[clang] [llvm] [WPD] Support speculative devirtualization with LTO (PR #233380)
Teresa Johnson via cfe-commits
cfe-commits at lists.llvm.org
Tue Nov 10 11:58:57 PST 2026
================
@@ -1152,13 +1135,27 @@ bool DevirtModule::tryFindVirtualCallTargets(
TargetsForSlot.push_back({GV, &TM});
}
+ // In most cases empty functions will be overridden by the
+ // implementation of the derived class, so we can skip them
+ // for speculative devirtualization.
+ if (HasPublicTarget)
+ erase_if(TargetsForSlot, [](const VirtualCallTarget &Target) {
+ const auto *Fn = dyn_cast<Function>(Target.Fn);
+ if (!Fn)
+ Fn = cast<Function>(cast<GlobalAlias>(Target.Fn)->getAliasee());
+ return Fn->getReturnType()->isVoidTy() && Fn->getInstructionCount() <= 1;
----------------
teresajohnson wrote:
getInstructionCount() is 0 for a declaration, so this also erases any void target that is only declared in this module. In hybrid ThinLTO (split LTO unit for CFI) the vtables are in the regular LTO module and all methods are declarations there, so no void method is ever speculatively devirtualized. When Fn->isDeclaration(), either keep the target, or look up its FunctionSummary in ExportSummary and use instCount() <= 1. A void method with a non-trivial body in devirt_speculative.ll under the hybrid RUN line would catch this.
https://github.com/llvm/llvm-project/pull/233380
More information about the cfe-commits
mailing list