8342103: C2 compiler support for Float16 type and associated operations

src/hotspot/cpu/x86/x86.ad

@@ -4514,18 +4514,43 @@ instruct vReplS_reg(vec dst, rRegI src) %{
 %}
 
 #ifdef _LP64
-instruct ReplS_imm(vec dst, immH con, rRegI rtmp) %{
-  predicate(VM_Version::supports_avx512_fp16() && Matcher::vector_element_basic_type(n) == T_SHORT);
+instruct ReplH_imm(vec dst, immH con, rRegI rtmp) %{
   match(Set dst (Replicate con));
   effect(TEMP rtmp);
-  format %{ "replicateS $dst, $con" %}
+  format %{ "replicateH $dst, $con \t! using $rtmp as TEMP" %}
   ins_encode %{
     int vlen_enc = vector_length_encoding(this);
+    BasicType bt = Matcher::vector_element_basic_type(this);
+    assert(VM_Version::supports_avx512_fp16() && bt == T_SHORT, "");
     __ movl($rtmp$$Register, $con$$constant);
     __ evpbroadcastw($dst$$XMMRegister, $rtmp$$Register, vlen_enc);
   %}
   ins_pipe( pipe_slow );
 %}
+
+instruct ReplH_short_reg(vec dst, rRegI src) %{
+  predicate(VM_Version::supports_avx512_fp16() && Matcher::vector_element_basic_type(n) == T_SHORT);
+  match(Set dst (Replicate (ReinterpretS2HF src)));
+  format %{ "replicateH $dst, $src" %}
+  ins_encode %{
+    int vlen_enc = vector_length_encoding(this);
+    __ evpbroadcastw($dst$$XMMRegister, $src$$Register, vlen_enc);
+  %}
+  ins_pipe( pipe_slow );
+%}
+
+instruct ReplH_reg(vec dst, regF src, rRegI rtmp) %{
+  predicate(VM_Version::supports_avx512_fp16() && Matcher::vector_element_basic_type(n) == T_SHORT);
+  match(Set dst (Replicate src));
+  effect(TEMP rtmp);
+  format %{ "replicateH $dst, $src \t! using $rtmp as TEMP" %}
+  ins_encode %{
+    int vlen_enc = vector_length_encoding(this);
+    __ vmovw($rtmp$$Register, $src$$XMMRegister);
+    __ evpbroadcastw($dst$$XMMRegister, $rtmp$$Register, vlen_enc);
+  %}
+  ins_pipe( pipe_slow );
+%}
 #endif
 
 instruct ReplS_mem(vec dst, memory mem) %{
@@ -10902,7 +10927,7 @@ instruct reinterpretH2S(rRegI dst, regF src)
   ins_pipe(pipe_slow);
 %}
 
-instruct scalar_sqrt_fp16(regF dst, regF src)
+instruct scalar_sqrt_fp16_reg(regF dst, regF src)
 %{
   match(Set dst (SqrtHF src));
   format %{ "vsqrtsh $dst, $src" %}
@@ -10912,7 +10937,7 @@ instruct scalar_sqrt_fp16(regF dst, regF src)
   ins_pipe(pipe_slow);
 %}
 
-instruct scalar_binOps_fp16(regF dst, regF src1, regF src2)
+instruct scalar_binOps_fp16_reg(regF dst, regF src1, regF src2)
 %{
   match(Set dst (AddHF src1 src2));
   match(Set dst (DivHF src1 src2));
@@ -10928,7 +10953,7 @@ instruct scalar_binOps_fp16(regF dst, regF src1, regF src2)
   ins_pipe(pipe_slow);
 %}
 
-instruct scalar_fma_fp16(regF dst, regF src1, regF src2)
+instruct scalar_fma_fp16_reg(regF dst, regF src1, regF src2)
 %{
   match(Set dst (FmaHF  src2 (Binary dst src1)));
   effect(DEF dst);
@@ -11008,7 +11033,7 @@ instruct vector_fma_fp16_reg(vec dst, vec src1, vec src2)
   ins_pipe( pipe_slow );
 %}
 
-instruct vector_fmah_fp16_mem(vec dst, memory src1, vec src2)
+instruct vector_fma_fp16_mem(vec dst, memory src1, vec src2)
 %{
   match(Set dst (FmaVHF src2 (Binary dst (VectorReinterpret (LoadVector src1)))));
   effect(DEF dst);

src/hotspot/share/adlc/output_h.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 1998, 2023, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 1998, 2024, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it

src/hotspot/share/opto/addnode.cpp

@@ -710,15 +710,15 @@ Node *AddFNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 //=============================================================================
 //------------------------------add_of_identity--------------------------------
 // Check for addition of the identity
-const Type *AddHFNode::add_of_identity( const Type *t1, const Type *t2 ) const {
+const Type *AddHFNode::add_of_identity(const Type *t1, const Type *t2) const {
   return nullptr;
 }
 
 //------------------------------add_ring---------------------------------------
 // Supplied function returns the sum of the inputs.
 // This also type-checks the inputs for sanity.  Guaranteed never to
 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
-const Type *AddHFNode::add_ring( const Type *t0, const Type *t1 ) const {
+const Type *AddHFNode::add_ring(const Type *t0, const Type *t1) const {
   if (!t0->isa_half_float_constant() || !t1->isa_half_float_constant()) {
     return bottom_type();
   }
@@ -1601,7 +1601,7 @@ Node* MaxNode::Identity(PhaseGVN* phase) {
 }
 
 //------------------------------add_ring---------------------------------------
-const Type* MinHFNode::add_ring(const Type* t0, const Type* t1 ) const {
+const Type* MinHFNode::add_ring(const Type* t0, const Type* t1) const {
   const TypeH* r0 = t0->isa_half_float_constant();
   const TypeH* r1 = t1->isa_half_float_constant();
   if (r0 == nullptr || r1 == nullptr) {
@@ -1676,9 +1676,9 @@ const Type* MinDNode::add_ring(const Type* t0, const Type* t1) const {
 }
 
 //------------------------------add_ring---------------------------------------
-const Type* MaxFNode::add_ring(const Type* t0, const Type* t1) const {
-  const TypeF* r0 = t0->isa_float_constant();
-  const TypeF* r1 = t1->isa_float_constant();
+const Type* MaxHFNode::add_ring(const Type* t0, const Type* t1) const {
+  const TypeH* r0 = t0->isa_half_float_constant();
+  const TypeH* r1 = t1->isa_half_float_constant();
   if (r0 == nullptr || r1 == nullptr) {
     return bottom_type();
   }
@@ -1700,10 +1700,11 @@ const Type* MaxFNode::add_ring(const Type* t0, const Type* t1) const {
   return (jint_cast(f0) > jint_cast(f1)) ? r0 : r1;
 }
 
+
 //------------------------------add_ring---------------------------------------
-const Type* MaxHFNode::add_ring(const Type* t0, const Type* t1) const {
-  const TypeH* r0 = t0->isa_half_float_constant();
-  const TypeH* r1 = t1->isa_half_float_constant();
+const Type* MaxFNode::add_ring(const Type* t0, const Type* t1) const {
+  const TypeF* r0 = t0->isa_float_constant();
+  const TypeF* r1 = t1->isa_float_constant();
   if (r0 == nullptr || r1 == nullptr) {
     return bottom_type();
   }

src/hotspot/share/opto/addnode.hpp

@@ -166,12 +166,12 @@ class AddDNode : public AddNode {
 // Add 2 half-precision floats
 class AddHFNode : public AddNode {
 public:
-  AddHFNode( Node *in1, Node *in2 ) : AddNode(in1,in2) {}
+  AddHFNode(Node *in1, Node *in2) : AddNode(in1,in2) {}
   virtual int Opcode() const;
-  virtual const Type *add_of_identity( const Type *t1, const Type *t2 ) const;
-  virtual const Type *add_ring( const Type *, const Type * ) const;
-  virtual const Type *add_id() const { return TypeH::ZERO; }
-  virtual const Type *bottom_type() const { return Type::HALF_FLOAT; }
+  virtual const Type* add_of_identity(const Type *t1, const Type *t2) const;
+  virtual const Type* add_ring(const Type*, const Type*) const;
+  virtual const Type* add_id() const { return TypeH::ZERO; }
+  virtual const Type* bottom_type() const { return Type::HALF_FLOAT; }
   int max_opcode() const { return Op_MaxHF; }
   int min_opcode() const { return Op_MinHF; }
   virtual Node* Identity(PhaseGVN* phase) { return this; }

src/hotspot/share/opto/connode.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 1997, 2023, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it

src/hotspot/share/opto/connode.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 1997, 2023, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it

src/hotspot/share/opto/constantTable.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020, 2023, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2020, 2024, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -88,7 +88,7 @@ static int constant_size(ConstantTable::Constant* con) {
     return type2aelembytes(con->type()) * con->get_array()->length();
   }
   switch (con->type()) {
-  case T_SHORT:   return sizeof(jshort );
+  case T_SHORT:   return sizeof(jint   );
   case T_INT:     return sizeof(jint   );
   case T_LONG:    return sizeof(jlong  );
   case T_FLOAT:   return sizeof(jfloat );

src/hotspot/share/opto/convertnode.cpp

@@ -257,7 +257,7 @@ Node* ConvF2HFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
   if (Float16NodeFactory::is_binary_oper(in(1)->Opcode()) &&
       in(1)->in(1)->Opcode() == Op_ConvHF2F &&
       in(1)->in(2)->Opcode() == Op_ConvHF2F) {
-    if (Matcher::match_rule_supported(in(1)->Opcode()) &&
+    if (Matcher::match_rule_supported(Float16NodeFactory::get_float16_binary_oper(in(1)->Opcode())) &&
         Matcher::match_rule_supported(Op_ReinterpretS2HF) &&
         Matcher::match_rule_supported(Op_ReinterpretHF2S)) {
       Node* in1 = phase->transform(new ReinterpretS2HFNode(in(1)->in(1)->in(1)));
@@ -959,6 +959,25 @@ bool Float16NodeFactory::is_binary_oper(int opc) {
   }
 }
 
+int Float16NodeFactory::get_float16_binary_oper(int opc) {
+  switch(opc) {
+    case Op_AddF:
+      return Op_AddHF;
+    case Op_SubF:
+      return Op_SubHF;
+    case Op_MulF:
+      return Op_MulHF;
+    case Op_DivF:
+      return Op_DivHF;
+    case Op_MaxF:
+      return Op_MaxHF;
+    case Op_MinF:
+      return Op_MinHF;
+    default:
+      return false;
+  }
+}
+
 Node* Float16NodeFactory::make(int opc, Node* c, Node* in1, Node* in2) {
   switch(opc) {
     case Op_AddF: return new AddHFNode(in1, in2);

src/hotspot/share/opto/convertnode.hpp

@@ -297,6 +297,7 @@ class RoundDoubleModeNode: public Node {
 class Float16NodeFactory {
   public:
   static bool is_binary_oper(int opc);
+  static int get_float16_binary_oper(int opc);
   static Node* make(int opc, Node* c, Node* in1, Node* in2);
 };
 

src/hotspot/share/opto/divnode.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 1997, 2023, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -753,15 +753,15 @@ Node *DivFNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 // prevent hoisting the divide above an unsafe test.
 const Type* DivHFNode::Value(PhaseGVN* phase) const {
   // Either input is TOP ==> the result is TOP
-  const Type *t1 = phase->type( in(1) );
-  const Type *t2 = phase->type( in(2) );
-  if( t1 == Type::TOP ) return Type::TOP;
-  if( t2 == Type::TOP ) return Type::TOP;
+  const Type* t1 = phase->type(in(1));
+  const Type* t2 = phase->type(in(2));
+  if(t1 == Type::TOP) return Type::TOP;
+  if(t2 == Type::TOP) return Type::TOP;
 
   // Either input is BOTTOM ==> the result is the local BOTTOM
-  const Type *bot = bottom_type();
-  if( (t1 == bot) || (t2 == bot) ||
-      (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
+  const Type* bot = bottom_type();
+  if((t1 == bot) || (t2 == bot) ||
+     (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM))
     return bot;
 
   // x/x == 1, we ignore 0/0.
@@ -772,20 +772,20 @@ const Type* DivHFNode::Value(PhaseGVN* phase) const {
     return TypeH::ONE;
   }
 
-  if( t2 == TypeH::ONE )
+  if(t2 == TypeH::ONE)
     return t1;
 
   // If divisor is a constant and not zero, divide them numbers
-  if( t1->base() == Type::HalfFloatCon &&
-      t2->base() == Type::HalfFloatCon &&
-      t2->getf() != 0.0 ) // could be negative zero
-    return TypeH::make( t1->getf()/t2->getf() );
+  if(t1->base() == Type::HalfFloatCon &&
+     t2->base() == Type::HalfFloatCon &&
+     t2->getf() != 0.0) // could be negative zero
+    return TypeH::make(t1->getf()/t2->getf());
 
   // If the dividend is a constant zero
   // Note: if t1 and t2 are zero then result is NaN (JVMS page 213)
   // Test TypeF::ZERO is not sufficient as it could be negative zero
 
-  if( t1 == TypeH::ZERO && !g_isnan(t2->getf()) && t2->getf() != 0.0 )
+  if(t1 == TypeH::ZERO && !g_isnan(t2->getf()) && t2->getf() != 0.0)
     return TypeH::ZERO;
 
   // Otherwise we give up all hope
@@ -801,36 +801,36 @@ Node* DivHFNode::Identity(PhaseGVN* phase) {
 
 
 //------------------------------Idealize---------------------------------------
-Node *DivHFNode::Ideal(PhaseGVN *phase, bool can_reshape) {
+Node *DivHFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
   if (in(0) && remove_dead_region(phase, can_reshape))  return this;
   // Don't bother trying to transform a dead node
-  if( in(0) && in(0)->is_top() )  return nullptr;
+  if(in(0) && in(0)->is_top())  return nullptr;
 
-  const Type *t2 = phase->type( in(2) );
-  if( t2 == TypeH::ONE )         // Identity?
-    return nullptr;              // Skip it
+  const Type* t2 = phase->type(in(2));
+  if(t2 == TypeH::ONE)         // Identity?
+    return nullptr;            // Skip it
 
-  const TypeH *tf = t2->isa_half_float_constant();
-  if( !tf ) return nullptr;
-  if( tf->base() != Type::HalfFloatCon ) return nullptr;
+  const TypeH* tf = t2->isa_half_float_constant();
+  if(!tf) return nullptr;
+  if(tf->base() != Type::HalfFloatCon) return nullptr;
 
   // Check for out of range values
-  if( tf->is_nan() || !tf->is_finite() ) return nullptr;
+  if(tf->is_nan() || !tf->is_finite()) return nullptr;
 
   // Get the value
   float f = tf->getf();
   int exp;
 
   // Only for special case of dividing by a power of 2
-  if( frexp((double)f, &exp) != 0.5 ) return nullptr;
+  if(frexp((double)f, &exp) != 0.5) return nullptr;
 
   // Limit the range of acceptable exponents
-  if( exp < -14 || exp > 15 ) return nullptr;
+  if(exp < -14 || exp > 15) return nullptr;
 
   // Compute the reciprocal
   float reciprocal = ((float)1.0) / f;
 
-  assert( frexp((double)reciprocal, &exp) == 0.5, "reciprocal should be power of 2" );
+  assert(frexp((double)reciprocal, &exp) == 0.5, "reciprocal should be power of 2");
 
   // return multiplication by the reciprocal
   return (new MulHFNode(in(1), phase->makecon(TypeH::make(reciprocal))));

src/hotspot/share/opto/divnode.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 1997, 2023, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -82,12 +82,12 @@ class DivFNode : public Node {
 // Half float division
 class DivHFNode : public Node {
 public:
-  DivHFNode( Node *c, Node *dividend, Node *divisor ) : Node(c, dividend, divisor) {}
+  DivHFNode(Node* c, Node* dividend, Node* divisor) : Node(c, dividend, divisor) {}
   virtual int Opcode() const;
   virtual Node* Identity(PhaseGVN* phase);
-  virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
+  virtual Node* Ideal(PhaseGVN* phase, bool can_reshape);
   virtual const Type* Value(PhaseGVN* phase) const;
-  virtual const Type *bottom_type() const { return Type::HALF_FLOAT; }
+  virtual const Type* bottom_type() const { return Type::HALF_FLOAT; }
   virtual uint ideal_reg() const { return Op_RegF; }
 };