8264104: Eliminate unnecessary vector mask conversion during VectorUn…

…box for floating point VectorMask

Reviewed-by: kvn, vlivanov

src/hotspot/cpu/aarch64/aarch64_neon.ad

@@ -3334,6 +3334,36 @@ instruct storemask2L(vecD dst, vecX src, immI_8 size)
   ins_pipe(pipe_slow);
 %}
 
+// vector mask cast
+
+instruct vmaskcastD(vecD dst)
+%{
+  predicate(n->bottom_type()->is_vect()->length_in_bytes() == 8 &&
+            n->in(1)->bottom_type()->is_vect()->length_in_bytes() == 8 &&
+            n->bottom_type()->is_vect()->length() == n->in(1)->bottom_type()->is_vect()->length());
+  match(Set dst (VectorMaskCast dst));
+  ins_cost(0);
+  format %{ "vmaskcast $dst\t# empty" %}
+  ins_encode %{
+    // empty
+  %}
+  ins_pipe(pipe_class_empty);
+%}
+
+instruct vmaskcastX(vecX dst)
+%{
+  predicate(n->bottom_type()->is_vect()->length_in_bytes() == 16 &&
+            n->in(1)->bottom_type()->is_vect()->length_in_bytes() == 16 &&
+            n->bottom_type()->is_vect()->length() == n->in(1)->bottom_type()->is_vect()->length());
+  match(Set dst (VectorMaskCast dst));
+  ins_cost(0);
+  format %{ "vmaskcast $dst\t# empty" %}
+  ins_encode %{
+    // empty
+  %}
+  ins_pipe(pipe_class_empty);
+%}
+
 //-------------------------------- LOAD_IOTA_INDICES----------------------------------
 
 instruct loadcon8B(vecD dst, immI0 src)

src/hotspot/cpu/aarch64/aarch64_neon_ad.m4

@@ -1258,6 +1258,27 @@ instruct storemask2L(vecD dst, vecX src, immI_8 size)
   ins_pipe(pipe_slow);
 %}
 
+// vector mask cast
+dnl
+define(`VECTOR_MASK_CAST', `
+instruct vmaskcast$1`'(vec$1 dst)
+%{
+  predicate(n->bottom_type()->is_vect()->length_in_bytes() == $2 &&
+            n->in(1)->bottom_type()->is_vect()->length_in_bytes() == $2 &&
+            n->bottom_type()->is_vect()->length() == n->in(1)->bottom_type()->is_vect()->length());
+  match(Set dst (VectorMaskCast dst));
+  ins_cost(0);
+  format %{ "vmaskcast $dst\t# empty" %}
+  ins_encode %{
+    // empty
+  %}
+  ins_pipe(pipe_class_empty);
+%}')dnl
+dnl              $1 $2
+VECTOR_MASK_CAST(D, 8)
+VECTOR_MASK_CAST(X, 16)
+dnl
+
 //-------------------------------- LOAD_IOTA_INDICES----------------------------------
 dnl
 define(`PREDICATE', `ifelse($1, 8,

src/hotspot/cpu/aarch64/aarch64_sve.ad

@@ -1760,3 +1760,18 @@ instruct vsubD(vReg dst, vReg src1, vReg src2) %{
   %}
   ins_pipe(pipe_slow);
 %}
+
+// vector mask cast
+
+instruct vmaskcast(vReg dst) %{
+  predicate(UseSVE > 0 && n->bottom_type()->is_vect()->length() == n->in(1)->bottom_type()->is_vect()->length() &&
+            n->bottom_type()->is_vect()->length_in_bytes() == n->in(1)->bottom_type()->is_vect()->length_in_bytes());
+  match(Set dst (VectorMaskCast dst));
+  ins_cost(0);
+  format %{ "vmaskcast $dst\t# empty (sve)" %}
+  ins_encode %{
+    // empty
+  %}
+  ins_pipe(pipe_class_empty);
+%}
+

src/hotspot/cpu/aarch64/aarch64_sve_ad.m4

@@ -904,3 +904,18 @@ BINARY_OP_UNPREDICATED(vsubI, SubVI, S, 4, sve_sub)
 BINARY_OP_UNPREDICATED(vsubL, SubVL, D, 2, sve_sub)
 BINARY_OP_UNPREDICATED(vsubF, SubVF, S, 4, sve_fsub)
 BINARY_OP_UNPREDICATED(vsubD, SubVD, D, 2, sve_fsub)
+
+// vector mask cast
+
+instruct vmaskcast(vReg dst) %{
+  predicate(UseSVE > 0 && n->bottom_type()->is_vect()->length() == n->in(1)->bottom_type()->is_vect()->length() &&
+            n->bottom_type()->is_vect()->length_in_bytes() == n->in(1)->bottom_type()->is_vect()->length_in_bytes());
+  match(Set dst (VectorMaskCast dst));
+  ins_cost(0);
+  format %{ "vmaskcast $dst\t# empty (sve)" %}
+  ins_encode %{
+    // empty
+  %}
+  ins_pipe(pipe_class_empty);
+%}
+

src/hotspot/share/adlc/formssel.cpp

@@ -4200,7 +4200,8 @@ bool MatchRule::is_vector() const {
     "FmaVD", "FmaVF","PopCountVI",
     // Next are not supported currently.
     "PackB","PackS","PackI","PackL","PackF","PackD","Pack2L","Pack2D",
-    "ExtractB","ExtractUB","ExtractC","ExtractS","ExtractI","ExtractL","ExtractF","ExtractD"
+    "ExtractB","ExtractUB","ExtractC","ExtractS","ExtractI","ExtractL","ExtractF","ExtractD",
+    "VectorMaskCast"
   };
   int cnt = sizeof(vector_list)/sizeof(char*);
   if (_rChild) {

src/hotspot/share/opto/classes.hpp

@@ -446,6 +446,7 @@ macro(VectorBoxAllocate)
 macro(VectorUnbox)
 macro(VectorMaskWrapper)
 macro(VectorMaskCmp)
+macro(VectorMaskCast)
 macro(VectorTest)
 macro(VectorBlend)
 macro(VectorRearrange)

src/hotspot/share/opto/vectornode.cpp

@@ -1232,6 +1232,13 @@ Node* VectorUnboxNode::Ideal(PhaseGVN* phase, bool can_reshape) {
         bool is_vector_mask    = vbox_klass->is_subclass_of(ciEnv::current()->vector_VectorMask_klass());
         bool is_vector_shuffle = vbox_klass->is_subclass_of(ciEnv::current()->vector_VectorShuffle_klass());
         if (is_vector_mask) {
+          if (in_vt->length_in_bytes() == out_vt->length_in_bytes() &&
+              Matcher::match_rule_supported_vector(Op_VectorMaskCast, out_vt->length(), out_vt->element_basic_type())) {
+            // Apply "VectorUnbox (VectorBox vmask) ==> VectorMaskCast (vmask)"
+            // directly. This could avoid the transformation ordering issue from
+            // "VectorStoreMask (VectorLoadMask vmask) => vmask".
+            return new VectorMaskCastNode(value, out_vt);
+          }
           // VectorUnbox (VectorBox vmask) ==> VectorLoadMask (VectorStoreMask vmask)
           value = phase->transform(VectorStoreMaskNode::make(*phase, value, in_vt->element_basic_type(), in_vt->length()));
           return new VectorLoadMaskNode(value, out_vt);

src/hotspot/share/opto/vectornode.hpp

@@ -1240,6 +1240,17 @@ class VectorStoreMaskNode : public VectorNode {
   static VectorStoreMaskNode* make(PhaseGVN& gvn, Node* in, BasicType in_type, uint num_elem);
 };
 
+class VectorMaskCastNode : public VectorNode {
+ public:
+  VectorMaskCastNode(Node* in, const TypeVect* vt) : VectorNode(in, vt) {
+    const TypeVect* in_vt = in->bottom_type()->is_vect();
+    assert(in_vt->length() == vt->length(), "vector length must match");
+    assert(type2aelembytes(in_vt->element_basic_type()) == type2aelembytes(vt->element_basic_type()), "element size must match");
+  }
+
+  virtual int Opcode() const;
+};
+
 // This is intended for use as a simple reinterpret node that has no cast.
 class VectorReinterpretNode : public VectorNode {
  private:

src/hotspot/share/runtime/vmStructs.cpp

@@ -1878,6 +1878,7 @@ typedef HashtableEntry<InstanceKlass*, mtClass>  KlassHashtableEntry;
   declare_c2_type(VectorInsertNode, VectorNode)                           \
   declare_c2_type(VectorUnboxNode, VectorNode)                            \
   declare_c2_type(VectorReinterpretNode, VectorNode)                      \
+  declare_c2_type(VectorMaskCastNode, VectorNode)                         \
   declare_c2_type(VectorBoxNode, Node)                                    \
   declare_c2_type(VectorBoxAllocateNode, CallStaticJavaNode)              \
   declare_c2_type(VectorTestNode, Node)                                   \