[LoopVectorize] Add support for tail folding using scalable vectors

[david-arm]

Summary

This patch fixes up an issue with InnerLoopVectorizer::getOrCreateVectorTripCount
whereby we weren't correctly generating the runtime trip count
for scalable vectors when tail-folding.

It also removes some asserts in the tail-folding path for cases when
the VF is not scalable.

In this patch I have only permitted tail-folding to be enabled
explicitly for scalable vectors when the user has specified one
of the following flags:

-prefer-predicate-over-epilogue=predicate-dont-vectorize
-prefer-predicate-over-epilogue=predicate-else-scalar-epilogue

For now it's best not to enable tail-folding with scalable vectors for
low trip counts or when optimising for code size, since there has been
no analysis on whether this is worth it.

Various tests have been added here:

Transforms/LoopVectorize/AArch64/sve-tail-folding.ll
Transforms/LoopVectorize/AArch64/sve-tail-folding-forced.ll

The tests cannot be target independent because they require masked
load/store support, i.e. TTI.isLegalMaskedLoad and TTI.isLegalMaskedStore
need to return true.

[SjoerdMeijer]

Since we are (also) talking about profitability of the transform:

For now it's best not to enable tail-folding with scalable vectors for low trip counts or when optimising for code size, since there has been no analysis on whether this is worth it.

I was expecting we would need to implement target hook preferPredicateOverEpilogue for AArch64/SVE that controls this, and add some of this logic there?

[david-arm]

In D113003#3102731, @SjoerdMeijer wrote:

Since we are (also) talking about profitability of the transform:

For now it's best not to enable tail-folding with scalable vectors for low trip counts or when optimising for code size, since there has been no analysis on whether this is worth it.

I was expecting we would need to implement target hook preferPredicateOverEpilogue for AArch64/SVE that controls this, and add some of this logic there?

Hi @SjoerdMeijer, at the moment we're not really worried about profitability since we don't intend to enable tail-folding by default anytime soon. I'm happy to look into the TTI hook (preferPredicateOverEpilogue), but I'm not sure what benefit it gives us at the moment? The main reason for supporting tail-folding for scalable vectors for AArch64 is to ensure we have the technical capability and nothing crashes with scalable vectors. In future we'd like to use predication for vector epilogues and avoid the scalar tail, but maintain an unpredicated main vector loop.

[david-arm]

Hi @SjoerdMeijer, I realised the title on the patch was misleading and I've changed it now to indicate we're not enabling this for AArch64, but just adding theoretical support for scalable vectors!

[SjoerdMeijer]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5487

Hi @david-arm , thanks for clarifying so far, that has helped. Moving the question about profitability to here, the place where it is probably relevant. Also note that I am asking a bunch of question just to reload to memory how things work here (and I am new to the SVE/scalable angle here).

Before we discuss the actual change, I don't think I understand this original comment + code here to be honest. The first sentence:

For scalable vectors, don't use tail folding as this is currently not yet supported.

suggest to me that we want to return CM_ScalarEpilogueAllowed in getScalarEpilogueLowering. And I don't know what the affect is of setting MaxFactors.ScalableVF to 0 below, but I guess that has somehow the affect?

To me it feels like that we do make a profitability call here "in the middle of something", which should be moved to a different place. I can also imagine that this is highly target dependent/specific, further supporting this.

I do remember though that there is a bit of an ordering problem: the decision to tail-fold or not is taken quite early, while not all information that we ideally would like to have are not yet available. Not sure that is the case here. But I guess that for now the decision "if scalable vector, then don't tail-fold" could be taken quite early. But yeah, I might be missing something here, so please enlighten me. :)

[david-arm]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5487

So when this code was originally added tail-folding just crashed for scalable vectors because a lot of the code related to tail-folding assumed fixed-width vectors. We've been trying to fix a lot of these issues and now we've got to the point where it is at least possible to fold the tail for most loops. However, it is still experimental for scalable vectors - scalable vectorisation has not yet been enabled by default for the vectoriser. Even after this patch lands there is still at least one known issue that we intend to fix related to tail-folding + scalable vectors. This is the reason why tail-folding is still hidden behind a flag for scalable vectors. If we start to allow tail-folding to be automatically enabled for low trip counts or when optimising for code size then we will be knowingly exposing users to these issues.

I do take your point about profitability and realise the original comment also suggested something related to profitability, but the aim of this patch is simply to add functionality that is not enabled by default. Perhaps I can change the comment to explain that we want to make this feature stable first for scalable vectors, then move on to questions of profitability?

To answer your other question about MaxFactors.ScalableVF it basically disables scalable vectorisation entirely if computeMaxVF has decided that we should use tail-folding for low trip counts or reducing code size. However, we are still be able to use fixed-width vectorisation.

[SjoerdMeijer]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5487

I do take your point about profitability and realise the original comment also suggested something related to profitability, but the aim of this patch is simply to add functionality that is not enabled by default. Perhaps I can change the comment to explain that we want to make this feature stable first for scalable vectors, then move on to questions of profitability?

Sound like a good approach to me.

[david-arm]

• Rebase, which required fixing up some tests.

[david-arm]

• Rebase, which required tweaking the tests slightly due to cost model changes.

[sdesmalen]

Seems like quite the straightforward change, LGTM.

llvm/test/Transforms/LoopVectorize/AArch64/sve-tail-folding-forced.ll

2

nit: this can be removed now that scalable vectorization is enabled by default for aarch64 when SVE is available.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

auto CreateScalarIV = [&](Value *&Step) -> Value * {
     return ScalarIV;
   };
 
   // Create the vector values from the scalar IV, in the absence of creating a
   // vector IV.
   auto CreateSplatIV = [&](Value *ScalarIV, Value *Step) {
     Value *Broadcasted = getBroadcastInstrs(ScalarIV);
     for (unsigned Part = 0; Part < UF; ++Part) {
-      assert(!State.VF.isScalable() && "scalable vectors not yet supported.");
       Value *StartIdx;
       if (Step->getType()->isFloatingPointTy())
         StartIdx =
             getRuntimeVFAsFloat(Builder, Step->getType(), State.VF * Part);
       else
         StartIdx = getRuntimeVF(Builder, Step->getType(), State.VF * Part);
 
       Value *EntryPart =
Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) {
   // up to a multiple of Step instead of rounding down. This is done by first
   // adding Step-1 and then rounding down. Note that it's ok if this addition
   // overflows: the vector induction variable will eventually wrap to zero given
   // that it starts at zero and its Step is a power of two; the loop will then
   // exit, with the last early-exit vector comparison also producing all-true.
   if (Cost->foldTailByMasking()) {
     assert(isPowerOf2_32(VF.getKnownMinValue() * UF) &&
            "VF*UF must be a power of 2 when folding tail by masking");
-    assert(!VF.isScalable() &&
-           "Tail folding not yet supported for scalable vectors");
+    Value *NumLanes = getRuntimeVF(Builder, Ty, VF * UF);
     TC = Builder.CreateAdd(
-        TC, ConstantInt::get(Ty, VF.getKnownMinValue() * UF - 1), "n.rnd.up");
+        TC, Builder.CreateSub(NumLanes, ConstantInt::get(Ty, 1)), "n.rnd.up");
   }
 
   // Now we need to generate the expression for the part of the loop that the
   // vectorized body will execute. This is equal to N - (N % Step) if scalar
   // iterations are not required for correctness, or N - Step, otherwise. Step
   // is equal to the vectorization factor (number of SIMD elements) times the
   // unroll factor (number of SIMD instructions).
   Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
const SCEV *Rem = SE->getURemExpr(
         SE->getConstant(BackedgeTakenCount->getType(), MaxVFtimesIC));
     if (Rem->isZero()) {
       // Accept MaxFixedVF if we do not have a tail.
       LLVM_DEBUG(dbgs() << "LV: No tail will remain for any chosen VF.\n");
       return MaxFactors;
     }
   }
 
-  // For scalable vectors, don't use tail folding as this is currently not yet
-  // supported. The code is likely to have ended up here if the tripcount is
-  // low, in which case it makes sense not to use scalable vectors.
-  if (MaxFactors.ScalableVF.isVector())
+  // For scalable vectors don't use tail folding for low trip counts or
+  // optimizing for code size. We only permit this if the user has explicitly
+  // requested it.
+  if (ScalarEpilogueStatus != CM_ScalarEpilogueNotNeededUsePredicate &&
+      ScalarEpilogueStatus != CM_ScalarEpilogueNotAllowedUsePredicate &&
+      MaxFactors.ScalableVF.isVector())
     MaxFactors.ScalableVF = ElementCount::getScalable(0);
 
   // If we don't know the precise trip count, or if the trip count that we
   // found modulo the vectorization factor is not zero, try to fold the tail
   // by masking.
   // FIXME: look for a smaller MaxVF that does divide TC rather than masking.
   if (Legal->prepareToFoldTailByMasking()) {
     FoldTailByMasking = true;

llvm/test/Transforms/LoopVectorize/AArch64/sve-tail-folding-forced.ll

+; RUN: opt -S -loop-vectorize < %s | FileCheck %s
+
+; These tests ensure that tail-folding is enabled when the predicate.enable
+; loop attribute is set to true.
+
+target triple = "aarch64-unknown-linux-gnu"
+
+
+define void @simple_memset(i32 %val, i32* %ptr, i64 %n) #0 {
+; CHECK-LABEL: @simple_memset(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 [[N:%.*]], i64 1)
+; CHECK-NEXT:    br i1 false, label %scalar.ph, label %vector.ph
+; CHECK:       vector.ph:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP1:%.*]] = mul i64 [[TMP0]], 4
+; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP3:%.*]] = mul i64 [[TMP2]], 4
+; CHECK-NEXT:    [[TMP4:%.*]] = sub i64 [[TMP3]], 1
+; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i64 [[UMAX]], [[TMP4]]
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP1]]
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
+; CHECK-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i64 [[UMAX]], 1
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TRIP_COUNT_MINUS_1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT5:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[VAL:%.*]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT6:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT5]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    br label %vector.body
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX1:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT2:%.*]], %vector.body ]
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[INDEX1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT4:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT3]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP5:%.*]] = call <vscale x 4 x i64> @llvm.experimental.stepvector.nxv4i64()
+; CHECK-NEXT:    [[TMP6:%.*]] = add <vscale x 4 x i64> [[TMP5]], zeroinitializer
+; CHECK-NEXT:    [[TMP7:%.*]] = mul <vscale x 4 x i64> [[TMP6]], shufflevector (<vscale x 4 x i64> insertelement (<vscale x 4 x i64> poison, i64 1, i32 0), <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer)
+; CHECK-NEXT:    [[INDUCTION:%.*]] = add <vscale x 4 x i64> [[BROADCAST_SPLAT4]], [[TMP7]]
+; CHECK-NEXT:    [[TMP8:%.*]] = add i64 [[INDEX1]], 0
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp ule <vscale x 4 x i64> [[INDUCTION]], [[BROADCAST_SPLAT]]
+; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr i32, i32* [[PTR:%.*]], i64 [[TMP8]]
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr i32, i32* [[TMP10]], i32 0
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast i32* [[TMP11]] to <vscale x 4 x i32>*
+; CHECK-NEXT:    call void @llvm.masked.store.nxv4i32.p0nxv4i32(<vscale x 4 x i32> [[BROADCAST_SPLAT6]], <vscale x 4 x i32>* [[TMP12]], i32 4, <vscale x 4 x i1> [[TMP9]])
+; CHECK-NEXT:    [[TMP13:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP14:%.*]] = mul i64 [[TMP13]], 4
+; CHECK-NEXT:    [[INDEX_NEXT2]] = add i64 [[INDEX1]], [[TMP14]]
+; CHECK-NEXT:    [[TMP15:%.*]] = icmp eq i64 [[INDEX_NEXT2]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP15]], label %middle.block, label %vector.body
+; CHECK:       middle.block:
+; CHECK-NEXT:    br i1 true, label %while.end.loopexit, label %scalar.ph
+;
+entry:
+  br label %while.body
+
+while.body:                                       ; preds = %while.body, %entry
+  %index = phi i64 [ %index.next, %while.body ], [ 0, %entry ]
+  %gep = getelementptr i32, i32* %ptr, i64 %index
+  store i32 %val, i32* %gep
+  %index.next = add nsw i64 %index, 1
+  %cmp10 = icmp ult i64 %index.next, %n
+  br i1 %cmp10, label %while.body, label %while.end.loopexit, !llvm.loop !0
+
+while.end.loopexit:                               ; preds = %while.body
+  ret void
+}
+
+
+attributes #0 = { "target-features"="+sve" }
+
+!0 = distinct !{!0, !1}
+!1 = !{!"llvm.loop.vectorize.predicate.enable", i1 true}

llvm/test/Transforms/LoopVectorize/AArch64/sve-tail-folding.ll

-; RUN: opt -S -loop-vectorize -prefer-predicate-over-epilogue=predicate-dont-vectorize < %s | FileCheck %s
+; RUN: opt -S -hints-allow-reordering=false -loop-vectorize -prefer-predicate-over-epilogue=predicate-dont-vectorize < %s | FileCheck %s
+; RUN: opt -S -hints-allow-reordering=false -loop-vectorize -prefer-predicate-over-epilogue=predicate-else-scalar-epilogue < %s | FileCheck %s
 
 ; CHECK-NOT: vector.body:
 
 target triple = "aarch64-unknown-linux-gnu"
 
-define void @tail_predication(i32 %init, i32* %ptr, i32 %val) #0 {
+
+define void @simple_memset(i32 %val, i32* %ptr, i64 %n) #0 {
+; CHECK-LABEL: @simple_memset(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 [[N:%.*]], i64 1)
+; CHECK-NEXT:    br i1 false, label %scalar.ph, label %vector.ph
+; CHECK:       vector.ph:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP1:%.*]] = mul i64 [[TMP0]], 4
+; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP3:%.*]] = mul i64 [[TMP2]], 4
+; CHECK-NEXT:    [[TMP4:%.*]] = sub i64 [[TMP3]], 1
+; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i64 [[UMAX]], [[TMP4]]
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP1]]
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
+; CHECK-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i64 [[UMAX]], 1
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TRIP_COUNT_MINUS_1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT5:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[VAL:%.*]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT6:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT5]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    br label %vector.body
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX1:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT2:%.*]], %vector.body ]
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[INDEX1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT4:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT3]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP5:%.*]] = call <vscale x 4 x i64> @llvm.experimental.stepvector.nxv4i64()
+; CHECK-NEXT:    [[TMP6:%.*]] = add <vscale x 4 x i64> [[TMP5]], zeroinitializer
+; CHECK-NEXT:    [[TMP7:%.*]] = mul <vscale x 4 x i64> [[TMP6]], shufflevector (<vscale x 4 x i64> insertelement (<vscale x 4 x i64> poison, i64 1, i32 0), <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer)
+; CHECK-NEXT:    [[INDUCTION:%.*]] = add <vscale x 4 x i64> [[BROADCAST_SPLAT4]], [[TMP7]]
+; CHECK-NEXT:    [[TMP8:%.*]] = add i64 [[INDEX1]], 0
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp ule <vscale x 4 x i64> [[INDUCTION]], [[BROADCAST_SPLAT]]
+; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr i32, i32* [[PTR:%.*]], i64 [[TMP8]]
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr i32, i32* [[TMP10]], i32 0
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast i32* [[TMP11]] to <vscale x 4 x i32>*
+; CHECK-NEXT:    call void @llvm.masked.store.nxv4i32.p0nxv4i32(<vscale x 4 x i32> [[BROADCAST_SPLAT6]], <vscale x 4 x i32>* [[TMP12]], i32 4, <vscale x 4 x i1> [[TMP9]])
+; CHECK-NEXT:    [[TMP13:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP14:%.*]] = mul i64 [[TMP13]], 4
+; CHECK-NEXT:    [[INDEX_NEXT2]] = add i64 [[INDEX1]], [[TMP14]]
+; CHECK-NEXT:    [[TMP15:%.*]] = icmp eq i64 [[INDEX_NEXT2]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP15]], label %middle.block, label %vector.body
+; CHECK:       middle.block:
+; CHECK-NEXT:    br i1 true, label %while.end.loopexit, label %scalar.ph
+;
 entry:
   br label %while.body
 
 while.body:                                       ; preds = %while.body, %entry
-  %index = phi i32 [ %index.dec, %while.body ], [ %init, %entry ]
-  %gep = getelementptr i32, i32* %ptr, i32 %index
+  %index = phi i64 [ %index.next, %while.body ], [ 0, %entry ]
+  %gep = getelementptr i32, i32* %ptr, i64 %index
   store i32 %val, i32* %gep
-  %index.dec = add nsw i32 %index, -1
-  %cmp10 = icmp sgt i32 %index, 0
-  br i1 %cmp10, label %while.body, label %while.end.loopexit
+  %index.next = add nsw i64 %index, 1
+  %cmp10 = icmp ult i64 %index.next, %n
+  br i1 %cmp10, label %while.body, label %while.end.loopexit, !llvm.loop !0
+
+while.end.loopexit:                               ; preds = %while.body
+  ret void
+}
+
+
+define void @simple_memcpy(i32* noalias %dst, i32* noalias %src, i64 %n) #0 {
+; CHECK-LABEL: @simple_memcpy(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 [[N:%.*]], i64 1)
+; CHECK-NEXT:    br i1 false, label %scalar.ph, label %vector.ph
+; CHECK:       vector.ph:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP1:%.*]] = mul i64 [[TMP0]], 4
+; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP3:%.*]] = mul i64 [[TMP2]], 4
+; CHECK-NEXT:    [[TMP4:%.*]] = sub i64 [[TMP3]], 1
+; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i64 [[UMAX]], [[TMP4]]
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP1]]
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
+; CHECK-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i64 [[UMAX]], 1
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TRIP_COUNT_MINUS_1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    br label %vector.body
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX1:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT2:%.*]], %vector.body ]
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[INDEX1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT4:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT3]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP5:%.*]] = call <vscale x 4 x i64> @llvm.experimental.stepvector.nxv4i64()
+; CHECK-NEXT:    [[TMP6:%.*]] = add <vscale x 4 x i64> [[TMP5]], zeroinitializer
+; CHECK-NEXT:    [[TMP7:%.*]] = mul <vscale x 4 x i64> [[TMP6]], shufflevector (<vscale x 4 x i64> insertelement (<vscale x 4 x i64> poison, i64 1, i32 0), <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer)
+; CHECK-NEXT:    [[INDUCTION:%.*]] = add <vscale x 4 x i64> [[BROADCAST_SPLAT4]], [[TMP7]]
+; CHECK-NEXT:    [[TMP8:%.*]] = add i64 [[INDEX1]], 0
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp ule <vscale x 4 x i64> [[INDUCTION]], [[BROADCAST_SPLAT]]
+; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr i32, i32* [[SRC:%.*]], i64 [[TMP8]]
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr i32, i32* [[TMP10]], i32 0
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast i32* [[TMP11]] to <vscale x 4 x i32>*
+; CHECK-NEXT:    [[WIDE_MASKED_LOAD:%.*]] = call <vscale x 4 x i32> @llvm.masked.load.nxv4i32.p0nxv4i32(<vscale x 4 x i32>* [[TMP12]], i32 4, <vscale x 4 x i1> [[TMP9]], <vscale x 4 x i32> poison)
+; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr i32, i32* [[DST:%.*]], i64 [[TMP8]]
+; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr i32, i32* [[TMP13]], i32 0
+; CHECK-NEXT:    [[TMP15:%.*]] = bitcast i32* [[TMP14]] to <vscale x 4 x i32>*
+; CHECK-NEXT:    call void @llvm.masked.store.nxv4i32.p0nxv4i32(<vscale x 4 x i32> [[WIDE_MASKED_LOAD]], <vscale x 4 x i32>* [[TMP15]], i32 4, <vscale x 4 x i1> [[TMP9]])
+; CHECK-NEXT:    [[TMP16:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP17:%.*]] = mul i64 [[TMP16]], 4
+; CHECK-NEXT:    [[INDEX_NEXT2]] = add i64 [[INDEX1]], [[TMP17]]
+; CHECK-NEXT:    [[TMP18:%.*]] = icmp eq i64 [[INDEX_NEXT2]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP18]], label %middle.block, label %vector.body, !llvm.loop [[LOOP4:![0-9]+]]
+; CHECK:       middle.block:
+; CHECK-NEXT:    br i1 true, label %while.end.loopexit, label %scalar.ph
+;
+entry:
+  br label %while.body
+
+while.body:                                       ; preds = %while.body, %entry
+  %index = phi i64 [ %index.next, %while.body ], [ 0, %entry ]
+  %gep1 = getelementptr i32, i32* %src, i64 %index
+  %val = load i32, i32* %gep1
+  %gep2 = getelementptr i32, i32* %dst, i64 %index
+  store i32 %val, i32* %gep2
+  %index.next = add nsw i64 %index, 1
+  %cmp10 = icmp ult i64 %index.next, %n
+  br i1 %cmp10, label %while.body, label %while.end.loopexit, !llvm.loop !0
+
+while.end.loopexit:                               ; preds = %while.body
+  ret void
+}
+
+
+define void @simple_gather_scatter(i32* noalias %dst, i32* noalias %src, i32* noalias %ind, i64 %n) #0 {
+; CHECK-LABEL: @simple_gather_scatter(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 [[N:%.*]], i64 1)
+; CHECK-NEXT:    br i1 false, label %scalar.ph, label %vector.ph
+; CHECK:       vector.ph:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP1:%.*]] = mul i64 [[TMP0]], 4
+; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP3:%.*]] = mul i64 [[TMP2]], 4
+; CHECK-NEXT:    [[TMP4:%.*]] = sub i64 [[TMP3]], 1
+; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i64 [[UMAX]], [[TMP4]]
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP1]]
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
+; CHECK-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i64 [[UMAX]], 1
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TRIP_COUNT_MINUS_1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX1:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT2:%.*]], %vector.body ]
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[INDEX1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT4:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT3]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP5:%.*]] = call <vscale x 4 x i64> @llvm.experimental.stepvector.nxv4i64()
+; CHECK-NEXT:    [[TMP6:%.*]] = add <vscale x 4 x i64> [[TMP5]], zeroinitializer
+; CHECK-NEXT:    [[TMP7:%.*]] = mul <vscale x 4 x i64> [[TMP6]], shufflevector (<vscale x 4 x i64> insertelement (<vscale x 4 x i64> poison, i64 1, i32 0), <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer)
+; CHECK-NEXT:    [[INDUCTION:%.*]] = add <vscale x 4 x i64> [[BROADCAST_SPLAT4]], [[TMP7]]
+; CHECK-NEXT:    [[TMP8:%.*]] = add i64 [[INDEX1]], 0
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp ule <vscale x 4 x i64> [[INDUCTION]], [[BROADCAST_SPLAT]]
+; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr i32, i32* [[IND:%.*]], i64 [[TMP8]]
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr i32, i32* [[TMP10]], i32 0
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast i32* [[TMP11]] to <vscale x 4 x i32>*
+; CHECK-NEXT:    [[WIDE_MASKED_LOAD:%.*]] = call <vscale x 4 x i32> @llvm.masked.load.nxv4i32.p0nxv4i32(<vscale x 4 x i32>* [[TMP12]], i32 4, <vscale x 4 x i1> [[TMP9]], <vscale x 4 x i32> poison)
+; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr i32, i32* [[SRC:%.*]], <vscale x 4 x i32> [[WIDE_MASKED_LOAD]]
+; CHECK-NEXT:    [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0i32(<vscale x 4 x i32*> [[TMP13]], i32 4, <vscale x 4 x i1> [[TMP9]], <vscale x 4 x i32> undef)
+; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr i32, i32* [[DST:%.*]], <vscale x 4 x i32> [[WIDE_MASKED_LOAD]]
+; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i32.nxv4p0i32(<vscale x 4 x i32> [[WIDE_MASKED_GATHER]], <vscale x 4 x i32*> [[TMP14]], i32 4, <vscale x 4 x i1> [[TMP9]])
+; CHECK-NEXT:    [[TMP15:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP16:%.*]] = mul i64 [[TMP15]], 4
+; CHECK-NEXT:    [[INDEX_NEXT2]] = add i64 [[INDEX1]], [[TMP16]]
+; CHECK-NEXT:    [[TMP17:%.*]] = icmp eq i64 [[INDEX_NEXT2]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP17]], label %middle.block, label %vector.body
+; CHECK:       middle.block:
+; CHECK-NEXT:    br i1 true, label %while.end.loopexit, label %scalar.ph
+;
+entry:
+  br label %while.body
+
+while.body:                                       ; preds = %while.body, %entry
+  %index = phi i64 [ %index.next, %while.body ], [ 0, %entry ]
+  %gep1 = getelementptr i32, i32* %ind, i64 %index
+  %ind_val = load i32, i32* %gep1
+  %gep2 = getelementptr i32, i32* %src, i32 %ind_val
+  %val = load i32, i32* %gep2
+  %gep3 = getelementptr i32, i32* %dst, i32 %ind_val
+  store i32 %val, i32* %gep3
+  %index.next = add nsw i64 %index, 1
+  %cmp10 = icmp ult i64 %index.next, %n
+  br i1 %cmp10, label %while.body, label %while.end.loopexit, !llvm.loop !0
 
 while.end.loopexit:                               ; preds = %while.body
   ret void
 }
 
+
+; The original loop had an unconditional uniform load. Let's make sure
+; we don't artificially create new predicated blocks for the load.
+define void @uniform_load(i32* noalias %dst, i32* noalias readonly %src, i64 %n) #0 {
+; CHECK-LABEL: @uniform_load(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br i1 false, label %scalar.ph, label %vector.ph
+; CHECK:       vector.ph:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP1:%.*]] = mul i64 [[TMP0]], 4
+; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP3:%.*]] = mul i64 [[TMP2]], 4
+; CHECK-NEXT:    [[TMP4:%.*]] = sub i64 [[TMP3]], 1
+; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i64 [[N:%.*]], [[TMP4]]
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP1]]
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
+; CHECK-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i64 [[N]], 1
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TRIP_COUNT_MINUS_1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    br label %vector.body
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT:%.*]], %vector.body ]
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[INDEX]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT1]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP5:%.*]] = call <vscale x 4 x i64> @llvm.experimental.stepvector.nxv4i64()
+; CHECK-NEXT:    [[TMP6:%.*]] = add <vscale x 4 x i64> [[TMP5]], zeroinitializer
+; CHECK-NEXT:    [[TMP7:%.*]] = mul <vscale x 4 x i64> [[TMP6]], shufflevector (<vscale x 4 x i64> insertelement (<vscale x 4 x i64> poison, i64 1, i32 0), <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer)
+; CHECK-NEXT:    [[INDUCTION:%.*]] = add <vscale x 4 x i64> [[BROADCAST_SPLAT2]], [[TMP7]]
+; CHECK-NEXT:    [[TMP8:%.*]] = add i64 [[INDEX]], 0
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp ule <vscale x 4 x i64> [[INDUCTION]], [[BROADCAST_SPLAT]]
+; CHECK-NEXT:    [[TMP10:%.*]] = load i32, i32* [[SRC:%.*]], align 4
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[TMP10]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT4:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT3]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr inbounds i32, i32* [[DST:%.*]], i64 [[TMP8]]
+; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i32, i32* [[TMP11]], i32 0
+; CHECK-NEXT:    [[TMP13:%.*]] = bitcast i32* [[TMP12]] to <vscale x 4 x i32>*
+; CHECK-NEXT:    call void @llvm.masked.store.nxv4i32.p0nxv4i32(<vscale x 4 x i32> [[BROADCAST_SPLAT4]], <vscale x 4 x i32>* [[TMP13]], i32 4, <vscale x 4 x i1> [[TMP9]])
+; CHECK-NEXT:    [[TMP14:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP15:%.*]] = mul i64 [[TMP14]], 4
+; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], [[TMP15]]
+; CHECK-NEXT:    [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP16]], label %middle.block, label %vector.body, !llvm.loop [[LOOP6:![0-9]+]]
+; CHECK:       middle.block:
+; CHECK-NEXT:    br i1 true, label %for.end, label %scalar.ph
+;
+
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %entry, %for.body
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
+  %val = load i32, i32* %src, align 4
+  %arrayidx = getelementptr inbounds i32, i32* %dst, i64 %indvars.iv
+  store i32 %val, i32* %arrayidx, align 4
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond.not = icmp eq i64 %indvars.iv.next, %n
+  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:                                          ; preds = %for.body, %entry
+  ret void
+}
+
+
+; The original loop had a conditional uniform load. In this case we actually
+; do need to perform conditional loads and so we end up using a gather instead.
+; However, we at least ensure the mask is the overlap of the loop predicate
+; and the original condition.
+define void @cond_uniform_load(i32* noalias %dst, i32* noalias readonly %src, i32* noalias readonly %cond, i64 %n) #0 {
+; CHECK-LABEL: @cond_uniform_load(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br i1 false, label %scalar.ph, label %vector.ph
+; CHECK:       vector.ph:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP1:%.*]] = mul i64 [[TMP0]], 4
+; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP3:%.*]] = mul i64 [[TMP2]], 4
+; CHECK-NEXT:    [[TMP4:%.*]] = sub i64 [[TMP3]], 1
+; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i64 [[N:%.*]], [[TMP4]]
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP1]]
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
+; CHECK-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i64 [[N]], 1
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TRIP_COUNT_MINUS_1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT5:%.*]] = insertelement <vscale x 4 x i32*> poison, i32* [[SRC:%.*]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT6:%.*]] = shufflevector <vscale x 4 x i32*> [[BROADCAST_SPLATINSERT5]], <vscale x 4 x i32*> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    br label %vector.body
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX1:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT2:%.*]], %vector.body ]
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[INDEX1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT4:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT3]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP5:%.*]] = call <vscale x 4 x i64> @llvm.experimental.stepvector.nxv4i64()
+; CHECK-NEXT:    [[TMP6:%.*]] = add <vscale x 4 x i64> [[TMP5]], zeroinitializer
+; CHECK-NEXT:    [[TMP7:%.*]] = mul <vscale x 4 x i64> [[TMP6]], shufflevector (<vscale x 4 x i64> insertelement (<vscale x 4 x i64> poison, i64 1, i32 0), <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer)
+; CHECK-NEXT:    [[INDUCTION:%.*]] = add <vscale x 4 x i64> [[BROADCAST_SPLAT4]], [[TMP7]]
+; CHECK-NEXT:    [[TMP8:%.*]] = add i64 [[INDEX1]], 0
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp ule <vscale x 4 x i64> [[INDUCTION]], [[BROADCAST_SPLAT]]
+; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr inbounds i32, i32* [[COND:%.*]], i64 [[TMP8]]
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr inbounds i32, i32* [[TMP10]], i32 0
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast i32* [[TMP11]] to <vscale x 4 x i32>*
+; CHECK-NEXT:    [[WIDE_MASKED_LOAD:%.*]] = call <vscale x 4 x i32> @llvm.masked.load.nxv4i32.p0nxv4i32(<vscale x 4 x i32>* [[TMP12]], i32 4, <vscale x 4 x i1> [[TMP9]], <vscale x 4 x i32> poison)
+; CHECK-NEXT:    [[TMP13:%.*]] = icmp eq <vscale x 4 x i32> [[WIDE_MASKED_LOAD]], zeroinitializer
+; CHECK-NEXT:    [[TMP14:%.*]] = xor <vscale x 4 x i1> [[TMP13]], shufflevector (<vscale x 4 x i1> insertelement (<vscale x 4 x i1> poison, i1 true, i32 0), <vscale x 4 x i1> poison, <vscale x 4 x i32> zeroinitializer)
+; CHECK-NEXT:    [[TMP15:%.*]] = select <vscale x 4 x i1> [[TMP9]], <vscale x 4 x i1> [[TMP14]], <vscale x 4 x i1> zeroinitializer
+; CHECK-NEXT:    [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0i32(<vscale x 4 x i32*> [[BROADCAST_SPLAT6]], i32 4, <vscale x 4 x i1> [[TMP15]], <vscale x 4 x i32> undef)
+; CHECK-NEXT:    [[TMP16:%.*]] = select <vscale x 4 x i1> [[TMP9]], <vscale x 4 x i1> [[TMP13]], <vscale x 4 x i1> zeroinitializer
+; CHECK-NEXT:    [[PREDPHI:%.*]] = select <vscale x 4 x i1> [[TMP16]], <vscale x 4 x i32> zeroinitializer, <vscale x 4 x i32> [[WIDE_MASKED_GATHER]]
+; CHECK-NEXT:    [[TMP17:%.*]] = getelementptr inbounds i32, i32* [[DST:%.*]], i64 [[TMP8]]
+; CHECK-NEXT:    [[TMP18:%.*]] = or <vscale x 4 x i1> [[TMP15]], [[TMP16]]
+; CHECK-NEXT:    [[TMP19:%.*]] = getelementptr inbounds i32, i32* [[TMP17]], i32 0
+; CHECK-NEXT:    [[TMP20:%.*]] = bitcast i32* [[TMP19]] to <vscale x 4 x i32>*
+; CHECK-NEXT:    call void @llvm.masked.store.nxv4i32.p0nxv4i32(<vscale x 4 x i32> [[PREDPHI]], <vscale x 4 x i32>* [[TMP20]], i32 4, <vscale x 4 x i1> [[TMP18]])
+; CHECK-NEXT:    [[TMP21:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP22:%.*]] = mul i64 [[TMP21]], 4
+; CHECK-NEXT:    [[INDEX_NEXT2]] = add i64 [[INDEX1]], [[TMP22]]
+; CHECK-NEXT:    [[TMP23:%.*]] = icmp eq i64 [[INDEX_NEXT2]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP23]], label %middle.block, label %vector.body
+; CHECK:       middle.block:
+; CHECK-NEXT:    br i1 true, label %for.end, label %scalar.ph
+;
+
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %entry, %if.end
+  %index = phi i64 [ %index.next, %if.end ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds i32, i32* %cond, i64 %index
+  %0 = load i32, i32* %arrayidx, align 4
+  %tobool.not = icmp eq i32 %0, 0
+  br i1 %tobool.not, label %if.end, label %if.then
+
+if.then:                                          ; preds = %for.body
+  %1 = load i32, i32* %src, align 4
+  br label %if.end
+
+if.end:                                           ; preds = %if.then, %for.body
+  %val.0 = phi i32 [ %1, %if.then ], [ 0, %for.body ]
+  %arrayidx1 = getelementptr inbounds i32, i32* %dst, i64 %index
+  store i32 %val.0, i32* %arrayidx1, align 4
+  %index.next = add nuw i64 %index, 1
+  %exitcond.not = icmp eq i64 %index.next, %n
+  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:                                          ; preds = %for.inc, %entry
+  ret void
+}
+
+
+define void @simple_fdiv(float* noalias %dst, float* noalias %src, i64 %n) #0 {
+; CHECK-LABEL: @simple_fdiv(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 [[N:%.*]], i64 1)
+; CHECK-NEXT:    br i1 false, label %scalar.ph, label %vector.ph
+; CHECK:       vector.ph:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP1:%.*]] = mul i64 [[TMP0]], 4
+; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP3:%.*]] = mul i64 [[TMP2]], 4
+; CHECK-NEXT:    [[TMP4:%.*]] = sub i64 [[TMP3]], 1
+; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i64 [[UMAX]], [[TMP4]]
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP1]]
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
+; CHECK-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i64 [[UMAX]], 1
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TRIP_COUNT_MINUS_1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    br label %vector.body
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX1:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT2:%.*]], %vector.body ]
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[INDEX1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT4:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT3]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP5:%.*]] = call <vscale x 4 x i64> @llvm.experimental.stepvector.nxv4i64()
+; CHECK-NEXT:    [[TMP6:%.*]] = add <vscale x 4 x i64> [[TMP5]], zeroinitializer
+; CHECK-NEXT:    [[TMP7:%.*]] = mul <vscale x 4 x i64> [[TMP6]], shufflevector (<vscale x 4 x i64> insertelement (<vscale x 4 x i64> poison, i64 1, i32 0), <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer)
+; CHECK-NEXT:    [[INDUCTION:%.*]] = add <vscale x 4 x i64> [[BROADCAST_SPLAT4]], [[TMP7]]
+; CHECK-NEXT:    [[TMP8:%.*]] = add i64 [[INDEX1]], 0
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp ule <vscale x 4 x i64> [[INDUCTION]], [[BROADCAST_SPLAT]]
+; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr float, float* [[SRC:%.*]], i64 [[TMP8]]
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr float, float* [[DST:%.*]], i64 [[TMP8]]
+; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr float, float* [[TMP10]], i32 0
+; CHECK-NEXT:    [[TMP13:%.*]] = bitcast float* [[TMP12]] to <vscale x 4 x float>*
+; CHECK-NEXT:    [[WIDE_MASKED_LOAD:%.*]] = call <vscale x 4 x float> @llvm.masked.load.nxv4f32.p0nxv4f32(<vscale x 4 x float>* [[TMP13]], i32 4, <vscale x 4 x i1> [[TMP9]], <vscale x 4 x float> poison)
+; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr float, float* [[TMP11]], i32 0
+; CHECK-NEXT:    [[TMP15:%.*]] = bitcast float* [[TMP14]] to <vscale x 4 x float>*
+; CHECK-NEXT:    [[WIDE_MASKED_LOAD5:%.*]] = call <vscale x 4 x float> @llvm.masked.load.nxv4f32.p0nxv4f32(<vscale x 4 x float>* [[TMP15]], i32 4, <vscale x 4 x i1> [[TMP9]], <vscale x 4 x float> poison)
+; CHECK-NEXT:    [[TMP16:%.*]] = fdiv <vscale x 4 x float> [[WIDE_MASKED_LOAD]], [[WIDE_MASKED_LOAD5]]
+; CHECK-NEXT:    [[TMP17:%.*]] = bitcast float* [[TMP14]] to <vscale x 4 x float>*
+; CHECK-NEXT:    call void @llvm.masked.store.nxv4f32.p0nxv4f32(<vscale x 4 x float> [[TMP16]], <vscale x 4 x float>* [[TMP17]], i32 4, <vscale x 4 x i1> [[TMP9]])
+; CHECK-NEXT:    [[TMP18:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP19:%.*]] = mul i64 [[TMP18]], 4
+; CHECK-NEXT:    [[INDEX_NEXT2]] = add i64 [[INDEX1]], [[TMP19]]
+; CHECK-NEXT:    [[TMP20:%.*]] = icmp eq i64 [[INDEX_NEXT2]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP20]], label %middle.block, label %vector.body, !llvm.loop [[LOOP18:![0-9]+]]
+; CHECK:       middle.block:
+; CHECK-NEXT:    br i1 true, label %while.end.loopexit, label %scalar.ph
+;
+entry:
+  br label %while.body
+
+while.body:                                       ; preds = %while.body, %entry
+  %index = phi i64 [ %index.next, %while.body ], [ 0, %entry ]
+  %gep1 = getelementptr float, float* %src, i64 %index
+  %gep2 = getelementptr float, float* %dst, i64 %index
+  %val1 = load float, float* %gep1
+  %val2 = load float, float* %gep2
+  %res = fdiv float %val1, %val2
+  store float %res, float* %gep2
+  %index.next = add nsw i64 %index, 1
+  %cmp10 = icmp ult i64 %index.next, %n
+  br i1 %cmp10, label %while.body, label %while.end.loopexit, !llvm.loop !0
+
+while.end.loopexit:                               ; preds = %while.body
+  ret void
+}
+
+
+define i32 @add_reduction_i32(i32* %ptr, i64 %n) #0 {
+; CHECK-LABEL: @add_reduction_i32(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 [[N:%.*]], i64 1)
+; CHECK-NEXT:    br i1 false, label %scalar.ph, label %vector.ph
+; CHECK:       vector.ph:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP1:%.*]] = mul i64 [[TMP0]], 4
+; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP3:%.*]] = mul i64 [[TMP2]], 4
+; CHECK-NEXT:    [[TMP4:%.*]] = sub i64 [[TMP3]], 1
+; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i64 [[UMAX]], [[TMP4]]
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP1]]
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
+; CHECK-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i64 [[UMAX]], 1
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TRIP_COUNT_MINUS_1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    br label %vector.body
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX1:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT2:%.*]], %vector.body ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <vscale x 4 x i32> [ insertelement (<vscale x 4 x i32> zeroinitializer, i32 0, i32 0), %vector.ph ], [ [[TMP16:%.*]], %vector.body ]
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[INDEX1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT4:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT3]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP5:%.*]] = call <vscale x 4 x i64> @llvm.experimental.stepvector.nxv4i64()
+; CHECK-NEXT:    [[TMP6:%.*]] = add <vscale x 4 x i64> [[TMP5]], zeroinitializer
+; CHECK-NEXT:    [[TMP7:%.*]] = mul <vscale x 4 x i64> [[TMP6]], shufflevector (<vscale x 4 x i64> insertelement (<vscale x 4 x i64> poison, i64 1, i32 0), <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer)
+; CHECK-NEXT:    [[INDUCTION:%.*]] = add <vscale x 4 x i64> [[BROADCAST_SPLAT4]], [[TMP7]]
+; CHECK-NEXT:    [[TMP8:%.*]] = add i64 [[INDEX1]], 0
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp ule <vscale x 4 x i64> [[INDUCTION]], [[BROADCAST_SPLAT]]
+; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr i32, i32* [[PTR:%.*]], i64 [[TMP8]]
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr i32, i32* [[TMP10]], i32 0
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast i32* [[TMP11]] to <vscale x 4 x i32>*
+; CHECK-NEXT:    [[WIDE_MASKED_LOAD:%.*]] = call <vscale x 4 x i32> @llvm.masked.load.nxv4i32.p0nxv4i32(<vscale x 4 x i32>* [[TMP12]], i32 4, <vscale x 4 x i1> [[TMP9]], <vscale x 4 x i32> poison)
+; CHECK-NEXT:    [[TMP13]] = add <vscale x 4 x i32> [[VEC_PHI]], [[WIDE_MASKED_LOAD]]
+; CHECK-NEXT:    [[TMP14:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP15:%.*]] = mul i64 [[TMP14]], 4
+; CHECK-NEXT:    [[INDEX_NEXT2]] = add i64 [[INDEX1]], [[TMP15]]
+; CHECK-NEXT:    [[TMP16:%.*]] = select <vscale x 4 x i1> [[TMP9]], <vscale x 4 x i32> [[TMP13]], <vscale x 4 x i32> [[VEC_PHI]]
+; CHECK-NEXT:    [[TMP17:%.*]] = icmp eq i64 [[INDEX_NEXT2]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP17]], label %middle.block, label %vector.body, !llvm.loop [[LOOP20:![0-9]+]]
+; CHECK:       middle.block:
+; CHECK-NEXT:    [[TMP18:%.*]] = call i32 @llvm.vector.reduce.add.nxv4i32(<vscale x 4 x i32> [[TMP16]])
+; CHECK-NEXT:    br i1 true, label %while.end.loopexit, label %scalar.ph
+;
+entry:
+  br label %while.body
+
+while.body:                                       ; preds = %while.body, %entry
+  %index = phi i64 [ %index.next, %while.body ], [ 0, %entry ]
+  %red = phi i32 [ %red.next, %while.body ], [ 0, %entry ]
+  %gep = getelementptr i32, i32* %ptr, i64 %index
+  %val = load i32, i32* %gep
+  %red.next = add i32 %red, %val
+  %index.next = add nsw i64 %index, 1
+  %cmp10 = icmp ult i64 %index.next, %n
+  br i1 %cmp10, label %while.body, label %while.end.loopexit, !llvm.loop !0
+
+while.end.loopexit:                               ; preds = %while.body
+  ret i32 %red.next
+}
+
+
+define float @add_reduction_f32(float* %ptr, i64 %n) #0 {
+; CHECK-LABEL: @add_reduction_f32(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 [[N:%.*]], i64 1)
+; CHECK-NEXT:    br i1 false, label %scalar.ph, label %vector.ph
+; CHECK:       vector.ph:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP1:%.*]] = mul i64 [[TMP0]], 4
+; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP3:%.*]] = mul i64 [[TMP2]], 4
+; CHECK-NEXT:    [[TMP4:%.*]] = sub i64 [[TMP3]], 1
+; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i64 [[UMAX]], [[TMP4]]
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP1]]
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
+; CHECK-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i64 [[UMAX]], 1
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TRIP_COUNT_MINUS_1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    br label %vector.body
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX1:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT2:%.*]], %vector.body ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi float [ 0.000000e+00, %vector.ph ], [ [[TMP14:%.*]], %vector.body ]
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[INDEX1]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT4:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT3]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP5:%.*]] = call <vscale x 4 x i64> @llvm.experimental.stepvector.nxv4i64()
+; CHECK-NEXT:    [[TMP6:%.*]] = add <vscale x 4 x i64> [[TMP5]], zeroinitializer
+; CHECK-NEXT:    [[TMP7:%.*]] = mul <vscale x 4 x i64> [[TMP6]], shufflevector (<vscale x 4 x i64> insertelement (<vscale x 4 x i64> poison, i64 1, i32 0), <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer)
+; CHECK-NEXT:    [[INDUCTION:%.*]] = add <vscale x 4 x i64> [[BROADCAST_SPLAT4]], [[TMP7]]
+; CHECK-NEXT:    [[TMP8:%.*]] = add i64 [[INDEX1]], 0
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp ule <vscale x 4 x i64> [[INDUCTION]], [[BROADCAST_SPLAT]]
+; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr float, float* [[PTR:%.*]], i64 [[TMP8]]
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr float, float* [[TMP10]], i32 0
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast float* [[TMP11]] to <vscale x 4 x float>*
+; CHECK-NEXT:    [[WIDE_MASKED_LOAD:%.*]] = call <vscale x 4 x float> @llvm.masked.load.nxv4f32.p0nxv4f32(<vscale x 4 x float>* [[TMP12]], i32 4, <vscale x 4 x i1> [[TMP9]], <vscale x 4 x float> poison)
+; CHECK-NEXT:    [[TMP13:%.*]] = select <vscale x 4 x i1> [[TMP9]], <vscale x 4 x float> [[WIDE_MASKED_LOAD]], <vscale x 4 x float> shufflevector (<vscale x 4 x float> insertelement (<vscale x 4 x float> poison, float -0.000000e+00, i32 0), <vscale x 4 x float> poison, <vscale x 4 x i32> zeroinitializer)
+; CHECK-NEXT:    [[TMP14]] = call float @llvm.vector.reduce.fadd.nxv4f32(float [[VEC_PHI]], <vscale x 4 x float> [[TMP13]])
+; CHECK-NEXT:    [[TMP15:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP16:%.*]] = mul i64 [[TMP15]], 4
+; CHECK-NEXT:    [[INDEX_NEXT2]] = add i64 [[INDEX1]], [[TMP16]]
+; CHECK-NEXT:    [[TMP17:%.*]] = icmp eq i64 [[INDEX_NEXT2]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP17]], label %middle.block, label %vector.body, !llvm.loop [[LOOP22:![0-9]+]]
+; CHECK:       middle.block:
+; CHECK-NEXT:    br i1 true, label %while.end.loopexit, label %scalar.ph
+;
+entry:
+  br label %while.body
+
+while.body:                                       ; preds = %while.body, %entry
+  %index = phi i64 [ %index.next, %while.body ], [ 0, %entry ]
+  %red = phi float [ %red.next, %while.body ], [ 0.000000, %entry ]
+  %gep = getelementptr float, float* %ptr, i64 %index
+  %val = load float, float* %gep
+  %red.next = fadd float %red, %val
+  %index.next = add nsw i64 %index, 1
+  %cmp10 = icmp ult i64 %index.next, %n
+  br i1 %cmp10, label %while.body, label %while.end.loopexit, !llvm.loop !0
+
+while.end.loopexit:                               ; preds = %while.body
+  ret float %red.next
+}
+
+
+; Negative tests where we don't expect tail-folding
+
+; Integer divides can throw exceptions and since we can't scalarize conditional
+; divides for scalable vectors we just don't bother vectorizing.
+define void @simple_idiv(i32* noalias %dst, i32* noalias %src, i64 %n) #0 {
+; CHECK-LABEL: @simple_idiv(
+; CHECK-NOT:   vector.body
+;
+entry:
+  br label %while.body
+
+while.body:                                       ; preds = %while.body, %entry
+  %index = phi i64 [ %index.next, %while.body ], [ 0, %entry ]
+  %gep1 = getelementptr i32, i32* %src, i64 %index
+  %gep2 = getelementptr i32, i32* %dst, i64 %index
+  %val1 = load i32, i32* %gep1
+  %val2 = load i32, i32* %gep2
+  %res = udiv i32 %val1, %val2
+  store i32 %res, i32* %gep2
+  %index.next = add nsw i64 %index, 1
+  %cmp10 = icmp ult i64 %index.next, %n
+  br i1 %cmp10, label %while.body, label %while.end.loopexit, !llvm.loop !0
+
+while.end.loopexit:                               ; preds = %while.body
+  ret void
+}
+
+!0 = distinct !{!0, !1, !2}
+!1 = !{!"llvm.loop.vectorize.width", i32 4}
+!2 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
+
 attributes #0 = { "target-features"="+sve" }