[clang][CodeGen] Fix crash on non-natural type in CheckAtomicAlignment llvm#141053

Author: dsalinas_amdeng

clang/lib/CodeGen/CGBuiltin.cpp

@@ -359,9 +359,10 @@ static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V,
 static Address CheckAtomicAlignment(CodeGenFunction &CGF, const CallExpr *E) {
   ASTContext &Ctx = CGF.getContext();
   Address Ptr = CGF.EmitPointerWithAlignment(E->getArg(0));
+  const llvm::DataLayout &DL = CGF.CGM.getDataLayout();
   unsigned Bytes = Ptr.getElementType()->isPointerTy()
                        ? Ctx.getTypeSizeInChars(Ctx.VoidPtrTy).getQuantity()
-                       : Ptr.getElementType()->getScalarSizeInBits() / 8;
+                       : DL.getTypeStoreSize(Ptr.getElementType());
   unsigned Align = Ptr.getAlignment().getQuantity();
   if (Align % Bytes != 0) {
     DiagnosticsEngine &Diags = CGF.CGM.getDiags();

clang/test/CodeGenOpenCL/check-atomic-alignment.cl

@@ -0,0 +1,41 @@
+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 5
+// RUN: %clang_cc1 -O0 -cl-std=CL2.0 -triple amdgcn-amd-amdhsa -target-cpu gfx942 \
+// RUN:   %s -emit-llvm -o - -disable-llvm-passes | FileCheck %s
+
+// REQUIRES: amdgpu-registered-target
+
+// `Ptr.getElementType()` in `CheckAtomicAlignment` returns
+//      %struct.__half2 = type { %union.anon }
+// Check we do not crash when handling that.
+
+typedef half  __attribute__((ext_vector_type(2))) half2;
+typedef short  __attribute__((ext_vector_type(2))) short2;
+
+struct __half2 {
+    union {
+        struct {
+            half x;
+            half y;
+        };
+        half2 data;
+    };
+};
+
+// CHECK-LABEL: define dso_local <2 x half> @test_flat_add_2f16(
+// CHECK-SAME: ptr noundef [[ADDR:%.*]], <2 x half> noundef [[VAL:%.*]]) #[[ATTR0:[0-9]+]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    [[ADDR_ADDR:%.*]] = alloca ptr, align 8, addrspace(5)
+// CHECK-NEXT:    [[VAL_ADDR:%.*]] = alloca <2 x half>, align 4, addrspace(5)
+// CHECK-NEXT:    store ptr [[ADDR]], ptr addrspace(5) [[ADDR_ADDR]], align 8
+// CHECK-NEXT:    store <2 x half> [[VAL]], ptr addrspace(5) [[VAL_ADDR]], align 4
+// CHECK-NEXT:    [[TMP0:%.*]] = load ptr, ptr addrspace(5) [[ADDR_ADDR]], align 8
+// CHECK-NEXT:    [[TMP1:%.*]] = load <2 x half>, ptr addrspace(5) [[VAL_ADDR]], align 4
+// CHECK-NEXT:    [[TMP2:%.*]] = atomicrmw fadd ptr [[TMP0]], <2 x half> [[TMP1]] syncscope("agent") monotonic, align 4, !amdgpu.no.fine.grained.memory [[META4:![0-9]+]]
+// CHECK-NEXT:    ret <2 x half> [[TMP2]]
+//
+half2 test_flat_add_2f16(short2 *addr, half2 val) {
+  return __builtin_amdgcn_flat_atomic_fadd_v2f16((struct __half2*)addr, val);
+}
+//.
+// CHECK: [[META4]] = !{}
+//.