[Bug c/46390] New: cpuid assembly instruction is optimized away in loop over threads.

[edwin at svi dot nl]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46390

           Summary: cpuid assembly instruction is optimized away in loop
                    over threads.
           Product: gcc
           Version: 4.4.3
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassig...@gcc.gnu.org
        ReportedBy: ed...@svi.nl


GCC seems to optimize the cpuid assembly instruction (used to retrieve the
current APIC ID) away, because the input for the instruction seems to be
independent from the loop:

    for (i = 0; i < CPU_SETSIZE; i++) {
        /* Add cpu i to the cpu set 'set'. */
        CPU_ZERO(&set);
        CPU_SET(i, &set);

        /* Try to switch the affinity and collect the APIC ID of
           thread i. Sleep gives the OS a chance to switch to the
           desired thread. */
        err = sched_setaffinity(pid, sizeof(cpu_set_t), &set);
        sleep(0);
        if (err >= 0) {
            asm("cpuid"
                : "=b" (additional)
                : "a" (1)
                : "%ecx", "%edx");
            id = (additional & APIC_ID_MASK) >> 24;
            list[(*cnt)++] = id;
        }
    }


It can be 'fixed' by using the loop variable i as a 'dummy input' for the
assembly instruction:

            asm("cpuid"
                : "=b" (additional)
                : "a" (1), "b" (i)
                : "%ecx", "%edx");


This method to retrieve APIC ID's is common practice:
http://software.intel.com/en-us/articles/optimal-performance-on-multithreaded-software-with-intel-tools/