void MacroAssembler::fast_log(XMMRegister xmm0, XMMRegister xmm1, XMMRegister xmm2, XMMRegister xmm3, XMMRegister xmm4, XMMRegister xmm5, XMMRegister xmm6, XMMRegister xmm7, Register eax, Register ecx, Register edx, Register tmp1, Register tmp2) {
Label L_2TAG_PACKET_0_0_2, L_2TAG_PACKET_1_0_2, L_2TAG_PACKET_2_0_2, L_2TAG_PACKET_3_0_2;
Label L_2TAG_PACKET_4_0_2, L_2TAG_PACKET_5_0_2, L_2TAG_PACKET_6_0_2, L_2TAG_PACKET_7_0_2;
Label L_2TAG_PACKET_8_0_2;
Label L_2TAG_PACKET_12_0_2, L_2TAG_PACKET_13_0_2, B1_3, B1_5, start;
assert_different_registers(tmp1, tmp2, eax, ecx, edx);
jmp(start);
address L_tbl = (address)_L_tbl;
address log2 = (address)_log2;
address coeff = (address)_coeff;
bind(start);
subq(rsp, 24);
movsd(Address(rsp, 0), xmm0);
mov64(rax, 0x3ff0000000000000);
movdq(xmm2, rax);
mov64(rdx, 0x77f0000000000000);
movdq(xmm3, rdx);
movl(ecx, 32768);
movdl(xmm4, rcx);
mov64(tmp1, 0xffffe00000000000);
movdq(xmm5, tmp1);
movdqu(xmm1, xmm0);
pextrw(eax, xmm0, 3);
por(xmm0, xmm2);
movl(ecx, 16352);
psrlq(xmm0, 27);
lea(tmp2, ExternalAddress(L_tbl));
psrld(xmm0, 2);
rcpps(xmm0, xmm0);
psllq(xmm1, 12);
pshufd(xmm6, xmm5, 228);
psrlq(xmm1, 12);
subl(eax, 16);
cmpl(eax, 32736);
jcc(Assembler::aboveEqual, L_2TAG_PACKET_0_0_2);
bind(L_2TAG_PACKET_1_0_2);
paddd(xmm0, xmm4);
por(xmm1, xmm3);
movdl(edx, xmm0);
psllq(xmm0, 29);
pand(xmm5, xmm1);
pand(xmm0, xmm6);
subsd(xmm1, xmm5);
mulpd(xmm5, xmm0);
andl(eax, 32752);
subl(eax, ecx);
cvtsi2sdl(xmm7, eax);
mulsd(xmm1, xmm0);
movq(xmm6, ExternalAddress(log2)); // 0xfefa3800UL, 0x3fa62e42UL
movdqu(xmm3, ExternalAddress(coeff)); // 0x92492492UL, 0x3fc24924UL, 0x00000000UL, 0xbfd00000UL
subsd(xmm5, xmm2);
andl(edx, 16711680);
shrl(edx, 12);
movdqu(xmm0, Address(tmp2, edx));
movdqu(xmm4, ExternalAddress(16 + coeff)); // 0x3d6fb175UL, 0xbfc5555eUL, 0x55555555UL, 0x3fd55555UL
addsd(xmm1, xmm5);
movdqu(xmm2, ExternalAddress(32 + coeff)); // 0x9999999aUL, 0x3fc99999UL, 0x00000000UL, 0xbfe00000UL
mulsd(xmm6, xmm7);
movddup(xmm5, xmm1);
mulsd(xmm7, ExternalAddress(8 + log2)); // 0x93c76730UL, 0x3ceef357UL
mulsd(xmm3, xmm1);
addsd(xmm0, xmm6);
mulpd(xmm4, xmm5);
mulpd(xmm5, xmm5);
movddup(xmm6, xmm0);
addsd(xmm0, xmm1);
addpd(xmm4, xmm2);
mulpd(xmm3, xmm5);
subsd(xmm6, xmm0);
mulsd(xmm4, xmm1);
pshufd(xmm2, xmm0, 238);
addsd(xmm1, xmm6);
mulsd(xmm5, xmm5);
addsd(xmm7, xmm2);
addpd(xmm4, xmm3);
addsd(xmm1, xmm7);
mulpd(xmm4, xmm5);
addsd(xmm1, xmm4);
pshufd(xmm5, xmm4, 238);
addsd(xmm1, xmm5);
addsd(xmm0, xmm1);
jmp(B1_5);
bind(L_2TAG_PACKET_0_0_2);
movq(xmm0, Address(rsp, 0));
movq(xmm1, Address(rsp, 0));
addl(eax, 16);
cmpl(eax, 32768);
jcc(Assembler::aboveEqual, L_2TAG_PACKET_2_0_2);
cmpl(eax, 16);
jcc(Assembler::below, L_2TAG_PACKET_3_0_2);
bind(L_2TAG_PACKET_4_0_2);
addsd(xmm0, xmm0);
jmp(B1_5);
bind(L_2TAG_PACKET_5_0_2);
jcc(Assembler::above, L_2TAG_PACKET_4_0_2);
cmpl(edx, 0);
jcc(Assembler::above, L_2TAG_PACKET_4_0_2);
jmp(L_2TAG_PACKET_6_0_2);
bind(L_2TAG_PACKET_3_0_2);
xorpd(xmm1, xmm1);
addsd(xmm1, xmm0);
movdl(edx, xmm1);
psrlq(xmm1, 32);
movdl(ecx, xmm1);
orl(edx, ecx);
cmpl(edx, 0);
jcc(Assembler::equal, L_2TAG_PACKET_7_0_2);
xorpd(xmm1, xmm1);
movl(eax, 18416);
pinsrw(xmm1, eax, 3);
mulsd(xmm0, xmm1);
movdqu(xmm1, xmm0);
pextrw(eax, xmm0, 3);
por(xmm0, xmm2);
psrlq(xmm0, 27);
movl(ecx, 18416);
psrld(xmm0, 2);
rcpps(xmm0, xmm0);
psllq(xmm1, 12);
pshufd(xmm6, xmm5, 228);
psrlq(xmm1, 12);
jmp(L_2TAG_PACKET_1_0_2);
bind(L_2TAG_PACKET_2_0_2);
movdl(edx, xmm1);
psrlq(xmm1, 32);
movdl(ecx, xmm1);
addl(ecx, ecx);
cmpl(ecx, -2097152);
jcc(Assembler::aboveEqual, L_2TAG_PACKET_5_0_2);
orl(edx, ecx);
cmpl(edx, 0);
jcc(Assembler::equal, L_2TAG_PACKET_7_0_2);
bind(L_2TAG_PACKET_6_0_2);
xorpd(xmm1, xmm1);
xorpd(xmm0, xmm0);
movl(eax, 32752);
pinsrw(xmm1, eax, 3);
mulsd(xmm0, xmm1);
movl(Address(rsp, 16), 3);
jmp(L_2TAG_PACKET_8_0_2);
bind(L_2TAG_PACKET_7_0_2);
xorpd(xmm1, xmm1);
xorpd(xmm0, xmm0);
movl(eax, 49136);
pinsrw(xmm0, eax, 3);
divsd(xmm0, xmm1);
movl(Address(rsp, 16), 2);
bind(L_2TAG_PACKET_8_0_2);
movq(Address(rsp, 8), xmm0);
bind(B1_3);
movq(xmm0, Address(rsp, 8));
bind(B1_5);
addq(rsp, 24);
}
void genDiv(const char*, const char*)
{
divsd(Xbyak::Xmm(regIdx_ - 1), Xbyak::Xmm(regIdx_));
regIdx_--;
}
void MacroAssembler::fast_log(XMMRegister xmm0, XMMRegister xmm1, XMMRegister xmm2, XMMRegister xmm3, XMMRegister xmm4, XMMRegister xmm5, XMMRegister xmm6, XMMRegister xmm7, Register eax, Register ecx, Register edx, Register tmp) {
Label L_2TAG_PACKET_0_0_2, L_2TAG_PACKET_1_0_2, L_2TAG_PACKET_2_0_2, L_2TAG_PACKET_3_0_2;
Label L_2TAG_PACKET_4_0_2, L_2TAG_PACKET_5_0_2, L_2TAG_PACKET_6_0_2, L_2TAG_PACKET_7_0_2;
Label L_2TAG_PACKET_8_0_2, L_2TAG_PACKET_9_0_2;
Label L_2TAG_PACKET_10_0_2, start;
assert_different_registers(tmp, eax, ecx, edx);
jmp(start);
address static_const_table = (address)_static_const_table_log;
bind(start);
subl(rsp, 104);
movl(Address(rsp, 40), tmp);
lea(tmp, ExternalAddress(static_const_table));
xorpd(xmm2, xmm2);
movl(eax, 16368);
pinsrw(xmm2, eax, 3);
xorpd(xmm3, xmm3);
movl(edx, 30704);
pinsrw(xmm3, edx, 3);
movsd(xmm0, Address(rsp, 112));
movapd(xmm1, xmm0);
movl(ecx, 32768);
movdl(xmm4, ecx);
movsd(xmm5, Address(tmp, 2128)); // 0x00000000UL, 0xffffe000UL
pextrw(eax, xmm0, 3);
por(xmm0, xmm2);
psllq(xmm0, 5);
movl(ecx, 16352);
psrlq(xmm0, 34);
rcpss(xmm0, xmm0);
psllq(xmm1, 12);
pshufd(xmm6, xmm5, 228);
psrlq(xmm1, 12);
subl(eax, 16);
cmpl(eax, 32736);
jcc(Assembler::aboveEqual, L_2TAG_PACKET_0_0_2);
bind(L_2TAG_PACKET_1_0_2);
paddd(xmm0, xmm4);
por(xmm1, xmm3);
movdl(edx, xmm0);
psllq(xmm0, 29);
pand(xmm5, xmm1);
pand(xmm0, xmm6);
subsd(xmm1, xmm5);
mulpd(xmm5, xmm0);
andl(eax, 32752);
subl(eax, ecx);
cvtsi2sdl(xmm7, eax);
mulsd(xmm1, xmm0);
movsd(xmm6, Address(tmp, 2064)); // 0xfefa3800UL, 0x3fa62e42UL
movdqu(xmm3, Address(tmp, 2080)); // 0x92492492UL, 0x3fc24924UL, 0x00000000UL, 0xbfd00000UL
subsd(xmm5, xmm2);
andl(edx, 16711680);
shrl(edx, 12);
movdqu(xmm0, Address(tmp, edx));
movdqu(xmm4, Address(tmp, 2096)); // 0x3d6fb175UL, 0xbfc5555eUL, 0x55555555UL, 0x3fd55555UL
addsd(xmm1, xmm5);
movdqu(xmm2, Address(tmp, 2112)); // 0x9999999aUL, 0x3fc99999UL, 0x00000000UL, 0xbfe00000UL
mulsd(xmm6, xmm7);
pshufd(xmm5, xmm1, 68);
mulsd(xmm7, Address(tmp, 2072)); // 0x93c76730UL, 0x3ceef357UL, 0x92492492UL, 0x3fc24924UL
mulsd(xmm3, xmm1);
addsd(xmm0, xmm6);
mulpd(xmm4, xmm5);
mulpd(xmm5, xmm5);
pshufd(xmm6, xmm0, 228);
addsd(xmm0, xmm1);
addpd(xmm4, xmm2);
mulpd(xmm3, xmm5);
subsd(xmm6, xmm0);
mulsd(xmm4, xmm1);
pshufd(xmm2, xmm0, 238);
addsd(xmm1, xmm6);
mulsd(xmm5, xmm5);
addsd(xmm7, xmm2);
addpd(xmm4, xmm3);
addsd(xmm1, xmm7);
mulpd(xmm4, xmm5);
addsd(xmm1, xmm4);
pshufd(xmm5, xmm4, 238);
addsd(xmm1, xmm5);
addsd(xmm0, xmm1);
jmp(L_2TAG_PACKET_2_0_2);
bind(L_2TAG_PACKET_0_0_2);
movsd(xmm0, Address(rsp, 112));
movdqu(xmm1, xmm0);
addl(eax, 16);
cmpl(eax, 32768);
jcc(Assembler::aboveEqual, L_2TAG_PACKET_3_0_2);
cmpl(eax, 16);
jcc(Assembler::below, L_2TAG_PACKET_4_0_2);
bind(L_2TAG_PACKET_5_0_2);
addsd(xmm0, xmm0);
jmp(L_2TAG_PACKET_2_0_2);
bind(L_2TAG_PACKET_6_0_2);
jcc(Assembler::above, L_2TAG_PACKET_5_0_2);
cmpl(edx, 0);
jcc(Assembler::above, L_2TAG_PACKET_5_0_2);
jmp(L_2TAG_PACKET_7_0_2);
bind(L_2TAG_PACKET_3_0_2);
movdl(edx, xmm1);
psrlq(xmm1, 32);
movdl(ecx, xmm1);
addl(ecx, ecx);
cmpl(ecx, -2097152);
jcc(Assembler::aboveEqual, L_2TAG_PACKET_6_0_2);
orl(edx, ecx);
cmpl(edx, 0);
jcc(Assembler::equal, L_2TAG_PACKET_8_0_2);
bind(L_2TAG_PACKET_7_0_2);
xorpd(xmm1, xmm1);
xorpd(xmm0, xmm0);
movl(eax, 32752);
pinsrw(xmm1, eax, 3);
movl(edx, 3);
mulsd(xmm0, xmm1);
bind(L_2TAG_PACKET_9_0_2);
movsd(Address(rsp, 0), xmm0);
movsd(xmm0, Address(rsp, 112));
fld_d(Address(rsp, 0));
jmp(L_2TAG_PACKET_10_0_2);
bind(L_2TAG_PACKET_8_0_2);
xorpd(xmm1, xmm1);
xorpd(xmm0, xmm0);
movl(eax, 49136);
pinsrw(xmm0, eax, 3);
divsd(xmm0, xmm1);
movl(edx, 2);
jmp(L_2TAG_PACKET_9_0_2);
bind(L_2TAG_PACKET_4_0_2);
movdl(edx, xmm1);
psrlq(xmm1, 32);
movdl(ecx, xmm1);
orl(edx, ecx);
cmpl(edx, 0);
jcc(Assembler::equal, L_2TAG_PACKET_8_0_2);
xorpd(xmm1, xmm1);
movl(eax, 18416);
pinsrw(xmm1, eax, 3);
mulsd(xmm0, xmm1);
movapd(xmm1, xmm0);
pextrw(eax, xmm0, 3);
por(xmm0, xmm2);
psllq(xmm0, 5);
movl(ecx, 18416);
psrlq(xmm0, 34);
rcpss(xmm0, xmm0);
psllq(xmm1, 12);
pshufd(xmm6, xmm5, 228);
psrlq(xmm1, 12);
jmp(L_2TAG_PACKET_1_0_2);
bind(L_2TAG_PACKET_2_0_2);
movsd(Address(rsp, 24), xmm0);
fld_d(Address(rsp, 24));
bind(L_2TAG_PACKET_10_0_2);
movl(tmp, Address(rsp, 40));
}
void genDiv()
{
divsd(Xbyak::Xmm(regIdx_ - 1), Xbyak::Xmm(regIdx_)); regIdx_--;
}
