8155980: ARM InterpreterMacroAssembler::get_method_counters() should not be saving caller saved registers

Make get_method_counters() only save registers specified by caller, not all of them

Reviewed-by: dlong, jiangli

hotspot/src/cpu/arm/vm/abstractInterpreter_arm.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2008, 2016, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2008, 2017, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -234,8 +234,15 @@ void AbstractInterpreter::layout_activation(Method* method,
 #ifdef AARCH64
   interpreter_frame->interpreter_frame_set_stack_top(stack_top);
 
+  // We have to add extra reserved slots to max_stack. There are 3 users of the extra slots,
+  // none of which are at the same time, so we just need to make sure there is enough room
+  // for the biggest user:
+  //   -reserved slot for exception handler
+  //   -reserved slots for JSR292. Method::extra_stack_entries() is the size.
+  //   -3 reserved slots so get_method_counters() can save some registers before call_VM().
+  int max_stack = method->constMethod()->max_stack() + MAX2(3, Method::extra_stack_entries());
   intptr_t* extended_sp = (intptr_t*) monbot  -
-    (method->max_stack() + 1) * Interpreter::stackElementWords - // +1 is reserved slot for exception handler
+    (max_stack * Interpreter::stackElementWords) -
     popframe_extra_args;
   extended_sp = (intptr_t*)round_down((intptr_t)extended_sp, StackAlignmentInBytes);
   interpreter_frame->interpreter_frame_set_extended_sp(extended_sp);

hotspot/src/cpu/arm/vm/interp_masm_arm.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2008, 2016, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2008, 2017, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -2195,75 +2195,42 @@ void InterpreterMacroAssembler::increment_mask_and_jump(Address counter_addr,
 
 void InterpreterMacroAssembler::get_method_counters(Register method,
                                                     Register Rcounters,
-                                                    Label& skip) {
+                                                    Label& skip,
+                                                    bool saveRegs,
+                                                    Register reg1,
+                                                    Register reg2,
+                                                    Register reg3) {
   const Address method_counters(method, Method::method_counters_offset());
   Label has_counters;
 
   ldr(Rcounters, method_counters);
   cbnz(Rcounters, has_counters);
 
+  if (saveRegs) {
+    // Save and restore in use caller-saved registers since they will be trashed by call_VM
+    assert(reg1 != noreg, "must specify reg1");
+    assert(reg2 != noreg, "must specify reg2");
 #ifdef AARCH64
-  const Register tmp = Rcounters;
+    assert(reg3 != noreg, "must specify reg3");
-  const int saved_regs_size = 20*wordSize;
+    stp(reg1, reg2, Address(Rstack_top, -2*wordSize, pre_indexed));
-
+    stp(reg3, ZR, Address(Rstack_top, -2*wordSize, pre_indexed));
-  // Note: call_VM will cut SP according to Rstack_top value before call, and restore SP to
-  // extended_sp value from frame after the call.
-  // So make sure there is enough stack space to save registers and adjust Rstack_top accordingly.
-  {
-    Label enough_stack_space;
-    check_extended_sp(tmp);
-    sub(Rstack_top, Rstack_top, saved_regs_size);
-    cmp(SP, Rstack_top);
-    b(enough_stack_space, ls);
-
-    align_reg(tmp, Rstack_top, StackAlignmentInBytes);
-    mov(SP, tmp);
-    str(tmp, Address(FP, frame::interpreter_frame_extended_sp_offset * wordSize));
-
-    bind(enough_stack_space);
-    check_stack_top();
-
-    int offset = 0;
-    stp(R0,  R1,  Address(Rstack_top, offset)); offset += 2*wordSize;
-    stp(R2,  R3,  Address(Rstack_top, offset)); offset += 2*wordSize;
-    stp(R4,  R5,  Address(Rstack_top, offset)); offset += 2*wordSize;
-    stp(R6,  R7,  Address(Rstack_top, offset)); offset += 2*wordSize;
-    stp(R8,  R9,  Address(Rstack_top, offset)); offset += 2*wordSize;
-    stp(R10, R11, Address(Rstack_top, offset)); offset += 2*wordSize;
-    stp(R12, R13, Address(Rstack_top, offset)); offset += 2*wordSize;
-    stp(R14, R15, Address(Rstack_top, offset)); offset += 2*wordSize;
-    stp(R16, R17, Address(Rstack_top, offset)); offset += 2*wordSize;
-    stp(R18, LR,  Address(Rstack_top, offset)); offset += 2*wordSize;
-    assert (offset == saved_regs_size, "should be");
-  }
 #else
-  push(RegisterSet(R0, R3) | RegisterSet(R12) | RegisterSet(R14));
+    assert(reg3 == noreg, "must not specify reg3");
-#endif // AARCH64
+    push(RegisterSet(reg1) | RegisterSet(reg2));
+#endif
+  }
 
   mov(R1, method);
-  call_VM(noreg, CAST_FROM_FN_PTR(address,
+  call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::build_method_counters), R1);
-          InterpreterRuntime::build_method_counters), R1);
 
+  if (saveRegs) {
 #ifdef AARCH64
-  {
+    ldp(reg3, ZR, Address(Rstack_top, 2*wordSize, post_indexed));
-    int offset = 0;
+    ldp(reg1, reg2, Address(Rstack_top, 2*wordSize, post_indexed));
-    ldp(R0,  R1,  Address(Rstack_top, offset)); offset += 2*wordSize;
-    ldp(R2,  R3,  Address(Rstack_top, offset)); offset += 2*wordSize;
-    ldp(R4,  R5,  Address(Rstack_top, offset)); offset += 2*wordSize;
-    ldp(R6,  R7,  Address(Rstack_top, offset)); offset += 2*wordSize;
-    ldp(R8,  R9,  Address(Rstack_top, offset)); offset += 2*wordSize;
-    ldp(R10, R11, Address(Rstack_top, offset)); offset += 2*wordSize;
-    ldp(R12, R13, Address(Rstack_top, offset)); offset += 2*wordSize;
-    ldp(R14, R15, Address(Rstack_top, offset)); offset += 2*wordSize;
-    ldp(R16, R17, Address(Rstack_top, offset)); offset += 2*wordSize;
-    ldp(R18, LR,  Address(Rstack_top, offset)); offset += 2*wordSize;
-    assert (offset == saved_regs_size, "should be");
-
-    add(Rstack_top, Rstack_top, saved_regs_size);
-  }
 #else
-  pop(RegisterSet(R0, R3) | RegisterSet(R12) | RegisterSet(R14));
+    pop(RegisterSet(reg1) | RegisterSet(reg2));
-#endif // AARCH64
+#endif
+  }
 
   ldr(Rcounters, method_counters);
   cbz(Rcounters, skip); // No MethodCounters created, OutOfMemory

hotspot/src/cpu/arm/vm/interp_masm_arm.hpp

@@ -349,7 +349,13 @@ class InterpreterMacroAssembler: public MacroAssembler {
 
   void trace_state(const char* msg) PRODUCT_RETURN;
 
-  void get_method_counters(Register method, Register Rcounters, Label& skip);
+void get_method_counters(Register method,
+                         Register Rcounters,
+                         Label& skip,
+                         bool saveRegs = false,
+                         Register reg1 = noreg,
+                         Register reg2 = noreg,
+                         Register reg3 = noreg);
 };
 
 #endif // CPU_ARM_VM_INTERP_MASM_ARM_HPP

hotspot/src/cpu/arm/vm/templateInterpreterGenerator_arm.cpp

@@ -1401,7 +1401,13 @@ address TemplateInterpreterGenerator::generate_normal_entry(bool synchronized) {
 #ifdef AARCH64
   // setup RmaxStack
   __ ldrh(RmaxStack, Address(RconstMethod, ConstMethod::max_stack_offset()));
-  __ add(RmaxStack, RmaxStack, MAX2(1, Method::extra_stack_entries())); // reserve slots for exception handler and JSR292 appendix argument
+  // We have to add extra reserved slots to max_stack. There are 3 users of the extra slots,
+  // none of which are at the same time, so we just need to make sure there is enough room
+  // for the biggest user:
+  //   -reserved slot for exception handler
+  //   -reserved slots for JSR292. Method::extra_stack_entries() is the size.
+  //   -3 reserved slots so get_method_counters() can save some registers before call_VM().
+  __ add(RmaxStack, RmaxStack, MAX2(3, Method::extra_stack_entries()));
 #endif // AARCH64
 
   // see if we've got enough room on the stack for locals plus overhead.

hotspot/src/cpu/arm/vm/templateTable_arm.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2008, 2016, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2008, 2017, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -2286,13 +2286,18 @@ void TemplateTable::branch(bool is_jsr, bool is_wide) {
       }
       __ bind(no_mdo);
       // Increment backedge counter in MethodCounters*
-      __ get_method_counters(Rmethod, Rcounters, dispatch);
+      // Note Rbumped_taken_count is a callee saved registers for ARM32, but caller saved for ARM64
+      __ get_method_counters(Rmethod, Rcounters, dispatch, true /*saveRegs*/,
+                             Rdisp, R3_bytecode,
+                             AARCH64_ONLY(Rbumped_taken_count) NOT_AARCH64(noreg));
       const Address mask(Rcounters, in_bytes(MethodCounters::backedge_mask_offset()));
       __ increment_mask_and_jump(Address(Rcounters, be_offset), increment, mask,
                                  Rcnt, R4_tmp, eq, &backedge_counter_overflow);
     } else {
-      // increment counter
+      // Increment backedge counter in MethodCounters*
-      __ get_method_counters(Rmethod, Rcounters, dispatch);
+      __ get_method_counters(Rmethod, Rcounters, dispatch, true /*saveRegs*/,
+                             Rdisp, R3_bytecode,
+                             AARCH64_ONLY(Rbumped_taken_count) NOT_AARCH64(noreg));
       __ ldr_u32(Rtemp, Address(Rcounters, be_offset));           // load backedge counter
       __ add(Rtemp, Rtemp, InvocationCounter::count_increment);   // increment counter
       __ str_32(Rtemp, Address(Rcounters, be_offset));            // store counter