src/hotspot/cpu/x86/x86.ad

*** 1448,1457 ****
--- 1448,1464 ----
        break;
      case Op_MulAddVS2VI:
        if (UseSSE < 2)
          ret_value = false;
        break;
+     case Op_MaxD:
+     case Op_MaxF:
+     case Op_MinD:
+     case Op_MinF:
+       if (UseAVX < 1) // enabled for AVX only
+         ret_value = false;
+       break;
    }
  
    return ret_value;  // Per default match rules are supported.
  }
  
*** 2839,2848 ****
--- 2846,2961 ----
      __ fmaf($c$$XMMRegister, $a$$XMMRegister, $b$$XMMRegister, $c$$XMMRegister);
    %}
    ins_pipe( pipe_slow );
  %}
  
+ // Following pseudo code describes the algorithm for max[FD]/min[FD]: 
+ //  if ( b < 0 )
+ //    swap(a, b)
+ //  Tmp  = Max_Float( a , b)
+ //  Mask = a == NaN ? 1 : 0
+ //  Res  = Mask ? a : Tmp 
+ instruct maxF_reg(legRegF dst, legRegF a, legRegF b, legRegF tmp, legRegF mask) %{
+   predicate(UseAVX > 0);
+   match(Set dst (MaxF a b));
+   effect(USE a, USE b, TEMP tmp, TEMP mask);
+   format %{ 
+      "blendvps         $tmp,$b,$a,$b   \n\t"
+      "blendvps         $a,$a,$b,$b     \n\t"
+      "movaps           $b,$tmp         \n\t"
+      "vmaxps           $tmp,$a,$b      \n\t"
+      "cmpps.unordered  $mask, $a, $a   \n\t"
+      "blendvps         $dst,$tmp,$a,$mask  \n\t"
+   %}
+   ins_encode %{
+     int vector_len = 0;
+     __ blendvps($tmp$$XMMRegister, $b$$XMMRegister, $a$$XMMRegister, $b$$XMMRegister, vector_len);
+     __ blendvps($a$$XMMRegister, $a$$XMMRegister, $b$$XMMRegister, $b$$XMMRegister, vector_len);
+     __ movflt($b$$XMMRegister , $tmp$$XMMRegister);
+     __ vmaxps($tmp$$XMMRegister, $a$$XMMRegister , $b$$XMMRegister);
+     __ cmpps($mask$$XMMRegister, $a$$XMMRegister, $a$$XMMRegister, 0x3, vector_len);
+     __ blendvps($dst$$XMMRegister, $tmp$$XMMRegister, $a$$XMMRegister, $mask$$XMMRegister, vector_len);
+   %}
+   ins_pipe( pipe_slow );
+ %}
+ 
+ // max = java.lang.Max(double a , double b)
+ instruct maxD_reg(legRegD dst, legRegD a, legRegD b, legRegD tmp, legRegD mask) %{
+   predicate(UseAVX > 0);
+   match(Set dst (MaxD a b));
+   effect(USE a, USE b, TEMP tmp, TEMP mask);
+   format %{ 
+      "blendvpd         $tmp,$b,$a,$b   \n\t"
+      "blendvpd         $a,$a,$b,$b     \n\t"
+      "movapd           $b,$tmp         \n\t"
+      "vmaxpd           $tmp,$a,$b      \n\t"
+      "cmppd.unordered  $mask, $a, $a   \n\t" 
+      "blendvpd         $dst,$tmp,$a,$mask  \n\t"
+   %}
+   ins_encode %{
+     int vector_len = 0;
+     __ blendvpd($tmp$$XMMRegister, $b$$XMMRegister, $a$$XMMRegister, $b$$XMMRegister, vector_len);
+     __ blendvpd($a$$XMMRegister, $a$$XMMRegister, $b$$XMMRegister, $b$$XMMRegister, vector_len);
+     __ movdbl($b$$XMMRegister , $tmp$$XMMRegister);
+     __ vmaxpd($tmp$$XMMRegister, $a$$XMMRegister , $b$$XMMRegister);
+     __ cmppd($mask$$XMMRegister, $a$$XMMRegister, $a$$XMMRegister, 0x3, vector_len);
+     __ blendvpd($dst$$XMMRegister, $tmp$$XMMRegister, $a$$XMMRegister, $mask$$XMMRegister, vector_len);
+   %}
+   ins_pipe( pipe_slow );
+ %}
+ 
+ 
+ // min = java.lang.Min(float a , float b)
+ instruct minF_reg(legRegF dst, legRegF a, legRegF b, legRegF tmp, legRegF mask) %{
+   predicate(UseAVX > 0);
+   match(Set dst (MinF a b));
+   effect(USE a, USE b, TEMP tmp, TEMP mask, DEF dst);
+   format %{ 
+      "blendvps         $tmp,$a,$b,$a   \n\t"
+      "blendvps         $b,$b,$a,$a     \n\t"
+      "movaps           $a,$tmp         \n\t"
+      "vminps           $tmp,$a,$b      \n\t"
+      "cmpps.unordered  $mask, $a, $a   \n\t" 
+      "blendvps         $dst,$tmp,$a,$mask  \n\t"
+   %}
+   ins_encode %{
+     int vector_len = 0;
+     __ blendvps($tmp$$XMMRegister, $a$$XMMRegister, $b$$XMMRegister, $a$$XMMRegister, vector_len);
+     __ blendvps($b$$XMMRegister, $b$$XMMRegister, $a$$XMMRegister, $a$$XMMRegister, vector_len);
+     __ movflt($a$$XMMRegister , $tmp$$XMMRegister);
+     __ vminps($tmp$$XMMRegister, $a$$XMMRegister , $b$$XMMRegister);
+     __ cmpps($mask$$XMMRegister, $a$$XMMRegister, $a$$XMMRegister, 0x3, vector_len);
+     __ blendvps($dst$$XMMRegister, $tmp$$XMMRegister, $a$$XMMRegister, $mask$$XMMRegister, vector_len);
+   %}
+   ins_pipe( pipe_slow );
+ %}
+ 
+ // min = java.lang.Min(double a , double b)
+ instruct minD_reg(legRegD dst, legRegD a, legRegD b, legRegD tmp, legRegD mask) %{
+   predicate(UseAVX > 0);
+   match(Set dst (MinD a b));
+   effect(USE a, USE b, TEMP tmp, TEMP mask, DEF dst);
+   format %{ 
+      "blendvpd         $tmp,$a,$b,$a   \n\t"
+      "blendvpd         $b,$b,$a,$a     \n\t"
+      "movapd           $a,$tmp         \n\t"
+      "vminpd           $tmp,$a,$b      \n\t"
+      "cmppd.unordered  $mask, $a, $a   \n\t" 
+      "blendvpd         $dst,$tmp,$a,$mask  \n\t"
+   %}
+   ins_encode %{
+     int vector_len = 0;
+     __ blendvpd($tmp$$XMMRegister, $a$$XMMRegister, $b$$XMMRegister, $a$$XMMRegister, vector_len);
+     __ blendvpd($b$$XMMRegister, $b$$XMMRegister, $a$$XMMRegister, $a$$XMMRegister, vector_len);
+     __ movdbl($a$$XMMRegister , $tmp$$XMMRegister);
+     __ vminpd($tmp$$XMMRegister, $a$$XMMRegister , $b$$XMMRegister);
+     __ cmppd($mask$$XMMRegister, $a$$XMMRegister, $a$$XMMRegister, 0x3, vector_len);
+     __ blendvpd($dst$$XMMRegister, $tmp$$XMMRegister, $a$$XMMRegister, $mask$$XMMRegister, vector_len);
+   %}
+   ins_pipe( pipe_slow );
+ %}
+ 
  // ====================VECTOR INSTRUCTIONS=====================================
  
  
  // Load vectors (4 bytes long)
  instruct loadV4(vecS dst, memory mem) %{