🚀 go-pugleaf

rlk@think.com (Robert L Krawitz) writes:

>Does anyone know how to make gcc NOT optimize this sort of expression:

>	    tempvar = *CURDEV.headaddr >> 16;
>	    tempvar = tempvar << 16;
>	    tempvar |= *CURDEV.patch_head_2 >> 16;
>	    break;

>The optimizer transforms this into various halfword loads and stores,
>but unfortunately the memory that this references is not addressable in
>quantities less than a full word, so the kernel (yup, this is kernel
>code) takes a bus error.

>Yes, it's possible to put temporary variables in, but this code seems to
>be all over the place.  I don't want to turn off optimization altogether
>because this is in an interrupt handler.

>This is gcc 2.3.3 running on a sparc.

First, I find it impressive that gcc optimizes this. Secondly, you might
try declaring CURDEV.headaddr as a pointer to volatile data, as memory
locations of special significance (such as memory mapped I/O) should
be declared usually, so that the comiler knows not to reorder and
reorganize accesses (just like it does here).

Whether gcc 2.3.3 sufficiently supports valatile as to make this
working correctly, I do not know. In addition, the above code will
(if it works) obviously reference the memory twice, so you might be
tempted to throw in the extra local variable anyway.

But declaring hardware locations volatile should be the first step, and
is really necessary anyhow if you have a highly optimizing compiler.
-- 
 David Kastrup        dak@pool.informatik.rwth-aachen.de          
 Tel: +49-241-72419 Fax: +49-241-79502
 Goethestr. 20, D-52064 Aachen