On Mon, 2016-05-30 at 09:08 +1000, Anton Blanchard via Linuxppc-dev wrote:

quoted

That is surprising, do we have any idea what specifically increases
the overhead so significantly ? Does gcc know about ldbrx/stdbrx ? I
notice in our io.h for example we still do manual ld/std + swap
because old processors didn't know these, we should fix that for
CONFIG_POWER8 (or is it POWER7 that brought these ?).

The futex issue seems to be __get_user_pages_fast():

        ld      r11,0(r6)
        ...
        rldicl  r8,r11,32,32
        rotlwi  r28,r11,24
        rlwimi  r28,r11,8,8,15
        rotlwi  r6,r8,24
        rlwimi  r28,r11,8,24,31
        rlwimi  r6,r8,8,8,15
        rlwimi  r6,r8,8,24,31
        rldicr  r28,r28,32,31
        or      r28,r28,r6
        cmpdi   cr7,r28,0
        beq     cr7,2428

That's a whole lot of work just to check if a pte is zero. I assume
the reason gcc can't replace this with a byte reversed load is that
we access the pte via the READ_ONCE() macro.

Did I mention we need a bswap instruction?

We can possibly improve some of them by doing the comparison on the raw value,
eg. see hash__pte_same().

The above is from pgd_none() ?

cheers