[WebAssembly] Implement %llvm.thread.pointer intrinsic (#117817)

We can simply use the `__tls_base` global for this which is guaranteed
to be non-zero and unique per thread.

Fixes: #117433

Author: sbc100

clang/test/CodeGen/builtins-wasm.c

@@ -740,3 +740,8 @@ __externref_t externref_null() {
   // WEBASSEMBLY: tail call ptr addrspace(10) @llvm.wasm.ref.null.extern()
   // WEBASSEMBLY-NEXT: ret
 }
+
+void *tp (void) {
+  return __builtin_thread_pointer ();
+  // WEBASSEMBLY: call {{.*}} @llvm.thread.pointer()
+}

llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

@@ -2089,6 +2089,17 @@ SDValue WebAssemblyTargetLowering::LowerIntrinsic(SDValue Op,
     }
     return DAG.getNode(WebAssemblyISD::SHUFFLE, DL, Op.getValueType(), Ops);
   }
+
+  case Intrinsic::thread_pointer: {
+    MVT PtrVT = getPointerTy(DAG.getDataLayout());
+    auto GlobalGet = PtrVT == MVT::i64 ? WebAssembly::GLOBAL_GET_I64
+                                       : WebAssembly::GLOBAL_GET_I32;
+    const char *TlsBase = MF.createExternalSymbolName("__tls_base");
+    return SDValue(
+        DAG.getMachineNode(GlobalGet, DL, PtrVT,
+                           DAG.getTargetExternalSymbol(TlsBase, PtrVT)),
+        0);
+  }
   }
 }
 

llvm/test/CodeGen/WebAssembly/thread_pointer.ll

@@ -0,0 +1,21 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=wasm32-unknown-unknown | FileCheck %s --check-prefix=WASM32
+; RUN: llc < %s -mtriple=wasm64-unknown-unknown | FileCheck %s --check-prefix=WASM64
+
+declare ptr @llvm.thread.pointer()
+
+define ptr @thread_pointer() nounwind {
+; WASM32-LABEL: thread_pointer:
+; WASM32:         .functype thread_pointer () -> (i32)
+; WASM32-NEXT:  # %bb.0:
+; WASM32-NEXT:    global.get __tls_base
+; WASM32-NEXT:    # fallthrough-return
+;
+; WASM64-LABEL: thread_pointer:
+; WASM64:         .functype thread_pointer () -> (i64)
+; WASM64-NEXT:  # %bb.0:
+; WASM64-NEXT:    global.get __tls_base
+; WASM64-NEXT:    # fallthrough-return
+  %1 = tail call ptr @llvm.thread.pointer()
+  ret ptr %1
+}