2 files changed, 125 insertions, 2 deletions
diff --git a/llvm/lib/Transforms/Scalar/LoopUnrollPass.cpp b/llvm/lib/Transforms/Scalar/LoopUnrollPass.cpp
index b692c08c9cc..a6d7849010c 100644
--- a/llvm/lib/Transforms/Scalar/LoopUnrollPass.cpp
+++ b/llvm/lib/Transforms/Scalar/LoopUnrollPass.cpp
@@ -1253,6 +1253,10 @@ PreservedAnalyses LoopUnrollPass::run(Function &F,
   auto &AC = AM.getResult<AssumptionAnalysis>(F);
   auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
 
+  LoopAnalysisManager *LAM = nullptr;
+  if (auto *LAMProxy = AM.getCachedResult<LoopAnalysisManagerFunctionProxy>(F))
+    LAM = &LAMProxy->getManager();
+
   const ModuleAnalysisManager &MAM =
       AM.getResult<ModuleAnalysisManagerFunctionProxy>(F).getManager();
   ProfileSummaryInfo *PSI =
@@ -1278,9 +1282,7 @@ PreservedAnalyses LoopUnrollPass::run(Function &F,
     // for unrolling is only needed to get optimization remarks emitted in
     // a forward order.
     Loop &L = *Worklist.pop_back_val();
-#ifndef NDEBUG
     Loop *ParentL = L.getParentLoop();
-#endif
 
     // The API here is quite complex to call, but there are only two interesting
     // states we support: partial and full (or "simple") unrolling. However, to
@@ -1305,6 +1307,20 @@ PreservedAnalyses LoopUnrollPass::run(Function &F,
     if (CurChanged && ParentL)
       ParentL->verifyLoop();
 #endif
+
+    // Walk the parent or top-level loops after unrolling to check whether we
+    // actually removed a loop, and if so clear any cached analysis results for
+    // it. We have to do this immediately as the next unrolling could allocate
+    // a new loop object that ends up with the same address as the deleted loop
+    // object causing cache collisions.
+    if (LAM) {
+      bool IsCurLoopValid =
+          ParentL
+              ? llvm::any_of(*ParentL, [&](Loop *SibL) { return SibL == &L; })
+              : llvm::any_of(LI, [&](Loop *SibL) { return SibL == &L; });
+      if (!IsCurLoopValid)
+        LAM->clear(L);
+    }
   }
 
   if (!Changed)
diff --git a/llvm/test/Transforms/LoopUnroll/unroll-loop-invalidation.ll b/llvm/test/Transforms/LoopUnroll/unroll-loop-invalidation.ll
new file mode 100644
index 00000000000..af4d54cbca7
--- /dev/null
+++ b/llvm/test/Transforms/LoopUnroll/unroll-loop-invalidation.ll
@@ -0,0 +1,107 @@
+; This test exercises that we don't corrupt a loop-analysis when running loop
+; unrolling in a way that deletes a loop. To do that, we first ensure the
+; analysis is cached, then unroll the loop (deleting it) and make sure that the
+; next function doesn't get a cache "hit" for this stale analysis result.
+;
+; RUN: opt -S -passes='loop(require<access-info>),unroll,loop(print-access-info)' -debug-pass-manager < %s 2>&1 | FileCheck %s
+;
+; CHECK: Starting llvm::Function pass manager run.
+; CHECK: Running pass: FunctionToLoopPassAdaptor
+; CHECK: Running analysis: LoopAnalysis
+; CHECK: Running analysis: InnerAnalysisManagerProxy<{{.*}}LoopAnalysisManager
+; CHECK: Starting Loop pass manager run.
+; CHECK: Running pass: RequireAnalysisPass<{{.*}}LoopAccessAnalysis
+; CHECK: Running analysis: LoopAccessAnalysis on inner1.header
+; CHECK: Finished Loop pass manager run.
+; CHECK: Starting Loop pass manager run.
+; CHECK: Running pass: RequireAnalysisPass<{{.*}}LoopAccessAnalysis
+; CHECK: Running analysis: LoopAccessAnalysis on inner2.header
+; CHECK: Finished Loop pass manager run.
+; CHECK: Starting Loop pass manager run.
+; CHECK: Running pass: RequireAnalysisPass<{{.*}}LoopAccessAnalysis
+; CHECK: Running analysis: LoopAccessAnalysis on outer.header
+; CHECK: Finished Loop pass manager run.
+; CHECK: Running pass: LoopUnrollPass
+; CHECK: Clearing all analysis results for: inner2.header
+; CHECK: Clearing all analysis results for: outer.header
+; CHECK: Invalidating all non-preserved analyses for: test
+; CHECK: Invalidating all non-preserved analyses for: inner1.header
+; CHECK: Invalidating analysis: LoopAccessAnalysis on inner1.header
+; CHECK: Invalidating all non-preserved analyses for: inner1.header.1
+; CHECK-NOT: Invalidating analysis: LoopAccessAnalysis on inner1.header.1
+; CHECK: Running pass: FunctionToLoopPassAdaptor
+; CHECK: Starting Loop pass manager run.
+; CHECK: Running pass: LoopAccessInfoPrinterPass
+; CHECK: Running analysis: LoopAccessAnalysis on inner1.header
+; CHECK: Loop access info in function 'test':
+; CHECK:   inner1.header:
+; CHECK: Finished Loop pass manager run.
+; CHECK: Starting Loop pass manager run.
+; CHECK: Running pass: LoopAccessInfoPrinterPass
+; CHECK: Running analysis: LoopAccessAnalysis on inner1.header.1
+; CHECK: Loop access info in function 'test':
+; CHECK:   inner1.header.1:
+; CHECK: Finished Loop pass manager run.
+
+target triple = "x86_64-unknown-linux-gnu"
+
+define void @test(i32 %inner1.count) {
+; CHECK-LABEL: define void @test(
+bb:
+  br label %outer.ph
+
+outer.ph:
+  br label %outer.header
+
+outer.header:
+  %outer.i = phi i32 [ 0, %outer.ph ], [ %outer.i.next, %outer.latch ]
+  br label %inner1.ph
+
+inner1.ph:
+  br label %inner1.header
+
+inner1.header:
+  %inner1.i = phi i32 [ 0, %inner1.ph ], [ %inner1.i.next, %inner1.header ]
+  %inner1.i.next = add i32 %inner1.i, 1
+  %inner1.cond = icmp eq i32 %inner1.i, %inner1.count
+  br i1 %inner1.cond, label %inner1.exit, label %inner1.header
+; We should have two unrolled copies of this loop and nothing else.
+;
+; CHECK-NOT:     icmp eq
+; CHECK-NOT:     br i1
+; CHECK:         %[[COND1:.*]] = icmp eq i32 %{{.*}}, %inner1.count
+; CHECK:         br i1 %[[COND1]],
+; CHECK-NOT:     icmp eq
+; CHECK-NOT:     br i1
+; CHECK:         %[[COND2:.*]] = icmp eq i32 %{{.*}}, %inner1.count
+; CHECK:         br i1 %[[COND2]],
+; CHECK-NOT:     icmp eq
+; CHECK-NOT:     br i1
+
+
+inner1.exit:
+  br label %inner2.ph
+
+inner2.ph:
+  br label %inner2.header
+
+inner2.header:
+  %inner2.i = phi i32 [ 0, %inner2.ph ], [ %inner2.i.next, %inner2.header ]
+  %inner2.i.next = add i32 %inner2.i, 1
+  %inner2.cond = icmp eq i32 %inner2.i, 4
+  br i1 %inner2.cond, label %inner2.exit, label %inner2.header
+
+inner2.exit:
+  br label %outer.latch
+
+outer.latch:
+  %outer.i.next = add i32 %outer.i, 1
+  %outer.cond = icmp eq i32 %outer.i.next, 2
+  br i1 %outer.cond, label %outer.exit, label %outer.header
+
+outer.exit:
+  br label %exit
+
+exit:
+  ret void
+}