peephole optimizations

peephole optimizations

[roy rosen]

Hi All,

I have tried to write some peephole patterns and I now have some
questions regarding the way it is working.

1. Is that true that if I try to match in the pattern two insns and in
my code between these insns there is another insn which does not have
any dependency connection to the other two, Is that true that the
peephole would not match in this case? (i.e. the insns to match must
come in the code in sequential order with no other insns between
them)?

2. I saw that the peephole2 pass is done after IRA. Many peepholes are
practically useless after register allocating.
Is it possible to run it before IRA? Is it possible to run it several times?

Thanks, Roy.

Re: peephole optimizations

[Ian Lance Taylor]

roy rosen <roy.1rosen@gmail.com> writes:

> 1. Is that true that if I try to match in the pattern two insns and in
> my code between these insns there is another insn which does not have
> any dependency connection to the other two, Is that true that the
> peephole would not match in this case? (i.e. the insns to match must
> come in the code in sequential order with no other insns between
> them)?

Yes.


> 2. I saw that the peephole2 pass is done after IRA. Many peepholes are
> practically useless after register allocating.
> Is it possible to run it before IRA? Is it possible to run it several times?

There are also many peepholes can not be run before register
allocation, because they depend on relationships between the
registers.

Before register allocation, you should normally write an insn pattern
which can be recognized by the combine pass, rather than writing a
peephole.  To do this, write a define_insn as usual, and use
appropriate operand and insn predicates to recognize the pattern you
are looking for.

That has the advantage that it works even if random insns appear in
between the insns you care about, the case you ask about in your first
question.  (However, it does not work across block boundaries.)

Ian

Re: peephole optimizations

[roy rosen]

Hi,

2010/5/3, Ian Lance Taylor <iant@google.com>:
> roy rosen <roy.1rosen@gmail.com> writes:
>
> > 1. Is that true that if I try to match in the pattern two insns and in
> > my code between these insns there is another insn which does not have
> > any dependency connection to the other two, Is that true that the
> > peephole would not match in this case? (i.e. the insns to match must
> > come in the code in sequential order with no other insns between
> > them)?
>
> Yes.
>

Was there any special problem implementing such peephole mechanism or
simply nobody tried to work on this yet?

I think it may be very beneficial since I guess that for most
patterns, there are other insns between the pattern's insns which
cause the peephole not to match...

>
> > 2. I saw that the peephole2 pass is done after IRA. Many peepholes are
> > practically useless after register allocating.
> > Is it possible to run it before IRA? Is it possible to run it several times?
>
> There are also many peepholes can not be run before register
> allocation, because they depend on relationships between the
> registers.
>
> Before register allocation, you should normally write an insn pattern
> which can be recognized by the combine pass, rather than writing a
> peephole.  To do this, write a define_insn as usual, and use
> appropriate operand and insn predicates to recognize the pattern you
> are looking for.
>
> That has the advantage that it works even if random insns appear in
> between the insns you care about, the case you ask about in your first
> question.  (However, it does not work across block boundaries.)
>
> Ian
>

Re: peephole optimizations

[Ian Lance Taylor]

roy rosen <roy.1rosen@gmail.com> writes:

> 2010/5/3, Ian Lance Taylor <iant@google.com>:
>> roy rosen <roy.1rosen@gmail.com> writes:
>>
>> > 1. Is that true that if I try to match in the pattern two insns and in
>> > my code between these insns there is another insn which does not have
>> > any dependency connection to the other two, Is that true that the
>> > peephole would not match in this case? (i.e. the insns to match must
>> > come in the code in sequential order with no other insns between
>> > them)?
>>
>> Yes.
>>
>
> Was there any special problem implementing such peephole mechanism or
> simply nobody tried to work on this yet?

It would have to be very conservative about whether the intervening
instructions affected the instructions being optimized.  I don't think
there is any special problem other than that.

The compiler doesn't rely on peepholes that much, since most of that
sort of thing is done in the combine pass.  The peepholes are mostly
used to optimize the output of the register allocator.  As such, the
instructions tend to be sequential anyhow.  Before starting a project
to permit skipping instructions in the middle of peepholes, I think it
would be useful to analyze some .s files and see how often the
peephole could be applied.

Ian

Re: peephole optimizations

[Jeff Law]

On 05/04/10 01:18, roy rosen wrote:
> Hi,
>
> 2010/5/3, Ian Lance Taylor<iant@google.com>:
>    
>> roy rosen<roy.1rosen@gmail.com>  writes:
>>
>>      
>>> 1. Is that true that if I try to match in the pattern two insns and in
>>> my code between these insns there is another insn which does not have
>>> any dependency connection to the other two, Is that true that the
>>> peephole would not match in this case? (i.e. the insns to match must
>>> come in the code in sequential order with no other insns between
>>> them)?
>>>        
>> Yes.
>>
>>      
> Was there any special problem implementing such peephole mechanism or
> simply nobody tried to work on this yet?
>
> I think it may be very beneficial since I guess that for most
> patterns, there are other insns between the pattern's insns which
> cause the peephole not to match...
>    
In general, peepholes (in the GCC world) are typically only used when 
there is no obvious data dependency between insns and thus the combiner 
phase of GCC doesn't know to try and optimize the two (or three) 
insns.    It has been our experience that peepholes are only 
occasionally useful and thus we haven't spent considerable time 
improving the infrastructure for the peephole optimizer.

Jeff