[ARM][MVE] canTailPredicateLoop

[SjoerdMeijer]

Summary

This implements TTI hook canTailPredicateLoop for MVE.
This is a first version, with the following restrictions:

• only operate on single block loops,
• only accept integer types, I will follow this up soon to add support for floats.

[samparker]

llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp

1004

• isHardwareLoopProfitable will already catch these and tries quite hard to get it right.
• isa<> not dyn_cast.
• The zext and truncs could likely be folded into memory options, so you're gonna need to check the uses / users.
• sext would be the same.

1016

If we're checking for casts within the loop, I don't see why this is necessary, because we shouldn't then also have to then check all the types used.

[SjoerdMeijer]

Thanks for taking a look! Will take this on board.

[samparker]

llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp

1009

... check uses/users.

1018

SmallVectorImpl<Instruction*>

1026

if (Stride != 1) return false

1051

isa

1070

Makes sense to just check the stride here and then we can exit early.

[SjoerdMeijer]

• strides: now allowing 1 and -1, and added a test case for -1.
• checking the operand of zext/sext/trunc to see if it is a load/store, and added a test case when trunc is actually allowed.

[samparker]

llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp

1010

You need to check the user of the load, not it's operand. Also the logic is wrong here, we need to accept extending loads and truncating stores.

1050

What is the complication of supporting half and float now?

1064

You don't need FirstStride, the uninitialised Stride holds the same info.

[SjoerdMeijer]

Thanks for checking. Got myself confused about the loads/stores, but I think we indeed want to accept extending loads and narrowing stores, so I've added checks and a bunch of tests:

• for SEXT/ZEXT, the operand needs to be a load,
• for TRUNC, the user needs to be a store.

And have also allowed the float types.

[SjoerdMeijer]

added one more FP test case.

[samparker]

Have you checked the current code generation for masked fpext/fptrunc load/stores? I have a strong recollection that the codegen is absolutely awful.

[dmgreen]

In an ideal world / in the long run, it would be better for the option of whether to predicate to be dependant on the vectorisation factor. I believe we will have cases where we can pick between a slower tail predicated loop or a quicker non-tail predicated loop, and we should be able to pick the quicker one. This implies that it should really be incorporated into the cost mechanism inside the vectoriser.

(Also the current method seems a bit fuzzy, even by vectoriser standards. It seems to essentially be trying to pre-computing what vectorisation factor it thinks will be produced, choosing whether to predicate and hoping that the vectoriser does end up picking the same factor. Without doing something it doesn't expect).

But I think you are correct that this is a sensible way forward, to get something working. We can adjust it as needed, and perhaps it will work OK for almost all cases.

It does need to option to turn it off though. I would expect that should be off by default for the moment until the backend part is in and we have verified that this is working well enough.

llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp

1065

Can you explain the reasoning for adding -1 strides? I would expect them to become vrev's and vmov's, both of which will be difficult to prove in the backend (if not outright incorrect).

Are you expecting any reverse shuffles to be cancelled out?

[SjoerdMeijer]

• fptrunc/fpextends: the loops are not that bad actually, but there's some quite terrible codegen around that, so let's indeed reject this for now.
• stride -1: I think that should be okay, but I do see that complicates things quite a bit for the backend, so let's indeed reject that for now too.
• this is now enabled/disabled by exisiting option -disable-mve-tail-predication.

[dmgreen]

llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp

1088

Nice one.

[samparker]

LGTM. Something I'll like to see in a follow-up patch would be preventing predication in the presence of fcmps.

[SjoerdMeijer]

Cheers, will do.

llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp

bool ARMTTIImpl::isHardwareLoopProfitable(Loop *L, ScalarEvolution &SE,
   HWLoopInfo.CounterInReg = true;
   HWLoopInfo.IsNestingLegal = false;
   HWLoopInfo.PerformEntryTest = true;
   HWLoopInfo.CountType = Type::getInt32Ty(C);
   HWLoopInfo.LoopDecrement = ConstantInt::get(HWLoopInfo.CountType, 1);
   return true;
 }
 
+static bool canTailPredicateInstruction(Instruction &I, int &ICmpCount) {
+  // We don't allow icmp's, and because we only look at single block loops,
+  // we simply count the icmps, i.e. there should only be 1 for the backedge.
+  if (isa<ICmpInst>(&I) && ++ICmpCount > 1)
+    return false;
+
+  // Extends have to be extending-loads
+  if (isa<SExtInst>(&I) || isa<ZExtInst>(&I) || isa<FPExtInst>(&I))
+    if (!I.getOperand(0)->hasOneUse() || !isa<LoadInst>(I.getOperand(0)))
+      return false;
+
+  // Truncs have to be narrowing-stores
+  if (isa<TruncInst>(&I))
+    if (!I.hasOneUse() || !isa<StoreInst>(*I.user_begin()))
+      return false;
+
+  return true;
+}
+
+// To set up a tail-predicated loop, we need to know the total number of
+// elements processed by that loop. Thus, we need to determine the element
+// size and:
+// 1) it should be uniform for all operations in the vector loop, so we
+//    e.g. don't want any widening/narrowing operations.
+// 2) it should be smaller than i64s because we don't have vector operations
+//    that work on i64s.
+// 3) we don't want elements to be reversed or shuffled, to make sure the
+//    tail-predication masks/predicates the right lanes.
+//
+static bool canTailPredicateLoop(Loop *L, LoopInfo *LI, ScalarEvolution &SE,
+                                 const DataLayout &DL,
+                                 const LoopAccessInfo *LAI) {
+  PredicatedScalarEvolution PSE = LAI->getPSE();
+  int ICmpCount = 0;
+  int Stride = 0;
+
+  LLVM_DEBUG(dbgs() << "tail-predication: checking allowed instructions\n");
+  SmallVector<Instruction *, 16> LoadStores;
+  for (BasicBlock *BB : L->blocks()) {
+    for (Instruction &I : BB->instructionsWithoutDebug()) {
+      if (isa<PHINode>(&I))
+        continue;
+      if (!canTailPredicateInstruction(I, ICmpCount)) {
+        LLVM_DEBUG(dbgs() << "Instruction not allowed: "; I.dump());
+        return false;
+      }
+
+      Type *T  = I.getType();
+      if (T->isPointerTy())
+        T = T->getPointerElementType();
+
+      if (T->getScalarSizeInBits() > 32) {
+        LLVM_DEBUG(dbgs() << "Unsupported Type: "; T->dump());
+        return false;
+      }
+
+      if (isa<StoreInst>(I) || isa<LoadInst>(I)) {
+        Value *Ptr = isa<LoadInst>(I) ? I.getOperand(0) : I.getOperand(1);
+        int64_t NextStride = getPtrStride(PSE, Ptr, L);
+        // TODO: for now only allow consecutive strides of 1 and -1. We
+        // could support other strides as long as it is uniform, but let's
+        // keep it simple for now.
+        if (Stride == 0 && (NextStride == 1 || NextStride == -1)) {
+          Stride = NextStride;
+          continue;
+        }
+        if (Stride != NextStride) {
+          LLVM_DEBUG(dbgs() << "Different strides found, can't "
+                               "tail-predicate\n.");
+          return false;
+        }
+      }
+    }
+  }
+
+  LLVM_DEBUG(dbgs() << "tail-predication: all instructions allowed!\n");
+  return true;
+}
+
 bool ARMTTIImpl::preferPredicateOverEpilogue(Loop *L, LoopInfo *LI,
                                              ScalarEvolution &SE,
                                              AssumptionCache &AC,
                                              TargetLibraryInfo *TLI,
                                              DominatorTree *DT,
                                              const LoopAccessInfo *LAI) {
   // Creating a predicated vector loop is the first step for generating a
   // tail-predicated hardware loop, for which we need the MVE masked
   // load/stores instructions:
   if (!ST->hasMVEIntegerOps())
     return false;
 
+  // For now, restrict this to single block loops.
+  if (L->getNumBlocks() > 1) {
+    LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: not a single block "
+                         "loop.\n");
+    return false;
+  }
+
+  assert(L->empty() && "preferPredicateOverEpilogue: inner-loop expected");
+
   HardwareLoopInfo HWLoopInfo(L);
   if (!HWLoopInfo.canAnalyze(*LI)) {
     LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: hardware-loop is not "
                          "analyzable.\n");
     return false;
   }
 
   // This checks if we have the low-overhead branch architecture
   // extension, and if we will create a hardware-loop:
   if (!isHardwareLoopProfitable(L, SE, AC, TLI, HWLoopInfo)) {
     LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: hardware-loop is not "
                          "profitable.\n");
     return false;
   }
 
   if (!HWLoopInfo.isHardwareLoopCandidate(SE, *LI, *DT)) {
     LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: hardware-loop is not "
                          "a candidate.\n");
     return false;
   }
 
-  // TODO: to set up a tail-predicated loop, which works by setting up
-  // the total number of elements processed by the loop, we need to
-  // determine the element size here, and if it is uniform for all operations
-  // in the vector loop. This means we will reject narrowing/widening
-  // operations, and don't want to predicate the vector loop, which is
-  // the main prep step for tail-predicated loops.
-
-  return false;
+  return canTailPredicateLoop(L, LI, SE, DL, LAI);
 }
 
 
 void ARMTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
                                          TTI::UnrollingPreferences &UP) {
   // Only currently enable these preferences for M-Class cores.
   if (!ST->isMClass())
     return BasicTTIImplBase::getUnrollingPreferences(L, SE, UP);

llvm/test/Transforms/LoopVectorize/ARM/prefer-tail-loop-folding.ll

 ; RUN: opt -mtriple=thumbv8.1m.main-arm-eabihf < %s -loop-vectorize -S | \
 ; RUN:  FileCheck %s -check-prefixes=CHECK,NO-FOLDING
 
 ; RUN: opt -mtriple=thumbv8.1m.main-arm-eabihf -mattr=-mve < %s -loop-vectorize -enable-arm-maskedldst=true -S | \
 ; RUN:  FileCheck %s -check-prefixes=CHECK,NO-FOLDING
 
 ; RUN: opt -mtriple=thumbv8.1m.main-arm-eabihf -mattr=+mve < %s -loop-vectorize -enable-arm-maskedldst=false -S | \
 ; RUN:  FileCheck %s -check-prefixes=CHECK,NO-FOLDING
 
 ; Disabling the low-overhead branch extension will make
 ; 'isHardwareLoopProfitable' return false, so that we test avoiding folding for
 ; these cases.
 ; RUN: opt -mtriple=thumbv8.1m.main-arm-eabihf -mattr=+mve,-lob < %s -loop-vectorize -enable-arm-maskedldst=true -S | \
 ; RUN:  FileCheck %s -check-prefixes=CHECK,NO-FOLDING
 
-; RUN: opt -mtriple=thumbv8.1m.main-arm-eabihf -mattr=+mve < %s -loop-vectorize -enable-arm-maskedldst=true -S | \
+; RUN: opt -mtriple=thumbv8.1m.main-arm-eabihf -mattr=+mve.fp < %s -loop-vectorize -enable-arm-maskedldst=true -S | \
 ; RUN:  FileCheck %s -check-prefixes=CHECK,PREFER-FOLDING
 
-define dso_local void @tail_folding(i32* noalias nocapture %A, i32* noalias nocapture readonly %B, i32* noalias nocapture readonly %C) {
-; CHECK-LABEL: tail_folding(
+define void @prefer_folding(i32* noalias nocapture %A, i32* noalias nocapture readonly %B, i32* noalias nocapture readonly %C) #0 {
+; CHECK-LABEL:    prefer_folding(
+; PREFER-FOLDING: vector.body:
+; PREFER-FOLDING: call <4 x i32> @llvm.masked.load.v4i32.p0v4i32
+; PREFER-FOLDING: call <4 x i32> @llvm.masked.load.v4i32.p0v4i32
+; PREFER-FOLDING: call void @llvm.masked.store.v4i32.p0v4i32
+; PREFER-FOLDING: br i1 %{{.*}}, label %{{.*}}, label %vector.body
 ;
-; NO-FOLDING-NOT:  call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(
-; NO-FOLDING-NOT:  call void @llvm.masked.store.v4i32.p0v4i32(
+; NO-FOLDING-NOT: call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(
+; NO-FOLDING-NOT: call void @llvm.masked.store.v4i32.p0v4i32(
+; NO-FOLDING:     br i1 %{{.*}}, label %{{.*}}, label %for.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.09 = phi i32 [ 0, %entry ], [ %add3, %for.body ]
+  %arrayidx = getelementptr inbounds i32, i32* %B, i32 %i.09
+  %0 = load i32, i32* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds i32, i32* %C, i32 %i.09
+  %1 = load i32, i32* %arrayidx1, align 4
+  %add = add nsw i32 %1, %0
+  %arrayidx2 = getelementptr inbounds i32, i32* %A, i32 %i.09
+  store i32 %add, i32* %arrayidx2, align 4
+  %add3 = add nuw nsw i32 %i.09, 1
+  %exitcond = icmp eq i32 %add3, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+}
+
+define void @mixed_types(i16* noalias nocapture %A, i16* noalias nocapture readonly %B, i16* noalias nocapture readonly %C, i32* noalias nocapture %D, i32* noalias nocapture readonly %E, i32* noalias nocapture readonly %F) #0 {
+; CHECK-LABEL:        mixed_types(
+; PREFER-FOLDING:     vector.body:
+; PREFER-FOLDING:     call <4 x i16> @llvm.masked.load.v4i16.p0v4i16
+; PREFER-FOLDING:     call <4 x i16> @llvm.masked.load.v4i16.p0v4i16
+; PREFER-FOLDING:     call void @llvm.masked.store.v4i16.p0v4i16
+; PREFER-FOLDING:     call <4 x i32> @llvm.masked.load.v4i32.p0v4i32
+; PREFER-FOLDING:     call <4 x i32> @llvm.masked.load.v4i32.p0v4i32
+; PREFER-FOLDING:     call void @llvm.masked.store.v4i32.p0v4i32
+; PREFER-FOLDING:     br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.018 = phi i32 [ 0, %entry ], [ %add9, %for.body ]
+  %arrayidx = getelementptr inbounds i16, i16* %B, i32 %i.018
+  %0 = load i16, i16* %arrayidx, align 2
+  %arrayidx1 = getelementptr inbounds i16, i16* %C, i32 %i.018
+  %1 = load i16, i16* %arrayidx1, align 2
+  %add = add i16 %1, %0
+  %arrayidx4 = getelementptr inbounds i16, i16* %A, i32 %i.018
+  store i16 %add, i16* %arrayidx4, align 2
+  %arrayidx5 = getelementptr inbounds i32, i32* %E, i32 %i.018
+  %2 = load i32, i32* %arrayidx5, align 4
+  %arrayidx6 = getelementptr inbounds i32, i32* %F, i32 %i.018
+  %3 = load i32, i32* %arrayidx6, align 4
+  %add7 = add nsw i32 %3, %2
+  %arrayidx8 = getelementptr inbounds i32, i32* %D, i32 %i.018
+  store i32 %add7, i32* %arrayidx8, align 4
+  %add9 = add nuw nsw i32 %i.018, 1
+  %exitcond = icmp eq i32 %add9, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+}
+
+define void @unsupported_i64_type(i64* noalias nocapture %A, i64* noalias nocapture readonly %B, i64* noalias nocapture readonly %C) #0 {
+; CHECK-LABEL:        unsupported_i64_type(
+; PREFER-FOLDING-NOT: vector.body:
+; PREFER-FOLDING-NOT: llvm.masked.load
+; PREFER-FOLDING-NOT: llvm.masked.store
+; PREFER-FOLDING:     for.body:
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.09 = phi i32 [ 0, %entry ], [ %add3, %for.body ]
+  %arrayidx = getelementptr inbounds i64, i64* %B, i32 %i.09
+  %0 = load i64, i64* %arrayidx, align 8
+  %arrayidx1 = getelementptr inbounds i64, i64* %C, i32 %i.09
+  %1 = load i64, i64* %arrayidx1, align 8
+  %add = add nsw i64 %1, %0
+  %arrayidx2 = getelementptr inbounds i64, i64* %A, i32 %i.09
+  store i64 %add, i64* %arrayidx2, align 8
+  %add3 = add nuw nsw i32 %i.09, 1
+  %exitcond = icmp eq i32 %add3, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+}
+
+define void @zero_extending_load_allowed(i32* noalias nocapture %A, i8* noalias nocapture readonly %B, i32* noalias nocapture readonly %C) #0 {
+; CHECK-LABEL:    zero_extending_load_allowed(
+; PREFER-FOLDING: vector.body:
+; PREFER-FOLDING: call <4 x i8> @llvm.masked.load.v4i8.p0v4i8
+; PREFER-FOLDING: call <4 x i32> @llvm.masked.load.v4i32.p0v4i32
+; PREFER-FOLDING: call void @llvm.masked.store.v4i32.p0v4i32
+; PREFER-FOLDING: br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.09 = phi i32 [ 0, %entry ], [ %add3, %for.body ]
+  %arrayidx = getelementptr inbounds i8, i8* %B, i32 %i.09
+  %0 = load i8, i8* %arrayidx, align 1
+  %conv = zext i8 %0 to i32
+  %arrayidx1 = getelementptr inbounds i32, i32* %C, i32 %i.09
+  %1 = load i32, i32* %arrayidx1, align 4
+  %add = add nsw i32 %1, %conv
+  %arrayidx2 = getelementptr inbounds i32, i32* %A, i32 %i.09
+  store i32 %add, i32* %arrayidx2, align 4
+  %add3 = add nuw nsw i32 %i.09, 1
+  %exitcond = icmp eq i32 %add3, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+}
+
+define void @sign_extending_load_allowed(i32* noalias nocapture %A, i8* noalias nocapture readonly %B, i32* noalias nocapture readonly %C) #0 {
+; CHECK-LABEL:    sign_extending_load_allowed(
+; PREFER-FOLDING: vector.body:
+; PREFER-FOLDING: call <4 x i8> @llvm.masked.load.v4i8.p0v4i8
+; PREFER-FOLDING: call <4 x i32> @llvm.masked.load.v4i32.p0v4i32
+; PREFER-FOLDING: call void @llvm.masked.store.v4i32.p0v4i32
+; PREFER-FOLDING: br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.09 = phi i32 [ 0, %entry ], [ %add3, %for.body ]
+  %arrayidx = getelementptr inbounds i8, i8* %B, i32 %i.09
+  %0 = load i8, i8* %arrayidx, align 1
+  %conv = sext i8 %0 to i32
+  %arrayidx1 = getelementptr inbounds i32, i32* %C, i32 %i.09
+  %1 = load i32, i32* %arrayidx1, align 4
+  %add = add nsw i32 %1, %conv
+  %arrayidx2 = getelementptr inbounds i32, i32* %A, i32 %i.09
+  store i32 %add, i32* %arrayidx2, align 4
+  %add3 = add nuw nsw i32 %i.09, 1
+  %exitcond = icmp eq i32 %add3, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+}
+
+define void @narrowing_load_not_allowed(i8* noalias nocapture %A, i8* noalias nocapture readonly %B, i16* noalias nocapture readonly %C) #0 {
+; CHECK-LABEL:        narrowing_load_not_allowed(
+; PREFER-FOLDING:     vector.body:
+; PREFER-FOLDING-NOT: llvm.masked.load
+; PREFER-FOLDING-NOT: llvm.masked.store
+; PREFER-FOLDING:     br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:                                 ; preds = %for.body
+  ret void
+
+for.body:                                         ; preds = %for.body, %entry
+  %i.012 = phi i32 [ 0, %entry ], [ %add6, %for.body ]
+  %arrayidx = getelementptr inbounds i16, i16* %C, i32 %i.012
+  %0 = load i16, i16* %arrayidx, align 2
+  %arrayidx1 = getelementptr inbounds i8, i8* %B, i32 %i.012
+  %1 = load i8, i8* %arrayidx1, align 1
+  %conv3 = trunc i16 %0 to i8
+  %add = add i8 %1, %conv3
+  %arrayidx5 = getelementptr inbounds i8, i8* %A, i32 %i.012
+  store i8 %add, i8* %arrayidx5, align 1
+  %add6 = add nuw nsw i32 %i.012, 1
+  %exitcond = icmp eq i32 %add6, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+}
+
+define void @narrowing_store_allowed(i8* noalias nocapture %A, i32* noalias nocapture readonly %B, i32* noalias nocapture readonly %C) #0 {
+; CHECK-LABEL:    narrowing_store_allowed(
+; PREFER-FOLDING: call void @llvm.masked.store.v4i8.p0v4i8
+; PREFER-FOLDING: br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.09 = phi i32 [ 0, %entry ], [ %add3, %for.body ]
+  %arrayidx = getelementptr inbounds i32, i32* %B, i32 %i.09
+  %0 = load i32, i32* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds i32, i32* %C, i32 %i.09
+  %1 = load i32, i32* %arrayidx1, align 4
+  %add = add nsw i32 %1, %0
+  %conv = trunc i32 %add to i8
+  %arrayidx2 = getelementptr inbounds i8, i8* %A, i32 %i.09
+  store i8 %conv, i8* %arrayidx2, align 1
+  %add3 = add nuw nsw i32 %i.09, 1
+  %exitcond = icmp eq i32 %add3, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+}
+
+; This is a trunc not connected to a store, so we don't allow this.
+; TODO: this is conservative, because the trunc is only used in the
+; loop control statements, and thus not affecting element sizes, so
+; we could allow this case.
+define void @trunc_not_allowed(i32* noalias nocapture %A, i32* noalias nocapture readonly %B, i32* noalias nocapture readonly %C) #0 {
+; CHECK-LABEL:        trunc_not_allowed(
+; PREFER-FOLDING:     vector.body:
+; PREFER-FOLDING-NOT: llvm.masked.load
+; PREFER-FOLDING-NOT: llvm.masked.store
+; PREFER-FOLDING:     br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.09 = phi i32 [ 0, %entry ], [ %add3, %for.body ]
+  %arrayidx = getelementptr inbounds i32, i32* %B, i32 %i.09
+  %0 = load i32, i32* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds i32, i32* %C, i32 %i.09
+  %1 = load i32, i32* %arrayidx1, align 4
+  %add = add nsw i32 %1, %0
+  %arrayidx2 = getelementptr inbounds i32, i32* %A, i32 %i.09
+  store i32 %add, i32* %arrayidx2, align 4
+  %add3 = add nuw nsw i32 %i.09, 1
+
+  %add.iv = trunc i32 %add3 to i16
+
+  %exitcond = icmp eq i16 %add.iv, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+}
+
+define void @trunc_not_allowed_different_vec_elemns(i32* noalias nocapture %A, i32* noalias nocapture readonly %B, i32* noalias nocapture readonly %C, i16* noalias nocapture %D) #0 {
+; CHECK-LABEL:        trunc_not_allowed_different_vec_elemns(
+; PREFER-FOLDING:     vector.body:
+; PREFER-FOLDING-NOT: llvm.masked.load
+; PREFER-FOLDING-NOT: llvm.masked.store
+; PREFER-FOLDING:     br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.021 = phi i32 [ 0, %entry ], [ %add9, %for.body ]
+  %arrayidx = getelementptr inbounds i32, i32* %B, i32 %i.021
+  %0 = load i32, i32* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds i32, i32* %C, i32 %i.021
+  %1 = load i32, i32* %arrayidx1, align 4
+  %add = add nsw i32 %1, %0
+  %arrayidx2 = getelementptr inbounds i32, i32* %A, i32 %i.021
+  store i32 %add, i32* %arrayidx2, align 4
+  %add.tr = trunc i32 %add to i16
+  %conv7 = shl i16 %add.tr, 1
+  %arrayidx8 = getelementptr inbounds i16, i16* %D, i32 %i.021
+  store i16 %conv7, i16* %arrayidx8, align 2
+  %add9 = add nuw nsw i32 %i.021, 1
+  %exitcond = icmp eq i32 %add9, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+}
+
+
+@tab = common global [32 x i8] zeroinitializer, align 1
+
+define i32 @icmp_not_allowed() #0 {
+; CHECK-LABEL:        icmp_not_allowed(
+; PREFER-FOLDING:     vector.body:
+; PREFER-FOLDING-NOT: llvm.masked.load
+; PREFER-FOLDING-NOT: llvm.masked.store
+; PREFER-FOLDING:     br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.body:
+  %i.08 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  %arrayidx = getelementptr inbounds [32 x i8], [32 x i8]* @tab, i32 0, i32 %i.08
+  %0 = load i8, i8* %arrayidx, align 1
+  %cmp1 = icmp eq i8 %0, 0
+  %. = select i1 %cmp1, i8 2, i8 1
+  store i8 %., i8* %arrayidx, align 1
+  %inc = add nsw i32 %i.08, 1
+  %exitcond = icmp slt i32 %inc, 1000
+  br i1 %exitcond, label %for.body, label %for.end
+
+for.end:
+  ret i32 0
+}
+
+define void @pragma_vect_predicate_disable(i32* noalias nocapture %A, i32* noalias nocapture readonly %B, i32* noalias nocapture readonly %C) #0 {
+; CHECK-LABEL: pragma_vect_predicate_disable(
 ;
-; TODO: this needs implementation of TTI::preferPredicateOverEpilogue,
-; then this will be tail-folded too:
-;
-; PREFER-FOLDING-NOT:  call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(
-; PREFER-FOLDING-NOT:  call void @llvm.masked.store.v4i32.p0v4i32(
+; FIXME:
+; respect loop hint predicate.enable = false, and don't tail-fold here:
 ;
+; PREFER-FOLDING: call <4 x i32> @llvm.masked.load.v4i32.p0v4i32
+; PREFER-FOLDING: call <4 x i32> @llvm.masked.load.v4i32.p0v4i32
+; PREFER-FOLDING: call void @llvm.masked.store.v4i32.p0v4i32
+; PREFER-FOLDING: br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.09 = phi i32 [ 0, %entry ], [ %add3, %for.body ]
+  %arrayidx = getelementptr inbounds i32, i32* %B, i32 %i.09
+  %0 = load i32, i32* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds i32, i32* %C, i32 %i.09
+  %1 = load i32, i32* %arrayidx1, align 4
+  %add = add nsw i32 %1, %0
+  %arrayidx2 = getelementptr inbounds i32, i32* %A, i32 %i.09
+  store i32 %add, i32* %arrayidx2, align 4
+  %add3 = add nuw nsw i32 %i.09, 1
+  %exitcond = icmp eq i32 %add3, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body, !llvm.loop !7
+}
+
+; Test directions for array indices i and N-1. I.e. check strides 1 and -1, and
+; force vectorisation with a loop hint.
+define void @strides_different_direction(i32* noalias nocapture %A, i32* noalias nocapture readonly %B, i32* noalias nocapture readonly %C, i32 %N) #0 {
+; CHECK-LABEL: strides_different_direction(
+; PREFER-FOLDING:     vector.body:
+; PREFER-FOLDING-NOT: llvm.masked.load
+; PREFER-FOLDING-NOT: llvm.masked.store
+; PREFER-FOLDING:     br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.09 = phi i32 [ 0, %entry ], [ %add3, %for.body ]
+  %arrayidx = getelementptr inbounds i32, i32* %B, i32 %i.09
+  %0 = load i32, i32* %arrayidx, align 4
+  %sub = sub nsw i32 %N, %i.09
+  %arrayidx1 = getelementptr inbounds i32, i32* %C, i32 %sub
+  %1 = load i32, i32* %arrayidx1, align 4
+  %add = add nsw i32 %1, %0
+  %arrayidx2 = getelementptr inbounds i32, i32* %A, i32 %i.09
+  store i32 %add, i32* %arrayidx2, align 4
+  %add3 = add nuw nsw i32 %i.09, 1
+  %exitcond = icmp eq i32 %add3, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body, !llvm.loop !10
+}
+
+define void @strides_minus_one(i32* noalias nocapture %A, i32* noalias nocapture readonly %B, i32* noalias nocapture readonly %C, i32 %N) #0 {
+; CHECK-LABEL:    strides_minus_one(
+; PREFER-FOLDING: call <4 x i32> @llvm.masked.load.v4i32.p0v4i32
+; PREFER-FOLDING: call <4 x i32> @llvm.masked.load.v4i32.p0v4i32
+; PREFER-FOLDING: call void @llvm.masked.store.v4i32.p0v4i32
+; PREFER-FOLDING: br i1 %{{.*}}, label %{{.*}}, label %vector.body
 entry:
   br label %for.body
 
 for.cond.cleanup:
   ret void
 
 for.body:
-  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
-  %arrayidx = getelementptr inbounds i32, i32* %B, i64 %indvars.iv
+  %i.013 = phi i32 [ 0, %entry ], [ %add5, %for.body ]
+  %sub = sub nsw i32 %N, %i.013
+  %arrayidx = getelementptr inbounds i32, i32* %B, i32 %sub
   %0 = load i32, i32* %arrayidx, align 4
-  %arrayidx2 = getelementptr inbounds i32, i32* %C, i64 %indvars.iv
+  %arrayidx2 = getelementptr inbounds i32, i32* %C, i32 %sub
   %1 = load i32, i32* %arrayidx2, align 4
   %add = add nsw i32 %1, %0
-  %arrayidx4 = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
+  %arrayidx4 = getelementptr inbounds i32, i32* %A, i32 %sub
   store i32 %add, i32* %arrayidx4, align 4
-  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
-  %exitcond = icmp eq i64 %indvars.iv.next, 430
+  %add5 = add nuw nsw i32 %i.013, 1
+  %exitcond = icmp eq i32 %add5, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body, !llvm.loop !10
+}
+
+
+define void @stride_4(i32* noalias nocapture %A, i32* noalias nocapture readonly %B, i32* noalias nocapture readonly %C) #0 {
+; CHECK-LABEL:        stride_4(
+; PREFER-FOLDING:     vector.body:
+; PREFER-FOLDING-NOT: llvm.masked.load
+; PREFER-FOLDING-NOT: llvm.masked.store
+; PREFER-FOLDING:     br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.09 = phi i32 [ 0, %entry ], [ %add3, %for.body ]
+  %arrayidx = getelementptr inbounds i32, i32* %B, i32 %i.09
+  %0 = load i32, i32* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds i32, i32* %C, i32 %i.09
+  %1 = load i32, i32* %arrayidx1, align 4
+  %add = add nsw i32 %1, %0
+  %arrayidx2 = getelementptr inbounds i32, i32* %A, i32 %i.09
+  store i32 %add, i32* %arrayidx2, align 4
+  %add3 = add nuw nsw i32 %i.09, 4
+  %cmp = icmp ult i32 %add3, 731
+  br i1 %cmp, label %for.body, label %for.cond.cleanup, !llvm.loop !5
+}
+
+define void @too_many_loop_blocks(i32* noalias nocapture %A, i32* noalias nocapture readonly %B, i32* noalias nocapture readonly %C) #0 {
+; CHECK-LABEL:        too_many_loop_blocks(
+; PREFER-FOLDING:     vector.body:
+; PREFER-FOLDING-NOT: llvm.masked.load
+; PREFER-FOLDING-NOT: llvm.masked.store
+; PREFER-FOLDING:     br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.09 = phi i32 [ 0, %entry ], [ %add3, %loopincr ]
+  %arrayidx = getelementptr inbounds i32, i32* %B, i32 %i.09
+  %0 = load i32, i32* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds i32, i32* %C, i32 %i.09
+  %1 = load i32, i32* %arrayidx1, align 4
+  %add = add nsw i32 %1, %0
+  %arrayidx2 = getelementptr inbounds i32, i32* %A, i32 %i.09
+  store i32 %add, i32* %arrayidx2, align 4
+  br label %loopincr
+
+loopincr:
+  %add3 = add nuw nsw i32 %i.09, 1
+  %exitcond = icmp eq i32 %add3, 431
   br i1 %exitcond, label %for.cond.cleanup, label %for.body
 }
+
+define dso_local void @half(half* noalias nocapture %A, half* noalias nocapture readonly %B, half* noalias nocapture readonly %C) #0 {
+; CHECK-LABEL:    half(
+; PREFER-FOLDING: vector.body:
+; PREFER-FOLDING: call <8 x half> @llvm.masked.load.v8f16.p0v8f16
+; PREFER-FOLDING: call <8 x half> @llvm.masked.load.v8f16.p0v8f16
+; PREFER-FOLDING: call void @llvm.masked.store.v8f16.p0v8f16
+; PREFER-FOLDING: br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.09 = phi i32 [ 0, %entry ], [ %add3, %for.body ]
+  %arrayidx = getelementptr inbounds half, half* %B, i32 %i.09
+  %0 = load half, half* %arrayidx, align 2
+  %arrayidx1 = getelementptr inbounds half, half* %C, i32 %i.09
+  %1 = load half, half* %arrayidx1, align 2
+  %add = fadd fast half %1, %0
+  %arrayidx2 = getelementptr inbounds half, half* %A, i32 %i.09
+  store half %add, half* %arrayidx2, align 2
+  %add3 = add nuw nsw i32 %i.09, 1
+  %exitcond = icmp eq i32 %add3, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+}
+
+define void @float(float* noalias nocapture %A, float* noalias nocapture readonly %B, float* noalias nocapture readonly %C) #0 {
+; CHECK-LABEL:    float(
+; PREFER-FOLDING: vector.body:
+; PREFER-FOLDING: call <4 x float> @llvm.masked.load.v4f32.p0v4f32
+; PREFER-FOLDING: call <4 x float> @llvm.masked.load.v4f32.p0v4f32
+; PREFER-FOLDING: call void @llvm.masked.store.v4f32.p0v4f32
+; PREFER-FOLDING: br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.09 = phi i32 [ 0, %entry ], [ %add3, %for.body ]
+  %arrayidx = getelementptr inbounds float, float* %B, i32 %i.09
+  %0 = load float, float* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds float, float* %C, i32 %i.09
+  %1 = load float, float* %arrayidx1, align 4
+  %add = fadd fast float %1, %0
+  %arrayidx2 = getelementptr inbounds float, float* %A, i32 %i.09
+  store float %add, float* %arrayidx2, align 4
+  %add3 = add nuw nsw i32 %i.09, 1
+  %exitcond = icmp eq i32 %add3, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body, !llvm.loop !10
+}
+
+define void @double(double* noalias nocapture %A, double* noalias nocapture readonly %B, double* noalias nocapture readonly %C) #0 {
+; CHECK-LABEL:        double(
+; PREFER-FOLDING:     for.body:
+; PREFER-FOLDING-NOT: vector.body:
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.09 = phi i32 [ 0, %entry ], [ %add3, %for.body ]
+  %arrayidx = getelementptr inbounds double, double* %B, i32 %i.09
+  %0 = load double, double* %arrayidx, align 8
+  %arrayidx1 = getelementptr inbounds double, double* %C, i32 %i.09
+  %1 = load double, double* %arrayidx1, align 8
+  %add = fadd fast double %1, %0
+  %arrayidx2 = getelementptr inbounds double, double* %A, i32 %i.09
+  store double %add, double* %arrayidx2, align 8
+  %add3 = add nuw nsw i32 %i.09, 1
+  %exitcond = icmp eq i32 %add3, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+}
+
+define void @fpext_allowed(float* noalias nocapture %A, half* noalias nocapture readonly %B, float* noalias nocapture readonly %C) #0 {
+; CHECK-LABEL:    fpext_allowed(
+; PREFER-FOLDING: vector.body:
+; PREFER-FOLDING: call <4 x half> @llvm.masked.load.v4f16.p0v4f16
+; PREFER-FOLDING: call <4 x float> @llvm.masked.load.v4f32.p0v4f32
+; PREFER-FOLDING: call void @llvm.masked.store.v4f32.p0v4f32
+; PREFER-FOLDING: br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.09 = phi i32 [ 0, %entry ], [ %add3, %for.body ]
+  %arrayidx = getelementptr inbounds half, half* %B, i32 %i.09
+  %0 = load half, half* %arrayidx, align 2
+  %conv = fpext half %0 to float
+  %arrayidx1 = getelementptr inbounds float, float* %C, i32 %i.09
+  %1 = load float, float* %arrayidx1, align 4
+  %add = fadd fast float %1, %conv
+  %arrayidx2 = getelementptr inbounds float, float* %A, i32 %i.09
+  store float %add, float* %arrayidx2, align 4
+  %add3 = add nuw nsw i32 %i.09, 1
+  %exitcond = icmp eq i32 %add3, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+}
+
+define void @fptrunc_allowed(half* noalias nocapture %A, float* noalias nocapture readonly %B, float* noalias nocapture readonly %C) #0 {
+; CHECK-LABEL:    fptrunc_allowed(
+; PREFER-FOLDING: vector.body:
+; PREFER-FOLDING: call <4 x float> @llvm.masked.load.v4f32.p0v4f32
+; PREFER-FOLDING: call <4 x float> @llvm.masked.load.v4f32.p0v4f32
+; PREFER-FOLDING: call void @llvm.masked.store.v4f16.p0v4f16
+; PREFER-FOLDING: br i1 %{{.*}}, label %{{.*}}, label %vector.body
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %i.09 = phi i32 [ 0, %entry ], [ %add3, %for.body ]
+  %arrayidx = getelementptr inbounds float, float* %B, i32 %i.09
+  %0 = load float, float* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds float, float* %C, i32 %i.09
+  %1 = load float, float* %arrayidx1, align 4
+  %add = fadd fast float %1, %0
+  %conv = fptrunc float %add to half
+  %arrayidx2 = getelementptr inbounds half, half* %A, i32 %i.09
+  store half %conv, half* %arrayidx2, align 2
+  %add3 = add nuw nsw i32 %i.09, 1
+  %exitcond = icmp eq i32 %add3, 431
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+}
+
+attributes #0 = { nofree norecurse nounwind "target-features"="+armv8.1-m.main,+mve.fp" }
+
+!5 = distinct !{!5, !6}
+!6 = !{!"llvm.loop.vectorize.enable", i1 true}
+
+!7 = distinct !{!7, !8}
+!8 = !{!"llvm.loop.vectorize.predicate.enable", i1 false}
+
+!10 = distinct !{!10, !11}
+!11 = !{!"llvm.loop.vectorize.width", i32 4}