[llvm-commits] [llvm] r166865 - /llvm/trunk/lib/Transforms/Vectorize/BBVectorize.cpp

[Hal Finkel]

----- Original Message -----
> From: "Nadav Rotem" <nrotem at apple.com>
> To: "Hal Finkel" <hfinkel at anl.gov>
> Cc: llvm-commits at cs.uiuc.edu
> Sent: Friday, October 26, 2012 11:42:55 PM
> Subject: Re: [llvm-commits] [llvm] r166865 - /llvm/trunk/lib/Transforms/Vectorize/BBVectorize.cpp
> 
> Hi Hal,
> 
> I am running some benchmarks and I see strange results.  The x86
> backend is weird. Expensive integer arithmetic ops <8 x i32> which
> are unsupported on AVX, are marked as 'custom' because they have
> custom lowering.  We don't have a good way of knowing that  'ADD <8
> x i32>' is actually slow on AVX.  On the other hand, easy operations
> such as FCMP <4 x float>, which should be straight forward, are
> marked as 'Expand', and they are lowered in a DAGCombine op.

To be fair, I think we expected this. I see two general solutions:
 1. Cover all such cases with target hooks
 2. When the cost of some operation is requested, actually compile it (make a small selection DAG and run instruction selection), and then use the result (we could just count the number of machine instructions, for example). We'd need to cache the results for this to be practical.

What do you think?

 -Hal

> 
> Nadav
> 
> On Oct 26, 2012, at 9:33 PM, Hal Finkel <hfinkel at anl.gov> wrote:
> 
> > Author: hfinkel
> > Date: Fri Oct 26 23:33:48 2012
> > New Revision: 166865
> > 
> > URL: http://llvm.org/viewvc/llvm-project?rev=166865&view=rev
> > Log:
> > Update BBVectorize to use the new VTTI instr. cost interfaces.
> > 
> > The monolithic interface for instruction costs has been split into
> > several functions. This is the corresponding change. No
> > functionality
> > change is intended.
> > 
> > Modified:
> >    llvm/trunk/lib/Transforms/Vectorize/BBVectorize.cpp
> > 
> > Modified: llvm/trunk/lib/Transforms/Vectorize/BBVectorize.cpp
> > URL:
> > http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Vectorize/BBVectorize.cpp?rev=166865&r1=166864&r2=166865&view=diff
> > ==============================================================================
> > --- llvm/trunk/lib/Transforms/Vectorize/BBVectorize.cpp (original)
> > +++ llvm/trunk/lib/Transforms/Vectorize/BBVectorize.cpp Fri Oct 26
> > 23:33:48 2012
> > @@ -484,6 +484,61 @@
> >       return 1;
> >     }
> > 
> > +    // Returns the cost of the provided instruction using VTTI.
> > +    // This does not handle loads and stores.
> > +    unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) {
> > +      switch (Opcode) {
> > +      default: break;
> > +      case Instruction::GetElementPtr:
> > +        // We mark this instruction as zero-cost because scalar
> > GEPs are usually
> > +        // lowered to the intruction addressing mode. At the
> > moment we don't
> > +        // generate vector GEPs.
> > +        return 0;
> > +      case Instruction::Br:
> > +        return VTTI->getCFInstrCost(Opcode);
> > +      case Instruction::PHI:
> > +        return 0;
> > +      case Instruction::Add:
> > +      case Instruction::FAdd:
> > +      case Instruction::Sub:
> > +      case Instruction::FSub:
> > +      case Instruction::Mul:
> > +      case Instruction::FMul:
> > +      case Instruction::UDiv:
> > +      case Instruction::SDiv:
> > +      case Instruction::FDiv:
> > +      case Instruction::URem:
> > +      case Instruction::SRem:
> > +      case Instruction::FRem:
> > +      case Instruction::Shl:
> > +      case Instruction::LShr:
> > +      case Instruction::AShr:
> > +      case Instruction::And:
> > +      case Instruction::Or:
> > +      case Instruction::Xor:
> > +        return VTTI->getArithmeticInstrCost(Opcode, T1);
> > +      case Instruction::Select:
> > +      case Instruction::ICmp:
> > +      case Instruction::FCmp:
> > +        return VTTI->getCmpSelInstrCost(Opcode, T1, T2);
> > +      case Instruction::ZExt:
> > +      case Instruction::SExt:
> > +      case Instruction::FPToUI:
> > +      case Instruction::FPToSI:
> > +      case Instruction::FPExt:
> > +      case Instruction::PtrToInt:
> > +      case Instruction::IntToPtr:
> > +      case Instruction::SIToFP:
> > +      case Instruction::UIToFP:
> > +      case Instruction::Trunc:
> > +      case Instruction::FPTrunc:
> > +      case Instruction::BitCast:
> > +        return VTTI->getCastInstrCost(Opcode, T1, T2);
> > +      }
> > +
> > +      return 1;
> > +    }
> > +
> >     // This determines the relative offset of two loads or stores,
> >     returning
> >     // true if the offset could be determined to be some constant
> >     value.
> >     // For example, if OffsetInElmts == 1, then J accesses the
> >     memory directly
> > @@ -834,11 +889,11 @@
> >         return false;
> >       }
> >     } else if (VTTI) {
> > -      unsigned ICost = VTTI->getInstrCost(I->getOpcode(), IT1,
> > IT2);
> > -      unsigned JCost = VTTI->getInstrCost(J->getOpcode(), JT1,
> > JT2);
> > +      unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
> > +      unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
> >       Type *VT1 = getVecTypeForPair(IT1, JT1),
> >            *VT2 = getVecTypeForPair(IT2, JT2);
> > -      unsigned VCost = VTTI->getInstrCost(I->getOpcode(), VT1,
> > VT2);
> > +      unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2);
> > 
> >       if (VCost > ICost + JCost)
> >         return false;
> > 
> > 
> > _______________________________________________
> > llvm-commits mailing list
> > llvm-commits at cs.uiuc.edu
> > http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> 
> 

-- 
Hal Finkel
Postdoctoral Appointee
Leadership Computing Facility
Argonne National Laboratory