Subzero: Basic Block Profiler.

src/IceCfg.cpp

 //===- subzero/src/IceCfg.cpp - Control flow graph implementation ---------===//
 //
 //                        The Subzero Code Generator
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 // This file implements the Cfg class, including constant pool
 // management.
 //
 //===----------------------------------------------------------------------===//
 
 #include "assembler.h"
 #include "IceCfg.h"
 #include "IceCfgNode.h"
 #include "IceClFlags.h"
 #include "IceDefs.h"
 #include "IceELFObjectWriter.h"
+#include "IceGlobalInits.h"
 #include "IceInst.h"
 #include "IceLiveness.h"
 #include "IceOperand.h"
 #include "IceTargetLowering.h"
 
 namespace Ice {
 
 ICE_TLS_DEFINE_FIELD(const Cfg *, Cfg, CurrentCfg);
 
 ArenaAllocator<> *getCurrentCfgAllocator() {
@@ skipping 37 matching lines @@
 
 void Cfg::addImplicitArg(Variable *Arg) {
   Arg->setIsImplicitArg();
   ImplicitArgs.push_back(Arg);
 }
 
 // Returns whether the stack frame layout has been computed yet.  This
 // is used for dumping the stack frame location of Variables.
 bool Cfg::hasComputedFrame() const { return getTarget()->hasComputedFrame(); }
 
+namespace {
+constexpr char BlockNameGlobalPrefix[] = ".Lblock_name_";
+constexpr char BlockStatsGlobalPrefix[] = ".Lprofile_";

[Jim Stichnoth, 2015/06/08 23:45:31]

Document somewhere that this is a "reserved" prefix?

Also, I think there would be less chance of label name collisions if reserved
prefixes begin with ".L$".  So, ".L$block_name_" and ".L$profile_".

[John, 2015/06/09 15:36:18]

What do you mean, "reserved"?

[Jim Stichnoth, 2015/06/09 16:39:45]

For a function named "foo", Subzero generates internal label strings as
  ".L" + "foo" + "$" + nodename
  ".L" + "foo" + "$local$__" + label_number
The first is for basic block labels, and the second is for intra-block branch
labels.

My concern here whether some bitcode could be conjured up where one of these
generated labels is the same as a generated profiling label, causing
-filetype=asm output to break, and possibly -filetype=obj as well.

I think we can count on function names not containing '$' (though not
necessarily node names), so labels starting with ".L$" should have no chance of
conflicting.  In that case, I think ".L$block_name_" and ".L$profile_" should be
good prefixes.  I would avoid the __Sz part unless it is to be an externally
visible symbol such as in the Subzero runtime, but in that case the __Sz should
be at the beginning of the symbol name, rather than the middle.

With all that, never mind the documentation request, maybe just tweak the prefix
strings.

[John, 2015/06/09 22:01:16]

To avoid name clashes (as long as $ is not a valid identifier, that is) I
changed the prefixes to

.L$profiler$block_name$ and
.L$profiler$block_info$ and

+
+VariableDeclaration *nodeNameDeclaration(const IceString &NodeAsmName) {
+  VariableDeclaration *Var = VariableDeclaration::create();
+  Var->setName(BlockNameGlobalPrefix + NodeAsmName);
+  Var->setIsConstant(true);
+  Var->addInitializer(new VariableDeclaration::DataInitializer(
+      NodeAsmName.data(), NodeAsmName.size() + 1));
+  Var->setAlignment(8); // Wasteful, 32-bit could use 4 bytes.
+  return Var;
+}
+
+VariableDeclaration *
+blockProfilingInfoDeclaration(const IceString &NodeAsmName,
+                              VariableDeclaration *NodeNameDeclaration) {
+  VariableDeclaration *Var = VariableDeclaration::create();
+  Var->setName(BlockStatsGlobalPrefix + NodeAsmName);
+  Var->addInitializer(new VariableDeclaration::ZeroInitializer(8));
+  Var->addInitializer(
+      new VariableDeclaration::RelocInitializer(NodeNameDeclaration, 0));
+  Var->setAlignment(8);
+  return Var;
+}
+
+} // end of namespace

[Jim Stichnoth, 2015/06/08 23:45:31]

end of anonymous namespace

[John, 2015/06/09 15:36:18]

Done.

+
+std::unique_ptr<VariableDeclarationList> Cfg::profileBlocks() {

[Jim Stichnoth, 2015/06/08 23:45:31]

Add light documentation of these two functions, either here or in the .h file.

[John, 2015/06/09 15:36:18]

Done.

+  std::unique_ptr<VariableDeclarationList> ProfilerInits(
+      new VariableDeclarationList());
+
+  for (CfgNode *Node : Nodes) {
+    IceString NodeAsmName = Node->getAsmName();
+    ProfilerInits->push_back(nodeNameDeclaration(NodeAsmName));
+    ProfilerInits->push_back(
+        blockProfilingInfoDeclaration(NodeAsmName, ProfilerInits->back()));
+    Node->profileExecutionCount(ProfilerInits->back());
+  }
+
+  return ProfilerInits;
+}
+
+bool Cfg::isProfileGlobal(const VariableDeclaration &Var) {
+  return Var.getName().find(BlockStatsGlobalPrefix) == 0;
+}
+
 void Cfg::translate() {
   if (hasError())
     return;
   // FunctionTimer conditionally pushes/pops a TimerMarker if
   // TimeEachFunction is enabled.
   std::unique_ptr<TimerMarker> FunctionTimer;
   if (ALLOW_DUMP) {
     const IceString &TimingFocusOn =
         getContext()->getFlags().getTimingFocusOn();
     const IceString &Name = getFunctionName();
@@ skipping 453 matching lines @@
     }
   }
   // Print each basic block
   for (CfgNode *Node : Nodes)
     Node->dump(this);
   if (isVerbose(IceV_Instructions))
     Str << "}\n";
 }
 
 } // end of namespace Ice