Structured Fortran - was Re: Self modifying code, lambda calculus

On 09/23/2015 01:39 AM, ANDY HOLT wrote: And I take your point. At least on supercomputers, it has pretty much been the pattern that vast amounts of code are spent on automatic optimization. There seems to be a situation where computers implement a set of features that allow for higher performance, but then the HLL has to have some way to express the use of these features. SIMD instructions can be very difficult to abstract. Initially, it's usually done through a subroutine/function library, then incorporated into the code generation and optimization. Then, the language committee haggles and comes up with language structures. Take F90, for example. One of the major focii of the effort was to incorporate vector extensions into the language. But how to cover the most ground and yet come up with something that a non-vector machine can do justice to. Whatever you come up with, it must be intelligible to humans (recall that the standard has a broad definition of "computer", which includes humans). Whatever the new extensions do, they must also not break the syntax and meaning of the already existing base language. So you have to have language features to define vectors of varying shapes, operations of various strides, control structures (e.g. sparse vectors, control vectors, etc.). This can engender some furious battles. I recall that IBM threatened to withdraw from X3J3 if its Vectran extensions were not implemented. Pointers were a hot topic in F90. There was a contingent who fiercely opposed the idea, as it can make automatic optimization very difficult--and it's equally difficult to see why it's necessary, as Fortran has survived for decades without them. And there were the problems that someone always had a neat hardware feature that no one else had who wanted a special language feature in the standard. The focus of FORTRAN has pretty much always been high performance numerical calculation. It can be used for other things, as can any other language, but even Wirth confessed that FORTRAN code consistently outran his Pascal on numerical problems. But the problem with any language, is that the more one adds to the language, the less comprehensible that language can become. C++ certainly has taken this route, as has Fortran. F2000 is so far removed it today, unless it was essential to my job. I know how to write C++ fairly well, but find myself relapsing to C for simple programs. Sometimes more isn't better. --Chuck