Add dynamic stack frame alignment to optimized functions with untagged doubles on the stack.

src/ia32/deoptimizer-ia32.cc

Index: src/ia32/deoptimizer-ia32.cc
diff --git a/src/ia32/deoptimizer-ia32.cc b/src/ia32/deoptimizer-ia32.cc
index 313b65f6e2168fe4fd0371083d103a90f493b83b..ddd593299621a42dbed4608d17cdf4b71e81eb9e 100644
--- a/src/ia32/deoptimizer-ia32.cc
+++ b/src/ia32/deoptimizer-ia32.cc
@@ -427,7 +427,14 @@ void Deoptimizer::DoComputeOsrOutputFrame() {
     output_[0]->SetPc(reinterpret_cast<uint32_t>(from_));
   } else {
     // Setup the frame pointer and the context pointer.
-    output_[0]->SetRegister(ebp.code(), input_->GetRegister(ebp.code()));
+    // All OSR stack frames are dynamically aligned to an 8-byte boundary.
+    int frame_pointer = input_->GetRegister(ebp.code());
+    if ((frame_pointer & 0x4) == 0) {
+      // Return address at FP + 4 should be aligned, so FP mod 8 should be 4.
+      frame_pointer -= kPointerSize;
+      has_alignment_padding_ = 1;
+    }
+    output_[0]->SetRegister(ebp.code(), frame_pointer);
     output_[0]->SetRegister(esi.code(), input_->GetRegister(esi.code()));
 
     unsigned pc_offset = data->OsrPcOffset()->value();
@@ -492,9 +499,11 @@ void Deoptimizer::DoComputeFrame(TranslationIterator* iterator,
   // top address and the current frame's size.
   uint32_t top_address;
   if (is_bottommost) {
-    // 2 = context and function in the frame.
-    top_address =
-        input_->GetRegister(ebp.code()) - (2 * kPointerSize) - height_in_bytes;
+    // If the optimized frame had alignment padding, adjust the frame pointer
+    // to point to the new position of the old frame pointer after padding
+    // is removed. Subtract 2 * kPointerSize for the context and function slots.
+    top_address = input_->GetRegister(ebp.code()) - (2 * kPointerSize) -
+        height_in_bytes + has_alignment_padding_ * kPointerSize;
   } else {
     top_address = output_[frame_index - 1]->GetTop() - output_frame_size;
   }
@@ -545,7 +554,9 @@ void Deoptimizer::DoComputeFrame(TranslationIterator* iterator,
   }
   output_frame->SetFrameSlot(output_offset, value);
   intptr_t fp_value = top_address + output_offset;
-  ASSERT(!is_bottommost || input_->GetRegister(ebp.code()) == fp_value);
+  ASSERT(!is_bottommost ||
+      input_->GetRegister(ebp.code()) + has_alignment_padding_ * kPointerSize
+      == fp_value);
   output_frame->SetFp(fp_value);
   if (is_topmost) output_frame->SetRegister(ebp.code(), fp_value);
   if (FLAG_trace_deopt) {
@@ -734,6 +745,17 @@ void Deoptimizer::EntryGenerator::Generate() {
   __ cmp(ecx, Operand(esp));
   __ j(not_equal, &pop_loop);
 
+  // If frame was dynamically aligned, pop padding.
+  Label sentinel, sentinel_done;
+  __ pop(Operand(ecx));
+  __ cmp(ecx, Operand(eax, Deoptimizer::frame_alignment_marker_offset()));
+  __ j(equal, &sentinel);
+  __ push(Operand(ecx));
+  __ jmp(&sentinel_done);
+  __ bind(&sentinel);
+  __ mov(Operand(eax, Deoptimizer::has_alignment_padding_offset()),
+         Immediate(1));
+  __ bind(&sentinel_done);
   // Compute the output frame in the deoptimizer.
   __ push(eax);
   __ PrepareCallCFunction(1, ebx);
@@ -745,6 +767,17 @@ void Deoptimizer::EntryGenerator::Generate() {
   }
   __ pop(eax);
 
+  if (type() == OSR) {
+    // If alignment padding is added, push the sentinel.
+    Label no_osr_padding;
+    __ cmp(Operand(eax, Deoptimizer::has_alignment_padding_offset()),
+           Immediate(0));
+    __ j(equal, &no_osr_padding, Label::kNear);
+    __ push(Operand(eax, Deoptimizer::frame_alignment_marker_offset()));
+    __ bind(&no_osr_padding);
+  }
+
+
   // Replace the current frame with the output frames.
   Label outer_push_loop, inner_push_loop;
   // Outer loop state: eax = current FrameDescription**, edx = one past the