[NFC][LoopVectorize] Allow getRuntimeVF to generate FP values

[david-arm]

Summary

Given there is already code in the vectoriser that sometimes needs
to create a floating-point value for the runtime VF, it makes sense
to move the logic for the int->FP conversions into getRuntimeVF
itself.

This will also be useful in future when we want to add tail-folding
support for scalable vectors.

[lebedev.ri]

Passing-by remark: i don't see how this makes sense.
VF is inherently an integral quantity, if something needs it as a float, should it not deal with the conversion?

[david-arm]

In D111873#3066465, @lebedev.ri wrote:

Passing-by remark: i don't see how this makes sense.
VF is inherently an integral quantity, if something needs it as a float, should it not deal with the conversion?

It's because this will become a frequent pattern in future work, where we have to follow the same pattern of 1) convert float type -> int type, 2) invoke getRuntimeVF, 3) convert result back to FP. I wrote this patch on the basis that we'd prefer to reduce code duplication scattered throughout the vectoriser. It will make future patches in the area of tail-folding smaller.

[paulwalker-arm]

What about making the patch additive only. So keep getRuntimeVF as is and add something like getRuntimeVFasFloat as the helper function? Just a thought. Is the patch which introduces the multiple uses big? If not then I don't see a reason for not introducing the helper function there, that way it'll be clear why you need it.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

1125

Should this be B.CreateUIToFP given VF is unsigned?

[david-arm]

Thanks for the comments @lebedev.ri and @paulwalker-arm, I'll have a think about possibly restructuring this. I will go back and double check which code paths will absolutely need a FP value.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

Constant *StepVal = ConstantInt::get(
       Step->getType(),
       cast<ConstantInt>(Step)->getSExtValue() * VF.getKnownMinValue());
   return VF.isScalable() ? B.CreateVScale(StepVal) : StepVal;
 }
 
 namespace llvm {
 
 /// Return the runtime value for VF.
-Value *getRuntimeVF(IRBuilder<> &B, Type *Ty, ElementCount VF) {
+Value *getRuntimeVF(IRBuilder<> &B, Type *OrigTy, ElementCount VF) {
+  Type *Ty = OrigTy;
+  if (Ty->isFloatingPointTy())
+    Ty = IntegerType::get(Ty->getContext(), Ty->getScalarSizeInBits());
+
   Constant *EC = ConstantInt::get(Ty, VF.getKnownMinValue());
-  return VF.isScalable() ? B.CreateVScale(EC) : EC;
+  Value *Res = VF.isScalable() ? B.CreateVScale(EC) : EC;
+
+  if (OrigTy->isFloatingPointTy())
+    Res = B.CreateSIToFP(Res, OrigTy);
+
+  return Res;
 }
 
 void reportVectorizationFailure(const StringRef DebugMsg,
                                 const StringRef OREMsg, const StringRef ORETag,
                                 OptimizationRemarkEmitter *ORE, Loop *TheLoop,
                                 Instruction *I) {
   LLVM_DEBUG(debugVectorizationMessage("Not vectorizing: ", DebugMsg, I));
   LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE);
if (Step->getType()->isIntegerTy()) {
     MulOp = Instruction::Mul;
   } else {
     AddOp = II.getInductionOpcode();
     MulOp = Instruction::FMul;
   }
 
   // Multiply the vectorization factor by the step using integer or
   // floating-point arithmetic as appropriate.
-  Type *StepType = Step->getType();
-  if (Step->getType()->isFloatingPointTy())
-    StepType = IntegerType::get(StepType->getContext(),
-                                StepType->getScalarSizeInBits());
-  Value *RuntimeVF = getRuntimeVF(Builder, StepType, VF);
-  if (Step->getType()->isFloatingPointTy())
-    RuntimeVF = Builder.CreateSIToFP(RuntimeVF, Step->getType());
+  Value *RuntimeVF = getRuntimeVF(Builder, Step->getType(), VF);
   Value *Mul = Builder.CreateBinOp(MulOp, Step, RuntimeVF);
 
   // Create a vector splat to use in the induction update.
   //
   // FIXME: If the step is non-constant, we create the vector splat with
   //        IRBuilder. IRBuilder can constant-fold the multiply, but it doesn't
   //        handle a constant vector splat.
   Value *SplatVF = isa<Constant>(Mul)