Core/DSPCore: Improve Interpreter address register add/sub, convert to

assembler for JIT. Replace JIT ToMask() with a different variant. Remove
 superfluous zeroWriteBackLog calls(added by me).
Core/Common: Don't bother creating a string and calling into a Logs trigger()
 when there is noone listening. Change AtomicLoadAcquire for gcc to just
 make the compiler not reorder memory accesses around it instead of doing
 a full memory barrier, per the comment in the win32 variant.
Core/AudioCommon: Fix a use of uninitialized variable inside libalsa.

Microbenchmarking results for ToMask variants:(1 000 000 000 iterations):
             cpu\variant| shifts | bit scan
intel mobile C2D@2.5GHz | 5.5s   | 4.0s
amd athlon64x2@3GHz     | 6.1s   | 6.4s
(including some constant overhead identical to both variants)



git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6667 8ced0084-cf51-0410-be5f-012b33b47a6e
This commit is contained in:
pierre 2010-12-27 15:05:18 +00:00
parent 0e737235a8
commit c33f46406e
9 changed files with 214 additions and 194 deletions

View File

@ -123,6 +123,7 @@ bool AlsaSound::AlsaInit()
return false;
}
dir = 0;
err = snd_pcm_hw_params_set_rate_near(handle, hwparams, &sample_rate, &dir);
if (err < 0)
{

View File

@ -57,8 +57,12 @@ inline u32 AtomicLoad(volatile u32& src) {
return src; // 32-bit reads are always atomic.
}
inline u32 AtomicLoadAcquire(volatile u32& src) {
__sync_synchronize(); // TODO: May not be necessary.
return src;
//keep the compiler from caching any memory references
u32 result = src; // 32-bit reads are always atomic.
//__sync_synchronize(); // TODO: May not be necessary.
// Compiler instruction only. x86 loads always have acquire semantics.
__asm__ __volatile__ ( "":::"memory" );
return result;
}
inline void AtomicOr(volatile u32& target, u32 value) {

View File

@ -116,7 +116,7 @@ void LogManager::Log(LogTypes::LOG_LEVELS level, LogTypes::LOG_TYPE type,
char msg[MAX_MSGLEN * 2];
LogContainer *log = m_Log[type];
if (! log->isEnable() || level > log->getLevel())
if (! log->isEnable() || level > log->getLevel() || ! log->hasListeners())
return;
CharArrayFromFormatV(temp, MAX_MSGLEN, format, args);

View File

@ -89,6 +89,7 @@ public:
void setLevel(LogTypes::LOG_LEVELS level) {
m_level = level;
}
bool hasListeners() const { return listeners.size() > 0; }
private:
char m_fullName[128];

View File

@ -259,7 +259,7 @@ private:
void Update_SR_Register(Gen::X64Reg val = Gen::EAX);
void ToMask(Gen::X64Reg value_reg = Gen::EDI, Gen::X64Reg temp_reg = Gen::ESI);
void ToMask(Gen::X64Reg value_reg = Gen::EDI);
void dsp_increment_one(Gen::X64Reg ar = Gen::EAX, Gen::X64Reg wr = Gen::EDX, Gen::X64Reg wr_pow = Gen::EDI, Gen::X64Reg temp_reg = Gen::ESI);
void dsp_decrement_one(Gen::X64Reg ar = Gen::EAX, Gen::X64Reg wr = Gen::EDX, Gen::X64Reg wr_pow = Gen::EDI, Gen::X64Reg temp_reg = Gen::ESI);
void get_long_prod(Gen::X64Reg long_prod = Gen::RAX);

View File

@ -68,10 +68,10 @@ inline u16 dsp_increase_addr_reg(u16 reg, s16 ix)
u16 m = ToMask(wr) | 1;
u16 nar = ar+ix;
if (ix >= 0) {
if((ar&m)+(ix&m) -m-1 >= 0)
if((ar&m) + (int)(ix&m) -(int)m-1 >= 0)
nar -= wr+1;
} else {
if((ar&m)+(ix&m) -m-1 < m-wr)
if((ar&m) + (int)(ix&m) -(int)m-1 < m-wr)
nar += wr+1;
}
return nar;
@ -82,13 +82,12 @@ inline u16 dsp_decrease_addr_reg(u16 reg, s16 ix)
u16 ar = g_dsp.r[reg];
u16 wr = g_dsp.r[reg+8];
u16 m = ToMask(wr) | 1;
ix = -ix-1;
u16 nar = ar+ix+1;
if (ix-1 >= 0) {
if((ar&m)+(ix&m) -m >= 0)
u16 nar = ar-ix;
if ((u16)ix > 0x8000) { // equiv: ix < 0 && ix != -0x8000
if((ar&m) - (int)(ix&m) >= 0)
nar -= wr+1;
} else {
if((ar&m)+(ix&m) -m < m-wr)
if((ar&m) - (int)(ix&m) < m-wr)
nar += wr+1;
}
return nar;

View File

@ -37,8 +37,8 @@ const DSPOPCTemplate opcodes[] =
{"DAR", 0x0004, 0xfffc, DSPInterpreter::dar, &DSPEmitter::dar, 1, 1, {{P_REG, 1, 0, 0, 0x0003}}, false, false, false, false, false},
{"IAR", 0x0008, 0xfffc, DSPInterpreter::iar, &DSPEmitter::iar, 1, 1, {{P_REG, 1, 0, 0, 0x0003}}, false, false, false, false, false},
{"SUBARN", 0x000c, 0xfffc, DSPInterpreter::subarn, NULL/*&DSPEmitter::subarn*/, 1, 1, {{P_REG, 1, 0, 0, 0x0003}}, false, false, false, false, false},
{"ADDARN", 0x0010, 0xfff0, DSPInterpreter::addarn, NULL/*&DSPEmitter::addarn*/, 1, 2, {{P_REG, 1, 0, 0, 0x0003}, {P_REG04, 1, 0, 2, 0x000c}}, false, false, false, false, false},
{"SUBARN", 0x000c, 0xfffc, DSPInterpreter::subarn, &DSPEmitter::subarn, 1, 1, {{P_REG, 1, 0, 0, 0x0003}}, false, false, false, false, false},
{"ADDARN", 0x0010, 0xfff0, DSPInterpreter::addarn, &DSPEmitter::addarn, 1, 2, {{P_REG, 1, 0, 0, 0x0003}, {P_REG04, 1, 0, 2, 0x000c}}, false, false, false, false, false},
{"HALT", 0x0021, 0xffff, DSPInterpreter::halt, NULL, 1, 0, {}, false, true, true, false, false},

View File

@ -370,7 +370,6 @@ void DSPEmitter::nx(const UDSPInstruction opc)
void DSPEmitter::dar(const UDSPInstruction opc)
{
// g_dsp.r[opc & 0x3] = dsp_decrement_addr_reg(opc & 0x3);
zeroWriteBackLog(opc);
decrement_addr_reg(opc & 0x3);
}
@ -381,7 +380,6 @@ void DSPEmitter::dar(const UDSPInstruction opc)
void DSPEmitter::iar(const UDSPInstruction opc)
{
// g_dsp.r[opc & 0x3] = dsp_increment_addr_reg(opc & 0x3);
zeroWriteBackLog(opc);
increment_addr_reg(opc & 0x3);
}
@ -393,7 +391,6 @@ void DSPEmitter::subarn(const UDSPInstruction opc)
{
// u8 dreg = opc & 0x3;
// g_dsp.r[dreg] = dsp_decrease_addr_reg(dreg, (s16)g_dsp.r[DSP_REG_IX0 + dreg]);
zeroWriteBackLog(opc);
decrease_addr_reg(opc & 0x3);
}
@ -408,7 +405,6 @@ void DSPEmitter::addarn(const UDSPInstruction opc)
// g_dsp.r[dreg] = dsp_increase_addr_reg(dreg, (s16)g_dsp.r[DSP_REG_IX0 + sreg]);
// From looking around it is always called with the matching index register
zeroWriteBackLog(opc);
increase_addr_reg(opc & 0x3);
}
@ -448,7 +444,6 @@ void DSPEmitter::sbclr(const UDSPInstruction opc)
{
u8 bit = (opc & 0x7) + 6;
zeroWriteBackLog(opc);
clrCompileSR(1 << bit);
}
@ -460,10 +455,10 @@ void DSPEmitter::sbset(const UDSPInstruction opc)
{
u8 bit = (opc & 0x7) + 6;
zeroWriteBackLog(opc);
setCompileSR(1 << bit);
}
// 1000 1bbb xxxx xxxx, bbb >= 010
// This is a bunch of flag setters, flipping bits in SR. So far so good,
// but it's harder to know exactly what effect they have.
void DSPEmitter::srbith(const UDSPInstruction opc)

View File

@ -27,126 +27,141 @@
using namespace Gen;
// Performs the hashing required by increment/increase/decrease_addr_reg
void DSPEmitter::ToMask(X64Reg value_reg, X64Reg temp_reg)
// clobbers RCX
void DSPEmitter::ToMask(X64Reg value_reg)
{
MOV(16, R(temp_reg), R(value_reg));
SHR(16, R(temp_reg), Imm8(8));
OR(16, R(value_reg), R(temp_reg));
MOV(16, R(temp_reg), R(value_reg));
SHR(16, R(temp_reg), Imm8(4));
OR(16, R(value_reg), R(temp_reg));
MOV(16, R(temp_reg), R(value_reg));
SHR(16, R(temp_reg), Imm8(2));
OR(16, R(value_reg), R(temp_reg));
MOV(16, R(temp_reg), R(value_reg));
SHR(16, R(temp_reg), Imm8(1));
OR(16, R(value_reg), R(temp_reg));
}
#if 0
MOV(16, R(CX), R(value_reg));
SHR(16, R(CX), Imm8(8));
OR(16, R(value_reg), R(CX));
MOV(16, R(CX), R(value_reg));
SHR(16, R(CX), Imm8(4));
OR(16, R(value_reg), R(CX));
MOV(16, R(CX), R(value_reg));
SHR(16, R(CX), Imm8(2));
OR(16, R(value_reg), R(CX));
MOV(16, R(CX), R(value_reg));
SHR(16, R(CX), Imm8(1));
OR(16, R(value_reg), R(CX));
MOVZX(32,16,value_reg, R(value_reg));
#else
BSR(16, CX, R(value_reg));
FixupBranch undef = J_CC(CC_Z); //CX is written, but undefined
// HORRIBLE UGLINESS, someone please fix.
// See http://code.google.com/p/dolphin-emu/source/detail?r=3125
void DSPEmitter::dsp_increment_one(X64Reg ar, X64Reg wr, X64Reg wr_pow, X64Reg temp_reg)
{
// if ((tmp & tmb) == tmb)
MOV(16, R(temp_reg), R(ar));
AND(16, R(temp_reg), R(wr_pow));
CMP(16, R(temp_reg), R(wr_pow));
FixupBranch not_equal = J_CC(CC_NE);
MOV(32, R(value_reg), Imm32(2));
SHL(32, R(value_reg), R(CL));
SUB(32, R(value_reg), Imm32(1));
//don't waste an instruction on jumping over an effective noop
// tmp -= wr_reg
SUB(16, R(ar), R(wr));
FixupBranch end = J();
SetJumpTarget(not_equal);
// else tmp++
ADD(16, R(ar), Imm16(1));
SetJumpTarget(end);
SetJumpTarget(undef);
#endif
OR(16, R(value_reg), Imm16(1));
XOR(64, R(RCX), R(RCX));
}
// EAX = g_dsp.r[reg]
// EDX = g_dsp.r[DSP_REG_WR0 + reg]
//clobbers RCX
void DSPEmitter::increment_addr_reg(int reg)
{
/*
u16 ar = g_dsp.r[reg];
u16 wr = g_dsp.r[reg+8];
u16 nar = ar+1;
//this works, because nar^ar will have all the bits from the highest
//changed bit downwards set(true only for +1!)
//based on an idea by Mylek
if((nar^ar)>=((wr<<1)|1))
nar -= wr+1;
*/
// s16 tmp = g_dsp.r[reg];
#ifdef _M_IX86 // All32
MOV(16, R(EAX), M(&g_dsp.r[reg]));
MOV(16, R(EDX), M(&g_dsp.r[DSP_REG_WR0 + reg]));
MOV(16, R(AX), M(&g_dsp.r[reg]));
MOV(16, R(DX), M(&g_dsp.r[DSP_REG_WR0 + reg]));
#else
MOV(64, R(R11), ImmPtr(&g_dsp.r));
MOV(16, R(EAX), MDisp(R11,reg*2));
MOV(16, R(EDX), MDisp(R11,(DSP_REG_WR0 + reg)*2));
MOV(16, R(AX), MDisp(R11,reg*2));
MOV(16, R(DX), MDisp(R11,(DSP_REG_WR0 + reg)*2));
#endif
// ToMask(WR0), calculating it into EDI
MOV(16, R(EDI), R(EDX));
ToMask(EDI);
MOV(16,R(DI), R(AX));
ADD(16,R(AX), Imm16(1));
XOR(16,R(DI), R(AX));
MOV(16,R(SI), R(DX));
dsp_increment_one(EAX, EDX, EDI);
SHL(16,R(SI), Imm8(1));
OR(16,R(SI), Imm16(3));
CMP(16,R(DI), R(SI));
FixupBranch nowrap = J_CC(CC_L);
SUB(16,R(AX), R(DX));
SUB(16,R(AX), Imm16(1));
SetJumpTarget(nowrap);
// g_dsp.r[reg] = tmp;
#ifdef _M_IX86 // All32
MOV(16, M(&g_dsp.r[reg]), R(EAX));
MOV(16, M(&g_dsp.r[reg]), R(AX));
#else
MOV(64, R(R11), ImmPtr(&g_dsp.r));
MOV(16, MDisp(R11,reg*2), R(EAX));
MOV(16, MDisp(R11,reg*2), R(AX));
#endif
}
// See http://code.google.com/p/dolphin-emu/source/detail?r=3125
void DSPEmitter::dsp_decrement_one(X64Reg ar, X64Reg wr, X64Reg wr_pow, X64Reg temp_reg)
{
// compute min from wr_pow and ar
// min = (tmb+1-ar)&tmb;
LEA(16, temp_reg, MDisp(wr_pow, 1));
SUB(16, R(temp_reg), R(ar));
AND(16, R(temp_reg), R(wr_pow));
// wr < min
CMP(16, R(wr), R(temp_reg));
FixupBranch wr_lt_min = J_CC(CC_B);
// !min
TEST(16, R(temp_reg), R(temp_reg));
FixupBranch min_zero = J_CC(CC_Z);
// ar--;
SUB(16, R(ar), Imm16(1));
FixupBranch end = J();
// ar += wr;
SetJumpTarget(wr_lt_min);
SetJumpTarget(min_zero);
ADD(16, R(ar), R(wr));
SetJumpTarget(end);
}
// EAX = g_dsp.r[reg]
// EDX = g_dsp.r[DSP_REG_WR0 + reg]
//clobbers RCX
void DSPEmitter::decrement_addr_reg(int reg)
{
/*
u16 ar = g_dsp.r[reg];
u16 wr = g_dsp.r[reg+8];
u16 m = ToMask(wr) | 1;
u16 nar = ar-1;
if((ar&m) - 1 < m-wr)
nar += wr+1;
return nar;
*/
// s16 ar = g_dsp.r[reg];
#ifdef _M_IX86 // All32
MOV(16, R(EAX), M(&g_dsp.r[reg]));
MOV(16, R(EDX), M(&g_dsp.r[DSP_REG_WR0 + reg]));
MOV(16, R(AX), M(&g_dsp.r[reg]));
MOVZX(32, 16, EDX, M(&g_dsp.r[DSP_REG_WR0 + reg]));
#else
MOV(64, R(R11), ImmPtr(&g_dsp.r));
MOV(16, R(EAX), MDisp(R11,reg*2));
MOV(16, R(EDX), MDisp(R11,(DSP_REG_WR0 + reg)*2));
MOV(16, R(AX), MDisp(R11,reg*2));
MOVZX(32, 16, EDX, MDisp(R11,(DSP_REG_WR0 + reg)*2));
#endif
// ToMask(WR0), calculating it into EDI
MOV(16, R(EDI), R(EDX));
ToMask(EDI);
//u16 m = ToMask(wr) | 1;
MOV(16, R(DI), R(DX));
ToMask(DI);
dsp_decrement_one(EAX, EDX, EDI);
// g_dsp.r[reg] = tmp;
//u16 nar = ar-1;
MOV(16, R(CX), R(AX));
SUB(16, R(AX), Imm16(1));
//(ar&m) - 1
AND(32, R(ECX), R(EDI));
SUB(32, R(ECX), Imm32(1));
//m-wr
SUB(32, R(EDI), R(EDX));
CMP(32, R(ECX), R(EDI));
FixupBranch out1 = J_CC(CC_GE);
ADD(16,R(AX),R(DX));
ADD(16,R(AX),Imm16(1));
SetJumpTarget(out1);
// g_dsp.r[reg] = tmp;
#ifdef _M_IX86 // All32
MOV(16, M(&g_dsp.r[reg]), R(EAX));
MOV(16, M(&g_dsp.r[reg]), R(AX));
#else
MOV(64, R(R11), ImmPtr(&g_dsp.r));
MOV(16, MDisp(R11,reg*2), R(EAX));
MOV(16, MDisp(R11,reg*2), R(AX));
#endif
}
@ -156,72 +171,71 @@ void DSPEmitter::decrement_addr_reg(int reg)
// EDX = g_dsp.r[DSP_REG_WR0 + reg]
// EDI = tomask(EDX)
void DSPEmitter::increase_addr_reg(int reg)
{
#ifdef _M_IX86 // All32
MOVZX(32, 16, ECX, M(&g_dsp.r[DSP_REG_IX0 + reg]));
#else
MOV(64, R(R11), ImmPtr(&g_dsp.r));
MOVZX(32, 16, ECX, MDisp(R11,(DSP_REG_IX0 + reg)*2));
#endif
// IX0 == 0, bail out
TEST(16, R(ECX), R(ECX));
// code too long for a 5-byte jump
// TODO: optimize a bit, maybe merge loops?
FixupBranch end = J_CC(CC_Z, true);
{
/*
u16 ar = g_dsp.r[reg];
u16 wr = g_dsp.r[reg+8];
u16 ix = g_dsp.r[reg+4];
u16 m = ToMask(wr) | 1;
u16 nar = ar+ix;
if (ix >= 0) {
if((ar&m) + (ix&m) -(int)m-1 >= 0)
nar -= wr+1;
} else {
if((ar&m) + (ix&m) -(int)m-1 < m-wr)
nar += wr+1;
}
return nar;
*/
#ifdef _M_IX86 // All32
MOV(16, R(EAX), M(&g_dsp.r[reg]));
MOV(16, R(EDX), M(&g_dsp.r[DSP_REG_WR0 + reg]));
MOV(16, R(SI), M(&g_dsp.r[DSP_REG_IX0 + reg]));
MOV(16, R(AX), M(&g_dsp.r[reg]));
MOVZX(32, 16, EDX, M(&g_dsp.r[DSP_REG_WR0 + reg]));
#else
MOV(16, R(EAX), MDisp(R11,reg*2));
MOV(16, R(EDX), MDisp(R11,(DSP_REG_WR0 + reg)*2));
MOV(64, R(R11), ImmPtr(&g_dsp.r));
MOV(16, R(SI), MDisp(R11,(DSP_REG_IX0 + reg)*2));
MOV(16, R(AX), MDisp(R11,reg*2));
MOVZX(32, 16, EDX, MDisp(R11,(DSP_REG_WR0 + reg)*2));
#endif
// ToMask(WR0), calculating it into EDI
MOV(16, R(EDI), R(EDX));
ToMask(EDI);
//u16 m = ToMask(wr) | 1;
MOV(16, R(DI), R(DX));
ToMask(DI);
// IX0 > 0
// TODO: ToMask flushes flags set by TEST,
// needs another CMP here.
CMP(16, R(ECX), Imm16(0));
FixupBranch negative = J_CC(CC_L);
//u16 nar = ar+ix;
MOV(16, R(CX), R(AX));
ADD(16, R(AX), R(SI));
JumpTarget loop_pos = GetCodePtr();
//(ar&m) + (ix&m) -(int)m-1
AND(32, R(ECX), R(EDI));
AND(32, R(ESI), R(EDI));
ADD(32, R(ECX), R(ESI));
SUB(32, R(ECX), R(EDI));
SUB(32, R(ECX), Imm32(1));
// dsp_increment
dsp_increment_one(EAX, EDX, EDI);
TEST(16,R(SI), Imm16(0x8000));
FixupBranch negative = J_CC(CC_NZ);
SUB(16, R(ECX), Imm16(1)); // value--
#ifdef _M_IX86 // All32
CMP(16, M(&g_dsp.r[DSP_REG_IX0 + reg]), Imm16(127));
#else
MOV(64, R(R11), ImmPtr(&g_dsp.r));
CMP(16, MDisp(R11,(DSP_REG_IX0 + reg)*2), Imm16(127));
#endif
FixupBranch dbg = J_CC(CC_NE);
CMP(16, R(ECX), Imm16(1));
FixupBranch dbg2 = J_CC(CC_NE);
INT3();
SetJumpTarget(dbg2);
SetJumpTarget(dbg);
CMP(16, R(ECX), Imm16(0)); // value > 0
J_CC(CC_G, loop_pos);
FixupBranch end_pos = J();
CMP(32, R(ECX), Imm32(0));
FixupBranch out1 = J_CC(CC_L);
SUB(16,R(AX),R(DX));
SUB(16,R(AX),Imm16(1));
FixupBranch out2 = J();
// else, IX0 < 0
SetJumpTarget(negative);
JumpTarget loop_neg = GetCodePtr();
// dsp_decrement
dsp_decrement_one(EAX, EDX, EDI);
//m-wr
SUB(32, R(EDI), R(EDX));
CMP(32, R(ECX), R(EDI));
FixupBranch out3 = J_CC(CC_GE);
ADD(16,R(AX),R(DX));
ADD(16,R(AX),Imm16(1));
ADD(16, R(ECX), Imm16(1)); // value++
CMP(16, R(ECX), Imm16(0)); // value < 0
J_CC(CC_L, loop_neg);
SetJumpTarget(end_pos);
SetJumpTarget(out1);
SetJumpTarget(out2);
SetJumpTarget(out3);
// g_dsp.r[reg] = tmp;
#ifdef _M_IX86 // All32
@ -230,8 +244,6 @@ void DSPEmitter::increase_addr_reg(int reg)
MOV(64, R(R11), ImmPtr(&g_dsp.r));
MOV(16, MDisp(R11,reg*2), R(EAX));
#endif
SetJumpTarget(end);
}
// Decrease addr register according to the correspond ix register
@ -241,58 +253,68 @@ void DSPEmitter::increase_addr_reg(int reg)
// EDI = tomask(EDX)
void DSPEmitter::decrease_addr_reg(int reg)
{
#ifdef _M_IX86 // All32
MOV(16, R(ECX), M(&g_dsp.r[DSP_REG_IX0 + reg]));
#else
MOV(64, R(R11), ImmPtr(&g_dsp.r));
MOV(16, R(ECX), MDisp(R11,(DSP_REG_IX0 + reg)*2));
#endif
// IX0 == 0, bail out
TEST(16, R(ECX), R(ECX));
// code too long for a 5-byte jump
// TODO: optimize a bit, maybe merge loops?
FixupBranch end = J_CC(CC_Z, true);
/*
u16 ar = g_dsp.r[reg];
u16 wr = g_dsp.r[reg+8];
u16 ix = g_dsp.r[reg+4];
u16 m = ToMask(wr) | 1;
u16 nar = ar-ix; //!!
if ((u16)ix > 0x8000) { // equiv: ix < 0 && ix != -0x8000 //!!
if((ar&m) - (int)(ix&m) >= 0) //!!
nar -= wr+1;
} else {
if((ar&m) - (int)(ix&m) < m-wr) //!!
nar += wr+1;
}
return nar;
*/
#ifdef _M_IX86 // All32
MOV(16, R(EAX), M(&g_dsp.r[reg]));
MOV(16, R(EDX), M(&g_dsp.r[DSP_REG_WR0 + reg]));
MOV(16, R(SI), M(&g_dsp.r[DSP_REG_IX0 + reg]));
MOV(16, R(AX), M(&g_dsp.r[reg]));
MOVZX(32, 16, EDX, M(&g_dsp.r[DSP_REG_WR0 + reg]));
#else
MOV(16, R(EAX), MDisp(R11,reg*2));
MOV(16, R(EDX), MDisp(R11,(DSP_REG_WR0 + reg)*2));
MOV(64, R(R11), ImmPtr(&g_dsp.r));
MOV(16, R(SI), MDisp(R11,(DSP_REG_IX0 + reg)*2));
MOV(16, R(AX), MDisp(R11,reg*2));
MOVZX(32, 16, EDX, MDisp(R11,(DSP_REG_WR0 + reg)*2));
#endif
// ToMask(WR0), calculating it into EDI
MOV(16, R(EDI), R(EDX));
ToMask(EDI);
//u16 m = ToMask(wr) | 1;
MOV(16, R(DI), R(DX));
ToMask(DI);
// IX0 > 0
// TODO: ToMask flushes flags set by TEST,
// needs another CMP here.
CMP(16, R(ECX), Imm16(0));
FixupBranch negative = J_CC(CC_L);
//u16 nar = ar-ix;
MOV(16, R(CX), R(AX));
SUB(16, R(AX), R(SI));
JumpTarget loop_pos = GetCodePtr();
//(ar&m) + (ix&m)
AND(32, R(ECX), R(EDI));
AND(32, R(ESI), R(EDI));
SUB(32, R(ECX), R(ESI));
// dsp_decrement
dsp_decrement_one(EAX, EDX, EDI);
CMP(16,R(SI), Imm16(0x8000));
FixupBranch negative = J_CC(CC_BE);
SUB(16, R(ECX), Imm16(1)); // value--
CMP(16, R(ECX), Imm16(0)); // value > 0
J_CC(CC_G, loop_pos);
FixupBranch end_pos = J();
CMP(32, R(ECX), Imm32(0));
FixupBranch out1 = J_CC(CC_L);
SUB(16,R(AX),R(DX));
SUB(16,R(AX),Imm16(1));
FixupBranch out2 = J();
// else, IX0 < 0
SetJumpTarget(negative);
JumpTarget loop_neg = GetCodePtr();
// dsp_increment
dsp_increment_one(EAX, EDX, EDI);
//m-wr
SUB(32, R(EDI), R(EDX));
CMP(32, R(ECX), R(EDI));
FixupBranch out3 = J_CC(CC_GE);
ADD(16,R(AX),R(DX));
ADD(16,R(AX),Imm16(1));
ADD(16, R(ECX), Imm16(1)); // value++
CMP(16, R(ECX), Imm16(0)); // value < 0
J_CC(CC_L, loop_neg);
SetJumpTarget(end_pos);
SetJumpTarget(out1);
SetJumpTarget(out2);
SetJumpTarget(out3);
// g_dsp.r[reg] = tmp;
#ifdef _M_IX86 // All32
@ -301,8 +323,6 @@ void DSPEmitter::decrease_addr_reg(int reg)
MOV(64, R(R11), ImmPtr(&g_dsp.r));
MOV(16, MDisp(R11,reg*2), R(EAX));
#endif
SetJumpTarget(end);
}