• Anton Blanchard's avatar
    powerpc: Optimise per cpu accesses on 64bit · ae01f84b
    Anton Blanchard authored
    Now we dynamically allocate the paca array, it takes an extra load
    whenever we want to access another cpu's paca. One place we do that a lot
    is per cpu variables. A simple example:
    
    DEFINE_PER_CPU(unsigned long, vara);
    unsigned long test4(int cpu)
    {
    	return per_cpu(vara, cpu);
    }
    
    This takes 4 loads, 5 if you include the actual load of the per cpu variable:
    
        ld r11,-32760(r30)  # load address of paca pointer
        ld r9,-32768(r30)   # load link address of percpu variable
        sldi r3,r29,9       # get offset into paca (each entry is 512 bytes)
        ld r0,0(r11)        # load paca pointer
        add r3,r0,r3        # paca + offset
        ld r11,64(r3)       # load paca[cpu].data_offset
    
        ldx r3,r9,r11       # load per cpu variable
    
    If we remove the ppc64 specific per_cpu_offset(), we get the generic one
    which indexes into a statically allocated array. This removes one load and
    one add:
    
        ld r11,-32760(r30)  # load address of __per_cpu_offset
        ld r9,-32768(r30)   # load link address of percpu variable
        sldi r3,r29,3       # get offset into __per_cpu_offset (each entry 8 bytes)
        ldx r11,r11,r3      # load __per_cpu_offset[cpu]
    
        ldx r3,r9,r11       # load per cpu variable
    
    Having all the offsets in one array also helps when iterating over a per cpu
    variable across a number of cpus, such as in the scheduler. Before we would
    need to load one paca cacheline when calculating each per cpu offset. Now we
    have 16 (128 / sizeof(long)) per cpu offsets in each cacheline.
    Signed-off-by: default avatarAnton Blanchard <anton@samba.org>
    Signed-off-by: default avatarBenjamin Herrenschmidt <benh@kernel.crashing.org>
    ae01f84b
setup_64.c 16.1 KB