Anton Blanchard wrote:

 >I changed the code a bit so that get_cpu() is now inline - this
 >represents our situation better. I think it is valid for gcc to cache
 >get_cpu across a function call in the below example because it knows
 >that get_cpu does not refer to any global variables.
 >
You are right: it seems that gcc internally decides if an inline
function is const or pure.

static inline int get_TR(void)
{
     int reg;
     __asm__ (
         "rep; nop;\n\t"
         "xor %0,%0\n\t"
         : "=r" (reg));
     return ret;
}

is always optimized as if get_TR is a function with __attribute__((const)).

Ben, is it intentional that get_TR is _not_ marked as inline? Your
version produces explicit function calls with -O2, and incorrect code
with -O99 (gcc decides to inline get_TR even without an inline
directive, and then optimizes away the calls after schedule.)

It seems that Anton's version generates the best code.
I've tested the attached version with egcs-1.1.2, gcc-2.96-98 and
gcc3-3.0.1-3, with -O0, -O2 and -O99.

--
     Manfred