[llvm] [NVPTX] Lower allocas to the local address space (PR #204346)

[Tim Besard via llvm-commits]

================
@@ -51,82 +56,228 @@ char NVPTXLowerAlloca::ID = 1;
 INITIALIZE_PASS(NVPTXLowerAlloca, "nvptx-lower-alloca", "Lower Alloca", false,
                 false)
 
+static Value *getOrCreateGenericPtr(Value *LocalPtr,
+                                    DenseMap<Value *, Value *> &GenericPtrs) {
+  auto It = GenericPtrs.find(LocalPtr);
+  if (It != GenericPtrs.end())
+    return It->second;
+
+  auto *LocalInst = cast<Instruction>(LocalPtr);
+  auto *GenericPtr = new AddrSpaceCastInst(
+      LocalPtr, PointerType::get(LocalPtr->getContext(), ADDRESS_SPACE_GENERIC),
+      "");
+  GenericPtr->insertAfter(LocalInst->getIterator());
+  GenericPtrs[LocalPtr] = GenericPtr;
+  return GenericPtr;
+}
+
+static void updateMemIntrinsicDeclaration(MemIntrinsic *MI) {
+  SmallVector<Type *, 3> Tys;
+  if (auto *MTI = dyn_cast<MemTransferInst>(MI)) {
+    Tys.push_back(MTI->getRawDest()->getType());
+    Tys.push_back(MTI->getRawSource()->getType());
+    Tys.push_back(MTI->getLength()->getType());
+  } else {
+    auto *MSI = cast<MemSetInst>(MI);
+    Tys.push_back(MSI->getRawDest()->getType());
+    Tys.push_back(MSI->getLength()->getType());
+  }
+
+  Function *Decl = Intrinsic::getOrInsertDeclaration(MI->getModule(),
+                                                     MI->getIntrinsicID(), Tys);
+  MI->setCalledFunction(Decl);
+}
+
+static void convertPointerUsersToLocal(Value *OldPtr, Value *LocalPtr,
+                                       DenseMap<Value *, Value *> &GenericPtrs,
+                                       SmallPtrSetImpl<Value *> &Visited) {
+  if (!Visited.insert(OldPtr).second)
+    return;
+
+  // Debug records aren't on the use-list visited below, so retarget them here;
----------------
maleadt wrote:

Yes, we only need to run NVPTXLowerAlloca now, but that only became legal after having it lower a generic alloca into an AS5 alloca + cast back to AS0. Which indeed can result in byte-by-byte memcpy, which I figured is acceptable on O0.

https://github.com/llvm/llvm-project/pull/204346