[ARM] Improve loop unrolling for Cortex-M

[samparker]

Summary

Override the default unroll count to 4 and use the UnrollRemainder option. Disable unrolling on Thumb1 only targets.

[samparker]

Slightly lowered the threshold for forced unrolling.

[efriedma]

Diff is missing context.

lib/Target/ARM/ARMTargetTransformInfo.cpp

596

getUserCost()?

623

Have you experimented with setting "UP.Runtime = false"?

[samparker]

lib/Target/ARM/ARMTargetTransformInfo.cpp

596

This has crossed my mind, I'll look into it.

623

Yes, but there's a lot of performance to be gained by enabling runtime. This patch is trying to get some of that improvement, while minimising the negative affect it can have.

[samparker]

Hi Eli,

Now using the TTI cost model for instruction counting. I've also now removed support for Thumb-1 targets as I am starting to loose the will trying to characterise their behaviour!

cheers,
sam

[efriedma]

lib/Target/ARM/ARMTargetTransformInfo.cpp

619

This seems like you're running into some sort of limitation of unrolling infrastructure. Maybe we need to add a feature to unroll remainder loops?

Also, which function in the test covers this codepath?

[samparker]

lib/Target/ARM/ARMTargetTransformInfo.cpp

619

The integer type check isn't actually tested and wasn't something that I was interested in, so I will remove it.

I'm not sure I understand what you mean. For clarity and posterity, the type check on the SCEV is not querying the number of iterations but whether the expression is based on int, float, etc... values. Currently, there is a TODO in the unroller to handle counts with pointer types. The runtime unroller creates a unrolled body and just uses an if-else statement to execute the correct loop, but the original loop is also called after the unrolled loop for the remaining iterations (N % unroll_count). The runtime unroller, by default, will only unroll the loops for which SCEV can produce a trip count because it can guarantee than the basic block can be duplicated and merged. Otherwise, the body can be duplicated but the basic blocks cannot be merged. The iterate_inc function is what tests this and hopefully highlights the problem that the loop count is dependent on the length of the linked list and SCEV cannot be expected to be able express this.

[samparker]

Hi Eli,

I've removed the integer type check of the trip expression. Please see my inline reply, sorry if I'm stating the obvious but I just wanted to try and avoid any unnecessary misunderstanding with delayed back-and-forth.

Many thanks,
sam

[efriedma]

Okay, I can try to expand on the "limitation of unrolling infrastructure" bit. The question is, given that I have a small loop, and have an expression for the trip count of the loop, what's the optimal way to generate code for the loop?

When the iteration count is generally large, and the CPU doesn't have special hardware for looping quickly, we want to unroll a bunch of times to reduce the iteration overhead as much as possible, then generate a tiny remainder loop (where the performance doesn't really matter).

Okay, but what happens when the iteration count is small, but the loop runs many times? The runtime-unrolled version of the loop is completely useless, and we end up spending most of our time in the remainder loop. So what can we do about that? One option is to fully unroll the remainder loop: we know the maximum trip count, so it's a straightforward unroll operation. Or maybe we could runtime-unroll the remainder loop (and generate a remainder loop for the remainder loop). Or maybe we could try to do something fancy with a switch. I'm not sure what option is best without actually testing it. But there are definitely options here, and we can probably do better than just setting "DefaultUnrollRuntimeCount = 2" to dodge the issue.

[samparker]

Thanks for the clarification and the ideas. This sounded like loop peeling to me so it's what I've used, could you take a look at D36309? I know it doesn't have a test case, I would like your feedback before I go too far. It solves the issues that I have observed though.

thanks,
sam

[samparker]

Hi Eli,

I've updated the patch to use the UnrollRemainder option and I've set the default count to 4 across the cores.

thanks,
sam

[efriedma]

LGTM.

lib/Target/ARM/ARMTargetTransformInfo.cpp

 
   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
                                            Alignment, AddressSpace);
 }
 
 void ARMTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
                                          TTI::UnrollingPreferences &UP) {
   // Only currently enable these preferences for M-Class cores.
-  if (!ST->isMClass() || L->getNumBlocks() != 1)
+  if (!ST->isMClass())
     return BasicTTIImplBase::getUnrollingPreferences(L, SE, UP);
 
+  // Only enable on Thumb-2 targets for simple loops.
+  if (!ST->isThumb2() || L->getNumBlocks() != 1)
+    return;
+
   // Disable loop unrolling for Oz and Os.
   UP.OptSizeThreshold = 0;
   UP.PartialOptSizeThreshold = 0;
+  BasicBlock *BB = L->getLoopLatch();
+  if (BB->getParent()->optForSize())
+    return;
 
   // Scan the loop: don't unroll loops with calls as this could prevent
   // inlining.
-  BasicBlock *BB = L->getLoopLatch();
+  unsigned Cost = 0;
   for (auto &I : *BB) {
     if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
       ImmutableCallSite CS(&I);
       if (const Function *F = CS.getCalledFunction()) {
         if (!isLoweredToCall(F))
           continue;
       }
       return;
     }
+    SmallVector<const Value*, 4> Operands(I.value_op_begin(),
+                                          I.value_op_end());
+    Cost += getUserCost(&I, Operands);
   }
 
-  // Enable partial and runtime unrolling, set the initial threshold based upon
-  // the number of registers available.
   UP.Partial = true;
   UP.Runtime = true;
-  UP.Threshold = ST->isThumb1Only() ? 75 : 150;
-  UP.PartialThreshold = ST->isThumb1Only() ? 75 : 150;
+  UP.UnrollRemainder = true;
+  UP.DefaultUnrollRuntimeCount = 4;
+
+  // Force unrolling small loops can be very useful because of the branch
+  // taken cost of the backedge.
+  if (Cost < 12)
+    UP.Force = true;
 }

test/Transforms/LoopUnroll/ARM/loop-unrolling.ll

-; RUN: opt -mtriple=armv7 -mcpu=cortex-a57 -loop-unroll -S %s -o - | FileCheck %s --check-prefix=CHECK-UNROLL-V7
-; RUN: opt -mtriple=thumbv7 -mcpu=cortex-a57 -loop-unroll -S %s -o - | FileCheck %s --check-prefix=CHECK-UNROLL-V7
-; RUN: opt -mtriple=thumbv8m -mcpu=cortex-m23 -loop-unroll -S %s -o - | FileCheck %s --check-prefix=CHECK-UNROLL-SMALL
-; RUN: opt -mtriple=thumbv7m -mcpu=cortex-m4 -loop-unroll -S %s -o - | FileCheck %s --check-prefix=CHECK-UNROLL
-; RUN: opt -mtriple=thumbv8m.main -mcpu=cortex-m33 -loop-unroll -S %s -o - | FileCheck %s --check-prefix=CHECK-UNROLL
+; RUN: opt -mtriple=armv7 -mcpu=cortex-a57 -loop-unroll -S %s -o - | FileCheck %s --check-prefix=CHECK-UNROLL-A
+; RUN: opt -mtriple=thumbv7 -mcpu=cortex-a57 -loop-unroll -S %s -o - | FileCheck %s --check-prefix=CHECK-UNROLL-A
+; RUN: opt -mtriple=thumbv8m -mcpu=cortex-m23 -loop-unroll -S %s -o - | FileCheck %s --check-prefix=CHECK-UNROLL-T1
+; RUN: opt -mtriple=thumbv8m.main -mcpu=cortex-m33 -loop-unroll -S %s -o - | FileCheck %s --check-prefix=CHECK-UNROLL-T2
+; RUN: opt -mtriple=thumbv7em -mcpu=cortex-m7 -loop-unroll -S %s -o - | FileCheck %s --check-prefix=CHECK-UNROLL-T2
 
 ; CHECK-LABEL: partial
 define arm_aapcs_vfpcc void @partial(i32* nocapture %C, i32* nocapture readonly %A, i32* nocapture readonly %B) local_unnamed_addr #0 {
 entry:
   br label %for.body
 
 ; CHECK-LABEL: for.body
 for.body:
 
-; CHECK-UNROLL-V7: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, %entry ], [ [[IV2:%[a-z.0-9]+]], %for.body ]
-; CHECK-UNROLL-V7: [[IV1:%[a-z.0-9]+]] = add nuw nsw i32 [[IV0]], 1
-; CHECK-UNROLL-V7: [[IV2]] = add nuw nsw i32 [[IV1]], 1
-; CHECK-UNROLL-V7: [[CMP:%[a-z.0-9]+]] = icmp eq i32 [[IV2]], 1024
-; CHECK-UNROLL-V7: br i1 [[CMP]], label [[END:%[a-z.]+]], label %for.body
-
-; CHECK-UNROLL-SMALL: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, %entry ], [ [[IV8:%[a-z.0-9]+]], %for.body ]
-; CHECK-UNROLL-SMALL: [[IV1:%[a-z.0-9]+]] = add nuw nsw i32 [[IV0]], 1
-; CHECK-UNROLL-SMALL: [[IV2:%[a-z.0-9]+]] = add nuw nsw i32 [[IV1]], 1
-; CHECK-UNROLL-SMALL: [[IV3:%[a-z.0-9]+]] = add nuw nsw i32 [[IV2]], 1
-; CHECK-UNROLL-SMALL: [[IV4:%[a-z.0-9]+]] = add nuw nsw i32 [[IV3]], 1
-; CHECK-UNROLL-SMALL: [[IV5:%[a-z.0-9]+]] = add nuw nsw i32 [[IV4]], 1
-; CHECK-UNROLL-SMALL: [[IV6:%[a-z.0-9]+]] = add nuw nsw i32 [[IV5]], 1
-; CHECK-UNROLL-SMALL: [[IV7:%[a-z.0-9]+]] = add nuw nsw i32 [[IV6]], 1
-; CHECK-UNROLL-SMALL: [[IV8]] = add nuw nsw i32 [[IV7]], 1
-; CHECK-UNROLL-SMALL: [[CMP:%[a-z.0-9]+]] = icmp eq i32 [[IV8]], 1024
-; CHECK-UNROLL-SMALL: br i1 [[CMP]], label [[END:%[a-z.]+]], label %for.body
-
-; CHECK-UNROLL: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, %entry ], [ [[IV16:%[a-z.0-9]+]], %for.body ]
-; CHECK-UNROLL: [[IV1:%[a-z.0-9]+]] = add nuw nsw i32 [[IV0]], 1
-; CHECK-UNROLL: [[IV2:%[a-z.0-9]+]] = add nuw nsw i32 [[IV1]], 1
-; CHECK-UNROLL: [[IV3:%[a-z.0-9]+]] = add nuw nsw i32 [[IV2]], 1
-; CHECK-UNROLL: [[IV4:%[a-z.0-9]+]] = add nuw nsw i32 [[IV3]], 1
-; CHECK-UNROLL: [[IV5:%[a-z.0-9]+]] = add nuw nsw i32 [[IV4]], 1
-; CHECK-UNROLL: [[IV6:%[a-z.0-9]+]] = add nuw nsw i32 [[IV5]], 1
-; CHECK-UNROLL: [[IV7:%[a-z.0-9]+]] = add nuw nsw i32 [[IV6]], 1
-; CHECK-UNROLL: [[IV8:%[a-z.0-9]+]] = add nuw nsw i32 [[IV7]], 1
-; CHECK-UNROLL: [[IV9:%[a-z.0-9]+]] = add nuw nsw i32 [[IV8]], 1
-; CHECK-UNROLL: [[IV10:%[a-z.0-9]+]] = add nuw nsw i32 [[IV9]], 1
-; CHECK-UNROLL: [[IV11:%[a-z.0-9]+]] = add nuw nsw i32 [[IV10]], 1
-; CHECK-UNROLL: [[IV12:%[a-z.0-9]+]] = add nuw nsw i32 [[IV11]], 1
-; CHECK-UNROLL: [[IV13:%[a-z.0-9]+]] = add nuw nsw i32 [[IV12]], 1
-; CHECK-UNROLL: [[IV14:%[a-z.0-9]+]] = add nuw nsw i32 [[IV13]], 1
-; CHECK-UNROLL: [[IV15:%[a-z.0-9]+]] = add nuw nsw i32 [[IV14]], 1
-; CHECK-UNROLL: [[IV16]] = add nuw nsw i32 [[IV15]], 1
-; CHECK-UNROLL: [[CMP:%[a-z.0-9]+]] = icmp eq i32 [[IV16]], 1024
-; CHECK-UNROLL: br i1 [[CMP]], label [[END:%[a-z.]+]], label %for.body
+; CHECK-UNROLL-A: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, %entry ], [ [[IV2:%[a-z.0-9]+]], %for.body ]
+; CHECK-UNROLL-A: [[IV1:%[a-z.0-9]+]] = add nuw nsw i32 [[IV0]], 1
+; CHECK-UNROLL-A: [[IV2]] = add nuw nsw i32 [[IV1]], 1
+; CHECK-UNROLL-A: [[CMP:%[a-z.0-9]+]] = icmp eq i32 [[IV2]], 1024
+; CHECK-UNROLL-A: br i1 [[CMP]], label [[END:%[a-z.]+]], label %for.body
+
+; CHECK-UNROLL-T1: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, %entry ], [ [[IV1:%[a-z.0-9]+]], %for.body ]
+; CHECK-UNROLL-T1: [[IV1]] = add nuw nsw i32 [[IV0]], 1
+; CHECK-UNROLL-T1: [[CMP:%[a-z.0-9]+]] = icmp eq i32 [[IV1]], 1024
+; CHECK-UNROLL-T1: br i1 [[CMP]], label [[END:%[a-z.]+]], label %for.body
+
+; CHECK-UNROLL-T2: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, %entry ], [ [[IV16:%[a-z.0-9]+]], %for.body ]
+; CHECK-UNROLL-T2: [[IV1:%[a-z.0-9]+]] = add nuw nsw i32 [[IV0]], 1
+; CHECK-UNROLL-T2: [[IV2:%[a-z.0-9]+]] = add nuw nsw i32 [[IV1]], 1
+; CHECK-UNROLL-T2: [[IV3:%[a-z.0-9]+]] = add nuw nsw i32 [[IV2]], 1
+; CHECK-UNROLL-T2: [[IV4:%[a-z.0-9]+]] = add nuw nsw i32 [[IV3]], 1
+; CHECK-UNROLL-T2: [[IV5:%[a-z.0-9]+]] = add nuw nsw i32 [[IV4]], 1
+; CHECK-UNROLL-T2: [[IV6:%[a-z.0-9]+]] = add nuw nsw i32 [[IV5]], 1
+; CHECK-UNROLL-T2: [[IV7:%[a-z.0-9]+]] = add nuw nsw i32 [[IV6]], 1
+; CHECK-UNROLL-T2: [[IV8:%[a-z.0-9]+]] = add nuw nsw i32 [[IV7]], 1
+; CHECK-UNROLL-T2: [[IV9:%[a-z.0-9]+]] = add nuw nsw i32 [[IV8]], 1
+; CHECK-UNROLL-T2: [[IV10:%[a-z.0-9]+]] = add nuw nsw i32 [[IV9]], 1
+; CHECK-UNROLL-T2: [[IV11:%[a-z.0-9]+]] = add nuw nsw i32 [[IV10]], 1
+; CHECK-UNROLL-T2: [[IV12:%[a-z.0-9]+]] = add nuw nsw i32 [[IV11]], 1
+; CHECK-UNROLL-T2: [[IV13:%[a-z.0-9]+]] = add nuw nsw i32 [[IV12]], 1
+; CHECK-UNROLL-T2: [[IV14:%[a-z.0-9]+]] = add nuw nsw i32 [[IV13]], 1
+; CHECK-UNROLL-T2: [[IV15:%[a-z.0-9]+]] = add nuw nsw i32 [[IV14]], 1
+; CHECK-UNROLL-T2: [[IV16]] = add nuw nsw i32 [[IV15]], 1
+; CHECK-UNROLL-T2: [[CMP:%[a-z.0-9]+]] = icmp eq i32 [[IV16]], 1024
+; CHECK-UNROLL-T2: br i1 [[CMP]], label [[END:%[a-z.]+]], label %for.body
 
   %i.08 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
   %arrayidx = getelementptr inbounds i32, i32* %A, i32 %i.08
   %0 = load i32, i32* %arrayidx, align 4
   %arrayidx1 = getelementptr inbounds i32, i32* %B, i32 %i.08
   %1 = load i32, i32* %arrayidx1, align 4
   %mul = mul nsw i32 %1, %0
   %arrayidx2 = getelementptr inbounds i32, i32* %C, i32 %i.08

 ; CHECK-LABEL: runtime
 define arm_aapcs_vfpcc void @runtime(i32* nocapture %C, i32* nocapture readonly %A, i32* nocapture readonly %B, i32 %N) local_unnamed_addr #0 {
 entry:
   %cmp8 = icmp eq i32 %N, 0
   br i1 %cmp8, label %for.cond.cleanup, label %for.body
 
 ; CHECK-LABEL: for.body
 for.body:
-; CHECK-UNROLL-V7: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, [[PRE:%[a-z.0-9]+]] ], [ [[IV8:%[a-z.0-9]+]], %for.body ]
-; CHECK-UNROLL-V7: [[IV1:%[a-z.0-9]+]] = add nuw nsw i32 [[IV0]], 1
-; CHECK-UNROLL-V7: [[IV2]] = add nuw  i32 [[IV1]], 1
-
-; CHECK-UNROLL-SMALL: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, [[PRE:%[a-z.0-9]+]] ], [ [[IV8:%[a-z.0-9]+]], %for.body ]
-; CHECK-UNROLL-SMALL: [[IV1:%[a-z.0-9]+]] = add nuw nsw i32 [[IV0]], 1
-; CHECK-UNROLL-SMALL: [[IV2:%[a-z.0-9]+]] = add nuw nsw i32 [[IV1]], 1
-; CHECK-UNROLL-SMALL: [[IV3:%[a-z.0-9]+]] = add nuw nsw i32 [[IV2]], 1
-; CHECK-UNROLL-SMALL: [[IV4:%[a-z.0-9]+]] = add nuw nsw i32 [[IV3]], 1
-; CHECK-UNROLL-SMALL: [[IV5:%[a-z.0-9]+]] = add nuw nsw i32 [[IV4]], 1
-; CHECK-UNROLL-SMALL: [[IV6:%[a-z.0-9]+]] = add nuw nsw i32 [[IV5]], 1
-; CHECK-UNROLL-SMALL: [[IV7:%[a-z.0-9]+]] = add nuw nsw i32 [[IV6]], 1
-; CHECK-UNROLL-SMALL: [[IV8]] = add nuw i32 [[IV7]], 1
-
-; CHECK-UNROLL: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, [[PRE:%[a-z.0-9]+]] ], [ [[IV8:%[a-z.0-9]+]], %for.body ]
-; CHECK-UNROLL: [[IV1:%[a-z.0-9]+]] = add nuw nsw i32 [[IV0]], 1
-; CHECK-UNROLL: [[IV2:%[a-z.0-9]+]] = add nuw nsw i32 [[IV1]], 1
-; CHECK-UNROLL: [[IV3:%[a-z.0-9]+]] = add nuw nsw i32 [[IV2]], 1
-; CHECK-UNROLL: [[IV4:%[a-z.0-9]+]] = add nuw nsw i32 [[IV3]], 1
-; CHECK-UNROLL: [[IV5:%[a-z.0-9]+]] = add nuw nsw i32 [[IV4]], 1
-; CHECK-UNROLL: [[IV6:%[a-z.0-9]+]] = add nuw nsw i32 [[IV5]], 1
-; CHECK-UNROLL: [[IV7:%[a-z.0-9]+]] = add nuw nsw i32 [[IV6]], 1
-; CHECK-UNROLL: [[IV8]] = add nuw i32 [[IV7]], 1
+; CHECK-UNROLL-A: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, [[PRE:%[a-z.0-9]+]] ], [ [[IV2:%[a-z.0-9]+]], %for.body ]
+; CHECK-UNROLL-A: [[IV1:%[a-z.0-9]+]] = add nuw nsw i32 [[IV0]], 1
+; CHECK-UNROLL-A: [[IV2]] = add nuw i32 [[IV1]], 1
+; CHECK-UNROLL-A: br
+
+; CHECK-UNROLL-T1: %i.09 = phi i32 [ %inc, %for.body ], [ 0
+; CHECK-UNROLL-T1: %inc = add nuw i32 %i.09, 1
+; CHECK-UNROLL-T1: %exitcond = icmp eq i32 %inc, %N
+; CHECK-UNROLL-T1: br
+
+; CHECK-UNROLL-T2: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, [[PRE:%[a-z.0-9]+]] ], [ [[IV4:%[a-z.0-9]+]], %for.body ]
+; CHECK-UNROLL-T2: [[IV1:%[a-z.0-9]+]] = add nuw nsw i32 [[IV0]], 1
+; CHECK-UNROLL-T2: [[IV2:%[a-z.0-9]+]] = add nuw nsw i32 [[IV1]], 1
+; CHECK-UNROLL-T2: [[IV3:%[a-z.0-9]+]] = add nuw nsw i32 [[IV2]], 1
+; CHECK-UNROLL-T2: [[IV4]] = add nuw i32 [[IV3]], 1
+; CHECK-UNROLL-T2: br
+
+; CHECK-UNROLL-T2: for.body.epil:
+; CHECK-UNROLL-T2: for.body.epil.1:
+; CHECK-UNROLL-T2: for.body.epil.2:
 
   %i.09 = phi i32 [ %inc, %for.body ], [ 0, %entry ]
   %arrayidx = getelementptr inbounds i32, i32* %A, i32 %i.09
   %0 = load i32, i32* %arrayidx, align 4
   %arrayidx1 = getelementptr inbounds i32, i32* %B, i32 %i.09
   %1 = load i32, i32* %arrayidx1, align 4
   %mul = mul nsw i32 %1, %0
   %arrayidx2 = getelementptr inbounds i32, i32* %C, i32 %i.09

 
 for.cond.cleanup3:
   %inc11 = add nuw i32 %h.026, 1
   %exitcond27 = icmp eq i32 %inc11, %N
   br i1 %exitcond27, label %for.cond.cleanup, label %for.body4.lr.ph
 
 ; CHECK-LABEL: for.body4
 for.body4:
-; CHECK-UNROLL-V7: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, [[PRE:%[a-z0-9.]+]] ], [ [[IV1:%[a-z.0-9]+]], %for.body4 ]
-; CHECK-UNROLL-V7: [[IV1:%[a-z.0-9]+]] = add nuw  i32 [[IV0]], 1
-
-; CHECK-UNROLL-SMALL: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, [[PRE:%[a-z0-9.]+]] ], [ [[IV4:%[a-z.0-9]+]], %for.body4 ]
-; CHECK-UNROLL-SMALL: [[IV1:%[a-z.0-9]+]] = add nuw nsw i32 [[IV0]], 1
-; CHECK-UNROLL-SMALL: [[IV2:%[a-z.0-9]+]] = add nuw nsw i32 [[IV1]], 1
-; CHECK-UNROLL-SMALL: [[IV3:%[a-z.0-9]+]] = add nuw nsw i32 [[IV2]], 1
-; CHECK-UNROLL-SMALL: [[IV4]] = add nuw i32 [[IV3]], 1
-
-; CHECK-UNROLL: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, [[PRE:%[a-z0-9.]+]] ], [ [[IV8:%[a-z.0-9]+]], %for.body4 ]
-; CHECK-UNROLL: [[IV1:%[a-z.0-9]+]] = add nuw nsw i32 [[IV0]], 1
-; CHECK-UNROLL: [[IV2:%[a-z.0-9]+]] = add nuw nsw i32 [[IV1]], 1
-; CHECK-UNROLL: [[IV3:%[a-z.0-9]+]] = add nuw nsw i32 [[IV2]], 1
-; CHECK-UNROLL: [[IV4:%[a-z.0-9]+]] = add nuw nsw i32 [[IV3]], 1
-; CHECK-UNROLL: [[IV5:%[a-z.0-9]+]] = add nuw nsw i32 [[IV4]], 1
-; CHECK-UNROLL: [[IV6:%[a-z.0-9]+]] = add nuw nsw i32 [[IV5]], 1
-; CHECK-UNROLL: [[IV7:%[a-z.0-9]+]] = add nuw nsw i32 [[IV6]], 1
-; CHECK-UNROLL: [[IV8:%[a-z.0-9]+]] = add nuw i32 [[IV7]], 1
+; CHECK-UNROLL-T1: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, [[PRE:%[a-z0-9.]+]] ], [ [[IV1:%[a-z.0-9]+]], %for.body4 ]
+; CHECK-UNROLL-T1: [[IV1]] = add nuw i32 [[IV0]], 1
+; CHECK-UNROLL-T1: br
+
+; CHECK-UNROLL-T2: for.body4.epil:
+; CHECK-UNROLL-T2: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, [[PRE:%[a-z0-9.]+]] ], [ [[IV4:%[a-z.0-9]+]], %for.body4 ]
+; CHECK-UNROLL-T2: [[IV1:%[a-z.0-9]+]] = add nuw nsw i32 [[IV0]], 1
+; CHECK-UNROLL-T2: [[IV2:%[a-z.0-9]+]] = add nuw nsw i32 [[IV1]], 1
+; CHECK-UNROLL-T2: [[IV3:%[a-z.0-9]+]] = add nuw nsw i32 [[IV2]], 1
+; CHECK-UNROLL-T2: [[IV4]] = add nuw i32 [[IV3]], 1
+; CHECK-UNROLL-T2: br
+; CHECK-UNROLL-T2: for.body4.epil.1:
+; CHECK-UNROLL-T2: for.body4.epil.2:
 
   %w.024 = phi i32 [ 0, %for.body4.lr.ph ], [ %inc, %for.body4 ]
   %add = add i32 %w.024, %mul
   %arrayidx = getelementptr inbounds i16, i16* %A, i32 %add
   %0 = load i16, i16* %arrayidx, align 2
   %conv = sext i16 %0 to i32
   %arrayidx5 = getelementptr inbounds i16, i16* %B, i32 %w.024
   %1 = load i16, i16* %arrayidx5, align 2

 entry:
   br label %for.body
 
 for.cond.cleanup:
   ret void
 
 ; CHECK-LABEL: for.body
 for.body:
-; CHECK-UNROLL-V7: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, %entry ], [ [[IV1:%[a-z.0-9]+]], %for.body ]
-; CHECK-UNROLL-V7: [[IV1]] = add nuw nsw i32 [[IV0]], 1
-; CHECK-UNROLL-V7: icmp eq i32 [[IV1]], 1024
-
-; CHECK-UNROLL-SMALL: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, %entry ], [ [[IV1:%[a-z.0-9]+]], %for.body ]
-; CHECK-UNROLL-SMALL: [[IV1]] = add nuw nsw i32 [[IV0]], 1
-; CHECK-UNROLL-SMALL: icmp eq i32 [[IV1]], 1024
-
-; CHECK-UNROLL: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, %entry ], [ [[IV1:%[a-z.0-9]+]], %for.body ]
-; CHECK-UNROLL: [[IV1]] = add nuw nsw i32 [[IV0]], 1
-; CHECK-UNROLL: icmp eq i32 [[IV1]], 1024
+; CHECK-UNROLL-A: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, %entry ], [ [[IV1:%[a-z.0-9]+]], %for.body ]
+; CHECK-UNROLL-A: [[IV1]] = add nuw nsw i32 [[IV0]], 1
+; CHECK-UNROLL-A: icmp eq i32 [[IV1]], 1024
+; CHECK-UNROLL-A: br
+
+; CHECK-UNROLL-T1: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, %entry ], [ [[IV1:%[a-z.0-9]+]], %for.body ]
+; CHECK-UNROLL-T1: [[IV1]] = add nuw nsw i32 [[IV0]], 1
+; CHECK-UNROLL-T1: icmp eq i32 [[IV1]], 1024
+; CHECK-UNROLL-T1: br
+
+; CHECK-UNROLL-T2: [[IV0:%[a-z.0-9]+]] = phi i32 [ 0, %entry ], [ [[IV1:%[a-z.0-9]+]], %for.body ]
+; CHECK-UNROLL-T2: [[IV1]] = add nuw nsw i32 [[IV0]], 1
+; CHECK-UNROLL-T2: icmp eq i32 [[IV1]], 1024
+; CHECK-UNROLL-T2: br
 
   %i.08 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
   %arrayidx = getelementptr inbounds i32, i32* %A, i32 %i.08
   %0 = load i32, i32* %arrayidx, align 4
   %arrayidx1 = getelementptr inbounds i32, i32* %B, i32 %i.08
   %1 = load i32, i32* %arrayidx1, align 4
   %call = tail call arm_aapcs_vfpcc i32 @some_func(i32 %0, i32 %1) #3
   %arrayidx2 = getelementptr inbounds i32, i32* %C, i32 %i.08
   store i32 %call, i32* %arrayidx2, align 4
   %inc = add nuw nsw i32 %i.08, 1
   %exitcond = icmp eq i32 %inc, 1024
   br i1 %exitcond, label %for.cond.cleanup, label %for.body
 }
 
+; CHECK-LABEL: iterate_inc
+; CHECK-UNROLL-A: %n.addr.04 = phi %struct.Node* [ %1, %while.body ], [ %n, %while.body.preheader ]
+; CHECK-UNROLL-A: %tobool = icmp eq %struct.Node* %1, null
+; CHECK-UNROLL-A: br i1 %tobool
+; CHECK-UNROLL-A-NOT: load
+
+; CHECK-UNROLL-T1: %n.addr.04 = phi %struct.Node* [ %1, %while.body ], [ %n, %while.body.preheader ]
+; CHECK-UNROLL-T1: %tobool = icmp eq %struct.Node* %1, null
+; CHECK-UNROLL-T1: br i1 %tobool
+; CHECK-UNROLL-T1-NOT: load
+
+; CHECK-UNROLL-T2: [[CMP0:%[a-z.0-9]+]] = icmp eq %struct.Node* [[VAR0:%[a-z.0-9]+]], null
+; CHECK-UNROLL-T2: br i1 [[CMP0]], label [[END:%[a-z.0-9]+]]
+; CHECK-UNROLL-T2: [[CMP1:%[a-z.0-9]+]] = icmp eq %struct.Node* [[VAR1:%[a-z.0-9]+]], null
+; CHECK-UNROLL-T2: br i1 [[CMP1]], label [[END]]
+; CHECK-UNROLL-T2: [[CMP2:%[a-z.0-9]+]] = icmp eq %struct.Node* [[VAR2:%[a-z.0-9]+]], null
+; CHECK-UNROLL-T2: br i1 [[CMP2]], label [[END]]
+; CHECK-UNROLL-T2: [[CMP3:%[a-z.0-9]+]] = icmp eq %struct.Node* [[VAR3:%[a-z.0-9]+]], null
+; CHECK-UNROLL-T2: br i1 [[CMP3]], label [[END]]
+; CHECK-UNROLL-T2: [[CMP4:%[a-z.0-9]+]] = icmp eq %struct.Node* [[VAR4:%[a-z.0-9]+]], null
+; CHECK-UNROLL-T2: br i1 [[CMP4]], label [[END]]
+; CHECK-UNROLL-T2-NOT: load
+
+%struct.Node = type { %struct.Node*, i32 }
+
+define arm_aapcscc void @iterate_inc(%struct.Node* %n) local_unnamed_addr #0 {
+entry:
+  %tobool3 = icmp eq %struct.Node* %n, null
+  br i1 %tobool3, label %while.end, label %while.body.preheader
+
+while.body.preheader:
+  br label %while.body
+
+while.body:
+  %n.addr.04 = phi %struct.Node* [ %1, %while.body ], [ %n, %while.body.preheader ]
+  %val = getelementptr inbounds %struct.Node, %struct.Node* %n.addr.04, i32 0, i32 1
+  %0 = load i32, i32* %val, align 4
+  %add = add nsw i32 %0, 1
+  store i32 %add, i32* %val, align 4
+  %next = getelementptr inbounds %struct.Node, %struct.Node* %n.addr.04, i32 0, i32 0
+  %1 = load %struct.Node*, %struct.Node** %next, align 4
+  %tobool = icmp eq %struct.Node* %1, null
+  br i1 %tobool, label %while.end, label %while.body
+
+while.end:
+  ret void
+}
+
 declare arm_aapcs_vfpcc i32 @some_func(i32, i32) local_unnamed_addr #2