[llvm] [CGP] Undo constant propagation of pointers across calls (PR #102926)
Nikita Popov via llvm-commits
llvm-commits at lists.llvm.org
Tue Aug 27 06:09:57 PDT 2024
================
@@ -2686,6 +2687,52 @@ bool CodeGenPrepare::optimizeCallInst(CallInst *CI, ModifyDT &ModifiedDT) {
return true;
}
+ // SCCP may have propagated, among other things, C++ static variables across
+ // calls. If this happens to be the case, we may want to undo it in order to
+ // avoid redundant pointer computation of the constant, as the function method
+ // returning the constant needs to be executed anyways.
+ auto GetUniformReturnValue = [](const Function *F) -> Constant * {
+ if (!F->getReturnType()->isPointerTy())
+ return nullptr;
+
+ Constant *UniformValue = nullptr;
+ for (auto &BB : *F) {
+ if (auto *RI = dyn_cast<ReturnInst>(BB.getTerminator())) {
+ if (auto *V = dyn_cast<GlobalVariable>(RI->getReturnValue())) {
+ if (!UniformValue)
+ UniformValue = V;
+ else if (V != UniformValue)
+ return nullptr;
+ } else {
+ return nullptr;
+ }
+ }
+ }
+
+ return UniformValue;
+ };
+
+ if (!Callee->isInterposable() && !Callee->hasDLLImportStorageClass()) {
+ if (Constant *RV = GetUniformReturnValue(Callee)) {
+ bool MadeChange = false;
+ const auto &DT = getDT(*CI->getFunction());
----------------
nikic wrote:
If we're in the same BB, you may as well skip the dominator tree and use comesBefore() instead.
https://github.com/llvm/llvm-project/pull/102926
More information about the llvm-commits
mailing list