1.1 --- a/src/cpu/x86/vm/stubGenerator_x86_64.cpp Mon Mar 21 14:58:37 2016 -0700
1.2 +++ b/src/cpu/x86/vm/stubGenerator_x86_64.cpp Wed Feb 17 13:40:12 2016 +0300
1.3 @@ -3743,6 +3743,107 @@
1.4 return start;
1.5 }
1.6
1.7 +/**
1.8 + * Arguments:
1.9 + *
1.10 + // Input:
1.11 + // c_rarg0 - x address
1.12 + // c_rarg1 - x length
1.13 + // c_rarg2 - z address
1.14 + // c_rarg3 - z lenth
1.15 + *
1.16 + */
1.17 + address generate_squareToLen() {
1.18 +
1.19 + __ align(CodeEntryAlignment);
1.20 + StubCodeMark mark(this, "StubRoutines", "squareToLen");
1.21 +
1.22 + address start = __ pc();
1.23 + // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
1.24 + // Unix: rdi, rsi, rdx, rcx (c_rarg0, c_rarg1, ...)
1.25 + const Register x = rdi;
1.26 + const Register len = rsi;
1.27 + const Register z = r8;
1.28 + const Register zlen = rcx;
1.29 +
1.30 + const Register tmp1 = r12;
1.31 + const Register tmp2 = r13;
1.32 + const Register tmp3 = r14;
1.33 + const Register tmp4 = r15;
1.34 + const Register tmp5 = rbx;
1.35 +
1.36 + BLOCK_COMMENT("Entry:");
1.37 + __ enter(); // required for proper stackwalking of RuntimeStub frame
1.38 +
1.39 + setup_arg_regs(4); // x => rdi, len => rsi, z => rdx
1.40 + // zlen => rcx
1.41 + // r9 and r10 may be used to save non-volatile registers
1.42 + __ movptr(r8, rdx);
1.43 + __ square_to_len(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx, rax);
1.44 +
1.45 + restore_arg_regs();
1.46 +
1.47 + __ leave(); // required for proper stackwalking of RuntimeStub frame
1.48 + __ ret(0);
1.49 +
1.50 + return start;
1.51 + }
1.52 +
1.53 + /**
1.54 + * Arguments:
1.55 + *
1.56 + * Input:
1.57 + * c_rarg0 - out address
1.58 + * c_rarg1 - in address
1.59 + * c_rarg2 - offset
1.60 + * c_rarg3 - len
1.61 + * not Win64
1.62 + * c_rarg4 - k
1.63 + * Win64
1.64 + * rsp+40 - k
1.65 + */
1.66 + address generate_mulAdd() {
1.67 + __ align(CodeEntryAlignment);
1.68 + StubCodeMark mark(this, "StubRoutines", "mulAdd");
1.69 +
1.70 + address start = __ pc();
1.71 + // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
1.72 + // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
1.73 + const Register out = rdi;
1.74 + const Register in = rsi;
1.75 + const Register offset = r11;
1.76 + const Register len = rcx;
1.77 + const Register k = r8;
1.78 +
1.79 + // Next registers will be saved on stack in mul_add().
1.80 + const Register tmp1 = r12;
1.81 + const Register tmp2 = r13;
1.82 + const Register tmp3 = r14;
1.83 + const Register tmp4 = r15;
1.84 + const Register tmp5 = rbx;
1.85 +
1.86 + BLOCK_COMMENT("Entry:");
1.87 + __ enter(); // required for proper stackwalking of RuntimeStub frame
1.88 +
1.89 + setup_arg_regs(4); // out => rdi, in => rsi, offset => rdx
1.90 + // len => rcx, k => r8
1.91 + // r9 and r10 may be used to save non-volatile registers
1.92 +#ifdef _WIN64
1.93 + // last argument is on stack on Win64
1.94 + __ movl(k, Address(rsp, 6 * wordSize));
1.95 +#endif
1.96 + __ movptr(r11, rdx); // move offset in rdx to offset(r11)
1.97 + __ mul_add(out, in, offset, len, k, tmp1, tmp2, tmp3, tmp4, tmp5, rdx, rax);
1.98 +
1.99 + restore_arg_regs();
1.100 +
1.101 + __ leave(); // required for proper stackwalking of RuntimeStub frame
1.102 + __ ret(0);
1.103 +
1.104 + return start;
1.105 + }
1.106 +
1.107 +
1.108 #undef __
1.109 #define __ masm->
1.110
1.111 @@ -3987,6 +4088,12 @@
1.112 if (UseMultiplyToLenIntrinsic) {
1.113 StubRoutines::_multiplyToLen = generate_multiplyToLen();
1.114 }
1.115 + if (UseSquareToLenIntrinsic) {
1.116 + StubRoutines::_squareToLen = generate_squareToLen();
1.117 + }
1.118 + if (UseMulAddIntrinsic) {
1.119 + StubRoutines::_mulAdd = generate_mulAdd();
1.120 + }
1.121 #endif
1.122 }
1.123
