Examples with AMD64MacroAssembler org.graalvm.compiler.asm.amd64.AMD64MacroAssembler used on opensource projects

Example 11 with AMD64MacroAssembler

use of org.graalvm.compiler.asm.amd64.AMD64MacroAssembler in project graal by oracle.

the class AMD64ArrayEqualsOp method emitCode.

@Override
public void emitCode(CompilationResultBuilder crb, AMD64MacroAssembler masm) {
    Register result = asRegister(resultValue);
    Register array1 = asRegister(temp1);
    Register array2 = asRegister(temp2);
    Register length = asRegister(temp3);
    Label trueLabel = new Label();
    Label falseLabel = new Label();
    Label done = new Label();
    // Load array base addresses.
    masm.leaq(array1, new AMD64Address(asRegister(array1Value), arrayBaseOffset));
    masm.leaq(array2, new AMD64Address(asRegister(array2Value), arrayBaseOffset));
    // Get array length in bytes.
    masm.movl(length, asRegister(lengthValue));
    if (arrayIndexScale > 1) {
        // scale length
        masm.shll(length, NumUtil.log2Ceil(arrayIndexScale));
    }
    // copy
    masm.movl(result, length);
    if (supportsAVX2(crb.target)) {
        emitAVXCompare(crb, masm, result, array1, array2, length, trueLabel, falseLabel);
    } else if (supportsSSE41(crb.target)) {
        // this code is used for AVX as well because our backend correctly ensures that
        // VEX-prefixed instructions are emitted if AVX is supported
        emitSSE41Compare(crb, masm, result, array1, array2, length, trueLabel, falseLabel);
    }
    emit8ByteCompare(crb, masm, result, array1, array2, length, trueLabel, falseLabel);
    emitTailCompares(masm, result, array1, array2, length, trueLabel, falseLabel);
    // Return true
    masm.bind(trueLabel);
    masm.movl(result, 1);
    masm.jmpb(done);
    // Return false
    masm.bind(falseLabel);
    masm.xorl(result, result);
    // That's it
    masm.bind(done);
}

Example 12 with AMD64MacroAssembler

use of org.graalvm.compiler.asm.amd64.AMD64MacroAssembler in project graal by oracle.

the class AMD64ArrayEqualsOp method emitFloatCompare.

/**
 * Emits code to compare if two floats are bitwise equal or both NaN.
 */
private void emitFloatCompare(AMD64MacroAssembler masm, Register base1, Register base2, Register index, int offset, Label falseLabel, boolean skipBitwiseCompare) {
    AMD64Address address1 = new AMD64Address(base1, index, Scale.Times1, offset);
    AMD64Address address2 = new AMD64Address(base2, index, Scale.Times1, offset);
    Label bitwiseEqual = new Label();
    if (!skipBitwiseCompare) {
        // Bitwise compare
        Register temp = asRegister(temp4);
        if (kind == JavaKind.Float) {
            masm.movl(temp, address1);
            masm.cmpl(temp, address2);
        } else {
            masm.movq(temp, address1);
            masm.cmpq(temp, address2);
        }
        masm.jccb(ConditionFlag.Equal, bitwiseEqual);
    }
    emitNaNCheck(masm, address1, falseLabel);
    emitNaNCheck(masm, address2, falseLabel);
    masm.bind(bitwiseEqual);
}

Example 13 with AMD64MacroAssembler

use of org.graalvm.compiler.asm.amd64.AMD64MacroAssembler in project graal by oracle.

the class AMD64ArrayEqualsOp method emitSSE41Compare.

/**
 * Emits code that uses SSE4.1 128-bit (16-byte) vector compares.
 */
private void emitSSE41Compare(CompilationResultBuilder crb, AMD64MacroAssembler masm, Register result, Register array1, Register array2, Register length, Label trueLabel, Label falseLabel) {
    assert supportsSSE41(crb.target);
    Register vector1 = asRegister(vectorTemp1, AMD64Kind.DOUBLE);
    Register vector2 = asRegister(vectorTemp2, AMD64Kind.DOUBLE);
    Label loop = new Label();
    Label compareTail = new Label();
    boolean requiresNaNCheck = kind.isNumericFloat();
    Label loopCheck = new Label();
    Label nanCheck = new Label();
    // Compare 16-byte vectors
    // tail count (in bytes)
    masm.andl(result, SSE4_1_VECTOR_SIZE - 1);
    // vector count (in bytes)
    masm.andl(length, ~(SSE4_1_VECTOR_SIZE - 1));
    masm.jcc(ConditionFlag.Zero, compareTail);
    masm.leaq(array1, new AMD64Address(array1, length, Scale.Times1, 0));
    masm.leaq(array2, new AMD64Address(array2, length, Scale.Times1, 0));
    masm.negq(length);
    // Align the main loop
    masm.align(crb.target.wordSize * 2);
    masm.bind(loop);
    masm.movdqu(vector1, new AMD64Address(array1, length, Scale.Times1, 0));
    masm.movdqu(vector2, new AMD64Address(array2, length, Scale.Times1, 0));
    masm.pxor(vector1, vector2);
    masm.ptest(vector1, vector1);
    masm.jcc(ConditionFlag.NotZero, requiresNaNCheck ? nanCheck : falseLabel);
    masm.bind(loopCheck);
    masm.addq(length, SSE4_1_VECTOR_SIZE);
    masm.jcc(ConditionFlag.NotZero, loop);
    masm.testl(result, result);
    masm.jcc(ConditionFlag.Zero, trueLabel);
    if (requiresNaNCheck) {
        Label unalignedCheck = new Label();
        masm.jmpb(unalignedCheck);
        masm.bind(nanCheck);
        emitFloatCompareWithinRange(crb, masm, array1, array2, length, 0, falseLabel, SSE4_1_VECTOR_SIZE);
        masm.jmpb(loopCheck);
        masm.bind(unalignedCheck);
    }
    /*
         * Compare the remaining bytes with an unaligned memory load aligned to the end of the
         * array.
         */
    masm.movdqu(vector1, new AMD64Address(array1, result, Scale.Times1, -SSE4_1_VECTOR_SIZE));
    masm.movdqu(vector2, new AMD64Address(array2, result, Scale.Times1, -SSE4_1_VECTOR_SIZE));
    masm.pxor(vector1, vector2);
    masm.ptest(vector1, vector1);
    if (requiresNaNCheck) {
        masm.jcc(ConditionFlag.Zero, trueLabel);
        emitFloatCompareWithinRange(crb, masm, array1, array2, result, -SSE4_1_VECTOR_SIZE, falseLabel, SSE4_1_VECTOR_SIZE);
    } else {
        masm.jcc(ConditionFlag.NotZero, falseLabel);
    }
    masm.jmp(trueLabel);
    masm.bind(compareTail);
    masm.movl(length, result);
}

Example 14 with AMD64MacroAssembler

use of org.graalvm.compiler.asm.amd64.AMD64MacroAssembler in project graal by oracle.

the class AMD64MathIntrinsicUnaryOp method expIntrinsic.

public void expIntrinsic(Register dest, Register value, CompilationResultBuilder crb, AMD64MacroAssembler masm) {
    ArrayDataPointerConstant onePtr = new ArrayDataPointerConstant(one, 16);
    ArrayDataPointerConstant cvExpPtr = new ArrayDataPointerConstant(cvExp, 16);
    ArrayDataPointerConstant shifterExpPtr = new ArrayDataPointerConstant(shifterExp, 8);
    ArrayDataPointerConstant mMaskExpPtr = new ArrayDataPointerConstant(mMaskExp, 16);
    ArrayDataPointerConstant biasExpPtr = new ArrayDataPointerConstant(biasExp, 16);
    ArrayDataPointerConstant tblAddrExpPtr = new ArrayDataPointerConstant(tblAddrExp, 16);
    ArrayDataPointerConstant expBiasPtr = new ArrayDataPointerConstant(expBias, 8);
    ArrayDataPointerConstant xMaxExpPtr = new ArrayDataPointerConstant(xMaxExp, 8);
    ArrayDataPointerConstant xMinExpPtr = new ArrayDataPointerConstant(xMinExp, 8);
    ArrayDataPointerConstant infExpPtr = new ArrayDataPointerConstant(infExp, 8);
    ArrayDataPointerConstant zeroExpPtr = new ArrayDataPointerConstant(zeroExp, 8);
    ArrayDataPointerConstant allOnesExpPtr = new ArrayDataPointerConstant(allOnesExp, 8);
    Label bb0 = new Label();
    Label bb1 = new Label();
    Label bb2 = new Label();
    Label bb3 = new Label();
    Label bb4 = new Label();
    Label bb5 = new Label();
    Label bb7 = new Label();
    Label bb8 = new Label();
    Label bb9 = new Label();
    Label bb10 = new Label();
    Label bb11 = new Label();
    Label bb12 = new Label();
    Label bb14 = new Label();
    Register gpr1 = asRegister(gpr1Temp, AMD64Kind.QWORD);
    Register gpr2 = asRegister(gpr2Temp, AMD64Kind.QWORD);
    Register gpr3 = asRegister(rcxTemp, AMD64Kind.QWORD);
    Register gpr4 = asRegister(gpr4Temp, AMD64Kind.QWORD);
    Register gpr5 = asRegister(gpr5Temp, AMD64Kind.QWORD);
    Register temp1 = asRegister(xmm1Temp, AMD64Kind.DOUBLE);
    Register temp2 = asRegister(xmm2Temp, AMD64Kind.DOUBLE);
    Register temp3 = asRegister(xmm3Temp, AMD64Kind.DOUBLE);
    Register temp4 = asRegister(xmm4Temp, AMD64Kind.DOUBLE);
    Register temp5 = asRegister(xmm5Temp, AMD64Kind.DOUBLE);
    Register temp6 = asRegister(xmm6Temp, AMD64Kind.DOUBLE);
    Register temp7 = asRegister(xmm7Temp, AMD64Kind.DOUBLE);
    Register temp8 = asRegister(xmm8Temp, AMD64Kind.DOUBLE);
    Register temp9 = asRegister(xmm9Temp, AMD64Kind.DOUBLE);
    Register temp10 = asRegister(xmm10Temp, AMD64Kind.DOUBLE);
    AMD64Address stackSlot = (AMD64Address) crb.asAddress(stackTemp);
    setCrb(crb);
    masm.movsd(stackSlot, value);
    if (dest.encoding != value.encoding) {
        masm.movdqu(dest, value);
    }
    // 0xffffffc0,
    masm.movdqu(temp9, externalAddress(mMaskExpPtr));
    // 0x00000000,
    // 0xffffffc0,
    // 0x00000000
    // 0x0000ffc0,
    masm.movdqu(temp10, externalAddress(biasExpPtr));
    // 0x00000000,
    // 0x0000ffc0,
    // 0x00000000
    masm.unpcklpd(dest, dest);
    masm.leaq(gpr5, stackSlot);
    masm.leaq(gpr2, externalAddress(cvExpPtr));
    // 0x652b82fe,
    masm.movdqu(temp1, new AMD64Address(gpr2, 0));
    // 0x40571547,
    // 0x652b82fe,
    // 0x40571547
    // 0x00000000,
    masm.movdqu(temp6, externalAddress(shifterExpPtr));
    // 0x43380000,
    // 0x00000000,
    // 0x43380000
    // 0xfefa0000,
    masm.movdqu(temp2, new AMD64Address(gpr2, 16));
    // 0x3f862e42,
    // 0xfefa0000,
    // 0x3f862e42
    // 0xbc9e3b3a,
    masm.movdqu(temp3, new AMD64Address(gpr2, 32));
    // 0x3d1cf79a,
    // 0xbc9e3b3a,
    // 0x3d1cf79a
    masm.pextrw(gpr1, dest, 3);
    masm.andl(gpr1, 32767);
    masm.movl(gpr4, 16527);
    masm.subl(gpr4, gpr1);
    masm.subl(gpr1, 15504);
    masm.orl(gpr4, gpr1);
    masm.cmpl(gpr4, Integer.MIN_VALUE);
    masm.jcc(ConditionFlag.AboveEqual, bb0);
    masm.leaq(gpr4, externalAddress(tblAddrExpPtr));
    // 0xfffffffe,
    masm.movdqu(temp8, new AMD64Address(gpr2, 48));
    // 0x3fdfffff,
    // 0xfffffffe,
    // 0x3fdfffff
    // 0xe3289860,
    masm.movdqu(temp4, new AMD64Address(gpr2, 64));
    // 0x3f56c15c,
    // 0x555b9e25,
    // 0x3fa55555
    // 0xc090cf0f,
    masm.movdqu(temp5, new AMD64Address(gpr2, 80));
    // 0x3f811115,
    // 0x55548ba1,
    // 0x3fc55555
    masm.mulpd(temp1, dest);
    masm.addpd(temp1, temp6);
    masm.movapd(temp7, temp1);
    masm.movdl(gpr1, temp1);
    masm.pand(temp7, temp9);
    masm.subpd(temp1, temp6);
    masm.mulpd(temp2, temp1);
    masm.mulpd(temp3, temp1);
    masm.paddq(temp7, temp10);
    masm.subpd(dest, temp2);
    masm.movl(gpr3, gpr1);
    masm.andl(gpr3, 63);
    masm.shll(gpr3, 4);
    masm.movdqu(temp2, new AMD64Address(gpr3, gpr4, Scale.Times1, 0));
    masm.sarl(gpr1, 6);
    masm.psllq(temp7, 46);
    masm.subpd(dest, temp3);
    masm.mulpd(temp4, dest);
    masm.movl(gpr4, gpr1);
    masm.movapd(temp6, dest);
    masm.movapd(temp1, dest);
    masm.mulpd(temp6, temp6);
    masm.mulpd(dest, temp6);
    masm.addpd(temp5, temp4);
    masm.mulsd(dest, temp6);
    masm.mulpd(temp6, temp8);
    masm.addsd(temp1, temp2);
    masm.unpckhpd(temp2, temp2);
    masm.mulpd(dest, temp5);
    masm.addsd(temp1, dest);
    masm.por(temp2, temp7);
    masm.unpckhpd(dest, dest);
    masm.addsd(dest, temp1);
    masm.addsd(dest, temp6);
    masm.addl(gpr4, 894);
    masm.cmpl(gpr4, 1916);
    masm.jcc(ConditionFlag.Above, bb1);
    masm.mulsd(dest, temp2);
    masm.addsd(dest, temp2);
    masm.jmp(bb14);
    masm.bind(bb1);
    // 0x00000000,
    masm.movdqu(temp6, externalAddress(expBiasPtr));
    // 0x3ff00000,
    // 0x00000000,
    // 0x3ff00000
    masm.xorpd(temp3, temp3);
    // 0xffffffff,
    masm.movdqu(temp4, externalAddress(allOnesExpPtr));
    // 0xffffffff,
    // 0xffffffff,
    // 0xffffffff
    masm.movl(gpr4, -1022);
    masm.subl(gpr4, gpr1);
    masm.movdl(temp5, gpr4);
    masm.psllq(temp4, temp5);
    masm.movl(gpr3, gpr1);
    masm.sarl(gpr1, 1);
    masm.pinsrw(temp3, gpr1, 3);
    masm.psllq(temp3, 4);
    masm.psubd(temp2, temp3);
    masm.mulsd(dest, temp2);
    masm.cmpl(gpr4, 52);
    masm.jcc(ConditionFlag.Greater, bb2);
    masm.pand(temp4, temp2);
    masm.paddd(temp3, temp6);
    masm.subsd(temp2, temp4);
    masm.addsd(dest, temp2);
    masm.cmpl(gpr3, 1023);
    masm.jcc(ConditionFlag.GreaterEqual, bb3);
    masm.pextrw(gpr3, dest, 3);
    masm.andl(gpr3, 32768);
    masm.orl(gpr4, gpr3);
    masm.cmpl(gpr4, 0);
    masm.jcc(ConditionFlag.Equal, bb4);
    masm.movapd(temp6, dest);
    masm.addsd(dest, temp4);
    masm.mulsd(dest, temp3);
    masm.pextrw(gpr3, dest, 3);
    masm.andl(gpr3, 32752);
    masm.cmpl(gpr3, 0);
    masm.jcc(ConditionFlag.Equal, bb5);
    masm.jmp(bb14);
    masm.bind(bb5);
    masm.mulsd(temp6, temp3);
    masm.mulsd(temp4, temp3);
    masm.movdqu(dest, temp6);
    masm.pxor(temp6, temp4);
    masm.psrad(temp6, 31);
    masm.pshufd(temp6, temp6, 85);
    masm.psllq(dest, 1);
    masm.psrlq(dest, 1);
    masm.pxor(dest, temp6);
    masm.psrlq(temp6, 63);
    masm.paddq(dest, temp6);
    masm.paddq(dest, temp4);
    masm.jmp(bb14);
    masm.bind(bb4);
    masm.addsd(dest, temp4);
    masm.mulsd(dest, temp3);
    masm.jmp(bb14);
    masm.bind(bb3);
    masm.addsd(dest, temp4);
    masm.mulsd(dest, temp3);
    masm.pextrw(gpr3, dest, 3);
    masm.andl(gpr3, 32752);
    masm.cmpl(gpr3, 32752);
    masm.jcc(ConditionFlag.AboveEqual, bb7);
    masm.jmp(bb14);
    masm.bind(bb2);
    masm.paddd(temp3, temp6);
    masm.addpd(dest, temp2);
    masm.mulsd(dest, temp3);
    masm.jmp(bb14);
    masm.bind(bb8);
    // 0xffffffff,
    masm.movsd(dest, externalAddress(xMaxExpPtr));
    // 0x7fefffff
    // 0x00000000,
    masm.movsd(temp8, externalAddress(xMinExpPtr));
    // 0x00100000
    masm.cmpl(gpr1, 2146435072);
    masm.jcc(ConditionFlag.AboveEqual, bb9);
    masm.movl(gpr1, new AMD64Address(gpr5, 4));
    masm.cmpl(gpr1, Integer.MIN_VALUE);
    masm.jcc(ConditionFlag.AboveEqual, bb10);
    masm.mulsd(dest, dest);
    masm.bind(bb7);
    masm.jmp(bb14);
    masm.bind(bb10);
    masm.mulsd(dest, temp8);
    masm.jmp(bb14);
    masm.bind(bb9);
    masm.movl(gpr4, stackSlot);
    masm.cmpl(gpr1, 2146435072);
    masm.jcc(ConditionFlag.Above, bb11);
    masm.cmpl(gpr4, 0);
    masm.jcc(ConditionFlag.NotEqual, bb11);
    masm.movl(gpr1, new AMD64Address(gpr5, 4));
    masm.cmpl(gpr1, 2146435072);
    masm.jcc(ConditionFlag.NotEqual, bb12);
    // 0x00000000,
    masm.movsd(dest, externalAddress(infExpPtr));
    // 0x7ff00000
    masm.jmp(bb14);
    masm.bind(bb12);
    // 0x00000000,
    masm.movsd(dest, externalAddress(zeroExpPtr));
    // 0x00000000
    masm.jmp(bb14);
    masm.bind(bb11);
    masm.movsd(dest, stackSlot);
    masm.addsd(dest, dest);
    masm.jmp(bb14);
    masm.bind(bb0);
    masm.movl(gpr1, new AMD64Address(gpr5, 4));
    masm.andl(gpr1, 2147483647);
    masm.cmpl(gpr1, 1083179008);
    masm.jcc(ConditionFlag.AboveEqual, bb8);
    // 0x00000000,
    masm.addsd(dest, externalAddress(onePtr));
    // 0x3ff00000
    masm.bind(bb14);
}

Example 15 with AMD64MacroAssembler

use of org.graalvm.compiler.asm.amd64.AMD64MacroAssembler in project graal by oracle.

the class AMD64MathIntrinsicUnaryOp method logIntrinsic.

/*
     * Copyright (c) 2014, 2016, Intel Corporation. All rights reserved. Intel Math Library (LIBM)
     * Source Code
     *
     * ALGORITHM DESCRIPTION - LOG() ---------------------
     *
     * x=2^k * mx, mx in [1,2)
     *
     * Get B~1/mx based on the output of rcpps instruction (B0) B = int((B0*2^7+0.5))/2^7
     *
     * Reduced argument: r=B*mx-1.0 (computed accurately in high and low parts)
     *
     * Result: k*log(2) - log(B) + p(r) if |x-1| >= small value (2^-6) and p(r) is a degree 7
     * polynomial -log(B) read from data table (high, low parts) Result is formed from high and low
     * parts.
     *
     * Special cases: log(NaN) = quiet NaN, and raise invalid exception log(+INF) = that INF log(0)
     * = -INF with divide-by-zero exception raised log(1) = +0 log(x) = NaN with invalid exception
     * raised if x < -0, including -INF
     *
     */
public void logIntrinsic(Register dest, Register value, CompilationResultBuilder crb, AMD64MacroAssembler masm) {
    ArrayDataPointerConstant logTwoTablePtr = new ArrayDataPointerConstant(logTwoTable, 16);
    ArrayDataPointerConstant logTwoDataPtr = new ArrayDataPointerConstant(logTwoData, 16);
    ArrayDataPointerConstant coeffLogTwoDataPtr = new ArrayDataPointerConstant(coeffLogTwoData, 16);
    Label bb0 = new Label();
    Label bb1 = new Label();
    Label bb2 = new Label();
    Label bb3 = new Label();
    Label bb4 = new Label();
    Label bb5 = new Label();
    Label bb6 = new Label();
    Label bb7 = new Label();
    Label bb8 = new Label();
    Register gpr1 = asRegister(gpr1Temp, AMD64Kind.QWORD);
    Register gpr2 = asRegister(gpr2Temp, AMD64Kind.QWORD);
    Register gpr3 = asRegister(rcxTemp, AMD64Kind.QWORD);
    Register gpr4 = asRegister(gpr4Temp, AMD64Kind.QWORD);
    Register temp1 = asRegister(xmm1Temp, AMD64Kind.DOUBLE);
    Register temp2 = asRegister(xmm2Temp, AMD64Kind.DOUBLE);
    Register temp3 = asRegister(xmm3Temp, AMD64Kind.DOUBLE);
    Register temp4 = asRegister(xmm4Temp, AMD64Kind.DOUBLE);
    Register temp5 = asRegister(xmm5Temp, AMD64Kind.DOUBLE);
    Register temp6 = asRegister(xmm6Temp, AMD64Kind.DOUBLE);
    Register temp7 = asRegister(xmm7Temp, AMD64Kind.DOUBLE);
    AMD64Address stackSlot = (AMD64Address) crb.asAddress(stackTemp);
    setCrb(crb);
    masm.movdq(stackSlot, value);
    if (dest.encoding != value.encoding) {
        masm.movdqu(dest, value);
    }
    masm.movq(gpr1, 0x3ff0000000000000L);
    masm.movdq(temp2, gpr1);
    masm.movq(gpr3, 0x77f0000000000000L);
    masm.movdq(temp3, gpr3);
    masm.movl(gpr2, 32768);
    masm.movdl(temp4, gpr2);
    masm.movq(gpr2, 0xffffe00000000000L);
    masm.movdq(temp5, gpr2);
    masm.movdqu(temp1, value);
    masm.pextrw(gpr1, dest, 3);
    masm.por(dest, temp2);
    masm.movl(gpr2, 16352);
    masm.psrlq(dest, 27);
    masm.leaq(gpr4, externalAddress(logTwoTablePtr));
    masm.psrld(dest, 2);
    masm.rcpps(dest, dest);
    masm.psllq(temp1, 12);
    masm.pshufd(temp6, temp5, 0xE4);
    masm.psrlq(temp1, 12);
    masm.subl(gpr1, 16);
    masm.cmpl(gpr1, 32736);
    masm.jcc(ConditionFlag.AboveEqual, bb0);
    masm.bind(bb1);
    masm.paddd(dest, temp4);
    masm.por(temp1, temp3);
    masm.movdl(gpr3, dest);
    masm.psllq(dest, 29);
    masm.pand(temp5, temp1);
    masm.pand(dest, temp6);
    masm.subsd(temp1, temp5);
    masm.mulpd(temp5, dest);
    masm.andl(gpr1, 32752);
    masm.subl(gpr1, gpr2);
    masm.cvtsi2sdl(temp7, gpr1);
    masm.mulsd(temp1, dest);
    // 0xfefa3800,
    masm.movdq(temp6, externalAddress(logTwoDataPtr));
    // 0x3fa62e42
    // 0x92492492,
    masm.movdqu(temp3, externalAddress(coeffLogTwoDataPtr));
    // 0x3fc24924,
    // 0x00000000,
    // 0xbfd00000
    masm.subsd(temp5, temp2);
    masm.andl(gpr3, 16711680);
    masm.shrl(gpr3, 12);
    masm.movdqu(dest, new AMD64Address(gpr4, gpr3, Scale.Times1, 0));
    masm.leaq(gpr4, externalAddress(coeffLogTwoDataPtr));
    // 0x3d6fb175,
    masm.movdqu(temp4, new AMD64Address(gpr4, 16));
    // 0xbfc5555e,
    // 0x55555555,
    // 0x3fd55555
    masm.addsd(temp1, temp5);
    // 0x9999999a,
    masm.movdqu(temp2, new AMD64Address(gpr4, 32));
    // 0x3fc99999,
    // 0x00000000,
    // 0xbfe00000
    masm.mulsd(temp6, temp7);
    if (masm.supports(CPUFeature.SSE3)) {
        masm.movddup(temp5, temp1);
    } else {
        masm.movdqu(temp5, temp1);
        masm.movlhps(temp5, temp5);
    }
    masm.leaq(gpr4, externalAddress(logTwoDataPtr));
    // 0x93c76730,
    masm.mulsd(temp7, new AMD64Address(gpr4, 8));
    // 0x3ceef357
    masm.mulsd(temp3, temp1);
    masm.addsd(dest, temp6);
    masm.mulpd(temp4, temp5);
    masm.mulpd(temp5, temp5);
    if (masm.supports(CPUFeature.SSE3)) {
        masm.movddup(temp6, dest);
    } else {
        masm.movdqu(temp6, dest);
        masm.movlhps(temp6, temp6);
    }
    masm.addsd(dest, temp1);
    masm.addpd(temp4, temp2);
    masm.mulpd(temp3, temp5);
    masm.subsd(temp6, dest);
    masm.mulsd(temp4, temp1);
    masm.pshufd(temp2, dest, 0xEE);
    masm.addsd(temp1, temp6);
    masm.mulsd(temp5, temp5);
    masm.addsd(temp7, temp2);
    masm.addpd(temp4, temp3);
    masm.addsd(temp1, temp7);
    masm.mulpd(temp4, temp5);
    masm.addsd(temp1, temp4);
    masm.pshufd(temp5, temp4, 0xEE);
    masm.addsd(temp1, temp5);
    masm.addsd(dest, temp1);
    masm.jmp(bb8);
    masm.bind(bb0);
    masm.movdq(dest, stackSlot);
    masm.movdq(temp1, stackSlot);
    masm.addl(gpr1, 16);
    masm.cmpl(gpr1, 32768);
    masm.jcc(ConditionFlag.AboveEqual, bb2);
    masm.cmpl(gpr1, 16);
    masm.jcc(ConditionFlag.Below, bb3);
    masm.bind(bb4);
    masm.addsd(dest, dest);
    masm.jmp(bb8);
    masm.bind(bb5);
    masm.jcc(ConditionFlag.Above, bb4);
    masm.cmpl(gpr3, 0);
    masm.jcc(ConditionFlag.Above, bb4);
    masm.jmp(bb6);
    masm.bind(bb3);
    masm.xorpd(temp1, temp1);
    masm.addsd(temp1, dest);
    masm.movdl(gpr3, temp1);
    masm.psrlq(temp1, 32);
    masm.movdl(gpr2, temp1);
    masm.orl(gpr3, gpr2);
    masm.cmpl(gpr3, 0);
    masm.jcc(ConditionFlag.Equal, bb7);
    masm.xorpd(temp1, temp1);
    masm.movl(gpr1, 18416);
    masm.pinsrw(temp1, gpr1, 3);
    masm.mulsd(dest, temp1);
    masm.movdqu(temp1, dest);
    masm.pextrw(gpr1, dest, 3);
    masm.por(dest, temp2);
    masm.psrlq(dest, 27);
    masm.movl(gpr2, 18416);
    masm.psrld(dest, 2);
    masm.rcpps(dest, dest);
    masm.psllq(temp1, 12);
    masm.pshufd(temp6, temp5, 0xE4);
    masm.psrlq(temp1, 12);
    masm.jmp(bb1);
    masm.bind(bb2);
    masm.movdl(gpr3, temp1);
    masm.psrlq(temp1, 32);
    masm.movdl(gpr2, temp1);
    masm.addl(gpr2, gpr2);
    masm.cmpl(gpr2, -2097152);
    masm.jcc(ConditionFlag.AboveEqual, bb5);
    masm.orl(gpr3, gpr2);
    masm.cmpl(gpr3, 0);
    masm.jcc(ConditionFlag.Equal, bb7);
    masm.bind(bb6);
    masm.xorpd(temp1, temp1);
    masm.xorpd(dest, dest);
    masm.movl(gpr1, 32752);
    masm.pinsrw(temp1, gpr1, 3);
    masm.mulsd(dest, temp1);
    masm.jmp(bb8);
    masm.bind(bb7);
    masm.xorpd(temp1, temp1);
    masm.xorpd(dest, dest);
    masm.movl(gpr1, 49136);
    masm.pinsrw(dest, gpr1, 3);
    masm.divsd(dest, temp1);
    masm.bind(bb8);
}