P3488 R1 Floating-Point Excess Precision

[mattkretz]

P3488R0 Floating-Point Excess Precision (Matthias Kretz)

[mattkretz]

This paper addresses #1584

[mattkretz]

P3488R0: Floating-Point Excess Precision

2024-11-19 SG6 Minutes

Chair: Matthias Kretz

Champion: Matthias Kretz

Polls

POLL: Constant expressions (not constant folding) may not use greater
precision and range than the type. => Reproducibility in constant expressions.
If you need more precision and range use std::float128_t in constant
expressions.

SF	F	N	A	SA
4	2	0	0	0

# of Participants: 6
Outcome: Strong consensus.

POLL: Constant folding behavior follows runtime evaluation behavior.

SF	F	N	A	SA
5	1	0	0	0

# of Participants: 6
Outcome: Strong consensus.

POLL: Literals are always rounded to their types. (Never can a higher
precision and range value be used in subsequent operations. 1.1f + x where
x is a double cannot be evaluated as 1.1 + x.)

SF	F	N	A	SA
3	3	0	0	0

# of Participants: 6
Outcome: Strong consensus.

POLL: We should consider not allowing excess precision in runtime
evaluation at all (except if we introduce an "attribute" that controls this, as
documented in 60559). Every binary operation on floating-point is rounded once
to the precision and range of its type (decltype(expr)).

SF	F	N	A	SA
2	3	0	1	0

# of Participants: 6
Outcome: Consensus in favor.

POLL: Excess precision in runtime evaluation should be a permissible
optimization for implementations. The implementation choice of what
intermediate type was used is reflected by FLT_EVAL_METHOD (potentially as a
different trait or with the extended values as defined in C23 Annex H).

SF	F	N	A	SA
1	2	1	2	0

# of Participants: 6
Outcome: No consensus.

POLL: Comparison operators should always round their operands to the
precision and range of their types in case excess precision is used. In
addition to operators we expect the <cmath> comparison functions to have the
same behavior (std::isless(a, b) etc...)

SF	F	N	A	SA
3	0	2	0	1

# of Participants: 6
Outcome: No consensus.
SA: Not worth differing from C in this. x87 does comparison on long double
types.