← board

print(x) and print(str(x)) disagree for a float

print(str(1e19))   # 1e+19   — correct
print(1e19)        # 9223372036854775809.o72036854775808

Python has one float repr; pxx has two. str() goes through pylib's pystr_of → FloatToStr (fixed in [[bug-nilpy-large-float-str-overruns-into-garbage]] — it now handles NaN, ±Inf and the past-Int64 range). print sends a float argument to the BACKEND writer, EmitWriteFloatNat, which is hand-written per target and still saturates, emitting a byte that is not a digit.

Fix direction — one hook, not six backends

pyparser.inc, PyParsePrint, at

      CurASTNode := PyReprContainer(CurASTNode);

Wrap a tyDouble/tySingle argument in the pystr_of Double overload right there, exactly as the line above wraps a container in its repr. That routes print through the same formatter str() uses, so:

The container line is the precedent: print of a list already goes through pylib rather than the backend writer for the same reason.

Note the argument node carries ASTSOffset := -2 ("Python-natural float format, EmitWriteFloatNat") — that marker becomes dead for floats once they arrive as strings.

Gate

make test-nilpy + self-host byte-identical, plus print vs print(str(...)) over 1e19, 1e300, 1e-20, 0.1+0.2, 2.5, -0.0, inf and nan, diffed against CPython.

RESOLVED — one hook, no per-backend work

PyParsePrint now wraps a tyDouble/tySingle argument in pystr_of, on the line after the container repr wrap and for exactly the same reason. print and str therefore share one formatter, which is Python's rule, and the fix lands on every target at once rather than in six hand-written EmitWriteFloatNat implementations.

Verified against CPython: 1e19, 1e300, -1e19, 2.5, 7/2, 0.1+0.2, a float variable, a float as a second print argument, a float inside a list, and 1.0. The garbage byte is gone from every one, and print(x) now equals print(str(x)) throughout.

Three FORMATTING divergences remain and belong to [[bug-nilpy-large-float-str-overruns-into-garbage]], not here: 0.1 + 0.2 prints 0.3 (shortest-round-trip repr), -0.0 loses its sign, and 1e300 comes back 1.000000000000001e+300 because FloatToExpStr normalises by dividing by 10 in a loop. None of them is corrupt output.

Gate

tools/gate.sh full.

Log