ARM: Make use of d16-d31 when available.

src/arm/deoptimizer-arm.cc

Index: src/arm/deoptimizer-arm.cc
diff --git a/src/arm/deoptimizer-arm.cc b/src/arm/deoptimizer-arm.cc
index ee2a581a57dcb443e22f1395dab540afa6e3927b..32d4d3e3c0bfa88b0fd4194b878b22428bf666e7 100644
--- a/src/arm/deoptimizer-arm.cc
+++ b/src/arm/deoptimizer-arm.cc
@@ -878,7 +878,7 @@ void Deoptimizer::FillInputFrame(Address tos, JavaScriptFrame* frame) {
   }
   input_->SetRegister(sp.code(), reinterpret_cast<intptr_t>(frame->sp()));
   input_->SetRegister(fp.code(), reinterpret_cast<intptr_t>(frame->fp()));
-  for (int i = 0; i < DoubleRegister::kNumAllocatableRegisters; i++) {
+  for (int i = 0; i < DoubleRegister::NumAllocatableRegisters(); i++) {
     input_->SetDoubleRegister(i, 0.0);
   }
 
@@ -906,23 +906,33 @@ void Deoptimizer::EntryGenerator::Generate() {
   RegList restored_regs = kJSCallerSaved | kCalleeSaved | ip.bit();
 
   const int kDoubleRegsSize =
-      kDoubleSize * DwVfpRegister::kNumAllocatableRegisters;
+      kDoubleSize * DwVfpRegister::NumAllocatableRegisters();
 
   // Save all VFP registers before messing with them.
   DwVfpRegister first = DwVfpRegister::FromAllocationIndex(0);
-  DwVfpRegister last =
-      DwVfpRegister::FromAllocationIndex(
-          DwVfpRegister::kNumAllocatableRegisters - 1);
+  DwVfpRegister last = DwVfpRegister::FromAllocationIndex(
+      DwVfpRegister::NumAllocatableRegisters() - 1);
   ASSERT(last.code() > first.code());
-  ASSERT((last.code() - first.code()) ==
-      (DwVfpRegister::kNumAllocatableRegisters - 1));
 #ifdef DEBUG
-  for (int i = 0; i <= (DwVfpRegister::kNumAllocatableRegisters - 1); i++) {
+  for (int i = 0; i <= (DwVfpRegister::NumAllocatableRegisters() - 1); i++) {
     ASSERT((DwVfpRegister::FromAllocationIndex(i).code() <= last.code()) &&
            (DwVfpRegister::FromAllocationIndex(i).code() >= first.code()));
   }
 #endif
-  __ vstm(db_w, sp, first, last);
+  if (last.code() > DwVfpRegister::ScratchReg().code()) {
+    // This happens when we also have d16-d31 available.
+    // Store the registers < ZeroReg and > ScratchReg separately.
+    ASSERT(DwVfpRegister::ScratchReg().code() ==
+           DwVfpRegister::ZeroReg().code() + 1);
+    __ vstm(db_w, sp, DwVfpRegister::from_code(
+        DwVfpRegister::ScratchReg().code() + 1), last);
+    __ vstm(db_w, sp, first, DwVfpRegister::from_code(
+        DwVfpRegister::ZeroReg().code() - 1));
+  } else {
+    ASSERT(last.code() - first.code() ==
+           DwVfpRegister::NumAllocatableRegisters() - 1);
+    __ vstm(db_w, sp, first, last);
+  }
 
   // Push all 16 registers (needed to populate FrameDescription::registers_).
   // TODO(1588) Note that using pc with stm is deprecated, so we should perhaps
@@ -984,7 +994,7 @@ void Deoptimizer::EntryGenerator::Generate() {
   // Copy VFP registers to
   // double_registers_[DoubleRegister::kNumAllocatableRegisters]
   int double_regs_offset = FrameDescription::double_registers_offset();
-  for (int i = 0; i < DwVfpRegister::kNumAllocatableRegisters; ++i) {
+  for (int i = 0; i < DwVfpRegister::NumAllocatableRegisters(); ++i) {
     int dst_offset = i * kDoubleSize + double_regs_offset;
     int src_offset = i * kDoubleSize + kNumberOfRegisters * kPointerSize;
     __ vldr(d0, sp, src_offset);
@@ -1027,6 +1037,7 @@ void Deoptimizer::EntryGenerator::Generate() {
         ExternalReference::compute_output_frames_function(isolate), 1);
   }
   __ pop(r0);  // Restore deoptimizer object (class Deoptimizer).
+  __ mov(r4, r0);

[Rodolph Perfetta, 2012/12/12 14:29:15]

minor tweak: the code below could be rewritten not to corrupt r0, removing the
need for this mov.

[hans, 2012/12/12 17:58:47]

Is it OK if I note this in a TODO instead, so as to keep the diff small? This
will be easy to fix in a separate patch.

 
   // Replace the current (input) frame with the output frames.
   Label outer_push_loop, inner_push_loop;
@@ -1050,6 +1061,16 @@ void Deoptimizer::EntryGenerator::Generate() {
   __ cmp(r0, r1);
   __ b(lt, &outer_push_loop);
 
+  // In case of OSR, we have to restore the d registers.
+  if (type() == OSR) {
+    __ ldr(r1, MemOperand(r4, Deoptimizer::input_offset()));
+    for (int i = 0; i < DwVfpRegister::NumAllocatableRegisters(); ++i) {
+      DwVfpRegister reg = DwVfpRegister::FromAllocationIndex(i);
+      int src_offset = double_regs_offset + i * kDoubleSize;
+      __ vldr(reg, r1, src_offset);
+    }
+  }
+
   // Push state, pc, and continuation from the last output frame.
   if (type() != OSR) {
     __ ldr(r6, MemOperand(r2, FrameDescription::state_offset()));