On Thursday, 1 August 2013 at 22:45:10 UTC, bearophile wrote:
Walter Bright:
But consider that optimizers are built to optimize typical
code patterns. Component programming is fairly non-existent in
C and C++, and is new in D. Hence, optimizers are not set up
to deal with those patterns (yet).
I agree.
GHC also works with a LLVM back-end, so those optimizations are
done in some kind of middle-end.
Probably a silly idea: perhaps we can collect some money, like
1000-2000 dollars, to pay for a 3 day long course for Walter
(total about 15 hours) about such matters.
Who's giving the course and where will it be held?
'Modern Compiler Design' by D. Grune, et al. & 'Compiler Design -
Analysis and Transformation' by H. Seidl, et al. discusses some
basic optimizations for functional programs. But I'm pretty sure
Walter is already familiar with these.
Taking an example from Ali's book "P' in D":
import std.stdio;
import std.algorithm;
void main()
{
auto values = [ 1, 2, 3, 4, 5 ];
writeln(values
.map!(a => a * 10)
.map!(a => a / 3)
.filter!(a => !(a % 2)));
}
As stated this implies 3 separate traversals of the list (or
array to be specific) which is what a naïve implementation would
do. But all three operations can run on the same traversal. Given
that all three operations are monotonic functions (preserving the
order and cardinality of the set, i.e. 1-1 mapping) they are also
inherently parallelizable (i.e. amenable to auto vectorization or
loop unrolling).
