[llvm] [IR] Allow uses of `llvm.global_ctors` and `llvm.global_dtors` (PR #96477)

[Daniil Kovalev via llvm-commits]

https://github.com/kovdan01 created https://github.com/llvm/llvm-project/pull/96477

With PAuth enabled, signed function pointers in init/fini arrays are going to be represented with `ptrauth` constants. To support address discrimination for such signed pointers, we need to fill the storage address with `getelementptr` referencing the array (`llvm.global_ctors` or `llvm.global_dtors`) itself.

Such uses of these special arrays were previously disallowed since `appendToGlobal{C|D}tors` did not update uses after construction of a new array (see b67c16ff6af557174e498d6d82df9936e333f3d6). This patch implements such update logic. This will be tested in a follow-up patch introducing PAuth support for function pointers in init/fini arrays since such signed pointers with address discrimination are currently the only entities which have such uses needed to be updated.

>From 219b873992b0a30afa796f61384c4ff26f15a45e Mon Sep 17 00:00:00 2001
From: Daniil Kovalev <dkovalev at accesssoftek.com>
Date: Mon, 24 Jun 2024 13:39:58 +0300
Subject: [PATCH] [IR] Allow uses of `llvm.global_ctors` and
 `llvm.global_dtors`

With PAuth enabled, signed function pointers in init/fini arrays are
going to be represented with `ptrauth` constants. To support address
discrimination for such signed pointers, we need to fill the storage
address with `getelementptr` referencing the array (`llvm.global_ctors`
or `llvm.global_dtors`) itself.

Such uses of these special arrays were previously disallowed since
`appendToGlobal{C|D}tors` did not update uses after construction of a new
array. This patch implements such update logic. This will be tested in a
follow-up patch introducing PAuth support for function pointers in init/fini
arrays since such signed pointers with address discrimination are
currently the only entities which have such uses needed to be updated.
---
 llvm/lib/IR/Verifier.cpp                      |  2 --
 llvm/lib/Transforms/Utils/ModuleUtils.cpp     | 14 ++++++++++----
 llvm/test/Verifier/global-ctors-dtors-uses.ll | 16 ----------------
 3 files changed, 10 insertions(+), 22 deletions(-)
 delete mode 100644 llvm/test/Verifier/global-ctors-dtors-uses.ll

diff --git a/llvm/lib/IR/Verifier.cpp b/llvm/lib/IR/Verifier.cpp
index 0abd5f76d449c..ee17d266e5057 100644
--- a/llvm/lib/IR/Verifier.cpp
+++ b/llvm/lib/IR/Verifier.cpp
@@ -850,8 +850,6 @@ void Verifier::visitGlobalVariable(const GlobalVariable &GV) {
                        GV.getName() == "llvm.global_dtors")) {
     Check(!GV.hasInitializer() || GV.hasAppendingLinkage(),
           "invalid linkage for intrinsic global variable", &GV);
-    Check(GV.materialized_use_empty(),
-          "invalid uses of intrinsic global variable", &GV);
 
     // Don't worry about emitting an error for it not being an array,
     // visitGlobalValue will complain on appending non-array.
diff --git a/llvm/lib/Transforms/Utils/ModuleUtils.cpp b/llvm/lib/Transforms/Utils/ModuleUtils.cpp
index 122279160cc7e..0656a094dfa65 100644
--- a/llvm/lib/Transforms/Utils/ModuleUtils.cpp
+++ b/llvm/lib/Transforms/Utils/ModuleUtils.cpp
@@ -35,7 +35,8 @@ static void appendToGlobalArray(StringRef ArrayName, Module &M, Function *F,
   // to the list.
   SmallVector<Constant *, 16> CurrentCtors;
   StructType *EltTy;
-  if (GlobalVariable *GVCtor = M.getNamedGlobal(ArrayName)) {
+  GlobalVariable *GVCtor = M.getNamedGlobal(ArrayName);
+  if (GVCtor) {
     EltTy = cast<StructType>(GVCtor->getValueType()->getArrayElementType());
     if (Constant *Init = GVCtor->getInitializer()) {
       unsigned n = Init->getNumOperands();
@@ -43,7 +44,7 @@ static void appendToGlobalArray(StringRef ArrayName, Module &M, Function *F,
       for (unsigned i = 0; i != n; ++i)
         CurrentCtors.push_back(cast<Constant>(Init->getOperand(i)));
     }
-    GVCtor->eraseFromParent();
+    GVCtor->removeFromParent();
   } else {
     EltTy = StructType::get(IRB.getInt32Ty(),
                             PointerType::get(FnTy, F->getAddressSpace()),
@@ -67,8 +68,13 @@ static void appendToGlobalArray(StringRef ArrayName, Module &M, Function *F,
 
   // Create the new global variable and replace all uses of
   // the old global variable with the new one.
-  (void)new GlobalVariable(M, NewInit->getType(), false,
-                           GlobalValue::AppendingLinkage, NewInit, ArrayName);
+  auto *NewGVCtor =
+      new GlobalVariable(M, NewInit->getType(), false,
+                         GlobalValue::AppendingLinkage, NewInit, ArrayName);
+  if (GVCtor)
+    for (Use &U : make_early_inc_range(GVCtor->uses()))
+      U.set(NewGVCtor);
+  delete GVCtor;
 }
 
 void llvm::appendToGlobalCtors(Module &M, Function *F, int Priority, Constant *Data) {
diff --git a/llvm/test/Verifier/global-ctors-dtors-uses.ll b/llvm/test/Verifier/global-ctors-dtors-uses.ll
deleted file mode 100644
index 1af4fb7ca9c0e..0000000000000
--- a/llvm/test/Verifier/global-ctors-dtors-uses.ll
+++ /dev/null
@@ -1,16 +0,0 @@
-; RUN: not llvm-as %s -o /dev/null 2>&1 | FileCheck %s
-
-; CHECK: invalid uses of intrinsic global variable
-; CHECK-NEXT: ptr @llvm.global_ctors
- at llvm.global_ctors = appending global [1 x { i32, ptr, ptr } ] [
-  { i32, ptr, ptr } { i32 65535, ptr null, ptr null }
-]
-
-; CHECK: invalid uses of intrinsic global variable
-; CHECK-NEXT: ptr @llvm.global_dtors
- at llvm.global_dtors = appending global [1 x { i32, ptr, ptr } ] [
-  { i32, ptr, ptr } { i32 65535, ptr null, ptr null }
-]
-
- at ctor_user = global ptr @llvm.global_ctors
- at dtor_user = global ptr @llvm.global_dtors