Re: Compiler optimization of division and remainder

hbaker@netcom.com (Henry G. Baker) writes:
>[an alleged optimization computes a/b and a%b with a single divide if both
>appear in a program]


>I can imagine how such an optimization would be done, but I can't
>imagine many other places -- except for possibly keeping track of a
>shifted out carry bit in a shift operation, or gathering both the high
>order and low order bits of a double-precision integer multiplication
>-- where such an optimization would be used.


        Some other places where two expressions should be evaluated together:


        a) The maximum and minimum of the same two arguments.
                [P.s. If your system has #define max(x, y) ((x) > (y) : (x) : (y))
                in a header file, then the definition of min(x, y) should
                be ((x) > (y) ? (y) : (x)) to facilitate recognition
                of the common subexpression (x) > (y) and allow the optimizer
                to generate one if-then-else.]


        b) The sine and cosine of the same argument. If sin(theta)
                and cos(theta) generate library calls, then any basic block
                referencing both functions should instead call a routine
                which evaluates both functions while doing common
                calculations such as range reduction only once.


        c) When there are two relationals on the same two operands.
                For example, the code


                                if (x > y) {
                                        ...
                                } else if (x < y) {
                                        ...
                                }


                should generate only one compare, whatever the types of x and y.
--
                Peter L. Montgomery pmontgom@cwi.nl San Rafael, California
--

Return to the comp.compilers page.
Search the comp.compilers archives again.