| 1 |
We propose three modes of floating point
|
| 2 |
strictfp: Reproducible results as is now default
|
| 3 |
default: Exploit natural hardware (extended exponent in Intel and fused multiply add)
|
| 4 |
associatefp: Allow conventional compiler optimizations
|
| 5 |
with new Intel algorithm, strictfp will be a little more than 2 times slower than default; associatefp gains will be algorithm dependent (regular problems will have greatest gain)
|
| 6 |
Less important are following niceties:
|
| 7 |
indigenous keyword specifies maximum precision format supported by hardware
|
| 8 |
anonymous {float, double, indigenous} specifies floating point format to be used in calculating expressions
-
C used to always use double internally
|
