Subzero: Decorate the text asm output with register availability info.

src/IceCfg.cpp

@@ -365,20 +365,15 @@ bool Cfg::validateLiveness() const {
   return Valid;
 }
 
-// Deletes redundant assignments like "var=var".  This includes
-// architecturally redundant moves like "var1:eax=var2:eax".  As such,
-// this needs to be done very late in the translation to avoid
-// liveness inconsistencies.
-void Cfg::deleteRedundantAssignments() {
-  for (CfgNode *Node : Nodes) {
-    // Ignore Phi instructions.
-    for (Inst *I : Node->getInsts())
-      if (I->isRedundantAssign())
-        I->setDeleted();
-  }
-}
-
 void Cfg::contractEmptyNodes() {
+  // If we're decorating the asm output with register liveness info,
+  // this information may become corrupted or incorrect after
+  // contracting nodes that contain only redundant assignments.  As
+  // such, we disable this pass when DecorateAsm is specified.  This
+  // may make the resulting code look more branchy, but it should have
+  // no effect on the register assignments.
+  if (Ctx->getFlags().DecorateAsm)
+    return;
   for (CfgNode *Node : Nodes) {
     Node->contractIfEmpty();
   }
@@ -397,6 +392,12 @@ void Cfg::doBranchOpt() {
 
 void Cfg::emit() {
   TimerMarker T(TimerStack::TT_emit, this);
+  if (Ctx->getFlags().DecorateAsm) {
+    renumberInstructions();
+    getVMetadata()->init(VMK_Uses);
+    liveness(Liveness_Basic);
+    dump("After recomputing liveness for -decorate-asm");
+  }
   Ostream &Str = Ctx->getStrEmit();
   if (!Ctx->testAndSetHasEmittedFirstMethod()) {
     // Print a helpful command for assembling the output.

src/IceCfg.h

@@ -121,7 +121,6 @@ public:
   void livenessLightweight();
   void liveness(LivenessMode Mode);
   bool validateLiveness() const;
-  void deleteRedundantAssignments();
   void contractEmptyNodes();
   void doBranchOpt();
 

src/IceCfgNode.cpp

@@ -756,9 +756,12 @@ void CfgNode::contractIfEmpty() {
     return;
   Inst *Branch = NULL;
   for (Inst *I : Insts) {
-    if (!I->isDeleted() && !I->isUnconditionalBranch())
-      return;
+    if (I->isDeleted())
+      continue;
+    if (I->isUnconditionalBranch())
       Branch = I;
+    else if (!I->isRedundantAssign())
+      return;
   }
   Branch->setDeleted();
   assert(OutEdges.size() == 1);
@@ -798,9 +801,73 @@ void CfgNode::doBranchOpt(const CfgNode *NextNode) {
 
 // ======================== Dump routines ======================== //
 
+namespace {
+
+// Helper functions for emit().
+
+void emitRegisterUsage(Ostream &Str, const Cfg *Func, const CfgNode *Node,
+                       bool IsLiveIn, std::vector<SizeT> &LiveRegCount) {
+  Liveness *Liveness = Func->getLiveness();
+  const LivenessBV *Live;
+  if (IsLiveIn) {
+    Live = &Liveness->getLiveIn(Node);
+    Str << "\t\t\t\t# LiveIn=";
+  } else {
+    Live = &Liveness->getLiveOut(Node);
+    Str << "\t\t\t\t# LiveOut=";
+  }
+  if (!Live->empty()) {
+    bool First = true;
+    for (SizeT i = 0; i < Live->size(); ++i) {
+      if ((*Live)[i]) {
+        Variable *Var = Liveness->getVariable(i, Node);
+        if (Var->hasReg()) {
+          if (IsLiveIn)
+            ++LiveRegCount[Var->getRegNum()];
+          if (!First)
+            Str << ",";
+          First = false;
+          Var->emit(Func);
+        }
+      }
+    }
+  }
+  Str << "\n";
+}
+
+void emitLiveRangesEnded(Ostream &Str, const Cfg *Func, const Inst *Instr,
+                         std::vector<SizeT> &LiveRegCount) {
+  bool First = true;
+  Variable *Dest = Instr->getDest();
+  if (Dest && Dest->hasReg())
+    ++LiveRegCount[Dest->getRegNum()];
+  for (SizeT I = 0; I < Instr->getSrcSize(); ++I) {
+    Operand *Src = Instr->getSrc(I);
+    SizeT NumVars = Src->getNumVars();
+    for (SizeT J = 0; J < NumVars; ++J) {
+      const Variable *Var = Src->getVar(J);
+      if (Var->hasReg()) {
+        if (Instr->isLastUse(Var) &&
+            --LiveRegCount[Var->getRegNum()] == 0) {
+          if (First)
+            Str << " \t# END=";
+          else
+            Str << ",";
+          Var->emit(Func);
+          First = false;
+        }
+      }
+    }
+  }
+}
+
+} // end of anonymous namespace
+
 void CfgNode::emit(Cfg *Func) const {
   Func->setCurrentNode(this);
   Ostream &Str = Func->getContext()->getStrEmit();
+  Liveness *Liveness = Func->getLiveness();
+  bool DecorateAsm = Liveness && Func->getContext()->getFlags().DecorateAsm;
   if (Func->getEntryNode() == this) {
     Str << Func->getContext()->mangleName(Func->getFunctionName()) << ":\n";
   }
@@ -809,6 +876,10 @@ void CfgNode::emit(Cfg *Func) const {
     Assembler *Asm = Func->getAssembler<Assembler>();
     Asm->BindCfgNodeLabel(getIndex());
   }
+  std::vector<SizeT> LiveRegCount(Func->getTarget()->getNumRegisters());
+  if (DecorateAsm)
+    emitRegisterUsage(Str, Func, this, true, LiveRegCount);
+
   for (InstPhi *Phi : Phis) {
     if (Phi->isDeleted())
       continue;
@@ -819,10 +890,18 @@ void CfgNode::emit(Cfg *Func) const {
   for (Inst *I : Insts) {
     if (I->isDeleted())
       continue;
+    if (I->isRedundantAssign()) {
+      Variable *Dest = I->getDest();
+      if (DecorateAsm && Dest->hasReg() && !I->isLastUse(I->getSrc(0)))
+        ++LiveRegCount[Dest->getRegNum()];
+      continue;
+    }
     if (Func->useIntegratedAssembler()) {
       I->emitIAS(Func);
     } else {
       I->emit(Func);
+      if (DecorateAsm)
+        emitLiveRangesEnded(Str, Func, I, LiveRegCount);
       Str << "\n";
     }
     // Update emitted instruction count, plus fill/spill count for
@@ -841,6 +920,8 @@ void CfgNode::emit(Cfg *Func) const {
       }
     }
   }
+  if (DecorateAsm)
+    emitRegisterUsage(Str, Func, this, false, LiveRegCount);
 }
 
 void CfgNode::dump(Cfg *Func) const {

src/IceClFlags.h

@@ -18,13 +18,13 @@
 
 namespace Ice {
 
-// TODO(stichnot) Move more command line flags into ClFlags.
 class ClFlags {
 public:
   ClFlags()
       : DisableInternal(false), SubzeroTimingEnabled(false),
         DisableTranslation(false), FunctionSections(false), DataSections(false),
-        UseIntegratedAssembler(false), UseSandboxing(false), DumpStats(false),
+        UseIntegratedAssembler(false), UseSandboxing(false),
+        PhiEdgeSplit(false), DecorateAsm(false), DumpStats(false),
         AllowUninitializedGlobals(false), TimeEachFunction(false),
         DefaultGlobalPrefix(""), DefaultFunctionPrefix(""), TimingFocusOn(""),
         VerboseFocusOn(""), TranslateOnly("") {}
@@ -36,6 +36,7 @@ public:
   bool UseIntegratedAssembler;
   bool UseSandboxing;
   bool PhiEdgeSplit;
+  bool DecorateAsm;
   bool DumpStats;
   bool AllowUninitializedGlobals;
   bool TimeEachFunction;

src/IceTargetLowering.h

@@ -153,6 +153,7 @@ public:
     return false;
   }
 
+  virtual SizeT getNumRegisters() const = 0;
   // Returns a variable pre-colored to the specified physical
   // register.  This is generally used to get very direct access to
   // the register such as in the prolog or epilog or for marking

src/IceTargetLoweringX8632.cpp

@@ -392,7 +392,6 @@ void TargetX8632::translateO2() {
     return;
   Func->dump("After stack frame mapping");
 
-  Func->deleteRedundantAssignments();
   Func->contractEmptyNodes();
   Func->reorderNodes();
 
@@ -435,8 +434,6 @@ void TargetX8632::translateOm1() {
     return;
   Func->dump("After stack frame mapping");
 
-  Func->deleteRedundantAssignments();
-
   // Nop insertion
   if (shouldDoNopInsertion()) {
     Func->doNopInsertion();

src/IceTargetLoweringX8632.h

@@ -35,6 +35,7 @@ public:
   void translateO2() override;
   bool doBranchOpt(Inst *I, const CfgNode *NextNode) override;
 
+  SizeT getNumRegisters() const override { return RegX8632::Reg_NUM; }
   Variable *getPhysicalRegister(SizeT RegNum, Type Ty = IceType_void) override;
   IceString getRegName(SizeT RegNum, Type Ty) const override;
   llvm::SmallBitVector getRegisterSet(RegSetMask Include,

src/llvm2ice.cpp

@@ -120,6 +120,10 @@ EnablePhiEdgeSplit("phi-edge-split",
                    cl::desc("Enable edge splitting for Phi lowering"),
                    cl::init(true));
 
+static cl::opt<bool> DecorateAsm(
+    "asm-verbose",
+    cl::desc("Decorate textual asm output with register liveness info"));
+
 static cl::opt<bool>
 DumpStats("stats",
           cl::desc("Print statistics after translating each function"));
@@ -256,6 +260,7 @@ int main(int argc, char **argv) {
   Flags.UseIntegratedAssembler = UseIntegratedAssembler;
   Flags.UseSandboxing = UseSandboxing;
   Flags.PhiEdgeSplit = EnablePhiEdgeSplit;
+  Flags.DecorateAsm = DecorateAsm;
   Flags.DumpStats = DumpStats;
   Flags.AllowUninitializedGlobals = AllowUninitializedGlobals;
   Flags.TimeEachFunction = TimeEachFunction;