llvm · hassnaaHamdi · Jun 20, 2025 · teresajohnson · Jul 8, 2025 · teresajohnson
diff --git a/clang/docs/UsersManual.rst b/clang/docs/UsersManual.rst
@@ -2275,9 +2275,13 @@ are listed below.
 
 .. option:: -fwhole-program-vtables
 
+   In LTO mode:
    Enable whole-program vtable optimizations, such as single-implementation
    devirtualization and virtual constant propagation, for classes with
-   :doc:`hidden LTO visibility <LTOVisibility>`. Requires ``-flto``.
+   :doc:`hidden LTO visibility <LTOVisibility>`.
+   In non-LTO mode:
+   Enables speculative devirtualization only without other features.
+   Doesn't require ``-flto`` or visibility.
 
 .. option:: -f[no]split-lto-unit
 
@@ -5170,7 +5174,7 @@ Execute ``clang-cl /?`` to see a list of supported options:
       -fstandalone-debug      Emit full debug info for all types used by the program
       -fstrict-aliasing	      Enable optimizations based on strict aliasing rules
       -fsyntax-only           Run the preprocessor, parser and semantic analysis stages
-      -fwhole-program-vtables Enables whole-program vtable optimization. Requires -flto
+      -fwhole-program-vtables Enables whole-program vtable optimization.
       -gcodeview-ghash        Emit type record hashes in a .debug$H section
       -gcodeview              Generate CodeView debug information
       -gline-directives-only  Emit debug line info directives only

diff --git a/clang/lib/CodeGen/BackendUtil.cpp b/clang/lib/CodeGen/BackendUtil.cpp
@@ -902,6 +902,7 @@ void EmitAssemblyHelper::RunOptimizationPipeline(
   // non-integrated assemblers don't recognize .cgprofile section.
   PTO.CallGraphProfile = !CodeGenOpts.DisableIntegratedAS;
   PTO.UnifiedLTO = CodeGenOpts.UnifiedLTO;
+  PTO.WholeProgramDevirt = CodeGenOpts.WholeProgramVTables;
 
   LoopAnalysisManager LAM;
   FunctionAnalysisManager FAM;

diff --git a/clang/lib/CodeGen/CGVTables.cpp b/clang/lib/CodeGen/CGVTables.cpp
@@ -1359,7 +1359,8 @@ void CodeGenModule::EmitVTableTypeMetadata(const CXXRecordDecl *RD,
   // Emit type metadata on vtables with LTO or IR instrumentation.
   // In IR instrumentation, the type metadata is used to find out vtable
   // definitions (for type profiling) among all global variables.
-  if (!getCodeGenOpts().LTOUnit && !getCodeGenOpts().hasProfileIRInstr())
+  if (!getCodeGenOpts().LTOUnit && !getCodeGenOpts().hasProfileIRInstr() &&
+      !getCodeGenOpts().WholeProgramVTables)
     return;
 
   CharUnits ComponentWidth = GetTargetTypeStoreSize(getVTableComponentType());

diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp
@@ -7847,8 +7847,12 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
         IsDeviceOffloadAction ? D.getLTOMode() : D.getOffloadLTOMode();
     auto OtherIsUsingLTO = OtherLTOMode != LTOK_None;
 
-    if ((!IsUsingLTO && !OtherIsUsingLTO) ||
-        (IsPS4 && !UnifiedLTO && (D.getLTOMode() != LTOK_Full)))
+    if (!IsUsingLTO && !OtherIsUsingLTO && !UnifiedLTO) {
+      if (const Arg *A = Args.getLastArg(options::OPT_O_Group))
+        if (!A->getOption().matches(options::OPT_O0))
+          CmdArgs.push_back("-fwhole-program-vtables");
+    } else if ((!IsUsingLTO && !OtherIsUsingLTO) ||
+               (IsPS4 && !UnifiedLTO && (D.getLTOMode() != LTOK_Full)))
       D.Diag(diag::err_drv_argument_only_allowed_with)
           << "-fwhole-program-vtables"
           << ((IsPS4 && !UnifiedLTO) ? "-flto=full" : "-flto");

diff --git a/clang/test/CodeGenCXX/devirt-single-impl.cpp b/clang/test/CodeGenCXX/devirt-single-impl.cpp
@@ -0,0 +1,56 @@
+// Check that speculative devirtualization works without the need for LTO or visibility.
+// RUN: %clang_cc1 -fwhole-program-vtables -O1 %s -emit-llvm -o - | FileCheck %s
+
+struct A {
+  A(){}
+  __attribute__((noinline))
+  virtual int virtual1(){return 20;}
+  __attribute__((noinline))
+  virtual void empty_virtual(){}
+};
+
+struct B : A {
+  B(){}
+  __attribute__((noinline))
+  virtual int virtual1() override {return 50;}
+  __attribute__((noinline))
+  virtual void empty_virtual() override {}
+};
+
+// Test that we can apply speculative devirtualization
+// without the need for LTO or visibility.
+__attribute__((noinline))
+int test_devirtual(A *a) {
+  // CHECK: %0 = load ptr, ptr %vtable, align 8
+  // CHECK-NEXT: %1 = icmp eq ptr %0, @_ZN1B8virtual1Ev
+  // CHECK-NEXT: br i1 %1, label %if.true.direct_targ, label %if.false.orig_indirect, !prof !12
+
+  // CHECK: if.true.direct_targ: ; preds = %entry
+  // CHECK-NEXT: %2 = tail call noundef i32 @_ZN1B8virtual1Ev(ptr noundef nonnull align 8 dereferenceable(8) %a)
+  // CHECK-NEXT: br label %if.end.icp
+
+  // CHECK: if.false.orig_indirect: ; preds = %entry
+  // CHECK-NEXT: %call = tail call noundef i32 %0(ptr noundef nonnull align 8 dereferenceable(8) %a)
+  // CHECK-NEXT: br label %if.end.icp
+
+  // CHECK: if.end.icp: ; preds = %if.false.orig_indirect, %if.true.direct_targ
+  // CHECK-NEXT: %3 = phi i32 [ %call, %if.false.orig_indirect ], [ %2, %if.true.direct_targ ]
+  // CHECK-NEXT: ret i32 %3
+
+  return a->virtual1();
+}
+
+// Test that we skip devirtualization for empty virtual functions as most probably
+// they are used for interfaces.
+__attribute__((noinline))
+void test_devirtual_empty_fn(A *a) {
+  // CHECK: load ptr, ptr %vfn, align 8
+  // CHECK-NEXT: tail call void %0(ptr noundef nonnull align 8 dereferenceable(8) %a)
+  a->empty_virtual();
+}
+
+void test() {
+  A *a = new B();
+  test_devirtual(a);
+  test_devirtual_empty_fn(a);
+}
diff --git a/clang/test/Driver/whole-program-vtables.c b/clang/test/Driver/whole-program-vtables.c
@@ -1,15 +1,11 @@
-// RUN: not %clang -target x86_64-unknown-linux -fwhole-program-vtables -### %s 2>&1 | FileCheck --check-prefix=NO-LTO %s
-// RUN: not %clang_cl --target=x86_64-pc-win32 -fwhole-program-vtables -### -- %s 2>&1 | FileCheck --check-prefix=NO-LTO %s
-// NO-LTO: invalid argument '-fwhole-program-vtables' only allowed with '-flto'
+// RUN: %clang -target x86_64-unknown-linux -fwhole-program-vtables -O1 -### %s 2>&1 | FileCheck --check-prefix=WPD-NO-LTO %s
+// RUN: %clang_cl --target=x86_64-pc-win32 -fwhole-program-vtables -O1 -### -- %s 2>&1 | FileCheck --check-prefix=WPD-NO-LTO %s
+// WPD-NO-LTO: "-fwhole-program-vtables"
 
 // RUN: %clang -target x86_64-unknown-linux -fwhole-program-vtables -flto -### %s 2>&1 | FileCheck --check-prefix=LTO %s
 // RUN: not %clang_cl --target=x86_64-pc-win32 -fwhole-program-vtables -flto -### -- %s 2>&1 | FileCheck --check-prefix=LTO %s
 // LTO: "-fwhole-program-vtables"
 
-/// -funified-lto does not imply -flto, so we still get an error that fwhole-program-vtables has no effect without -flto
-// RUN: not %clang --target=x86_64-pc-linux-gnu -fwhole-program-vtables -funified-lto -### %s 2>&1 | FileCheck --check-prefix=NO-LTO %s
-// RUN: not %clang --target=x86_64-pc-linux-gnu -fwhole-program-vtables -fno-unified-lto -### %s 2>&1 | FileCheck --check-prefix=NO-LTO %s
-
 // RUN: %clang -target x86_64-unknown-linux -fwhole-program-vtables -fno-whole-program-vtables -flto -### %s 2>&1 | FileCheck --check-prefix=LTO-DISABLE %s
 // RUN: not %clang_cl --target=x86_64-pc-win32 -fwhole-program-vtables -fno-whole-program-vtables -flto -### -- %s 2>&1 | FileCheck --check-prefix=LTO-DISABLE %s
 // LTO-DISABLE-NOT: "-fwhole-program-vtables"
diff --git a/llvm/include/llvm/Passes/PassBuilder.h b/llvm/include/llvm/Passes/PassBuilder.h
@@ -98,6 +98,12 @@ class PipelineTuningOptions {
   // analyses after various module->function or cgscc->function adaptors in the
   // default pipelines.
   bool EagerlyInvalidateAnalyses;
+
+  /// Tuning option to enable/disable whole program devirtualization.
+  /// Its default value is false.
+  /// This is controlled by the `-whole-program-vtables` flag.
+  /// Used only in non-LTO mode.
+  bool WholeProgramDevirt;
 };
 
 /// This class provides access to building LLVM's passes.

diff --git a/llvm/include/llvm/Transforms/IPO/WholeProgramDevirt.h b/llvm/include/llvm/Transforms/IPO/WholeProgramDevirt.h
@@ -226,11 +226,15 @@ struct WholeProgramDevirtPass : public PassInfoMixin<WholeProgramDevirtPass> {
   ModuleSummaryIndex *ExportSummary;
   const ModuleSummaryIndex *ImportSummary;
   bool UseCommandLine = false;
+  const bool InLTOMode;
   WholeProgramDevirtPass()
-      : ExportSummary(nullptr), ImportSummary(nullptr), UseCommandLine(true) {}
+      : ExportSummary(nullptr), ImportSummary(nullptr), UseCommandLine(true),
+        InLTOMode(true) {}
   WholeProgramDevirtPass(ModuleSummaryIndex *ExportSummary,
-                         const ModuleSummaryIndex *ImportSummary)
-      : ExportSummary(ExportSummary), ImportSummary(ImportSummary) {
+                         const ModuleSummaryIndex *ImportSummary,
+                         bool InLTOMode = true)
+      : ExportSummary(ExportSummary), ImportSummary(ImportSummary),
+        InLTOMode(InLTOMode) {
     assert(!(ExportSummary && ImportSummary));
   }
   LLVM_ABI PreservedAnalyses run(Module &M, ModuleAnalysisManager &);

diff --git a/llvm/lib/Passes/PassBuilderPipelines.cpp b/llvm/lib/Passes/PassBuilderPipelines.cpp
@@ -321,6 +321,7 @@ PipelineTuningOptions::PipelineTuningOptions() {
   MergeFunctions = EnableMergeFunctions;
   InlinerThreshold = -1;
   EagerlyInvalidateAnalyses = EnableEagerlyInvalidateAnalyses;
+  WholeProgramDevirt = false;
 }
 
 namespace llvm {
@@ -1629,6 +1630,23 @@ PassBuilder::buildModuleOptimizationPipeline(OptimizationLevel Level,
   if (!LTOPreLink)
     MPM.addPass(RelLookupTableConverterPass());
 
+  if (PTO.WholeProgramDevirt && LTOPhase == ThinOrFullLTOPhase::None) {
+    MPM.addPass(WholeProgramDevirtPass(/*ExportSummary*/ nullptr,
+                                       /*ImportSummary*/ nullptr,
+                                       /*InLTOMode=*/false));
+    MPM.addPass(LowerTypeTestsPass(nullptr, nullptr,
+                                   lowertypetests::DropTestKind::Assume));
+    if (EnableModuleInliner) {
+      MPM.addPass(ModuleInlinerPass(getInlineParamsFromOptLevel(Level),
+                                    UseInlineAdvisor,
+                                    ThinOrFullLTOPhase::None));
+    } else {
+      MPM.addPass(ModuleInlinerWrapperPass(
+          getInlineParamsFromOptLevel(Level),
+          /* MandatoryFirst */ true,
+          InlineContext{ThinOrFullLTOPhase::None, InlinePass::CGSCCInliner}));
+    }
+  }
   return MPM;
 }
 

diff --git a/llvm/lib/Transforms/IPO/WholeProgramDevirt.cpp b/llvm/lib/Transforms/IPO/WholeProgramDevirt.cpp
@@ -24,7 +24,8 @@
 //   returns 0, or a single vtable's function returns 1, replace each virtual
 //   call with a comparison of the vptr against that vtable's address.
 //
-// This pass is intended to be used during the regular and thin LTO pipelines:
+// This pass is intended to be used during the regular/thinLTO and non-LTO
+// pipelines:
 //
 // During regular LTO, the pass determines the best optimization for each
 // virtual call and applies the resolutions directly to virtual calls that are
@@ -48,6 +49,13 @@
 //   is supported.
 // - Import phase: (same as with hybrid case above).
 //
+// In non-LTO mode:
+// - The pass apply speculative devirtualization without requiring any type of
+// visibility.
+// - Skips other features like virtual constant propagation, uniform return
+// value
+//   optimization, unique return value optimization, branch funnels to minimize
+//   the drawbacks of wrong speculation.
 //===----------------------------------------------------------------------===//
 
 #include "llvm/Transforms/IPO/WholeProgramDevirt.h"
@@ -60,7 +68,9 @@
 #include "llvm/ADT/Statistic.h"
 #include "llvm/Analysis/AssumptionCache.h"
 #include "llvm/Analysis/BasicAliasAnalysis.h"
+#include "llvm/Analysis/ModuleSummaryAnalysis.h"
 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
+#include "llvm/Analysis/ProfileSummaryInfo.h"
 #include "llvm/Analysis/TypeMetadataUtils.h"
 #include "llvm/Bitcode/BitcodeReader.h"
 #include "llvm/Bitcode/BitcodeWriter.h"
@@ -798,6 +808,21 @@ PreservedAnalyses WholeProgramDevirtPass::run(Module &M,
       return PreservedAnalyses::all();
     return PreservedAnalyses::none();
   }
+  std::optional<ModuleSummaryIndex> Index;
+  // Force Fallback mode as it's safe in case it's non-LTO mode where
+  // we don't have hidden visibility.
+  if (!InLTOMode) {
+    DevirtCheckMode = WPDCheckMode::Fallback;
+    // In non-LTO mode, we don't have an ExportSummary, so we
+    // build the ExportSummary from the module.
+    assert(!ExportSummary &&
+           "ExportSummary is expected to be empty in non-LTO mode");
+    if (DevirtCheckMode == WPDCheckMode::Fallback && !ExportSummary) {
+      ProfileSummaryInfo PSI(M);
+      Index.emplace(buildModuleSummaryIndex(M, nullptr, &PSI));
+      ExportSummary = Index.has_value() ? &Index.value() : nullptr;
+    }
+  }
   if (!DevirtModule(M, AARGetter, OREGetter, LookupDomTree, ExportSummary,
                     ImportSummary)
            .run())
@@ -1091,10 +1116,12 @@ bool DevirtModule::tryFindVirtualCallTargets(
     if (!TM.Bits->GV->isConstant())
       return false;
 
-    // We cannot perform whole program devirtualization analysis on a vtable
-    // with public LTO visibility.
-    if (TM.Bits->GV->getVCallVisibility() ==
-        GlobalObject::VCallVisibilityPublic)
+    // If speculative devirtualization is NOT enabled, it's not safe to perform
+    // whole program devirtualization
+    //  analysis on a vtable with public LTO visibility.
+    if (DevirtCheckMode != WPDCheckMode::Fallback &&
+        TM.Bits->GV->getVCallVisibility() ==
+            GlobalObject::VCallVisibilityPublic)
       return false;
 
     Function *Fn = nullptr;
@@ -1112,6 +1139,11 @@ bool DevirtModule::tryFindVirtualCallTargets(
     // calls to pure virtuals are UB.
     if (Fn->getName() == "__cxa_pure_virtual")
       continue;
+    // In Most cases empty functions will be overridden by the
+    // implementation of the derived class, so we can skip them.
+    if (DevirtCheckMode == WPDCheckMode::Fallback &&
+        Fn->getReturnType()->isVoidTy() && Fn->getInstructionCount() <= 1)
+      continue;
 
     // We can disregard unreachable functions as possible call targets, as
     // unreachable functions shouldn't be called.
@@ -1333,10 +1365,11 @@ bool DevirtModule::trySingleImplDevirt(
   if (!IsExported)
     return false;
 
-  // If the only implementation has local linkage, we must promote to external
-  // to make it visible to thin LTO objects. We can only get here during the
-  // ThinLTO export phase.
-  if (TheFn->hasLocalLinkage()) {
+  // In case of non-speculative devirtualization, If the only implementation has
+  // local linkage, we must promote to external
+  //  to make it visible to thin LTO objects. We can only get here during the
+  //  ThinLTO export phase.
+  if (DevirtCheckMode != WPDCheckMode::Fallback && TheFn->hasLocalLinkage()) {
     std::string NewName = (TheFn->getName() + ".llvm.merged").str();
 
     // Since we are renaming the function, any comdats with the same name must
@@ -2315,6 +2348,11 @@ bool DevirtModule::run() {
 
   Function *TypeTestFunc =
       Intrinsic::getDeclarationIfExists(&M, Intrinsic::type_test);
+  // If we are applying speculative devirtualization, we can work on the public
+  // type test intrinsics.
+  if (!TypeTestFunc && DevirtCheckMode == WPDCheckMode::Fallback)
+    TypeTestFunc =
+        Intrinsic::getDeclarationIfExists(&M, Intrinsic::public_type_test);
   Function *TypeCheckedLoadFunc =
       Intrinsic::getDeclarationIfExists(&M, Intrinsic::type_checked_load);
   Function *TypeCheckedLoadRelativeFunc = Intrinsic::getDeclarationIfExists(
@@ -2437,12 +2475,18 @@ bool DevirtModule::run() {
                  .WPDRes[S.first.ByteOffset];
     if (tryFindVirtualCallTargets(TargetsForSlot, TypeMemberInfos,
                                   S.first.ByteOffset, ExportSummary)) {
-
-      if (!trySingleImplDevirt(ExportSummary, TargetsForSlot, S.second, Res)) {
-        DidVirtualConstProp |=
-            tryVirtualConstProp(TargetsForSlot, S.second, Res, S.first);
-
-        tryICallBranchFunnel(TargetsForSlot, S.second, Res, S.first);
+      trySingleImplDevirt(ExportSummary, TargetsForSlot, S.second, Res);
+      // In Speculative devirt mode, we skip virtual constant propagation
+      // and branch funneling to minimize the drawback if we got wrong
+      // speculation during devirtualization.
+      if (DevirtCheckMode != WPDCheckMode::Fallback) {
+        if (!trySingleImplDevirt(ExportSummary, TargetsForSlot, S.second,
+                                 Res)) {
+          DidVirtualConstProp |=
+              tryVirtualConstProp(TargetsForSlot, S.second, Res, S.first);
+
+          tryICallBranchFunnel(TargetsForSlot, S.second, Res, S.first);
+        }
       }
 
       // Collect functions devirtualized at least for one call site for stats.