Track live instance and allocation counts for classes

runtime/vm/assembler_ia32.cc

Index: runtime/vm/assembler_ia32.cc
diff --git a/runtime/vm/assembler_ia32.cc b/runtime/vm/assembler_ia32.cc
index 8f6164fcfef9614c1d1274517dbbf0993e02246e..71337266955d9c97c02a1b4ceb1bbe71fd3895dd 100644
--- a/runtime/vm/assembler_ia32.cc
+++ b/runtime/vm/assembler_ia32.cc
@@ -8,6 +8,7 @@
 #include "vm/assembler.h"
 #include "vm/code_generator.h"
 #include "vm/heap.h"
+#include "vm/heap_class_stats.h"
 #include "vm/memory_region.h"
 #include "vm/runtime_entry.h"
 #include "vm/stack_frame.h"
@@ -2263,6 +2264,23 @@ void Assembler::Bind(Label* label) {
 }
 
 
+// Updates the allocation count for cid.
+void Assembler::BumpAllocationCount(intptr_t cid,
+                                    Register temp_reg) {
+  ASSERT(temp_reg != kNoRegister);
+  HeapClassStatistics* heap_class_stats =
+      Isolate::Current()->heap_class_stats();
+  // temp_reg gets address of class table pointer.
+  movl(temp_reg, Address::Absolute(heap_class_stats->ClassStatsTableAddress()));
+  // Offset into class table.
+  const intptr_t offset = cid * sizeof(HeapClassData);  // NOLINT
+  // temp_reg points at HeapClassData for cid + offset to allocation count.
+  leal(temp_reg, Address(temp_reg, offset));
+  // Increment allocation count.
+  addl(Address(temp_reg, HeapClassData::allocated_since_gc_offset()),
+       Immediate(1));
+}
+
 void Assembler::TryAllocate(const Class& cls,
                             Label* failure,
                             bool near_jump,
@@ -2279,6 +2297,9 @@ void Assembler::TryAllocate(const Class& cls,
     // Successfully allocated the object, now update top to point to
     // next object start and store the class in the class field of object.
     movl(Address::Absolute(heap->TopAddress()), instance_reg);
+    pushl(instance_reg);

[Florian Schneider, 2013/10/31 09:56:01]

I suspect that this will make bump allocation a _lot_ slower:
Before: 2 loads, 2 stores.
After:  4 loads, 4 stores.
How will this cost be amortized?

[Cutch, 2013/10/31 11:20:02]

I will run some tests tomorrow against patch set 2 (which uses temp registers).
In some places, like intrinsics, we have registers to use. 

+    BumpAllocationCount(cls.id(), instance_reg);
+    popl(instance_reg);
     ASSERT(instance_size >= kHeapObjectTag);
     subl(instance_reg, Immediate(instance_size - kHeapObjectTag));
     uword tags = 0;