Use the `@fastmath` macro. This dynamically enables floating-point optimizations in the expression (or function) to which you apply it. Compare `@inbounds`, `@simd`.
-erik On Sun, Dec 6, 2015 at 5:23 PM, Meik Hellmund <meik.hellm...@gmail.com> wrote: > Hi, > > It seems that julia does not "optimize away" constant expressions in > functions: > > julia> f(x)=x+2-2 > f (generic function with 1 method) > > julia> f(1.e-18) > 0.0 > > This is of course correct in Float64 arithmetics. > But I wonder: In languages like Fortran and C the result of such code > depends on the "optimization flags" used when compiling. > With optimization, the compiler would reduce the function to f(x)=x. > Is there something comparable in Julia? > > Another test. Compare > f(x)=x+sin(.34567) > > to > > const sn=sin(.34567) > g(x)=x+sn > > > julia> y=0; @time(for i in 1:10^9; y=f(y); end) > 22.489526 seconds (2.00 G allocations: 29.802 GB, 3.50% gc time) > > julia> y=0; @time(for i in 1:10^9; y=g(y); end) > 16.268512 seconds (1000.00 M allocations: 14.901 GB, 2.61% gc time) > > > It looks like Julia does not optimize even the simplest constant > expressions so that they are evaluated only once. > And why, by all means, does it allocate so much memory for that? > Is there something I overlook? > > Best wishes, > Meik > > -- Erik Schnetter <schnet...@gmail.com> http://www.perimeterinstitute.ca/personal/eschnetter/
