[llvm-commits] [llvm] r136583 - /llvm/trunk/lib/Analysis/BranchProbabilityInfo.cpp

[Jakub Staszak]

-Kuba

On Jul 30, 2011, at 8:53 PM, Eli Friedman wrote:

> On Sat, Jul 30, 2011 at 8:27 PM, Jakub Staszak <jstaszak at apple.com> wrote:
>> Author: kuba
>> Date: Sat Jul 30 22:27:24 2011
>> New Revision: 136583
>> 
>> URL: http://llvm.org/viewvc/llvm-project?rev=136583&view=rev
>> Log:
>> Add Zero Heurestics to BranchProbabilityInfo. If we compare value to zero we
>> decide whether condition is likely to be true this way:
>> 
>> x == 0  ->  false
>> x <  0  ->  false
>> x <= 0  ->  false
>> x != 0  ->  true
>> x >  0  ->  true
>> x >= 0  ->  true
> 
> What's the justification for these heuristics?

It is one of the heuristics proposed in the paper "Static Branch Frequency and Program Profile Analysis". The other heuristics we have right now were also proposed there.

>> Modified:
>>    llvm/trunk/lib/Analysis/BranchProbabilityInfo.cpp
>> 
>> Modified: llvm/trunk/lib/Analysis/BranchProbabilityInfo.cpp
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/BranchProbabilityInfo.cpp?rev=136583&r1=136582&r2=136583&view=diff
>> ==============================================================================
>> --- llvm/trunk/lib/Analysis/BranchProbabilityInfo.cpp (original)
>> +++ llvm/trunk/lib/Analysis/BranchProbabilityInfo.cpp Sat Jul 30 22:27:24 2011
>> @@ -11,6 +11,7 @@
>>  //
>>  //===----------------------------------------------------------------------===//
>> 
>> +#include "llvm/Constants.h"
>>  #include "llvm/Instructions.h"
>>  #include "llvm/Analysis/BranchProbabilityInfo.h"
>>  #include "llvm/Analysis/LoopInfo.h"
>> @@ -72,6 +73,9 @@
>>   static const uint32_t PH_TAKEN_WEIGHT = 20;
>>   static const uint32_t PH_NONTAKEN_WEIGHT = 12;
>> 
>> +  static const uint32_t ZH_TAKEN_WEIGHT = 20;
>> +  static const uint32_t ZH_NONTAKEN_WEIGHT = 12;
>> +
>>   // Standard weight value. Used when none of the heuristics set weight for
>>   // the edge.
>>   static const uint32_t NORMAL_WEIGHT = 16;
>> @@ -125,6 +129,9 @@
>>   // Loop Branch Heuristics
>>   bool calcLoopBranchHeuristics(BasicBlock *BB);
>> 
>> +  // Zero Heurestics
>> +  bool calcZeroHeuristics(BasicBlock *BB);
>> +
>>   bool runOnFunction(Function &F);
>>  };
>>  } // end anonymous namespace
>> @@ -270,6 +277,86 @@
>>   return true;
>>  }
>> 
>> +bool BranchProbabilityAnalysis::calcZeroHeuristics(BasicBlock *BB) {
>> +  BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
>> +  if (!BI || !BI->isConditional())
>> +    return false;
>> +
>> +  Value *Cond = BI->getCondition();
>> +  ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
>> +  if (!CI)
>> +    return false;
>> +
>> +  Value *LHS = CI->getOperand(0);
>> +  Value *RHS = CI->getOperand(1);
>> +
>> +  bool hasZero = false;
>> +  bool lhsZero = false;
>> +  if (ConstantInt *CI = dyn_cast<ConstantInt>(LHS)) {
>> +    hasZero = CI->isZero();
>> +    lhsZero = true;
>> +  }
> 
> The LHS of a compare will never be zero after instcombine runs... so
> this check strikes me as unlikely to be useful.
> 
So instcombiner always moves zero (or constants in general) to RHS?

> -Eli

- Kuba

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20110730/caaa7040/attachment.html>