Re: [PATCH, PR 10474] Shedule pass_cprop_hardreg before pass_thread_prologue_and_epilogue

[Vladimir Makarov <vmakarov@redhat.com>]

On 13-04-19 4:21 AM, Richard Biener wrote:
> On Fri, Apr 19, 2013 at 1:08 AM, Martin Jambor <mjambor@suse.cz> wrote:
>> Hi,
>>
>> On Fri, Apr 19, 2013 at 12:37:58AM +0200, Steven Bosscher wrote:
>>> On Fri, Apr 19, 2013 at 12:09 AM, Martin Jambor wrote:
>>>> I also have not tried scheduling the hard register copy propagation
>>>> pass twice and measuring the impact on compile times.  Any suggestion
>>>> what might be a good testcase for that?
>>> I think a better question is when this would be useful in the first
>>> place, and why. In other words: If you propagate hardregs before
>>> shrink wrapping, what could be a source of new opportunities after
>>> shrink wrapping?
>> Yes, we also did that and neither I nor Honza could think of any
>> potential problems there.  And of course, I'd also measure how many
>> statements the second run of the pass changed.  I'll probably do that
>> tomorrow anyway.
>>
>>>
>>> The only things I can think of, given the current pass order, are:
>>>
>>> * different basic block order due to shrink wrapping
>>>   regcprop's effectiveness depends on the order of the basic blocks
>>> (unfortunately)
>>>
>>> * different basic block contents due to head/tail merging (pass_jump2)
>>>   Head/tail merging extends some basic blocks and shortens others. The
>>> elongated basic blocks may present new opportunities (regcprop is a
>>> local pass).
>>>
>>> * different basic block contents due to dead store elimination (pass_dse2)
>>>   A removed dead store may also make an address calculation redundant,
>>> changing the regcprop value chains.
>>>
>>> * different basic block contents due to peephole2
>>>   A peephole2 may present new regcprop opportunities, peephole2 misses
>>> the context to avoid trivial copies.
>>>
>>>
>>> On the other hand, running regcprop earlier also helps some other
>>> passes. For example, I think regcprop before jump2 may result in more
>>> successful head/tail merging attempts by making more input operands
>>> match, but it could hurt if_after_reload by extending live times of
>>> registers.
>>>
>>>
>>> But wouldn't it be better to avoid these argument-register pseudos
>>> being assigned to callee-saved registers? Perhaps splitting the live
>>> range of the pseudos before the first call on each path will do the
>>> trick, and let IRA pick the right registers for you instead.
>> First, where can I have a look how a live range is split?  ;-)
> Insert a copy and adjust all dominated uses:
>
>   (set (new-pseudo old-pseudo))
>
> ... adjust downstream uses of old-pseudo to use new-pseudo ...
>
>> But second, if such a call is in a loop (or accessible by more than
>> one path), I wouldn't it be wrong to do that?  To avoid that, I
>> suppose I might end up doing another shrink-wrapping-like search for
>> the right edge for prologue and actually coming up with a very similar
>> result to the propagation and shrink-wrapping preparation.  But I'm
>> willing to try.
> I suppose splitting life-ranges in general before argument setup might
> make sense - I see hardreg copyprop as a hack around limitations
> in register allocation.  Note that life-range splitting is undone by
> regular copy propagation.
>
> ISTR IRA splits life-ranges in loop code, so there must be already
> some helpers that could be used.  Vlad?
>
>
   I'd not recommend to reuse this code as actual live-range splitting 
is buried under a lot of code to modify IR of IRA as we need the IR 
after live-range splitting.  Long ago I used splitting and rebuilding IR 
but building IR is a very time consuming procedure (I guess 1/2 of IRA) 
therefore the current solution is used.

There is an alternative simpler code for this in IRA.  The code was 
written by Bernd:

http://gcc.gnu.org/ml/gcc-patches/2011-12/msg01531.html

   By the way, I  have plans to do a separate register pressure 
decreasing pass based on live-range shrinkage and rematerialization.   I 
found that we need this as more optimizations have a tendency to deal 
with this issue by themselves.  May be I'll find time to do this in this 
year (but most probably not for gcc4.9).  I am not sure even for release 
next after gcc4.9.