Re: Q: division vs multiplication

hbaker@netcom.com (Henry Baker) writes:
>The problem isn't in the compiler optimization phase, but in the compiler
>input phase. There are some compiler input phases that don't properly
>read decimal constants, resulting in mangled constants. If you contact
>the compiler vendor, they usually don't consider that a 'bug'. However,
>if you provide an _integer_ and the compiler input phase mangles it, then
>you can usually convince the compiler vendor to fix it.
>
>For this reason, I have found a/1048576 to be more robust/portable than
>multiplying by its inverse.


Have you tried to define the needed constant as a long, and then
casting the binary value to a float when you need it?


As long as both integer and floating point values are stored with the same
endian-ness, this should work. (Assuming normal IEEE fp encodings.)


I.e., for the Pentium FPATAN workaround I needed to be able to multiply by
1/16, which I encoded directly as:


static unsigned long one_sixteenth = 0x3d800000;


That evaluates to a biased exponent of -4, and a mantissa of 1.0, when
accessed as:


    *(float *) &one_sixteenth;


I do agree that a/1048576 is probably a _lot_ more portable than a cast
of (0x35800000) to (float)! :-)


-Terje Mathisen (include std disclaimer) <Terje.Mathisen@hda.hydro.com>
--

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