8347901: C2 should remove unused leaf / pure runtime calls

Reviewed-by: thartmann, vlivanov

Author: marc-chevalier

src/hotspot/share/opto/callnode.cpp

@@ -918,7 +918,7 @@ Node *CallNode::result_cast() {
 }
 
 
-void CallNode::extract_projections(CallProjections* projs, bool separate_io_proj, bool do_asserts) {
+void CallNode::extract_projections(CallProjections* projs, bool separate_io_proj, bool do_asserts) const {
   projs->fallthrough_proj      = nullptr;
   projs->fallthrough_catchproj = nullptr;
   projs->fallthrough_ioproj    = nullptr;
@@ -1303,6 +1303,57 @@ void CallLeafVectorNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_
 
 
 //=============================================================================
+bool CallLeafPureNode::is_unused() const {
+  return proj_out_or_null(TypeFunc::Parms) == nullptr;
+}
+
+bool CallLeafPureNode::is_dead() const {
+  return proj_out_or_null(TypeFunc::Control) == nullptr;
+}
+
+/* We make a tuple of the global input state + TOP for the output values.
+ * We use this to delete a pure function that is not used: by replacing the call with
+ * such a tuple, we let output Proj's idealization pick the corresponding input of the
+ * pure call, so jumping over it, and effectively, removing the call from the graph.
+ * This avoids doing the graph surgery manually, but leaves that to IGVN
+ * that is specialized for doing that right. We need also tuple components for output
+ * values of the function to respect the return arity, and in case there is a projection
+ * that would pick an output (which shouldn't happen at the moment).
+ */
+TupleNode* CallLeafPureNode::make_tuple_of_input_state_and_top_return_values(const Compile* C) const {
+  // Transparently propagate input state but parameters
+  TupleNode* tuple = TupleNode::make(
+      tf()->range(),
+      in(TypeFunc::Control),
+      in(TypeFunc::I_O),
+      in(TypeFunc::Memory),
+      in(TypeFunc::FramePtr),
+      in(TypeFunc::ReturnAdr));
+
+  // And add TOPs for the return values
+  for (uint i = TypeFunc::Parms; i < tf()->range()->cnt(); i++) {
+    tuple->set_req(i, C->top());
+  }
+
+  return tuple;
+}
+
+Node* CallLeafPureNode::Ideal(PhaseGVN* phase, bool can_reshape) {
+  if (is_dead()) {
+    return nullptr;
+  }
+
+  // We need to wait until IGVN because during parsing, usages might still be missing
+  // and we would remove the call immediately.
+  if (can_reshape && is_unused()) {
+    // The result is not used. We remove the call by replacing it with a tuple, that
+    // is later disintegrated by the projections.
+    return make_tuple_of_input_state_and_top_return_values(phase->C);
+  }
+
+  return CallRuntimeNode::Ideal(phase, can_reshape);
+}
+
 #ifndef PRODUCT
 void CallLeafNode::dump_spec(outputStream *st) const {
   st->print("# ");

src/hotspot/share/opto/callnode.hpp

@@ -738,7 +738,7 @@ class CallNode : public SafePointNode {
   // Collect all the interesting edges from a call for use in
   // replacing the call by something else.  Used by macro expansion
   // and the late inlining support.
-  void extract_projections(CallProjections* projs, bool separate_io_proj, bool do_asserts = true);
+  void extract_projections(CallProjections* projs, bool separate_io_proj, bool do_asserts = true) const;
 
   virtual uint match_edge(uint idx) const;
 
@@ -914,6 +914,33 @@ class CallLeafNode : public CallRuntimeNode {
 #endif
 };
 
+/* A pure function call, they are assumed not to be safepoints, not to read or write memory,
+ * have no exception... They just take parameters, return a value without side effect. It is
+ * always correct to create some, or remove them, if the result is not used.
+ *
+ * They still have control input to allow easy lowering into other kind of calls that require
+ * a control, but this is more a technical than a moral constraint.
+ *
+ * Pure calls must have only control and data input and output: I/O, Memory and so on must be top.
+ * Nevertheless, pure calls can typically be expensive math operations so care must be taken
+ * when letting the node float.
+ */
+class CallLeafPureNode : public CallLeafNode {
+protected:
+  bool is_unused() const;
+  bool is_dead() const;
+  TupleNode* make_tuple_of_input_state_and_top_return_values(const Compile* C) const;
+
+public:
+  CallLeafPureNode(const TypeFunc* tf, address addr, const char* name,
+                   const TypePtr* adr_type)
+      : CallLeafNode(tf, addr, name, adr_type) {
+    init_class_id(Class_CallLeafPure);
+  }
+  int Opcode() const override;
+  Node* Ideal(PhaseGVN* phase, bool can_reshape) override;
+};
+
 //------------------------------CallLeafNoFPNode-------------------------------
 // CallLeafNode, not using floating point or using it in the same manner as
 // the generated code

src/hotspot/share/opto/classes.hpp

@@ -61,6 +61,7 @@ macro(CallDynamicJava)
 macro(CallJava)
 macro(CallLeaf)
 macro(CallLeafNoFP)
+macro(CallLeafPure)
 macro(CallLeafVector)
 macro(CallRuntime)
 macro(CallStaticJava)
@@ -372,6 +373,7 @@ macro(SubI)
 macro(SubL)
 macro(TailCall)
 macro(TailJump)
+macro(Tuple)
 macro(MacroLogicV)
 macro(ThreadLocal)
 macro(Unlock)

src/hotspot/share/opto/compile.cpp

@@ -3303,6 +3303,25 @@ void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& f
   case Op_Opaque1:              // Remove Opaque Nodes before matching
     n->subsume_by(n->in(1), this);
     break;
+  case Op_CallLeafPure: {
+    // If the pure call is not supported, then lower to a CallLeaf.
+    if (!Matcher::match_rule_supported(Op_CallLeafPure)) {
+      CallNode* call = n->as_Call();
+      CallNode* new_call = new CallLeafNode(call->tf(), call->entry_point(),
+                                            call->_name, TypeRawPtr::BOTTOM);
+      new_call->init_req(TypeFunc::Control, call->in(TypeFunc::Control));
+      new_call->init_req(TypeFunc::I_O, C->top());
+      new_call->init_req(TypeFunc::Memory, C->top());
+      new_call->init_req(TypeFunc::ReturnAdr, C->top());
+      new_call->init_req(TypeFunc::FramePtr, C->top());
+      for (unsigned int i = TypeFunc::Parms; i < call->tf()->domain()->cnt(); i++) {
+        new_call->init_req(i, call->in(i));
+      }
+      n->subsume_by(new_call, this);
+    }
+    frc.inc_call_count();
+    break;
+  }
   case Op_CallStaticJava:
   case Op_CallJava:
   case Op_CallDynamicJava:

src/hotspot/share/opto/divnode.cpp

@@ -42,19 +42,19 @@
 
 #include <math.h>
 
-ModFloatingNode::ModFloatingNode(Compile* C, const TypeFunc* tf, const char* name) : CallLeafNode(tf, nullptr, name, TypeRawPtr::BOTTOM) {
+ModFloatingNode::ModFloatingNode(Compile* C, const TypeFunc* tf, address addr, const char* name) : CallLeafPureNode(tf, addr, name, TypeRawPtr::BOTTOM) {
   add_flag(Flag_is_macro);
   C->add_macro_node(this);
 }
 
-ModDNode::ModDNode(Compile* C, Node* a, Node* b) : ModFloatingNode(C, OptoRuntime::Math_DD_D_Type(), "drem") {
+ModDNode::ModDNode(Compile* C, Node* a, Node* b) : ModFloatingNode(C, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::drem), "drem") {
   init_req(TypeFunc::Parms + 0, a);
   init_req(TypeFunc::Parms + 1, C->top());
   init_req(TypeFunc::Parms + 2, b);
   init_req(TypeFunc::Parms + 3, C->top());
 }
 
-ModFNode::ModFNode(Compile* C, Node* a, Node* b) : ModFloatingNode(C, OptoRuntime::modf_Type(), "frem") {
+ModFNode::ModFNode(Compile* C, Node* a, Node* b) : ModFloatingNode(C, OptoRuntime::modf_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::frem), "frem") {
   init_req(TypeFunc::Parms + 0, a);
   init_req(TypeFunc::Parms + 1, b);
 }
@@ -1511,137 +1511,109 @@ const Type* UModLNode::Value(PhaseGVN* phase) const {
   return unsigned_mod_value<TypeLong, julong, jlong>(phase, this);
 }
 
-Node* ModFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
-  if (!can_reshape) {
-    return nullptr;
-  }
-  PhaseIterGVN* igvn = phase->is_IterGVN();
-
-  bool result_is_unused = proj_out_or_null(TypeFunc::Parms) == nullptr;
-  bool not_dead = proj_out_or_null(TypeFunc::Control) != nullptr;
-  if (result_is_unused && not_dead) {
-    return replace_with_con(igvn, TypeF::make(0.));
-  }
-
-  // Either input is TOP ==> the result is TOP
-  const Type* t1 = phase->type(dividend());
-  const Type* t2 = phase->type(divisor());
-  if (t1 == Type::TOP || t2 == Type::TOP) {
-    return phase->C->top();
-  }
-
+const Type* ModFNode::get_result_if_constant(const Type* dividend, const Type* divisor) const {
   // If either number is not a constant, we know nothing.
-  if ((t1->base() != Type::FloatCon) || (t2->base() != Type::FloatCon)) {
+  if ((dividend->base() != Type::FloatCon) || (divisor->base() != Type::FloatCon)) {
     return nullptr; // note: x%x can be either NaN or 0
   }
 
-  float f1 = t1->getf();
-  float f2 = t2->getf();
-  jint x1 = jint_cast(f1); // note:  *(int*)&f1, not just (int)f1
-  jint x2 = jint_cast(f2);
+  float dividend_f = dividend->getf();
+  float divisor_f = divisor->getf();
+  jint dividend_i = jint_cast(dividend_f); // note:  *(int*)&f1, not just (int)f1
+  jint divisor_i = jint_cast(divisor_f);
 
   // If either is a NaN, return an input NaN
-  if (g_isnan(f1)) {
-    return replace_with_con(igvn, t1);
+  if (g_isnan(dividend_f)) {
+    return dividend;
   }
-  if (g_isnan(f2)) {
-    return replace_with_con(igvn, t2);
+  if (g_isnan(divisor_f)) {
+    return divisor;
   }
 
   // If an operand is infinity or the divisor is +/- zero, punt.
-  if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jint) {
+  if (!g_isfinite(dividend_f) || !g_isfinite(divisor_f) || divisor_i == 0 || divisor_i == min_jint) {
     return nullptr;
   }
 
   // We must be modulo'ing 2 float constants.
   // Make sure that the sign of the fmod is equal to the sign of the dividend
-  jint xr = jint_cast(fmod(f1, f2));
-  if ((x1 ^ xr) < 0) {
+  jint xr = jint_cast(fmod(dividend_f, divisor_f));
+  if ((dividend_i ^ xr) < 0) {
     xr ^= min_jint;
   }
 
-  return replace_with_con(igvn, TypeF::make(jfloat_cast(xr)));
+  return TypeF::make(jfloat_cast(xr));
 }
 
-Node* ModDNode::Ideal(PhaseGVN* phase, bool can_reshape) {
-  if (!can_reshape) {
-    return nullptr;
-  }
-  PhaseIterGVN* igvn = phase->is_IterGVN();
-
-  bool result_is_unused = proj_out_or_null(TypeFunc::Parms) == nullptr;
-  bool not_dead = proj_out_or_null(TypeFunc::Control) != nullptr;
-  if (result_is_unused && not_dead) {
-    return replace_with_con(igvn, TypeD::make(0.));
-  }
-
-  // Either input is TOP ==> the result is TOP
-  const Type* t1 = phase->type(dividend());
-  const Type* t2 = phase->type(divisor());
-  if (t1 == Type::TOP || t2 == Type::TOP) {
-    return nullptr;
-  }
-
+const Type* ModDNode::get_result_if_constant(const Type* dividend, const Type* divisor) const {
   // If either number is not a constant, we know nothing.
-  if ((t1->base() != Type::DoubleCon) || (t2->base() != Type::DoubleCon)) {
+  if ((dividend->base() != Type::DoubleCon) || (divisor->base() != Type::DoubleCon)) {
     return nullptr; // note: x%x can be either NaN or 0
   }
 
-  double f1 = t1->getd();
-  double f2 = t2->getd();
-  jlong x1 = jlong_cast(f1); // note:  *(long*)&f1, not just (long)f1
-  jlong x2 = jlong_cast(f2);
+  double dividend_d = dividend->getd();
+  double divisor_d = divisor->getd();
+  jlong dividend_l = jlong_cast(dividend_d); // note:  *(long*)&f1, not just (long)f1
+  jlong divisor_l = jlong_cast(divisor_d);
 
   // If either is a NaN, return an input NaN
-  if (g_isnan(f1)) {
-    return replace_with_con(igvn, t1);
+  if (g_isnan(dividend_d)) {
+    return dividend;
   }
-  if (g_isnan(f2)) {
-    return replace_with_con(igvn, t2);
+  if (g_isnan(divisor_d)) {
+    return divisor;
   }
 
   // If an operand is infinity or the divisor is +/- zero, punt.
-  if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jlong) {
+  if (!g_isfinite(dividend_d) || !g_isfinite(divisor_d) || divisor_l == 0 || divisor_l == min_jlong) {
     return nullptr;
   }
 
   // We must be modulo'ing 2 double constants.
   // Make sure that the sign of the fmod is equal to the sign of the dividend
-  jlong xr = jlong_cast(fmod(f1, f2));
-  if ((x1 ^ xr) < 0) {
+  jlong xr = jlong_cast(fmod(dividend_d, divisor_d));
+  if ((dividend_l ^ xr) < 0) {
     xr ^= min_jlong;
   }
 
-  return replace_with_con(igvn, TypeD::make(jdouble_cast(xr)));
+  return TypeD::make(jdouble_cast(xr));
 }
 
-Node* ModFloatingNode::replace_with_con(PhaseIterGVN* phase, const Type* con) {
-  Compile* C = phase->C;
-  Node* con_node = phase->makecon(con);
-  CallProjections projs;
-  extract_projections(&projs, false, false);
-  phase->replace_node(projs.fallthrough_proj, in(TypeFunc::Control));
-  if (projs.fallthrough_catchproj != nullptr) {
-    phase->replace_node(projs.fallthrough_catchproj, in(TypeFunc::Control));
-  }
-  if (projs.fallthrough_memproj != nullptr) {
-    phase->replace_node(projs.fallthrough_memproj, in(TypeFunc::Memory));
-  }
-  if (projs.catchall_memproj != nullptr) {
-    phase->replace_node(projs.catchall_memproj, C->top());
-  }
-  if (projs.fallthrough_ioproj != nullptr) {
-    phase->replace_node(projs.fallthrough_ioproj, in(TypeFunc::I_O));
-  }
-  assert(projs.catchall_ioproj == nullptr, "no exceptions from floating mod");
-  assert(projs.catchall_catchproj == nullptr, "no exceptions from floating mod");
-  if (projs.resproj != nullptr) {
-    phase->replace_node(projs.resproj, con_node);
+Node* ModFloatingNode::Ideal(PhaseGVN* phase, bool can_reshape) {
+  if (can_reshape) {
+    PhaseIterGVN* igvn = phase->is_IterGVN();
+
+    // Either input is TOP ==> the result is TOP
+    const Type* dividend_type = phase->type(dividend());
+    const Type* divisor_type = phase->type(divisor());
+    if (dividend_type == Type::TOP || divisor_type == Type::TOP) {
+      return phase->C->top();
+    }
+    const Type* constant_result = get_result_if_constant(dividend_type, divisor_type);
+    if (constant_result != nullptr) {
+      return make_tuple_of_input_state_and_constant_result(igvn, constant_result);
+    }
   }
-  phase->replace_node(this, C->top());
-  C->remove_macro_node(this);
-  disconnect_inputs(C);
-  return nullptr;
+
+  return CallLeafPureNode::Ideal(phase, can_reshape);
+}
+
+/* Give a tuple node for ::Ideal to return, made of the input state (control to return addr)
+ * and the given constant result. Idealization of projections will make sure to transparently
+ * propagate the input state and replace the result by the said constant.
+ */
+TupleNode* ModFloatingNode::make_tuple_of_input_state_and_constant_result(PhaseIterGVN* phase, const Type* con) const {
+  Node* con_node = phase->makecon(con);
+  TupleNode* tuple = TupleNode::make(
+      tf()->range(),
+      in(TypeFunc::Control),
+      in(TypeFunc::I_O),
+      in(TypeFunc::Memory),
+      in(TypeFunc::FramePtr),
+      in(TypeFunc::ReturnAdr),
+      con_node);
+
+  return tuple;
 }
 
 //=============================================================================