Emulation/ppsspp
1
0
Fork 0
mirror of https://github.com/hrydgard/ppsspp.git synced 2025-04-02 11:01:50 -04:00
Code Issues Projects Releases Packages Wiki Activity

More RIP elimination

Browse source
This commit is contained in:
Henrik Rydgård 2017-07-05 15:06:44 +02:00
parent 7c1ae5b3e6
commit 7c3b37c561
8 changed files with 97 additions and 65 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
Common/x64Emitter.h
View file

@ -1073,7 +1073,7 @@ public:
class XCodeBlock : public CodeBlock<XEmitter> {
public:
void PoisonMemory(int offset) override;
bool RipAccessible(void *ptr) const {
bool RipAccessible(const void *ptr) const {
#ifdef _M_IX86
return true;
#else

2
Core/MIPS/MIPS.h
View file

@ -232,6 +232,8 @@ public:
u32 intBranchExit;
u32 jitBranchExit;
u32 savedPC;
static const u32 FCR0_VALUE = 0x00003351;
#if defined(PPSSPP_ARCH_X86) || defined(PPSSPP_ARCH_AMD64)

21
Core/MIPS/x86/Asm.cpp
View file

@ -114,15 +114,10 @@ void Jit::GenerateFixedCode(JitOptions &jo) {
// This is the most common situation.
TEST(32, MIPSSTATE_VAR(fcr31), Imm32(0x01000003));
FixupBranch skip = J_CC(CC_Z);
#ifdef _M_X64
// TODO: Move the hasSetRounding flag somewhere we can reach it through the context pointer, or something.
MOV(64, R(RAX), Imm64((uintptr_t)&js.hasSetRounding));
MOV(PTRBITS, R(RAX), ImmPtr(&js.hasSetRounding));
MOV(8, MatR(RAX), Imm8(1));
#else
MOV(8, M(&js.hasSetRounding), Imm8(1));
#endif
SetJumpTarget(skip);
RET();
}
@ -153,7 +148,12 @@ void Jit::GenerateFixedCode(JitOptions &jo) {
FixupBranch bailCoreState = J_CC(CC_S, true);
SetJumpTarget(skipToCoreStateCheck);
CMP(32, M(&coreState), Imm32(0));
if (RipAccessible((const void *)&coreState)) {
CMP(32, M(&coreState), Imm32(0));
} else {
MOV(PTRBITS, R(RAX), ImmPtr((const void *)&coreState));
CMP(32, MatR(RAX), Imm32(0));
}
FixupBranch badCoreState = J_CC(CC_NZ, true);
FixupBranch skipToRealDispatch2 = J(); //skip the sync and compare first time
@ -210,7 +210,12 @@ void Jit::GenerateFixedCode(JitOptions &jo) {
SetJumpTarget(bail);
SetJumpTarget(bailCoreState);
CMP(32, M(&coreState), Imm32(0));
if (RipAccessible((const void *)&coreState)) {
CMP(32, M(&coreState), Imm32(0));
} else {
MOV(PTRBITS, R(RAX), ImmPtr((const void *)&coreState));
CMP(32, MatR(RAX), Imm32(0));
}
J_CC(CC_Z, outerLoop, true);
SetJumpTarget(badCoreState);

13
Core/MIPS/x86/CompBranch.cpp
View file

@ -658,8 +658,6 @@ void Jit::Comp_Jump(MIPSOpcode op) {
js.compiling = false;
}
static u32 savedPC;
void Jit::Comp_JumpReg(MIPSOpcode op)
{
CONDITIONAL_LOG;
@ -725,21 +723,18 @@ void Jit::Comp_JumpReg(MIPSOpcode op)
MOV(32, R(EAX), gpr.R(rs));
}
FlushAll();
}
else
{
} else {
// Latch destination now - save it in memory.
gpr.MapReg(rs, true, false);
MOV(32, M(&savedPC), gpr.R(rs));
MOV(32, MIPSSTATE_VAR(savedPC), gpr.R(rs));
if (andLink)
gpr.SetImm(rd, GetCompilerPC() + 8);
CompileDelaySlot(DELAYSLOT_NICE);
MOV(32, R(EAX), M(&savedPC));
MOV(32, R(EAX), MIPSSTATE_VAR(savedPC));
FlushAll();
}
switch (op & 0x3f)
{
switch (op & 0x3f) {
case 8: //jr
break;
case 9: //jalr

6
Core/MIPS/x86/CompFPU.cpp
View file

@ -98,8 +98,6 @@ void Jit::Comp_FPU3op(MIPSOpcode op) {
}
}
static u32 MEMORY_ALIGNED16(ssLoadStoreTemp);
void Jit::Comp_FPULS(MIPSOpcode op) {
CONDITIONAL_DISABLE;
s32 offset = _IMM16;
@ -137,8 +135,8 @@ void Jit::Comp_FPULS(MIPSOpcode op) {
MOVSS(dest, fpr.RX(ft));
if (safe.PrepareSlowWrite())
{
MOVSS(M(&ssLoadStoreTemp), fpr.RX(ft));
safe.DoSlowWrite(safeMemFuncs.writeU32, M(&ssLoadStoreTemp));
MOVSS(MIPSSTATE_VAR(temp), fpr.RX(ft));
safe.DoSlowWrite(safeMemFuncs.writeU32, MIPSSTATE_VAR(temp));
}
safe.Finish();

3
Core/MIPS/x86/CompReplace.cpp
View file

@ -32,7 +32,8 @@ int Jit::Replace_fabsf() {
fpr.SpillLock(0, 12);
fpr.MapReg(0, false, true);
MOVSS(fpr.RX(0), fpr.R(12));
ANDPS(fpr.RX(0), M(&ssNoSignMask));
MOV(PTRBITS, R(RAX), ImmPtr(&ssNoSignMask));
ANDPS(fpr.RX(0), MatR(RAX));
fpr.ReleaseSpillLocks();
return 4; // Number of instructions in the MIPS function
}

108
Core/MIPS/x86/CompVFPU.cpp
View file

@ -15,6 +15,9 @@
// Official git repository and contact information can be found at
// https://github.com/hrydgard/ppsspp and http://www.ppsspp.org/.
// Table 13.10 in http://agner.org/optimize/optimizing_assembly.pdf is cool - generate constants with
// short instruction sequences. Surprisingly many are possible.
#include "ppsspp_config.h"
#if PPSSPP_ARCH(X86) || PPSSPP_ARCH(AMD64)
@ -66,9 +69,6 @@ const u32 MEMORY_ALIGNED16( noSignMask[4] ) = {0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFF
const u32 MEMORY_ALIGNED16( signBitAll[4] ) = {0x80000000, 0x80000000, 0x80000000, 0x80000000};
const u32 MEMORY_ALIGNED16( signBitLower[4] ) = {0x80000000, 0, 0, 0};
const float MEMORY_ALIGNED16( oneOneOneOne[4] ) = {1.0f, 1.0f, 1.0f, 1.0f};
const u32 MEMORY_ALIGNED16( solidOnes[4] ) = {0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF};
const u32 MEMORY_ALIGNED16( lowOnes[4] ) = {0xFFFFFFFF, 0x00000000, 0x00000000, 0x00000000};
const u32 MEMORY_ALIGNED16( lowZeroes[4] ) = {0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF};
const u32 MEMORY_ALIGNED16( fourinfnan[4] ) = {0x7F800000, 0x7F800000, 0x7F800000, 0x7F800000};
const float MEMORY_ALIGNED16( identityMatrix[4][4]) = { { 1.0f, 0, 0, 0 }, { 0, 1.0f, 0, 0 }, { 0, 0, 1.0f, 0 }, { 0, 0, 0, 1.0f} };
@ -176,13 +176,15 @@ void Jit::ApplyPrefixD(const u8 *vregs, VectorSize sz) {
ANDNPS(XMM0, fpr.V(vregs[i]));
// Retain a NAN in XMM0 (must be second operand.)
MOVSS(fpr.VX(vregs[i]), M(&one));
MOV(PTRBITS, R(TEMPREG), ImmPtr(&one));
MOVSS(fpr.VX(vregs[i]), MatR(TEMPREG));
MINSS(fpr.VX(vregs[i]), R(XMM0));
} else if (sat == 3) {
fpr.MapRegV(vregs[i], MAP_DIRTY);
// Check for < -1.0f, but careful of NANs.
MOVSS(XMM1, M(&minus_one));
MOV(PTRBITS, R(TEMPREG), ImmPtr(&minus_one));
MOVSS(XMM1, MatR(TEMPREG));
MOVSS(R(XMM0), fpr.VX(vregs[i]));
CMPLESS(XMM0, R(XMM1));
// If it was NOT less, the three ops below do nothing.
@ -192,7 +194,8 @@ void Jit::ApplyPrefixD(const u8 *vregs, VectorSize sz) {
ORPS(XMM0, R(XMM1));
// Retain a NAN in XMM0 (must be second operand.)
MOVSS(fpr.VX(vregs[i]), M(&one));
MOV(PTRBITS, R(TEMPREG), ImmPtr(&one));
MOVSS(fpr.VX(vregs[i]), MatR(TEMPREG));
MINSS(fpr.VX(vregs[i]), R(XMM0));
}
}
@ -236,12 +239,10 @@ void Jit::Comp_SV(MIPSOpcode op) {
JitSafeMem safe(this, rs, imm);
safe.SetFar();
OpArg src;
if (safe.PrepareRead(src, 4))
{
if (safe.PrepareRead(src, 4)) {
MOVSS(fpr.VX(vt), safe.NextFastAddress(0));
}
if (safe.PrepareSlowRead(safeMemFuncs.readU32))
{
if (safe.PrepareSlowRead(safeMemFuncs.readU32)) {
MOVD_xmm(fpr.VX(vt), R(EAX));
}
safe.Finish();
@ -260,14 +261,12 @@ void Jit::Comp_SV(MIPSOpcode op) {
JitSafeMem safe(this, rs, imm);
safe.SetFar();
OpArg dest;
if (safe.PrepareWrite(dest, 4))
{
if (safe.PrepareWrite(dest, 4)) {
MOVSS(safe.NextFastAddress(0), fpr.VX(vt));
}
if (safe.PrepareSlowWrite())
{
MOVSS(M(&ssLoadStoreTemp), fpr.VX(vt));
safe.DoSlowWrite(safeMemFuncs.writeU32, M(&ssLoadStoreTemp), 0);
if (safe.PrepareSlowWrite()) {
MOVSS(MIPSSTATE_VAR(temp), fpr.VX(vt));
safe.DoSlowWrite(safeMemFuncs.writeU32, MIPSSTATE_VAR(temp), 0);
}
safe.Finish();
@ -453,9 +452,9 @@ void Jit::Comp_SVQ(MIPSOpcode op) {
if (safe.PrepareSlowWrite()) {
MOVAPS(XMM0, fpr.VS(vregs));
for (int i = 0; i < 4; i++) {
MOVSS(M(&ssLoadStoreTemp), XMM0);
MOVSS(MIPSSTATE_VAR(temp), XMM0);
SHUFPS(XMM0, R(XMM0), _MM_SHUFFLE(3, 3, 2, 1));
safe.DoSlowWrite(safeMemFuncs.writeU32, M(&ssLoadStoreTemp), i * 4);
safe.DoSlowWrite(safeMemFuncs.writeU32, MIPSSTATE_VAR(temp), i * 4);
}
}
safe.Finish();
@ -476,8 +475,8 @@ void Jit::Comp_SVQ(MIPSOpcode op) {
}
if (safe.PrepareSlowWrite()) {
for (int i = 0; i < 4; i++) {
MOVSS(M(&ssLoadStoreTemp), fpr.VX(vregs[i]));
safe.DoSlowWrite(safeMemFuncs.writeU32, M(&ssLoadStoreTemp), i * 4);
MOVSS(MIPSSTATE_VAR(temp), fpr.VX(vregs[i]));
safe.DoSlowWrite(safeMemFuncs.writeU32, MIPSSTATE_VAR(temp), i * 4);
}
}
safe.Finish();
@ -508,7 +507,12 @@ void Jit::Comp_VVectorInit(MIPSOpcode op) {
if (type == 6) {
XORPS(fpr.VSX(dregs), fpr.VS(dregs));
} else if (type == 7) {
MOVAPS(fpr.VSX(dregs), M(&oneOneOneOne));
if (RipAccessible(&oneOneOneOne)) {
MOVAPS(fpr.VSX(dregs), M(&oneOneOneOne));
} else {
MOV(PTRBITS, R(TEMPREG), ImmPtr(&oneOneOneOne));
MOVAPS(fpr.VSX(dregs), MatR(TEMPREG));
}
} else {
DISABLE;
}
@ -522,7 +526,12 @@ void Jit::Comp_VVectorInit(MIPSOpcode op) {
XORPS(XMM0, R(XMM0));
break;
case 7: // v=ones; break; //vone
MOVSS(XMM0, M(&one));
if (RipAccessible(&one)) {
MOVSS(XMM0, M(&one));
} else {
MOV(PTRBITS, R(TEMPREG), ImmPtr(&one));
MOVSS(XMM0, MatR(TEMPREG));
}
break;
default:
DISABLE;
@ -558,7 +567,12 @@ void Jit::Comp_VIdt(MIPSOpcode op) {
}
XORPS(XMM0, R(XMM0));
MOVSS(XMM1, M(&one));
if (RipAccessible(&one)) {
MOVSS(XMM1, M(&one));
} else {
MOV(PTRBITS, R(TEMPREG), ImmPtr(&one));
MOVSS(XMM1, MatR(TEMPREG));
}
fpr.MapRegsV(dregs, sz, MAP_NOINIT | MAP_DIRTY);
switch (sz) {
case V_Pair:
@ -1426,13 +1440,16 @@ void Jit::Comp_Vcmp(MIPSOpcode op) {
// Finalize the comparison for ES/NS.
if (cond == VC_ES || cond == VC_NS) {
ANDPS(XMM0, M(&fourinfnan));
PCMPEQD(XMM0, M(&fourinfnan)); // Integer comparison
MOV(PTRBITS, R(TEMPREG), ImmPtr(&fourinfnan));
ANDPS(XMM0, MatR(TEMPREG));
PCMPEQD(XMM0, MatR(TEMPREG)); // Integer comparison
// It's inversed below for NS.
}
if (inverse) {
XORPS(XMM0, M(&solidOnes));
// The canonical way to generate a bunch of ones, see https://stackoverflow.com/questions/35085059/what-are-the-best-instruction-sequences-to-generate-vector-constants-on-the-fly
PCMPEQW(XMM1, R(XMM1));
XORPS(XMM0, R(XMM1));
}
ANDPS(XMM0, M(vcmpMask[n - 1]));
MOVAPS(M(vcmpResult), XMM0);
@ -1451,8 +1468,9 @@ void Jit::Comp_Vcmp(MIPSOpcode op) {
} else {
// Finalize the comparison for ES/NS.
if (cond == VC_ES || cond == VC_NS) {
ANDPS(XMM0, M(&fourinfnan));
PCMPEQD(XMM0, M(&fourinfnan)); // Integer comparison
MOV(PTRBITS, R(TEMPREG), ImmPtr(&fourinfnan));
ANDPS(XMM0, MatR(TEMPREG));
PCMPEQD(XMM0, MatR(TEMPREG)); // Integer comparison
// It's inversed below for NS.
}
@ -1983,26 +2001,22 @@ void Jit::Comp_Vocp(MIPSOpcode op) {
fpr.SimpleRegsV(dregs, sz, MAP_NOINIT | MAP_DIRTY);
X64Reg tempxregs[4];
for (int i = 0; i < n; ++i)
{
if (!IsOverlapSafeAllowS(dregs[i], i, n, sregs))
{
for (int i = 0; i < n; ++i) {
if (!IsOverlapSafeAllowS(dregs[i], i, n, sregs)) {
int reg = fpr.GetTempV();
fpr.MapRegV(reg, MAP_NOINIT | MAP_DIRTY);
fpr.SpillLockV(reg);
tempxregs[i] = fpr.VX(reg);
}
else
{
} else {
fpr.MapRegV(dregs[i], dregs[i] == sregs[i] ? MAP_DIRTY : MAP_NOINIT);
fpr.SpillLockV(dregs[i]);
tempxregs[i] = fpr.VX(dregs[i]);
}
}
MOVSS(XMM1, M(&one));
for (int i = 0; i < n; ++i)
{
MOV(PTRBITS, R(TEMPREG), ImmPtr(&one));
MOVSS(XMM1, MatR(TEMPREG));
for (int i = 0; i < n; ++i) {
MOVSS(XMM0, R(XMM1));
SUBSS(XMM0, fpr.V(sregs[i]));
MOVSS(tempxregs[i], R(XMM0));
@ -2274,13 +2288,23 @@ void Jit::Comp_VV2Op(MIPSOpcode op) {
MINSS(tempxregs[i], R(XMM0));
break;
case 16: // d[i] = 1.0f / s[i]; break; //vrcp
MOVSS(XMM0, M(&one));
if (RipAccessible(&one)) {
MOVSS(XMM0, M(&one));
} else {
MOV(PTRBITS, R(TEMPREG), ImmPtr(&one));
MOVSS(XMM0, MatR(TEMPREG));
}
DIVSS(XMM0, fpr.V(sregs[i]));
MOVSS(tempxregs[i], R(XMM0));
break;
case 17: // d[i] = 1.0f / sqrtf(s[i]); break; //vrsq
SQRTSS(XMM0, fpr.V(sregs[i]));
MOVSS(tempxregs[i], M(&one));
if (RipAccessible(&one)) {
MOVSS(tempxregs[i], M(&one));
} else {
MOV(PTRBITS, R(TEMPREG), ImmPtr(&one));
MOVSS(tempxregs[i], MatR(TEMPREG));
}
DIVSS(tempxregs[i], R(XMM0));
break;
case 18: // d[i] = sinf((float)M_PI_2 * s[i]); break; //vsin
@ -2306,7 +2330,9 @@ void Jit::Comp_VV2Op(MIPSOpcode op) {
MOVSS(tempxregs[i], M(&sincostemp[0]));
break;
case 24: // d[i] = -1.0f / s[i]; break; // vnrcp
MOVSS(XMM0, M(&minus_one));
// Rare so let's not bother checking for RipAccessible.
MOV(PTRBITS, R(TEMPREG), ImmPtr(&minus_one));
MOVSS(XMM0, MatR(TEMPREG));
DIVSS(XMM0, fpr.V(sregs[i]));
MOVSS(tempxregs[i], R(XMM0));
break;

7
Core/MIPS/x86/Jit.cpp
View file

@ -374,7 +374,12 @@ const u8 *Jit::DoJit(u32 em_address, JitBlock *b) {
// If we're rewinding, CORE_NEXTFRAME should not cause a rewind.
// It doesn't really matter either way if we're not rewinding.
// CORE_RUNNING is <= CORE_NEXTFRAME.
CMP(32, M(&coreState), Imm32(CORE_NEXTFRAME));
if (RipAccessible((const void *)coreState)) {
CMP(32, M(&coreState), Imm32(CORE_NEXTFRAME));
} else {
MOV(PTRBITS, R(RAX), ImmPtr((const void *)&coreState));
CMP(32, MatR(RAX), Imm32(CORE_NEXTFRAME));
}
FixupBranch skipCheck = J_CC(CC_LE);
if (js.afterOp & JitState::AFTER_REWIND_PC_BAD_STATE)
MOV(32, MIPSSTATE_VAR(pc), Imm32(GetCompilerPC()));