Direct call api functions (arm implementation)

src/arm/macro-assembler-arm.cc

Index: src/arm/macro-assembler-arm.cc
===================================================================
--- src/arm/macro-assembler-arm.cc	(revision 6213)
+++ src/arm/macro-assembler-arm.cc	(working copy)
@@ -1397,6 +1397,133 @@
 }
 
 
+MaybeObject* MacroAssembler::TryTailCallStub(CodeStub* stub, Condition cond) {
+  ASSERT(allow_stub_calls());  // stub calls are not allowed in some stubs
+  Object* result;
+  { MaybeObject* maybe_result = stub->TryGetCode();
+    if (!maybe_result->ToObject(&result)) return maybe_result;
+  }
+  Jump(stub->GetCode(), RelocInfo::CODE_TARGET, cond);
+  return result;
+}
+
+
+void MacroAssembler::PrepareCallApiFunction(int arg_stack_space,
+                                            int unwind_space,
+                                            Register scratch) {
+  mov(scratch, Operand(unwind_space));

[antonm, 2011/01/11 14:11:33]

it looks like a lot of code below is shared with MacroAssembler::EnterExitFrame.
 Is it possible to refactor the common code and make it more like ia32
implementation (which just invokes EnterApiExitFrame + some magic for some
platforms?

[Zaheer, 2011/01/11 15:44:35]

The behavior is different from EnterExitFrame
- v8::arguments has to be allocated above exit frame to miss gc
- The alignment has to happen after allocating space for the args and it should
handle aligned/unaligned case based on arg_stack_space (unlike EnterExitFrame
which knows the number of pushes - 5)
- exit frame pc patching has to be handled 
- argv and builtin function is n/a in this case

[antonm, 2011/01/11 19:27:40]

Zaheer,

I am not sure it's something which cannot be solved.

I do agree with Sergey and think we should allocate v8::Arguments below exit
frame (exactly like we do for ia32)---one of the reasons I asked you to refactor
the code was to make these things immediately visible.

We deal with alignment on x64 as well and it looks fine (in the worst case you
would need the same and_ trick, see PrepareCallCFunction).

And so on.

On 2011/01/11 15:44:35, zaheer wrote:

[Zaheer, 2011/01/12 13:20:03]

split the code as EnterApiExitFramePrologue/Epilogue and yes, the arguments is
below the exit frame
exit frame
arguments
<stack grows here>
I can reuse PrepareCallCFunction (instead of EnterExitFrameEpilogue) but it
seems to be creating one extra space and storing restore sp which is not
required (i.e. two instructions redundant)

+  add(ip, sp, Operand(scratch, LSL, kPointerSizeLog2));

[SeRya, 2011/01/11 16:34:50]

add(ip, sp, Operand(unwind_space * kPointerSize)); ?

[Zaheer, 2011/01/12 13:20:03]

Done

+
+  stm(db_w, sp, fp.bit() | ip.bit() | lr.bit());
+  mov(fp, Operand(sp));
+
+  mov(ip, Operand(CodeObject()));
+  push(ip);
+  push(ip);  // Exit Frame pc patched before call
+
+  // create space for the args
+  sub(sp, sp, Operand(arg_stack_space * kPointerSize));
+
+  int frame_alignment = ActivationFrameAlignment();
+  int frame_alignment_mask = frame_alignment - 1;
+  if (frame_alignment > kPointerSize) {
+    mov(scratch, Operand(0));
+    tst(sp, Operand(frame_alignment_mask));
+    push(scratch, nz);

[SeRya, 2011/01/11 16:34:50]

ASSERT(frame_alignment == 2 * kPointerSize);

[Zaheer, 2011/01/12 13:20:03]

Done

+  }
+
+  mov(ip, Operand(ExternalReference(Top::k_c_entry_fp_address)));

[SeRya, 2011/01/11 16:34:50]

ia32 code allocates C arguments below c_entry_fp_address. It allows to not care
that C arguments could be reached by GC. I think this semantic should be
preserved here as well.

By the way, you put argc into the argumets stack space what could be
misinterpreted as an object reference (if argc is odd) and crash GC if it
happens in the called function.

Also stack alignment placeholder don't need to be initialized if it's below
c_entry_fp_address.

[Zaheer, 2011/01/12 13:20:03]

i do allocate the args on top of the exit frame similar to ia32. Do i miss
something
[c args]
[Exit Frame] <- c_entry_fp
[fast api args]
[js args]
Maybe the ordering of code is confusing? i have split it in to
EnterApiExitFramePrologue/Epilogue
I did verify in Filecycler that there is no issue with gc invoked from called
function.
removed scratch argument. pushed ip (assumed dummy)

+  str(fp, MemOperand(ip));
+  mov(ip, Operand(ExternalReference(Top::k_context_address)));
+  str(cp, MemOperand(ip));
+}
+
+
+MaybeObject* MacroAssembler::TryCallApiFunctionAndReturn(
+    ApiFunction* function) {
+  ExternalReference next_address =
+      ExternalReference::handle_scope_next_address();
+  ExternalReference limit_address =
+      ExternalReference::handle_scope_limit_address();
+  ExternalReference level_address =
+      ExternalReference::handle_scope_level_address();
+
+  // Allocate HandleScope in callee-save registers.
+  mov(ip, Operand(next_address));

[SeRya, 2011/01/11 16:34:50]

x64 implementation use offsets to eliminate 2 of 3 ldr instructions:

static int Offset(ExternalReference ref0, ExternalReference ref1) {
  int64_t offset = (ref0.address() - ref1.address());
  // Check that fits into int.
  ASSERT(static_cast<int>(offset) == offset);
  return static_cast<int>(offset);
}

  ExternalReference next_address =
      ExternalReference::handle_scope_next_address();
  const int kNextOffset = 0;
  const int kLimitOffset = Offset(
      ExternalReference::handle_scope_limit_address(),
      next_address);
  const int kLevelOffset = Offset(
      ExternalReference::handle_scope_level_address(),
      next_address);

May be it would work here as well?

[Zaheer, 2011/01/12 13:20:03]

Done! but i could remove only the mov not ldr instructions. Also i have fixed r7
for next address removing two mov's.  

+  ldr(r4, MemOperand(ip));
+  mov(ip, Operand(limit_address));
+  ldr(r5, MemOperand(ip));
+  mov(ip, Operand(level_address));
+  ldr(r6, MemOperand(ip));
+  add(r6, r6, Operand(1));
+  str(r6, MemOperand(ip));
+
+  // patch the exit frame pc and Call the api function!

[antonm, 2011/01/11 14:11:33]

nit: [P]atch and [c]all.

[antonm, 2011/01/11 14:11:33]

once again, can we factor out frame handling?

[Zaheer, 2011/01/11 15:44:35]

Done

[Zaheer, 2011/01/11 15:44:35]

same issues as mentioned before for refactoring

+  add(ip, pc, Operand(8));
+  str(ip, MemOperand(fp, -2 * kPointerSize));

[antonm, 2011/01/11 14:11:33]

do not we have a named constant for -2?

[Zaheer, 2011/01/11 15:44:35]

check ExitFrame::FillState(..) it seems to be hard coded (pc = sp - 1 = fp -2). 

This value is already used for kMarkerOffset so i wasnt sure to introduce a new
constant.

+  Call(function->address(), RelocInfo::RUNTIME_ENTRY);
+
+  Label empty_handle;
+  Label prologue;
+  Label promote_scheduled_exception;
+  Label delete_allocated_handles;
+  Label leave_exit_frame;
+
+  // Check if the result handle holds 0.
+  cmp(r0, Operand(0));
+  b(eq, &empty_handle);
+  // It was non-zero.  Dereference to get the result value.
+  ldr(r0, MemOperand(r0));

[SeRya, 2011/01/11 16:34:50]

LoadRoot(r0, Heap::kUndefinedValueRootIndex, eq);
ldr(r0, MemOperand(r0), ne);

How about it?

[Zaheer, 2011/01/12 13:20:03]

Done! Thanks for the pointer, it simplifies the code very much.

+  bind(&prologue);
+  // No more valid handles (the result handle was the last one). Restore
+  // previous handle scope.
+  mov(ip, Operand(next_address));
+  str(r4, MemOperand(ip));
+  mov(ip, Operand(level_address));
+  ldr(r6, MemOperand(ip));
+  sub(r6, r6, Operand(1));
+  str(r6, MemOperand(ip));
+  mov(ip, Operand(limit_address));
+  ldr(ip, MemOperand(ip));
+  cmp(r5, ip);
+  b(ne, &delete_allocated_handles);
+
+  // Check if the function scheduled an exception.
+  bind(&leave_exit_frame);
+  mov(ip, Operand(ExternalReference::the_hole_value_location()));
+  ldr(r4, MemOperand(ip));

[antonm, 2011/01/11 14:11:33]

Shouldn't it be LoadRoot(r4, Heap::kTheHoleValueRootIndex)?

[Zaheer, 2011/01/11 15:44:35]

Done

+  mov(ip, Operand(ExternalReference(Top::k_pending_exception_address)));
+  ldr(r5, MemOperand(ip));
+  cmp(r4, r5);
+  b(ne, &promote_scheduled_exception);
+  LeaveExitFrame(0);

[antonm, 2011/01/11 14:11:33]

LeaveExitFrame apparently takes bool, not int

[Zaheer, 2011/01/11 15:44:35]

done

+
+  bind(&promote_scheduled_exception);
+  MaybeObject* result = TryTailCallExternalReference(
+      ExternalReference(Runtime::kPromoteScheduledException), 0, 1);
+  if (result->IsFailure()) {
+    return result;
+  }
+
+  bind(&empty_handle);
+  // It was zero; the result is undefined.
+  LoadRoot(r0, Heap::kUndefinedValueRootIndex);
+  b(&prologue);
+
+  // HandleScope limit has changed. Delete allocated extensions.
+  bind(&delete_allocated_handles);
+  mov(ip, Operand(limit_address));
+  str(r5, MemOperand(ip));
+  mov(r4, r0);
+  PrepareCallCFunction(0, r5);
+  CallCFunction(ExternalReference::delete_handle_scope_extensions(), 0);
+  mov(r0, r4);
+  jmp(&leave_exit_frame);
+
+  return result;
+}
+
+
 void MacroAssembler::IllegalOperation(int num_arguments) {
   if (num_arguments > 0) {
     add(sp, sp, Operand(num_arguments * kPointerSize));
@@ -1649,6 +1776,15 @@
   JumpToExternalReference(ext);
 }
 
+MaybeObject* MacroAssembler::TryTailCallExternalReference(
+    const ExternalReference& ext, int num_arguments, int result_size) {
+  // TODO(1236192): Most runtime routines don't need the number of
+  // arguments passed in because it is constant. At some point we
+  // should remove this need and make the runtime routine entry code
+  // smarter.
+  mov(r0, Operand(num_arguments));
+  return TryJumpToExternalReference(ext);
+}
 
 void MacroAssembler::TailCallRuntime(Runtime::FunctionId fid,
                                      int num_arguments,
@@ -1667,6 +1803,16 @@
   Jump(stub.GetCode(), RelocInfo::CODE_TARGET);
 }
 
+MaybeObject* MacroAssembler::TryJumpToExternalReference(
+    const ExternalReference& builtin) {
+#if defined(__thumb__)
+  // Thumb mode builtin.
+  ASSERT((reinterpret_cast<intptr_t>(builtin.address()) & 1) == 1);
+#endif
+  mov(r1, Operand(builtin));
+  CEntryStub stub(1);
+  return TryTailCallStub(&stub);
+}
 
 void MacroAssembler::InvokeBuiltin(Builtins::JavaScript id,
                                    InvokeJSFlags flags) {