[PATCH] ARM: Runtime patch udiv/sdiv instructions into __aeabi_{u}idiv()

[Nicolas Pitre]

On Thu, 7 Jan 2016, Stephen Boyd wrote:

> On 01/04, Nicolas Pitre wrote:
> > On Mon, 4 Jan 2016, Stephen Boyd wrote:
> > > 
> > > I can update the patches to be based on this patch here and
> > > handle the conditional branches and tail call optimization cases
> > > by adding some safety checks like we have for the ftrace branch
> > > patching. But I'd rather not do that work unless we all agree
> > > that it's worthwhile pursuing it.
> > > 
> > > Is there still any concern about the benefit of patching each
> > > call site vs. patching the functions? The micro benchmark seems
> > > to show some theoretical improvement on cortex-a7 and I can run
> > > it on Scorpion and Krait processors to look for any potential
> > > benefits there, but I'm not sure of any good kernel benchmark for
> > > this. If it will be rejected due to complexity vs. benefit
> > > arguments I'd rather work on something else.
> > 
> > You could run the benchmark on Scorpion and Krait to start with. If 
> > there is no improvement what so ever like on A15's then the answer might 
> > be rather simple.
> > 
> 
> So running the benchmark on Scorpion is not useful because we
> don't have the idiv instruction there. On Krait I get the
> following results. I ran this on a dragonboard apq8074 with
> maxcpus=1 on the kernel command line.
> 
> Testing INLINE_DIV ...
> real    0m 13.56s
> user    0m 13.56s
> sys     0m 0.00s
> 
> Testing PATCHED_DIV ...
> real    0m 15.15s
> user    0m 15.14s
> sys     0m 0.00s
> 
> Testing OUTOFLINE_DIV ...
> real    0m 18.09s
> user    0m 18.09s
> sys     0m 0.00s
> 
> Testing LIBGCC_DIV ...
> real    0m 24.26s
> user    0m 24.25s
> sys     0m 0.00s
> 
> It looks like the branch actually costs us some time here.
> Patching isn't as good as the compiler inserting the instruction
> itself, but it is better than branching to the division routine.

I think it is up to you at this point whether or not you feel the call 
site patching complexity is worth it. Personally I'd rather go directly 
for the compiler flag solution to close the final performance gap, but I 
won't stand in the way of a call site patching patch.


Nicolas