[f-cpu] A simple SIMD extension for C(++)

[Michael Riepe <michael@stud.uni-hannover.de>]

I read an old article about the shortcomings of C/C++ the other day. The
author complained that there is no way to apply a function to a number
of arguments (like the lisp functions `map', `mapcar' & friends do).
He didn't mention it, but a `map' function would be valuable for SIMD
processors, too: it explicitly tells the compiler that the same function
is going to be applied to many arguments in a uniform way. E.g. an
instruction like

	__map__(result_array, function, count, array_1, array_2);

(or similar) would be largely equivalent to

	for (int i = 0; i < count; i++) {
		result_array[i] = function(array_1[i], array_2[i]);
	}

except that the former may apply the function to the array elements in any
order (that is, even in parallel -- exactly what SIMD does). Of course
the construct should not be limited to functions with two operands.
But the number and element types of the array arguments passed to
__map__ must match the prototype of the function. Therefore, __map__
itself can't be implemented as a C function or preprocessor macro.

Note that the extension is fully backwards compatible: the compiler may,
at its discretion, replace the mapping with a conventional `for' loop
without changing the semantics of the program. On the other hand, it can
easily find out that <function> ist to be applied to <count> successive
elements of <array_1> and <array_2>, which enables it to perform SIMD
optimizations (and/or loop unrolling) without prior data flow analysis.

-- 
 Michael "Tired" Riepe <Michael.Riepe@stud.uni-hannover.de>
 "All I wanna do is have a little fun before I die"
*************************************************************
To unsubscribe, send an e-mail to majordomo@seul.org with
unsubscribe f-cpu       in the body. http://f-cpu.seul.org/