Repository navigation
semantic: a numeric literal outside its own type's range is not checked before codegen — 1e400, 1e39f are SPY0599 behind Roslyn, 1e29m is a SPY0909 ICE #2340
Description
Class tracker. Found by the P68 verify round (sibling-cell prober, C1 × literal magnitude), measured @ 0d58024 with sharpyc run.
Contract
A numeric literal whose value its own type cannot hold gets a typed refusal in semantic analysis, or an explicit spec rule (CPython gives inf for a float). It must never reach code generation. Integers already work this way: an unsuffixed literal past int64 is SPY0210 "too large for 'long'". The float32 narrowing twin does too: float_literals.md:82 says x: float32 = 1e40 is SPY0220. The literal forms below are not checked. They reach Roslyn (SPY0599) or crash the compiler (SPY0909).
Cells
| Literal type | Program | Observed @ 0d58024 | CPython 3.12 |
|---|---|---|---|
| float64 (unsuffixed) | print(1e400) (also -1e400, 1e309, 1.8e308) |
SPY0599 "generated C# contains 1 syntax error(s): Floating-point constant is outside the range of type 'double'" | inf / -inf |
float64 (d suffix) |
print(1e400d) |
SPY0599, same message | n/a (no suffix) |
| float64 in a typed slot | x: float = 1e400; b: bool = 1.0 < 1e400 |
SPY0599 | inf; True |
float32 (f suffix) |
print(1e39f) (also 3.5e38f) |
SPY0599 "outside the range of type 'float'" (compare: x: float32 = 1e40 is SPY0220) |
n/a |
decimal (m suffix) |
print(1e29m) (also 1e400m, 79228162514264337593543950336m) |
SPY0909 internal compiler error (InvalidOperationException): "The decimal literal '1e29' … could not be parsed as a decimal. Semantic analysis admits a literal into a decimal slot only after checking its range…" | n/a |
Minimal repro:
def main():
print(1e400)
Notes
- The lexer is right.
NumericLiteralTwinTests(tests: lexer observation totality — the Stryker pilot's seven survivor classes C1–C7 (numeric literals, token positions, line endings, f-string arms, LiteralLoss scanner, end of file, keyword roster) #2300 C1) compares lexer values only, and the lexer yieldsFloat '1e400'. The missing check is semantic, keyed on the literal's resolved type: float64, float32 and decimal. - Open ruling: should a float64 literal past
double.MaxValuebe refused, or beinfas in CPython? The spec says nothing for float64.float_literals.md:82refuses silent infinity for float32. Either way, SPY0599/SPY0909 is a bug. - Not in this class: int64/uint64 overflow is already SPY0210, which is by design.
Activity
antonsynd commentedon Oct 22, 2026
Batching 2026-10-21 @ e7e0a0e — charter
.claude/plans/remediation-round-2026-09-20-batching.md§3.0, chores band (order 0, one commit, no plan; Semantic expression typing only; seam-free with P21e, rides its gate). Ruling R-HU (owner, 2026-10-21): (a) REFUSE every numeric literal outside the range of its own resolved type at ONE semantic seam — the SPY0210 family. Ruling R-HV (owner, 2026-10-21): (a) a chore in the current band with a closed-axis theory.Measured @ e7e0a0e (built CLI, lock-free):
print(1e400)→ SPY0599 "generated C# contains 1 syntax error(s): Floating-point constant is outside the range of type 'double'" (Roslyn CS0594 behind the net);print(1e29m)→ SPY0909 internal compiler errorInvalidOperationExceptionwhose own message (CodeGen/RoslynEmitter.Expressions.cs:796) says "Semantic analysis admits a literal into a decimal slot only after checking its range" — the emitter holds an invariant semantic does not keep;1e39f→ SPY0599 'float'. Controls:x: float32 = 1e40→ SPY0220 (float_literals.md:82, "not a silent infinity");print(99999999999999999999)→ SPY0210 "too large for any integer type". CPython 3.12.13:1e400→inf. Lane / reach: ice · program (rare shape). Class: the integer check and the float32 slot check exist; float64,fandmare unchecked — contract §1's position-missing-shared-check meta-class.Mechanism (R-HU a): one range check keyed on the literal's RESOLVED type; the SPY0210 family gains "too large for 'double' / 'float' / 'decimal'";
float_literals.mdgains the sentence beside line 82 anddecimallikewise;DiagnosticExplanationsfor any new code. Guard (R-HV a): theory type {int32→int64 promotion, int64, uint32, uint64, float64, float32, decimal} × spelling {unsuffixed, suffixed, slot-narrowed} × magnitude {max, max + one ulp, far, negative}; mutation (drop the float64 arm → its rows red) in the commit body; direction record: NO program that ran changes (every cell is SPY0599/SPY0909 today) — recorded in the commit body, no fixture moves. Blast radius: Semantic → the whole gate. Class tracker: this issue.Rejected — R-HU (b): float64 unsuffixed →
inf(CPython's value,double.PositiveInfinity) withf/mrefusing — two rules for one surface, andx: float = 1e400then carries infinity into arithmetic silently; a runtime value is one-way, a refusal is relaxable to (b) later; Type Safety > Python syntax, and the spec already chose refusal for float32. R-HV (b): a P37 row — P37 is slot-DIRECTED operand totality (Some(v)/None()against a typed other operand);print(1e400)and1e29mhave no slot, so P37's generator would gain a foreign axis and an ICE would wait behind P70. Ordering two-way.fix(semantic): a numeric literal outside its own type's range is SPY0…
antonsynd commentedon Oct 22, 2026
Landed (chores band 2026-10-22) on
dev@ 4d22212, PR #2346 to mainline. Whole-solution gate 81669/0/7 @ 4d22212 (measured), against 81093/0/7 @ be4664e.Commit 64ff07e (R-HU a, R-HV a).
FloatLiteralClassifier.Fits(text, double|float|decimal)is the one range question. The literal-typing arm refuses a miss with SPY0210 ("too large for 'double'/'float'/'decimal'"), and the float32 and decimal narrowing predicates ask the sameFits. The SPY0210 constant is renamedNumericLiteralOutOfRange.float_literals.mdgains "A Literal Must Fit Its Own Type".Direction:
1e400,-1e400,1e400d,1e39f, the slot and pattern positions, and the parameter-default and const positions.1e29m,79228162514264337593543950336m.-1UL→ SPY0223, andx: uint32 = -1U→ SPY0220.Matrix:
NumericLiteralRangeMatrixTests, 84 rows plus a 13-text Roslyn differential.Mutations:
trueturned 14 rows red.NegatesAsLiteralreturningtrueturned 3 rows red.Sibling filed: #2344. A non-zero decimal literal below 1e-28 still gets two verdicts.
1e-30mis stored as zero, whiled: decimal = 1e-30is refused. The narrowing path keeps an underflow check the suffixed path never asks. That is a magnitude axis this issue's contract did not list, so it has its own tracker.Every cell in this issue's table is met. Closing.
Merge pull request #2346 from antonsynd/dev