Generate some intrinsics using our IR.

runtime/vm/flow_graph_compiler.cc

Index: runtime/vm/flow_graph_compiler.cc
===================================================================
--- runtime/vm/flow_graph_compiler.cc	(revision 39726)
+++ runtime/vm/flow_graph_compiler.cc	(working copy)
@@ -98,6 +98,7 @@
           GrowableObjectArray::New())),
       is_optimizing_(is_optimizing),
       may_reoptimize_(false),
+      intrinsic_mode_(false),
       double_class_(Class::ZoneHandle(
           isolate_->object_store()->double_class())),
       mint_class_(Class::ZoneHandle(
@@ -547,6 +548,7 @@
 void FlowGraphCompiler::AddDeoptIndexAtCall(intptr_t deopt_id,
                                             intptr_t token_pos) {
   ASSERT(is_optimizing());
+  ASSERT(!intrinsic_mode());
   CompilerDeoptInfo* info =
       new CompilerDeoptInfo(deopt_id,
                             ICData::kDeoptAtCall,
@@ -727,6 +729,8 @@
 
 Label* FlowGraphCompiler::AddDeoptStub(intptr_t deopt_id,
                                        ICData::DeoptReasonId reason) {
+  if (intrinsic_mode()) return &intrinsic_deopt_label_;

[srdjan, 2014/09/02 18:48:21]

This feels fragile as we must guaranteed that the lifetime of FlowGraphCompiler
exceeds the lifetime of the new 'owner' of the  intrinsic_deopt_label_.

[Florian Schneider, 2014/09/02 21:03:20]

Can you elaborate? There is only one instance of FlowGraphCompiler. It is either
in intrinsic mode, or in normal mode. Intrinsic mode is active before generating
the normal body (slow-path) of the function. When it is active, all
deopt-branches go to this label instead of a real deopt-stub.

[srdjan, 2014/09/08 17:34:06]

I was mainly worried that the lifetime of a FlowGraphCompiler instance may be
shorter than the use of the label. Looking at the code again, my worry is
probably unnecessary

+
   ASSERT(is_optimizing_);
   CompilerDeoptInfoWithStub* stub =
       new CompilerDeoptInfoWithStub(deopt_id,
@@ -855,9 +859,15 @@
       }
     }
   }
-  // Even if an intrinsified version of the function was successfully
-  // generated, it may fall through to the non-intrinsified method body.
-  Intrinsifier::Intrinsify(parsed_function().function(), assembler());
+
+  EnterIntrinsicMode();
+
+  Intrinsifier::Intrinsify(parsed_function(), this);
+
+  ExitIntrinsicMode();
+  // "Deoptimization" from intrinsic continues here.

[srdjan, 2014/09/02 18:48:21]

We do not deoptimize, this is actually the slow-path, isn't it?

[Florian Schneider, 2014/09/02 21:03:20]

That's why it's in quotes. Intrinsic code uses the same IL instructions as the
optimizing compiler. Check instructions that deoptimize in regular optimized
code just branch to here.

+  ASSERT(!intrinsic_deopt_label_.IsBound());
+  assembler()->Bind(&intrinsic_deopt_label_);
 }
 
 
@@ -1290,13 +1300,17 @@
 
 intptr_t ParallelMoveResolver::AllocateScratchRegister(
     Location::Kind kind,
-    intptr_t blocked,
+    uword blocked_mask,
     intptr_t first_free_register,
     intptr_t last_free_register,
     bool* spilled) {
+  COMPILE_ASSERT(static_cast<intptr_t>(sizeof(blocked_mask)) * kBitsPerByte >=
+                 kNumberOfFpuRegisters);
+  COMPILE_ASSERT(static_cast<intptr_t>(sizeof(blocked_mask)) * kBitsPerByte >=
+                 kNumberOfCpuRegisters);
   intptr_t scratch = -1;
   for (intptr_t reg = first_free_register; reg <= last_free_register; reg++) {
-    if ((blocked != reg) &&
+    if ((((1u << reg) & blocked_mask) == 0) &&
         IsScratchLocation(Location::MachineRegisterLocation(kind, reg))) {
       scratch = reg;
       break;
@@ -1306,7 +1320,7 @@
   if (scratch == -1) {
     *spilled = true;
     for (intptr_t reg = first_free_register; reg <= last_free_register; reg++) {
-      if (blocked != reg) {
+      if (((1u << reg) & blocked_mask) == 0) {
         scratch = reg;
         break;
       }
@@ -1326,7 +1340,7 @@
       spilled_(false) {
   reg_ = static_cast<FpuRegister>(
       resolver_->AllocateScratchRegister(Location::kFpuRegister,
-                                         blocked,
+                                         1u << blocked,
                                          0,
                                          kNumberOfFpuRegisters - 1,
                                          &spilled_));
@@ -1349,9 +1363,15 @@
     : resolver_(resolver),
       reg_(kNoRegister),
       spilled_(false) {
+  uword blocked_mask = 1u << blocked;
+  if (resolver->compiler_->intrinsic_mode()) {
+    // Block additional registers that must be preserved for intrinsics.
+    blocked_mask |= 1u << ICREG;
+    blocked_mask |= 1u << ARGS_DESC_REG;
+  }
   reg_ = static_cast<Register>(
       resolver_->AllocateScratchRegister(Location::kRegister,
-                                         blocked,
+                                         1u << blocked,
                                          kFirstFreeCpuRegister,
                                          kLastFreeCpuRegister,
                                          &spilled_));