[LLVMdev] Suggestion About Adding Target Dependent Decision in LSR Please

[Andrew Trick]

On Mar 14, 2013, at 2:21 PM, Yin Ma <yinma at codeaurora.org> wrote:

> Hi Andy,
>  
> Actually, if we just add hooks that preserves the existing behavior,
> It is not difficult. For example,
>  
> For case one, we can define one function like
>   virtual const SCEV* getTargetPreferredWinnerReg(const SCEV*& ScaledReg,
>            SmallVector<const SCEV *, 4>& BaseRegs, GlobalValue*& BaseGV) const;
>  
> In NarrowSearchSpaceByPickingWinnerRegs, we can preserves the winner
> reg from target and winner reg from the original algorithm if this function
> returns NULL, it is just like before
>  
> For case two, we can define a general cost from TTI function, like
>   virtual int getLSRFormulaCost(const unsigned NumRegs,
>                             const unsigned AddRecCost, const unsigned NumIVMuls,
>                             const unsigned NumBaseAdds, const unsigned ImmCost,
>                             const unsigned SetupCost) const;
> Then we do something like
>   int thisCost = TTI->getLSRFormulaCost(NumRegs, AddRecCost, NumIVMuls,
>                                            NumBaseAdds, ImmCost, SetupCost);
>   if (thisCost >= 0) {
>     int otherCost = TTI->getLSRFormulaCost(Other.NumRegs, Other.AddRecCost,
>                                             Other.NumIVMuls, Other.NumBaseAdds,
>                                             Other.ImmCost, Other.SetupCost);
>     if (otherCost >= 0)
>       return thisCost < otherCost;
>   }
> In bool Cost::operator<(const Cost &Other) const

Exposing the internals of LSR to TTI is cheating. This might actually be acceptable though as long as it would be rare for a target to specialize at this level, and doing so implies that the target may be broken by major LSR changes.

If you post your implementation of these hooks, we may be able to see a way to form a better abstraction.

> We could have more decision from target backend.
>  
> However, from the problem I am dealing with, which has a lot of branches in multiple level
> Loop nests. LSR is still not able to perform the best because
> 1.       LSR is not control flow sensitive. It treats all USE equally, which doesn’t care which
> USE is on critical path and which USE is on a branch. Without these kind of information,
> We cannot predict AddRec precisely because we only can assume all USEs can be post
> Increment or all not.
> 2.       The most occurred winner regs pruning may not be the best approach. Because target
> may prefer certain regs than others, even some registers do occur more. Specially,
> register with small computation is more likely to occur in formulas. However, register
> with small computation may not always be a best choice if the content in register are
> loop invariant.
>  
> Therefore,  We may need a systemic agreement or plan to address the existing LSR problems. I
> would like to ask if any party has any improvement plan about LSR? So we can come together
> to have an unified solution to handle all known problem in one round?

I'm open to redesign or total replacement of LSR. I don't have any simple fixes in mind for the current design other than to improve the bailout logic so we fall back to the original code in more cases.

Before speculating about the right design, I would first like to see opt -loop-reduce test cases for whatever we think is important. Hopefully you can checkin all the TTI hooks for your target so we can have working unit tests. Meanwhile, attaching examples to a PR would be good.

It sounds like your loops have a large number of IV users. I'm surprised LSR is able to find the best solution given its current set of heuristics. It often prunes the best solution or simply fails to find a solution. Are you sure that adding complexity to the heuristics will lead it to the best solution? Or can you imagine a different way to approach the problem that doesn't involve a search space that grows exponentially with the loop size?

-Andy

> From: Andrew Trick [mailto:atrick at apple.com] 
> Sent: Thursday, March 14, 2013 9:42 AM
> To: Yin Ma
> Cc: llvmdev at cs.uiuc.edu
> Subject: Re: [LLVMdev] Suggestion About Adding Target Dependent Decision in LSR Please
>  
>  
> On Mar 13, 2013, at 4:37 PM, Yin Ma <yinma at codeaurora.org> wrote:
> 
> 
> Hi All,
>  
> In the target I am working, we comes cross a situation that the loop strength reduction
> could deliver a better result but currently not, because
> 1.       the algorithm narrows search space by winner registers without considering
> the target preferred format. (NarrowSearchSpaceByPickingWinnerRegs)
> 2.       Cost comparison solely favors the number register without considering other
> Impacts.
>  
> For the case one,
> NarrowSearchSpaceByPickingWinnerRegs filters by most occurred registers.
> ld(basereg, immediate) is a target preferred addressing mode. However, it may
> be deleted because basereg is very likely not to be the most occurred register
> because the less opportunity in a combination.
>  
> For the case two, by observing the cost comparison equation
> bool Cost::operator<(const Cost &Other) const {
>   if (NumRegs != Other.NumRegs)                            return NumRegs < Other.NumRegs;
>   if (AddRecCost != Other.AddRecCost)                  return AddRecCost < Other.AddRecCost;
>   if (NumIVMuls != Other.NumIVMuls)                   return NumIVMuls < Other.NumIVMuls;
>   if (NumBaseAdds != Other.NumBaseAdds)       return NumBaseAdds < Other.NumBaseAdds;
>   if (ImmCost != Other.ImmCost)                               return ImmCost < Other.ImmCost;
>   if (SetupCost != Other.SetupCost)                         return SetupCost < Other.SetupCost;
>   return false;
> }
>  
> If we have a case to compare
> Cost at 5 regs, with addrec cost 1, plus 15 base adds, plus 1 imm cost, plus 4 setup cost.
> Cost at 4 regs, with addrec cost 1, plus 28 base adds, plus 1 imm cost, plus 2 setup cost.
> The current mode will select 4 regs case even there are 14 more base adds. And base
> Adds matters in our targets.
>  
> So I think the current LSR should be pushing more decision into target dependent backend.
> Like calling new functions in TargetTransformInfo. At least, in narrow search space and cost
> comparison phase, or more in cost rating phase. LSR can be tightly cooped with the target
> attributes in order to get the most beneficial result.
>  
> Yes. LSR decisions are tightly coupled with the target architecture and potentially the subtarget microarcthitecture. As you figured out, the way to handle it is to communicate more information to LSR through TTI. It's easy to do this to solve individual benchmarks on your target. It's hard to know if you have a general solution that works across targets. But if you can add hooks in a way that preserves existing behavior on other targets it shouldn't be a problem. We want to design general hooks, but leave it up to someone doing the benchmarking to tune them for a particular target.
>  
> -Andy

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20130319/3d66d604/attachment.html>