[clang][IR] Overload @llvm.thread.pointer to support non-AS0 targets (#132489)

Thread-local globals live, by default, in the default globals address
space, which may not be 0, so we need to overload @llvm.thread.pointer
to support other address spaces, and use the default globals address
space in Clang.

Author: jrtc27

clang/lib/CodeGen/CGBuiltin.cpp

@@ -6115,8 +6115,9 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
   case Builtin::BI__builtin_thread_pointer: {
     if (!getContext().getTargetInfo().isTLSSupported())
       CGM.ErrorUnsupported(E, "__builtin_thread_pointer");
-    // Fall through - it's already mapped to the intrinsic by ClangBuiltin.
-    break;
+
+    return RValue::get(Builder.CreateIntrinsic(llvm::Intrinsic::thread_pointer,
+                                               {GlobalsInt8PtrTy}, {}));
   }
   case Builtin::BI__builtin_os_log_format:
     return emitBuiltinOSLogFormat(*E);

clang/test/CodeGen/builtins-arm64.c

@@ -10,7 +10,7 @@ void f0(void *a, void *b) {
 
 void *tp (void) {
   return __builtin_thread_pointer ();
-// CHECK-LINUX: call {{.*}} @llvm.thread.pointer()
+// CHECK-LINUX: call {{.*}} @llvm.thread.pointer.p0()
 }
 
 // CHECK: call {{.*}} @llvm.bitreverse.i32(i32 %a)

clang/test/CodeGen/builtins-wasm.c

@@ -749,5 +749,5 @@ int externref_is_null(__externref_t arg) {
 
 void *tp (void) {
   return __builtin_thread_pointer ();
-  // WEBASSEMBLY: call {{.*}} @llvm.thread.pointer()
+  // WEBASSEMBLY: call {{.*}} @llvm.thread.pointer.p0()
 }

llvm/docs/LangRef.rst

@@ -3133,6 +3133,9 @@ as follows:
     program memory space defaults to the default address space of 0,
     which corresponds to a Von Neumann architecture that has code
     and data in the same space.
+
+.. _globals_addrspace:
+
 ``G<address space>``
     Specifies the address space to be used by default when creating global
     variables. If omitted, the globals address space defaults to the default
@@ -15061,7 +15064,8 @@ Syntax:
 
 ::
 
-      declare ptr @llvm.thread.pointer()
+      declare ptr @llvm.thread.pointer.p0()
+      declare ptr addrspace(5) @llvm.thread.pointer.p5()
 
 Overview:
 """""""""
@@ -15078,7 +15082,8 @@ specific: it may point to the start of TLS area, to the end, or somewhere
 in the middle.  Depending on the target, this intrinsic may read a register,
 call a helper function, read from an alternate memory space, or perform
 other operations necessary to locate the TLS area.  Not all targets support
-this intrinsic.
+this intrinsic.  The address space must be the :ref:`globals address space
+<globals_addrspace>`.
 
 '``llvm.call.preallocated.setup``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

llvm/include/llvm/IR/Intrinsics.td

@@ -903,7 +903,7 @@ def int_stackrestore  : DefaultAttrsIntrinsic<[], [llvm_anyptr_ty]>,
 
 def int_get_dynamic_area_offset : DefaultAttrsIntrinsic<[llvm_anyint_ty]>;
 
-def int_thread_pointer : DefaultAttrsIntrinsic<[llvm_ptr_ty], [], [IntrNoMem]>,
+def int_thread_pointer : DefaultAttrsIntrinsic<[llvm_anyptr_ty], [], [IntrNoMem]>,
                          ClangBuiltin<"__builtin_thread_pointer">;
 
 // IntrInaccessibleMemOrArgMemOnly is a little more pessimistic than strictly

llvm/lib/IR/AutoUpgrade.cpp

@@ -646,8 +646,8 @@ static bool upgradeArmOrAarch64IntrinsicFunction(bool IsArm, Function *F,
 
   if (Name == "thread.pointer") {
     // '(arm|aarch64).thread.pointer'.
-    NewFn = Intrinsic::getOrInsertDeclaration(F->getParent(),
-                                              Intrinsic::thread_pointer);
+    NewFn = Intrinsic::getOrInsertDeclaration(
+        F->getParent(), Intrinsic::thread_pointer, F->getReturnType());
     return true;
   }
 
@@ -1475,6 +1475,14 @@ static bool upgradeIntrinsicFunction1(Function *F, Function *&NewFn,
     }
     break;
 
+  case 't':
+    if (Name == "thread.pointer") {
+      NewFn = Intrinsic::getDeclaration(
+          F->getParent(), Intrinsic::thread_pointer, F->getReturnType());
+      return true;
+    }
+    break;
+
   case 'v': {
     if (Name == "var.annotation" && F->arg_size() == 4) {
       rename(F);

llvm/lib/IR/Verifier.cpp

@@ -6553,6 +6553,14 @@ void Verifier::visitIntrinsicCall(Intrinsic::ID ID, CallBase &Call) {
           &Call);
     break;
   }
+  case Intrinsic::thread_pointer: {
+    Check(Call.getType()->getPointerAddressSpace() ==
+              DL.getDefaultGlobalsAddressSpace(),
+          "llvm.thread.pointer intrinsic return type must be for the globals "
+          "address space",
+          &Call);
+    break;
+  }
   case Intrinsic::threadlocal_address: {
     const Value &Arg0 = *Call.getArgOperand(0);
     Check(isa<GlobalValue>(Arg0),

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -28169,8 +28169,8 @@ bool AArch64TargetLowering::shouldNormalizeToSelectSequence(LLVMContext &,
 
 static Value *UseTlsOffset(IRBuilderBase &IRB, unsigned Offset) {
   Module *M = IRB.GetInsertBlock()->getParent()->getParent();
-  Function *ThreadPointerFunc =
-      Intrinsic::getOrInsertDeclaration(M, Intrinsic::thread_pointer);
+  Function *ThreadPointerFunc = Intrinsic::getOrInsertDeclaration(
+      M, Intrinsic::thread_pointer, IRB.getPtrTy());
   return IRB.CreatePointerCast(
       IRB.CreateConstGEP1_32(IRB.getInt8Ty(), IRB.CreateCall(ThreadPointerFunc),
                              Offset),

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -23431,8 +23431,8 @@ bool RISCVTargetLowering::preferScalarizeSplat(SDNode *N) const {
 
 static Value *useTpOffset(IRBuilderBase &IRB, unsigned Offset) {
   Module *M = IRB.GetInsertBlock()->getModule();
-  Function *ThreadPointerFunc =
-      Intrinsic::getOrInsertDeclaration(M, Intrinsic::thread_pointer);
+  Function *ThreadPointerFunc = Intrinsic::getOrInsertDeclaration(
+      M, Intrinsic::thread_pointer, IRB.getPtrTy());
   return IRB.CreateConstGEP1_32(IRB.getInt8Ty(),
                                 IRB.CreateCall(ThreadPointerFunc), Offset);
 }

llvm/lib/Transforms/Utils/MemoryTaggingSupport.cpp

@@ -290,8 +290,9 @@ Value *getAndroidSlotPtr(IRBuilder<> &IRB, int Slot) {
   Module *M = IRB.GetInsertBlock()->getParent()->getParent();
   // Android provides a fixed TLS slot for sanitizers. See TLS_SLOT_SANITIZER
   // in Bionic's libc/private/bionic_tls.h.
-  Function *ThreadPointerFunc =
-      Intrinsic::getOrInsertDeclaration(M, Intrinsic::thread_pointer);
+  Function *ThreadPointerFunc = Intrinsic::getOrInsertDeclaration(
+      M, Intrinsic::thread_pointer,
+      IRB.getPtrTy(M->getDataLayout().getDefaultGlobalsAddressSpace()));
   return IRB.CreateConstGEP1_32(IRB.getInt8Ty(),
                                 IRB.CreateCall(ThreadPointerFunc), 8 * Slot);
 }