Subzero: Basic Block Profiler.

src/IceCfg.cpp

Index: src/IceCfg.cpp
diff --git a/src/IceCfg.cpp b/src/IceCfg.cpp
index 50aa5ee1dbbd25974869e0b766092876a0830a08..73d59ac39bfeedb6e3e8ee75742dd82e78560f9b 100644
--- a/src/IceCfg.cpp
+++ b/src/IceCfg.cpp
@@ -18,6 +18,7 @@
 #include "IceClFlags.h"
 #include "IceDefs.h"
 #include "IceELFObjectWriter.h"
+#include "IceGlobalInits.h"
 #include "IceInst.h"
 #include "IceLiveness.h"
 #include "IceOperand.h"
@@ -75,6 +76,53 @@ void Cfg::addImplicitArg(Variable *Arg) {
 // 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;