src/cpu/x86/vm/macroAssembler_x86.cpp

*** 9427,9436 ****
--- 9427,9437 ----
  }
  
  void MacroAssembler::vectorized_mismatch(Register obja, Register objb, Register length, Register log2_array_indxscale,
    Register result, Register tmp1, Register tmp2, XMMRegister rymm0, XMMRegister rymm1, XMMRegister rymm2){
    assert(UseSSE42Intrinsics, "SSE4.2 must be enabled.");
+   Label VECTOR64_LOOP, VECTOR64_TAIL, VECTOR64_NOT_EQUAL, VECTOR32_TAIL;
    Label VECTOR32_LOOP, VECTOR16_LOOP, VECTOR8_LOOP, VECTOR4_LOOP;
    Label VECTOR16_TAIL, VECTOR8_TAIL, VECTOR4_TAIL;
    Label VECTOR32_NOT_EQUAL, VECTOR16_NOT_EQUAL, VECTOR8_NOT_EQUAL, VECTOR4_NOT_EQUAL;
    Label SAME_TILL_END, DONE;
    Label BYTES_LOOP, BYTES_TAIL, BYTES_NOT_EQUAL;
*** 9439,9453 ****
    ShortBranchVerifier sbv(this);
  
    shlq(length);
    xorq(result, result);
  
    cmpq(length, 8);
    jcc(Assembler::equal, VECTOR8_LOOP);
    jcc(Assembler::less, VECTOR4_TAIL);
  
!   if (UseAVX >= 2){
  
      cmpq(length, 16);
      jcc(Assembler::equal, VECTOR16_LOOP);
      jcc(Assembler::less, VECTOR8_LOOP);
  
--- 9440,9505 ----
    ShortBranchVerifier sbv(this);
  
    shlq(length);
    xorq(result, result);
  
+   if ((UseAVX > 2) &&
+       VM_Version::supports_avx512vlbw()) {
+     set_programmed_mask_reg();  // opening of the stub context for programming mask registers
+     cmpq(length, 64);
+     jcc(Assembler::less, VECTOR32_TAIL);
+     movq(tmp1, length);
+     andq(tmp1, 0x3F);      // tail count
+     andq(length, ~(0x3F)); //vector count
+ 
+     bind(VECTOR64_LOOP);
+     // AVX512 code to compare 64 byte vectors.
+     evmovdqub(rymm0, Address(obja, result), Assembler::AVX_512bit);
+     evpcmpeqb(k7, rymm0, Address(objb, result), Assembler::AVX_512bit);
+     kortestql(k7, k7);
+     jcc(Assembler::aboveEqual, VECTOR64_NOT_EQUAL);     // mismatch
+     addq(result, 64);
+     subq(length, 64);
+     jccb(Assembler::notZero, VECTOR64_LOOP);
+ 
+     //bind(VECTOR64_TAIL);
+     testq(tmp1, tmp1);
+     jcc(Assembler::zero, SAME_TILL_END);
+ 
+     bind(VECTOR64_TAIL);
+     // AVX512 code to compare upto 63 byte vectors.
+     // Save k1
+     kmovql(k3, k1);
+     mov64(tmp2, 0xFFFFFFFFFFFFFFFF);
+     shlxq(tmp2, tmp2, tmp1);
+     notq(tmp2);
+     kmovql(k1, tmp2);
+ 
+     evmovdqub(k1, false, rymm0, Address(obja, result), Assembler::AVX_512bit);
+     evpcmpeqb(k1, false, k7, rymm0, Address(objb, result), Assembler::AVX_512bit);
+ 
+     ktestql(k7, k1);
+     // Restore k1
+     kmovql(k1, k3);
+     jcc(Assembler::below, SAME_TILL_END);     // not mismatch
+ 
+     bind(VECTOR64_NOT_EQUAL);
+     kmovql(tmp1, k7);
+     notq(tmp1);
+     tzcntq(tmp1, tmp1);
+     addq(result, tmp1);
+     shrq(result);
+     jmp(DONE);
+     bind(VECTOR32_TAIL);
+     clear_programmed_mask_reg();   // closing of the stub context for programming mask registers
+   }
+ 
    cmpq(length, 8);
    jcc(Assembler::equal, VECTOR8_LOOP);
    jcc(Assembler::less, VECTOR4_TAIL);
  
!   if (UseAVX >= 2) {
  
      cmpq(length, 16);
      jcc(Assembler::equal, VECTOR16_LOOP);
      jcc(Assembler::less, VECTOR8_LOOP);
  
*** 9551,9561 ****
    xorl(tmp1, tmp2);
    testl(tmp1, tmp1);
    jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
    jmpb(SAME_TILL_END);
  
!   if (UseAVX >= 2){
      bind(VECTOR32_NOT_EQUAL);
      vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_256bit);
      vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_256bit);
      vpxor(rymm0, rymm0, rymm2, Assembler::AVX_256bit);
      vpmovmskb(tmp1, rymm0);
--- 9603,9613 ----
    xorl(tmp1, tmp2);
    testl(tmp1, tmp1);
    jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
    jmpb(SAME_TILL_END);
  
!   if (UseAVX >= 2) {
      bind(VECTOR32_NOT_EQUAL);
      vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_256bit);
      vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_256bit);
      vpxor(rymm0, rymm0, rymm2, Assembler::AVX_256bit);
      vpmovmskb(tmp1, rymm0);
*** 9564,9574 ****
      shrq(result);
      jmpb(DONE);
    }
  
    bind(VECTOR16_NOT_EQUAL);
!   if (UseAVX >= 2){
      vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_128bit);
      vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_128bit);
      pxor(rymm0, rymm2);
    } else {
      pcmpeqb(rymm2, rymm2);
--- 9616,9626 ----
      shrq(result);
      jmpb(DONE);
    }
  
    bind(VECTOR16_NOT_EQUAL);
!   if (UseAVX >= 2) {
      vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_128bit);
      vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_128bit);
      pxor(rymm0, rymm2);
    } else {
      pcmpeqb(rymm2, rymm2);
*** 9595,9605 ****
    mov64(result, -1);
  
    bind(DONE);
  }
  
- 
  //Helper functions for square_to_len()
  
  /**
   * Store the squares of x[], right shifted one bit (divided by 2) into z[]
   * Preserves x and z and modifies rest of the registers.
--- 9647,9656 ----