Subzero: Add a few performance measurement tools.

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 "IceCfg.h"
 #include "IceCfgNode.h"
 #include "IceClFlags.h"
 #include "IceDefs.h"
 #include "IceInst.h"
 #include "IceLiveness.h"
 #include "IceOperand.h"
 #include "IceTargetLowering.h"
 
 namespace Ice {
 
 Cfg::Cfg(GlobalContext *Ctx)
     : Ctx(Ctx), FunctionName(""), ReturnType(IceType_void),
-      IsInternalLinkage(false), HasError(false), ErrorMessage(""), Entry(NULL),
-      NextInstNumber(1), Live(nullptr),
+      IsInternalLinkage(false), HasError(false), FocusedTiming(false),
+      ErrorMessage(""), Entry(NULL), NextInstNumber(1), Live(nullptr),
       Target(TargetLowering::createLowering(Ctx->getTargetArch(), this)),
       VMetadata(new VariablesMetadata(this)),
       TargetAssembler(
           TargetLowering::createAssembler(Ctx->getTargetArch(), this)),
       CurrentNode(NULL) {}
 
 Cfg::~Cfg() {}
 
 void Cfg::setError(const IceString &Message) {
   HasError = true;
@@ skipping 22 matching lines @@
   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(); }
 
 void Cfg::translate() {
   if (hasError())
     return;
+  VerboseMask OldVerboseMask = getContext()->getVerbose();
+  const IceString &TimingFocusOn = getContext()->getFlags().TimingFocusOn;
+  if (TimingFocusOn == "*" || TimingFocusOn == getFunctionName())
+    setFocusedTiming();
+  bool VerboseFocus = (getContext()->getFlags().VerboseFocusOn == getFunctionName());

[jvoung (off chromium), 2014/10/06 16:30:56]

80 col ?

[Jim Stichnoth, 2014/10/06 21:29:57]

Done.

+  if (VerboseFocus)
+    getContext()->setVerbose(IceV_All);
   static TimerIdT IDtranslate = GlobalContext::getTimerID("translate");
-  TimerMarker T(IDtranslate, getContext());
+  TimerMarker T(IDtranslate, this);
 
   dump("Initial CFG");
 
   // The set of translation passes and their order are determined by
   // the target.
   getTarget()->translate();
 
   dump("Final output");
+  if (getFocusedTiming())
+    getContext()->dumpTimers();
+  if (VerboseFocus)
+    getContext()->setVerbose(OldVerboseMask);
 }
 
 void Cfg::computePredecessors() {
   for (CfgNode *Node : Nodes)
     Node->computePredecessors();
 }
 
 void Cfg::renumberInstructions() {
   static TimerIdT IDrenumberInstructions =
       GlobalContext::getTimerID("renumberInstructions");
-  TimerMarker T(IDrenumberInstructions, getContext());
+  TimerMarker T(IDrenumberInstructions, this);
   NextInstNumber = 1;
   for (CfgNode *Node : Nodes)
     Node->renumberInstructions();
 }
 
 // placePhiLoads() must be called before placePhiStores().
 void Cfg::placePhiLoads() {
   static TimerIdT IDplacePhiLoads = GlobalContext::getTimerID("placePhiLoads");
-  TimerMarker T(IDplacePhiLoads, getContext());
+  TimerMarker T(IDplacePhiLoads, this);
   for (CfgNode *Node : Nodes)
     Node->placePhiLoads();
 }
 
 // placePhiStores() must be called after placePhiLoads().
 void Cfg::placePhiStores() {
   static TimerIdT IDplacePhiStores =
       GlobalContext::getTimerID("placePhiStores");
-  TimerMarker T(IDplacePhiStores, getContext());
+  TimerMarker T(IDplacePhiStores, this);
   for (CfgNode *Node : Nodes)
     Node->placePhiStores();
 }
 
 void Cfg::deletePhis() {
   static TimerIdT IDdeletePhis = GlobalContext::getTimerID("deletePhis");
-  TimerMarker T(IDdeletePhis, getContext());
+  TimerMarker T(IDdeletePhis, this);
   for (CfgNode *Node : Nodes)
     Node->deletePhis();
 }
 
 void Cfg::doArgLowering() {
   static TimerIdT IDdoArgLowering = GlobalContext::getTimerID("doArgLowering");
-  TimerMarker T(IDdoArgLowering, getContext());
+  TimerMarker T(IDdoArgLowering, this);
   getTarget()->lowerArguments();
 }
 
 void Cfg::doAddressOpt() {
   static TimerIdT IDdoAddressOpt = GlobalContext::getTimerID("doAddressOpt");
-  TimerMarker T(IDdoAddressOpt, getContext());
+  TimerMarker T(IDdoAddressOpt, this);
   for (CfgNode *Node : Nodes)
     Node->doAddressOpt();
 }
 
 void Cfg::doNopInsertion() {
   static TimerIdT IDdoNopInsertion =
       GlobalContext::getTimerID("doNopInsertion");
-  TimerMarker T(IDdoNopInsertion, getContext());
+  TimerMarker T(IDdoNopInsertion, this);
   for (CfgNode *Node : Nodes)
     Node->doNopInsertion();
 }
 
 void Cfg::genCode() {
   static TimerIdT IDgenCode = GlobalContext::getTimerID("genCode");
-  TimerMarker T(IDgenCode, getContext());
+  TimerMarker T(IDgenCode, this);
   for (CfgNode *Node : Nodes)
     Node->genCode();
 }
 
 // Compute the stack frame layout.
 void Cfg::genFrame() {
   static TimerIdT IDgenFrame = GlobalContext::getTimerID("genFrame");
-  TimerMarker T(IDgenFrame, getContext());
+  TimerMarker T(IDgenFrame, this);
   getTarget()->addProlog(Entry);
   // TODO: Consider folding epilog generation into the final
   // emission/assembly pass to avoid an extra iteration over the node
   // list.  Or keep a separate list of exit nodes.
   for (CfgNode *Node : Nodes)
     if (Node->getHasReturn())
       getTarget()->addEpilog(Node);
 }
 
 // This is a lightweight version of live-range-end calculation.  Marks
 // the last use of only those variables whose definition and uses are
 // completely with a single block.  It is a quick single pass and
 // doesn't need to iterate until convergence.
 void Cfg::livenessLightweight() {
   static TimerIdT IDlivenessLightweight =
       GlobalContext::getTimerID("livenessLightweight");
-  TimerMarker T(IDlivenessLightweight, getContext());
+  TimerMarker T(IDlivenessLightweight, this);
   getVMetadata()->init();
   for (CfgNode *Node : Nodes)
     Node->livenessLightweight();
 }
 
 void Cfg::liveness(LivenessMode Mode) {
   static TimerIdT IDliveness = GlobalContext::getTimerID("liveness");
-  TimerMarker T(IDliveness, getContext());
+  TimerMarker T(IDliveness, this);
   Live.reset(new Liveness(this, Mode));
   getVMetadata()->init();
   Live->init();
   // Initialize with all nodes needing to be processed.
   llvm::BitVector NeedToProcess(Nodes.size(), true);
   while (NeedToProcess.any()) {
     // Iterate in reverse topological order to speed up convergence.
     // TODO(stichnot): Use llvm::make_range with LLVM 3.5.
     for (auto I = Nodes.rbegin(), E = Nodes.rend(); I != E; ++I) {
       CfgNode *Node = *I;
@@ skipping 13 matching lines @@
     // Reset each variable's live range.
     for (Variable *Var : Variables)
       Var->resetLiveRange();
   }
   // Collect timing for just the portion that constructs the live
   // range intervals based on the end-of-live-range computation, for a
   // finer breakdown of the cost.
   // Make a final pass over instructions to delete dead instructions
   // and build each Variable's live range.
   static TimerIdT IDliveRange = GlobalContext::getTimerID("liveRange");
-  TimerMarker T1(IDliveRange, getContext());
+  TimerMarker T1(IDliveRange, this);
   for (CfgNode *Node : Nodes)
     Node->livenessPostprocess(Mode, getLiveness());
   if (Mode == Liveness_Intervals) {
     // Special treatment for live in-args.  Their liveness needs to
     // extend beyond the beginning of the function, otherwise an arg
     // whose only use is in the first instruction will end up having
     // the trivial live range [1,1) and will *not* interfere with
     // other arguments.  So if the first instruction of the method is
     // "r=arg1+arg2", both args may be assigned the same register.
     for (SizeT I = 0; I < Args.size(); ++I) {
@@ skipping 27 matching lines @@
     }
     dump();
   }
 }
 
 // Traverse every Variable of every Inst and verify that it
 // appears within the Variable's computed live range.
 bool Cfg::validateLiveness() const {
   static TimerIdT IDvalidateLiveness =
       GlobalContext::getTimerID("validateLiveness");
-  TimerMarker T(IDvalidateLiveness, getContext());
+  TimerMarker T(IDvalidateLiveness, this);
   bool Valid = true;
   Ostream &Str = Ctx->getStrDump();
   for (CfgNode *Node : Nodes) {
     for (Inst *Inst : Node->getInsts()) {
       if (Inst->isDeleted())
         continue;
       if (llvm::isa<InstFakeKill>(Inst))
         continue;
       InstNumberT InstNumber = Inst->getNumber();
       Variable *Dest = Inst->getDest();
@@ skipping 23 matching lines @@
           }
         }
       }
     }
   }
   return Valid;
 }
 
 void Cfg::doBranchOpt() {
   static TimerIdT IDdoBranchOpt = GlobalContext::getTimerID("doBranchOpt");
-  TimerMarker T(IDdoBranchOpt, getContext());
+  TimerMarker T(IDdoBranchOpt, this);
   for (auto I = Nodes.begin(), E = Nodes.end(); I != E; ++I) {
     auto NextNode = I;
     ++NextNode;
     (*I)->doBranchOpt(NextNode == E ? NULL : *NextNode);
   }
 }
 
 // ======================== Dump routines ======================== //
 
 void Cfg::emit() {
   static TimerIdT IDemit = GlobalContext::getTimerID("emit");
-  TimerMarker T(IDemit, getContext());
+  TimerMarker T(IDemit, this);
   Ostream &Str = Ctx->getStrEmit();
   if (!Ctx->testAndSetHasEmittedFirstMethod()) {
     // Print a helpful command for assembling the output.
     // TODO: have the Target emit the header
     // TODO: need a per-file emit in addition to per-CFG
     Str << "# $LLVM_BIN_PATH/llvm-mc"
         << " -arch=x86"
         << " -x86-asm-syntax=intel"
         << " -filetype=obj"
         << " -o=MyObj.o"
@@ skipping 53 matching lines @@
     }
   }
   // Print each basic block
   for (CfgNode *Node : Nodes)
     Node->dump(this);
   if (getContext()->isVerbose(IceV_Instructions))
     Str << "}\n";
 }
 
 } // end of namespace Ice