the FULL exact decimal expansion — for a value with a large negative binary exponent that is dozens to hundreds of digits, built one at a time into an AnsiString — and the formatter then keeps a handful of them. The work is proportional to the expansion, not to prec.

Where the fast path is

Not a guard band on the scaled arithmetic: "the remainder is not near 0.5" is exactly the judgement the bug was about, and a band wide enough to be safe for large values is wide enough to catch real values. Two directions that are actually provable:

1. Stop the expansion early. The formatter needs decExp + 1 + prec digits plus enough to decide the round (one digit, plus whether any nonzero digit follows). PyExDecOfMant could take a digit budget and stop, with a "rest is nonzero" flag — the same information PyExDecRound already computes by scanning the tail. That keeps the answer exact and makes the cost proportional to prec.
2. A power-of-two shortcut. A double whose exponent makes it an exact multiple of 10^-prec (all the x.5, x.25, x.125 cases, and every integer) cannot be near a midpoint by construction, and its digits are the scaled integer. Cheap to test on the mantissa's trailing zero count.

(1) is the general one and probably the whole job.

Gate

The bug-nilpy-float-formatting-manufactures-ties-by-scaling test still exact (it carries the midpoint rows), the dense sweep in that ticket still matching CPython, and the benchmark above back under ~2x of pinned.