[CGData][Merger] Avoid merging the attached call target

[kyulee-com]

For global function merging, the target of the arc-attached call must be a constant and cannot be parameterized.
This change adds a check to bypass this case in canParameterizeCallOperand().

[kyulee-com]

cc. @nocchijiang

[llvmbot]

@llvm/pr-subscribers-backend-aarch64

Author: Kyungwoo Lee (kyulee-com)

Changes

For global function merging, the target of the arc-attached call must be a constant and cannot be parameterized.
This change adds a check to bypass this case in canParameterizeCallOperand().

---

Full diff: https://github.com/llvm/llvm-project/pull/121030.diff

2 Files Affected:

• (modified) llvm/lib/CodeGen/GlobalMergeFunctions.cpp (+9-3)
• (added) llvm/test/CodeGen/AArch64/cgdata-no-merge-attached-call-garget.ll (+37)

diff --git a/llvm/lib/CodeGen/GlobalMergeFunctions.cpp b/llvm/lib/CodeGen/GlobalMergeFunctions.cpp
index c8f1b98c9a18e9..b5863e0551b78c 100644
--- a/llvm/lib/CodeGen/GlobalMergeFunctions.cpp
+++ b/llvm/lib/CodeGen/GlobalMergeFunctions.cpp
@@ -60,11 +60,17 @@ static bool canParameterizeCallOperand(const CallBase *CI, unsigned OpIdx) {
     if (Name.starts_with("__dtrace"))
       return false;
   }
-  if (isCalleeOperand(CI, OpIdx) &&
-      CI->getOperandBundle(LLVMContext::OB_ptrauth).has_value()) {
+  if (isCalleeOperand(CI, OpIdx)) {
     // The operand is the callee and it has already been signed. Ignore this
     // because we cannot add another ptrauth bundle to the call instruction.
-    return false;
+    if (CI->getOperandBundle(LLVMContext::OB_ptrauth).has_value())
+      return false;
+  } else {
+    // The target of the arc-attached call must be a constant and cannot be
+    // parameterized.
+    if (CI->isOperandBundleOfType(llvm::LLVMContext::OB_clang_arc_attachedcall,
+                                  OpIdx))
+      return false;
   }
   return true;
 }
diff --git a/llvm/test/CodeGen/AArch64/cgdata-no-merge-attached-call-garget.ll b/llvm/test/CodeGen/AArch64/cgdata-no-merge-attached-call-garget.ll
new file mode 100644
index 00000000000000..1163314062df1b
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/cgdata-no-merge-attached-call-garget.ll
@@ -0,0 +1,37 @@
+; This test verifies that two similar functions, f1 and f2, are not merged
+; when their attached call targets differ, since these targets cannot be parameterized.
+
+; RUN: llc -mtriple=arm64-apple-darwin -enable-global-merge-func=true < %s | FileCheck %s
+
+; CHECK-NOT: _f1.Tgm
+; CHECK-NOT: _f2.Tgm
+
+target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
+target triple = "arm64-apple-darwin"
+
+define i64 @f1(ptr %0) {
+  %2 = call ptr @g1(ptr %0, i32 0) minsize [ "clang.arc.attachedcall"(ptr @llvm.objc.unsafeClaimAutoreleasedReturnValue) ]
+  tail call void (...) @llvm.objc.clang.arc.noop.use(ptr %2)
+  %3 = call i64 @g2(ptr %2)
+  tail call void @objc_release(ptr %2)
+  %4 = tail call i64 @g3(i64 %3)
+  ret i64 %4
+}
+
+define i64 @f2(ptr %0) {
+  %2 = call ptr @g1(ptr %0, i32 0) minsize [ "clang.arc.attachedcall"(ptr @llvm.objc.retainAutoreleasedReturnValue) ]
+  tail call void (...) @llvm.objc.clang.arc.noop.use(ptr %2)
+  %3 = call i64 @g2(ptr %2)
+  tail call void @objc_release(ptr %2)
+  %4 = tail call i64 @g3(i64 %3)
+  ret i64 %4
+}
+
+declare ptr @g1(ptr, i32)
+declare i64 @g2(ptr)
+declare i64 @g3(i64)
+
+declare void @llvm.objc.clang.arc.noop.use(...)
+declare ptr @llvm.objc.unsafeClaimAutoreleasedReturnValue(ptr)
+declare ptr @llvm.objc.retainAutoreleasedReturnValue(ptr)
+declare void @objc_release(ptr)

[nocchijiang]

Suggested change

	if (CI->isOperandBundleOfType(llvm::LLVMContext::OB_clang_arc_attachedcall,
	if (CI->isOperandBundleOfType(LLVMContext::OB_clang_arc_attachedcall,

to be consistent with CI->getOperandBundle(LLVMContext::OB_ptrauth) above.

[nocchijiang]

Is there a general method to determine whether a bundle is unsafe to optimize? If so, we can avoid enumerating all of them.

[kyulee-com]

Not familiar with the bundle. Conservatively, we could bail out all bundle cases, but for now I bail out the known case only. I think if we run into invalid cases, the compilation will fail/crash so we can follow up.

[kyulee-com]

Any other comments? This is relatively a straightforward fix to unblock compiler crash.

[nocchijiang]

LGTM, but hope that someone else with commit access to review & approve before merging.

[kyulee-com]

@ellishg @thevinster Do you have any comments?