Index: llvm/trunk/lib/Passes/PassBuilder.cpp
===================================================================
--- llvm/trunk/lib/Passes/PassBuilder.cpp
+++ llvm/trunk/lib/Passes/PassBuilder.cpp
@@ -147,6 +147,9 @@
 
 using namespace llvm;
 
+static cl::opt<unsigned> MaxDevirtIterations("pm-max-devirt-iterations",
+                                             cl::ReallyHidden, cl::init(4));
+
 static Regex DefaultAliasRegex("^(default|lto-pre-link|lto)<(O[0123sz])>$");
 
 static bool isOptimizingForSize(PassBuilder::OptimizationLevel Level) {
@@ -465,8 +468,14 @@
   MainCGPipeline.addPass(createCGSCCToFunctionPassAdaptor(
       buildFunctionSimplificationPipeline(Level, DebugLogging)));
 
+  // We wrap the CGSCC pipeline in a devirtualization repeater. This will try
+  // to detect when we devirtualize indirect calls and iterate the SCC passes
+  // in that case to try and catch knock-on inlining or function attrs
+  // opportunities. Then we add it to the module pipeline by walking the SCCs
+  // in postorder (or bottom-up).
   MPM.addPass(
-      createModuleToPostOrderCGSCCPassAdaptor(std::move(MainCGPipeline)));
+      createModuleToPostOrderCGSCCPassAdaptor(createDevirtSCCRepeatedPass(
+          std::move(MainCGPipeline), MaxDevirtIterations, DebugLogging)));
 
   // This ends the canonicalization and simplification phase of the pipeline.
   // At this point, we expect to have canonical and simple IR which we begin
Index: llvm/trunk/test/Other/cgscc-devirt-iteration.ll
===================================================================
--- llvm/trunk/test/Other/cgscc-devirt-iteration.ll
+++ llvm/trunk/test/Other/cgscc-devirt-iteration.ll
@@ -7,6 +7,9 @@
 ; RUN: opt -aa-pipeline=basic-aa -passes='cgscc(function-attrs,function(gvn,instcombine))' -S < %s | FileCheck %s --check-prefix=CHECK --check-prefix=BEFORE
 ; RUN: opt -aa-pipeline=basic-aa -passes='cgscc(devirt<1>(function-attrs,function(gvn,instcombine)))' -S < %s | FileCheck %s --check-prefix=CHECK --check-prefix=AFTER --check-prefix=AFTER1
 ; RUN: opt -aa-pipeline=basic-aa -passes='cgscc(devirt<2>(function-attrs,function(gvn,instcombine)))' -S < %s | FileCheck %s --check-prefix=CHECK --check-prefix=AFTER --check-prefix=AFTER2
+;
+; We also verify that the real O2 pipeline catches these cases.
+; RUN: opt -aa-pipeline=basic-aa -passes='default<O2>' -S < %s | FileCheck %s --check-prefix=CHECK --check-prefix=AFTER --check-prefix=AFTER2
 
 declare void @readnone() readnone
 ; CHECK: Function Attrs: readnone
@@ -93,8 +96,7 @@
 }
 
 declare i8* @memcpy(i8*, i8*, i64)
-; CHECK-NOT: Function Attrs
-; CHECK: declare i8* @memcpy(i8*, i8*, i64)
+; CHECK: declare i8* @memcpy(
 
 ; The @test3 function checks that when we refine an indirect call to an
 ; intrinsic we still revisit the SCC pass. This also covers cases where the
@@ -112,3 +114,15 @@
 ; CHECK: call void @llvm.memcpy
   ret void
 }
+
+; A boring function that just keeps our declarations around.
+define void @keep(i8** %sink) {
+; CHECK-NOT: Function Attrs
+; CHECK: define void @keep(
+entry:
+  store volatile i8* bitcast (void ()* @readnone to i8*), i8** %sink
+  store volatile i8* bitcast (void ()* @unknown to i8*), i8** %sink
+  store volatile i8* bitcast (i8* (i8*, i8*, i64)* @memcpy to i8*), i8** %sink
+  call void @unknown()
+  ret void
+}