[llvm-commits] [llvm] r169025 - /llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp

[Michael Ilseman]

On Nov 30, 2012, at 1:00 PM, David Blaikie <dblaikie at gmail.com> wrote:

> On Fri, Nov 30, 2012 at 11:19 AM, Michael Ilseman <milseman at apple.com> wrote:
>> Very minor tweak:
>> 
>>> +      Instruction *FMul = dyn_cast<Instruction>(FMulVal);
>>> +      assert(FMul && "Must be instruction as Log2 is instruction");
>> 
>> You can write this as:
>> 
>> assert(isa<Instruction>(FMulVal) && "Must be instruction as Log2 is instruction");
>> Instruction *FMul = cast<Instruction>(FMulVal);
> 
> & in this case you can drop the assert entirely, since cast already
> does the assert.
> 

That's true. Having the assert here lets you specify a more specific/helpful message, but it's not necessary unless there's a very specific assumption you're trying to point out.

>> 
>> dyn_cast consists of both an isa<> followed by a cast<>. If you're just asserting, then you can split it out yourself to make non-assert builds just do the cast<>.
>> 
>> Thanks for the fast-math contribution!
>> 
>> On Nov 30, 2012, at 11:09 AM, Pedro Artigas <partigas at apple.com> wrote:
>> 
>>> Author: partigas
>>> Date: Fri Nov 30 13:09:41 2012
>>> New Revision: 169025
>>> 
>>> URL: http://llvm.org/viewvc/llvm-project?rev=169025&view=rev
>>> Log:
>>> Add fast math inst combine X*log2(Y*0.5)-->X*log2(Y)-X
>>> 
>>> reviewed by Michael Ilseman <milseman at apple.com>
>>> 
>>> 
>>> Modified:
>>>   llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
>>> 
>>> Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp?rev=169025&r1=169024&r2=169025&view=diff
>>> ==============================================================================
>>> --- llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp (original)
>>> +++ llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp Fri Nov 30 13:09:41 2012
>>> @@ -284,6 +284,83 @@
>>>    if (Value *Op1v = dyn_castFNegVal(Op1))
>>>      return BinaryOperator::CreateFMul(Op0v, Op1v);
>>> 
>>> +  // Under unsafe algebra do:
>>> +  // X * log2(0.5*Y) = X*log2(Y) - X
>>> +  if (I.hasUnsafeAlgebra()) {
>>> +    Value *OpX = NULL;
>>> +    Value *OpY = NULL;
>>> +    IntrinsicInst *Log2;
>>> +    if (Op0->hasOneUse()) {
>>> +      if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
>>> +        if (II->getIntrinsicID() == Intrinsic::log2 &&
>>> +            II->hasUnsafeAlgebra())
>>> +        {
>>> +          Log2 = II;
>>> +          Value *OpLog2Of = II->getArgOperand(0);
>>> +          if (OpLog2Of->hasOneUse()) {
>>> +            if (Instruction *I = dyn_cast<Instruction>(OpLog2Of)) {
>>> +              if (I->getOpcode() == Instruction::FMul &&
>>> +                  I->hasUnsafeAlgebra())
>>> +              {
>>> +                ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(0));
>>> +                if (CFP && CFP->isExactlyValue(0.5)) {
>>> +                  OpY = I->getOperand(1);
>>> +                  OpX = Op1;
>>> +                } else {
>>> +                  CFP = dyn_cast<ConstantFP>(I->getOperand(1));
>>> +                  if (CFP && CFP->isExactlyValue(0.5)) {
>>> +                    OpY = I->getOperand(0);
>>> +                    OpX = Op1;
>>> +                  }
>>> +                }
>>> +              }
>>> +            }
>>> +          }
>>> +        }
>>> +      }
>>> +    }
>>> +    if (Op1->hasOneUse()) {
>>> +      if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op1)) {
>>> +        if (II->getIntrinsicID() == Intrinsic::log2 &&
>>> +            II->hasUnsafeAlgebra())
>>> +        {
>>> +          Log2 = II;
>>> +          Value *OpLog2Of = II->getArgOperand(0);
>>> +          if (OpLog2Of->hasOneUse()) {
>>> +            if (Instruction *I = dyn_cast<Instruction>(OpLog2Of)) {
>>> +              if (I->getOpcode() == Instruction::FMul &&
>>> +                  I->hasUnsafeAlgebra())
>>> +              {
>>> +                ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(0));
>>> +                if (CFP && CFP->isExactlyValue(0.5)) {
>>> +                  OpY = I->getOperand(1);
>>> +                  OpX = Op0;
>>> +                } else {
>>> +                  CFP = dyn_cast<ConstantFP>(I->getOperand(1));
>>> +                  if (CFP && CFP->isExactlyValue(0.5)) {
>>> +                    OpY = I->getOperand(0);
>>> +                    OpX = Op0;
>>> +                  }
>>> +                }
>>> +              }
>>> +            }
>>> +          }
>>> +        }
>>> +      }
>>> +    }
>>> +    // if pattern detected emit alternate sequence
>>> +    if (OpX && OpY) {
>>> +      Log2->setArgOperand(0, OpY);
>>> +      Value *FMulVal = Builder->CreateFMul(OpX, Log2);
>>> +      Instruction *FMul = dyn_cast<Instruction>(FMulVal);
>>> +      assert(FMul && "Must be instruction as Log2 is instruction");
>>> +      FMul->copyFastMathFlags(Log2);
>>> +      Instruction *FSub = BinaryOperator::CreateFSub(FMulVal, OpX);
>>> +      FSub->copyFastMathFlags(Log2);
>>> +      return FSub;
>>> +    }
>>> +  }
>>> +
>>>  return Changed ? &I : 0;
>>> }
>>> 
>>> 
>>> 
>>> _______________________________________________
>>> llvm-commits mailing list
>>> llvm-commits at cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>> 
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits