On 2008-11-30 11:35:45 -0800, Tim Prince wrote: > gfortran is permitted to extract common factors this way, but it will > recognize only a few of them from the original version. The C standard > doesn't permit automatic optimization of common factors. The C standard permits such kind of optimization unless contraction of floating expressions is disabled. This can have bad side effects, though. I only know that GCC can automatically contract a multiply and an add to a FMA, on targets where the FMA is available (even when contraction is disabled -- but this is a bug). C code can also exploit that, but you should know what you're doing. BTW, I've shown an example of a bad side effect due to the contraction to a FMA: http://www.vinc17.org/research/fptest.en.html#contract2fma > This points up an exception to the common factors optimization rule: > x + x2*x should not be written as x*(1 + x2), as this would not save > an operation, When a FMA is available, you may even waste an operation with x * (1 + x2), depending on the context. > and would reduce accuracy for the desirable case where the x > term is larger than the trailing terms. (x+x) is preferred to x*2, but > optimizers should be capable of making these choices automatically. This also depends on the context (still the FMA...). -- Vincent Lefèvre <vincent@xxxxxxxxxx> - Web: <http://www.vinc17.org/> 100% accessible validated (X)HTML - Blog: <http://www.vinc17.org/blog/> Work: CR INRIA - computer arithmetic / Arenaire project (LIP, ENS-Lyon)
