Re: [PATCH] lower-subreg, expr: Mitigate inefficiencies derived from "(clobber (reg X))" followed by "(set (subreg (reg X)) (...))"

[Takayuki 'January June' Suwa <jjsuwa_sys3175@yahoo.co.jp>]

(sorry repost due to the lack of cc here)
Hi!

On 2022/08/04 18:49, Richard Sandiford wrote:
> Takayuki 'January June' Suwa <jjsuwa_sys3175@yahoo.co.jp> writes:
>> Thanks for your response.
>>
>> On 2022/08/03 16:52, Richard Sandiford wrote:
>>> Takayuki 'January June' Suwa via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
>>>> Emitting "(clobber (reg X))" before "(set (subreg (reg X)) (...))" keeps
>>>> data flow consistent, but it also increases register allocation pressure
>>>> and thus often creates many unwanted register-to-register moves that
>>>> cannot be optimized away.
>>>
>>> There are two things here:
>>>
>>> - If emit_move_complex_parts emits a clobber of a hard register,
>>>   then that's probably a bug/misfeature.  The point of the clobber is
>>>   to indicate that the register has no useful contents.  That's useful
>>>   for wide pseudos that are written to in parts, since it avoids the
>>>   need to track the liveness of each part of the pseudo individually.
>>>   But it shouldn't be necessary for hard registers, since subregs of
>>>   hard registers are simplified to hard registers wherever possible
>>>   (which on most targets is "always").
>>>
>>>   So I think the emit_move_complex_parts clobber should be restricted
>>>   to !HARD_REGISTER_P, like the lower-subreg clobber is.  If that helps
>>>   (if only partly) then it would be worth doing as its own patch.
>>>
>>> - I think it'd be worth looking into more detail why a clobber makes
>>>   a difference to register pressure.  A clobber of a pseudo register R
>>>   shouldn't make R conflict with things that are live at the point of
>>>   the clobber.
>>
>> I agree with its worth.
>> In fact, aside from other ports, on the xtensa one, RA in code with frequent D[FC]mode pseudos is terribly bad.
>> For example, in __muldc3 on libgcc2, the size of the stack frame reserved will almost double depending on whether or not this patch is applied.
> 
> Yeah, that's a lot.

So lots, but almost double might be an overstatement :)

BTW after some quick experimentation, I found that turning on -fsplit-wide-types-early would roughly (but not completely) solve the problem.  Surely, the output was not so bad in the past...

> 
>>>>  It seems just analogous to partial register
>>>> stall which is a famous problem on processors that do register renaming.
>>>>
>>>> In my opinion, when the register to be clobbered is a composite of hard
>>>> ones, we should clobber the individual elements separetely, otherwise
>>>> clear the entire to zero prior to use as the "init-regs" pass does (like
>>>> partial register stall workarounds on x86 CPUs).  Such redundant zero
>>>> constant assignments will be removed later in the "cprop_hardreg" pass.
>>>
>>> I don't think we should rely on the zero being optimised away later.
>>>
>>> Emitting the zero also makes it harder for the register allocator
>>> to elide the move.  For example, if we have:
>>>
>>>   (set (subreg:SI (reg:DI P) 0) (reg:SI R0))
>>>   (set (subreg:SI (reg:DI P) 4) (reg:SI R1))
>>>
>>> then there is at least a chance that the RA could assign hard registers
>>> R0:R1 to P, which would turn the moves into nops.  If we emit:
>>>
>>>   (set (reg:DI P) (const_int 0))
>>>
>>> beforehand then that becomes impossible, since R0 and R1 would then
>>> conflict with P.
>>
>> Ah, surely, as you pointed out for targets where "(reg: DI)" corresponds to one hard register.
> 
> I was thinking here about the case where (reg:DI …) corresponds to
> 2 hard registers.  Each subreg move is then a single hard register
> copy, but assigning P to the combination R0:R1 can remove both of
> the subreg moves.
> 
>>> TBH I'm surprised we still run init_regs for LRA.  I thought there was
>>> a plan to stop doing that, but perhaps I misremember.
>>
>> Sorry I am not sure about the status of LRA... because the xtensa port is still using reload.
> 
> Ah, hadn't realised that.  If you have time to work on it, it would be
> really good to move over to LRA.  There are plans to remove old reload.

Alas you do overestimate me :) I've only been working about the GCC development for a little over a year.
Well it's a lie that I'm not interested in it, but too much for me.

> 
> It might be that old reload *does* treat a pseudo clobber as a conflict.
> I can't remember now.  If so, then zeroing the register wouldn't be
> too bad (for old reload only).
> 
>> As conclusion, trying to tweak the common code side may have been a bit premature.
>> I'll consider if I can deal with those issues on the side of the target-specific code.
> 
> It's likely to be at least partly a target-independent issue, so tweaking
> the common code makes sense in principle.
> 
> Does adding !HARD_REGISTER_P (x) to:
> 
>   /* Show the output dies here.  This is necessary for SUBREGs
>      of pseudos since we cannot track their lifetimes correctly;
>      hard regs shouldn't appear here except as return values.  */
>   if (!reload_completed && !reload_in_progress
>       && REG_P (x) && !reg_overlap_mentioned_p (x, y))
>     emit_clobber (x);
> 
> in emit_move_complex_parts help?  If so, I think we should do at

Probably yes.  Quick test says the abovementioned mod makes the ad-hoc fix I posted earlier (https://gcc.gnu.org/pipermail/gcc-patches/2022-June/596626.html) a thing of the past.

> least that much.
> 
> Thanks,
> Richard