[WPD] Allow virtual calls to be analyzed with multiple type tests

Summary:
In D52514 I had fixed a bug with WPD after indirect call promotion, by
checking that a type test being analyzed dominates potential virtual
calls. With that fix I included a small effiency enhancement to avoid
processing a devirt candidate multiple times (when there are multiple
type tests). This latter change wasn't in response to any measured
efficiency issues, it was merely theoretical. Unfortuantely, it turns
out to limit optimization opportunities after inlining.

Specifically, consider code that looks like:

class A {
  virtual void foo();
};
class B : public A {
  void foo();
}
void callee(A *a) {
  a->foo(); // Call 1
}
void caller(B *b) {
  b->foo(); // Call 2
  callee(b);
}

After inlining callee into caller, because of the existing call to
b->foo() in caller there will be 2 type tests in caller for the vtable
pointer of b: the original type test against B from Call 2, and the
inlined type test against A from Call 1. If the code was compiled with
-fstrict-vtable-pointers, then after optimization WPD will see that
both type tests are associated with the inlined virtual Call 1.
With my earlier change to only process a virtual call against one type
test, we may only consider virtual Call 1 against the base class A type
test, which can't be devirtualized. With my change here to remove this
restriction, it also gets considered for the type test against the
derived class B type test, where it can be devirtualized.

Note that if caller didn't include it's own earlier virtual call
b->foo() we will not be able to devirtualize after inlining callee even
after this fix, since there would not be a type test against B in the
IR. As a future enhancement we can consider inserting type tests at call
sites that pass pointers to classes with virtual calls, to enable
context-sensitive devirtualization after inlining.

Reviewers: pcc, vitalybuka, evgeny777

Subscribers: Prazek, hiraditya, steven_wu, dexonsmith, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D79235

llvm/lib/Transforms/IPO/WholeProgramDevirt.cpp

@@ -1713,7 +1713,6 @@ void DevirtModule::scanTypeTestUsers(Function *TypeTestFunc) {
   // points to a member of the type identifier %md. Group calls by (type ID,
   // offset) pair (effectively the identity of the virtual function) and store
   // to CallSlots.
-  DenseSet<CallBase *> SeenCallSites;
   for (auto I = TypeTestFunc->use_begin(), E = TypeTestFunc->use_end();
        I != E;) {
     auto CI = dyn_cast<CallInst>(I->getUser());
@@ -1732,15 +1731,8 @@ void DevirtModule::scanTypeTestUsers(Function *TypeTestFunc) {
       Metadata *TypeId =
           cast<MetadataAsValue>(CI->getArgOperand(1))->getMetadata();
       Value *Ptr = CI->getArgOperand(0)->stripPointerCasts();
-      for (DevirtCallSite Call : DevirtCalls) {
-        // Only add this CallSite if we haven't seen it before. The vtable
-        // pointer may have been CSE'd with pointers from other call sites,
-        // and we don't want to process call sites multiple times. We can't
-        // just skip the vtable Ptr if it has been seen before, however, since
-        // it may be shared by type tests that dominate different calls.
-        if (SeenCallSites.insert(&Call.CB).second)
-          CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CB, nullptr);
-      }
+      for (DevirtCallSite Call : DevirtCalls)
+        CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CB, nullptr);
     }
 
     // We no longer need the assumes or the type test.

llvm/test/ThinLTO/X86/devirt_multiple_type_test.ll

@@ -0,0 +1,80 @@
+; Test to ensure that devirtualization will succeed when there is an earlier
+; type test also corresponding to the same vtable (when indicated by invariant
+; load metadata), that provides a more refined type. This could happen in
+; after inlining into a caller passing a derived type.
+
+; RUN: opt -module-summary %s -o %t.o
+; RUN: llvm-lto2 run -o %t.out %t.o \
+; RUN:	 -pass-remarks=wholeprogramdevirt \
+; RUN:	 -r %t.o,_ZN1A3fooEv,px \
+; RUN:	 -r %t.o,_ZN1B3fooEv,px \
+; RUN:	 -r %t.o,_Z6callerP1B,px \
+; RUN:	 -r %t.o,_ZTV1A,px \
+; RUN:	 -r %t.o,_ZTV1B,px \
+; RUN:	 -save-temps 2>&1 | FileCheck %s
+
+; CHECK-COUNT-2: single-impl: devirtualized a call to _ZN1B3fooEv
+
+; RUN: llvm-dis %t.out.1.4.opt.bc -o - | FileCheck %s --check-prefix=IR
+; IR-NOT: tail call void %
+
+; ModuleID = 'devirt_multiple_type_test.o'
+source_filename = "devirt_multiple_type_test.cc"
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+%class.A = type { i32 (...)** }
+%class.B = type { %class.A }
+
+@_ZTV1A = hidden unnamed_addr constant { [3 x i8*] } { [3 x i8*] [i8* null, i8* undef, i8* bitcast (void (%class.A*)* @_ZN1A3fooEv to i8*)] }, align 8, !type !0, !vcall_visibility !2
+@_ZTV1B = hidden unnamed_addr constant { [3 x i8*] } { [3 x i8*] [i8* null, i8* undef, i8* bitcast (void (%class.B*)* @_ZN1B3fooEv to i8*)] }, align 8, !type !0, !type !3, !vcall_visibility !2
+
+declare void @_ZN1A3fooEv(%class.A* nocapture %this)
+
+define hidden void @_ZN1B3fooEv(%class.B* nocapture %this) {
+entry:
+  ret void
+}
+
+; Function Attrs: nounwind readnone willreturn
+declare i1 @llvm.type.test(i8*, metadata)
+
+; Function Attrs: nounwind willreturn
+declare void @llvm.assume(i1)
+
+; Function Attrs: uwtable
+define hidden void @_Z6callerP1B(%class.B* %b) local_unnamed_addr {
+entry:
+  %0 = bitcast %class.B* %b to void (%class.B*)***
+  %vtable = load void (%class.B*)**, void (%class.B*)*** %0, align 8, !tbaa !12, !invariant.group !15
+  %1 = bitcast void (%class.B*)** %vtable to i8*
+  %2 = tail call i1 @llvm.type.test(i8* %1, metadata !"_ZTS1B")
+  tail call void @llvm.assume(i1 %2)
+  %3 = load void (%class.B*)*, void (%class.B*)** %vtable, align 8, !invariant.load !15
+  tail call void %3(%class.B* %b)
+  %4 = getelementptr %class.B, %class.B* %b, i64 0, i32 0
+  %5 = bitcast void (%class.B*)** %vtable to i8*
+  %6 = tail call i1 @llvm.type.test(i8* nonnull %5, metadata !"_ZTS1A")
+  tail call void @llvm.assume(i1 %6)
+  %7 = bitcast void (%class.B*)* %3 to void (%class.A*)*
+  tail call void %7(%class.A* %4)
+  ret void
+}
+
+!llvm.module.flags = !{!5, !6, !8, !9, !10}
+!llvm.ident = !{!11}
+
+!0 = !{i64 16, !"_ZTS1A"}
+!2 = !{i64 1}
+!3 = !{i64 16, !"_ZTS1B"}
+!5 = !{i32 1, !"StrictVTablePointers", i32 1}
+!6 = !{i32 3, !"StrictVTablePointersRequirement", !7}
+!7 = !{!"StrictVTablePointers", i32 1}
+!8 = !{i32 1, !"wchar_size", i32 4}
+!9 = !{i32 1, !"Virtual Function Elim", i32 0}
+!10 = !{i32 1, !"EnableSplitLTOUnit", i32 0}
+!11 = !{!"clang version 11.0.0 (https://github.com/llvm/llvm-project.git 85247c1e898f88d65154b9a437b4bd83fcad8d52)"}
+!12 = !{!13, !13, i64 0}
+!13 = !{!"vtable pointer", !14, i64 0}
+!14 = !{!"Simple C++ TBAA"}
+!15 = !{}