From mboxrd@z Thu Jan 1 00:00:00 1970 From: Toon Moene To: Richard Henderson Cc: egcs@cygnus.com Subject: Re: More than you ever wanted to know about Fortran array indexing;-) Date: Mon, 06 Oct 1997 07:02:00 -0000 Message-id: <9710061346.AA05616@moene.indiv.nluug.nl> References: <9710051153.AA02261@moene.indiv.nluug.nl> <19971005174628.31661@dot.cygnus.com> X-SW-Source: 1997-10/msg00220.html > The appended patch, I believe, does this. The result is > relatively good code for all of the examples that have > been posted so far. Yep, the timing for bs3dvw.f also is around the same value as with the FFECOM_FASTER_ARRAY_REFS "solution". BTW, it seems to be very hard to get consistent timings on a Linux Alpha - in five runs I got anywhere between 3.6 and 4.3 seconds, whereas on my machine it is consistent up to the tenth of seconds (and it takes 90 seconds !). > Only, I would like to see loop take y[i-1], y[i], y[i+1], > make one induction variable, and use -4(R), 0(R), 4(R) > in the loop. Hmm... This is a nice follow-on project :-) Note that lots of computational physics code is based on finite difference or finite element algorithms, which all have these opportunities for optimisation. Or recognise that, on loop unrolling, you've got the following equivalences: 1 y(i-1) y(i) y(i+1) 2 y(i-1) y(i) y(i+1) 3 y(i-1) y(i) y(i+1) 4 y(i-1) y(i) y(i+1) Anyway, we can consider this problem "solved" (apply Jeff's `use' patch, your `expr' patch, my `fold' patch and your `frontend' patch), unless Craig objects against the frontend changes. There is namely a pedantic-standards-interpretation question hiding in your frontend patch: Should the compiler evaluate each and every expression in the mode it is written in ? Note that currently, g77+backend as a whole already violate that rule because of the `sizetype' promotion `expr.c' does. What your frontend patch does is lift this uncleanliness up into the frontend ... We'll see - Craig has the last word on this. Cheers, Toon. PS: I'm starting a test now with all those patches applied to egcs-970929, to see if it still works on my Fortran code ;-) PS2: My `fold' patch has a typo; somewhere it says: + && multiple_of_p (Ttype, REE_OPERAND (xarg0, 1), arg1)) Obviously, this should be: + && multiple_of_p (type, TREE_OPERAND (xarg0, 1), arg1))