[LoopUnroll] Prevent LoopFullUnrollPass to perform partial/runtime unrolling

[yassingh]

Summary

FullLoopUnroll was performing runtime unrolling in certain cases when
'#pragma unroll' was specified. Patch to fix this by introducing new parameter
to tryToUnrollLoop() to differentiate between LoopUnrollPass and
FullLoopUnrollPass. Based on the discussion here
(https://discourse.llvm.org/t/loop-unroller-fails-to-unroll-loop/69834)

[yassingh]

LLVM :: Transforms/LoopUnroll/full-unroll-keep-first-exit.ll
LLVM :: Transforms/LoopUnroll/revisit.ll

These are the 2 tests that I think should be updated.

[nikic]

llvm/lib/Transforms/Scalar/LoopUnrollPass.cpp

1127

isFullUnroll -> OnlyFullUnroll?

1309

This probably isn't quite the right condition. I can think of two problematic cases here:

• We know the trip count, but it's very large, so we will only perform a partial unroll. We don't want to do that during FullUnroll.
• We know the max trip count and unroll to the max trip count, removing the loop. I think we consider this a full unroll as well.

I believe our criterion for what counts as a full unroll is basically "is the unroll count >= max trip count".

1311

parital -> partial

[yassingh]

• Addressed review comments

[nikic]

LGTM, assuming this passes tests.

[arsenm]

llvm/lib/Transforms/Scalar/LoopUnrollPass.cpp

1312

No reason to break the newline out of the string unless you are going to print additional information before it

[yassingh]

• updated failing tests
• fixed debug string

[yassingh]

llvm/test/Transforms/LoopUnroll/revisit.ll

11

How can I add an appropriate test for this option "unroll-revisit-child-loops"?

80–156

Removed this test, as full-unroll no longer performs partial unrolling when forced.

[nikic]

llvm/test/Transforms/LoopUnroll/full-unroll-keep-first-exit.ll

3

Can you please restore this check line, possibly with an additional prefix?

llvm/test/Transforms/LoopUnroll/revisit.ll

11

I think you can just drop the option entirely, it no longer makes sense. We can probably assert(!IsCurrentLoopValid) now: Full unrolling should always invalidate the current loop.

[yassingh]

• Review comments

[yassingh]

llvm/test/Transforms/LoopUnroll/revisit.ll

11

Removed this run line. I can submit another change to entirely drop this cl option.

[nikic]

llvm/test/Transforms/LoopUnroll/full-unroll-keep-first-exit.ll

3

This should use --check-prefixes=CHECK,CHECK-FULL to deduplicate the cases that are the same between full/non-full.

llvm/test/Transforms/LoopUnroll/revisit.ll

11

That's fine.

[yassingh]

• Updated test, didn't need to update the check prefix. Passing now.

[nikic]

LGTM

llvm/test/Transforms/LoopUnroll/full-unroll-keep-first-exit.ll

436

Spurious change. Sounds like you can just leave this test alone.

llvm/lib/Transforms/Scalar/LoopUnrollPass.cpp

#endif
   return ExplicitUnroll;
 }
 
 static LoopUnrollResult
 tryToUnrollLoop(Loop *L, DominatorTree &DT, LoopInfo *LI, ScalarEvolution &SE,
                 const TargetTransformInfo &TTI, AssumptionCache &AC,
                 OptimizationRemarkEmitter &ORE, BlockFrequencyInfo *BFI,
                 ProfileSummaryInfo *PSI, bool PreserveLCSSA, int OptLevel,
-                bool OnlyWhenForced, bool ForgetAllSCEV,
+                bool OnlyFullUnroll, bool OnlyWhenForced, bool ForgetAllSCEV,
                 std::optional<unsigned> ProvidedCount,
                 std::optional<unsigned> ProvidedThreshold,
                 std::optional<bool> ProvidedAllowPartial,
                 std::optional<bool> ProvidedRuntime,
                 std::optional<bool> ProvidedUpperBound,
                 std::optional<bool> ProvidedAllowPeeling,
                 std::optional<bool> ProvidedAllowProfileBasedPeeling,
                 std::optional<unsigned> ProvidedFullUnrollMaxCount) {
+
   LLVM_DEBUG(dbgs() << "Loop Unroll: F["
                     << L->getHeader()->getParent()->getName() << "] Loop %"
                     << L->getHeader()->getName() << "\n");
   TransformationMode TM = hasUnrollTransformation(L);
   if (TM & TM_Disable)
     return LoopUnrollResult::Unmodified;
 
   // If this loop isn't forced to be unrolled, avoid unrolling it when the
if (peelLoop(L, PP.PeelCount, LI, &SE, DT, &AC, PreserveLCSSA, VMap)) {
       // we had, so we don't want to unroll or peel again.
       if (PP.PeelProfiledIterations)
         L->setLoopAlreadyUnrolled();
       return LoopUnrollResult::PartiallyUnrolled;
     }
     return LoopUnrollResult::Unmodified;
   }
 
+  // Do not attempt partial/runtime unrolling in FullLoopUnrolling
+  if (OnlyFullUnroll && !(UP.Count >= MaxTripCount)) {
+    LLVM_DEBUG(
+        dbgs() << "Not attempting partial/runtime unroll in FullLoopUnroll.\n");
+    return LoopUnrollResult::Unmodified;
+  }
+
   // At this point, UP.Runtime indicates that run-time unrolling is allowed.
   // However, we only want to actually perform it if we don't know the trip
   // count and the unroll count doesn't divide the known trip multiple.
   // TODO: This decision should probably be pushed up into
   // computeUnrollCount().
   UP.Runtime &= TripCount == 0 && TripMultiple % UP.Count != 0;
 
   // Save loop properties before it is transformed.
bool runOnLoop(Loop *L, LPPassManager &LPM) override {
     // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
     // pass.  Function analyses need to be preserved across loop transformations
     // but ORE cannot be preserved (see comment before the pass definition).
     OptimizationRemarkEmitter ORE(&F);
     bool PreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
 
     LoopUnrollResult Result = tryToUnrollLoop(
         L, DT, LI, SE, TTI, AC, ORE, nullptr, nullptr, PreserveLCSSA, OptLevel,
-        OnlyWhenForced, ForgetAllSCEV, ProvidedCount, ProvidedThreshold,
-        ProvidedAllowPartial, ProvidedRuntime, ProvidedUpperBound,
-        ProvidedAllowPeeling, ProvidedAllowProfileBasedPeeling,
-        ProvidedFullUnrollMaxCount);
+        /*OnlyFullUnroll*/ false, OnlyWhenForced, ForgetAllSCEV, ProvidedCount,
+        ProvidedThreshold, ProvidedAllowPartial, ProvidedRuntime,
+        ProvidedUpperBound, ProvidedAllowPeeling,
+        ProvidedAllowProfileBasedPeeling, ProvidedFullUnrollMaxCount);
 
     if (Result == LoopUnrollResult::FullyUnrolled)
       LPM.markLoopAsDeleted(*L);
 
     return Result != LoopUnrollResult::Unmodified;
   }
 
   /// This transformation requires natural loop information & requires that
PreservedAnalyses LoopFullUnrollPass::run(Loop &L, LoopAnalysisManager &AM,
   else
     OldLoops.insert(AR.LI.begin(), AR.LI.end());
 
   std::string LoopName = std::string(L.getName());
 
   bool Changed =
       tryToUnrollLoop(&L, AR.DT, &AR.LI, AR.SE, AR.TTI, AR.AC, ORE,
                       /*BFI*/ nullptr, /*PSI*/ nullptr,
-                      /*PreserveLCSSA*/ true, OptLevel, OnlyWhenForced,
-                      ForgetSCEV, /*Count*/ std::nullopt,
+                      /*PreserveLCSSA*/ true, OptLevel, /*OnlyFullUnroll*/ true,
+                      OnlyWhenForced, ForgetSCEV, /*Count*/ std::nullopt,
                       /*Threshold*/ std::nullopt, /*AllowPartial*/ false,
                       /*Runtime*/ false, /*UpperBound*/ false,
                       /*AllowPeeling*/ true,
                       /*AllowProfileBasedPeeling*/ false,
                       /*FullUnrollMaxCount*/ std::nullopt) !=
       LoopUnrollResult::Unmodified;
   if (!Changed)
     return PreservedAnalyses::all();
#endif
     std::optional<bool> LocalAllowPeeling = UnrollOpts.AllowPeeling;
     if (PSI && PSI->hasHugeWorkingSetSize())
       LocalAllowPeeling = false;
     std::string LoopName = std::string(L.getName());
     // The API here is quite complex to call and we allow to select some
     // flavors of unrolling during construction time (by setting UnrollOpts).
     LoopUnrollResult Result = tryToUnrollLoop(
         &L, DT, &LI, SE, TTI, AC, ORE, BFI, PSI,
-        /*PreserveLCSSA*/ true, UnrollOpts.OptLevel, UnrollOpts.OnlyWhenForced,
-        UnrollOpts.ForgetSCEV, /*Count*/ std::nullopt,
+        /*PreserveLCSSA*/ true, UnrollOpts.OptLevel, /*OnlyFullUnroll*/ false,
+        UnrollOpts.OnlyWhenForced, UnrollOpts.ForgetSCEV,
+        /*Count*/ std::nullopt,
         /*Threshold*/ std::nullopt, UnrollOpts.AllowPartial,
         UnrollOpts.AllowRuntime, UnrollOpts.AllowUpperBound, LocalAllowPeeling,
         UnrollOpts.AllowProfileBasedPeeling, UnrollOpts.FullUnrollMaxCount);
     Changed |= Result != LoopUnrollResult::Unmodified;
 
     // The parent must not be damaged by unrolling!
 #ifndef NDEBUG
     if (Result != LoopUnrollResult::Unmodified && ParentL)

llvm/test/Transforms/LoopUnroll/full-unroll-avoid-partial.ll

+; RUN: opt -S -passes=loop-unroll --debug-only=loop-unroll < %s 2>&1 | FileCheck %s -check-prefix=LOOP-UNROLL
+; RUN: opt -S -passes='require<opt-remark-emit>,loop(loop-unroll-full)' --debug-only=loop-unroll < %s 2>&1 | FileCheck %s -check-prefix=LOOP-UNROLL-FULL
+
+; REQUIRES: asserts
+
+%struct.HIP_vector_type = type {  %union.anon }
+%union.anon = type { <2 x float> }
+
+
+; LOOP-UNROLL-LABEL: Loop Unroll: F[pragma_unroll] Loop %for.body
+; LOOP-UNROLL-NEXT: Loop Size = 9
+; LOOP-UNROLL-NEXT: runtime unrolling with count: 8
+; LOOP-UNROLL-NEXT: Exiting block %for.body: TripCount=0, TripMultiple=1, BreakoutTrip=1
+; LOOP-UNROLL-NEXT: Trying runtime unrolling on Loop:
+; LOOP-UNROLL-NEXT: Loop at depth 1 containing: %for.body<header><latch><exiting>
+; LOOP-UNROLL-NEXT: Using epilog remainder.
+; LOOP-UNROLL-NEXT: UNROLLING loop %for.body by 8 with run-time trip count!
+
+; LOOP-UNROLL-FULL-LABEL: Loop Unroll: F[pragma_unroll] Loop %for.body
+; LOOP-UNROLL-FULL-NEXT: Loop Size = 9
+; LOOP-UNROLL-FULL-NEXT:  runtime unrolling with count: 8
+; LOOP-UNROLL-FULL-NEXT: Not attempting partial/runtime unroll in FullLoopUnroll
+define void @pragma_unroll(ptr %queue, i32 %num_elements) {
+entry:
+  %cmp5 = icmp sgt i32 %num_elements, 0
+  br i1 %cmp5, label %for.body.preheader, label %for.cond.cleanup
+
+for.body.preheader:                               ; preds = %entry
+  br label %for.body
+
+for.cond.cleanup.loopexit:                        ; preds = %for.body
+  br label %for.cond.cleanup
+
+for.cond.cleanup:                                 ; preds = %for.cond.cleanup.loopexit, %entry
+  ret void
+
+for.body:                                         ; preds = %for.body.preheader, %for.body
+  %i.06 = phi i32 [ %add, %for.body ], [ 0, %for.body.preheader ]
+  %add = add nuw nsw i32 %i.06, 1
+  %idxprom = zext i32 %add to i64
+  %arrayidx = getelementptr inbounds %struct.HIP_vector_type, ptr %queue, i64 %idxprom
+  %idxprom1 = zext i32 %i.06 to i64
+  %arrayidx2 = getelementptr inbounds %struct.HIP_vector_type, ptr %queue, i64 %idxprom1
+  %0 = load i64, ptr %arrayidx, align 8
+  store i64 %0, ptr %arrayidx2, align 8
+  %exitcond = icmp ne i32 %add, %num_elements
+  br i1 %exitcond, label %for.body, label %for.cond.cleanup.loopexit, !llvm.loop !1
+}
+
+; LOOP-UNROLL-LABEL: Loop Unroll: F[pragma_unroll_count1] Loop %for.body
+; LOOP-UNROLL-NEXT: Loop Size = 9
+; LOOP-UNROLL-NEXT: Exiting block %for.body: TripCount=0, TripMultiple=1, BreakoutTrip=1
+; LOOP-UNROLL-NEXT: Trying runtime unrolling on Loop:
+; LOOP-UNROLL-NEXT: Loop at depth 1 containing: %for.body<header><latch><exiting>
+; LOOP-UNROLL-NEXT: Using epilog remainder.
+; LOOP-UNROLL-NEXT: UNROLLING loop %for.body by 5 with run-time trip count!
+
+; LOOP-UNROLL-FULL-LABEL: Loop Unroll: F[pragma_unroll_count1] Loop %for.body
+; LOOP-UNROLL-FULL-NEXT: Loop Size = 9
+; LOOP-UNROLL-FULL-NEXT: Not attempting partial/runtime unroll in FullLoopUnroll
+define void @pragma_unroll_count1(ptr %queue, i32 %num_elements) {
+entry:
+  %cmp5 = icmp sgt i32 %num_elements, 0
+  br i1 %cmp5, label %for.body.preheader, label %for.cond.cleanup
+
+for.body.preheader:                               ; preds = %entry
+  br label %for.body
+
+for.cond.cleanup.loopexit:                        ; preds = %for.body
+  br label %for.cond.cleanup
+
+for.cond.cleanup:                                 ; preds = %for.cond.cleanup.loopexit, %entry
+  ret void
+
+for.body:                                         ; preds = %for.body.preheader, %for.body
+  %i.06 = phi i32 [ %add, %for.body ], [ 0, %for.body.preheader ]
+  %add = add nuw nsw i32 %i.06, 1
+  %idxprom = zext i32 %add to i64
+  %arrayidx = getelementptr inbounds %struct.HIP_vector_type, ptr %queue, i64 %idxprom
+  %idxprom1 = zext i32 %i.06 to i64
+  %arrayidx2 = getelementptr inbounds %struct.HIP_vector_type, ptr %queue, i64 %idxprom1
+  %0 = load i64, ptr %arrayidx, align 8
+  store i64 %0, ptr %arrayidx2, align 8
+  %exitcond = icmp ne i32 %add, %num_elements
+  br i1 %exitcond, label %for.body, label %for.cond.cleanup.loopexit, !llvm.loop !3
+}
+
+; LOOP-UNROLL: llvm.loop.unroll.disable
+; LOOP-UNROLL-FULL: llvm.loop.unroll.enable
+!0 = !{!"llvm.loop.unroll.enable"}
+!1 = distinct !{!1, !0}
+
+!2 = !{!"llvm.loop.unroll.count", i32 5}
+!3 = distinct !{!3, !2}

llvm/test/Transforms/LoopUnroll/full-unroll-keep-first-exit.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt -S -passes=loop-unroll < %s | FileCheck %s
-; RUN: opt -S -passes='require<opt-remark-emit>,loop(loop-unroll-full)' < %s | FileCheck %s
+; RUN: opt -S -passes='require<opt-remark-emit>,loop(loop-unroll-full)' < %s | FileCheck %s -check-prefix=CHECK-FULL
 
 ; Unroll twice, with first loop exit kept
 define void @s32_max1(i32 %n, ptr %p) {
 ;
 ; CHECK-LABEL: @s32_max1(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[ADD:%.*]] = add i32 [[N:%.*]], 1
 ; CHECK-NEXT:    br label [[DO_BODY:%.*]]
 ; CHECK:       do.body:
 ; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 [[N]]
 ; CHECK-NEXT:    store i32 [[N]], ptr [[ARRAYIDX]], align 4
 ; CHECK-NEXT:    [[INC:%.*]] = add i32 [[N]], 1
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i32 [[N]], [[ADD]]
 ; CHECK-NEXT:    br i1 [[CMP]], label [[DO_BODY_1:%.*]], label [[DO_END:%.*]]
 ; CHECK:       do.body.1:
 ; CHECK-NEXT:    [[ARRAYIDX_1:%.*]] = getelementptr i32, ptr [[P]], i32 [[INC]]
 ; CHECK-NEXT:    store i32 [[INC]], ptr [[ARRAYIDX_1]], align 4
 ; CHECK-NEXT:    br label [[DO_END]]
 ; CHECK:       do.end:
 ; CHECK-NEXT:    ret void
 ;
+; CHECK-FULL-LABEL: @s32_max1(
+; CHECK-FULL-NEXT:  entry:
+; CHECK-FULL-NEXT:    [[ADD:%.*]] = add i32 [[N:%.*]], 1
+; CHECK-FULL-NEXT:    br label [[DO_BODY:%.*]]
+; CHECK-FULL:       do.body:
+; CHECK-FULL-NEXT:    [[ARRAYIDX:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 [[N]]
+; CHECK-FULL-NEXT:    store i32 [[N]], ptr [[ARRAYIDX]], align 4
+; CHECK-FULL-NEXT:    [[INC:%.*]] = add i32 [[N]], 1
+; CHECK-FULL-NEXT:    [[CMP:%.*]] = icmp slt i32 [[N]], [[ADD]]
+; CHECK-FULL-NEXT:    br i1 [[CMP]], label [[DO_BODY_1:%.*]], label [[DO_END:%.*]]
+; CHECK-FULL:       do.body.1:
+; CHECK-FULL-NEXT:    [[ARRAYIDX_1:%.*]] = getelementptr i32, ptr [[P]], i32 [[INC]]
+; CHECK-FULL-NEXT:    store i32 [[INC]], ptr [[ARRAYIDX_1]], align 4
+; CHECK-FULL-NEXT:    br label [[DO_END]]
+; CHECK-FULL:       do.end:
+; CHECK-FULL-NEXT:    ret void
+;
 entry:
   %add = add i32 %n, 1
   br label %do.body
 
 do.body:
   %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
   %arrayidx = getelementptr i32, ptr %p, i32 %i.0
   store i32 %i.0, ptr %arrayidx, align 4

 ; CHECK-NEXT:    store i32 [[INC]], ptr [[ARRAYIDX_1]], align 4
 ; CHECK-NEXT:    [[INC_1:%.*]] = add i32 [[INC]], 1
 ; CHECK-NEXT:    [[ARRAYIDX_2:%.*]] = getelementptr i32, ptr [[P]], i32 [[INC_1]]
 ; CHECK-NEXT:    store i32 [[INC_1]], ptr [[ARRAYIDX_2]], align 4
 ; CHECK-NEXT:    br label [[DO_END]]
 ; CHECK:       do.end:
 ; CHECK-NEXT:    ret void
 ;
+; CHECK-FULL-LABEL: @s32_max2(
+; CHECK-FULL-NEXT:  entry:
+; CHECK-FULL-NEXT:    [[ADD:%.*]] = add i32 [[N:%.*]], 2
+; CHECK-FULL-NEXT:    br label [[DO_BODY:%.*]]
+; CHECK-FULL:       do.body:
+; CHECK-FULL-NEXT:    [[ARRAYIDX:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 [[N]]
+; CHECK-FULL-NEXT:    store i32 [[N]], ptr [[ARRAYIDX]], align 4
+; CHECK-FULL-NEXT:    [[INC:%.*]] = add i32 [[N]], 1
+; CHECK-FULL-NEXT:    [[CMP:%.*]] = icmp slt i32 [[N]], [[ADD]]
+; CHECK-FULL-NEXT:    br i1 [[CMP]], label [[DO_BODY_1:%.*]], label [[DO_END:%.*]]
+; CHECK-FULL:       do.body.1:
+; CHECK-FULL-NEXT:    [[ARRAYIDX_1:%.*]] = getelementptr i32, ptr [[P]], i32 [[INC]]
+; CHECK-FULL-NEXT:    store i32 [[INC]], ptr [[ARRAYIDX_1]], align 4
+; CHECK-FULL-NEXT:    [[INC_1:%.*]] = add i32 [[INC]], 1
+; CHECK-FULL-NEXT:    [[ARRAYIDX_2:%.*]] = getelementptr i32, ptr [[P]], i32 [[INC_1]]
+; CHECK-FULL-NEXT:    store i32 [[INC_1]], ptr [[ARRAYIDX_2]], align 4
+; CHECK-FULL-NEXT:    br label [[DO_END]]
+; CHECK-FULL:       do.end:
+; CHECK-FULL-NEXT:    ret void
+;
 entry:
   %add = add i32 %n, 2
   br label %do.body
 
 do.body:
   %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
   %arrayidx = getelementptr i32, ptr %p, i32 %i.0
   store i32 %i.0, ptr %arrayidx, align 4

 ; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 [[I_0]]
 ; CHECK-NEXT:    store i32 [[I_0]], ptr [[ARRAYIDX]], align 4
 ; CHECK-NEXT:    [[INC]] = add i32 [[I_0]], 1
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i32 [[I_0]], [[ADD]]
 ; CHECK-NEXT:    br i1 [[CMP]], label [[DO_BODY]], label [[DO_END:%.*]]
 ; CHECK:       do.end:
 ; CHECK-NEXT:    ret void
 ;
+; CHECK-FULL-LABEL: @s32_maxx(
+; CHECK-FULL-NEXT:  entry:
+; CHECK-FULL-NEXT:    [[ADD:%.*]] = add i32 [[X:%.*]], [[N:%.*]]
+; CHECK-FULL-NEXT:    br label [[DO_BODY:%.*]]
+; CHECK-FULL:       do.body:
+; CHECK-FULL-NEXT:    [[I_0:%.*]] = phi i32 [ [[N]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[DO_BODY]] ]
+; CHECK-FULL-NEXT:    [[ARRAYIDX:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 [[I_0]]
+; CHECK-FULL-NEXT:    store i32 [[I_0]], ptr [[ARRAYIDX]], align 4
+; CHECK-FULL-NEXT:    [[INC]] = add i32 [[I_0]], 1
+; CHECK-FULL-NEXT:    [[CMP:%.*]] = icmp slt i32 [[I_0]], [[ADD]]
+; CHECK-FULL-NEXT:    br i1 [[CMP]], label [[DO_BODY]], label [[DO_END:%.*]]
+; CHECK-FULL:       do.end:
+; CHECK-FULL-NEXT:    ret void
+;
 entry:
   %add = add i32 %x, %n
   br label %do.body
 
 do.body:
   %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
   %arrayidx = getelementptr i32, ptr %p, i32 %i.0
   store i32 %i.0, ptr %arrayidx, align 4

 ; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 [[I_0]]
 ; CHECK-NEXT:    store i32 [[I_0]], ptr [[ARRAYIDX]], align 4
 ; CHECK-NEXT:    [[INC]] = add i32 [[I_0]], 1
 ; CHECK-NEXT:    [[CMP1:%.*]] = icmp slt i32 [[I_0]], [[ADD]]
 ; CHECK-NEXT:    br i1 [[CMP1]], label [[DO_BODY]], label [[DO_END]]
 ; CHECK:       do.end:
 ; CHECK-NEXT:    ret void
 ;
+; CHECK-FULL-LABEL: @s32_max2_unpredictable_exit(
+; CHECK-FULL-NEXT:  entry:
+; CHECK-FULL-NEXT:    [[ADD:%.*]] = add i32 [[N:%.*]], 2
+; CHECK-FULL-NEXT:    br label [[DO_BODY:%.*]]
+; CHECK-FULL:       do.body:
+; CHECK-FULL-NEXT:    [[I_0:%.*]] = phi i32 [ [[N]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[IF_END:%.*]] ]
+; CHECK-FULL-NEXT:    [[CMP:%.*]] = icmp eq i32 [[I_0]], [[X:%.*]]
+; CHECK-FULL-NEXT:    br i1 [[CMP]], label [[DO_END:%.*]], label [[IF_END]]
+; CHECK-FULL:       if.end:
+; CHECK-FULL-NEXT:    [[ARRAYIDX:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 [[I_0]]
+; CHECK-FULL-NEXT:    store i32 [[I_0]], ptr [[ARRAYIDX]], align 4
+; CHECK-FULL-NEXT:    [[INC]] = add i32 [[I_0]], 1
+; CHECK-FULL-NEXT:    [[CMP1:%.*]] = icmp slt i32 [[I_0]], [[ADD]]
+; CHECK-FULL-NEXT:    br i1 [[CMP1]], label [[DO_BODY]], label [[DO_END]]
+; CHECK-FULL:       do.end:
+; CHECK-FULL-NEXT:    ret void
+;
 entry:
   %add = add i32 %n, 2
   br label %do.body
 
 do.body:
   %i.0 = phi i32 [ %n, %entry ], [ %inc, %if.end ]
   %cmp = icmp eq i32 %i.0, %x
   br i1 %cmp, label %do.end, label %if.end ; unpredictable

 ; CHECK-NEXT:    br i1 [[CMP]], label [[DO_BODY_1:%.*]], label [[DO_END:%.*]]
 ; CHECK:       do.body.1:
 ; CHECK-NEXT:    [[ARRAYIDX_1:%.*]] = getelementptr i32, ptr [[P]], i32 [[INC]]
 ; CHECK-NEXT:    store i32 [[INC]], ptr [[ARRAYIDX_1]], align 4
 ; CHECK-NEXT:    br label [[DO_END]]
 ; CHECK:       do.end:
 ; CHECK-NEXT:    ret void
 ;
+; CHECK-FULL-LABEL: @u32_max1(
+; CHECK-FULL-NEXT:  entry:
+; CHECK-FULL-NEXT:    [[ADD:%.*]] = add i32 [[N:%.*]], 1
+; CHECK-FULL-NEXT:    br label [[DO_BODY:%.*]]
+; CHECK-FULL:       do.body:
+; CHECK-FULL-NEXT:    [[ARRAYIDX:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 [[N]]
+; CHECK-FULL-NEXT:    store i32 [[N]], ptr [[ARRAYIDX]], align 4
+; CHECK-FULL-NEXT:    [[INC:%.*]] = add i32 [[N]], 1
+; CHECK-FULL-NEXT:    [[CMP:%.*]] = icmp ult i32 [[N]], [[ADD]]
+; CHECK-FULL-NEXT:    br i1 [[CMP]], label [[DO_BODY_1:%.*]], label [[DO_END:%.*]]
+; CHECK-FULL:       do.body.1:
+; CHECK-FULL-NEXT:    [[ARRAYIDX_1:%.*]] = getelementptr i32, ptr [[P]], i32 [[INC]]
+; CHECK-FULL-NEXT:    store i32 [[INC]], ptr [[ARRAYIDX_1]], align 4
+; CHECK-FULL-NEXT:    br label [[DO_END]]
+; CHECK-FULL:       do.end:
+; CHECK-FULL-NEXT:    ret void
+;
 entry:
   %add = add i32 %n, 1
   br label %do.body
 
 do.body:
   %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
   %arrayidx = getelementptr i32, ptr %p, i32 %i.0
   store i32 %i.0, ptr %arrayidx, align 4

 ; CHECK-NEXT:    store i32 [[INC]], ptr [[ARRAYIDX_1]], align 4
 ; CHECK-NEXT:    [[INC_1:%.*]] = add i32 [[INC]], 1
 ; CHECK-NEXT:    [[ARRAYIDX_2:%.*]] = getelementptr i32, ptr [[P]], i32 [[INC_1]]
 ; CHECK-NEXT:    store i32 [[INC_1]], ptr [[ARRAYIDX_2]], align 4
 ; CHECK-NEXT:    br label [[DO_END]]
 ; CHECK:       do.end:
 ; CHECK-NEXT:    ret void
 ;
+; CHECK-FULL-LABEL: @u32_max2(
+; CHECK-FULL-NEXT:  entry:
+; CHECK-FULL-NEXT:    [[ADD:%.*]] = add i32 [[N:%.*]], 2
+; CHECK-FULL-NEXT:    br label [[DO_BODY:%.*]]
+; CHECK-FULL:       do.body:
+; CHECK-FULL-NEXT:    [[ARRAYIDX:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 [[N]]
+; CHECK-FULL-NEXT:    store i32 [[N]], ptr [[ARRAYIDX]], align 4
+; CHECK-FULL-NEXT:    [[INC:%.*]] = add i32 [[N]], 1
+; CHECK-FULL-NEXT:    [[CMP:%.*]] = icmp ult i32 [[N]], [[ADD]]
+; CHECK-FULL-NEXT:    br i1 [[CMP]], label [[DO_BODY_1:%.*]], label [[DO_END:%.*]]
+; CHECK-FULL:       do.body.1:
+; CHECK-FULL-NEXT:    [[ARRAYIDX_1:%.*]] = getelementptr i32, ptr [[P]], i32 [[INC]]
+; CHECK-FULL-NEXT:    store i32 [[INC]], ptr [[ARRAYIDX_1]], align 4
+; CHECK-FULL-NEXT:    [[INC_1:%.*]] = add i32 [[INC]], 1
+; CHECK-FULL-NEXT:    [[ARRAYIDX_2:%.*]] = getelementptr i32, ptr [[P]], i32 [[INC_1]]
+; CHECK-FULL-NEXT:    store i32 [[INC_1]], ptr [[ARRAYIDX_2]], align 4
+; CHECK-FULL-NEXT:    br label [[DO_END]]
+; CHECK-FULL:       do.end:
+; CHECK-FULL-NEXT:    ret void
+;
 entry:
   %add = add i32 %n, 2
   br label %do.body
 
 do.body:
   %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
   %arrayidx = getelementptr i32, ptr %p, i32 %i.0
   store i32 %i.0, ptr %arrayidx, align 4

 ; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 [[I_0]]
 ; CHECK-NEXT:    store i32 [[I_0]], ptr [[ARRAYIDX]], align 4
 ; CHECK-NEXT:    [[INC]] = add i32 [[I_0]], 1
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i32 [[I_0]], [[ADD]]
 ; CHECK-NEXT:    br i1 [[CMP]], label [[DO_BODY]], label [[DO_END:%.*]]
 ; CHECK:       do.end:
 ; CHECK-NEXT:    ret void
 ;
+; CHECK-FULL-LABEL: @u32_maxx(
+; CHECK-FULL-NEXT:  entry:
+; CHECK-FULL-NEXT:    [[ADD:%.*]] = add i32 [[X:%.*]], [[N:%.*]]
+; CHECK-FULL-NEXT:    br label [[DO_BODY:%.*]]
+; CHECK-FULL:       do.body:
+; CHECK-FULL-NEXT:    [[I_0:%.*]] = phi i32 [ [[N]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[DO_BODY]] ]
+; CHECK-FULL-NEXT:    [[ARRAYIDX:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 [[I_0]]
+; CHECK-FULL-NEXT:    store i32 [[I_0]], ptr [[ARRAYIDX]], align 4
+; CHECK-FULL-NEXT:    [[INC]] = add i32 [[I_0]], 1
+; CHECK-FULL-NEXT:    [[CMP:%.*]] = icmp ult i32 [[I_0]], [[ADD]]
+; CHECK-FULL-NEXT:    br i1 [[CMP]], label [[DO_BODY]], label [[DO_END:%.*]]
+; CHECK-FULL:       do.end:
+; CHECK-FULL-NEXT:    ret void
+;
 entry:
   %add = add i32 %x, %n
   br label %do.body
 
 do.body:
   %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
   %arrayidx = getelementptr i32, ptr %p, i32 %i.0
   store i32 %i.0, ptr %arrayidx, align 4

 ; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 [[I_0]]
 ; CHECK-NEXT:    store i32 [[I_0]], ptr [[ARRAYIDX]], align 4
 ; CHECK-NEXT:    [[INC]] = add i32 [[I_0]], 1
 ; CHECK-NEXT:    [[CMP1:%.*]] = icmp ult i32 [[I_0]], [[ADD]]
 ; CHECK-NEXT:    br i1 [[CMP1]], label [[DO_BODY]], label [[DO_END]]
 ; CHECK:       do.end:
 ; CHECK-NEXT:    ret void
 ;
+; CHECK-FULL-LABEL: @u32_max2_unpredictable_exit(
+; CHECK-FULL-NEXT:  entry:
+; CHECK-FULL-NEXT:    [[ADD:%.*]] = add i32 [[N:%.*]], 2
+; CHECK-FULL-NEXT:    br label [[DO_BODY:%.*]]
+; CHECK-FULL:       do.body:
+; CHECK-FULL-NEXT:    [[I_0:%.*]] = phi i32 [ [[N]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[IF_END:%.*]] ]
+; CHECK-FULL-NEXT:    [[CMP:%.*]] = icmp eq i32 [[I_0]], [[X:%.*]]
+; CHECK-FULL-NEXT:    br i1 [[CMP]], label [[DO_END:%.*]], label [[IF_END]]
+; CHECK-FULL:       if.end:
+; CHECK-FULL-NEXT:    [[ARRAYIDX:%.*]] = getelementptr i32, ptr [[P:%.*]], i32 [[I_0]]
+; CHECK-FULL-NEXT:    store i32 [[I_0]], ptr [[ARRAYIDX]], align 4
+; CHECK-FULL-NEXT:    [[INC]] = add i32 [[I_0]], 1
+; CHECK-FULL-NEXT:    [[CMP1:%.*]] = icmp ult i32 [[I_0]], [[ADD]]
+; CHECK-FULL-NEXT:    br i1 [[CMP1]], label [[DO_BODY]], label [[DO_END]]
+; CHECK-FULL:       do.end:
+; CHECK-FULL-NEXT:    ret void
+;
 entry:
   %add = add i32 %n, 2
   br label %do.body
 
 do.body:
   %i.0 = phi i32 [ %n, %entry ], [ %inc, %if.end ]
   %cmp = icmp eq i32 %i.0, %x
   br i1 %cmp, label %do.end, label %if.end ; unpredictable
 
 if.end:
   %arrayidx = getelementptr i32, ptr %p, i32 %i.0
   store i32 %i.0, ptr %arrayidx, align 4
   %inc = add i32 %i.0, 1
   %cmp1 = icmp ult i32 %i.0, %add
   br i1 %cmp1, label %do.body, label %do.end ; taken either 0 or 2 times
 
 do.end:
   ret void
 }

llvm/test/Transforms/LoopUnroll/revisit.ll

 ; This test checks that nested loops are revisited in various scenarios when
 ; unrolling. Note that if we ever start doing outer loop peeling a test case
-; for that should be added here that will look essentially like a hybrid of the
-; current two cases.
+; for that should be added here.
 ;
 ; RUN: opt < %s -disable-output -debug-pass-manager 2>&1 \
 ; RUN: -passes='require<opt-remark-emit>,loop(loop-unroll-full)' \
 ; RUN:     | FileCheck %s
-;
-; Also run in a special mode that visits children.
-; RUN: opt < %s -disable-output -debug-pass-manager -unroll-revisit-child-loops 2>&1 \
-; RUN: -passes='require<opt-remark-emit>,loop(loop-unroll-full)' \
-; RUN:     | FileCheck %s --check-prefixes=CHECK,CHECK-CHILDREN
-
+; 
 ; Basic test is fully unrolled and we revisit the post-unroll new sibling
 ; loops, including the ones that used to be child loops.
 define void @full_unroll(ptr %ptr) {
 ; CHECK-LABEL: OptimizationRemarkEmitterAnalysis on full_unroll
 ; CHECK-NOT: LoopFullUnrollPass
 
 entry:
   br label %l0

 l0.latch:
   br label %l0
 ; CHECK: LoopFullUnrollPass on l0
 ; CHECK-NOT: LoopFullUnrollPass
 
 exit:
   ret void
 }
-
-; Now we test forced runtime partial unrolling with metadata. Here we end up
-; duplicating child loops without changing their structure and so they aren't by
-; default visited, but will be visited with a special parameter.
-define void @partial_unroll(i32 %count, ptr %ptr) {
-; CHECK-LABEL: OptimizationRemarkEmitterAnalysis on partial_unroll
-; CHECK-NOT: LoopFullUnrollPass
-
-entry:
-  br label %l0
-
-l0:
-  %cond.0 = load volatile i1, ptr %ptr
-  br i1 %cond.0, label %l0.0.ph, label %exit
-
-l0.0.ph:
-  br label %l0.0
-
-l0.0:
-  %iv = phi i32 [ %iv.next, %l0.0.latch ], [ 0, %l0.0.ph ]
-  %iv.next = add i32 %iv, 1
-  br label %l0.0.0.ph
-
-l0.0.0.ph:
-  br label %l0.0.0
-
-l0.0.0:
-  %cond.0.0.0 = load volatile i1, ptr %ptr
-  br i1 %cond.0.0.0, label %l0.0.0, label %l0.0.1.ph
-; CHECK: LoopFullUnrollPass on l0.0.0
-; CHECK-NOT: LoopFullUnrollPass
-
-l0.0.1.ph:
-  br label %l0.0.1
-
-l0.0.1:
-  %cond.0.0.1 = load volatile i1, ptr %ptr
-  br i1 %cond.0.0.1, label %l0.0.1, label %l0.0.latch
-; CHECK: LoopFullUnrollPass on l0.0.1
-; CHECK-NOT: LoopFullUnrollPass
-
-l0.0.latch:
-  %cmp = icmp slt i32 %iv.next, %count
-  br i1 %cmp, label %l0.0, label %l0.latch, !llvm.loop !1
-; CHECK: LoopFullUnrollPass on l0.0
-; CHECK-NOT: LoopFullUnrollPass
-;
-; Partial unrolling occurs which introduces both new child loops and new sibling
-; loops. We only visit the child loops in a special mode, not by default.
-; CHECK-CHILDREN: LoopFullUnrollPass on l0.0.0
-; CHECK-CHILDREN-NOT: LoopFullUnrollPass
-; CHECK-CHILDREN: LoopFullUnrollPass on l0.0.1
-; CHECK-CHILDREN-NOT: LoopFullUnrollPass
-; CHECK-CHILDREN: LoopFullUnrollPass on l0.0.0.1
-; CHECK-CHILDREN-NOT: LoopFullUnrollPass
-; CHECK-CHILDREN: LoopFullUnrollPass on l0.0.1.1
-; CHECK-CHILDREN-NOT: LoopFullUnrollPass
-;
-; When we revisit children, we also revisit the current loop.
-; CHECK-CHILDREN: LoopFullUnrollPass on l0.0
-; CHECK-CHILDREN-NOT: LoopFullUnrollPass
-;
-; Revisit the children of the outer loop that are part of the epilogue.
-;
-; CHECK: LoopFullUnrollPass on l0.0.1.epil
-; CHECK-NOT: LoopFullUnrollPass
-; CHECK: LoopFullUnrollPass on l0.0.0.epil
-; CHECK-NOT: LoopFullUnrollPass
-l0.latch:
-  br label %l0
-; CHECK: LoopFullUnrollPass on l0
-; CHECK-NOT: LoopFullUnrollPass
-
-exit:
-  ret void
-}
-!1 = !{!1, !2}
-!2 = !{!"llvm.loop.unroll.count", i32 2}