Commit a392dafd authored by Benjamin Bouvier's avatar Benjamin Bouvier
Browse files

Bug 1499198 - Use a constant memory load instead of a word reference in... 

Bug 1499198 - Use a constant memory load instead of a word reference in convertUint64ToDouble. r=luke, a=RyanVM

--HG--
extra : source : 01368a04b48aeca172e2153cd0f8401e5162226a
parent a48467f0

js/src/jit/x86/Assembler-x86.h

+2 −11
Original line number Diff line number Diff line
@@ -473,20 +473,11 @@ class Assembler : public AssemblerX86Shared 
        MOZ_ASSERT(dest.size() == 16);

        masm.vhaddpd_rr(src.encoding(), dest.encoding());

    }

    void vsubpd(const Operand& src1, FloatRegister src0, FloatRegister dest) {

    void vsubpd(FloatRegister src1, FloatRegister src0, FloatRegister dest) {

        MOZ_ASSERT(HasSSE2());

        MOZ_ASSERT(src0.size() == 16);

        MOZ_ASSERT(dest.size() == 16);

        switch (src1.kind()) {

          case Operand::MEM_REG_DISP:

            masm.vsubpd_mr(src1.disp(), src1.base(), src0.encoding(), dest.encoding());

            break;

          case Operand::MEM_ADDRESS32:

            masm.vsubpd_mr(src1.address(), src0.encoding(), dest.encoding());

            break;

          default:

            MOZ_CRASH("unexpected operand kind");

        }

        masm.vsubpd_rr(src1.encoding(), src0.encoding(), dest.encoding());

    }



    void vpunpckldq(FloatRegister src1, FloatRegister src0, FloatRegister dest) {

js/src/jit/x86/BaseAssembler-x86.h

+0 −8
Original line number Diff line number Diff line
@@ -176,14 +176,6 @@ class BaseAssemblerX86 : public BaseAssembler 
    {

        twoByteOpSimd("vsubpd", VEX_PD, OP2_SUBPS_VpsWps, src1, src0, dst);

    }

    void vsubpd_mr(int32_t offset, RegisterID base, XMMRegisterID src0, XMMRegisterID dst)

    {

        twoByteOpSimd("vsubpd", VEX_PD, OP2_SUBPS_VpsWps, offset, base, src0, dst);

    }

    void vsubpd_mr(const void* address, XMMRegisterID src0, XMMRegisterID dst)

    {

        twoByteOpSimd("vsubpd", VEX_PD, OP2_SUBPS_VpsWps, address, src0, dst);

    }



    void vpunpckldq_rr(XMMRegisterID src1, XMMRegisterID src0, XMMRegisterID dst) {

        twoByteOpSimd("vpunpckldq", VEX_PD, OP2_PUNPCKLDQ, src1, src0, dst);

js/src/jit/x86/MacroAssembler-x86.cpp

+19 −12
Original line number Diff line number Diff line
@@ -1147,15 +1147,6 @@ MacroAssembler::wasmTruncateFloat32ToUInt64(FloatRegister input, Register64 outp 
// ========================================================================

// Convert floating point.



// vpunpckldq requires 16-byte boundary for memory operand.

// See convertUInt64ToDouble for the details.

MOZ_ALIGNED_DECL(static const uint64_t, 16) TO_DOUBLE[4] = {

    0x4530000043300000LL,

    0x0LL,

    0x4330000000000000LL,

    0x4530000000000000LL

};



bool

MacroAssembler::convertUInt64ToDoubleNeedsTemp()

{
@@ -1214,8 +1205,16 @@ MacroAssembler::convertUInt64ToDouble(Register64 src, FloatRegister dest, Regist 
    // here, each 64-bit part of dest represents following double:

    //   HI(dest) = 0x 1.00000HHHHHHHH * 2**84 == 2**84 + 0x HHHHHHHH 00000000

    //   LO(dest) = 0x 1.00000LLLLLLLL * 2**52 == 2**52 + 0x 00000000 LLLLLLLL

    movePtr(ImmWord((uintptr_t)TO_DOUBLE), temp);

    vpunpckldq(Operand(temp, 0), dest128, dest128);

    // See convertUInt64ToDouble for the details.

    static const int32_t CST1[4] = {

        0x43300000,

        0x45300000,

        0x0,

        0x0,

    };



    loadConstantSimd128Int(SimdConstant::CreateX4(CST1), ScratchSimd128Reg);

    vpunpckldq(ScratchSimd128Reg, dest128, dest128);



    // Subtract a constant C2 from dest, for each 64-bit part:

    //   C2       = 0x 45300000 00000000  43300000 00000000
@@ -1225,7 +1224,15 @@ MacroAssembler::convertUInt64ToDouble(Register64 src, FloatRegister dest, Regist 
    // after the operation each 64-bit part of dest represents following:

    //   HI(dest) = double(0x HHHHHHHH 00000000)

    //   LO(dest) = double(0x 00000000 LLLLLLLL)

    vsubpd(Operand(temp, sizeof(uint64_t) * 2), dest128, dest128);

    static const int32_t CST2[4] = {

        0x0,

        0x43300000,

        0x0,

        0x45300000,

    };



    loadConstantSimd128Int(SimdConstant::CreateX4(CST2), ScratchSimd128Reg);

    vsubpd(ScratchSimd128Reg, dest128, dest128);



    // Add HI(dest) and LO(dest) in double and store it into LO(dest),

    //   LO(dest) = double(0x HHHHHHHH 00000000) + double(0x 00000000 LLLLLLLL)