Loop vectorization with FP induction variables

[delena]

Summary

Handling FP IVs without SCEV. Covered simple cases with constant and loop invariant step.

float *A;
float x = init;
for (int i=0; i < N; ++i) {

  A[i] = x;
  x -= fp_inc;
}

[sanjoy]

Some comments inline.

I cannot review the LoopVectorize.cpp changes.

lib/Analysis/ScalarEvolutionExpander.cpp

1610 ↗

(On Diff #60711)

Why do we need this?

lib/Transforms/Utils/LoopUtils.cpp

758 ↗

(On Diff #60711)

Looking at the users of isInductionPHI, looks like it should be easy to pass in the Loop * directly? If so, can we just have this check be: if (L->getHeader() != Phi->getParent()) return false; ? That way we won't have to add the getLoopFor interface to SCEV (which does not look like it belongs there).

[delena]

lib/Analysis/ScalarEvolutionExpander.cpp

1610 ↗

(On Diff #60711)

I defined Step as unknown SCEV, but underlying value is of floating point type.
In this case expand() just returns the underlying value, but it fails on one of these two lines.

lib/Transforms/Utils/LoopUtils.cpp

758 ↗

(On Diff #60711)

Yes, of course. I'll change.

[mkuper]

lib/Transforms/Utils/LoopUtils.cpp

675 ↗

(On Diff #60711)

Maybe ((IK == IK_FpInduction) || Step->getType()->isIntegerTy()) would be clearer.

678 ↗

(On Diff #60711)

Maybe put this on line 668, next to the two asserts for the other types?

733 ↗

(On Diff #60711)

As long as you're touching this - maybe hoist this assert for all 3 cases above the switch?

739 ↗

(On Diff #60711)

Why do we need this?

767 ↗

(On Diff #60711)

Some variable naming nitpicking:
Bb -> either BB or bb?
BEValueV -> BEValue
StartValueV -> StartValue

772 ↗

(On Diff #60711)

Do we really need the if here? We expect L->contains(BB) to be true for one of the incoming values, and false for the other, right? So if we're in the "else", we're already in a bad shape, regardless of whether BEValueV == V or not.
Or did I misunderstand?

776 ↗

(On Diff #60711)

Same as above.

780 ↗

(On Diff #60711)

Can we replace this with an early exit?

801 ↗

(On Diff #60711)

Are you sure this is always a safe place to insert? E.g. what if Addend is a PHI in an outer loop?

827 ↗

(On Diff #60711)

Why not isFloatingPointTy() here?

lib/Transforms/Vectorize/LoopVectorize.cpp

2154 ↗

(On Diff #60711)

ITy here stands for "IntegerTy", I guess. Perhaps rename, now that ITy can be float?
(Unless ITy stands for IndexTy, which also make sense, and in which case we should also probably rename. :-) )

2178 ↗

(On Diff #60711)

This looks a bit odd.
If I understand correctly, you're relying on the step being FNeg to distinguish whether the original direction of the loop was positive or negative (by transforming a phi that feeds an fsub by a phi that feeds an fadd with an fneg).
What happens if the scalar loop has an fadd, where the original step is a loop-invariant fneg?

[mkuper]

lib/Analysis/ScalarEvolutionExpander.cpp

1610 ↗

(On Diff #60711)

That sounds a bit weird to me.
It looks like the contract for expandCodeFor is that if you pass a type, you get the expansion cast to this type. What happens if you pass a type, but get a result that's from a different type than you passed (because it's not scevable)?

[delena]

Michael, thanks a lot for your comments. I'll upload a new patch.

lib/Analysis/ScalarEvolutionExpander.cpp

1610 ↗

(On Diff #60711)

I don't know why do we need to pass the type.
May be for casting between i64 and PointerType?
But an Unknown SCEV may have any type, right?

lib/Transforms/Utils/LoopUtils.cpp

678 ↗

(On Diff #60711)

All these checks belong to the Step type. See comment #669.

733 ↗

(On Diff #60711)

Ok

739 ↗

(On Diff #60711)

I should cover fsub operation somehow.
; for (int i=0; i < N; ++i) {
; A[i] = x;
; x -= fp_inc;
; }
I keep Step as Fneg(fp_inc) in this case.

772 ↗

(On Diff #60711)

I just copied this code from integer Phi together with weird variable names.
Can we have multiple bbs inside loop?

801 ↗

(On Diff #60711)

This FNeg disappears after transformation. It is just a way to distinguish between FAdd and FSub.

827 ↗

(On Diff #60711)

isFloatingPointTy() includes more than half, float and double. I'm not sure I want to cover other types.

lib/Transforms/Vectorize/LoopVectorize.cpp

2178 ↗

(On Diff #60711)

I don't see any problem. x + (-a) is equal to x - a.

[mkuper]

Thanks, Elena.

I guess there's one thing I don't understand about the fnegs - why are you using an IR instruction as the marker of whether the original induction had an fadd or an fsub, instead of a property of the IV? It would make sense to me if you actually used the fneg to feed the vector induction, thus simplifying the code (not having to special-case the sub), but instead you special-case it anyway by looking through the fneg.

Another thing I forgot earlier - this should only fire when we're in unsafe/fast math mode, right? Is there a check for that?

lib/Transforms/Utils/LoopUtils.cpp

739 ↗

(On Diff #60711)

I understand, what I mean is - let's say the step is an FNeg.
Why can't you feed the FNeg directly into the CreateFMul? Do we get worse code?

772 ↗

(On Diff #60711)

Right now, I think not. (and that needs to be fixed.) :-\
Anyway, if this is a verbatim copy from the int case, leave it be for now. Should probably be fixed separately.
(Of course, that raises the question - can the int and the fp case share the code? Or is it not worth it?)

801 ↗

(On Diff #60711)

When does it disappear? If it disappears in later clean-up, we still don't want to insert it at an illegal location (like between two phis).
If it's guaranteed to be deleted during the run of the vectorizer, I guess this will technically work, but I'd still prefer to avoid it.

lib/Transforms/Vectorize/LoopVectorize.cpp

2178 ↗

(On Diff #60711)

What I'm trying to say is that it's weird that the behavior would be different based on whether the step is an fneg. It could be an fneg because you added an fneg, it could be an fneg because there was already an fneg. Why does this code look through the fneg? It doesn't seem like this should be the vectorizer's job.
Although if the fneg won't get cleaned up later, this is probably the right thing to do.

[delena]

In D21330#462655, @mkuper wrote:

I guess there's one thing I don't understand about the fnegs - why are you using an IR instruction as the marker of whether the original induction had an fadd or an fsub, instead of a property of the IV? It would make sense to me if you actually used the fneg to feed the vector induction, thus simplifying the code (not having to special-case the sub), but instead you special-case it anyway by looking through the fneg.

What do you mean by "property of IV" ? Do you suggest to add a special field to InductionDescriptor?
Fneg should be a part of Step. But Step is a SCEV and I was not allowed to implement FP SCEV.

[delena]

In D21330#462655, @mkuper wrote:

Another thing I forgot earlier - this should only fire when we're in unsafe/fast math mode, right? Is there a check for that?

FP reduction is allowed for +/- operators. Only max/min checks for unsafe.

[mkuper]

What do you mean by "property of IV" ? Do you suggest to add a special field to InductionDescriptor?
Fneg should be a part of Step. But Step is a SCEV and I was not allowed to implement FP SCEV.

You know, I probably just don't understand the SCEV situation well enough.
Sanjoy, any chance you could take a look, even though part of it is in LoopVectorize.cpp? :-)

[delena]

1. Changed handling of FSUB operation. I keep the original binary operation inside Induction Descriptor.
2. Allow FP induction in fast-math mode only.

[delena]

Ping *

[sanjoy]

Hi Elena,

Sorry for the delay, I was out on vacation. I'm catching up on email now, and I will try to get to review this this week.

Thanks!

[sanjoy]

Some comments inline. Please let me know if you want me to look at something specific, but I'm not familiar enough with the code this patch touches to lgtm it.

../lib/Analysis/ScalarEvolutionExpander.cpp

1614

This is odd -- is it just to help keep the Step as a SCEV *? If so, I'd suggest solving that within InductionDescriptor itself (i.e. maybe support having the step as either a SCEV * or a Value *, depending on the type of the InductionDescriptor?).

../lib/Transforms/Utils/LoopUtils.cpp

762

(Not for fixing in this change) looks like a better interface would be to return an Optional< InductionDescriptor>?

787

Maybe use a dyn_cast here?

811

The condition looks inverted?

[delena]

../lib/Analysis/ScalarEvolutionExpander.cpp

1614

I't dropping this change, I don't need it anymore.

../lib/Transforms/Utils/LoopUtils.cpp

787

Ok.

811

The hasUnsafeAlgebra() means that instruction itself has "fast" attribute. In this case we don't need additional check. But if the BOp does not have the "fast" attribute, the legality of FP transformation should be allowed on function level. I'll add a comment.

[delena]

Some changes according to Sanjoy's comment. Thanks Sanjoy and Michael for review.
Still looking for somebody who can accept this patch. Added @sbaranga, who made similar changes in induction variables.

[mkuper]

I want to go over the code again, after the changes, but the reason I don't feel like I can accept the patch is because I wasn't part of the original FP SCEV discussion, and I'm not sure I understand the design considerations.

If someone - e.g. sanjoy - OK's the design, I can LGTM the LV code change.

[mkuper]

../include/llvm/Transforms/Utils/LoopUtils.h

274

I think Instruction::BinaryOpsEnd may be better for an explicitly invalid BinaryOp. Not sure that's a good choice, but pretty sure 0 isn't.

../lib/Transforms/Utils/LoopUtils.cpp

787

I think what sanjoy meant was:

BinaryOperator *BOp = dyn_cast<BinaryOperator>(BEValue);
if (!BOp)
  return false;

../lib/Transforms/Vectorize/LoopVectorize.cpp

4347

Main -> Primary (I think we use that consistently)

6659

Can you add a test for this? All of the tests you added force UF == 1.

6662

Are you sure about this? I mean, it's true for vectorizing, but is it true here as well?
(I'm not saying it isn't, just making sure this is intentional)

[delena]

In D21330#487162, @mkuper wrote:

I want to go over the code again, after the changes, but the reason I don't feel like I can accept the patch is because I wasn't part of the original FP SCEV discussion, and I'm not sure I understand the design considerations.

The bottom line of the FP SCEV discussion was the point that FP SCEV is overkill for "secondary" IV (like in the example above). We'll need FP SCEV for primary FP IV like
for (float f =0.0; f < g; f+=0.5) {}.
But such loops are rare and most of them can be re-mapped to integers.
The suggestion was to include FP IV to the current InductionDescriptor.

[sanjoy]

In D21330#487357, @delena wrote:

In D21330#487162, @mkuper wrote:

I want to go over the code again, after the changes, but the reason I don't feel like I can accept the patch is because I wasn't part of the original FP SCEV discussion, and I'm not sure I understand the design considerations.

The bottom line of the FP SCEV discussion was the point that FP SCEV is overkill for "secondary" IV (like in the example above). We'll need FP SCEV for primary FP IV like
for (float f =0.0; f < g; f+=0.5) {}.
But such loops are rare and most of them can be re-mapped to integers.
The suggestion was to include FP IV to the current InductionDescriptor.

+1

[delena]

../include/llvm/Transforms/Utils/LoopUtils.h

274

Thanks, I'll fix.

../lib/Transforms/Vectorize/LoopVectorize.cpp

6662

Even if you have only unrolling, and VF is 1, the value of FP induction is calculated as:
sitofp(PrimaryIV) * Increment.

for (int i=0; i<N; i++) {
  fp_ind += fp_inc;
}

is transferred to something like this:

init = fp_inc;
for (int i=0; i<N; i++) {
  fp_ind = init + i*fp_inc; 
}

In this case we need unsafe math.
I added tests for unrolling.

[delena]

Updated the code according to Michael's comments.

[mkuper]

LGTM

[anemet]

In D21330#488606, @mkuper wrote:

LGTM

I've just started looking at this too. Please give me a few mins. So far I only encountered minor things.

../include/llvm/Transforms/Utils/LoopUtils.h

308

Fp->FP

328–329

binary op -> induction op is better everywhere. Also I am assuming this is the op that advances the induction variable. You may want to spell this out somewhere.

../lib/Transforms/Utils/LoopUtils.cpp

770

this function is returning a bool

[mkuper]

Sorry, didn't realize anyone else is still interested in looking at this, given how long this patch has been up.
Ignore my LGTM. :-)

[anemet]

In D21330#488622, @mkuper wrote:

Sorry, didn't realize anyone else is still interested in looking at this, given how long this patch has been up.
Ignore my LGTM. :-)

No, *I am* sorry for chiming in this late. I felt obliged because you mentioned that no one looked at this from the original llvm-dev thread and I felt bad ;). And thanks for reviewing it, this is looking pretty good.

[anemet]

So with these it should LGTM too. I haven't checked everything (most notably the unsafe math part).

I just wanted to see whether this was in line with the direction set in the original llvm-dev thread and it is! Thanks to all of you and sorry about the delay again.

../lib/Transforms/Utils/LoopUtils.cpp

816–818

I think that a better interface would be to take BOp (step instruction) optionally and then derive DI::hasUnsafeAlgebra and the opcode from that. This is OK as a follow-up if you prefer.

../lib/Transforms/Vectorize/LoopVectorize.cpp

4354–4355

by adding StepVal, you mean

[delena]

Minor changes according to Adam's comments.

[delena]

../include/llvm/Transforms/Utils/LoopUtils.h

328–329

Ok. Changing it to getInductionBinOp()

../lib/Transforms/Utils/LoopUtils.cpp

770

Fixed. Thanks.

816–818

I did it at the beginning and then prefer to follow Reduction implementation, just to be consistent with existing interface.

../lib/Transforms/Vectorize/LoopVectorize.cpp

4354–4355

I fixed the comment. thanks.

[delena]

Added one more test where init value and step are constants.

[anemet]

You should probably also have a test for the case where there is not "fast" attribute on the step instruction but we can still vectorize with the hints.

../include/llvm/Transforms/Utils/LoopUtils.h

321–323

This comment seems incorrect/misleading. I think this flag "allows" a relaxed FP model not "requires" it. Do you agree?

328–329

I think you missed the part about "everywhere", i.e. making the member variable consistent with this name too: BinaryOp -> InductionOp

343–344

This comment is also ambiguous. This is again the induction/step instruction *iff* the instruction allows for a relaxed FP model.

../lib/Transforms/Utils/LoopUtils.cpp

816–818

Fair but it does not make sense to pass a different instructions for UAI and BO and the current interface allows for that. At least then change the ctor to take a single instruction and derive UAI and BO from that.

../test/Transforms/LoopVectorize/float-induction.ll

13–14

Where is the result of these guys used? Would be good to check too.

17–18

Why not also match the result of the fsub and check it in the store?

[delena]

Updated the patch according to Adam's comments.

[anemet]

../include/llvm/Transforms/Utils/LoopUtils.h

343

Comments are full sentences, please end with a period.

../test/Transforms/LoopVectorize/float-induction.ll

5

I *think* you can only have this under LoopVectorize/X86 (what if the X86 backend is not enabled in a build?). But more importantly, I don't understand why you need to formulate the non-fast-math case as an x86 test.

[delena]

../include/llvm/Transforms/Utils/LoopUtils.h

343

ok.

../test/Transforms/LoopVectorize/float-induction.ll

5

The "unsafe" function attribute works only for auto-vec. If you specify -force-vector-width=4 the loop will be vectorized anyway. So I added "X86" tests just to check combination of function attribute and "safe" FP induction in the auto-vectorization mode.

[anemet]

LGTM with the x86 test split into its own test under Transform/LoopVectorize/X86. Thanks!

../test/Transforms/LoopVectorize/float-induction.ll

5

Hmm, that is somewhat unexpected. I thought that the -force* stuff only overruled the cost model... @hfinkel, is it intentional/desired that -force-vector-width=>1 overrules (some of) the legality checks? We have different ways to bypass the legality checks so perhaps keeping -force-vector-width as a way to overrule the profitability checks is more desired?

I guess one concerns is that this internal flag was advertised on Nadav's Auto-vectorizer blog post so perhaps changing the behavior is not trivial at this point.

Anyhow, coming back to this patch, you need then to split this part of the test into a Transform/LoopVectorize/X86 test.

../include/llvm/Transforms/Utils/LoopUtils.h

 
 /// A struct for saving information about induction variables.
 class InductionDescriptor {
 public:
   /// This enum represents the kinds of inductions that we support.
   enum InductionKind {
     IK_NoInduction,  ///< Not an induction variable.
     IK_IntInduction, ///< Integer induction variable. Step = C.
-    IK_PtrInduction  ///< Pointer induction var. Step = C / sizeof(elem).
+    IK_PtrInduction, ///< Pointer induction var. Step = C / sizeof(elem).
+    IK_FpInduction   ///< Floating point induction variable.
   };
 
 public:
   /// Default constructor - creates an invalid induction.
   InductionDescriptor()
-      : StartValue(nullptr), IK(IK_NoInduction), Step(nullptr) {}
+    : StartValue(nullptr), IK(IK_NoInduction), Step(nullptr),
+    UnsafeAlgebraInst(nullptr), BinaryOp((Instruction::BinaryOps)0) {}
 
   /// Get the consecutive direction. Returns:
   ///   0 - unknown or non-consecutive.
   ///   1 - consecutive and increasing.
   ///  -1 - consecutive and decreasing.
   int getConsecutiveDirection() const;
 
   /// Compute the transformed value of Index at offset StartValue using step
   /// StepValue.
   /// For integer induction, returns StartValue + Index * StepValue.
   /// For pointer induction, returns StartValue[Index * StepValue].
   /// FIXME: The newly created binary instructions should contain nsw/nuw
   /// flags, which can be found from the original scalar operations.
   Value *transform(IRBuilder<> &B, Value *Index, ScalarEvolution *SE,
                    const DataLayout& DL) const;
 
   Value *getStartValue() const { return StartValue; }
   InductionKind getKind() const { return IK; }
   const SCEV *getStep() const { return Step; }
   ConstantInt *getConstIntStepValue() const;
 
-  /// Returns true if \p Phi is an induction. If \p Phi is an induction,
-  /// the induction descriptor \p D will contain the data describing this
-  /// induction. If by some other means the caller has a better SCEV
+  /// Returns true if \p Phi is an induction in the loop \p L. If \p Phi is an
+  /// induction, the induction descriptor \p D will contain the data describing
+  /// this induction. If by some other means the caller has a better SCEV
   /// expression for \p Phi than the one returned by the ScalarEvolution
   /// analysis, it can be passed through \p Expr.
-  static bool isInductionPHI(PHINode *Phi, ScalarEvolution *SE,
+  static bool isInductionPHI(PHINode *Phi, const Loop* L, ScalarEvolution *SE,
                              InductionDescriptor &D,
                              const SCEV *Expr = nullptr);
 
-  /// Returns true if \p Phi is an induction, in the context associated with
-  /// the run-time predicate of PSE. If \p Assume is true, this can add further
-  /// SCEV predicates to \p PSE in order to prove that \p Phi is an induction.
+  /// Returns true if \p Phi is a floating point induction in the loop \p L.
+  /// If \p Phi is an induction, the induction descriptor \p D will contain 
+  /// the data describing this induction.
+  static bool isFpInductionPHI(PHINode *Phi, const Loop* L,
+                               ScalarEvolution *SE, InductionDescriptor &D);
+
+  /// Returns true if \p Phi is a loop \p L induction, in the context associated
+  /// with the run-time predicate of PSE. If \p Assume is true, this can add
+  /// further SCEV predicates to \p PSE in order to prove that \p Phi is an
+  /// induction.
   /// If \p Phi is an induction, \p D will contain the data describing this
   /// induction.
-  static bool isInductionPHI(PHINode *Phi, PredicatedScalarEvolution &PSE,
+  static bool isInductionPHI(PHINode *Phi, const Loop* L,
+                             PredicatedScalarEvolution &PSE,
                              InductionDescriptor &D, bool Assume = false);
 
+  /// Returns true if the induction has unsafe algebra which requires a relaxed
+  /// floating-point model.
+  bool hasUnsafeAlgebra() { return UnsafeAlgebraInst != nullptr; }
+
+  /// Returns unsafe algebra instruction.
+  Instruction *getUnsafeAlgebraInst() { return UnsafeAlgebraInst; }
+
+  /// Returns binary opcode.
+  Instruction::BinaryOps getBinaryOpcode() const { return BinaryOp; }
+
 private:
   /// Private constructor - used by \c isInductionPHI.
-  InductionDescriptor(Value *Start, InductionKind K, const SCEV *Step);
+  InductionDescriptor(Value *Start, InductionKind K, const SCEV *Step,
+                      Instruction *UAI = nullptr,
+                      Instruction::BinaryOps BOp = (Instruction::BinaryOps)0);
 
   /// Start value.
   TrackingVH<Value> StartValue;
   /// Induction kind.
   InductionKind IK;
   /// Step value.
   const SCEV *Step;
+  // Induction has unsafe algebra.
+  Instruction *UnsafeAlgebraInst;
+  // Original Induction opcode.
+  Instruction::BinaryOps BinaryOp;
 };
 
 BasicBlock *InsertPreheaderForLoop(Loop *L, DominatorTree *DT, LoopInfo *LI,
                                    bool PreserveLCSSA);
 
 /// \brief Simplify each loop in a loop nest recursively.
 ///
 /// This takes a potentially un-simplified loop L (and its children) and turns

../lib/Analysis/ScalarEvolutionExpander.cpp

Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty,
   assert(IP);
   Builder.SetInsertPoint(IP);
   return expandCodeFor(SH, Ty);
 }
 
 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) {
   // Expand the code for this SCEV.
   Value *V = expand(SH);
-  if (Ty) {
+  if (Ty && SE.isSCEVable(Ty)) {
     assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
            "non-trivial casts should be done with the SCEVs directly!");
     V = InsertNoopCastOfTo(V, Ty);
   }
   return V;
 }
 
 Value *SCEVExpander::FindValueInExprValueMap(const SCEV *S,

../lib/Transforms/Scalar/LoopInterchange.cpp

bool LoopInterchangeLegality::findInductionAndReductions(
     Loop *L, SmallVector<PHINode *, 8> &Inductions,
     SmallVector<PHINode *, 8> &Reductions) {
   if (!L->getLoopLatch() || !L->getLoopPredecessor())
     return false;
   for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
     RecurrenceDescriptor RD;
     InductionDescriptor ID;
     PHINode *PHI = cast<PHINode>(I);
-    if (InductionDescriptor::isInductionPHI(PHI, SE, ID))
+    if (InductionDescriptor::isInductionPHI(PHI, L, SE, ID))
       Inductions.push_back(PHI);
     else if (RecurrenceDescriptor::isReductionPHI(PHI, L, RD))
       Reductions.push_back(PHI);
     else {
       DEBUG(
           dbgs() << "Failed to recognize PHI as an induction or reduction.\n");
       return false;
     }

../lib/Transforms/Utils/LoopUtils.cpp

Value *RecurrenceDescriptor::createMinMaxOp(IRBuilder<> &Builder,
   else
     Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
 
   Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
   return Select;
 }
 
 InductionDescriptor::InductionDescriptor(Value *Start, InductionKind K,
-                                         const SCEV *Step)
-  : StartValue(Start), IK(K), Step(Step) {
+                                         const SCEV *Step, Instruction *UAI,
+                                         Instruction::BinaryOps BOp)
+  : StartValue(Start), IK(K), Step(Step), UnsafeAlgebraInst(UAI),
+    BinaryOp(BOp) {
   assert(IK != IK_NoInduction && "Not an induction");
 
   // Start value type should match the induction kind and the value
   // itself should not be null.
   assert(StartValue && "StartValue is null");
   assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
          "StartValue is not a pointer for pointer induction");
   assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
          "StartValue is not an integer for integer induction");
 
   // Check the Step Value. It should be non-zero integer value.
   assert((!getConstIntStepValue() || !getConstIntStepValue()->isZero()) &&
          "Step value is zero");
 
   assert((IK != IK_PtrInduction || getConstIntStepValue()) &&
          "Step value should be constant for pointer induction");
-  assert(Step->getType()->isIntegerTy() && "StepValue is not an integer");
+  assert((IK == IK_FpInduction || Step->getType()->isIntegerTy()) &&
+         "StepValue is not an integer");
+
+  assert((IK != IK_FpInduction || Step->getType()->isFloatingPointTy()) &&
+         "StepValue is not FP for FpInduction");
+  assert((IK != IK_FpInduction || BinaryOp != 0) &&
+         "Binary opcode should be specified for FP induction");
 }
 
 int InductionDescriptor::getConsecutiveDirection() const {
   ConstantInt *ConstStep = getConstIntStepValue();
   if (ConstStep && (ConstStep->isOne() || ConstStep->isMinusOne()))
     return ConstStep->getSExtValue();
   return 0;
 }
 
 ConstantInt *InductionDescriptor::getConstIntStepValue() const {
   if (isa<SCEVConstant>(Step))
     return dyn_cast<ConstantInt>(cast<SCEVConstant>(Step)->getValue());
   return nullptr;
 }
 
 Value *InductionDescriptor::transform(IRBuilder<> &B, Value *Index,
                                       ScalarEvolution *SE,
                                       const DataLayout& DL) const {
 
   SCEVExpander Exp(*SE, DL, "induction");
+  assert(Index->getType() == Step->getType() &&
+         "Index type does not match StepValue type");
   switch (IK) {
   case IK_IntInduction: {
     assert(Index->getType() == StartValue->getType() &&
            "Index type does not match StartValue type");
 
     // FIXME: Theoretically, we can call getAddExpr() of ScalarEvolution
     // and calculate (Start + Index * Step) for all cases, without
     // special handling for "isOne" and "isMinusOne".
     // But in the real life the result code getting worse. We mix SCEV
     // expressions and ADD/SUB operations and receive redundant
     // intermediate values being calculated in different ways and
     // Instcombine is unable to reduce them all.
 
     if (getConstIntStepValue() &&
         getConstIntStepValue()->isMinusOne())
       return B.CreateSub(StartValue, Index);
     if (getConstIntStepValue() &&
         getConstIntStepValue()->isOne())
       return B.CreateAdd(StartValue, Index);
     const SCEV *S = SE->getAddExpr(SE->getSCEV(StartValue),
                                    SE->getMulExpr(Step, SE->getSCEV(Index)));
     return Exp.expandCodeFor(S, StartValue->getType(), &*B.GetInsertPoint());
   }
   case IK_PtrInduction: {
-    assert(Index->getType() == Step->getType() &&
-           "Index type does not match StepValue type");
     assert(isa<SCEVConstant>(Step) &&
            "Expected constant step for pointer induction");
     const SCEV *S = SE->getMulExpr(SE->getSCEV(Index), Step);
     Index = Exp.expandCodeFor(S, Index->getType(), &*B.GetInsertPoint());
     return B.CreateGEP(nullptr, StartValue, Index);
   }
+  case IK_FpInduction: {
+    assert(Step->getType()->isFloatingPointTy() && "Expected FP Step value");
+    assert(BinaryOp &&
+           "Original bin op should be defined for FP induction");
+
+    Value *StepValue = cast<SCEVUnknown>(Step)->getValue();
+
+    // Floating point operations had to be 'fast' to enable the induction.
+    FastMathFlags Flags;
+    Flags.setUnsafeAlgebra();
+
+    Value *MulExp = B.CreateFMul(StepValue, Index);
+    cast<Instruction>(MulExp)->setFastMathFlags(Flags);
+
+    Value *BOp = B.CreateBinOp(BinaryOp, StartValue,
+                               MulExp, "induction");
+    cast<Instruction>(BOp)->setFastMathFlags(Flags);
+
+    return BOp;
+  }
   case IK_NoInduction:
     return nullptr;
   }
   llvm_unreachable("invalid enum");
 }
 
-bool InductionDescriptor::isInductionPHI(PHINode *Phi,
+bool InductionDescriptor::isFpInductionPHI(PHINode *Phi, const Loop *TheLoop,
+                                           ScalarEvolution *SE,
+                                           InductionDescriptor &D) {
+
+  // Here we only handle FP induction variables.
+  assert(Phi->getType()->isFloatingPointTy() && "Unexpected Phi type");
+
+  if (TheLoop->getHeader() != Phi->getParent())
+    return nullptr;
+
+  // The loop may have multiple entrances or multiple exits; we can analyze
+  // this phi if it has a unique entry value and a unique backedge value.
+  if (Phi->getNumIncomingValues() != 2)
+    return false;
+  Value *BEValue = nullptr, *StartValue = nullptr;
+  if (TheLoop->contains(Phi->getIncomingBlock(0))) {
+    BEValue = Phi->getIncomingValue(0);
+    StartValue = Phi->getIncomingValue(1);
+  } else {
+    assert(TheLoop->contains(Phi->getIncomingBlock(1)) &&
+           "Unexpected Phi node in the loop"); 
+    BEValue = Phi->getIncomingValue(1);
+    StartValue = Phi->getIncomingValue(0);
+  }
+
+  if (!isa<BinaryOperator>(BEValue))
+    return false;
+
+  BinaryOperator *BOp = cast<BinaryOperator>(BEValue);
+  Value *Addend = nullptr;
+  if (BOp->getOpcode() == Instruction::FAdd) {
+    if (BOp->getOperand(0) == Phi)
+      Addend = BOp->getOperand(1);
+    else if (BOp->getOperand(1) == Phi)
+      Addend = BOp->getOperand(0);
+  } else if (BOp->getOpcode() == Instruction::FSub)
+    if (BOp->getOperand(0) == Phi)
+      Addend = BOp->getOperand(1);
+  
+  if (!Addend)
+    return false;
+
+  // The addend should be loop invariant
+  if (auto *I = dyn_cast<Instruction>(Addend))
+    if (TheLoop->contains(I))
+      return false;
+
+  const SCEV *Step = SE->getUnknown(Addend);
+
+  Instruction *UAI = !BOp->hasUnsafeAlgebra() ? BOp : nullptr;
+  D = InductionDescriptor(StartValue, IK_FpInduction, Step, UAI,
+                          BOp->getOpcode());
+  return true;
+}
+
+bool InductionDescriptor::isInductionPHI(PHINode *Phi, const Loop *TheLoop,
                                          PredicatedScalarEvolution &PSE,
                                          InductionDescriptor &D,
                                          bool Assume) {
   Type *PhiTy = Phi->getType();
-  // We only handle integer and pointer inductions variables.
-  if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
+
+  // Handle integer and pointer inductions variables.
+  // Now we handle also FP induction but not trying to make a
+  // recurrent expression from the PHI node in-place.
+
+  if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy() &&
+      !PhiTy->isFloatTy() && !PhiTy->isDoubleTy() && !PhiTy->isHalfTy())
     return false;
 
+  if (PhiTy->isFloatingPointTy())
+    return isFpInductionPHI(Phi, TheLoop, PSE.getSE(), D);
+
   const SCEV *PhiScev = PSE.getSCEV(Phi);
   const auto *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
 
   // We need this expression to be an AddRecExpr.
   if (Assume && !AR)
     AR = PSE.getAsAddRec(Phi);
 
   if (!AR) {
     DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
     return false;
   }
 
-  return isInductionPHI(Phi, PSE.getSE(), D, AR);
+  return isInductionPHI(Phi, TheLoop, PSE.getSE(), D, AR);
 }
 
-bool InductionDescriptor::isInductionPHI(PHINode *Phi,
+bool InductionDescriptor::isInductionPHI(PHINode *Phi, const Loop *TheLoop,
                                          ScalarEvolution *SE,
                                          InductionDescriptor &D,
                                          const SCEV *Expr) {
   Type *PhiTy = Phi->getType();
   // We only handle integer and pointer inductions variables.
   if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
     return false;
 
   // Check that the PHI is consecutive.
   const SCEV *PhiScev = Expr ? Expr : SE->getSCEV(Phi);
   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
 
   if (!AR) {
     DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
     return false;
   }
 
-  assert(AR->getLoop()->getHeader() == Phi->getParent() &&
+  assert(TheLoop->getHeader() == Phi->getParent() &&
          "PHI is an AddRec for a different loop?!");
   Value *StartValue =
     Phi->getIncomingValueForBlock(AR->getLoop()->getLoopPreheader());
   const SCEV *Step = AR->getStepRecurrence(*SE);
   // Calculate the pointer stride and check if it is consecutive.
   // The stride may be a constant or a loop invariant integer value.
   const SCEVConstant *ConstStep = dyn_cast<SCEVConstant>(Step);
-  if (!ConstStep && !SE->isLoopInvariant(Step, AR->getLoop()))
+  if (!ConstStep && !SE->isLoopInvariant(Step, TheLoop))
     return false;
 
   if (PhiTy->isIntegerTy()) {
     D = InductionDescriptor(StartValue, IK_IntInduction, Step);
     return true;
   }
 
   assert(PhiTy->isPointerTy() && "The PHI must be a pointer");

../lib/Transforms/Vectorize/LoopVectorize.cpp

protected:
   /// instruction (shuffle) for loop invariant values and for the induction
   /// value. If this is the induction variable then we extend it to N, N+1, ...
   /// this is needed because each iteration in the loop corresponds to a SIMD
   /// element.
   virtual Value *getBroadcastInstrs(Value *V);
 
   /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
   /// to each vector element of Val. The sequence starts at StartIndex.
-  virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
+  /// \p Opcode is relevant for FP induction variable.
+  virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step,
+                               Instruction::BinaryOps Opcode =
+                               (Instruction::BinaryOps)0);
 
   /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
   /// to each vector element of Val. The sequence starts at StartIndex.
   /// Step is a SCEV. In order to get StepValue it takes the existing value
   /// from SCEV or creates a new using SCEVExpander.
-  virtual Value *getStepVector(Value *Val, int StartIdx, const SCEV *Step);
+  /// \p Opcode is relevant for FP induction variable.
+  virtual Value *getStepVector(Value *Val, int StartIdx, const SCEV *Step,
+                               Instruction::BinaryOps Opcode =
+                               (Instruction::BinaryOps)0);
 
   /// Create a vector induction variable based on an existing scalar one.
   /// Currently only works for integer induction variables with a constant
   /// step.
   /// If TruncType is provided, instead of widening the original IV, we
   /// widen a version of the IV truncated to TruncType.
   void widenInductionVariable(const InductionDescriptor &II, VectorParts &Entry,
                               IntegerType *TruncType = nullptr);
InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
       : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, 1,
                             UnrollFactor) {}
 
 private:
   void scalarizeInstruction(Instruction *Instr,
                             bool IfPredicateStore = false) override;
   void vectorizeMemoryInstruction(Instruction *Instr) override;
   Value *getBroadcastInstrs(Value *V) override;
-  Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
-  Value *getStepVector(Value *Val, int StartIdx, const SCEV *StepSCEV) override;
+  Value *getStepVector(Value *Val, int StartIdx, Value *Step,
+                       Instruction::BinaryOps Opcode =
+                       (Instruction::BinaryOps)0) override;
+  Value *getStepVector(Value *Val, int StartIdx, const SCEV *StepSCEV,
+                       Instruction::BinaryOps Opcode =
+                       (Instruction::BinaryOps)0) override;
   Value *reverseVector(Value *Vec) override;
 };
 
 /// \brief Look for a meaningful debug location on the instruction or it's
 /// operands.
 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
   if (!I)
     return I;
Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
 
   // Broadcast the scalar into all locations in the vector.
   Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
 
   return Shuf;
 }
 
 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
-                                          const SCEV *StepSCEV) {
+                                          const SCEV *StepSCEV,
+                                          Instruction::BinaryOps BinOp) {
   const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
   SCEVExpander Exp(*PSE.getSE(), DL, "induction");
   Value *StepValue = Exp.expandCodeFor(StepSCEV, StepSCEV->getType(),
                                        &*Builder.GetInsertPoint());
-  return getStepVector(Val, StartIdx, StepValue);
+  return getStepVector(Val, StartIdx, StepValue, BinOp);
 }
 
 void InnerLoopVectorizer::widenInductionVariable(const InductionDescriptor &II,
                                                  VectorParts &Entry,
                                                  IntegerType *TruncType) {
   Value *Start = II.getStartValue();
   ConstantInt *Step = II.getConstIntStepValue();
   assert(Step && "Can not widen an IV with a non-constant step");
for (unsigned Part = 0; Part < UF; ++Part) {
     LastInduction = Builder.CreateAdd(LastInduction, SplatVF, "step.add");
   }
 
   VecInd->addIncoming(SteppedStart, LoopVectorPreHeader);
   VecInd->addIncoming(LastInduction, LoopVectorBody);
 }
 
 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
-                                          Value *Step) {
+                                          Value *Step,
+                                          Instruction::BinaryOps BinOp) {
   assert(Val->getType()->isVectorTy() && "Must be a vector");
-  assert(Val->getType()->getScalarType()->isIntegerTy() &&
-         "Elem must be an integer");
+  assert((Val->getType()->getScalarType()->isIntegerTy() ||
+    Val->getType()->getScalarType()->isFloatingPointTy()) &&
+    "Induction Step must be an integer or FP");
   assert(Step->getType() == Val->getType()->getScalarType() &&
          "Step has wrong type");
   // Create the types.
-  Type *ITy = Val->getType()->getScalarType();
+  Type *STy = Val->getType()->getScalarType();
   VectorType *Ty = cast<VectorType>(Val->getType());
   int VLen = Ty->getNumElements();
   SmallVector<Constant *, 8> Indices;
+  auto CurrIP = Builder.saveIP();
+  Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
 
+  if (Step->getType()->isIntegerTy()) {
     // Create a vector of consecutive numbers from zero to VF.
     for (int i = 0; i < VLen; ++i)
-    Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
+      Indices.push_back(ConstantInt::get(STy, StartIdx + i));
 
-  // Add the consecutive indices to the vector value.
     Constant *Cv = ConstantVector::get(Indices);
     assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
     Step = Builder.CreateVectorSplat(VLen, Step);
     assert(Step->getType() == Val->getType() && "Invalid step vec");
     // FIXME: The newly created binary instructions should contain nsw/nuw flags,
     // which can be found from the original scalar operations.
     Step = Builder.CreateMul(Cv, Step);
+    Builder.restoreIP(CurrIP);
     return Builder.CreateAdd(Val, Step, "induction");
   }
 
+  // Floating point induction.
+  assert(BinOp && "Binary Opcode should be specified for FP induction");
+  // Create a vector of consecutive numbers from zero to VF.
+  for (int i = 0; i < VLen; ++i)
+    Indices.push_back(ConstantFP::get(STy, (double)(StartIdx + i)));
+
+  // Add the consecutive indices to the vector value.
+  Constant *Cv = ConstantVector::get(Indices);
+
+  Step = Builder.CreateVectorSplat(VLen, Step);
+
+  // Floating point operations had to be 'fast' to enable the induction.
+  FastMathFlags Flags;
+  Flags.setUnsafeAlgebra();
+
+  Value *MulOp = Builder.CreateFMul(Cv, Step);
+  if (isa<Instruction>(MulOp))
+    cast<Instruction>(MulOp)->setFastMathFlags(Flags);
+  Builder.restoreIP(CurrIP);
+
+  Value *BOp = Builder.CreateBinOp(BinOp, Val, MulOp, "induction");
+  cast<Instruction>(BOp)->setFastMathFlags(Flags);
+  return BOp;
+}
+
 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
   assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
   auto *SE = PSE.getSE();
   // Make sure that the pointer does not point to structs.
   if (Ptr->getType()->getPointerElementType()->isAggregateType())
     return 0;
 
   // If this value is a pointer induction variable, we know it is consecutive.
for (I = List->begin(), E = List->end(); I != E; ++I) {
     PHINode *BCResumeVal = PHINode::Create(
         OrigPhi->getType(), 3, "bc.resume.val", ScalarPH->getTerminator());
     Value *EndValue;
     if (OrigPhi == OldInduction) {
       // We know what the end value is.
       EndValue = CountRoundDown;
     } else {
       IRBuilder<> B(LoopBypassBlocks.back()->getTerminator());
-      Value *CRD = B.CreateSExtOrTrunc(CountRoundDown,
-                                       II.getStep()->getType(), "cast.crd");
+      Value *CRD;
+      if (II.getStep()->getType()->isIntegerTy())
+        CRD = B.CreateSExtOrTrunc(CountRoundDown, II.getStep()->getType(),
+                                  "cast.crd");
+      else
+        CRD = B.CreateCast(Instruction::SIToFP, CountRoundDown,
       const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
       EndValue = II.transform(B, CRD, PSE.getSE(), DL);
       EndValue->setName("ind.end");
     }
 
     // The new PHI merges the original incoming value, in case of a bypass,
     // or the value at the end of the vectorized loop.
     BCResumeVal->addIncoming(EndValue, MiddleBlock);
if (VF == 1 || P->getType() != Induction->getType() ||
         Entry[part] = getStepVector(Broadcasted, VF * part, II.getStep());
     } else {
       // Instead of re-creating the vector IV by splatting the scalar IV
       // in each iteration, we can make a new independent vector IV.
       widenInductionVariable(II, Entry);
     }
     return;
   }
-  case InductionDescriptor::IK_PtrInduction:
+  case InductionDescriptor::IK_PtrInduction: {
     // Handle the pointer induction variable case.
     assert(P->getType()->isPointerTy() && "Unexpected type.");
     // This is the normalized GEP that starts counting at zero.
     Value *PtrInd = Induction;
     PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStep()->getType());
     // This is the vector of results. Notice that we don't generate
     // vector geps because scalar geps result in better code.
     for (unsigned part = 0; part < UF; ++part) {
for (unsigned part = 0; part < UF; ++part) {
         SclrGep->setName("next.gep");
         VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
                                              Builder.getInt32(i), "insert.gep");
       }
       Entry[part] = VecVal;
     }
     return;
   }
+  case InductionDescriptor::IK_FpInduction: {
+    assert(P->getType() == II.getStartValue()->getType() &&
+           "Types must match");
+    // Handle other induction variables that are now based on the
+    // canonical one.
+    assert(P != OldInduction && "Main induction can be integer only");
+
+    Value *V = Builder.CreateCast(Instruction::SIToFP, Induction, P->getType());
+    V = II.transform(Builder, V, PSE.getSE(), DL);
+    V->setName("fp.offset.idx");
+
+    Value *Broadcasted = getBroadcastInstrs(V);
+    // After broadcasting the induction variable we need to make the vector
+    // consecutive by adding 0, 1, 2, etc.
+    for (unsigned part = 0; part < UF; ++part)
+      Entry[part] = getStepVector(Broadcasted, VF * part, II.getStep(),
+                                  II.getBinaryOpcode());
+    return;
+  }
+  }
 }
 
 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
   // For each instruction in the old loop.
   for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
     VectorParts &Entry = WidenMap.get(&*it);
 
     switch (it->getOpcode()) {

 void LoopVectorizationLegality::addInductionPhi(
     PHINode *Phi, const InductionDescriptor &ID,
     SmallPtrSetImpl<Value *> &AllowedExit) {
   Inductions[Phi] = ID;
   Type *PhiTy = Phi->getType();
   const DataLayout &DL = Phi->getModule()->getDataLayout();
 
   // Get the widest type.
+  if (!PhiTy->isFloatingPointTy()) {
     if (!WidestIndTy)
       WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
     else
       WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
+  }
 
   // Int inductions are special because we only allow one IV.
   if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
       ID.getConstIntStepValue() &&
       ID.getConstIntStepValue()->isOne() &&
       isa<Constant>(ID.getStartValue()) &&
       cast<Constant>(ID.getStartValue())->isNullValue()) {
 
for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
           if (RedDes.hasUnsafeAlgebra())
             Requirements->addUnsafeAlgebraInst(RedDes.getUnsafeAlgebraInst());
           AllowedExit.insert(RedDes.getLoopExitInstr());
           Reductions[Phi] = RedDes;
           continue;
         }
 
         InductionDescriptor ID;
-        if (InductionDescriptor::isInductionPHI(Phi, PSE, ID)) {
+        if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID)) {
           addInductionPhi(Phi, ID, AllowedExit);
+          if (ID.hasUnsafeAlgebra())
+            Requirements->addUnsafeAlgebraInst(ID.getUnsafeAlgebraInst());
           continue;
         }
 
         if (RecurrenceDescriptor::isFirstOrderRecurrence(Phi, TheLoop, DT)) {
           FirstOrderRecurrences.insert(Phi);
           continue;
         }
 
         // As a last resort, coerce the PHI to a AddRec expression
         // and re-try classifying it a an induction PHI.
-        if (InductionDescriptor::isInductionPHI(Phi, PSE, ID, true)) {
+        if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID, true)) {
           addInductionPhi(Phi, ID, AllowedExit);        
           continue;
         }
 
         emitAnalysis(VectorizationReport(&*it)
                      << "value that could not be identified as "
                         "reduction is used outside the loop");
         DEBUG(dbgs() << "LV: Found an unidentified PHI." << *Phi << "\n");
void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
   return scalarizeInstruction(Instr, IfPredicateStore);
 }
 
 Value *InnerLoopUnroller::reverseVector(Value *Vec) { return Vec; }
 
 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { return V; }
 
 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx,
-                                        const SCEV *StepSCEV) {
+                                        const SCEV *StepSCEV,
+                                        Instruction::BinaryOps BinOp) {
   const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
   SCEVExpander Exp(*PSE.getSE(), DL, "induction");
   Value *StepValue = Exp.expandCodeFor(StepSCEV, StepSCEV->getType(),
                                        &*Builder.GetInsertPoint());
   return getStepVector(Val, StartIdx, StepValue);
 }
 
-Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
+Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step,
+                                        Instruction::BinaryOps BinOp) {
   // When unrolling and the VF is 1, we only need to add a simple scalar.
-  Type *ITy = Val->getType();
-  assert(!ITy->isVectorTy() && "Val must be a scalar");
-  Constant *C = ConstantInt::get(ITy, StartIdx);
+  Type *Ty = Val->getType();
+  assert(!Ty->isVectorTy() && "Val must be a scalar");
+
+  if (Ty->isFloatingPointTy()) {
+    Constant *C = ConstantFP::get(Ty, (double)StartIdx);
+
+    // Floating point operations had to be 'fast' to enable the unrolling.
+    FastMathFlags Flags;
+    Flags.setUnsafeAlgebra();
+
+    Value *MulOp = Builder.CreateFMul(C, Step);
+    if (isa<Instruction>(MulOp))
+      cast<Instruction>(MulOp)->setFastMathFlags(Flags);
+    Value *BOp = Builder.CreateBinOp(BinOp, Val, MulOp);
+    cast<Instruction>(BOp)->setFastMathFlags(Flags);
+    return BOp;
+  }
+  Constant *C = ConstantInt::get(Ty, StartIdx);
   return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");
 }

../test/Transforms/LoopVectorize/float-induction.ll

+; RUN: opt < %s  -loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -dce -instcombine -S | FileCheck %s
+
+
+; CHECK-LABEL: @fp_iv_loop1(
+; CHECK: vector.body:
+; CHECK:  %[[INDEX:.*]] = sitofp i64 {{.*}} to float
+; CHECK:  %[[VEC_INCR:.*]] = fmul fast float %{{.*}}, %[[INDEX]]
+; CHECK:  fsub fast float %init, %[[VEC_INCR]]
+; CHECK:  store <4 x float> 
+
+@fp_inc = common global float 0.000000e+00, align 4
+
+;void fp_iv_loop1(float init, float * __restrict__ A, int N) {
+;  float x = init;
+;  for (int i=0; i < N; ++i) {
+;    A[i] = x;
+;    x -= fp_inc;
+;  }
+;}
+
+define void @fp_iv_loop1(float %init, float* noalias nocapture %A, i32 %N) #0 {
+entry:
+  %cmp4 = icmp sgt i32 %N, 0
+  br i1 %cmp4, label %for.body.lr.ph, label %for.end
+
+for.body.lr.ph:                                   ; preds = %entry
+  %0 = load float, float* @fp_inc, align 4
+  br label %for.body
+
+for.body:                                         ; preds = %for.body, %for.body.lr.ph
+  %indvars.iv = phi i64 [ 0, %for.body.lr.ph ], [ %indvars.iv.next, %for.body ]
+  %x.05 = phi float [ %init, %for.body.lr.ph ], [ %add, %for.body ]
+  %arrayidx = getelementptr inbounds float, float* %A, i64 %indvars.iv
+  store float %x.05, float* %arrayidx, align 4
+  %add = fsub fast float %x.05, %0
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, %N
+  br i1 %exitcond, label %for.end.loopexit, label %for.body
+
+for.end.loopexit:                                 ; preds = %for.body
+  br label %for.end
+
+for.end:                                          ; preds = %for.end.loopexit, %entry
+  ret void
+}
+
+;void fp_iv_loop2(float init, float * __restrict__ A, int N) {
+;  float x = init;
+;  for (int i=0; i < N; ++i) {
+;    A[i] = x;
+;    x += 0.5;
+;  }
+;}
+
+; CHECK-LABEL: @fp_iv_loop2(
+; CHECK: vector.body
+; CHECK:  %[[index:.*]] = phi i64 [ 0, %vector.ph ]
+; CHECK: sitofp i64 %[[index]] to float
+; CHECK: %[[VAR1:.*]] = fmul fast float {{.*}}, 5.000000e-01
+; CHECK: %[[VAR2:.*]] = fadd fast float %[[VAR1]]
+; CHECK:  insertelement <4 x float> undef, float %[[VAR2]], i32 0
+; CHECK:  shufflevector <4 x float> {{.*}}, <4 x float> undef, <4 x i32> zeroinitializer
+; CHECK:  fadd fast <4 x float> {{.*}}, <float 0.000000e+00, float 5.000000e-01, float 1.000000e+00, float 1.500000e+00>
+; CHECK:  store <4 x float> 
+
+define void @fp_iv_loop2(float %init, float* noalias nocapture %A, i32 %N) #0 {
+entry:
+  %cmp4 = icmp sgt i32 %N, 0
+  br i1 %cmp4, label %for.body.preheader, label %for.end
+
+for.body.preheader:                               ; preds = %entry
+  br label %for.body
+
+for.body:                                         ; preds = %for.body.preheader, %for.body
+  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %for.body.preheader ]
+  %x.06 = phi float [ %conv1, %for.body ], [ %init, %for.body.preheader ]
+  %arrayidx = getelementptr inbounds float, float* %A, i64 %indvars.iv
+  store float %x.06, float* %arrayidx, align 4
+  %conv1 = fadd fast float %x.06, 5.000000e-01
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, %N
+  br i1 %exitcond, label %for.end.loopexit, label %for.body
+
+for.end.loopexit:                                 ; preds = %for.body
+  br label %for.end
+
+for.end:                                          ; preds = %for.end.loopexit, %entry
+  ret void
+}
+
+;void fp_iv_loop3(float init, float * __restrict__ A, float * __restrict__ B, float * __restrict__ C, int N) {
+;  int i = 0;
+;  float x = init;
+;  float y = 0.1;
+;  for (; i < N; ++i) {
+;    A[i] = x;
+;    x += fp_inc;
+;    y -= 0.5;
+;    B[i] = x + y;
+;    C[i] = y;
+;  }
+;}
+; CHECK-LABEL: @fp_iv_loop3(
+; CHECK: vector.body
+; CHECK:  %[[index:.*]] = phi i64 [ 0, %vector.ph ]
+; CHECK: sitofp i64 %[[index]] to float
+; CHECK: %[[VAR1:.*]] = fmul fast float {{.*}}, -5.000000e-01
+; CHECK:  fadd fast float %[[VAR1]]
+; CHECK:  fadd fast <4 x float> {{.*}}, <float -5.000000e-01, float -1.000000e+00, float -1.500000e+00, float -2.000000e+00>
+; CHECK:  store <4 x float>
+
+define void @fp_iv_loop3(float %init, float* noalias nocapture %A, float* noalias nocapture %B, float* noalias nocapture %C, i32 %N) #0 {
+entry:
+  %cmp9 = icmp sgt i32 %N, 0
+  br i1 %cmp9, label %for.body.lr.ph, label %for.end
+
+for.body.lr.ph:                                   ; preds = %entry
+  %0 = load float, float* @fp_inc, align 4
+  br label %for.body
+
+for.body:                                         ; preds = %for.body, %for.body.lr.ph
+  %indvars.iv = phi i64 [ 0, %for.body.lr.ph ], [ %indvars.iv.next, %for.body ]
+  %y.012 = phi float [ 0x3FB99999A0000000, %for.body.lr.ph ], [ %conv1, %for.body ]
+  %x.011 = phi float [ %init, %for.body.lr.ph ], [ %add, %for.body ]
+  %arrayidx = getelementptr inbounds float, float* %A, i64 %indvars.iv
+  store float %x.011, float* %arrayidx, align 4
+  %add = fadd fast float %x.011, %0
+  %conv1 = fadd fast float %y.012, -5.000000e-01
+  %add2 = fadd fast float %conv1, %add
+  %arrayidx4 = getelementptr inbounds float, float* %B, i64 %indvars.iv
+  store float %add2, float* %arrayidx4, align 4
+  %arrayidx6 = getelementptr inbounds float, float* %C, i64 %indvars.iv
+  store float %conv1, float* %arrayidx6, align 4
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, %N
+  br i1 %exitcond, label %for.end.loopexit, label %for.body
+
+for.end.loopexit:                                 
+  br label %for.end
+
+for.end: 
+  ret void
+}

../test/Transforms/LoopVectorize/induction-step.ll

 ; RUN: opt < %s -loop-vectorize -force-vector-interleave=1 -force-vector-width=8 -S | FileCheck %s
 
 ; int int_inc;
 ;
 ;int induction_with_global(int init, int *restrict A, int N) {
 ;  int x = init;
 ;  for (int i=0;i<N;i++){
 ;    A[i] = x;
 ;    x += int_inc;
 ;  }
 ;  return x;
 ;}
 
 ; CHECK-LABEL: @induction_with_global(
 ; CHECK: %[[INT_INC:.*]] = load i32, i32* @int_inc, align 4
-; CHECK: vector.body:
+; CHECK: vector.ph:
 ; CHECK:  %[[VAR1:.*]] = insertelement <8 x i32> undef, i32 %[[INT_INC]], i32 0
 ; CHECK:  %[[VAR2:.*]] = shufflevector <8 x i32> %[[VAR1]], <8 x i32> undef, <8 x i32> zeroinitializer
 ; CHECK:  mul <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>, %[[VAR2]]
 
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 
 
 @int_inc = common global i32 0, align 4

 ;  }
 ;  return x;
 ;}
 
 ; CHECK-LABEL: @induction_with_loop_inv(
 ; CHECK: for.cond1.preheader:                            
 ; CHECK: %[[INDVAR0:.*]] = phi i32 [ 0,
 ; CHECK: %[[INDVAR1:.*]] = phi i32 [ 0,
-; CHECK: vector.body:
+; CHECK: vector.ph:
 ; CHECK:  %[[VAR1:.*]] = insertelement <8 x i32> undef, i32 %[[INDVAR1]], i32 0
 ; CHECK:  %[[VAR2:.*]] = shufflevector <8 x i32> %[[VAR1]], <8 x i32> undef, <8 x i32> zeroinitializer
 ; CHECK:  mul <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>, %[[VAR2]]
 
 define i32 @induction_with_loop_inv(i32 %init, i32* noalias nocapture %A, i32 %N, i32 %M) {
 entry:
   %cmp10 = icmp sgt i32 %M, 0
   br i1 %cmp10, label %for.cond1.preheader.lr.ph, label %for.end6