[LoopVectorizer] Inloop vector reductions

[dmgreen]

Summary

Arm MVE has multiple instructions such as VMLAVA.s8, which (in this case) can take two 128bit vectors, sign extend the inputs to i32, multiplying them together and sum the result into a 32bit general purpose register. So taking 16 i8's as inputs, they can multiply and accumulate the result into a single i32 without any rounding/truncating along the way. There are also reduction instructions for plain integer add and min/max, and operations that sum into a pair of 32bit registers together treated as a 64bit integer (even though MVE does not have a plain 64bit addition instruction). So giving the vectorizer the ability to use these instructions both enables us to vectorize at higher bitwidths, and to vectorize things we previously could not.

In order to do that we need a way to represent that the "reduction" operation, specified with a llvm.experimental.vector.reduce when vectorizing for Arm, occurs inside the loop not after it like most reductions. This patch attempts to do that, teaching the vectorizer about in-loop reductions. It does this through a vplan recipe representing the reductions that the original chain of reduction operations is replaced by. Cost modelling is currently just done through a "prefersInloopReduction" TTI hook (which follows in a later patch).

[fhahn]

This patch does not work like that yet though, as I am unsure how that should really look, and it did not seem simple to create a vplan from another yet. The def-use relations are usually not in place for example. Neither is the costing of vplans against one another, which makes sense.

That's correct. Work on modelling the def-use relationships exclusively in VPlan is currently underway, but it probably does not make sense to wait with this patch until the transition is complete.

Some small high-level comments:

• I think it's best to move the TTI hook into a separate change
• Do you know how 'experimental' the reduction intrinsics still are? It would be good to move them out of experimental at some point, especially when making use of them in LV.

I'm currently traveling, but I try to take a closer look next week or the week after (please ping in case I forget ;))

[dmgreen]

Thanks for taking a look!

I noticed a comment in VPValue that says:

DESIGN PRINCIPLE: Access to the underlying IR must be strictly limited to
the front-end and back-end of VPlan so that the middle-end is as
independent as possible of the underlying IR. We grant access to the
underlying IR using friendship. In that way, we should be able to use VPlan
for multiple underlying IRs (Polly?) by providing a new VPlan front-end,
back-end and analysis information for the new IR.

That's the part that confused me about optimising vplans in-place. Does that mean that we have to replicate all the opcode/flags data into vplan too? Not just the use-defs. So it knows what is an Add and what is a Mul (and if they are nsw etc). What would you say counts as "middle-end" for vplan? I would presume the costmodelling and creating vplans from one another, but I didn't follow vplans early days very closely.

I think it's best to move the TTI hook into a separate change

Can do. I have some other changes that I was trying out too about adding a VPReductionPhi. Not sure it makes sense yet though.

Do you know how 'experimental' the reduction intrinsics still are? It would be good to move them out of experimental at some point, especially when making use of them in LV.

As far as I understand, on architectures that request it (AArch64 and Arm MVE now), we have been using the "experimental" reduction intrinsics since they were added, in the code generated after the vector body. There is a useReductionIntrinsic which the targets can override to prefer them over the IR equivalents.

I would like to add a mask to the reduction intrinsics. That will be important for getting tail-predicated inloop reductions to work. Whether this is done by adjusting the existing ones or adding new ones, I'm not sure. Obviously some architectures do not handle the mask, only handling a all-one mask. After that is sorted out then I agree it would be nice to drop the "experimental".

[dmgreen]

Taken out the TTI change and simplified the tests, just shoing the diffs here now.

[dmgreen]

Rebase and a bit of a cleanup. Now uses VPValues for the VPRecipe operands.

[dmgreen]

Ping. Anyone have any thought/comments about this one? Does it at least look like its going in the right direction, or does anyone have some better suggestions?

[Ayal]

In D75069#2041073, @dmgreen wrote:

Ping. Anyone have any thought/comments about this one? Does it at least look like its going in the right direction, or does anyone have some better suggestions?

Yes, this direction does look right!
Would be good to clarify that the unrolled Parts are still reduced out of loop, and document the current limitations / future TODO extensions, including min/max, conditional bumps, truncate/extend.
Added various minor comments.

llvm/lib/Analysis/IVDescriptors.cpp

812

Looking for a chain of hasOneUse op's would be easier starting from the Phi and going downwards, until reaching LoopExitInstr?

Note that when extended to handle reductions with conditional bumps, some ops will have more than one use.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

269

"Force" >> "Prefer"? If a reduction cannot be handled in loop, it is vectorized out of loop, rather than bailing out.

1045

collectInLoopReductions()? Perhaps worth holding a map of in loop reduction phi's with their chains.

1315

isInLoopReduction(Phi)?

1320

Simply call InLoopReductions.empty() directly?

1323

hasOutOfLoopReductions()? Outerloop stands for a non innermost loop in a nest.

3769

isInLoopReductionPhi

3787

This is dead code if cmp/select chains are not recognized yet, as noted above.

3865

Checking if !UseInLoopReductions is redundant, as current in loop reductions have phi's of the same type as the recurrence; right?
May leave it as an assert, until in loop reductions handle truncation/extension.

3916–3918

Worth adding a comment that in loop reductions create their target reductions inside the loop using a recipe, rather that here in the middle block.

6559

Move this invariant check out as another early-exit?

7314

Iterate over in loop reductions?

7449

In loop reductions are currently disabled in fold tail. Checking also if hasOuterloopReductions() is an independent optimization that should be done separately of this patch.

7451

Iterate over in loop reductions?
Comment that they are ordered top-down, starting from the phi.

7454

This is currently dead code, no in loop reductions under fold tail.

7671

NewChain, NewAcc >> PrevInChain, NextInChain?

llvm/lib/Transforms/Vectorize/VPlan.h

240

Too bad this requires passing TTI through the State everywhere.
Perhaps storing TTI in the recipe would be somewhat better.

1030

Comment is redundant.

1032

private is redundant.

llvm/lib/Transforms/Vectorize/VPlanValue.h

47 ↗

(On Diff #263027)

Why is this friendship needed?

llvm/test/Transforms/LoopVectorize/icmp-uniforms.ll

43 ↗

(On Diff #263027)

Optimizing this comparison away is independent of this patch.

llvm/test/Transforms/LoopVectorize/reduction-inloop.ll

2

It is indeed good to examine the diff in this review, but existing non -force-inloop-reductions RUN should be retained; e.g., by cloning the file for a separate -force-inloop-reductions RUN.

Would be good to also test with -force-vector-interleave greater than 1.

[dmgreen]

Thanks for the review! You rock. I'll update this as per the comments.

[fhahn]

llvm/lib/Analysis/IVDescriptors.cpp

799

nit: would be good to document what this function does

802

nit: would be good to add a message

llvm/lib/Transforms/Vectorize/VPlan.cpp

787

The ::print methods do not need to emit the " +\n" at the start and the "\\l\" parts after 4c8285c750bc

llvm/lib/Transforms/Vectorize/VPlan.h

43

Would a forward decl of ReductionDesciptor be sufficient or is there another need to include LoopUtils.h?

1034

would be good to document the members.

llvm/test/Transforms/LoopVectorize/reduction-inloop.ll

6–52

I think it would be also good to add a few additional simpler test cases, e.g. reductions with only 1 or 2 reduction steps in the loop, some with constant scalar operands. And with some of the (unnecessary trounces stripped). Also the test diff can be slightly reduced by having constant trip counts.

It's good to have the more complex ones, but it would also be good to start with simpler cases. That way it is probably easier to follow what's going on. And it would reduce the diff quite a bit if the tests for the various different opcodes would be a bit more compact.

[fhahn]

llvm/lib/Analysis/IVDescriptors.cpp

812

Instead of doing a recursive traversal, would it be simpler to just do the traversal iteratively, at least as long as we are only using at a single use chain?

[dmgreen]

Round one. Lots of renaming, plenty of cleanup. I've tried to add Min/Max handling too, and added/cleaned up some more test. Hopefully addressed most of the comments.

[dmgreen]

llvm/lib/Analysis/IVDescriptors.cpp

812

Yeah, that direction makes it a lot simpler. Thanks.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

3787

I've added the code to handle minmax too (but not tested it a lot yet. I will try that now).

MVE has instructions for integer min/max reductions, but they can be slow enough to make them not worth using over a normal vmin/vmax. Adds are always not-slower-enough to warrant the inloop reduction (and have other advantages like handling higher type sizes and folding in more instructions.)

My point is that min/max, like some of the other fadd/mul/and/etc might not be used by MVE yet. If you think the code is more hassle than it deserves, then we could take them out for the time being. I'd like to leave them in for consistency though, even if it's not used straight away.

3865

Right. We look from the phi down (now), so can't go through the 'and' we would need to detect the different bitwidth. That would leave 'and' reduction, which I think would not got through that codepath to create type promoted phis?

I've put an explicit check in for it to be sure, and added some testing for those cases.

6559

This does look a little strange here on it's own. The followup patch to add the TTI hook makes it look like:

if (!PreferInloopReductions &&
      !TTI.preferInloopReduction(Opcode, Phi->getType(),
                                 TargetTransformInfo::ReductionFlags()))
    continue;

7314

Do you mean adding an iterator for iterating over reductions, stepping over the ones not inloop?

It would seem like it's similar to the existing code, but as a new iterator class. My gut says the current code is simpler and clearer what is going on?

llvm/lib/Transforms/Vectorize/VPlan.h

240

I've changed it to be stored there. It does mean multiple things are holding TTI. Let me know what you think.

llvm/lib/Transforms/Vectorize/VPlanValue.h

47 ↗

(On Diff #263027)

Ah. Old code no longer needed.

llvm/test/Transforms/LoopVectorize/reduction-inloop.ll

2

This testcase is new for this patch (although that wasn't obvious from how I presented it). I was just trying to show the diffs here and the option didn't exist before this patch, hence it wasn't already there.

Are you suggesting we should still have these same test for without -force-inloop-reductions? I think those should already be tested elsewhere, but I can add it if you think it's useful.

6–52

(unnecessary trounces stripped)?

The const case is going to look a little odd until there is a instcombine type fold for constant reductions. I'll see about adding that.

Otherwise it is now hopefully a bit leaner. Even if I have added a few extra tests.

[dmgreen]

ping

[Ayal]

The proposed RecurrenceDescriptor::getReductionOpChain() method identifies horizontal reductions, a subset of the vertical reductions identified by RecurrenceDescriptor::isReductionPHI(). Being two totally independent methods, it's unclear what the latter supports that the former does not. Would it be better to have isReductionPHI() also take care of recognizing horizontal reductions, and record them as in CastsToIgnore? See also TODO commented inline.

llvm/lib/Analysis/IVDescriptors.cpp

439

Perhaps the above "better way" would also help recognize and record horizontal reductions?

848

|| !Cur->hasNUses(ExpectedUses) ?

nit: can alternatively let getNextInstruction check its result and return only valid ones, e.g.:

bool RedOpIsCmp = (RedOp == Instruction::ICmp || RedOp == Instruction::FCmp);
unsigned ExpectedUses = RedOpIsCmp ? 2 : 1;

auto getNextInstruction = [&](Instruction *Cur) {
  if (!Cur->hasNUses(ExpectedUses))
    return nullptr;
  auto *FirstUser = Cur->user_begin();
  if (!RedOpIsCmp)
    return FirstUser->getOpcode() == RedOp ? FirstUser : nullptr;
  // Handle cmp/select pair:
  auto *Sel = dyn_cast<SelectInst>(*FirstUser) ||
    dyn_cast<SelectInst>(*std::next(FirstUser));
  if (SelectPatternResult::isMinOrMax(matchSelectPattern(Sel, LHS, RHS).Flavor))
    return Sel;
  return nullptr;
}

for (auto *Cur = getNextInstruction(Phi); Cur && Cur != LoopExitInstr; Cur = getNextInstruction(Cur))
  ReductionOperations.push_back(Cur);

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

1463

... along with their associated chains of reduction operations,
in program order from top (PHI) to bottom?

3787

Would be good to make sure code is being exercised and tested. Could inloop min/max (and/or other reductions) help reduce code size, and be applied when vectorizing under optsize?

3865

'and' reduction may or may not be performed in a smaller type, iinm.

4215

Check in-loop reduction along with VF == 1, instead of modifying VF, as in other places? E.g.,

bool ScalarPHI = (VF == 1) || Cost->isInLoopReduction(cast<PHINode>(PN));

Type *VecTy =
    ScalarPHI ? PN->getType() : VectorType::get(PN->getType(), VF);

?

6577

The "else" clause will be empty in non-debug mode.
Can hoist the LLVM_DEBUG's, e.g.,

bool InLoop = !ReductionOperations.empty();
LLVM_DEBUG(dbgs() << "LV: Using " << (InLoop ? "inloop" : "out of loop") << " reduction for phi: " << *Phi << "\n");

7314

Suggestion was to iterate over the PHIs/elements of InloopReductionChains, rather than over all reduction PHIs of Legal->getReductionVars().

(Better early-exit via "if (!CM.isInLoopReduction(Reduction.first)) continue;")

7323

Add a comment that this is the compare operand of the select, whose recipe will also be discarded.

7447

Consider outlining this part, unless it can be shortened.

7449

Checking also if hasOuterloopReductions() is an independent optimization that should be done separately of this patch.

7451

Suggestion was to iterate over the PHIs/elements of InloopReductionChains, rather than over all reduction PHIs of Legal->getReductionVars().

7478

ChainOp can be simply set:

VPValue *ChainOp = Plan->getVPValue(Chain);

leaving only VecOp to figure out which operand index to use. Can preset the options before the loop:

RecurrenceDescriptor::RecurrenceKind Kind = RdxDesc.getRecurrenceKind();
bool InMinMax = (Kind == RecurrenceDescriptor::RK_IntegerMinMax || Kind == RecurrenceDescriptor::RK_FloatMinMax);
FirstOpId = IsMinMax ? 1 : 0;
SecondOpId = FirstOpId + 1;

and then do

auto *FirstOp = R->getOperand(FirstOpId);
auto *SecondOp =  R->getOperand(SecondOpId);
VPValue *VecOp = Plan->getVPValue(FirstOp == Chain ? SecondOp : FirstOp);

llvm/test/Transforms/LoopVectorize/reduction-inloop.ll

2

ok, understood; no need to retain non-existing tests.

[dmgreen]

llvm/lib/Analysis/IVDescriptors.cpp

439

Hmm. The reason I kept them separate was that this method is already pretty complex. I was trying to keep thing simpler. Adding the ability to detect a single chain of operations from Phi to LoopExitValue that can be used for horizontal reductions looks.. difficult. And error prone. :) If you think it's worth it then I can certainly give it a go! I like the separation of concerns in keeping them separate though.

The extra things that AddReductionVar will detect that getReductionOpChain will not are:

• Phi/select predicated reductions like in if-conversion-reductions.ll and if-reduction.ll. These would need some form of predicated reduction intrinsic.
• Narrow bitwidths. This one I could add.
• Subs/FSubs are treated like Adds/FAdds for vertical reductions.

848

This is the loop exit instr, so can have as many uses as it likes I believe.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

3787

-Os sounds like a good plan. It will take some backend work to make it efficient enough first though. And predicated reductions?

3865

The check for lookThroughAnd in AddReductionVar is guarded by isArithmeticRecurrenceKind, so at least currently it cannot be an And reduction.

7314

I believe that InloopReductionChains would not iterate in a deterministic order, which is why I avoided it.

Perhaps that would not matter here? The reductions should be independent anyway. Seems safer to try and use deterministic ordering anyway if we can.

7449

Ah sorry. I split it off but apparently didn't upstream the other part yet. Now in D81415.

[dmgreen]

Updates

[dmgreen]

ping

[dmgreen]

ping :)

[Ayal]

llvm/lib/Analysis/IVDescriptors.cpp

439

Would be good to see if above TODO can be addressed - providing the set of all instructions that take part in the reduction. This set could then be used for checking in-loop reductions. Hopefully this could help simplify both, and keep them in some sync. But could be done later, possibly with another TODO..

812

Is treating sub as an add reduction something in-loop reduction could support as a future extension?

848

Ahh, ok. (It should have ExpectedUses+1 users being in lcssa.)

"instruciton"

856

Is Subs the only issue?.
Can check this earlier, before traversing the chain, although it is pushed back last, here.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

1045

Thanks. Worth updating the comment.

3787

Hoisting the horizontal reduction from the middle block into the loop could potentially eliminate the middle block (as in tests below), so could presumably lead to code of smaller size? At-least for in-loop chains of a single link.

And predicated reductions?

These are yet to be handled in-loop, right?

6559

Then better placed above right after defining Phi?

7314

Agreed it would be better to use deterministic ordering. How about letting InloopReductionChains be a MapVector and iterate over
for (auto &Reduction : CM.getInloopReductions())?
The number of reductions is expected to be small, w/o removals.

7516

This is the other potential use of for (auto &Reduction : CM.getInloopReductions()).

[dmgreen]

Thanks for taking another look.

llvm/lib/Analysis/IVDescriptors.cpp

812

Hmm. I don't want to say never. A normal inloop reduction looks like:

p = PHI(0, a)
l = VLDR (..)
a = VADDVA(p, l)

Where the VADDV is an across-vector reductions, and the extra A means also add p. Reducing a sub would need to become:

p = PHI(0, a)
l = VLDR (..)
a = VADDV(l)
p = SUB(p, a)

With the SUB as a separate scalar instruction, which would be quite slow on some hardware (getting a value over from the VADDV to the SUB). So this would almost certainly be slower than a out-of-loop reduction.

But if we could end up using a higher vector factor for the reduction, or end up vectorizing loops that would previously not be vectorized.. that may lead to a gain overall to overcome the extra cost of adding the sub to the loop. It will require some very careful costing I think. And maybe the ability to create multiple vplans and cost them against one another :)

856

I believe sub is the only issue, unless I am forgetting something.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

3787

And predicated reductions?

These are yet to be handled in-loop, right?

Yep. It will need a predicated reduction intrinsic. A vecreduce that takes a mask. That will allow us to tail-fold the reductions with trip counts that do not divide the vector factor, which will make them look a lot better under -Os. And nice in general I think once it all starts being tail predicated.

The backend work I was mentioning was that we need to more efficiently transform

x = min(vecreduce.min(z), y)

into

x = VMINV(y, z)

Where y is (confusingly) accumulated in the case (even though the instruction doesn't have an A suffix). We currently generate

x = min(VMINV(UINT_MAX, z), y)

Once that is sorted out then, yep, using these for Os sounds like a good plan.

6559

I can put it somewhere sensible in this patch and move it in the next :)

7314

MapVector sounds good. I've changed it to use that and tried to use that in a few more places. Let me know what you think.

[Ayal]

llvm/lib/Analysis/IVDescriptors.cpp

812

An original sub code, say, acc -= a[i], can be treated as acc += (-a[i]). This could be in-loop reduced by first negating a[i]'s, at LV's LLVM-IR level, presumably lowered later to something like

p = PHI(0, a)
l = VLDR (..)
s = VSUBV (zero, l)
a = VADDVA(p, s)

, right?

842

rephrase sentence so it parses

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

3769

Ahh, this should actually be capitalized IsInLoopReductionPhi

3787

Re: predicated reductions - could they be handled by replacing masked-off elements with Identity using a select prior to reduction? To be potentially folded later by suitable targets into a predicated reduction operation which they may support.
Somewhat akin to "passthru" values of masked loads.

7314

Uses as MapVector look good to me, thanks. Can also retain isInLoopReduction(PHI).

7320

recode >> record

[dmgreen]

Reinstate isInLoopReduction and fixup some wording/typos/capitalisation.

llvm/lib/Analysis/IVDescriptors.cpp

812

Yep. We would have the option to trading a scalar instruction for a vector instruction + an extra register (to hold the 0, we only have 8 registers!)

Unfortunately both would be slower than in out-of-loop reduction unless we were vectorizing at a higher factor, though.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

3787

Oh. So select

s = select m, a, 0
v = vecreduce.add s

to a predicated vaddv?

Yeah, sounds interesting. I'll look into that as an alternative to predicated intrinsics. Nice suggestion.

[Ayal]

llvm/lib/Analysis/IVDescriptors.cpp

812

ok, so sub's can be handled in-loop, but doing so is expected to be more costly than out-of-loop, at-least if a horizontal add operation is to be used rather than a horizontal subtract; probably worth a comment.
If a reduction chain has only sub's, they could all sink - negating the sum once after the loop, using VADDVA inside. Doing so however will retain the middle block, i.e., w/o decreasing code size.

[dmgreen]

Update a FIXME with sub info.

[gilr]

llvm/lib/Transforms/Vectorize/VPlan.h

240

It seems that TTI is only used later for deciding whether to use a shuffle sequence or an intrinsic based on data available during planning. If so, then it would be best if the Planner calls TTI->useReductionIntrinsic() and records that boolean decision in the Recipe. This is also required in order to estimate in-loop reduction cost. This could be done separately.

[dmgreen]

llvm/lib/Transforms/Vectorize/VPlan.h

240

Do you mean to change the interface to createTargetReduction, to take a bool instead? Yeah I think that sounds good. I'd prefer to do it as a separate review as it does involve changing the interface. I will put a patch together.

I was imagining that we would change the cost to use getArithmeticReductionCost, which hopefully handles the details of how the target lowers reductions. I haven't looked deeply into the details yet though. That is on the list of things to do.

[Ayal]

This looks good to me, thanks! with last couple of nits.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

6563

nit: "... that leads from the loop exit value back.." - chain is now found top-down.

7326

nit: can record the recipe of Phi first, just to follow chain order.

7554

nit: can assert CompareRecipe->getVPRecipeID()

[dmgreen]

Thanks. I will update the patch, but I will wait until at least after the branch before I commit it.

D83646 is an attempt at changing the interface to createTargetReduction, so that we don't need to store TTI in the reduction recipe.

[dmgreen]

Updates for those last nitpicks.

[rupprecht]

FYI, reverted in 3c39db0c4452218c967a8ac3ad48144fbf1159ff. See http://lab.llvm.org:8011/builders/clang-cmake-armv7-quick/builds/19573/steps/build%20stage%201/logs/stdio for a breakage that's shown up in buildbots.

[dmgreen]

Thanks. Sorry about that, it appears clang doesn't like this code as much as gcc did and then my internet went out whilst I was trying to figure out what was wrong.

I will try to run some extra testing using different compilers.

llvm/include/llvm/Analysis/IVDescriptors.h

public:
 
   /// Returns a reference to the instructions used for type-promoting the
   /// recurrence.
   SmallPtrSet<Instruction *, 8> &getCastInsts() { return CastInsts; }
 
   /// Returns true if all source operands of the recurrence are SExtInsts.
   bool isSigned() { return IsSigned; }
 
+  /// Attempts to find a chain of operations from Phi to LoopExitInst that can
+  /// be treated as a set of reductions instructions for in-loop reductions.
+  SmallVector<Instruction *, 4> getReductionOpChain(PHINode *Phi,
+                                                    Loop *L) const;
+
 private:
   // The starting value of the recurrence.
   // It does not have to be zero!
   TrackingVH<Value> StartValue;
   // The instruction who's value is used outside the loop.
   Instruction *LoopExitInstr = nullptr;
   // The kind of the recurrence.
   RecurrenceKind Kind = RK_NoRecurrence;

llvm/lib/Analysis/IVDescriptors.cpp

if (Start != Phi) {
     // The recurrence expression will be represented in a narrower type. If
     // there are any cast instructions that will be unnecessary, collect them
     // in CastInsts. Note that the 'and' instruction was already included in
     // this list.
     //
     // TODO: A better way to represent this may be to tag in some way all the
     //       instructions that are a part of the reduction. The vectorizer cost
     //       model could then apply the recurrence type to these instructions,
     //       without needing a white list of instructions to ignore.
     collectCastsToIgnore(TheLoop, ExitInstruction, RecurrenceType, CastInsts);
   }
 
   // We found a reduction var if we have reached the original phi node and we
   // only have a single instruction with out-of-loop users.
 
   // The ExitInstruction(Instruction which is allowed to have out-of-loop users)
   // is saved as part of the RecurrenceDescriptor.
case RK_IntegerMinMax:
     return Instruction::ICmp;
   case RK_FloatMinMax:
     return Instruction::FCmp;
   default:
     llvm_unreachable("Unknown recurrence operation");
   }
 }
 
+SmallVector<Instruction *, 4>
+RecurrenceDescriptor::getReductionOpChain(PHINode *Phi, Loop *L) const {
+  SmallVector<Instruction *, 8> ReductionOperations;
+  unsigned RedOp = getRecurrenceBinOp(Kind);
+
+  // Search down from the Phi to the LoopExitInstr, looking for instructions
+  // with a single user of the correct type for the reduction.
+
+  // Note that we check that the type of the operand is correct for each item in
+  // the chain, including the last (the loop exit value). This can come up from
+  // sub, which would otherwise be treated as an add reduction. MinMax also need
+  // to check for a pair of icmp/select, for which we use getNextInstruction and
+  // isCorrectOpcode functions to step the right number of instruction, and
+  // check the icmp/select pair.
+  // FIXME: We also do not attempt to look through Phi/Select's yet, which might
+  // be part of the reduction chain, or attempt to looks through And's to find a
+  // smaller bitwidth.
+  unsigned ExpectedUses = 1;
+  if (RedOp == Instruction::ICmp || RedOp == Instruction::FCmp)
+    ExpectedUses = 2;
+
+  auto getNextInstruction = [&](Instruction *Cur) {
+    if (RedOp == Instruction::ICmp || RedOp == Instruction::FCmp) {
+      // We are expecting a icmp/select pair, which we go to the next select
+      // instruction if we can. We already know that Cur has 2 uses.
+      if (isa<SelectInst>(*Cur->user_begin()))
+        return cast<Instruction>(*Cur->user_begin());
+      else
+        return cast<Instruction>(*std::next(Cur->user_begin()));
+    }
+    return cast<Instruction>(*Cur->user_begin());
+  };
+  auto isCorrectOpcode = [&](Instruction *Cur) {
+    if (RedOp == Instruction::ICmp || RedOp == Instruction::FCmp) {
+      Value *LHS, *RHS;
+      return SelectPatternResult::isMinOrMax(
+          matchSelectPattern(Cur, LHS, RHS).Flavor);
+    }
+    return Cur->getOpcode() == RedOp;
+  };
+
+  // Check that the Phi has one (or two for min/max) uses.
+  if (!Phi->hasNUses(ExpectedUses))
+    return {};
+  Instruction *Cur = getNextInstruction(Phi);
+
+  // Each other instruciton in the chain should have the expected number of uses
+  // and be the correct opcode.
+  while (Cur != LoopExitInstr) {
+    if (!isCorrectOpcode(Cur) || !Cur->hasNUses(ExpectedUses))
+      return {};
+
+    ReductionOperations.push_back(Cur);
+    Cur = getNextInstruction(Cur);
+  }
+
+  // The loop exit instruction also needs to be the same opcode. We dont allow
+  // them to be Subs.
+  if (!isCorrectOpcode(Cur))
+    return {};
+
+  ReductionOperations.push_back(Cur);
+  return ReductionOperations;
+}
+
 InductionDescriptor::InductionDescriptor(Value *Start, InductionKind K,
                                          const SCEV *Step, BinaryOperator *BOp,
                                          SmallVectorImpl<Instruction *> *Casts)
     : StartValue(Start), IK(K), Step(Step), InductionBinOp(BOp) {
   assert(IK != IK_NoInduction && "Not an induction");
 
   // Start value type should match the induction kind and the value
   // itself should not be null.

llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h

 #include "llvm/Analysis/TargetLibraryInfo.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 
 namespace llvm {
 
 class LoopVectorizationLegality;
 class LoopVectorizationCostModel;
 class PredicatedScalarEvolution;
+class VPRecipeBuilder;
 
 /// VPlan-based builder utility analogous to IRBuilder.
 class VPBuilder {
   VPBasicBlock *BB = nullptr;
   VPBasicBlock::iterator InsertPt = VPBasicBlock::iterator();
 
   VPInstruction *createInstruction(unsigned Opcode,
                                    ArrayRef<VPValue *> Operands) {
VPlanPtr buildVPlanWithVPRecipes(
       VFRange &Range, SmallPtrSetImpl<Value *> &NeedDef,
       SmallPtrSetImpl<Instruction *> &DeadInstructions,
       const DenseMap<Instruction *, Instruction *> &SinkAfter);
 
   /// Build VPlans for power-of-2 VF's between \p MinVF and \p MaxVF inclusive,
   /// according to the information gathered by Legal when it checked if it is
   /// legal to vectorize the loop. This method creates VPlans using VPRecipes.
   void buildVPlansWithVPRecipes(unsigned MinVF, unsigned MaxVF);
+
+  /// Adjust the recipes for any inloop reductions. The chain of instructions
+  /// leading from the loop exit instr to the phi need to be converted to
+  /// reductions, with one operand being vector and the other being the scalar
+  /// reduction chain.
+  void adjustRecipesForInloopReductions(VPlanPtr &Plan,
+                                        VPRecipeBuilder &RecipeBuilder);
 };
 
 } // namespace llvm
 
 #endif // LLVM_TRANSFORMS_VECTORIZE_LOOPVECTORIZATIONPLANNER_H

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

static cl::opt<bool> EnableCondStoresVectorization(
     "enable-cond-stores-vec", cl::init(true), cl::Hidden,
     cl::desc("Enable if predication of stores during vectorization."));
 
 static cl::opt<unsigned> MaxNestedScalarReductionIC(
     "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
     cl::desc("The maximum interleave count to use when interleaving a scalar "
              "reduction in a nested loop."));
 
+static cl::opt<bool>
+    PreferInloopReductions("prefer-inloop-reductions", cl::init(false),
+                           cl::Hidden,
+                           cl::desc("Prefer in-loop vector reductions, "
+                                    "overriding the targets preference."));
+
 cl::opt<bool> EnableVPlanNativePath(
     "enable-vplan-native-path", cl::init(false), cl::Hidden,
     cl::desc("Enable VPlan-native vectorization path with "
              "support for outer loop vectorization."));
 
 // FIXME: Remove this switch once we have divergence analysis. Currently we
 // assume divergent non-backedge branches when this switch is true.
 cl::opt<bool> EnableVPlanPredication(
public:
 
   /// \return Returns information about the register usages of the loop for the
   /// given vectorization factors.
   SmallVector<RegisterUsage, 8> calculateRegisterUsage(ArrayRef<unsigned> VFs);
 
   /// Collect values we want to ignore in the cost model.
   void collectValuesToIgnore();
 
+  /// Split reductions into those that happen in the loop, and those that happen
+  /// outside.
+  void collectInLoopReductions();
+
   /// \returns The smallest bitwidth each instruction can be represented with.
   /// The vector equivalents of these instructions should be truncated to this
   /// type.
   const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const {
     return MinBWs;
   }
 
   /// \returns True if it is more profitable to scalarize instruction \p I for
public:
 
   /// Returns true if all loop blocks should be masked to fold tail loop.
   bool foldTailByMasking() const { return FoldTailByMasking; }
 
   bool blockNeedsPredication(BasicBlock *BB) {
     return foldTailByMasking() || Legal->blockNeedsPredication(BB);
   }
 
+  /// Returns true if the Phi is part of an inloop reduction.
+  bool isInLoopReduction(PHINode *Phi) const {
+    return InloopReductionChains.count(Phi);
+  }
+
+  /// Return the chain of instructions representing an inloop reduction.
+  const SmallVector<Instruction *, 4> &
+  getInLoopReductionChain(PHINode *Phi) const {
+    assert(InloopReductionChains.count(Phi) &&
+           "Expected to find an inloop reduction chain!");
+    return InloopReductionChains.find(Phi)->second;
+  }
+
+  /// Returns true if there are any inloop reductions.
+  bool hasInloopReductions() const { return !InloopReductionChains.empty(); }
+
   /// Returns true if there are any outside-loop reductions.
   bool hasOutOfLoopReductions() const {
-    return !Legal->getReductionVars().empty();
+    return InloopReductionChains.size() < Legal->getReductionVars().size();
   }
 
   /// Estimate cost of an intrinsic call instruction CI if it were vectorized
   /// with factor VF.  Return the cost of the instruction, including
   /// scalarization overhead if it's needed.
   unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF);
 
   /// Estimate cost of a call instruction CI if it were vectorized with factor
private:
   /// Holds the instructions known to be scalar after vectorization.
   /// The data is collected per VF.
   DenseMap<unsigned, SmallPtrSet<Instruction *, 4>> Scalars;
 
   /// Holds the instructions (address computations) that are forced to be
   /// scalarized.
   DenseMap<unsigned, SmallPtrSet<Instruction *, 4>> ForcedScalars;
 
+  /// PHINodes of the reductions that should be expanded in-loop along with
+  /// their associated chains of reduction operations, in program order from top
+  /// (PHI) to bottom
+  DenseMap<PHINode *, SmallVector<Instruction *, 4>> InloopReductionChains;
+
   /// Returns the expected difference in cost from scalarizing the expression
   /// feeding a predicated instruction \p PredInst. The instructions to
   /// scalarize and their scalar costs are collected in \p ScalarCosts. A
   /// non-negative return value implies the expression will be scalarized.
   /// Currently, only single-use chains are considered for scalarization.
   int computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts,
                               unsigned VF);
 
void InnerLoopVectorizer::fixReduction(PHINode *Phi) {
   RecurrenceDescriptor RdxDesc = Legal->getReductionVars()[Phi];
 
   RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
   TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
   Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
   RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
     RdxDesc.getMinMaxRecurrenceKind();
   setDebugLocFromInst(Builder, ReductionStartValue);
+  bool isInLoopReductionPhi = Cost->isInLoopReduction(Phi);
 
   // We need to generate a reduction vector from the incoming scalar.
   // To do so, we need to generate the 'identity' vector and override
   // one of the elements with the incoming scalar reduction. We need
   // to do it in the vector-loop preheader.
   Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
 
   // This is the vector-clone of the value that leaves the loop.
   Type *VecTy = getOrCreateVectorValue(LoopExitInst, 0)->getType();
 
   // Find the reduction identity variable. Zero for addition, or, xor,
   // one for multiplication, -1 for And.
   Value *Identity;
   Value *VectorStart;
   if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
       RK == RecurrenceDescriptor::RK_FloatMinMax) {
     // MinMax reduction have the start value as their identify.
-    if (VF == 1) {
+    if (VF == 1 || isInLoopReductionPhi) {
       VectorStart = Identity = ReductionStartValue;
     } else {
       VectorStart = Identity =
         Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
     }
   } else {
     // Handle other reduction kinds:
     Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
         RK, VecTy->getScalarType());
-    if (VF == 1) {
+    if (VF == 1 || isInLoopReductionPhi) {
       Identity = Iden;
       // This vector is the Identity vector where the first element is the
       // incoming scalar reduction.
       VectorStart = ReductionStartValue;
     } else {
       Identity = ConstantVector::getSplat({VF, false}, Iden);
 
       // This vector is the Identity vector where the first element is the
for (unsigned Part = 0; Part < UF; ++Part) {
       VectorLoopValueMap.resetVectorValue(LoopExitInst, Part, Sel);
     }
   }
 
   // If the vector reduction can be performed in a smaller type, we truncate
   // then extend the loop exit value to enable InstCombine to evaluate the
   // entire expression in the smaller type.
   if (VF > 1 && Phi->getType() != RdxDesc.getRecurrenceType()) {
+    assert(!isInLoopReductionPhi && "Unexpected truncated inloop reduction!");
     Type *RdxVecTy = FixedVectorType::get(RdxDesc.getRecurrenceType(), VF);
     Builder.SetInsertPoint(
         LI->getLoopFor(LoopVectorBody)->getLoopLatch()->getTerminator());
     VectorParts RdxParts(UF);
     for (unsigned Part = 0; Part < UF; ++Part) {
       RdxParts[Part] = VectorLoopValueMap.getVectorValue(LoopExitInst, Part);
       Value *Trunc = Builder.CreateTrunc(RdxParts[Part], RdxVecTy);
       Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
if (Op != Instruction::ICmp && Op != Instruction::FCmp)
           Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxPart,
                               ReducedPartRdx, "bin.rdx"),
           RdxDesc.getFastMathFlags());
     else
       ReducedPartRdx = createMinMaxOp(Builder, MinMaxKind, ReducedPartRdx,
                                       RdxPart);
   }
 
-  if (VF > 1) {
+  // Create the reduction after the loop. Note that inloop reductions create the
+  // target reduction in the loop using a Reduction recipe.
+  if (VF > 1 && !isInLoopReductionPhi) {
     bool NoNaN = Legal->hasFunNoNaNAttr();
     ReducedPartRdx =
         createTargetReduction(Builder, TTI, RdxDesc, ReducedPartRdx, NoNaN);
     // If the reduction can be performed in a smaller type, we need to extend
     // the reduction to the wider type before we branch to the original loop.
     if (Phi->getType() != RdxDesc.getRecurrenceType())
       ReducedPartRdx =
         RdxDesc.isSigned()
void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, unsigned UF,
 
   // In order to support recurrences we need to be able to vectorize Phi nodes.
   // Phi nodes have cycles, so we need to vectorize them in two stages. This is
   // stage #1: We create a new vector PHI node with no incoming edges. We'll use
   // this value when we vectorize all of the instructions that use the PHI.
   if (Legal->isReductionVariable(P) || Legal->isFirstOrderRecurrence(P)) {
     for (unsigned Part = 0; Part < UF; ++Part) {
       // This is phase one of vectorizing PHIs.
+      bool ScalarPHI = (VF == 1) || Cost->isInLoopReduction(cast<PHINode>(PN));
       Type *VecTy =
-          (VF == 1) ? PN->getType() : FixedVectorType::get(PN->getType(), VF);
+          ScalarPHI ? PN->getType() : FixedVectorType::get(PN->getType(), VF);
       Value *EntryPart = PHINode::Create(
           VecTy, 2, "vec.phi", &*LoopVectorBody->getFirstInsertionPt());
       VectorLoopValueMap.setVectorValue(P, Part, EntryPart);
     }
     return;
   }
 
   setDebugLocFromInst(Builder, P);
void LoopVectorizationCostModel::collectValuesToIgnore() {
   // detection.
   for (auto &Induction : Legal->getInductionVars()) {
     InductionDescriptor &IndDes = Induction.second;
     const SmallVectorImpl<Instruction *> &Casts = IndDes.getCastInsts();
     VecValuesToIgnore.insert(Casts.begin(), Casts.end());
   }
 }
 
+void LoopVectorizationCostModel::collectInLoopReductions() {
+  // For the moment, without predicated reduction instructions, we do not
+  // support inloop reductions whilst folding the tail, and hence in those cases
+  // all reductions are currently out of the loop.
+  if (foldTailByMasking())
+    return;
+
+  for (auto &Reduction : Legal->getReductionVars()) {
+    PHINode *Phi = Reduction.first;
+    RecurrenceDescriptor &RdxDesc = Reduction.second;
+
+    // We don't collect reductions that are type promoted (yet).
+    if (RdxDesc.getRecurrenceType() != Phi->getType())
+      continue;
+
+    // If the target would prefer this reduction to happen "in-loop", then we
+    // want to record it as such.
+    if (!PreferInloopReductions)
+      continue;
+
+    // Check that we can correctly put the reductions into the loop, by
+    // finding the chain of operations that leads from the loop exit value back
+    // to the phi.
+    SmallVector<Instruction *, 4> ReductionOperations =
+        RdxDesc.getReductionOpChain(Phi, TheLoop);
+    bool InLoop = !ReductionOperations.empty();
+    if (InLoop)
+      InloopReductionChains[Phi] = ReductionOperations;
+    LLVM_DEBUG(dbgs() << "LV: Using " << (InLoop ? "inloop" : "out of loop")
+                      << " reduction for phi: " << *Phi << "\n");
+  }
+}
+
 // TODO: we could return a pair of values that specify the max VF and
 // min VF, to be used in `buildVPlans(MinVF, MaxVF)` instead of
 // `buildVPlans(VF, VF)`. We cannot do it because VPLAN at the moment
 // doesn't have a cost model that can choose which plan to execute if
 // more than one is generated.
 static unsigned determineVPlanVF(const unsigned WidestVectorRegBits,
                                  LoopVectorizationCostModel &CM) {
   unsigned WidestType;
Optional<VectorizationFactor> LoopVectorizationPlanner::plan(unsigned UserVF,
   }
 
   if (UserVF) {
     LLVM_DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
     assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
     // Collect the instructions (and their associated costs) that will be more
     // profitable to scalarize.
     CM.selectUserVectorizationFactor(UserVF);
+    CM.collectInLoopReductions();
     buildVPlansWithVPRecipes(UserVF, UserVF);
     LLVM_DEBUG(printPlans(dbgs()));
     return {{UserVF, 0}};
   }
 
   unsigned MaxVF = MaybeMaxVF.getValue();
   assert(MaxVF != 0 && "MaxVF is zero.");
 
   for (unsigned VF = 1; VF <= MaxVF; VF *= 2) {
     // Collect Uniform and Scalar instructions after vectorization with VF.
     CM.collectUniformsAndScalars(VF);
 
     // Collect the instructions (and their associated costs) that will be more
     // profitable to scalarize.
     if (VF > 1)
       CM.collectInstsToScalarize(VF);
   }
 
+  CM.collectInLoopReductions();
+
   buildVPlansWithVPRecipes(1, MaxVF);
   LLVM_DEBUG(printPlans(dbgs()));
   if (MaxVF == 1)
     return VectorizationFactor::Disabled();
 
   // Select the optimal vectorization factor.
   return CM.selectVectorizationFactor(MaxVF);
 }
VPlanPtr LoopVectorizationPlanner::buildVPlanWithVPRecipes(
   // ---------------------------------------------------------------------------
 
   // Mark instructions we'll need to sink later and their targets as
   // ingredients whose recipe we'll need to record.
   for (auto &Entry : SinkAfter) {
     RecipeBuilder.recordRecipeOf(Entry.first);
     RecipeBuilder.recordRecipeOf(Entry.second);
   }
+  for (auto &Reduction : Legal->getReductionVars()) {
+    if (!CM.isInLoopReduction(Reduction.first))
+      continue;
+    PHINode *Phi = Reduction.first;
+    RecurrenceDescriptor::RecurrenceKind Kind =
+        Reduction.second.getRecurrenceKind();
+    const SmallVector<Instruction *, 4> &ReductionOperations =
+        CM.getInLoopReductionChain(Phi);
+    for (auto &R : ReductionOperations) {
+      RecipeBuilder.recordRecipeOf(R);
+      // For min/max reducitons, where we have a pair of icmp/select, we also
+      // need to recode the ICmp recipe, so it can be removed later.
+      if (Kind == RecurrenceDescriptor::RK_IntegerMinMax ||
+          Kind == RecurrenceDescriptor::RK_FloatMinMax) {
+        RecipeBuilder.recordRecipeOf(cast<Instruction>(R->getOperand(0)));
+      }
+    }
+    RecipeBuilder.recordRecipeOf(Phi);
+  }
 
   // For each interleave group which is relevant for this (possibly trimmed)
   // Range, add it to the set of groups to be later applied to the VPlan and add
   // placeholders for its members' Recipes which we'll be replacing with a
   // single VPInterleaveRecipe.
   for (InterleaveGroup<Instruction> *IG : IAI.getInterleaveGroups()) {
     auto applyIG = [IG, this](unsigned VF) -> bool {
       return (VF >= 2 && // Query is illegal for VF == 1
(new VPInterleaveRecipe(IG, Recipe->getAddr(), Recipe->getMask()))
         ->insertBefore(Recipe);
 
     for (unsigned i = 0; i < IG->getFactor(); ++i)
       if (Instruction *Member = IG->getMember(i)) {
         RecipeBuilder.getRecipe(Member)->eraseFromParent();
       }
   }
 
+  // Adjust the recipes for any inloop reductions.
+  if (Range.Start > 1)
+    adjustRecipesForInloopReductions(Plan, RecipeBuilder);
+
   // Finally, if tail is folded by masking, introduce selects between the phi
   // and the live-out instruction of each reduction, at the end of the latch.
   if (CM.foldTailByMasking() && CM.hasOutOfLoopReductions()) {
     Builder.setInsertPoint(VPBB);
     auto *Cond = RecipeBuilder.createBlockInMask(OrigLoop->getHeader(), Plan);
     for (auto &Reduction : Legal->getReductionVars()) {
+      assert(!CM.isInLoopReduction(Reduction.first) &&
+             "Didn't expect inloop tail folded reduction yet!");
       VPValue *Phi = Plan->getVPValue(Reduction.first);
       VPValue *Red = Plan->getVPValue(Reduction.second.getLoopExitInstr());
       Builder.createNaryOp(Instruction::Select, {Cond, Red, Phi});
     }
   }
 
   std::string PlanName;
   raw_string_ostream RSO(PlanName);
   unsigned VF = Range.Start;
   Plan->addVF(VF);
   RSO << "Initial VPlan for VF={" << VF;
   for (VF *= 2; VF < Range.End; VF *= 2) {
     Plan->addVF(VF);
     RSO << "," << VF;
   }
   RSO << "},UF>=1";
   RSO.flush();
   Plan->setName(PlanName);
 
   return Plan;
 }
 
 VPlanPtr LoopVectorizationPlanner::buildVPlan(VFRange &Range) {
   // Outer loop handling: They may require CFG and instruction level
   // transformations before even evaluating whether vectorization is profitable.
   // Since we cannot modify the incoming IR, we need to build VPlan upfront in
   // the vectorization pipeline.
   assert(!OrigLoop->empty());
   assert(EnableVPlanNativePath && "VPlan-native path is not enabled.");
 
   // Create new empty VPlan
   auto Plan = std::make_unique<VPlan>();
VPlanPtr LoopVectorizationPlanner::buildVPlan(VFRange &Range) {
   }
 
   SmallPtrSet<Instruction *, 1> DeadInstructions;
   VPlanTransforms::VPInstructionsToVPRecipes(
       OrigLoop, Plan, Legal->getInductionVars(), DeadInstructions);
   return Plan;
 }
 
+// Adjust the recipes for any inloop reductions. The chain of instructions
+// leading from the loop exit instr to the phi need to be converted to
+// reductions, with one operand being vector and the other being the scalar
+// reduction chain.
+void LoopVectorizationPlanner::adjustRecipesForInloopReductions(
+    VPlanPtr &Plan, VPRecipeBuilder &RecipeBuilder) {
+  for (auto &Reduction : Legal->getReductionVars()) {
+    PHINode *Phi = Reduction.first;
+    RecurrenceDescriptor &RdxDesc = Reduction.second;
+    if (!CM.isInLoopReduction(Phi))
+      continue;
+
+    const SmallVector<Instruction *, 4> &ReductionOperations =
+        CM.getInLoopReductionChain(Phi);
+
+    // ReductionOperations are orders top-down from the phi's use to the
+    // LoopExitValue. We keep a track of the previous item (the Chain) to tell
+    // which of the two operands will remain scalar and which will be reduced.
+    // For minmax the chain will be the select instructions.
+    Instruction *Chain = Phi;
+    for (Instruction *R : ReductionOperations) {
+      VPRecipeBase *WidenRecipe = RecipeBuilder.getRecipe(R);
+      RecurrenceDescriptor::RecurrenceKind Kind = RdxDesc.getRecurrenceKind();
+
+      VPValue *ChainOp = Plan->getVPValue(Chain);
+      unsigned FirstOpId;
+      if (Kind == RecurrenceDescriptor::RK_IntegerMinMax ||
+          Kind == RecurrenceDescriptor::RK_FloatMinMax) {
+        assert(WidenRecipe->getVPRecipeID() ==
+                    VPRecipeBase::VPWidenSelectSC &&
+                "Expected to replace a VPWidenSelectSC");
+        FirstOpId = 1;
+      } else {
+        assert(WidenRecipe->getVPRecipeID() == VPRecipeBase::VPWidenSC &&
+                "Expected to replace a VPWidenSC");
+        FirstOpId = 0;
+      }
+      unsigned VecOpId =
+          R->getOperand(FirstOpId) == Chain ? FirstOpId + 1 : FirstOpId;
+      VPValue *VecOp = Plan->getVPValue(R->getOperand(VecOpId));
+
+      VPReductionRecipe *RedRecipe = new VPReductionRecipe(
+          &RdxDesc, R, ChainOp, VecOp, Legal->hasFunNoNaNAttr(), TTI);
+      WidenRecipe->getParent()->insert(RedRecipe, WidenRecipe->getIterator());
+      WidenRecipe->removeFromParent();
+
+      if (Kind == RecurrenceDescriptor::RK_IntegerMinMax ||
+          Kind == RecurrenceDescriptor::RK_FloatMinMax) {
+        VPRecipeBase *CompareRecipe =
+            RecipeBuilder.getRecipe(cast<Instruction>(R->getOperand(0)));
+        CompareRecipe->removeFromParent();
+      }
+      Chain = R;
+    }
+  }
+}
+
 Value* LoopVectorizationPlanner::VPCallbackILV::
 getOrCreateVectorValues(Value *V, unsigned Part) {
       return ILV.getOrCreateVectorValue(V, Part);
 }
 
 Value *LoopVectorizationPlanner::VPCallbackILV::getOrCreateScalarValue(
     Value *V, const VPIteration &Instance) {
   return ILV.getOrCreateScalarValue(V, Instance);
for (unsigned Part = 0; Part < State.UF; ++Part)
     State.ValueMap.setVectorValue(Phi, Part, Entry[Part]);
 }
 
 void VPInterleaveRecipe::execute(VPTransformState &State) {
   assert(!State.Instance && "Interleave group being replicated.");
   State.ILV->vectorizeInterleaveGroup(IG, State, getAddr(), getMask());
 }
 
+void VPReductionRecipe::execute(VPTransformState &State) {
+  assert(!State.Instance && "Reduction being replicated.");
+  for (unsigned Part = 0; Part < State.UF; ++Part) {
+    unsigned Kind = RdxDesc->getRecurrenceKind();
+    Value *NewVecOp = State.get(VecOp, Part);
+    Value *NewRed =
+        createTargetReduction(State.Builder, TTI, *RdxDesc, NewVecOp, NoNaN);
+    Value *PrevInChain = State.get(ChainOp, Part);
+    Value *NextInChain;
+    if (Kind == RecurrenceDescriptor::RK_IntegerMinMax ||
+        Kind == RecurrenceDescriptor::RK_FloatMinMax) {
+      NextInChain =
+          createMinMaxOp(State.Builder, RdxDesc->getMinMaxRecurrenceKind(),
+                         NewRed, PrevInChain);
+    } else {
+      NextInChain = State.Builder.CreateBinOp(
+          (Instruction::BinaryOps)I->getOpcode(), NewRed, PrevInChain);
+    }
+    State.ValueMap.setVectorValue(I, Part, NextInChain);
+  }
+}
+
 void VPReplicateRecipe::execute(VPTransformState &State) {
   if (State.Instance) { // Generate a single instance.
     State.ILV->scalarizeInstruction(Ingredient, User, *State.Instance,
                                     IsPredicated, State);
     // Insert scalar instance packing it into a vector.
     if (AlsoPack && State.VF > 1) {
       // If we're constructing lane 0, initialize to start from undef.
       if (State.Instance->Lane == 0) {

llvm/lib/Transforms/Vectorize/VPlan.h

 #include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/SmallSet.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/Twine.h"
 #include "llvm/ADT/ilist.h"
 #include "llvm/ADT/ilist_node.h"
 #include "llvm/Analysis/VectorUtils.h"
 #include "llvm/IR/IRBuilder.h"
 #include <algorithm>
 #include <cassert>
 #include <cstddef>
 #include <map>
 #include <string>
 
 namespace llvm {
 
 class BasicBlock;
 class DominatorTree;
 class InnerLoopVectorizer;
 template <class T> class InterleaveGroup;
 class LoopInfo;
 class raw_ostream;
+class RecurrenceDescriptor;
 class Value;
 class VPBasicBlock;
 class VPRegionBlock;
 class VPSlotTracker;
 class VPlan;
 class VPlanSlp;
 
 /// A range of powers-of-2 vectorization factors with fixed start and

 };
 
 /// VPTransformState holds information passed down when "executing" a VPlan,
 /// needed for generating the output IR.
 struct VPTransformState {
   VPTransformState(unsigned VF, unsigned UF, LoopInfo *LI, DominatorTree *DT,
                    IRBuilder<> &Builder, VectorizerValueMap &ValueMap,
                    InnerLoopVectorizer *ILV, VPCallback &Callback)
       : VF(VF), UF(UF), Instance(), LI(LI), DT(DT), Builder(Builder),
         ValueMap(ValueMap), ILV(ILV), Callback(Callback) {}
 
   /// The chosen Vectorization and Unroll Factors of the loop being vectorized.
   unsigned VF;
   unsigned UF;
 
   /// Hold the indices to generate specific scalar instructions. Null indicates
   /// that all instances are to be generated, using either scalar or vector
public:
   /// SubclassID field of the VPRecipeBase objects. They are used for concrete
   /// type identification.
   using VPRecipeTy = enum {
     VPBlendSC,
     VPBranchOnMaskSC,
     VPInstructionSC,
     VPInterleaveSC,
     VPPredInstPHISC,
+    VPReductionSC,
     VPReplicateSC,
     VPWidenCallSC,
     VPWidenCanonicalIVSC,
     VPWidenGEPSC,
     VPWidenIntOrFpInductionSC,
     VPWidenMemoryInstructionSC,
     VPWidenPHISC,
     VPWidenSC,
public:
 
   /// Print the recipe.
   void print(raw_ostream &O, const Twine &Indent,
              VPSlotTracker &SlotTracker) const override;
 
   const InterleaveGroup<Instruction> *getInterleaveGroup() { return IG; }
 };
 
+/// A recipe to represent inloop reduction operations, performing a reduction on
+/// a vector operand into a scalar value, and adding the result to a chain.
+class VPReductionRecipe : public VPRecipeBase {
+  /// The recurrence decriptor for the reduction in question.
+  RecurrenceDescriptor *RdxDesc;
+  /// The original instruction being converted to a reduction.
+  Instruction *I;
+  /// The VPValue of the vector value to be reduced.
+  VPValue *VecOp;
+  /// The VPValue of the scalar Chain being accumulated.
+  VPValue *ChainOp;
+  /// Fast math flags to use for the resulting reduction operation.
+  bool NoNaN;
+  /// Pointer to the TTI, needed to create the target reduction
+  const TargetTransformInfo *TTI;
+
+public:
+  VPReductionRecipe(RecurrenceDescriptor *R, Instruction *I, VPValue *ChainOp,
+                    VPValue *VecOp, bool NoNaN, const TargetTransformInfo *TTI)
+      : VPRecipeBase(VPReductionSC), RdxDesc(R), I(I), VecOp(VecOp),
+        ChainOp(ChainOp), NoNaN(NoNaN), TTI(TTI) {}
+
+  ~VPReductionRecipe() override = default;
+
+  /// Method to support type inquiry through isa, cast, and dyn_cast.
+  static inline bool classof(const VPRecipeBase *V) {
+    return V->getVPRecipeID() == VPRecipeBase::VPReductionSC;
+  }
+
+  /// Generate the reduction in the loop
+  void execute(VPTransformState &State) override;
+
+  /// Print the recipe.
+  void print(raw_ostream &O, const Twine &Indent,
+             VPSlotTracker &SlotTracker) const override;
+};
+
 /// VPReplicateRecipe replicates a given instruction producing multiple scalar
 /// copies of the original scalar type, one per lane, instead of producing a
 /// single copy of widened type for all lanes. If the instruction is known to be
 /// uniform only one copy, per lane zero, will be generated.
 class VPReplicateRecipe : public VPRecipeBase {
   /// The instruction being replicated.
   Instruction *Ingredient;
 

llvm/lib/Transforms/Vectorize/VPlan.cpp

for (unsigned I = 0, E = getNumIncomingValues(); I < E; ++I) {
       O << " ";
       getIncomingValue(I)->printAsOperand(O, SlotTracker);
       O << "/";
       getMask(I)->printAsOperand(O, SlotTracker);
     }
   }
 }
 
+void VPReductionRecipe::print(raw_ostream &O, const Twine &Indent,
+                              VPSlotTracker &SlotTracker) const {
+  O << "\"REDUCE of" << *I << " as ";
+  ChainOp->printAsOperand(O, SlotTracker);
+  O << " + reduce(";
+  VecOp->printAsOperand(O, SlotTracker);
+  O << ")";
+}
+
 void VPReplicateRecipe::print(raw_ostream &O, const Twine &Indent,
                               VPSlotTracker &SlotTracker) const {
   O << "\"" << (IsUniform ? "CLONE " : "REPLICATE ")
     << VPlanIngredient(Ingredient);
   if (AlsoPack)
     O << " (S->V)";
 }
 

llvm/test/Transforms/LoopVectorize/reduction-inloop-uf4.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
-; RUN: opt < %s  -loop-vectorize -force-vector-interleave=4 -force-vector-width=4 -force-reduction-intrinsics -dce -instcombine -S | FileCheck %s
+; RUN: opt < %s  -loop-vectorize -force-vector-interleave=4 -force-vector-width=4 -prefer-inloop-reductions -force-reduction-intrinsics -dce -instcombine -S | FileCheck %s
 
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
 
 define i32 @reduction_sum_single(i32* noalias nocapture %A) {
 ; CHECK-LABEL: @reduction_sum_single(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP8:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI2:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP10:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI3:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP11:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP11:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI2:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP13:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI3:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP15:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[TMP0]], i64 4
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD4:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
 ; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i32, i32* [[TMP0]], i64 8
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32* [[TMP4]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD5:%.*]] = load <4 x i32>, <4 x i32>* [[TMP5]], align 4
 ; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr inbounds i32, i32* [[TMP0]], i64 12
 ; CHECK-NEXT:    [[TMP7:%.*]] = bitcast i32* [[TMP6]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD6:%.*]] = load <4 x i32>, <4 x i32>* [[TMP7]], align 4
-; CHECK-NEXT:    [[TMP8]] = add <4 x i32> [[VEC_PHI]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP9]] = add <4 x i32> [[VEC_PHI1]], [[WIDE_LOAD4]]
-; CHECK-NEXT:    [[TMP10]] = add <4 x i32> [[VEC_PHI2]], [[WIDE_LOAD5]]
-; CHECK-NEXT:    [[TMP11]] = add <4 x i32> [[VEC_PHI3]], [[WIDE_LOAD6]]
+; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[TMP9]] = add i32 [[TMP8]], [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP10:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD4]])
+; CHECK-NEXT:    [[TMP11]] = add i32 [[TMP10]], [[VEC_PHI1]]
+; CHECK-NEXT:    [[TMP12:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD5]])
+; CHECK-NEXT:    [[TMP13]] = add i32 [[TMP12]], [[VEC_PHI2]]
+; CHECK-NEXT:    [[TMP14:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD6]])
+; CHECK-NEXT:    [[TMP15]] = add i32 [[TMP14]], [[VEC_PHI3]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 16
-; CHECK-NEXT:    [[TMP12:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP12]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !0
+; CHECK-NEXT:    [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP16]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !0
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[BIN_RDX:%.*]] = add <4 x i32> [[TMP9]], [[TMP8]]
-; CHECK-NEXT:    [[BIN_RDX7:%.*]] = add <4 x i32> [[TMP10]], [[BIN_RDX]]
-; CHECK-NEXT:    [[BIN_RDX8:%.*]] = add <4 x i32> [[TMP11]], [[BIN_RDX7]]
-; CHECK-NEXT:    [[TMP13:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX8]])
+; CHECK-NEXT:    [[BIN_RDX:%.*]] = add i32 [[TMP11]], [[TMP9]]
+; CHECK-NEXT:    [[BIN_RDX7:%.*]] = add i32 [[TMP13]], [[BIN_RDX]]
+; CHECK-NEXT:    [[BIN_RDX8:%.*]] = add i32 [[TMP15]], [[BIN_RDX7]]
 ; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
 ; CHECK:       .lr.ph:
 ; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], !llvm.loop !2
 ; CHECK:       ._crit_edge:
-; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP13]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[BIN_RDX8]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[SUM_0_LCSSA]]
 ;
 entry:
   br label %.lr.ph
 
 .lr.ph:                                           ; preds = %entry, %.lr.ph
   %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]
   %sum.02 = phi i32 [ %l7, %.lr.ph ], [ 0, %entry ]

llvm/test/Transforms/LoopVectorize/reduction-inloop.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
-; RUN: opt < %s  -loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -force-reduction-intrinsics -dce -instcombine -S | FileCheck %s
+; RUN: opt < %s  -loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -prefer-inloop-reductions -force-reduction-intrinsics -dce -instcombine -S | FileCheck %s
 
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
 
 define i32 @reduction_sum_single(i32* noalias nocapture %A) {
 ; CHECK-LABEL: @reduction_sum_single(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP2:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP3:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
-; CHECK-NEXT:    [[TMP2]] = add <4 x i32> [[VEC_PHI]], [[WIDE_LOAD]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[TMP3]] = add i32 [[TMP2]], [[VEC_PHI]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP3:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP3]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !0
+; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP4]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !0
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP4:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[TMP2]])
 ; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
 ; CHECK:       .lr.ph:
 ; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], !llvm.loop !2
 ; CHECK:       ._crit_edge:
-; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP4]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP3]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[SUM_0_LCSSA]]
 ;
 entry:
   br label %.lr.ph
 
 .lr.ph:                                           ; preds = %entry, %.lr.ph
   %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]
   %sum.02 = phi i32 [ %l7, %.lr.ph ], [ 0, %entry ]
   %l2 = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
   %l3 = load i32, i32* %l2, align 4
   %l7 = add i32 %sum.02, %l3
   %indvars.iv.next = add i64 %indvars.iv, 1
   %lftr.wideiv = trunc i64 %indvars.iv.next to i32
   %exitcond = icmp eq i32 %lftr.wideiv, 256
   br i1 %exitcond, label %._crit_edge, label %.lr.ph
 
 ._crit_edge:                                      ; preds = %.lr.ph
   %sum.0.lcssa = phi i32 [ %l7, %.lr.ph ]
   ret i32 %sum.0.lcssa
 }
 
 define i32 @reduction_sum(i32* noalias nocapture %A, i32* noalias nocapture %B) {
 ; CHECK-LABEL: @reduction_sum(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[VEC_IND2:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT3:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
-; CHECK-NEXT:    [[TMP4:%.*]] = add <4 x i32> [[VEC_PHI]], [[VEC_IND2]]
-; CHECK-NEXT:    [[TMP5:%.*]] = add <4 x i32> [[TMP4]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP6]] = add <4 x i32> [[TMP5]], [[WIDE_LOAD1]]
+; CHECK-NEXT:    [[TMP4:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[VEC_IND2]])
+; CHECK-NEXT:    [[TMP5:%.*]] = add i32 [[TMP4]], [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[TMP7:%.*]] = add i32 [[TMP6]], [[TMP5]]
+; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD1]])
+; CHECK-NEXT:    [[TMP9]] = add i32 [[TMP8]], [[TMP7]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
 ; CHECK-NEXT:    [[VEC_IND_NEXT3]] = add <4 x i32> [[VEC_IND2]], <i32 4, i32 4, i32 4, i32 4>
-; CHECK-NEXT:    [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !4
+; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !4
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[TMP6]])
 ; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
 ; CHECK:       .lr.ph:
 ; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], !llvm.loop !5
 ; CHECK:       ._crit_edge:
-; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP8]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP9]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[SUM_0_LCSSA]]
 ;
 entry:
   br label %.lr.ph
 
 .lr.ph:                                           ; preds = %entry, %.lr.ph
   %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]
   %sum.02 = phi i32 [ %l9, %.lr.ph ], [ 0, %entry ]

 define i32 @reduction_sum_const(i32* noalias nocapture %A) {
 ; CHECK-LABEL: @reduction_sum_const(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP3:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP5:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
-; CHECK-NEXT:    [[TMP2:%.*]] = add <4 x i32> [[VEC_PHI]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP3]] = add <4 x i32> [[TMP2]], <i32 3, i32 3, i32 3, i32 3>
+; CHECK-NEXT:    [[TMP2:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[TMP3:%.*]] = add i32 [[TMP2]], [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP4:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> <i32 3, i32 3, i32 3, i32 3>)
+; CHECK-NEXT:    [[TMP5]] = add i32 [[TMP4]], [[TMP3]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP4]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !6
+; CHECK-NEXT:    [[TMP6:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP6]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !6
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP5:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[TMP3]])
 ; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
 ; CHECK:       .lr.ph:
 ; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], !llvm.loop !7
 ; CHECK:       ._crit_edge:
 ; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP5]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[SUM_0_LCSSA]]

 define i32 @reduction_prod(i32* noalias nocapture %A, i32* noalias nocapture %B) {
 ; CHECK-LABEL: @reduction_prod(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ <i32 1, i32 1, i32 1, i32 1>, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 1, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[VEC_IND2:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT3:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
-; CHECK-NEXT:    [[TMP4:%.*]] = mul <4 x i32> [[VEC_PHI]], [[VEC_IND2]]
-; CHECK-NEXT:    [[TMP5:%.*]] = mul <4 x i32> [[TMP4]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP6]] = mul <4 x i32> [[TMP5]], [[WIDE_LOAD1]]
+; CHECK-NEXT:    [[TMP4:%.*]] = call i32 @llvm.experimental.vector.reduce.mul.v4i32(<4 x i32> [[VEC_IND2]])
+; CHECK-NEXT:    [[TMP5:%.*]] = mul i32 [[TMP4]], [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.experimental.vector.reduce.mul.v4i32(<4 x i32> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[TMP7:%.*]] = mul i32 [[TMP6]], [[TMP5]]
+; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.experimental.vector.reduce.mul.v4i32(<4 x i32> [[WIDE_LOAD1]])
+; CHECK-NEXT:    [[TMP9]] = mul i32 [[TMP8]], [[TMP7]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
 ; CHECK-NEXT:    [[VEC_IND_NEXT3]] = add <4 x i32> [[VEC_IND2]], <i32 4, i32 4, i32 4, i32 4>
-; CHECK-NEXT:    [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !8
+; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !8
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.experimental.vector.reduce.mul.v4i32(<4 x i32> [[TMP6]])
 ; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
 ; CHECK:       .lr.ph:
 ; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], !llvm.loop !9
 ; CHECK:       ._crit_edge:
-; CHECK-NEXT:    [[PROD_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP8]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[PROD_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP9]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[PROD_0_LCSSA]]
 ;
 entry:
   br label %.lr.ph
 
 .lr.ph:                                           ; preds = %entry, %.lr.ph
   %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]
   %prod.02 = phi i32 [ %l9, %.lr.ph ], [ 1, %entry ]

 define i32 @reduction_mix(i32* noalias nocapture %A, i32* noalias nocapture %B) {
 ; CHECK-LABEL: @reduction_mix(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP8:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[VEC_IND2:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT3:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
 ; CHECK-NEXT:    [[TMP4:%.*]] = mul nsw <4 x i32> [[WIDE_LOAD1]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP5:%.*]] = add <4 x i32> [[VEC_PHI]], [[VEC_IND2]]
-; CHECK-NEXT:    [[TMP6]] = add <4 x i32> [[TMP5]], [[TMP4]]
+; CHECK-NEXT:    [[TMP5:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[VEC_IND2]])
+; CHECK-NEXT:    [[TMP6:%.*]] = add i32 [[TMP5]], [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP7:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[TMP4]])
+; CHECK-NEXT:    [[TMP8]] = add i32 [[TMP7]], [[TMP6]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
 ; CHECK-NEXT:    [[VEC_IND_NEXT3]] = add <4 x i32> [[VEC_IND2]], <i32 4, i32 4, i32 4, i32 4>
-; CHECK-NEXT:    [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !10
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP9]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !10
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[TMP6]])
 ; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
 ; CHECK:       .lr.ph:
 ; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], !llvm.loop !11
 ; CHECK:       ._crit_edge:
 ; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP8]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[SUM_0_LCSSA]]

 define i32 @reduction_mul(i32* noalias nocapture %A, i32* noalias nocapture %B) {
 ; CHECK-LABEL: @reduction_mul(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ <i32 19, i32 1, i32 1, i32 1>, [[VECTOR_PH]] ], [ [[TMP5:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 19, [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
-; CHECK-NEXT:    [[TMP4:%.*]] = mul <4 x i32> [[VEC_PHI]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP5]] = mul <4 x i32> [[TMP4]], [[WIDE_LOAD1]]
+; CHECK-NEXT:    [[TMP4:%.*]] = call i32 @llvm.experimental.vector.reduce.mul.v4i32(<4 x i32> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[TMP5:%.*]] = mul i32 [[TMP4]], [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.experimental.vector.reduce.mul.v4i32(<4 x i32> [[WIDE_LOAD1]])
+; CHECK-NEXT:    [[TMP7]] = mul i32 [[TMP6]], [[TMP5]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP6:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP6]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !12
+; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !12
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP7:%.*]] = call i32 @llvm.experimental.vector.reduce.mul.v4i32(<4 x i32> [[TMP5]])
 ; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
 ; CHECK:       .lr.ph:
 ; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], !llvm.loop !13
 ; CHECK:       ._crit_edge:
 ; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP7]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[SUM_0_LCSSA]]

 define i32 @start_at_non_zero(i32* nocapture %in, i32* nocapture %coeff, i32* nocapture %out) {
 ; CHECK-LABEL: @start_at_non_zero(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ <i32 120, i32 0, i32 0, i32 0>, [[VECTOR_PH]] ], [ [[TMP5:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 120, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[IN:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[COEFF:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
 ; CHECK-NEXT:    [[TMP4:%.*]] = mul nsw <4 x i32> [[WIDE_LOAD1]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP5]] = add <4 x i32> [[TMP4]], [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP5:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[TMP4]])
+; CHECK-NEXT:    [[TMP6]] = add i32 [[TMP5]], [[VEC_PHI]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP6:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP6]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !14
+; CHECK-NEXT:    [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !14
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP7:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[TMP5]])
 ; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], !llvm.loop !15
 ; CHECK:       for.end:
-; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[TMP7]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[TMP6]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[SUM_0_LCSSA]]
 ;
 entry:
   br label %for.body
 
 for.body:                                         ; preds = %entry, %for.body
   %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
   %sum.09 = phi i32 [ %add, %for.body ], [ 120, %entry ]

 define i32 @reduction_and(i32* nocapture %A, i32* nocapture %B) {
 ; CHECK-LABEL: @reduction_and(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ <i32 -1, i32 -1, i32 -1, i32 -1>, [[VECTOR_PH]] ], [ [[TMP5:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ -1, [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
-; CHECK-NEXT:    [[TMP4:%.*]] = and <4 x i32> [[VEC_PHI]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP5]] = and <4 x i32> [[TMP4]], [[WIDE_LOAD1]]
+; CHECK-NEXT:    [[TMP4:%.*]] = call i32 @llvm.experimental.vector.reduce.and.v4i32(<4 x i32> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[TMP5:%.*]] = and i32 [[TMP4]], [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.experimental.vector.reduce.and.v4i32(<4 x i32> [[WIDE_LOAD1]])
+; CHECK-NEXT:    [[TMP7]] = and i32 [[TMP6]], [[TMP5]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP6:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP6]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !16
+; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !16
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP7:%.*]] = call i32 @llvm.experimental.vector.reduce.and.v4i32(<4 x i32> [[TMP5]])
 ; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], !llvm.loop !17
 ; CHECK:       for.end:
 ; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[TMP7]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[RESULT_0_LCSSA]]

 define i32 @reduction_or(i32* nocapture %A, i32* nocapture %B) {
 ; CHECK-LABEL: @reduction_or(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP5:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
 ; CHECK-NEXT:    [[TMP4:%.*]] = add nsw <4 x i32> [[WIDE_LOAD1]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP5]] = or <4 x i32> [[TMP4]], [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP5:%.*]] = call i32 @llvm.experimental.vector.reduce.or.v4i32(<4 x i32> [[TMP4]])
+; CHECK-NEXT:    [[TMP6]] = or i32 [[TMP5]], [[VEC_PHI]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP6:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP6]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !18
+; CHECK-NEXT:    [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !18
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP7:%.*]] = call i32 @llvm.experimental.vector.reduce.or.v4i32(<4 x i32> [[TMP5]])
 ; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], !llvm.loop !19
 ; CHECK:       for.end:
-; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[TMP7]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[TMP6]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[RESULT_0_LCSSA]]
 ;
 entry:
   br label %for.body
 
 for.body:                                         ; preds = %entry, %for.body
   %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
   %result.08 = phi i32 [ %or, %for.body ], [ 0, %entry ]

 define i32 @reduction_xor(i32* nocapture %A, i32* nocapture %B) {
 ; CHECK-LABEL: @reduction_xor(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP5:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
 ; CHECK-NEXT:    [[TMP4:%.*]] = add nsw <4 x i32> [[WIDE_LOAD1]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP5]] = xor <4 x i32> [[TMP4]], [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP5:%.*]] = call i32 @llvm.experimental.vector.reduce.xor.v4i32(<4 x i32> [[TMP4]])
+; CHECK-NEXT:    [[TMP6]] = xor i32 [[TMP5]], [[VEC_PHI]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP6:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP6]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !20
+; CHECK-NEXT:    [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !20
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP7:%.*]] = call i32 @llvm.experimental.vector.reduce.xor.v4i32(<4 x i32> [[TMP5]])
 ; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], !llvm.loop !21
 ; CHECK:       for.end:
-; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[TMP7]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[TMP6]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[RESULT_0_LCSSA]]
 ;
 entry:
   br label %for.body
 
 for.body:                                         ; preds = %entry, %for.body
   %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
   %result.08 = phi i32 [ %xor, %for.body ], [ 0, %entry ]

 define float @reduction_fadd(float* nocapture %A, float* nocapture %B) {
 ; CHECK-LABEL: @reduction_fadd(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x float> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP5:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi float [ 0.000000e+00, [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds float, float* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast float* [[TMP0]] to <4 x float>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, <4 x float>* [[TMP1]], align 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds float, float* [[B:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float* [[TMP2]] to <4 x float>*
 ; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x float>, <4 x float>* [[TMP3]], align 4
-; CHECK-NEXT:    [[TMP4:%.*]] = fadd fast <4 x float> [[VEC_PHI]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP5]] = fadd fast <4 x float> [[TMP4]], [[WIDE_LOAD1]]
+; CHECK-NEXT:    [[TMP4:%.*]] = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float 0.000000e+00, <4 x float> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[TMP5:%.*]] = fadd float [[TMP4]], [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP6:%.*]] = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float 0.000000e+00, <4 x float> [[WIDE_LOAD1]])
+; CHECK-NEXT:    [[TMP7]] = fadd float [[TMP6]], [[TMP5]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP6:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP6]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !22
+; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !22
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP7:%.*]] = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float 0.000000e+00, <4 x float> [[TMP5]])
 ; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], !llvm.loop !23
 ; CHECK:       for.end:
 ; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi float [ undef, [[FOR_BODY]] ], [ [[TMP7]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret float [[RESULT_0_LCSSA]]

 define float @reduction_fmul(float* nocapture %A, float* nocapture %B) {
 ; CHECK-LABEL: @reduction_fmul(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x float> [ <float 0.000000e+00, float 1.000000e+00, float 1.000000e+00, float 1.000000e+00>, [[VECTOR_PH]] ], [ [[TMP5:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi float [ 0.000000e+00, [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds float, float* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast float* [[TMP0]] to <4 x float>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, <4 x float>* [[TMP1]], align 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds float, float* [[B:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float* [[TMP2]] to <4 x float>*
 ; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x float>, <4 x float>* [[TMP3]], align 4
-; CHECK-NEXT:    [[TMP4:%.*]] = fmul fast <4 x float> [[VEC_PHI]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP5]] = fmul fast <4 x float> [[TMP4]], [[WIDE_LOAD1]]
+; CHECK-NEXT:    [[TMP4:%.*]] = call fast float @llvm.experimental.vector.reduce.v2.fmul.f32.v4f32(float 1.000000e+00, <4 x float> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[TMP5:%.*]] = fmul float [[TMP4]], [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP6:%.*]] = call fast float @llvm.experimental.vector.reduce.v2.fmul.f32.v4f32(float 1.000000e+00, <4 x float> [[WIDE_LOAD1]])
+; CHECK-NEXT:    [[TMP7]] = fmul float [[TMP6]], [[TMP5]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP6:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP6]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !24
+; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !24
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP7:%.*]] = call fast float @llvm.experimental.vector.reduce.v2.fmul.f32.v4f32(float 1.000000e+00, <4 x float> [[TMP5]])
 ; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], !llvm.loop !25
 ; CHECK:       for.end:
 ; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi float [ undef, [[FOR_BODY]] ], [ [[TMP7]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret float [[RESULT_0_LCSSA]]

 define i32 @reduction_min(i32* nocapture %A, i32* nocapture %B) {
 ; CHECK-LABEL: @reduction_min(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ <i32 1000, i32 1000, i32 1000, i32 1000>, [[VECTOR_PH]] ], [ [[TMP3:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 1000, [[VECTOR_PH]] ], [ [[RDX_MINMAX_SELECT:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
-; CHECK-NEXT:    [[TMP2:%.*]] = icmp slt <4 x i32> [[VEC_PHI]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP3]] = select <4 x i1> [[TMP2]], <4 x i32> [[VEC_PHI]], <4 x i32> [[WIDE_LOAD]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call i32 @llvm.experimental.vector.reduce.smin.v4i32(<4 x i32> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[RDX_MINMAX_CMP:%.*]] = icmp slt i32 [[TMP2]], [[VEC_PHI]]
+; CHECK-NEXT:    [[RDX_MINMAX_SELECT]] = select i1 [[RDX_MINMAX_CMP]], i32 [[TMP2]], i32 [[VEC_PHI]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP4]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !26
+; CHECK-NEXT:    [[TMP3:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP3]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !26
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP5:%.*]] = call i32 @llvm.experimental.vector.reduce.smin.v4i32(<4 x i32> [[TMP3]])
 ; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], !llvm.loop !27
 ; CHECK:       for.end:
-; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[TMP5]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[RDX_MINMAX_SELECT]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[RESULT_0_LCSSA]]
 ;
 entry:
   br label %for.body
 
 for.body:                                         ; preds = %entry, %for.body
   %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
   %result.08 = phi i32 [ %v0, %for.body ], [ 1000, %entry ]

 define i32 @reduction_max(i32* nocapture %A, i32* nocapture %B) {
 ; CHECK-LABEL: @reduction_max(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ <i32 1000, i32 1000, i32 1000, i32 1000>, [[VECTOR_PH]] ], [ [[TMP3:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 1000, [[VECTOR_PH]] ], [ [[RDX_MINMAX_SELECT:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
-; CHECK-NEXT:    [[TMP2:%.*]] = icmp ugt <4 x i32> [[VEC_PHI]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP3]] = select <4 x i1> [[TMP2]], <4 x i32> [[VEC_PHI]], <4 x i32> [[WIDE_LOAD]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call i32 @llvm.experimental.vector.reduce.umax.v4i32(<4 x i32> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[RDX_MINMAX_CMP:%.*]] = icmp ugt i32 [[TMP2]], [[VEC_PHI]]
+; CHECK-NEXT:    [[RDX_MINMAX_SELECT]] = select i1 [[RDX_MINMAX_CMP]], i32 [[TMP2]], i32 [[VEC_PHI]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP4]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !28
+; CHECK-NEXT:    [[TMP3:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP3]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !28
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP5:%.*]] = call i32 @llvm.experimental.vector.reduce.umax.v4i32(<4 x i32> [[TMP3]])
 ; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], !llvm.loop !29
 ; CHECK:       for.end:
-; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[TMP5]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[RDX_MINMAX_SELECT]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[RESULT_0_LCSSA]]
 ;
 entry:
   br label %for.body
 
 for.body:                                         ; preds = %entry, %for.body
   %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
   %result.08 = phi i32 [ %v0, %for.body ], [ 1000, %entry ]

 define i32 @reduction_sum_multiuse(i32* noalias nocapture %A, i32* noalias nocapture %B) {
 ; CHECK-LABEL: @reduction_sum_multiuse(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[VEC_IND2:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT3:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
-; CHECK-NEXT:    [[TMP4:%.*]] = add <4 x i32> [[VEC_PHI]], [[VEC_IND2]]
-; CHECK-NEXT:    [[TMP5:%.*]] = add <4 x i32> [[TMP4]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP6]] = add <4 x i32> [[TMP5]], [[WIDE_LOAD1]]
+; CHECK-NEXT:    [[TMP4:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[VEC_IND2]])
+; CHECK-NEXT:    [[TMP5:%.*]] = add i32 [[TMP4]], [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[TMP7:%.*]] = add i32 [[TMP6]], [[TMP5]]
+; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD1]])
+; CHECK-NEXT:    [[TMP9]] = add i32 [[TMP8]], [[TMP7]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
 ; CHECK-NEXT:    [[VEC_IND_NEXT3]] = add <4 x i32> [[VEC_IND2]], <i32 4, i32 4, i32 4, i32 4>
-; CHECK-NEXT:    [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !34
+; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !34
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[TMP6]])
 ; CHECK-NEXT:    br i1 true, label [[END:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
 ; CHECK:       .lr.ph:
 ; CHECK-NEXT:    br i1 undef, label [[END]], label [[DOTLR_PH]], !llvm.loop !35
 ; CHECK:       end:
-; CHECK-NEXT:    [[F1:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP8]], [[MIDDLE_BLOCK]] ]
-; CHECK-NEXT:    [[F2:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP8]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[F1:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP9]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[F2:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP9]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    [[FINAL:%.*]] = add i32 [[F1]], [[F2]]
 ; CHECK-NEXT:    ret i32 [[FINAL]]
 ;
 entry:
   br label %.lr.ph
 
 .lr.ph:                                           ; preds = %entry, %.lr.ph
   %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]

 define i32 @reduction_predicated(i32* noalias nocapture %A, i32* noalias nocapture %B) {
 ; CHECK-LABEL: @reduction_predicated(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[VEC_IND2:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT3:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
-; CHECK-NEXT:    [[TMP4:%.*]] = add <4 x i32> [[VEC_PHI]], [[VEC_IND2]]
-; CHECK-NEXT:    [[TMP5:%.*]] = add <4 x i32> [[TMP4]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP6]] = add <4 x i32> [[TMP5]], [[WIDE_LOAD1]]
+; CHECK-NEXT:    [[TMP4:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[VEC_IND2]])
+; CHECK-NEXT:    [[TMP5:%.*]] = add i32 [[TMP4]], [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[TMP7:%.*]] = add i32 [[TMP6]], [[TMP5]]
+; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD1]])
+; CHECK-NEXT:    [[TMP9]] = add i32 [[TMP8]], [[TMP7]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
 ; CHECK-NEXT:    [[VEC_IND_NEXT3]] = add <4 x i32> [[VEC_IND2]], <i32 4, i32 4, i32 4, i32 4>
-; CHECK-NEXT:    [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !36
+; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !36
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> [[TMP6]])
 ; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
 ; CHECK:       .lr.ph:
 ; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], !llvm.loop !37
 ; CHECK:       ._crit_edge:
-; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP8]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP9]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[SUM_0_LCSSA]]
 ;
 entry:
   br label %.lr.ph
 
 .lr.ph:                                           ; preds = %entry, %.lr.ph
   %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]
   %sum.02 = phi i32 [ %l9, %.lr.ph ], [ 0, %entry ]

 define i8 @reduction_and_trunc(i8* noalias nocapture %A) {
 ; CHECK-LABEL: @reduction_and_trunc(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ <i32 255, i32 -1, i32 -1, i32 -1>, [[VECTOR_PH]] ], [ [[TMP4:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[TMP0:%.*]] = sext i32 [[INDEX]] to i64
-; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i8, i8* [[A:%.*]], i64 [[TMP0]]
-; CHECK-NEXT:    [[TMP2:%.*]] = bitcast i8* [[TMP1]] to <4 x i8>*
-; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i8>, <4 x i8>* [[TMP2]], align 4
-; CHECK-NEXT:    [[TMP3:%.*]] = zext <4 x i8> [[WIDE_LOAD]] to <4 x i32>
-; CHECK-NEXT:    [[TMP4]] = and <4 x i32> [[VEC_PHI]], [[TMP3]]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 255, [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[TMP0:%.*]] = call i32 @llvm.experimental.vector.reduce.and.v4i32(<4 x i32> <i32 255, i32 255, i32 255, i32 255>)
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[TMP0]], [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP2:%.*]] = sext i32 [[INDEX]] to i64
+; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i8, i8* [[A:%.*]], i64 [[TMP2]]
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast i8* [[TMP3]] to <4 x i8>*
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i8>, <4 x i8>* [[TMP4]], align 4
+; CHECK-NEXT:    [[TMP5:%.*]] = zext <4 x i8> [[WIDE_LOAD]] to <4 x i32>
+; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.experimental.vector.reduce.and.v4i32(<4 x i32> [[TMP5]])
+; CHECK-NEXT:    [[TMP7]] = and i32 [[TMP6]], [[TMP1]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i32 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i32 [[INDEX_NEXT]], 256
-; CHECK-NEXT:    br i1 [[TMP5]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !40
+; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i32 [[INDEX_NEXT]], 256
+; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !40
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.experimental.vector.reduce.and.v4i32(<4 x i32> [[TMP4]])
 ; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
 ; CHECK:       .lr.ph:
 ; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], !llvm.loop !41
 ; CHECK:       ._crit_edge:
-; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP6]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP7]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    [[RET:%.*]] = trunc i32 [[SUM_0_LCSSA]] to i8
 ; CHECK-NEXT:    ret i8 [[RET]]
 ;
 entry:
   br label %.lr.ph
 
 .lr.ph:                                           ; preds = %entry, %.lr.ph
   %indvars.iv = phi i32 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]