constant folding with infinities is wrong

main = let big = 99999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999
                  :: Double in
  print (big * big - big * big)

prints NaN when compiled without optimizations but 0.0 when compiled with optimizations.

Or, print (big * big / 2) prints Infinity when compiled without optimizations but 8.98846567431158e307 when compiled with optimizations (this number is fromRational (toRational (1/0) / 2)).

The cause is the conversions that go on between Rational and Double in PrelRules (in the functions doubleOp2, mkDoubleVal, convFloating).

Ideally, we would have an accurate model of floating-point arithmetic on the target machine, but for now it would be an improvement to just not constant fold when the result is not a finite floating-point number.

Trac metadata

Trac field	Value
Version	7.8.3
Type	Bug
TypeOfFailure	OtherFailure
Priority	low
Resolution	Unresolved
Component	Compiler
Test case
Differential revisions
BlockedBy
Related
Blocking
CC
Operating system
Architecture