----- On Jul 9, 2020, at 4:46 PM, Segher Boessenkool seg...@kernel.crashing.org
wrote:
> On Thu, Jul 09, 2020 at 01:56:19PM -0400, Mathieu Desnoyers wrote:
>> > Just to make sure I understand your recommendation. So rather than
>> > hard coding r17 as the temporary registers, we could explicitly
>> > declare the temporary register as a C variable, pass it as an
>> > input operand to the inline asm, and then refer to it by operand
>> > name in the macros using it. This way the compiler would be free
>> > to perform its own register allocation.
>> >
>> > If that is what you have in mind, then yes, I think it makes a
>> > lot of sense.
>>
>> Except that asm goto have this limitation with gcc: those cannot
>> have any output operand, only inputs, clobbers and target labels.
>> We cannot modify a temporary register received as input operand. So I don't
>> see how to get a temporary register allocated by the compiler considering
>> this limitation.
>
> Heh, yet another reason not to obfuscate your inline asm: it didn't
> register this is asm goto.
>
> A clobber is one way, yes (those *are* allowed in asm goto). Another
> way is to not actually change that register: move the original value
> back into there at the end of the asm! (That isn't always easy to do,
> it depends on your code). So something like
>
> long start = ...;
> long tmp = start;
> asm("stuff that modifies %0; ...; mr %0,%1" : : "r"(tmp), "r"(start));
>
> is just fine: %0 isn't actually modified at all, as far as GCC is
> concerned, and this isn't lying to it!
It appears to be at the cost of adding one extra instruction on the fast-path
to restore the register to its original value. I'll leave Boqun whom authored
the original rseq-ppc code to figure out what works best performance-wise
(when he finds time).
Thanks for the pointers!
Mathieu
--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
