Add support for XMM registers in SSA code generation pipeline.

runtime/vm/flow_graph_compiler_x64.cc

 // Copyright (c) 2012, 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"  // Needed here to get TARGET_ARCH_X64.
 #if defined(TARGET_ARCH_X64)
 
 #include "vm/flow_graph_compiler.h"
 
 #include "lib/error.h"
@@ skipping 867 matching lines @@
   __ testq(RAX, Immediate(kSmiTagMask));
   __ j(ZERO, &smi_to_double);
   __ CompareClassId(RAX, kDoubleCid);
   __ j(NOT_EQUAL, &call_method);
   __ movsd(XMM1, FieldAddress(RAX, Double::value_offset()));
   __ Bind(&double_op);
   __ sqrtsd(XMM0, XMM1);
   AssemblerMacros::TryAllocate(assembler_,
                                double_class_,
                                &call_method,
+                               Assembler::kNearJump,
                                RAX);  // Result register.
   __ movsd(FieldAddress(RAX, Double::value_offset()), XMM0);
   __ Drop(1);
   __ jmp(done);
   __ Bind(&smi_to_double);
   __ SmiUntag(RAX);
   __ cvtsi2sd(XMM1, RAX);
   __ jmp(&double_op);
   __ Bind(&call_method);
 }
@@ skipping 209 matching lines @@
   __ movsd(result, FieldAddress(reg, Double::value_offset()));
   __ jmp(&done);
   __ Bind(&is_smi);
   __ movq(temp, reg);
   __ SmiUntag(temp);
   __ cvtsi2sd(result, temp);
   __ Bind(&done);
 }
 
 
+void FlowGraphCompiler::SaveLiveRegisters(LocationSummary* locs) {
+  // TODO(vegorov): consider saving only caller save (volatile) registers.
+  const intptr_t xmm_regs_count = locs->live_registers()->xmm_regs_count();
+  if (xmm_regs_count > 0) {
+    // Pointer maps don't support pushed untagged values so we reserve spill
+    // slots at the top of the spill slot area for live XMM registers.
+    intptr_t stack_offs = (StackSize() + 1) * kWordSize;
+    for (intptr_t reg_idx = 0; reg_idx < kNumberOfXmmRegisters; ++reg_idx) {
+      XmmRegister xmm_reg = static_cast<XmmRegister>(reg_idx);
+      if (locs->live_registers()->ContainsXmmRegister(xmm_reg)) {
+        __ movsd(Address(RBP, -stack_offs), xmm_reg);
+        stack_offs += kDoubleSize;
+      }
+    }
+  }
+
+  for (intptr_t reg_idx = 0; reg_idx < kNumberOfCpuRegisters; ++reg_idx) {
+    Register reg = static_cast<Register>(reg_idx);
+    if (locs->live_registers()->ContainsRegister(reg)) {
+      __ pushq(reg);
+    }
+  }
+}
+
+
+void FlowGraphCompiler::RestoreLiveRegisters(LocationSummary* locs) {
+  for (intptr_t reg_idx = kNumberOfCpuRegisters - 1; reg_idx >= 0; --reg_idx) {
+    Register reg = static_cast<Register>(reg_idx);
+    if (locs->live_registers()->ContainsRegister(reg)) {
+      __ popq(reg);
+    }
+  }
+
+  const intptr_t xmm_regs_count = locs->live_registers()->xmm_regs_count();
+  if (xmm_regs_count > 0) {
+    // Pointer maps don't support pushed untagged values so we reserve spill
+    // slots at the top of the spill slot area for live XMM registers.
+    intptr_t stack_offs = (StackSize() + 1) * kWordSize;
+    for (intptr_t reg_idx = 0; reg_idx < kNumberOfXmmRegisters; ++reg_idx) {
+      XmmRegister xmm_reg = static_cast<XmmRegister>(reg_idx);
+      if (locs->live_registers()->ContainsXmmRegister(xmm_reg)) {
+        __ movsd(xmm_reg, Address(RBP, -stack_offs));
+        stack_offs += kDoubleSize;
+      }
+    }
+  }
+}
+
+
 #undef __
 #define __ compiler_->assembler()->
 
 
 static Address ToStackSlotAddress(Location loc) {
-  ASSERT(loc.IsStackSlot());
   const intptr_t index = loc.stack_index();
   if (index < 0) {
     const intptr_t offset = (1 - index)  * kWordSize;
     return Address(RBP, offset);
   } else {
     const intptr_t offset =
         (ParsedFunction::kFirstLocalSlotIndex - index) * kWordSize;
     return Address(RBP, offset);
   }
 }
@@ skipping 12 matching lines @@
       __ movq(ToStackSlotAddress(destination), source.reg());
     }
   } else if (source.IsStackSlot()) {
     if (destination.IsRegister()) {
       __ movq(destination.reg(), ToStackSlotAddress(source));
     } else {
       ASSERT(destination.IsStackSlot());
       MoveMemoryToMemory(ToStackSlotAddress(destination),
                          ToStackSlotAddress(source));
     }
+  } else if (source.IsXmmRegister()) {
+    if (destination.IsXmmRegister()) {
+      // Optimization manual recommends using MOVAPS for register
+      // to register moves.
+      __ movaps(destination.xmm_reg(), source.xmm_reg());
+    } else {
+      ASSERT(destination.IsDoubleStackSlot());
+      __ movsd(ToStackSlotAddress(destination), source.xmm_reg());
+    }
+  } else if (source.IsDoubleStackSlot()) {
+    if (destination.IsXmmRegister()) {
+      __ movsd(destination.xmm_reg(), ToStackSlotAddress(source));
+    } else {
+      ASSERT(destination.IsDoubleStackSlot());
+      __ movsd(XMM0, ToStackSlotAddress(source));
+      __ movsd(ToStackSlotAddress(destination), XMM0);
+    }
   } else {
     ASSERT(source.IsConstant());
     if (destination.IsRegister()) {
       __ LoadObject(destination.reg(), source.constant());
     } else {
       ASSERT(destination.IsStackSlot());
       StoreObject(ToStackSlotAddress(destination), source.constant());
     }
   }
 
   move->Eliminate();
 }
 
 
 void ParallelMoveResolver::EmitSwap(int index) {
   MoveOperands* move = moves_[index];
   const Location source = move->src();
   const Location destination = move->dest();
 
   if (source.IsRegister() && destination.IsRegister()) {
     __ xchgq(destination.reg(), source.reg());
   } else if (source.IsRegister() && destination.IsStackSlot()) {
     Exchange(source.reg(), ToStackSlotAddress(destination));
   } else if (source.IsStackSlot() && destination.IsRegister()) {
     Exchange(destination.reg(), ToStackSlotAddress(source));
   } else if (source.IsStackSlot() && destination.IsStackSlot()) {
     Exchange(ToStackSlotAddress(destination), ToStackSlotAddress(source));
+  } else if (source.IsXmmRegister() && destination.IsXmmRegister()) {
+    __ movaps(XMM0, source.xmm_reg());
+    __ movaps(source.xmm_reg(), destination.xmm_reg());
+    __ movaps(destination.xmm_reg(), XMM0);
+  } else if (source.IsXmmRegister() || destination.IsXmmRegister()) {
+    ASSERT(destination.IsDoubleStackSlot() || source.IsDoubleStackSlot());
+    XmmRegister reg = source.IsXmmRegister() ? source.xmm_reg()
+                                             : destination.xmm_reg();
+    Address slot_address =
+        ToStackSlotAddress(source.IsXmmRegister() ? destination : source);
+
+    __ movsd(XMM0, slot_address);
+    __ movsd(slot_address, reg);
+    __ movaps(reg, XMM0);
   } else {
     UNREACHABLE();
   }
 
   // The swap of source and destination has executed a move from source to
   // destination.
   move->Eliminate();
 
   // Any unperformed (including pending) move with a source of either
   // this move's source or destination needs to have their source
@@ skipping 28 matching lines @@
 void ParallelMoveResolver::Exchange(const Address& mem1, const Address& mem2) {
   __ Exchange(mem1, mem2);
 }
 
 
 #undef __
 
 }  // namespace dart
 
 #endif  // defined TARGET_ARCH_X64