diff --git a/llvm/lib/Transforms/Utils/LoopUnrollRuntime.cpp b/llvm/lib/Transforms/Utils/LoopUnrollRuntime.cpp
--- a/llvm/lib/Transforms/Utils/LoopUnrollRuntime.cpp
+++ b/llvm/lib/Transforms/Utils/LoopUnrollRuntime.cpp
@@ -50,6 +50,10 @@
     "unroll-runtime-multi-exit", cl::init(false), cl::Hidden,
     cl::desc("Allow runtime unrolling for loops with multiple exits, when "
              "epilog is generated"));
+static cl::opt<bool> UnrollRuntimeAtMostTwoExits(
+    "unroll-runtime-at-most-two-exits", cl::init(false), cl::Hidden,
+    cl::desc("Allow runtime unrolling for loops with at most two exits, when "
+             "epilog is generated"));
 
 /// Connect the unrolling prolog code to the original loop.
 /// The unrolling prolog code contains code to execute the
@@ -493,6 +497,15 @@
   if (ExitingBlocks.size() > 2)
     return false;
 
+  // Only enable unrolling of loops that have a single side exit other than the
+  // normal LatchExit.
+  if (OtherExits.size() > 1)
+    return false;
+
+  // At this point, the loop must have at most two exiting/exit blocks.
+  if (UnrollRuntimeAtMostTwoExits)
+    return true;
+
   // The second heuristic is that L has one exit other than the latchexit and
   // that exit is a deoptimize block. We know that deoptimize blocks are rarely
   // taken, which also implies the branch leading to the deoptimize block is
diff --git a/llvm/test/Transforms/LoopUnroll/runtime-loop-at-most-two-exits.ll b/llvm/test/Transforms/LoopUnroll/runtime-loop-at-most-two-exits.ll
new file mode 100644
--- /dev/null
+++ b/llvm/test/Transforms/LoopUnroll/runtime-loop-at-most-two-exits.ll
@@ -0,0 +1,130 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -loop-unroll -unroll-runtime=true -unroll-count=2 -unroll-runtime-epilog=true -unroll-runtime-at-most-two-exits=true -verify-loop-lcssa -verify-dom-info -verify-loop-info -S | FileCheck %s --check-prefix=ENABLED
+; RUN: opt < %s -loop-unroll -unroll-runtime=true -unroll-count=2 -unroll-runtime-epilog=true -unroll-runtime-at-most-two-exits=false -verify-loop-lcssa -verify-dom-info -verify-loop-info -S | FileCheck %s --check-prefix=DISABLED
+
+define i32 @test(i32* nocapture %a, i64 %n) {
+; ENABLED-LABEL: @test(
+; ENABLED-NEXT:  entry:
+; ENABLED-NEXT:    [[TMP0:%.*]] = add i64 [[N:%.*]], -1
+; ENABLED-NEXT:    [[XTRAITER:%.*]] = and i64 [[N]], 1
+; ENABLED-NEXT:    [[TMP1:%.*]] = icmp ult i64 [[TMP0]], 1
+; ENABLED-NEXT:    br i1 [[TMP1]], label [[FOR_END_UNR_LCSSA:%.*]], label [[ENTRY_NEW:%.*]]
+; ENABLED:       entry.new:
+; ENABLED-NEXT:    [[UNROLL_ITER:%.*]] = sub i64 [[N]], [[XTRAITER]]
+; ENABLED-NEXT:    br label [[HEADER:%.*]]
+; ENABLED:       header:
+; ENABLED-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY_NEW]] ], [ [[INDVARS_IV_NEXT_1:%.*]], [[FOR_BODY_1:%.*]] ]
+; ENABLED-NEXT:    [[SUM_02:%.*]] = phi i32 [ 0, [[ENTRY_NEW]] ], [ [[ADD_1:%.*]], [[FOR_BODY_1]] ]
+; ENABLED-NEXT:    [[NITER:%.*]] = phi i64 [ [[UNROLL_ITER]], [[ENTRY_NEW]] ], [ [[NITER_NSUB_1:%.*]], [[FOR_BODY_1]] ]
+; ENABLED-NEXT:    [[CMP:%.*]] = icmp eq i64 [[N]], 42
+; ENABLED-NEXT:    br i1 [[CMP]], label [[FOR_EXIT2_LOOPEXIT:%.*]], label [[FOR_BODY:%.*]]
+; ENABLED:       for.body:
+; ENABLED-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDVARS_IV]]
+; ENABLED-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+; ENABLED-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP2]], [[SUM_02]]
+; ENABLED-NEXT:    [[INDVARS_IV_NEXT:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 1
+; ENABLED-NEXT:    [[NITER_NSUB:%.*]] = sub i64 [[NITER]], 1
+; ENABLED-NEXT:    [[CMP_1:%.*]] = icmp eq i64 [[N]], 42
+; ENABLED-NEXT:    br i1 [[CMP_1]], label [[FOR_EXIT2_LOOPEXIT]], label [[FOR_BODY_1]]
+; ENABLED:       for.end.unr-lcssa.loopexit:
+; ENABLED-NEXT:    [[SUM_0_LCSSA_PH_PH:%.*]] = phi i32 [ [[ADD_1]], [[FOR_BODY_1]] ]
+; ENABLED-NEXT:    [[INDVARS_IV_UNR_PH:%.*]] = phi i64 [ [[INDVARS_IV_NEXT_1]], [[FOR_BODY_1]] ]
+; ENABLED-NEXT:    [[SUM_02_UNR_PH:%.*]] = phi i32 [ [[ADD_1]], [[FOR_BODY_1]] ]
+; ENABLED-NEXT:    br label [[FOR_END_UNR_LCSSA]]
+; ENABLED:       for.end.unr-lcssa:
+; ENABLED-NEXT:    [[SUM_0_LCSSA_PH:%.*]] = phi i32 [ undef, [[ENTRY:%.*]] ], [ [[SUM_0_LCSSA_PH_PH]], [[FOR_END_UNR_LCSSA_LOOPEXIT:%.*]] ]
+; ENABLED-NEXT:    [[INDVARS_IV_UNR:%.*]] = phi i64 [ 0, [[ENTRY]] ], [ [[INDVARS_IV_UNR_PH]], [[FOR_END_UNR_LCSSA_LOOPEXIT]] ]
+; ENABLED-NEXT:    [[SUM_02_UNR:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[SUM_02_UNR_PH]], [[FOR_END_UNR_LCSSA_LOOPEXIT]] ]
+; ENABLED-NEXT:    [[LCMP_MOD:%.*]] = icmp ne i64 [[XTRAITER]], 0
+; ENABLED-NEXT:    br i1 [[LCMP_MOD]], label [[HEADER_EPIL_PREHEADER:%.*]], label [[FOR_END:%.*]]
+; ENABLED:       header.epil.preheader:
+; ENABLED-NEXT:    br label [[HEADER_EPIL:%.*]]
+; ENABLED:       header.epil:
+; ENABLED-NEXT:    [[INDVARS_IV_EPIL:%.*]] = phi i64 [ [[INDVARS_IV_UNR]], [[HEADER_EPIL_PREHEADER]] ]
+; ENABLED-NEXT:    [[SUM_02_EPIL:%.*]] = phi i32 [ [[SUM_02_UNR]], [[HEADER_EPIL_PREHEADER]] ]
+; ENABLED-NEXT:    [[CMP_EPIL:%.*]] = icmp eq i64 [[N]], 42
+; ENABLED-NEXT:    br i1 [[CMP_EPIL]], label [[FOR_EXIT2:%.*]], label [[FOR_BODY_EPIL:%.*]]
+; ENABLED:       for.body.epil:
+; ENABLED-NEXT:    [[ARRAYIDX_EPIL:%.*]] = getelementptr inbounds i32, i32* [[A]], i64 [[INDVARS_IV_EPIL]]
+; ENABLED-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX_EPIL]], align 4
+; ENABLED-NEXT:    [[ADD_EPIL:%.*]] = add nsw i32 [[TMP3]], [[SUM_02_EPIL]]
+; ENABLED-NEXT:    [[INDVARS_IV_NEXT_EPIL:%.*]] = add i64 [[INDVARS_IV_EPIL]], 1
+; ENABLED-NEXT:    [[EXITCOND_EPIL:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT_EPIL]], [[N]]
+; ENABLED-NEXT:    br label [[FOR_END_EPILOG_LCSSA:%.*]]
+; ENABLED:       for.end.epilog-lcssa:
+; ENABLED-NEXT:    br label [[FOR_END]]
+; ENABLED:       for.end:
+; ENABLED-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ [[SUM_0_LCSSA_PH]], [[FOR_END_UNR_LCSSA]] ], [ [[ADD_EPIL]], [[FOR_END_EPILOG_LCSSA]] ]
+; ENABLED-NEXT:    ret i32 [[SUM_0_LCSSA]]
+; ENABLED:       for.exit2.loopexit:
+; ENABLED-NEXT:    [[RETVAL_PH:%.*]] = phi i32 [ [[SUM_02]], [[HEADER]] ], [ [[ADD]], [[FOR_BODY]] ]
+; ENABLED-NEXT:    br label [[FOR_EXIT2]]
+; ENABLED:       for.exit2:
+; ENABLED-NEXT:    [[RETVAL:%.*]] = phi i32 [ [[SUM_02_EPIL]], [[HEADER_EPIL]] ], [ [[RETVAL_PH]], [[FOR_EXIT2_LOOPEXIT]] ]
+; ENABLED-NEXT:    ret i32 [[RETVAL]]
+; ENABLED:       for.body.1:
+; ENABLED-NEXT:    [[ARRAYIDX_1:%.*]] = getelementptr inbounds i32, i32* [[A]], i64 [[INDVARS_IV_NEXT]]
+; ENABLED-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX_1]], align 4
+; ENABLED-NEXT:    [[ADD_1]] = add nsw i32 [[TMP4]], [[ADD]]
+; ENABLED-NEXT:    [[INDVARS_IV_NEXT_1]] = add i64 [[INDVARS_IV_NEXT]], 1
+; ENABLED-NEXT:    [[NITER_NSUB_1]] = sub i64 [[NITER_NSUB]], 1
+; ENABLED-NEXT:    [[NITER_NCMP_1:%.*]] = icmp eq i64 [[NITER_NSUB_1]], 0
+; ENABLED-NEXT:    br i1 [[NITER_NCMP_1]], label [[FOR_END_UNR_LCSSA_LOOPEXIT]], label [[HEADER]], [[LOOP0:!llvm.loop !.*]]
+;
+; DISABLED-LABEL: @test(
+; DISABLED-NEXT:  entry:
+; DISABLED-NEXT:    br label [[HEADER:%.*]]
+; DISABLED:       header:
+; DISABLED-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT_1:%.*]], [[FOR_BODY_1:%.*]] ]
+; DISABLED-NEXT:    [[SUM_02:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[ADD_1:%.*]], [[FOR_BODY_1]] ]
+; DISABLED-NEXT:    [[CMP:%.*]] = icmp eq i64 [[N:%.*]], 42
+; DISABLED-NEXT:    br i1 [[CMP]], label [[FOR_EXIT2:%.*]], label [[FOR_BODY:%.*]]
+; DISABLED:       for.body:
+; DISABLED-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDVARS_IV]]
+; DISABLED-NEXT:    [[TMP0:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+; DISABLED-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP0]], [[SUM_02]]
+; DISABLED-NEXT:    [[INDVARS_IV_NEXT:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 1
+; DISABLED-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
+; DISABLED-NEXT:    br i1 [[EXITCOND]], label [[FOR_END:%.*]], label [[HEADER_1:%.*]]
+; DISABLED:       for.end:
+; DISABLED-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ [[ADD]], [[FOR_BODY]] ], [ [[ADD_1]], [[FOR_BODY_1]] ]
+; DISABLED-NEXT:    ret i32 [[SUM_0_LCSSA]]
+; DISABLED:       for.exit2:
+; DISABLED-NEXT:    [[RETVAL:%.*]] = phi i32 [ [[SUM_02]], [[HEADER]] ], [ [[ADD]], [[HEADER_1]] ]
+; DISABLED-NEXT:    ret i32 [[RETVAL]]
+; DISABLED:       header.1:
+; DISABLED-NEXT:    [[CMP_1:%.*]] = icmp eq i64 [[N]], 42
+; DISABLED-NEXT:    br i1 [[CMP_1]], label [[FOR_EXIT2]], label [[FOR_BODY_1]]
+; DISABLED:       for.body.1:
+; DISABLED-NEXT:    [[ARRAYIDX_1:%.*]] = getelementptr inbounds i32, i32* [[A]], i64 [[INDVARS_IV_NEXT]]
+; DISABLED-NEXT:    [[TMP1:%.*]] = load i32, i32* [[ARRAYIDX_1]], align 4
+; DISABLED-NEXT:    [[ADD_1]] = add nsw i32 [[TMP1]], [[ADD]]
+; DISABLED-NEXT:    [[INDVARS_IV_NEXT_1]] = add i64 [[INDVARS_IV_NEXT]], 1
+; DISABLED-NEXT:    [[EXITCOND_1:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT_1]], [[N]]
+; DISABLED-NEXT:    br i1 [[EXITCOND_1]], label [[FOR_END]], label [[HEADER]], [[LOOP0:!llvm.loop !.*]]
+;
+entry:
+  br label %header
+
+header:
+  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
+  %sum.02 = phi i32 [ %add, %for.body ], [ 0, %entry ]
+  %cmp = icmp eq i64 %n, 42
+  br i1 %cmp, label %for.exit2, label %for.body
+
+for.body:
+  %arrayidx = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
+  %0 = load i32, i32* %arrayidx, align 4
+  %add = add nsw i32 %0, %sum.02
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %exitcond = icmp eq i64 %indvars.iv.next, %n
+  br i1 %exitcond, label %for.end, label %header
+
+for.end:
+  %sum.0.lcssa = phi i32 [ %add, %for.body ]
+  ret i32 %sum.0.lcssa
+
+for.exit2:
+  %retval = phi i32 [ %sum.02, %header ]
+  ret i32 %retval
+}