6340864: Implement vectorization optimizations in hotspot-server

Summary: Added asm encoding and mach nodes for vector arithmetic instructions on x86.
Reviewed-by: roland

src/cpu/x86/vm/assembler_x86.hpp

@@ -617,6 +617,7 @@ private:
                    VexSimdPrefix pre, VexOpcode opc = VEX_OPCODE_0F) {
     simd_prefix(dst, xnoreg, src, pre, opc);
   }
+
   void simd_prefix(Address dst, XMMRegister src, VexSimdPrefix pre) {
     simd_prefix(src, dst, pre);
   }
@@ -626,16 +627,10 @@ private:
     simd_prefix(dst, nds, src, pre, VEX_OPCODE_0F, rex_w);
   }
 
-
   int simd_prefix_and_encode(XMMRegister dst, XMMRegister nds, XMMRegister src,
                              VexSimdPrefix pre, VexOpcode opc = VEX_OPCODE_0F,
                              bool rex_w = false, bool vector256 = false);
 
-  int simd_prefix_and_encode(XMMRegister dst, XMMRegister src,
-                             VexSimdPrefix pre, VexOpcode opc = VEX_OPCODE_0F) {
-    return simd_prefix_and_encode(dst, xnoreg, src, pre, opc);
-  }
-
   // Move/convert 32-bit integer value.
   int simd_prefix_and_encode(XMMRegister dst, XMMRegister nds, Register src,
                              VexSimdPrefix pre) {
@@ -677,6 +672,15 @@ private:
   void emit_arith(int op1, int op2, Register dst, jobject obj);
   void emit_arith(int op1, int op2, Register dst, Register src);
 
+  void emit_simd_arith(int opcode, XMMRegister dst, Address src, VexSimdPrefix pre);
+  void emit_simd_arith(int opcode, XMMRegister dst, XMMRegister src, VexSimdPrefix pre);
+  void emit_simd_arith_nonds(int opcode, XMMRegister dst, Address src, VexSimdPrefix pre);
+  void emit_simd_arith_nonds(int opcode, XMMRegister dst, XMMRegister src, VexSimdPrefix pre);
+  void emit_vex_arith(int opcode, XMMRegister dst, XMMRegister nds,
+                      Address src, VexSimdPrefix pre, bool vector256);
+  void emit_vex_arith(int opcode, XMMRegister dst, XMMRegister nds,
+                      XMMRegister src, VexSimdPrefix pre, bool vector256);
+
   void emit_operand(Register reg,
                     Register base, Register index, Address::ScaleFactor scale,
                     int disp,
@@ -891,12 +895,6 @@ private:
   void andq(Register dst, Address src);
   void andq(Register dst, Register src);
 
-  // Bitwise Logical AND of Packed Double-Precision Floating-Point Values
-  void andpd(XMMRegister dst, XMMRegister src);
-
-  // Bitwise Logical AND of Packed Single-Precision Floating-Point Values
-  void andps(XMMRegister dst, XMMRegister src);
-
   void bsfl(Register dst, Register src);
   void bsrl(Register dst, Register src);
 
@@ -1436,10 +1434,6 @@ private:
   void prefetcht2(Address src);
   void prefetchw(Address src);
 
-  // POR - Bitwise logical OR
-  void por(XMMRegister dst, XMMRegister src);
-  void por(XMMRegister dst, Address src);
-
   // Shuffle Packed Doublewords
   void pshufd(XMMRegister dst, XMMRegister src, int mode);
   void pshufd(XMMRegister dst, Address src,     int mode);
@@ -1448,9 +1442,6 @@ private:
   void pshuflw(XMMRegister dst, XMMRegister src, int mode);
   void pshuflw(XMMRegister dst, Address src,     int mode);
 
-  // Shift Right by bits Logical Quadword Immediate
-  void psrlq(XMMRegister dst, int shift);
-
   // Shift Right by bytes Logical DoubleQuadword Immediate
   void psrldq(XMMRegister dst, int shift);
 
@@ -1475,10 +1466,6 @@ private:
 
   void pushq(Address src);
 
-  // Xor Packed Byte Integer Values
-  void pxor(XMMRegister dst, Address src);
-  void pxor(XMMRegister dst, XMMRegister src);
-
   void rcll(Register dst, int imm8);
 
   void rclq(Register dst, int imm8);
@@ -1601,15 +1588,10 @@ private:
   void xorq(Register dst, Address src);
   void xorq(Register dst, Register src);
 
-  // Bitwise Logical XOR of Packed Double-Precision Floating-Point Values
-  void xorpd(XMMRegister dst, XMMRegister src);
-
-  // Bitwise Logical XOR of Packed Single-Precision Floating-Point Values
-  void xorps(XMMRegister dst, XMMRegister src);
-
   void set_byte_if_not_zero(Register dst); // sets reg to 1 if not zero, otherwise 0
 
   // AVX 3-operands scalar instructions (encoded with VEX prefix)
+
   void vaddsd(XMMRegister dst, XMMRegister nds, Address src);
   void vaddsd(XMMRegister dst, XMMRegister nds, XMMRegister src);
   void vaddss(XMMRegister dst, XMMRegister nds, Address src);
@@ -1627,14 +1609,147 @@ private:
   void vsubss(XMMRegister dst, XMMRegister nds, Address src);
   void vsubss(XMMRegister dst, XMMRegister nds, XMMRegister src);
 
-  // AVX Vector instrucitons.
-  void vandpd(XMMRegister dst, XMMRegister nds, Address src);
-  void vandps(XMMRegister dst, XMMRegister nds, Address src);
-  void vxorpd(XMMRegister dst, XMMRegister nds, Address src);
-  void vxorps(XMMRegister dst, XMMRegister nds, Address src);
+
+  //====================VECTOR ARITHMETIC=====================================
+
+  // Add Packed Floating-Point Values
+  void addpd(XMMRegister dst, XMMRegister src);
+  void addps(XMMRegister dst, XMMRegister src);
+  void vaddpd(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vaddps(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vaddpd(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+  void vaddps(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+
+  // Subtract Packed Floating-Point Values
+  void subpd(XMMRegister dst, XMMRegister src);
+  void subps(XMMRegister dst, XMMRegister src);
+  void vsubpd(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vsubps(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vsubpd(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+  void vsubps(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+
+  // Multiply Packed Floating-Point Values
+  void mulpd(XMMRegister dst, XMMRegister src);
+  void mulps(XMMRegister dst, XMMRegister src);
+  void vmulpd(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vmulps(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vmulpd(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+  void vmulps(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+
+  // Divide Packed Floating-Point Values
+  void divpd(XMMRegister dst, XMMRegister src);
+  void divps(XMMRegister dst, XMMRegister src);
+  void vdivpd(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vdivps(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vdivpd(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+  void vdivps(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+
+  // Bitwise Logical AND of Packed Floating-Point Values
+  void andpd(XMMRegister dst, XMMRegister src);
+  void andps(XMMRegister dst, XMMRegister src);
+  void vandpd(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vandps(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vandpd(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+  void vandps(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+
+  // Bitwise Logical XOR of Packed Floating-Point Values
+  void xorpd(XMMRegister dst, XMMRegister src);
+  void xorps(XMMRegister dst, XMMRegister src);
   void vxorpd(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
   void vxorps(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vxorpd(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+  void vxorps(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+
+  // Add packed integers
+  void paddb(XMMRegister dst, XMMRegister src);
+  void paddw(XMMRegister dst, XMMRegister src);
+  void paddd(XMMRegister dst, XMMRegister src);
+  void paddq(XMMRegister dst, XMMRegister src);
+  void vpaddb(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vpaddw(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vpaddd(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vpaddq(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vpaddb(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+  void vpaddw(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+  void vpaddd(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+  void vpaddq(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+
+  // Sub packed integers
+  void psubb(XMMRegister dst, XMMRegister src);
+  void psubw(XMMRegister dst, XMMRegister src);
+  void psubd(XMMRegister dst, XMMRegister src);
+  void psubq(XMMRegister dst, XMMRegister src);
+  void vpsubb(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vpsubw(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vpsubd(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vpsubq(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vpsubb(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+  void vpsubw(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+  void vpsubd(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+  void vpsubq(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+
+  // Multiply packed integers (only shorts and ints)
+  void pmullw(XMMRegister dst, XMMRegister src);
+  void pmulld(XMMRegister dst, XMMRegister src);
+  void vpmullw(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vpmulld(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vpmullw(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+  void vpmulld(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+
+  // Shift left packed integers
+  void psllw(XMMRegister dst, int shift);
+  void pslld(XMMRegister dst, int shift);
+  void psllq(XMMRegister dst, int shift);
+  void psllw(XMMRegister dst, XMMRegister shift);
+  void pslld(XMMRegister dst, XMMRegister shift);
+  void psllq(XMMRegister dst, XMMRegister shift);
+  void vpsllw(XMMRegister dst, XMMRegister src, int shift, bool vector256);
+  void vpslld(XMMRegister dst, XMMRegister src, int shift, bool vector256);
+  void vpsllq(XMMRegister dst, XMMRegister src, int shift, bool vector256);
+  void vpsllw(XMMRegister dst, XMMRegister src, XMMRegister shift, bool vector256);
+  void vpslld(XMMRegister dst, XMMRegister src, XMMRegister shift, bool vector256);
+  void vpsllq(XMMRegister dst, XMMRegister src, XMMRegister shift, bool vector256);
+
+  // Logical shift right packed integers
+  void psrlw(XMMRegister dst, int shift);
+  void psrld(XMMRegister dst, int shift);
+  void psrlq(XMMRegister dst, int shift);
+  void psrlw(XMMRegister dst, XMMRegister shift);
+  void psrld(XMMRegister dst, XMMRegister shift);
+  void psrlq(XMMRegister dst, XMMRegister shift);
+  void vpsrlw(XMMRegister dst, XMMRegister src, int shift, bool vector256);
+  void vpsrld(XMMRegister dst, XMMRegister src, int shift, bool vector256);
+  void vpsrlq(XMMRegister dst, XMMRegister src, int shift, bool vector256);
+  void vpsrlw(XMMRegister dst, XMMRegister src, XMMRegister shift, bool vector256);
+  void vpsrld(XMMRegister dst, XMMRegister src, XMMRegister shift, bool vector256);
+  void vpsrlq(XMMRegister dst, XMMRegister src, XMMRegister shift, bool vector256);
+
+  // Arithmetic shift right packed integers (only shorts and ints, no instructions for longs)
+  void psraw(XMMRegister dst, int shift);
+  void psrad(XMMRegister dst, int shift);
+  void psraw(XMMRegister dst, XMMRegister shift);
+  void psrad(XMMRegister dst, XMMRegister shift);
+  void vpsraw(XMMRegister dst, XMMRegister src, int shift, bool vector256);
+  void vpsrad(XMMRegister dst, XMMRegister src, int shift, bool vector256);
+  void vpsraw(XMMRegister dst, XMMRegister src, XMMRegister shift, bool vector256);
+  void vpsrad(XMMRegister dst, XMMRegister src, XMMRegister shift, bool vector256);
+
+  // And packed integers
+  void pand(XMMRegister dst, XMMRegister src);
+  void vpand(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vpand(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+
+  // Or packed integers
+  void por(XMMRegister dst, XMMRegister src);
+  void vpor(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vpor(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+
+  // Xor packed integers
+  void pxor(XMMRegister dst, XMMRegister src);
   void vpxor(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256);
+  void vpxor(XMMRegister dst, XMMRegister nds, Address src, bool vector256);
+
+  // Copy low 128bit into high 128bit of YMM registers.
   void vinsertf128h(XMMRegister dst, XMMRegister nds, XMMRegister src);
   void vinserti128h(XMMRegister dst, XMMRegister nds, XMMRegister src);
 
@@ -2532,11 +2647,13 @@ public:
   void vaddss(XMMRegister dst, XMMRegister nds, Address src)     { Assembler::vaddss(dst, nds, src); }
   void vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src);
 
-  void vandpd(XMMRegister dst, XMMRegister nds, Address src)     { Assembler::vandpd(dst, nds, src); }
-  void vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+  void vandpd(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256) { Assembler::vandpd(dst, nds, src, vector256); }
+  void vandpd(XMMRegister dst, XMMRegister nds, Address src, bool vector256)     { Assembler::vandpd(dst, nds, src, vector256); }
+  void vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256);
 
-  void vandps(XMMRegister dst, XMMRegister nds, Address src)     { Assembler::vandps(dst, nds, src); }
-  void vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+  void vandps(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256) { Assembler::vandps(dst, nds, src, vector256); }
+  void vandps(XMMRegister dst, XMMRegister nds, Address src, bool vector256)     { Assembler::vandps(dst, nds, src, vector256); }
+  void vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256);
 
   void vdivsd(XMMRegister dst, XMMRegister nds, XMMRegister src) { Assembler::vdivsd(dst, nds, src); }
   void vdivsd(XMMRegister dst, XMMRegister nds, Address src)     { Assembler::vdivsd(dst, nds, src); }
@@ -2565,12 +2682,12 @@ public:
   // AVX Vector instructions
 
   void vxorpd(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256) { Assembler::vxorpd(dst, nds, src, vector256); }
-  void vxorpd(XMMRegister dst, XMMRegister nds, Address src) { Assembler::vxorpd(dst, nds, src); }
-  void vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+  void vxorpd(XMMRegister dst, XMMRegister nds, Address src, bool vector256) { Assembler::vxorpd(dst, nds, src, vector256); }
+  void vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256);
 
   void vxorps(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256) { Assembler::vxorps(dst, nds, src, vector256); }
-  void vxorps(XMMRegister dst, XMMRegister nds, Address src) { Assembler::vxorps(dst, nds, src); }
-  void vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+  void vxorps(XMMRegister dst, XMMRegister nds, Address src, bool vector256) { Assembler::vxorps(dst, nds, src, vector256); }
+  void vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256);
 
   void vpxor(XMMRegister dst, XMMRegister nds, XMMRegister src, bool vector256) {
     if (UseAVX > 1 || !vector256) // vpxor 256 bit is available only in AVX2
@@ -2578,6 +2695,12 @@ public:
     else
       Assembler::vxorpd(dst, nds, src, vector256);
   }
+  void vpxor(XMMRegister dst, XMMRegister nds, Address src, bool vector256) {
+    if (UseAVX > 1 || !vector256) // vpxor 256 bit is available only in AVX2
+      Assembler::vpxor(dst, nds, src, vector256);
+    else
+      Assembler::vxorpd(dst, nds, src, vector256);
+  }
 
   // Move packed integer values from low 128 bit to hign 128 bit in 256 bit vector.
   void vinserti128h(XMMRegister dst, XMMRegister nds, XMMRegister src) {

src/cpu/x86/vm/x86_32.ad

@@ -1367,22 +1367,6 @@ int emit_deopt_handler(CodeBuffer& cbuf) {
   return offset;
 }
 
-
-const bool Matcher::match_rule_supported(int opcode) {
-  if (!has_match_rule(opcode))
-    return false;
-
-  switch (opcode) {
-    case Op_PopCountI:
-    case Op_PopCountL:
-      if (!UsePopCountInstruction)
-        return false;
-    break;
-  }
-  
-  return true;  // Per default match rules are supported.
-}
-
 int Matcher::regnum_to_fpu_offset(int regnum) {
   return regnum - 32; // The FP registers are in the second chunk
 }

src/cpu/x86/vm/x86_64.ad

@@ -1513,22 +1513,6 @@ int emit_deopt_handler(CodeBuffer& cbuf)
   return offset;
 }
 
-
-const bool Matcher::match_rule_supported(int opcode) {
-  if (!has_match_rule(opcode))
-    return false;
-
-  switch (opcode) {
-    case Op_PopCountI:
-    case Op_PopCountL:
-      if (!UsePopCountInstruction)
-        return false;
-    break;
-  }
-
-  return true;  // Per default match rules are supported.
-}
-
 int Matcher::regnum_to_fpu_offset(int regnum)
 {
   return regnum - 32; // The FP registers are in the second chunk
@@ -6427,6 +6411,31 @@ instruct castP2X(rRegL dst, rRegP src)
   ins_pipe(ialu_reg_reg); // XXX
 %}
 
+// Convert oop into int for vectors alignment masking
+instruct convP2I(rRegI dst, rRegP src)
+%{
+  match(Set dst (ConvL2I (CastP2X src)));
+
+  format %{ "movl    $dst, $src\t# ptr -> int" %}
+  ins_encode %{
+    __ movl($dst$$Register, $src$$Register);
+  %}
+  ins_pipe(ialu_reg_reg); // XXX
+%}
+
+// Convert compressed oop into int for vectors alignment masking
+// in case of 32bit oops (heap < 4Gb).
+instruct convN2I(rRegI dst, rRegN src)
+%{
+  predicate(Universe::narrow_oop_shift() == 0);
+  match(Set dst (ConvL2I (CastP2X (DecodeN src))));
+
+  format %{ "movl    $dst, $src\t# compressed ptr -> int" %}
+  ins_encode %{
+    __ movl($dst$$Register, $src$$Register);
+  %}
+  ins_pipe(ialu_reg_reg); // XXX
+%}
 
 // Convert oop pointer into compressed form
 instruct encodeHeapOop(rRegN dst, rRegP src, rFlagsReg cr) %{
@@ -10049,11 +10058,10 @@ instruct MoveD2L_reg_reg(rRegL dst, regD src) %{
   ins_pipe( pipe_slow );
 %}
 
-// The next instructions have long latency and use Int unit. Set high cost.
 instruct MoveI2F_reg_reg(regF dst, rRegI src) %{
   match(Set dst (MoveI2F src));
   effect(DEF dst, USE src);
-  ins_cost(300);
+  ins_cost(100);
   format %{ "movd    $dst,$src\t# MoveI2F" %}
   ins_encode %{
     __ movdl($dst$$XMMRegister, $src$$Register);
@@ -10064,7 +10072,7 @@ instruct MoveI2F_reg_reg(regF dst, rRegI src) %{
 instruct MoveL2D_reg_reg(regD dst, rRegL src) %{
   match(Set dst (MoveL2D src));
   effect(DEF dst, USE src);
-  ins_cost(300);
+  ins_cost(100);
   format %{ "movd    $dst,$src\t# MoveL2D" %}
   ins_encode %{
      __ movdq($dst$$XMMRegister, $src$$Register);

src/share/vm/opto/classes.hpp

@@ -256,6 +256,8 @@ macro(SubVI)
 macro(SubVL)
 macro(SubVF)
 macro(SubVD)
+macro(MulVS)
+macro(MulVI)
 macro(MulVF)
 macro(MulVD)
 macro(DivVF)
@@ -263,9 +265,15 @@ macro(DivVD)
 macro(LShiftVB)
 macro(LShiftVS)
 macro(LShiftVI)
+macro(LShiftVL)
 macro(RShiftVB)
 macro(RShiftVS)
 macro(RShiftVI)
+macro(RShiftVL)
+macro(URShiftVB)
+macro(URShiftVS)
+macro(URShiftVI)
+macro(URShiftVL)
 macro(AndV)
 macro(OrV)
 macro(XorV)

src/share/vm/opto/loopnode.cpp

@@ -1773,6 +1773,8 @@ void IdealLoopTree::dump_head( ) const {
     if (stride_con > 0) tty->print("+");
     tty->print("%d", stride_con);
 
+    tty->print(" (%d iters) ", (int)cl->profile_trip_cnt());
+
     if (cl->is_pre_loop ()) tty->print(" pre" );
     if (cl->is_main_loop()) tty->print(" main");
     if (cl->is_post_loop()) tty->print(" post");

src/share/vm/opto/superword.cpp

@@ -1357,6 +1357,12 @@ void SuperWord::output() {
         // Promote operands to vector
         Node* in1 = vector_opd(p, 1);
         Node* in2 = vector_opd(p, 2);
+        if (VectorNode::is_invariant_vector(in1) && (n->is_Add() || n->is_Mul())) {
+          // Move invariant vector input into second position to avoid register spilling.
+          Node* tmp = in1;
+          in1 = in2;
+          in2 = tmp;
+        }
         vn = VectorNode::make(_phase->C, opc, in1, in2, vlen, velt_basic_type(n));
       } else {
         ShouldNotReachHere();
@@ -1400,6 +1406,36 @@ Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
     if (opd->is_Vector() || opd->is_LoadVector()) {
       return opd; // input is matching vector
     }
+    if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
+      // No vector is needed for shift count.
+      // Vector instructions do not mask shift count, do it here.
+      Compile* C = _phase->C;
+      Node* cnt = opd;
+      juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
+      const TypeInt* t = opd->find_int_type();
+      if (t != NULL && t->is_con()) {
+        juint shift = t->get_con();
+        if (shift > mask) { // Unsigned cmp
+          cnt = ConNode::make(C, TypeInt::make(shift & mask));
+        }
+      } else {
+        if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
+          cnt = ConNode::make(C, TypeInt::make(mask));
+          _phase->_igvn.register_new_node_with_optimizer(cnt);
+          cnt = new (C, 3) AndINode(opd, cnt);
+          _phase->_igvn.register_new_node_with_optimizer(cnt);
+          _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
+        }
+        assert(opd->bottom_type()->isa_int(), "int type only");
+        // Move non constant shift count into XMM register.
+        cnt = new (_phase->C, 2) MoveI2FNode(cnt);
+      }
+      if (cnt != opd) {
+        _phase->_igvn.register_new_node_with_optimizer(cnt);
+        _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
+      }
+      return cnt;
+    }
     assert(!opd->is_StoreVector(), "such vector is not expected here");
     // Convert scalar input to vector with the same number of elements as
     // p0's vector. Use p0's type because size of operand's container in
@@ -1718,26 +1754,17 @@ void SuperWord::compute_vector_element_type() {
   for (int i = _block.length() - 1; i >= 0; i--) {
     Node* n = _block.at(i);
     // Only integer types need be examined
-    if (n->bottom_type()->isa_int()) {
+    const Type* vt = velt_type(n);
+    if (vt->basic_type() == T_INT) {
       uint start, end;
       vector_opd_range(n, &start, &end);
       const Type* vt = velt_type(n);
 
       for (uint j = start; j < end; j++) {
         Node* in  = n->in(j);
-        // Don't propagate through a type conversion
-        if (n->bottom_type() != in->bottom_type())
-          continue;
-        switch(in->Opcode()) {
-        case Op_AddI:    case Op_AddL:
-        case Op_SubI:    case Op_SubL:
-        case Op_MulI:    case Op_MulL:
-        case Op_AndI:    case Op_AndL:
-        case Op_OrI:     case Op_OrL:
-        case Op_XorI:    case Op_XorL:
-        case Op_LShiftI: case Op_LShiftL:
-        case Op_CMoveI:  case Op_CMoveL:
-          if (in_bb(in)) {
+        // Don't propagate through a memory
+        if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
+            data_size(n) < data_size(in)) {
           bool same_type = true;
           for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
             Node *use = in->fast_out(k);
@@ -1753,7 +1780,6 @@ void SuperWord::compute_vector_element_type() {
       }
     }
   }
-  }
 #ifndef PRODUCT
   if (TraceSuperWord && Verbose) {
     for (int i = 0; i < _block.length(); i++) {
@@ -1792,10 +1818,8 @@ const Type* SuperWord::container_type(Node* n) {
   }
   const Type* t = _igvn.type(n);
   if (t->basic_type() == T_INT) {
-    if (t->higher_equal(TypeInt::BOOL))  return TypeInt::BOOL;
-    if (t->higher_equal(TypeInt::BYTE))  return TypeInt::BYTE;
-    if (t->higher_equal(TypeInt::CHAR))  return TypeInt::CHAR;
-    if (t->higher_equal(TypeInt::SHORT)) return TypeInt::SHORT;
+    // A narrow type of arithmetic operations will be determined by
+    // propagating the type of memory operations.
     return TypeInt::INT;
   }
   return t;
@@ -1940,7 +1964,7 @@ void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
   //     lim0 == original pre loop limit
   //     V == v_align (power of 2)
   //     invar == extra invariant piece of the address expression
-  //     e == k [ +/- invar ]
+  //     e == offset [ +/- invar ]
   //
   // When reassociating expressions involving '%' the basic rules are:
   //     (a - b) % k == 0   =>  a % k == b % k
@@ -1993,13 +2017,12 @@ void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
   int elt_size = align_to_ref_p.memory_size();
   int v_align  = vw / elt_size;
   assert(v_align > 1, "sanity");
-  int k        = align_to_ref_p.offset_in_bytes() / elt_size;
-
-  Node *kn   = _igvn.intcon(k);
+  int offset   = align_to_ref_p.offset_in_bytes() / elt_size;
+  Node *offsn  = _igvn.intcon(offset);
 
-  Node *e = kn;
+  Node *e = offsn;
   if (align_to_ref_p.invar() != NULL) {
-    // incorporate any extra invariant piece producing k +/- invar >>> log2(elt)
+    // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
     Node* aref     = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt);
     _phase->_igvn.register_new_node_with_optimizer(aref);
@@ -2014,15 +2037,15 @@ void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
   }
   if (vw > ObjectAlignmentInBytes) {
     // incorporate base e +/- base && Mask >>> log2(elt)
-    Node* mask = _igvn.MakeConX(~(-1 << exact_log2(vw)));
     Node* xbase = new(_phase->C, 2) CastP2XNode(NULL, align_to_ref_p.base());
     _phase->_igvn.register_new_node_with_optimizer(xbase);
-    Node* masked_xbase  = new (_phase->C, 3) AndXNode(xbase, mask);
-    _phase->_igvn.register_new_node_with_optimizer(masked_xbase);
 #ifdef _LP64
-    masked_xbase  = new (_phase->C, 2) ConvL2INode(masked_xbase);
-    _phase->_igvn.register_new_node_with_optimizer(masked_xbase);
+    xbase  = new (_phase->C, 2) ConvL2INode(xbase);
+    _phase->_igvn.register_new_node_with_optimizer(xbase);
 #endif
+    Node* mask = _igvn.intcon(vw-1);
+    Node* masked_xbase  = new (_phase->C, 3) AndINode(xbase, mask);
+    _phase->_igvn.register_new_node_with_optimizer(masked_xbase);
     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
     Node* bref     = new (_phase->C, 3) URShiftINode(masked_xbase, log2_elt);
     _phase->_igvn.register_new_node_with_optimizer(bref);

src/share/vm/opto/vectornode.cpp

@@ -69,6 +69,15 @@ int VectorNode::opcode(int sopc, uint vlen, BasicType bt) {
   case Op_SubD:
     assert(bt == T_DOUBLE, "must be");
     return Op_SubVD;
+  case Op_MulI:
+    switch (bt) {
+    case T_BOOLEAN:
+    case T_BYTE:   return 0;   // Unimplemented
+    case T_CHAR:
+    case T_SHORT:  return Op_MulVS;
+    case T_INT:    return Matcher::match_rule_supported(Op_MulVI) ? Op_MulVI : 0; // SSE4_1
+    }
+    ShouldNotReachHere();
   case Op_MulF:
     assert(bt == T_FLOAT, "must be");
     return Op_MulVF;
@@ -90,6 +99,9 @@ int VectorNode::opcode(int sopc, uint vlen, BasicType bt) {
     case T_INT:    return Op_LShiftVI;
     }
     ShouldNotReachHere();
+  case Op_LShiftL:
+    assert(bt == T_LONG, "must be");
+    return Op_LShiftVL;
   case Op_RShiftI:
     switch (bt) {
     case T_BOOLEAN:
@@ -99,6 +111,21 @@ int VectorNode::opcode(int sopc, uint vlen, BasicType bt) {
     case T_INT:    return Op_RShiftVI;
     }
     ShouldNotReachHere();
+  case Op_RShiftL:
+    assert(bt == T_LONG, "must be");
+    return Op_RShiftVL;
+  case Op_URShiftI:
+    switch (bt) {
+    case T_BOOLEAN:
+    case T_BYTE:   return Op_URShiftVB;
+    case T_CHAR:
+    case T_SHORT:  return Op_URShiftVS;
+    case T_INT:    return Op_URShiftVI;
+    }
+    ShouldNotReachHere();
+  case Op_URShiftL:
+    assert(bt == T_LONG, "must be");
+    return Op_URShiftVL;
   case Op_AndI:
   case Op_AndL:
     return Op_AndV;
@@ -140,6 +167,34 @@ bool VectorNode::implemented(int opc, uint vlen, BasicType bt) {
   return false;
 }
 
+bool VectorNode::is_shift(Node* n) {
+  switch (n->Opcode()) {
+  case Op_LShiftI:
+  case Op_LShiftL:
+  case Op_RShiftI:
+  case Op_RShiftL:
+  case Op_URShiftI:
+  case Op_URShiftL:
+    return true;
+  }
+  return false;
+}
+
+// Check if input is loop invarient vector.
+bool VectorNode::is_invariant_vector(Node* n) {
+  // Only Replicate vector nodes are loop invarient for now.
+  switch (n->Opcode()) {
+  case Op_ReplicateB:
+  case Op_ReplicateS:
+  case Op_ReplicateI:
+  case Op_ReplicateL:
+  case Op_ReplicateF:
+  case Op_ReplicateD:
+    return true;
+  }
+  return false;
+}
+
 // Return the vector version of a scalar operation node.
 VectorNode* VectorNode::make(Compile* C, int opc, Node* n1, Node* n2, uint vlen, BasicType bt) {
   const TypeVect* vt = TypeVect::make(bt, vlen);
@@ -160,6 +215,8 @@ VectorNode* VectorNode::make(Compile* C, int opc, Node* n1, Node* n2, uint vlen,
   case Op_SubVF: return new (C, 3) SubVFNode(n1, n2, vt);
   case Op_SubVD: return new (C, 3) SubVDNode(n1, n2, vt);
 
+  case Op_MulVS: return new (C, 3) MulVSNode(n1, n2, vt);
+  case Op_MulVI: return new (C, 3) MulVINode(n1, n2, vt);
   case Op_MulVF: return new (C, 3) MulVFNode(n1, n2, vt);
   case Op_MulVD: return new (C, 3) MulVDNode(n1, n2, vt);
 
@@ -169,10 +226,17 @@ VectorNode* VectorNode::make(Compile* C, int opc, Node* n1, Node* n2, uint vlen,
   case Op_LShiftVB: return new (C, 3) LShiftVBNode(n1, n2, vt);
   case Op_LShiftVS: return new (C, 3) LShiftVSNode(n1, n2, vt);
   case Op_LShiftVI: return new (C, 3) LShiftVINode(n1, n2, vt);
+  case Op_LShiftVL: return new (C, 3) LShiftVLNode(n1, n2, vt);
 
   case Op_RShiftVB: return new (C, 3) RShiftVBNode(n1, n2, vt);
   case Op_RShiftVS: return new (C, 3) RShiftVSNode(n1, n2, vt);
   case Op_RShiftVI: return new (C, 3) RShiftVINode(n1, n2, vt);
+  case Op_RShiftVL: return new (C, 3) RShiftVLNode(n1, n2, vt);
+
+  case Op_URShiftVB: return new (C, 3) URShiftVBNode(n1, n2, vt);
+  case Op_URShiftVS: return new (C, 3) URShiftVSNode(n1, n2, vt);
+  case Op_URShiftVI: return new (C, 3) URShiftVINode(n1, n2, vt);
+  case Op_URShiftVL: return new (C, 3) URShiftVLNode(n1, n2, vt);
 
   case Op_AndV: return new (C, 3) AndVNode(n1, n2, vt);
   case Op_OrV:  return new (C, 3) OrVNode (n1, n2, vt);

src/share/vm/opto/vectornode.hpp

@@ -46,6 +46,7 @@ class VectorNode : public TypeNode {
 
   const TypeVect* vect_type() const { return type()->is_vect(); }
   uint length() const { return vect_type()->length(); } // Vector length
+  uint length_in_bytes() const { return vect_type()->length_in_bytes(); }
 
   virtual int Opcode() const;
 
@@ -57,7 +58,8 @@ class VectorNode : public TypeNode {
 
   static int  opcode(int opc, uint vlen, BasicType bt);
   static bool implemented(int opc, uint vlen, BasicType bt);
-
+  static bool is_shift(Node* n);
+  static bool is_invariant_vector(Node* n);
 };
 
 //===========================Vector=ALU=Operations====================================
@@ -158,6 +160,22 @@ class SubVDNode : public VectorNode {
   virtual int Opcode() const;
 };
 
+//------------------------------MulVSNode---------------------------------------
+// Vector multiply short
+class MulVSNode : public VectorNode {
+ public:
+  MulVSNode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
+  virtual int Opcode() const;
+};
+
+//------------------------------MulVINode---------------------------------------
+// Vector multiply int
+class MulVINode : public VectorNode {
+ public:
+  MulVINode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
+  virtual int Opcode() const;
+};
+
 //------------------------------MulVFNode---------------------------------------
 // Vector multiply float
 class MulVFNode : public VectorNode {
@@ -191,7 +209,7 @@ class DivVDNode : public VectorNode {
 };
 
 //------------------------------LShiftVBNode---------------------------------------
-// Vector lshift byte
+// Vector left shift bytes
 class LShiftVBNode : public VectorNode {
  public:
   LShiftVBNode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
@@ -199,7 +217,7 @@ class LShiftVBNode : public VectorNode {
 };
 
 //------------------------------LShiftVSNode---------------------------------------
-// Vector lshift shorts
+// Vector left shift shorts
 class LShiftVSNode : public VectorNode {
  public:
   LShiftVSNode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
@@ -207,39 +225,88 @@ class LShiftVSNode : public VectorNode {
 };
 
 //------------------------------LShiftVINode---------------------------------------
-// Vector lshift ints
+// Vector left shift ints
 class LShiftVINode : public VectorNode {
  public:
   LShiftVINode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
   virtual int Opcode() const;
 };
 
-//------------------------------URShiftVBNode---------------------------------------
-// Vector urshift bytes
+//------------------------------LShiftVLNode---------------------------------------
+// Vector left shift longs
+class LShiftVLNode : public VectorNode {
+ public:
+  LShiftVLNode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
+  virtual int Opcode() const;
+};
+
+//------------------------------RShiftVBNode---------------------------------------
+// Vector right arithmetic (signed) shift bytes
 class RShiftVBNode : public VectorNode {
  public:
   RShiftVBNode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
   virtual int Opcode() const;
 };
 
-//------------------------------URShiftVSNode---------------------------------------
-// Vector urshift shorts
+//------------------------------RShiftVSNode---------------------------------------
+// Vector right arithmetic (signed) shift shorts
 class RShiftVSNode : public VectorNode {
  public:
   RShiftVSNode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
   virtual int Opcode() const;
 };
 
-//------------------------------URShiftVINode---------------------------------------
-// Vector urshift ints
+//------------------------------RShiftVINode---------------------------------------
+// Vector right arithmetic (signed) shift ints
 class RShiftVINode : public VectorNode {
  public:
   RShiftVINode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
   virtual int Opcode() const;
 };
 
+//------------------------------RShiftVLNode---------------------------------------
+// Vector right arithmetic (signed) shift longs
+class RShiftVLNode : public VectorNode {
+ public:
+  RShiftVLNode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
+  virtual int Opcode() const;
+};
+
+//------------------------------URShiftVBNode---------------------------------------
+// Vector right logical (unsigned) shift bytes
+class URShiftVBNode : public VectorNode {
+ public:
+  URShiftVBNode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
+  virtual int Opcode() const;
+};
+
+//------------------------------URShiftVSNode---------------------------------------
+// Vector right logical (unsigned) shift shorts
+class URShiftVSNode : public VectorNode {
+ public:
+  URShiftVSNode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
+  virtual int Opcode() const;
+};
+
+//------------------------------URShiftVINode---------------------------------------
+// Vector right logical (unsigned) shift ints
+class URShiftVINode : public VectorNode {
+ public:
+  URShiftVINode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
+  virtual int Opcode() const;
+};
+
+//------------------------------URShiftVLNode---------------------------------------
+// Vector right logical (unsigned) shift longs
+class URShiftVLNode : public VectorNode {
+ public:
+  URShiftVLNode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
+  virtual int Opcode() const;
+};
+
+
 //------------------------------AndVNode---------------------------------------
-// Vector and
+// Vector and integer
 class AndVNode : public VectorNode {
  public:
   AndVNode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
@@ -247,7 +314,7 @@ class AndVNode : public VectorNode {
 };
 
 //------------------------------OrVNode---------------------------------------
-// Vector or
+// Vector or integer
 class OrVNode : public VectorNode {
  public:
   OrVNode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}
@@ -255,7 +322,7 @@ class OrVNode : public VectorNode {
 };
 
 //------------------------------XorVNode---------------------------------------
-// Vector xor
+// Vector xor integer
 class XorVNode : public VectorNode {
  public:
   XorVNode(Node* in1, Node* in2, const TypeVect* vt) : VectorNode(in1,in2,vt) {}