Finish implementing lazy deoptimization (ia32, x64). Ran tests with --deoptimize-alot.

runtime/vm/stub_code_ia32.cc

 // Copyright (c) 2011, 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_IA32)
 
 #include "vm/assembler.h"
 #include "vm/assembler_macros.h"
 #include "vm/compiler.h"
 #include "vm/flow_graph_compiler.h"
+#include "vm/instructions.h"
 #include "vm/object_store.h"
 #include "vm/pages.h"
 #include "vm/resolver.h"
 #include "vm/scavenger.h"
 #include "vm/stub_code.h"
 
 
 #define __ assembler->
 
 namespace dart {
@@ skipping 452 matching lines @@
   __ ret();
 }
 
 
 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 EAX 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:
 //   +------------------+
-//   | 0 as PC marker   | <- TOS
-//   +------------------+
-//   | Saved FP         |
+//   | Saved FP         | <- TOS
 //   +------------------+
 //   | return-address   |  (deoptimization point)
 //   +------------------+
 //   | optimized frame  |
 //   |  ...             |
 //
-void StubCode::GenerateDeoptimizeStub(Assembler* assembler) {
+// Parts of the code cannot GC, part of the code can GC.
+static void GenerateDeoptimizationSequence(Assembler* assembler,
+                                           bool preserve_eax) {
   __ EnterFrame(0);
+  // The code in this frame may not cause GC. kDeoptimizeCopyFrameRuntimeEntry
+  // and kDeoptimizeFillFrameRuntimeEntry are leaf runtime calls.
+  const intptr_t saved_eax_offset_from_ebp = -(kNumberOfCpuRegisters - EAX);
+  // Result in EAX is preserved as part of pushing all registers below.
+
   // Push registers in their enumeration order: lowest register number at
   // lowest address.
   for (intptr_t i = kNumberOfCpuRegisters - 1; i >= 0; i--) {
     __ pushl(static_cast<Register>(i));
   }
   __ subl(ESP, Immediate(kNumberOfXmmRegisters * kDoubleSize));
   intptr_t offset = 0;
   for (intptr_t reg_idx = 0; reg_idx < kNumberOfXmmRegisters; ++reg_idx) {
     XmmRegister xmm_reg = static_cast<XmmRegister>(reg_idx);
     __ movsd(Address(ESP, offset), xmm_reg);
     offset += kDoubleSize;
   }
 
   __ movl(ECX, ESP);  // Saved saved registers block.
   __ ReserveAlignedFrameSpace(1 * kWordSize);
   __ SmiUntag(EAX);
   __ movl(Address(ESP, 0), ECX);  // Start of register block.
   __ CallRuntime(kDeoptimizeCopyFrameRuntimeEntry);
   // Result (EAX) is stack-size (FP - SP) in bytes, incl. the return address.
+
+  if (preserve_eax) {
+    // Restore result into EBX temporarily.
+    __ movl(EBX, Address(EBP, saved_eax_offset_from_ebp * kWordSize));
+  }
+
   __ LeaveFrame();
   __ popl(EDX);  // Preserve return address.
   __ movl(ESP, EBP);
   __ subl(ESP, EAX);
   __ movl(Address(ESP, 0), EDX);
 
   __ EnterFrame(0);
   __ movl(ECX, ESP);  // Get last FP address.
+  if (preserve_eax) {
+    __ pushl(EBX);  // Preserve result.
+  }
   __ ReserveAlignedFrameSpace(1 * kWordSize);
   __ movl(Address(ESP, 0), ECX);
   __ CallRuntime(kDeoptimizeFillFrameRuntimeEntry);
+
+  if (preserve_eax) {
+    // Restore result into EAX.
+    __ movl(EAX, Address(EBP, -1 * kWordSize));
+  }
+  // Code above cannot cause GC.
   __ LeaveFrame();
 
   // 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.
   __ EnterFrame(0);
+  if (preserve_eax) {
+    __ pushl(EAX);  // Preserve result, it will be GC-d here.
+  }
   __ CallRuntime(kDeoptimizeMaterializeDoublesRuntimeEntry);
+  if (preserve_eax) {
+    __ popl(EAX);  // Restore result.
+  }
   __ LeaveFrame();
   __ ret();
 }
 
+// TOS: return address + call-instruction-size (5 bytes).
+// EAX: result, must be preserved
+void StubCode::GenerateDeoptimizeLazyStub(Assembler* assembler) {
+  // Correct return address to point just after the call that is being
+  // deoptimized.
+  __ popl(EBX);
+  __ subl(EBX, Immediate(CallPattern::InstructionLength()));
+  __ pushl(EBX);
+  GenerateDeoptimizationSequence(assembler, true);  // Preserve EAX.
+}
+
+
+void StubCode::GenerateDeoptimizeStub(Assembler* assembler) {
+  GenerateDeoptimizationSequence(assembler, false);  // Don't preserve EAX.
+}
+
 
 // Called for inline allocation of arrays.
 // Input parameters:
 //   EDX : Array length as Smi.
 //   ECX : array element type (either NULL or an instantiated type).
 // Uses EAX, EBX, ECX, EDI  as temporary registers.
 // NOTE: EDX cannot be clobbered here as the caller relies on it being saved.
 // The newly allocated object is returned in EAX.
 void StubCode::GenerateAllocateArrayStub(Assembler* assembler) {
   Label slow_case;
@@ skipping 1346 matching lines @@
   __ movl(EAX, Address(ESP, 4 * kWordSize));  // Load error object.
   __ movl(EBP, Address(ESP, 3 * kWordSize));  // Load target frame_pointer.
   __ movl(EBX, Address(ESP, 1 * kWordSize));  // Load target PC into EBX.
   __ movl(ESP, Address(ESP, 2 * kWordSize));  // Load target stack_pointer.
   __ jmp(EBX);  // Jump to the exception handler code.
 }
 
 }  // namespace dart
 
 #endif  // defined TARGET_ARCH_IA32