8252848: Optimize small primitive arrayCopy operations through partial inlining using AVX-512 masked instructions.

Author: Jatin Bhateja

src/hotspot/cpu/x86/assembler_x86.cpp

@@ -2589,6 +2589,38 @@ void Assembler::evmovdqub(XMMRegister dst, KRegister mask, Address src, int vect
   emit_operand(dst, src);
 }
 
+void Assembler::evmovdqu(XMMRegister dst, KRegister mask, Address src, int vector_len, int type) {
+  assert(VM_Version::supports_avx512vlbw(), "");
+  InstructionMark im(this);
+  bool wide = type == T_SHORT || type == T_LONG || type == T_CHAR;
+  bool bwinstr = type == T_BYTE ||  type == T_SHORT || type == T_CHAR;
+  InstructionAttr attributes(vector_len, /* vex_w */ wide, /* legacy_mode */ false, /* no_mask_reg */ false, /* uses_vl */ true);
+  attributes.set_address_attributes(/* tuple_type */ EVEX_FVM, /* input_size_in_bits */ EVEX_NObit);
+  attributes.set_embedded_opmask_register_specifier(mask);
+  attributes.set_is_evex_instruction();
+  int prefix = bwinstr ? VEX_SIMD_F2 : VEX_SIMD_F3;
+  vex_prefix(src, 0, dst->encoding(), (Assembler::VexSimdPrefix)prefix, VEX_OPCODE_0F, &attributes);
+  emit_int8(0x6F);
+  emit_operand(dst, src);
+}
+
+void Assembler::evmovdqu(Address dst, KRegister mask, XMMRegister src, int vector_len, int type) {
+  assert(VM_Version::supports_avx512vlbw(), "");
+  assert(src != xnoreg, "sanity");
+  InstructionMark im(this);
+  bool wide = type == T_SHORT || type == T_LONG || type == T_CHAR;
+  bool bwinstr = type == T_BYTE ||  type == T_SHORT || type == T_CHAR;
+  InstructionAttr attributes(vector_len, /* vex_w */ wide, /* legacy_mode */ false, /* no_mask_reg */ false, /* uses_vl */ true);
+  attributes.set_address_attributes(/* tuple_type */ EVEX_FVM, /* input_size_in_bits */ EVEX_NObit);
+  attributes.reset_is_clear_context();
+  attributes.set_embedded_opmask_register_specifier(mask);
+  attributes.set_is_evex_instruction();
+  int prefix = bwinstr ? VEX_SIMD_F2 : VEX_SIMD_F3;
+  vex_prefix(dst, 0, src->encoding(), (Assembler::VexSimdPrefix)prefix, VEX_OPCODE_0F, &attributes);
+  emit_int8(0x7F);
+  emit_operand(src, dst);
+}
+
 void Assembler::evmovdquw(XMMRegister dst, Address src, int vector_len) {
   assert(VM_Version::supports_evex(), "");
   InstructionMark im(this);
@@ -7803,6 +7835,13 @@ void Assembler::shlxq(Register dst, Register src1, Register src2) {
   emit_int16((unsigned char)0xF7, (0xC0 | encode));
 }
 
+void Assembler::shrxq(Register dst, Register src1, Register src2) {
+  assert(VM_Version::supports_bmi2(), "");
+  InstructionAttr attributes(AVX_128bit, /* vex_w */ true, /* legacy_mode */ true, /* no_mask_reg */ true, /* uses_vl */ true);
+  int encode = vex_prefix_and_encode(dst->encoding(), src2->encoding(), src1->encoding(), VEX_SIMD_F2, VEX_OPCODE_0F_38, &attributes);
+  emit_int16((unsigned char)0xF7, (0xC0 | encode));
+}
+
 #ifndef _LP64
 
 void Assembler::incl(Register dst) {

src/hotspot/cpu/x86/assembler_x86.hpp

@@ -826,7 +826,6 @@ class Assembler : public AbstractAssembler  {
 
   void decl(Register dst);
   void decl(Address dst);
-  void decq(Register dst);
   void decq(Address dst);
 
   void incl(Register dst);
@@ -911,6 +910,7 @@ class Assembler : public AbstractAssembler  {
   void popa_uncached();
 #endif
   void vzeroupper_uncached();
+  void decq(Register dst);
 
   void pusha();
   void popa();
@@ -1519,6 +1519,10 @@ class Assembler : public AbstractAssembler  {
   void evmovdquq(XMMRegister dst, Address src, int vector_len);
   void evmovdquq(XMMRegister dst, XMMRegister src, int vector_len);
 
+  // Generic move instructions.
+  void evmovdqu(Address dst, KRegister mask, XMMRegister src, int vector_len, int type);
+  void evmovdqu(XMMRegister dst, KRegister mask, Address src, int vector_len, int type);
+
   // Move lower 64bit to high 64bit in 128bit register
   void movlhps(XMMRegister dst, XMMRegister src);
 
@@ -2021,6 +2025,8 @@ class Assembler : public AbstractAssembler  {
 
   void shlxl(Register dst, Register src1, Register src2);
   void shlxq(Register dst, Register src1, Register src2);
+  void shrxq(Register dst, Register src1, Register src2);
+
 
   //====================VECTOR ARITHMETIC=====================================
 

src/hotspot/cpu/x86/c2_MacroAssembler_x86.cpp

@@ -1218,6 +1218,20 @@ void C2_MacroAssembler::reduce8L(int opcode, Register dst, Register src1, XMMReg
   reduce_operation_256(opcode, vtmp2, vtmp2, src2);
   reduce4L(opcode, dst, src1, vtmp2, vtmp1, vtmp2);
 }
+
+void C2_MacroAssembler::genmask(Register dst, Register len, Register temp) {
+  if (ArrayCopyPartialInlineSize <= 32) {
+    mov64(dst, 1);
+    shlxq(dst, dst, len);
+    decq(dst);
+  } else {
+    mov64(dst, -1);
+    movq(temp, len);
+    negptr(temp);
+    addptr(temp, 64);
+    shrxq(dst, dst, temp);
+  }
+}
 #endif // _LP64
 
 void C2_MacroAssembler::reduce2F(int opcode, XMMRegister dst, XMMRegister src, XMMRegister vtmp) {

src/hotspot/cpu/x86/c2_MacroAssembler_x86.hpp

@@ -93,6 +93,7 @@
   void reduceI(int opcode, int vlen, Register dst, Register src1, XMMRegister src2, XMMRegister vtmp1, XMMRegister vtmp2);
 #ifdef _LP64
   void reduceL(int opcode, int vlen, Register dst, Register src1, XMMRegister src2, XMMRegister vtmp1, XMMRegister vtmp2);
+  void genmask(Register dst, Register len, Register temp);
 #endif // _LP64
 
   // dst = reduce(op, src2) using vtmp as temps

src/hotspot/cpu/x86/vm_version_x86.cpp

@@ -761,6 +761,8 @@ void VM_Version::get_processor_features() {
   if (is_intel()) { // Intel cpus specific settings
     if (is_knights_family()) {
       _features &= ~CPU_VZEROUPPER;
+      _features &= ~CPU_AVX512BW;
+      _features &= ~CPU_AVX512VL;
     }
   }
 
@@ -1162,7 +1164,7 @@ void VM_Version::get_processor_features() {
 #endif // COMPILER2 && ASSERT
 
   if (!FLAG_IS_DEFAULT(AVX3Threshold)) {
-    if (!is_power_of_2(AVX3Threshold)) {
+    if (AVX3Threshold !=0 && !is_power_of_2(AVX3Threshold)) {
       warning("AVX3Threshold must be a power of 2");
       FLAG_SET_DEFAULT(AVX3Threshold, 4096);
     }
@@ -1411,6 +1413,29 @@ void VM_Version::get_processor_features() {
         MaxLoopPad = 11;
       }
 #endif // COMPILER2
+
+      if (FLAG_IS_DEFAULT(ArrayCopyPartialInlineSize) ||
+          (!FLAG_IS_DEFAULT(ArrayCopyPartialInlineSize) &&
+           ArrayCopyPartialInlineSize != 0 &&
+           ArrayCopyPartialInlineSize != 32 &&
+           ArrayCopyPartialInlineSize != 64)) {
+        int pi_size = 0;
+        if (MaxVectorSize > 32 && AVX3Threshold == 0) {
+          pi_size = 64;
+        } else if (MaxVectorSize >= 32) {
+          pi_size = 32;
+        }
+        if(!FLAG_IS_DEFAULT(ArrayCopyPartialInlineSize)) {
+          warning("Setting ArrayCopyPartialInlineSize as %d", pi_size);
+        }
+        ArrayCopyPartialInlineSize = pi_size;
+      }
+
+      if (ArrayCopyPartialInlineSize > MaxVectorSize) {
+        ArrayCopyPartialInlineSize = MaxVectorSize;
+        warning("Setting ArrayCopyPartialInlineSize as MaxVectorSize");
+      }
+
       if (FLAG_IS_DEFAULT(UseXMMForArrayCopy)) {
         UseXMMForArrayCopy = true; // use SSE2 movq on new Intel cpus
       }

src/hotspot/cpu/x86/x86.ad

@@ -1405,6 +1405,13 @@ const bool Matcher::match_rule_supported(int opcode) {
         return false;
       }
       break;
+    case Op_VectorMaskGen:
+    case Op_VectorMaskedLoad:
+    case Op_VectorMaskedStore:
+      if (UseAVX < 3) {
+        return false;
+      }
+      break;
 #ifndef _LP64
     case Op_AddReductionVF:
     case Op_AddReductionVD:
@@ -1477,6 +1484,16 @@ const bool Matcher::match_rule_supported_vector(int opcode, int vlen, BasicType
         return false;
       }
       break;
+    case Op_VectorMaskGen:
+    case Op_VectorMaskedLoad:
+    case Op_VectorMaskedStore:
+      if (!VM_Version::supports_avx512bw()) {
+        return false;
+      }
+      if ((size_in_bits != 512) && !VM_Version::supports_avx512vl()) {
+        return false;
+      }
+      break;
     case Op_CMoveVD:
       if (vlen != 4) {
         return false; // implementation limitation (only vcmov4D_reg is present)
@@ -5444,3 +5461,52 @@ instruct vprorate(vec dst, vec src, vec shift) %{
   ins_pipe( pipe_slow );
 %}
 
+#ifdef _LP64
+// ---------------------------------- Masked Block Copy ------------------------------------
+
+instruct vmasked_load64(vec dst, memory mem, rRegL mask) %{
+  match(Set dst (VectorMaskedLoad mem mask));
+  format %{ "vector_masked_load $dst, $mem, $mask \t! vector masked copy" %}
+  ins_encode %{
+    BasicType elmType =  this->bottom_type()->is_vect()->element_basic_type();
+    int vector_len = vector_length_encoding(this);
+    //TODO: KRegister to be made valid "bound" operand to promote sharing.
+    __ kmovql(k2, $mask$$Register);
+    __ evmovdqu($dst$$XMMRegister, k2, $mem$$Address, vector_len, elmType);
+  %}
+  ins_pipe( pipe_slow );
+%}
+
+instruct vmask_gen(rRegL dst, rRegL len, rRegL tempLen) %{
+  match(Set dst (VectorMaskGen len));
+  effect(TEMP_DEF dst, TEMP tempLen);
+  format %{ "vector_mask_gen $len \t! vector mask generator" %}
+  ins_encode %{
+    __ genmask($dst$$Register, $len$$Register, $tempLen$$Register);
+  %}
+  ins_pipe( pipe_slow );
+%}
+
+instruct vmask_gen_imm(rRegL dst, immL len) %{
+  match(Set dst (VectorMaskGen len));
+  format %{ "vector_mask_gen $len \t! vector mask generator" %}
+  ins_encode %{
+    __ mov64($dst$$Register, (1L << ($len$$constant & 63)) -1);
+  %}
+  ins_pipe( pipe_slow );
+%}
+
+instruct vmasked_store64(memory mem, vec src, rRegL mask) %{
+  match(Set mem (VectorMaskedStore mem (Binary src mask)));
+  format %{ "vector_masked_store $mem, $src, $mask \t! vector masked store" %}
+  ins_encode %{
+    const MachNode* src_node = static_cast<const MachNode*>(this->in(this->operand_index($src)));
+    BasicType elmType =  src_node->bottom_type()->is_vect()->element_basic_type();
+    int vector_len = vector_length_encoding(src_node);
+    //TODO: KRegister to be made valid "bound" operand to promote sharing.
+    __ kmovql(k2, $mask$$Register);
+    __ evmovdqu($mem$$Address, k2, $src$$XMMRegister, vector_len, elmType);
+  %}
+  ins_pipe( pipe_slow );
+%}
+#endif // _LP64

src/hotspot/share/adlc/forms.cpp

@@ -268,6 +268,7 @@ Form::DataType Form::is_load_from_memory(const char *opType) const {
   if( strcmp(opType,"LoadRange")==0 )  return Form::idealI;
   if( strcmp(opType,"LoadS")==0 )  return Form::idealS;
   if( strcmp(opType,"LoadVector")==0 )  return Form::idealV;
+  if( strcmp(opType,"VectorMaskedLoad")==0 )  return Form::idealV;
   assert( strcmp(opType,"Load") != 0, "Must type Loads" );
   return Form::none;
 }
@@ -284,6 +285,7 @@ Form::DataType Form::is_store_to_memory(const char *opType) const {
   if( strcmp(opType,"StoreN")==0)  return Form::idealN;
   if( strcmp(opType,"StoreNKlass")==0)  return Form::idealNKlass;
   if( strcmp(opType,"StoreVector")==0 )  return Form::idealV;
+  if( strcmp(opType,"VectorMaskedStore")==0 )  return Form::idealV;
   assert( strcmp(opType,"Store") != 0, "Must type Stores" );
   return Form::none;
 }

src/hotspot/share/adlc/formssel.cpp

@@ -779,6 +779,7 @@ bool InstructForm::captures_bottom_type(FormDict &globals) const {
        !strcmp(_matrule->_rChild->_opType,"ShenandoahCompareAndExchangeP") ||
        !strcmp(_matrule->_rChild->_opType,"ShenandoahCompareAndExchangeN") ||
 #endif
+       !strcmp(_matrule->_rChild->_opType,"VectorMaskGen")||
        !strcmp(_matrule->_rChild->_opType,"CompareAndExchangeP") ||
        !strcmp(_matrule->_rChild->_opType,"CompareAndExchangeN"))) return true;
   else if ( is_ideal_load() == Form::idealP )                return true;
@@ -3484,7 +3485,7 @@ int MatchNode::needs_ideal_memory_edge(FormDict &globals) const {
     "StoreB","StoreC","Store" ,"StoreFP",
     "LoadI", "LoadL", "LoadP" ,"LoadN", "LoadD" ,"LoadF"  ,
     "LoadB" , "LoadUB", "LoadUS" ,"LoadS" ,"Load" ,
-    "StoreVector", "LoadVector",
+    "StoreVector", "LoadVector", "VectorMaskedLoad", "VectorMaskedStore",
     "LoadRange", "LoadKlass", "LoadNKlass", "LoadL_unaligned", "LoadD_unaligned",
     "LoadPLocked",
     "StorePConditional", "StoreIConditional", "StoreLConditional",
@@ -4168,8 +4169,8 @@ bool MatchRule::is_vector() const {
     "RShiftVB","RShiftVS","RShiftVI","RShiftVL",
     "URShiftVB","URShiftVS","URShiftVI","URShiftVL",
     "ReplicateB","ReplicateS","ReplicateI","ReplicateL","ReplicateF","ReplicateD",
-    "RoundDoubleModeV","RotateLeftV" , "RotateRightV", "LoadVector","StoreVector",
-    "FmaVD", "FmaVF","PopCountVI",
+    "RoundDoubleModeV", "RotateLeftV" , "RotateRightV", "LoadVector","StoreVector",
+    "FmaVD", "FmaVF","PopCountVI","VectorMaskedLoad","VectorMaskedStore",
     // Next are not supported currently.
     "PackB","PackS","PackI","PackL","PackF","PackD","Pack2L","Pack2D",
     "ExtractB","ExtractUB","ExtractC","ExtractS","ExtractI","ExtractL","ExtractF","ExtractD"

src/hotspot/share/opto/arraycopynode.cpp

@@ -85,6 +85,7 @@ void ArrayCopyNode::connect_outputs(GraphKit* kit, bool deoptimize_on_exception)
   kit->set_all_memory_call(this);
 }
 
+
 #ifndef PRODUCT
 const char* ArrayCopyNode::_kind_names[] = {"arraycopy", "arraycopy, validated arguments", "clone", "oop array clone", "CopyOf", "CopyOfRange"};
 
@@ -670,13 +671,28 @@ bool ArrayCopyNode::may_modify(const TypeOopPtr *t_oop, MemBarNode* mb, PhaseTra
   CallNode* call = NULL;
   guarantee(c != NULL, "step_over_gc_barrier failed, there must be something to step to.");
   if (c->is_Region()) {
+    PhiNode* phi = NULL;
     for (uint i = 1; i < c->req(); i++) {
       if (c->in(i) != NULL) {
         Node* n = c->in(i)->in(0);
         if (may_modify_helper(t_oop, n, phase, call)) {
           ac = call->isa_ArrayCopy();
           assert(c == mb->in(0), "only for clone");
           return true;
+        } else if (n != NULL && n->is_Region() &&
+                   (phi = n->as_Region()->has_phi()) &&
+                    phi->in(1)->Opcode() == Op_VectorMaskedStore) {
+          return true;
+        } else  {
+          for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
+            Node* phi = c->fast_out(i);
+            if (phi->is_Phi()) {
+              assert(phi->in(0) == c, "phi region validation");
+              if(phi->in(1) && phi->in(1)->Opcode() == Op_VectorMaskedStore) {
+                return true;
+              }
+            }
+          }
         }
       }
     }
@@ -734,3 +750,16 @@ bool ArrayCopyNode::modifies(intptr_t offset_lo, intptr_t offset_hi, PhaseTransf
   }
   return false;
 }
+
+// As an optimization, choose optimum vector size for copy length known at compile time.
+int ArrayCopyNode::get_partial_inline_vector_lane_count(BasicType type, int con_len) {
+  int lane_count = ArrayCopyPartialInlineSize/type2aelembytes(type);
+  if (con_len > 0) {
+    int size_in_bytes = con_len * type2aelembytes(type);
+    if (size_in_bytes <= 16)
+      lane_count = 16/type2aelembytes(type);
+    else if (size_in_bytes > 16 && size_in_bytes <= 32)
+      lane_count = 32/type2aelembytes(type);
+  }
+  return lane_count;
+}

src/hotspot/share/opto/arraycopynode.hpp

@@ -180,6 +180,9 @@ class ArrayCopyNode : public CallNode {
   bool has_negative_length_guard() const { return _has_negative_length_guard; }
 
   static bool may_modify(const TypeOopPtr *t_oop, MemBarNode* mb, PhaseTransform *phase, ArrayCopyNode*& ac);
+
+  static int get_partial_inline_vector_lane_count(BasicType type, int con_len);
+
   bool modifies(intptr_t offset_lo, intptr_t offset_hi, PhaseTransform* phase, bool must_modify) const;
 
 #ifndef PRODUCT