8388170: [s390x] compiler/inlining/LateInlineQueueDrainTest.java#id0 fail due to invalid ret_addr_offset

Reviewed-by: amitkumar, aph

src/hotspot/cpu/aarch64/aarch64.ad

@@ -1662,19 +1662,19 @@ bool needs_acquiring_load_exclusive(const Node *n)
 //       from the start of the call to the point where the return address
 //       will point.
 
-int MachCallStaticJavaNode::ret_addr_offset()
+int MachCallStaticJavaNode::ret_addr_offset() const
 {
   // call should be a simple bl
   int off = 4;
   return off;
 }
 
-int MachCallDynamicJavaNode::ret_addr_offset()
+int MachCallDynamicJavaNode::ret_addr_offset() const
 {
   return 16; // movz, movk, movk, bl
 }
 
-int MachCallRuntimeNode::ret_addr_offset() {
+int MachCallRuntimeNode::ret_addr_offset() const {
   // for generated stubs the call will be
   //   bl(addr)
   // or with far branches

src/hotspot/cpu/arm/arm_32.ad

@@ -430,18 +430,18 @@ OptoRegPair c2::return_value(int ideal_reg) {
 //       from the start of the call to the point where the return address
 //       will point.
 
-int MachCallStaticJavaNode::ret_addr_offset() {
+int MachCallStaticJavaNode::ret_addr_offset() const {
   bool far = (_method == nullptr) ? maybe_far_call(this) : !cache_reachable();
   return (far ? 3 : 1) * NativeInstruction::instruction_size;
 }
 
-int MachCallDynamicJavaNode::ret_addr_offset() {
+int MachCallDynamicJavaNode::ret_addr_offset() const {
   bool far = !cache_reachable();
   // mov_oop is always 2 words
   return (2 + (far ? 3 : 1)) * NativeInstruction::instruction_size;
 }
 
-int MachCallRuntimeNode::ret_addr_offset() {
+int MachCallRuntimeNode::ret_addr_offset() const {
   // bl or movw; movt; blx
   bool far = maybe_far_call(this);
   return (far ? 3 : 1) * NativeInstruction::instruction_size;

src/hotspot/cpu/ppc/ppc.ad

@@ -1171,16 +1171,16 @@ bool followed_by_acquire(const Node *load) {
 
 // PPC port: Removed use of lazy constant construct.
 
-int MachCallStaticJavaNode::ret_addr_offset() {
+int MachCallStaticJavaNode::ret_addr_offset() const {
   // It's only a single branch-and-link instruction.
   return 4;
 }
 
-int MachCallDynamicJavaNode::ret_addr_offset() {
+int MachCallDynamicJavaNode::ret_addr_offset() const {
   return 12;
 }
 
-int MachCallRuntimeNode::ret_addr_offset() {
+int MachCallRuntimeNode::ret_addr_offset() const {
   if (rule() == CallRuntimeDirect_rule) {
     // CallRuntimeDirectNode uses call_c.
 #if defined(ABI_ELFv2)

src/hotspot/cpu/riscv/riscv.ad

@@ -1208,17 +1208,17 @@ bool needs_acquiring_load_reserved(const Node *n)
 //       from the start of the call to the point where the return address
 //       will point.
 
-int MachCallStaticJavaNode::ret_addr_offset()
+int MachCallStaticJavaNode::ret_addr_offset() const
 {
   return 3 * NativeInstruction::instruction_size; // auipc + ld + jalr
 }
 
-int MachCallDynamicJavaNode::ret_addr_offset()
+int MachCallDynamicJavaNode::ret_addr_offset() const
 {
   return NativeMovConstReg::movptr2_instruction_size + (3 * NativeInstruction::instruction_size); // movptr2, auipc + ld + jal
 }
 
-int MachCallRuntimeNode::ret_addr_offset() {
+int MachCallRuntimeNode::ret_addr_offset() const {
   // For address inside the code cache the call will be:
   //   auipc + jalr
   // For real runtime callouts it will be 8 instructions

src/hotspot/cpu/s390/s390.ad

@@ -886,29 +886,59 @@ int MachNode::compute_padding(int current_offset) const {
   return 0;
 }
 
-int MachCallStaticJavaNode::ret_addr_offset() {
+int MachCallStaticJavaNode::ret_addr_offset() const {
   if (_method) {
-    return 8;
+    return MacroAssembler::call_far_pcrelative_size();
   } else {
     return MacroAssembler::call_far_patchable_ret_addr_offset();
   }
 }
 
-int MachCallDynamicJavaNode::ret_addr_offset() {
+int MachCallDynamicJavaNode::ret_addr_offset() const {
   // Consider size of receiver type profiling (C2 tiers).
-  int profile_receiver_type_size = 0;
 
   int vtable_index = this->_vtable_index;
   if (vtable_index == -4) {
-    return 14 + profile_receiver_type_size;
+    return MacroAssembler::load_const_from_toc_size()
+         + MacroAssembler::call_far_pcrelative_size();
   } else {
     assert(!UseInlineCaches, "expect vtable calls only if not using ICs");
-    return 36 + profile_receiver_type_size;
+    // This should return the size of instructions in vtable dispatch
+    // branch of z_enc_java_dynamic_call
+    int offset = 0;
+
+    //     __ load_klass(Z_method, Z_R2);
+    if (UseCompactObjectHeaders) {
+      // load_narrow_klass_compact (z_lg z_srlg)
+      offset += 6 // z_lg
+              + 6; // z_srlg;
+    } else {
+      offset += 6; // z_llgf
+    }
+    offset += MacroAssembler::instr_size_for_decode_klass_not_null();
+
+    // check if displacement is valid, as it will generate different
+    // instructions:
+    int entry_offset = in_bytes(Klass::vtable_start_offset()) +
+                       vtable_index * vtableEntry::size_in_bytes();
+    int v_off = entry_offset + in_bytes(vtableEntry::method_offset());
+    if (!Displacement::is_validDisp(v_off)) {
+        offset += MacroAssembler::load_const_size(); // emits iihf + iilf
+    }
+    // both generate z_lg
+    offset += 6; // z_lg (z_method, v_off | Address(Z_method, Z_R1_scratch))
+    // common footer
+    offset += 6; // z_lg(Z_R1_scratch, Method::from_compiled_offset())
+    offset += 2; // z_basr
+
+    return offset;
   }
 }
 
-int MachCallRuntimeNode::ret_addr_offset() {
-  return 12 + MacroAssembler::call_far_patchable_ret_addr_offset();
+int MachCallRuntimeNode::ret_addr_offset() const {
+  return 6 // get_PC()  (LARL)
+    + 6 // save_return_pc() (STG)
+    + MacroAssembler::call_far_patchable_ret_addr_offset();
 }
 
 // Compute padding required for nodes which need alignment
@@ -994,7 +1024,6 @@ static inline void z_assert_aligned(C2_MacroAssembler *masm, int disp, Register
 int emit_call_reloc(C2_MacroAssembler *masm, intptr_t entry_point, relocInfo::relocType rtype,
                     PhaseRegAlloc* ra_, bool is_native_call = false) {
   __ set_inst_mark(); // Used in z_enc_java_static_call() and emit_java_to_interp().
-  address old_mark = __ inst_mark();
   unsigned int start_off = __ offset();
 
   if (is_native_call) {
@@ -1024,7 +1053,6 @@ int emit_call_reloc(C2_MacroAssembler *masm, intptr_t entry_point, relocInfo::re
 
 static int emit_call_reloc(C2_MacroAssembler *masm, intptr_t entry_point, RelocationHolder const& rspec) {
   __ set_inst_mark(); // Used in z_enc_java_static_call() and emit_java_to_interp().
-  address old_mark = __ inst_mark();
   unsigned int start_off = __ offset();
 
   relocInfo::relocType rtype = rspec.type();
@@ -2344,8 +2372,7 @@ encode %{
     // callee doesn't.
     unsigned int start_off = __ offset();
     // Compute size of "larl + stg + call_c_opt".
-    const int size_of_code = 6 + 6 + MacroAssembler::call_far_patchable_size();
-    __ get_PC(Z_R14, size_of_code);
+    __ get_PC(Z_R14, ret_addr_offset());
     __ save_return_pc();
     assert(__ offset() - start_off == 12, "bad prelude len: %d", __ offset() - start_off);
 
@@ -2356,30 +2383,28 @@ encode %{
       return;
     }
 
-#ifdef ASSERT
-    // Plausibility check for size_of_code assumptions.
-    unsigned int actual_ret_off = __ offset();
-    assert(start_off + size_of_code == actual_ret_off, "wrong return_pc");
-#endif
+    assert(__ offset() - start_off == (uint)ret_addr_offset(),
+            "z_enc_java_to_runtime_call return offset mismatch: emitted %d bytes, ret_addr_offset()=%d",
+            __ offset() - start_off, ret_addr_offset());
     __ post_call_nop();
   %}
 
   enc_class z_enc_java_static_call(method meth) %{
+    unsigned int start_off = __ offset();
     // Call to fixup routine. Fixup routine uses ScopeDesc info to determine
     // whom we intended to call.
-    int ret_offset = 0;
 
     if (!_method) {
-      ret_offset = emit_call_reloc(masm, $meth$$method,
-                                   relocInfo::runtime_call_w_cp_type, ra_);
+      emit_call_reloc(masm, $meth$$method,
+                      relocInfo::runtime_call_w_cp_type, ra_);
     } else {
       int method_index = resolved_method_index(masm);
       if (_optimized_virtual) {
-        ret_offset = emit_call_reloc(masm, $meth$$method,
-                                     opt_virtual_call_Relocation::spec(method_index));
+        emit_call_reloc(masm, $meth$$method,
+                        opt_virtual_call_Relocation::spec(method_index));
       } else {
-        ret_offset = emit_call_reloc(masm, $meth$$method,
-                                     static_call_Relocation::spec(method_index));
+        emit_call_reloc(masm, $meth$$method,
+                        static_call_Relocation::spec(method_index));
       }
     }
     assert(__ inst_mark() != nullptr, "emit_call_reloc must set_inst_mark()");
@@ -2394,6 +2419,9 @@ encode %{
     }
 
     __ clear_inst_mark();
+    assert(__ offset() - start_off == (uint)ret_addr_offset(),
+            "z_enc_java_static_call return offset mismatch: emitted %d bytes, ret_addr_offset()=%d",
+            __ offset() - start_off, ret_addr_offset());
     __ post_call_nop();
   %}
 
@@ -2418,9 +2446,8 @@ encode %{
       // to determine who we intended to call.
       int method_index = resolved_method_index(masm);
       __ relocate(virtual_call_Relocation::spec(virtual_call_oop_addr, method_index));
-      unsigned int ret_off = __ offset();
       assert(__ offset() - start_off == 6, "bad prelude len: %d", __ offset() - start_off);
-      ret_off += emit_call_reloc(masm, $meth$$method, relocInfo::none, ra_);
+      emit_call_reloc(masm, $meth$$method, relocInfo::none, ra_);
       __ clear_inst_mark();
       assert(_method, "lazy_constant may be wrong when _method==null");
     } else {
@@ -2449,8 +2476,11 @@ encode %{
       __ z_lg(Z_R1_scratch, Address(Z_method, Method::from_compiled_offset()));
       // Call target. Either compiled code or C2I adapter.
       __ z_basr(Z_R14, Z_R1_scratch);
-      unsigned int ret_off = __ offset();
     }
+    assert(__ offset() - start_off == (uint)ret_addr_offset(),
+            "z_enc_java_dynamic_call return offset mismatch: emitted %d bytes, ret_addr_offset()=%d",
+            __ offset() - start_off, ret_addr_offset());
+
     __ post_call_nop();
   %}
 

src/hotspot/cpu/x86/x86.ad

@@ -1634,21 +1634,21 @@ static int clear_avx_size() {
 // !!!!! Special hack to get all types of calls to specify the byte offset
 //       from the start of the call to the point where the return address
 //       will point.
-int MachCallStaticJavaNode::ret_addr_offset()
+int MachCallStaticJavaNode::ret_addr_offset() const
 {
   int offset = 5; // 5 bytes from start of call to where return address points
   offset += clear_avx_size();
   return offset;
 }
 
-int MachCallDynamicJavaNode::ret_addr_offset()
+int MachCallDynamicJavaNode::ret_addr_offset() const
 {
   int offset = 15; // 15 bytes from start of call to where return address points
   offset += clear_avx_size();
   return offset;
 }
 
-int MachCallRuntimeNode::ret_addr_offset() {
+int MachCallRuntimeNode::ret_addr_offset() const {
   int offset = 13; // movq r10,#addr; callq (r10)
   if (this->ideal_Opcode() != Op_CallLeafVector) {
     offset += clear_avx_size();

src/hotspot/share/opto/machnode.hpp

@@ -937,7 +937,7 @@ public:
   virtual bool  pinned() const { return false; }
   virtual const Type* Value(PhaseGVN* phase) const;
   virtual const RegMask &in_RegMask(uint) const;
-  virtual int ret_addr_offset() { return 0; }
+  virtual int ret_addr_offset() const { return 0; }
 
   NOT_LP64(bool return_value_is_used() const;)
 
@@ -998,7 +998,7 @@ public:
   // If this is an uncommon trap, return the request code, else zero.
   int uncommon_trap_request() const;
 
-  virtual int ret_addr_offset();
+  virtual int ret_addr_offset() const;
 #ifndef PRODUCT
   virtual void dump_spec(outputStream *st) const;
   void dump_trap_args(outputStream *st) const;
@@ -1014,7 +1014,7 @@ public:
     init_class_id(Class_MachCallDynamicJava);
     DEBUG_ONLY(_vtable_index = -99);  // throw an assert if uninitialized
   }
-  virtual int ret_addr_offset();
+  virtual int ret_addr_offset() const;
 #ifndef PRODUCT
   virtual void dump_spec(outputStream *st) const;
 #endif
@@ -1032,7 +1032,7 @@ public:
   MachCallRuntimeNode() : MachCallNode() {
     init_class_id(Class_MachCallRuntime);
   }
-  virtual int ret_addr_offset();
+  virtual int ret_addr_offset() const;
 #ifndef PRODUCT
   virtual void dump_spec(outputStream *st) const;
 #endif