AMDGPU: Avoid creating vector extracts if we aren't going to do anything

Try to avoid expensive checks failures from reporting no changes
when some dead instructions were introduced.

Author: arsenm

llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp

@@ -2182,19 +2182,18 @@ bool AMDGPUCodeGenPrepareImpl::visitSqrt(IntrinsicInst &Sqrt) {
       isOneOrNegOne(FDiv->getOperand(0)))
     return false;
 
-  IRBuilder<> Builder(&Sqrt);
   Value *SrcVal = Sqrt.getOperand(0);
-
-  SmallVector<Value *, 4> SrcVals;
-  extractValues(Builder, SrcVals, SrcVal);
-
   bool CanTreatAsDAZ = canIgnoreDenormalInput(SrcVal, &Sqrt);
 
   // The raw instruction is 1 ulp, but the correction for denormal handling
   // brings it to 2.
   if (!CanTreatAsDAZ && ReqdAccuracy < 2.0f)
     return false;
 
+  IRBuilder<> Builder(&Sqrt);
+  SmallVector<Value *, 4> SrcVals;
+  extractValues(Builder, SrcVals, SrcVal);
+
   SmallVector<Value *, 4> ResultVals(SrcVals.size());
   for (int I = 0, E = SrcVals.size(); I != E; ++I) {
     if (CanTreatAsDAZ)