Enables optimization on MIPS

runtime/vm/stub_code_mips.cc

 // Copyright (c) 2013, the Dart project authors.  Please see the AUTHORS file
 // for details. All rights reserved. Use of this source code is governed by a
 // BSD-style license that can be found in the LICENSE file.
 
 #include "vm/globals.h"
 #if defined(TARGET_ARCH_MIPS)
 
 #include "vm/assembler.h"
 #include "vm/code_generator.h"
 #include "vm/compiler.h"
@@ skipping 93 matching lines @@
   __ mov(CTX, A2);
 
   __ mov(SP, FP);
   __ lw(RA, Address(SP, 1 * kWordSize));
   __ lw(FP, Address(SP, 0 * kWordSize));
   __ Ret();
   __ delay_slot()->addiu(SP, SP, Immediate(2 * kWordSize));
 }
 
 
+// Print the stop message.
+DEFINE_LEAF_RUNTIME_ENTRY(void, PrintStopMessage, const char* message) {
+  OS::Print("Stop message: %s\n", message);
+}
+END_LEAF_RUNTIME_ENTRY
+
+
+// Input parameters:
+//   A0 : stop message (const char*).
+// Must preserve all registers.
 void StubCode::GeneratePrintStopMessageStub(Assembler* assembler) {
-  __ Unimplemented("PrintStopMessage stub");
+  __ EnterCallRuntimeFrame(0);
+  // Call the runtime leaf function. A0 already contains the parameter.
+  __ CallRuntime(kPrintStopMessageRuntimeEntry);
+  __ LeaveCallRuntimeFrame();
+  __ Ret();
 }
 
 
 // Input parameters:
 //   RA : return address.
 //   SP : address of return value.
 //   T5 : address of the native function to call.
 //   A2 : address of first argument in argument array.
 //   A1 : argc_tag including number of arguments and function kind.
 void StubCode::GenerateCallNativeCFunctionStub(Assembler* assembler) {
@@ skipping 106 matching lines @@
   __ LeaveStubFrameAndReturn(T0);
 }
 
 
 // Called from a static call only when an invalid code has been entered
 // (invalid because its function was optimized or deoptimized).
 // S4: arguments descriptor array.
 void StubCode::GenerateFixCallersTargetStub(Assembler* assembler) {
   // Create a stub frame as we are pushing some objects on the stack before
   // calling into the runtime.
+  __ TraceSimMsg("FixCallersTarget");
   __ EnterStubFrame();
   // Setup space on stack for return value and preserve arguments descriptor.
   __ addiu(SP, SP, Immediate(-2 * kWordSize));
   __ sw(S4, Address(SP, 1 * kWordSize));
   __ sw(NULLREG, Address(SP, 0 * kWordSize));
   __ CallRuntime(kFixCallersTargetRuntimeEntry);
   // Get Code object result and restore arguments descriptor array.
   __ lw(T0, Address(SP, 0 * kWordSize));
   __ lw(S4, Address(SP, 1 * kWordSize));
   __ addiu(SP, SP, Immediate(2 * kWordSize));
@@ skipping 65 matching lines @@
   __ addiu(SP, SP, Immediate(-4 * kWordSize));
   __ sw(NULLREG, Address(SP, 3 * kWordSize));
   __ sw(T1, Address(SP, 2 * kWordSize));
   __ sw(S5, Address(SP, 1 * kWordSize));
   __ sw(S4, Address(SP, 0 * kWordSize));
 
   // A1: Smi-tagged arguments array length.
   PushArgumentsArray(assembler);
   __ TraceSimMsg("InstanceFunctionLookupStub return");
 
-  // Stack:
-  // TOS + 0: argument array.
-  // TOS + 1: arguments descriptor array.
-  // TOS + 2: IC data object.
-  // TOS + 3: Receiver.
-  // TOS + 4: place for result from the call.
-  // TOS + 5: saved FP of previous frame.
-  // TOS + 6: dart code return address
-  // TOS + 7: pc marker (0 for stub).
-  // TOS + 8: last argument of caller.
-  // ....
   __ CallRuntime(kInstanceFunctionLookupRuntimeEntry);
 
   __ lw(V0, Address(SP, 4 * kWordSize));  // Get result into V0.
   __ addiu(SP, SP, Immediate(5 * kWordSize));    // Remove arguments.
 
   __ LeaveStubFrameAndReturn();
 }
 
 
 DECLARE_LEAF_RUNTIME_ENTRY(intptr_t, DeoptimizeCopyFrame,
                            intptr_t deopt_reason,
                            uword saved_registers_address);
 
 DECLARE_LEAF_RUNTIME_ENTRY(void, DeoptimizeFillFrame, uword last_fp);
 
 
+
+
 // Used by eager and lazy deoptimization. Preserve result in V0 if necessary.
 // This stub translates optimized frame into unoptimized frame. The optimized
 // frame can contain values in registers and on stack, the unoptimized
 // frame contains all values on stack.
 // Deoptimization occurs in following steps:
 // - Push all registers that can contain values.
 // - Call C routine to copy the stack and saved registers into temporary buffer.
 // - Adjust caller's frame to correct unoptimized frame size.
 // - Fill the unoptimized frame.
 // - Materialize objects that require allocation (e.g. Double instances).
 // GC can occur only after frame is fully rewritten.
 // Stack after EnterFrame(...) below:
 //   +------------------+
-//   | Saved FP         | <- TOS
+//   | Saved PP         | <- TOS
 //   +------------------+
-//   | return-address   |  (deoptimization point)
+//   | Saved FP         | <- FP of stub
 //   +------------------+
-//   | optimized frame  |
-//   |  ...             |
+//   | Saved LR         |  (deoptimization point)
+//   +------------------+
+//   | PC marker        |
+//   +------------------+
+//   | ...              | <- SP of optimized frame
 //
 // Parts of the code cannot GC, part of the code can GC.
 static void GenerateDeoptimizationSequence(Assembler* assembler,
                                            bool preserve_result) {
+  __ TraceSimMsg("GenerateDeoptimizationSequence");
   const intptr_t kPushedRegistersSize =
       kNumberOfCpuRegisters * kWordSize +
-      2 * kWordSize +  // FP and RA.
+      4 * kWordSize +  // PP, FP, RA, PC marker.
       kNumberOfFRegisters * kWordSize;
 
+  // DeoptimizeCopyFrame expects a Dart frame, i.e. EnterDartFrame(0), but there
+  // is no need to set the correct PC marker or load PP, since they get patched.
   __ addiu(SP, SP, Immediate(-kPushedRegistersSize * kWordSize));
-  __ sw(RA, Address(SP, kPushedRegistersSize - 1 * kWordSize));
-  __ sw(FP, Address(SP, kPushedRegistersSize - 2 * kWordSize));
-  __ addiu(FP, SP, Immediate(kPushedRegistersSize - 2 * kWordSize));
-
+  __ sw(ZR, Address(SP, kPushedRegistersSize - 1 * kWordSize));
+  __ sw(RA, Address(SP, kPushedRegistersSize - 2 * kWordSize));
+  __ sw(FP, Address(SP, kPushedRegistersSize - 3 * kWordSize));
+  __ sw(PP, Address(SP, kPushedRegistersSize - 4 * kWordSize));
+  __ addiu(FP, SP, Immediate(kPushedRegistersSize - 3 * kWordSize));
 
   // The code in this frame may not cause GC. kDeoptimizeCopyFrameRuntimeEntry
   // and kDeoptimizeFillFrameRuntimeEntry are leaf runtime calls.
-  const intptr_t saved_v0_offset_from_fp = -(kNumberOfCpuRegisters - V0);
+  const intptr_t saved_result_slot_from_fp =
+      kFirstLocalSlotFromFp + 1 - (kNumberOfCpuRegisters - V0);
   // Result in V0 is preserved as part of pushing all registers below.
 
+  // TODO(regis): Should we align the stack before pushing the fpu registers?
+  // If we do, saved_result_slot_from_fp is not constant anymore.
+
   // Push registers in their enumeration order: lowest register number at
   // lowest address.
   for (int i = 0; i < kNumberOfCpuRegisters; i++) {
-    const int slot = 2 + kNumberOfCpuRegisters - i;
+    const int slot = 4 + kNumberOfCpuRegisters - i;
     Register reg = static_cast<Register>(i);
     __ sw(reg, Address(SP, kPushedRegistersSize - slot * kWordSize));
   }
   for (int i = 0; i < kNumberOfFRegisters; i++) {
     // These go below the CPU registers.
-    const int slot = 2 + kNumberOfCpuRegisters + kNumberOfFRegisters - i;
+    const int slot = 4 + kNumberOfCpuRegisters + kNumberOfFRegisters - i;
     FRegister reg = static_cast<FRegister>(i);
     __ swc1(reg, Address(SP, kPushedRegistersSize - slot * kWordSize));
   }
 
   __ mov(A0, SP);  // Pass address of saved registers block.
   __ ReserveAlignedFrameSpace(0);
   __ CallRuntime(kDeoptimizeCopyFrameRuntimeEntry);
   // Result (V0) is stack-size (FP - SP) in bytes, incl. the return address.
 
   if (preserve_result) {
     // Restore result into T1 temporarily.
-    __ lw(T1, Address(FP, saved_v0_offset_from_fp * kWordSize));
+    __ lw(T1, Address(FP, saved_result_slot_from_fp * kWordSize));
   }
 
-  __ mov(SP, FP);
-  __ lw(FP, Address(SP, 0 * kWordSize));
-  __ lw(RA, Address(SP, 1 * kWordSize));
-  __ addiu(SP, SP, Immediate(2 * kWordSize));
-
+  __ addiu(SP, FP, Immediate(-kWordSize));
+  __ lw(RA, Address(SP, 2 * kWordSize));
+  __ lw(FP, Address(SP, 1 * kWordSize));
+  __ lw(PP, Address(SP, 0 * kWordSize));
   __ subu(SP, FP, V0);
 
-  __ addiu(SP, SP, Immediate(-2 * kWordSize));
-  __ sw(RA, Address(SP, 1 * kWordSize));
-  __ sw(FP, Address(SP, 0 * kWordSize));
-  __ mov(FP, SP);
+  // DeoptimizeFillFrame expects a Dart frame, i.e. EnterDartFrame(0), but there
+  // is no need to set the correct PC marker or load PP, since they get patched.
+  __ addiu(SP, SP, Immediate(-4 * kWordSize));
+  __ sw(ZR, Address(SP, 3 * kWordSize));
+  __ sw(RA, Address(SP, 2 * kWordSize));
+  __ sw(FP, Address(SP, 1 * kWordSize));
+  __ sw(PP, Address(SP, 0 * kWordSize));
+  __ addiu(FP, SP, Immediate(kWordSize));
 
-  __ mov(A0, SP);  // Get last FP address.
+  __ mov(A0, FP);  // Get last FP address.
   if (preserve_result) {
-    __ Push(T1);  // Preserve result.
+    __ Push(T1);  // Preserve result as first local.
   }
   __ ReserveAlignedFrameSpace(0);
   __ CallRuntime(kDeoptimizeFillFrameRuntimeEntry);  // Pass last FP in A0.
-  // Result (V0) is our FP.
   if (preserve_result) {
     // Restore result into T1.
-    __ lw(T1, Address(FP, -1 * kWordSize));
+    __ lw(T1, Address(FP, kFirstLocalSlotFromFp * kWordSize));
   }
   // Code above cannot cause GC.
-  __ mov(SP, FP);
-  __ lw(FP, Address(SP, 0 * kWordSize));
-  __ lw(RA, Address(SP, 1 * kWordSize));
-  __ addiu(SP, SP, Immediate(2 * kWordSize));
-  __ mov(FP, V0);
+  __ addiu(SP, FP, Immediate(-kWordSize));
+  __ lw(RA, Address(SP, 2 * kWordSize));
+  __ lw(FP, Address(SP, 1 * kWordSize));
+  __ lw(PP, Address(SP, 0 * kWordSize));
+  __ addiu(SP, SP, Immediate(4 * kWordSize));
 
   // Frame is fully rewritten at this point and it is safe to perform a GC.
   // Materialize any objects that were deferred by FillFrame because they
   // require allocation.
   __ EnterStubFrame();
   if (preserve_result) {
     __ Push(T1);  // Preserve result, it will be GC-d here.
   }
   __ PushObject(Smi::ZoneHandle());  // Space for the result.
   __ CallRuntime(kDeoptimizeMaterializeRuntimeEntry);
   // Result tells stub how many bytes to remove from the expression stack
   // of the bottom-most frame. They were used as materialization arguments.
   __ Pop(T1);
-  __ SmiUntag(T1);
   if (preserve_result) {
     __ Pop(V0);  // Restore result.
   }
   __ LeaveStubFrame();
-
-  // Return.
-  __ jr(RA);
-  __ delay_slot()->addu(SP, SP, T1);  // Remove materialization arguments.
+  // Remove materialization arguments.
+  __ SmiUntag(T1);
+  __ addu(SP, SP, T1);
+  __ Ret();
 }
 
 
 void StubCode::GenerateDeoptimizeLazyStub(Assembler* assembler) {
-  __ Unimplemented("DeoptimizeLazy stub");
+  // Correct return address to point just after the call that is being
+  // deoptimized.
+  __ AddImmediate(RA, -CallPattern::kFixedLengthInBytes);
+  GenerateDeoptimizationSequence(assembler, true);  // Preserve V0.
 }
 
 
 void StubCode::GenerateDeoptimizeStub(Assembler* assembler) {
   GenerateDeoptimizationSequence(assembler, false);  // Don't preserve V0.
 }
 
 
 void StubCode::GenerateMegamorphicMissStub(Assembler* assembler) {
   __ Unimplemented("MegamorphicMiss stub");
+  return;
 }
 
 
 // Called for inline allocation of arrays.
 // Input parameters:
 //   RA: return address.
 //   A1: Array length as Smi.
 //   A0: array element type (either NULL or an instantiated type).
 // NOTE: A1 cannot be clobbered here as the caller relies on it being saved.
 // The newly allocated object is returned in V0.
@@ skipping 366 matching lines @@
   __ LeaveStubFrameAndReturn();
 }
 
 
 // Called for inline allocation of contexts.
 // Input:
 //   T1: number of context variables.
 // Output:
 //   V0: new allocated RawContext object.
 void StubCode::GenerateAllocateContextStub(Assembler* assembler) {
+  __ TraceSimMsg("AllocateContext");
   if (FLAG_inline_alloc) {
     const Class& context_class = Class::ZoneHandle(Object::context_class());
     Label slow_case;
     Heap* heap = Isolate::Current()->heap();
     // First compute the rounded instance size.
     // T1: number of context variables.
     intptr_t fixed_size = sizeof(RawContext) + kObjectAlignment - 1;
     __ LoadImmediate(T2, fixed_size);
     __ sll(T0, T1, 2);
     __ addu(T2, T2, T0);
@@ skipping 871 matching lines @@
   __ LeaveDartFrameAndReturn();
 }
 
 
 //  RA: return address (Dart code).
 //  S5: Inline cache data array.
 //  S4: Arguments descriptor array.
 void StubCode::GenerateBreakpointDynamicStub(Assembler* assembler) {
   // Create a stub frame as we are pushing some objects on the stack before
   // calling into the runtime.
+  __ TraceSimMsg("BreakpointDynamicStub");
   __ EnterStubFrame();
   __ addiu(SP, SP, Immediate(-2 * kWordSize));
   __ sw(S5, Address(SP, 1 * kWordSize));
   __ sw(S4, Address(SP, 0 * kWordSize));
   __ CallRuntime(kBreakpointDynamicHandlerRuntimeEntry);
   __ lw(S4, Address(SP, 0 * kWordSize));
   __ lw(S5, Address(SP, 1 * kWordSize));
   __ addiu(SP, SP, Immediate(2 * kWordSize));
   __ LeaveStubFrame();
 
@@ skipping 360 matching lines @@
   __ CallRuntime(kBigintCompareRuntimeEntry);
   __ TraceSimMsg("IdenticalWithNumberCheckStub return");
   // Result in V0, 0 means equal.
   __ LeaveStubFrame();
   __ b(&done);
   __ delay_slot()->mov(CMPRES, V0);
 
   __ Bind(&reference_compare);
   __ subu(ret, left, right);
   __ Bind(&done);
+  // A branch or test after this comparison will check CMPRES == TMP1.
+  __ mov(TMP1, ZR);
   __ lw(T0, Address(SP, 0 * kWordSize));
   __ lw(T1, Address(SP, 1 * kWordSize));
   __ Ret();
   __ delay_slot()->addiu(SP, SP, Immediate(2 * kWordSize));
 }
 
 }  // namespace dart
 
 #endif  // defined TARGET_ARCH_MIPS