[arm][stubs] Fix d16-d31 preservation on stub failure

src/arm64/deoptimizer-arm64.cc

 // Copyright 2013 the V8 project authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "src/arm64/frames-arm64.h"
 #include "src/codegen.h"
 #include "src/deoptimizer.h"
 #include "src/full-codegen/full-codegen.h"
 #include "src/register-configuration.h"
 #include "src/safepoint-table.h"
@@ skipping 149 matching lines @@
   CPURegList copy_to_input = saved_registers;
   for (int i = 0; i < saved_registers.Count(); i++) {
     __ Peek(x2, i * kPointerSize);
     CPURegister current_reg = copy_to_input.PopLowestIndex();
     int offset = (current_reg.code() * kPointerSize) +
         FrameDescription::registers_offset();
     __ Str(x2, MemOperand(x1, offset));
   }
 
   // Copy FP registers to the input frame.
+  CPURegList copy_fp_to_input = saved_fp_registers;
   for (int i = 0; i < saved_fp_registers.Count(); i++) {
-    int dst_offset = FrameDescription::double_registers_offset() +
-        (i * kDoubleSize);
     int src_offset = kFPRegistersOffset + (i * kDoubleSize);
     __ Peek(x2, src_offset);
+    CPURegister reg = copy_fp_to_input.PopLowestIndex();
+    int dst_offset = FrameDescription::double_registers_offset() +
+                     (reg.code() * kDoubleSize);

[Jakob Kummerow, 2016/03/09 11:07:25]

The change here is s/i/reg.code()/. We must not use compact storage; non-saved
registers must leave gaps in the FrameDescription in order for
Deoptimizer::CopyDoubleRegisters (lines 79ff above) to work correctly. This is
consistent with other platforms.

     __ Str(x2, MemOperand(x1, dst_offset));
   }
 
   // Remove the bailout id and the saved registers from the stack.
   __ Drop(1 + (kSavedRegistersAreaSize / kXRegSize));
 
   // Compute a pointer to the unwinding limit in register x2; that is
   // the first stack slot not part of the input frame.
   Register unwind_limit = x2;
   __ Ldr(unwind_limit, MemOperand(x1, FrameDescription::frame_size_offset()));
@@ skipping 48 matching lines @@
 
   __ Add(x0, x0, kPointerSize);
   __ Bind(&outer_loop_header);
   __ Cmp(x0, x1);
   __ B(lt, &outer_push_loop);
 
   __ Ldr(x1, MemOperand(x4, Deoptimizer::input_offset()));
   DCHECK(!saved_fp_registers.IncludesAliasOf(crankshaft_fp_scratch) &&
          !saved_fp_registers.IncludesAliasOf(fp_zero) &&
          !saved_fp_registers.IncludesAliasOf(fp_scratch));
-  int src_offset = FrameDescription::double_registers_offset();
   while (!saved_fp_registers.IsEmpty()) {
     const CPURegister reg = saved_fp_registers.PopLowestIndex();
+    int src_offset = FrameDescription::double_registers_offset() +
+                     (reg.code() * kDoubleSize);
     __ Ldr(reg, MemOperand(x1, src_offset));
-    src_offset += kDoubleSize;
   }
 
   // Push state from the last output frame.
   __ Ldr(x6, MemOperand(current_frame, FrameDescription::state_offset()));
   __ Push(x6);
 
   // TODO(all): ARM copies a lot (if not all) of the last output frame onto the
   // stack, then pops it all into registers. Here, we try to load it directly
   // into the relevant registers. Is this correct? If so, we should improve the
   // ARM code.
@@ skipping 77 matching lines @@
 void FrameDescription::SetCallerConstantPool(unsigned offset, intptr_t value) {
   // No embedded constant pool support.
   UNREACHABLE();
 }
 
 
 #undef __
 
 }  // namespace internal
 }  // namespace v8