[LLVMdev] Any plan to add MIN/MAX isd node?

Duncan Sands baldrick at free.fr
Tue Oct 30 12:10:15 PDT 2012


Hi Yin,

>      To use select, usually, there is a compare before select.
> Presence of comparison will disable some opportunities to
> optimize some code. Select and Compare is not associative
> neither.

at the IR level LLVM already has pattern matching helpers for
identifying min/max idioms, here is part of a transform using
this, from InstructionSimplify.cpp:

   // Signed variants on "max(a,b)>=a -> true".
   if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
     if (A != RHS) std::swap(A, B); // smax(A, B) pred A.
     EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
     // We analyze this as smax(A, B) pred A.
     P = Pred;
   }

At the codegen level, the same could be done.  Also, if a target has a
max/min instruction it can transform compare+select to max/min, in fact
I'm pretty sure some targets do this already.

Ciao, Duncan.

>
> Thanks,
>
>                         Yin
>
> -----Original Message-----
> From: llvmdev-bounces at cs.uiuc.edu [mailto:llvmdev-bounces at cs.uiuc.edu] On
> Behalf Of Duncan Sands
> Sent: Tuesday, October 30, 2012 12:38 AM
> To: llvmdev at cs.uiuc.edu
> Subject: Re: [LLVMdev] Any plan to add MIN/MAX isd node?
>
> Hi Yin,
>
>>         Do you have any plan to add MIN/MAX in ISD node?
>>
>> Atomic operations have Min/Max operator already. This should
>>
>> be easy to add Min/Max node in ISD list. Because after
>>
>> converting control flow into Min/Max operator, it will
>>
>> result in better optimization because associative attribute
>>
>> and can be moved around.
>>
>>        Could you give some opinions on this issue?
>
> what are the advantages over using "select" (or one of its codegen
> variants)?
>
> Ciao, Duncan.
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>




More information about the llvm-dev mailing list