From: Rildo P. <rpr...@ac...> - 1999-12-13 17:23:57
|
Hi Jim, On Mon, 13 Dec 1999, Noeth, Jim wrote: > I have been working on the run time routines to facilitate the cobol move > verb, and have them very close to being finished. My concern, though, is > that these routines have considerable overhead (especially when moving > numeric fields) to figure out what type of move is involved, how much to > pad, how much to truncate, etc.. I'm afraid that our move routines are going > to be very innefficient. What I would like to do is to modify the 'gen_move' > routine to generate the code directly, all of the information to do these > 'overhead' calculations is available at compile time, and we could generate > code that is very specific to each move statement. We can do a simple 2-way classification. If the fields are both binaries, we generate a simple move at compile-time. If not, we call your routine. What you think? There are some optimizations I have already introduced, but I can try to do more. (or you can too, if you want. Look for the functions gen_move, gen_loadvar, gen_loadloc at the file htcobgen.c) I really think we are more concerned with having something usable fast. Then we make gradual performance improvements. > I realize that this will make the compiler less portable (and a little more > difficult to debug), I believe that I can write the code generation routines > such that they can be easily modified for different platforms. Let me know > what you think. We can leeave the more efficient code generation for the next version of the compiler. We need a beta version soon, to have more contributors helping us. We only will have a really portable compiler when we output C language code (or some intermediary language), perhaps with a multi-pass compiler. regards, ---------------------------------------------------------------- Rildo Pragana FPGA/uControllers * Linux * tcl/tk P.O. Box 721 Camaragibe PE http://members.xoom.com/rpragana Brazil 54792-990 +55-81-459-1776 |