Repository navigation
CWG2752 Excess-precision floating-point literals #1584
Description
Activity
SG6 input is desired, please assign to EWG afterwards.
Chair: Matthias Kretz
Minute Taker: Philip Craig
Summary
The issue is asking
What is the design intent for handling excess-precision floating-point
computations in C++?Nobody in SG6 felt qualified to answer this question from an "original design
intent" perspective. So the question was interpreted as stating design intent
as required by numerics users. There was no consensus for a desired behavior
and it appears that a more thorough exploration in a paper would be necessary
to aid the discussion.Polls
- Compilers need to be allowed to use the x87 FPU for computation and
comparison (without "too much" intermediate rounding)
POLL: C++ should not deviate from C in how floating-point behaves (excess
precision, fp contraction, literals)SF F N A SA 0 3 1 1 1 # of Participants: 7
Outcome: No consensus for change
Comments: We could really use a paper to make a more informed decision. The
paper would need to explore the consequences of all the different options.Polls considered, but not taken without a better understanding of the
consequences:POLL: The C++ standard should allow floating-point contraction into FMAs by
default.(We disagreed on whether the current wording in [expr.pre] allows fp
contraction. Therefore, there was no clear "status-quo" position.)POLL: There needs to be a standard-conforming way to control floating-point
contraction.- Compilers need to be allowed to use the x87 FPU for computation and
Seen in EWG on Wednesday:
Poll: CWG2752 / P3488: Floating-Point Excess Precision,
FLT_EVAL_METHODmust always be0.SF F N A SA 2 5 5 4 4 Result: not consensus
Poll: CWG2752 / P3488: Floating-Point Excess Precision,
FLT_EVAL_METHODcan have any value allowed by C (this poll requires fixing some Core wording).SF F N A SA 0 5 8 4 4 Result: not consensus
In C23, floating-point literals (called "constants" there) can be represented with excess precision, meaning a
floatliteral is treated as adoubleorlong doublevalue (without affecting the type system, though).In C++, there is no such permission.
Both C++ and C permit floating-point computations (including operands) to use excess precision. A footnote hints that a conversion will drop the excess precision (e.g.
(float)(314.f/100.f)yields a value withfloatprecision) (like C does), but there is no normative wording to make that happen.What is the design intent for handling excess-precision floating-point computations in C++?
See CWG2752 for details.