archive/dolphin
1
0
Fork 0
mirror of https://github.com/dolphin-emu/dolphin.git synced 2024-09-21 03:41:42 +02:00
Code Issues Projects Releases Packages Wiki Activity

Merge pull request #9857 from JosJuice/jitarm64-cr-analysis

Browse source 
PPCAnalyst: Allow more reordering of CR operations
This commit is contained in:
Mai 2023-11-28 21:01:07 -05:00 committed by GitHub
commit a7216a3035
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
7 changed files with 113 additions and 73 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3
Source/Core/Core/PowerPC/JitArm64/Jit.cpp
View file

@ -205,6 +205,7 @@ void JitArm64::FallBackToInterpreter(UGeckoInstruction inst)
// we must mark them as no longer discarded // we must mark them as no longer discarded
gpr.ResetRegisters(js.op->regsOut); gpr.ResetRegisters(js.op->regsOut);
fpr.ResetRegisters(js.op->GetFregsOut()); fpr.ResetRegisters(js.op->GetFregsOut());
gpr.ResetCRRegisters(js.op->crOut);
if (js.op->opinfo->flags & FL_ENDBLOCK) if (js.op->opinfo->flags & FL_ENDBLOCK)
{ {
@ -1222,9 +1223,11 @@ bool JitArm64::DoJit(u32 em_address, JitBlock* b, u32 nextPC)
{ {
gpr.DiscardRegisters(op.gprDiscardable); gpr.DiscardRegisters(op.gprDiscardable);
fpr.DiscardRegisters(op.fprDiscardable); fpr.DiscardRegisters(op.fprDiscardable);
gpr.DiscardCRRegisters(op.crDiscardable);
} }
gpr.StoreRegisters(~op.gprInUse & (op.regsIn | op.regsOut)); gpr.StoreRegisters(~op.gprInUse & (op.regsIn | op.regsOut));
fpr.StoreRegisters(~op.fprInUse & (op.fregsIn | op.GetFregsOut())); fpr.StoreRegisters(~op.fprInUse & (op.fregsIn | op.GetFregsOut()));
gpr.StoreCRRegisters(~op.crInUse & (op.crIn | op.crOut));
if (opinfo->flags & FL_LOADSTORE) if (opinfo->flags & FL_LOADSTORE)
++js.numLoadStoreInst; ++js.numLoadStoreInst;

21
Source/Core/Core/PowerPC/JitArm64/JitArm64_RegCache.cpp
View file

@ -129,6 +129,8 @@ void Arm64RegCache::DiscardRegister(size_t preg)
{ {
OpArg& reg = m_guest_registers[preg]; OpArg& reg = m_guest_registers[preg];
ARM64Reg host_reg = reg.GetReg(); ARM64Reg host_reg = reg.GetReg();
if (!IsVector(host_reg))
host_reg = EncodeRegTo32(host_reg);
reg.Discard(); reg.Discard();
if (host_reg != ARM64Reg::INVALID_REG) if (host_reg != ARM64Reg::INVALID_REG)
@ -288,6 +290,25 @@ void Arm64GPRCache::FlushCRRegisters(BitSet8 regs, bool maintain_state, ARM64Reg
} }
} }
void Arm64GPRCache::DiscardCRRegisters(BitSet8 regs)
{
for (int i : regs)
DiscardRegister(GUEST_CR_OFFSET + i);
}
void Arm64GPRCache::ResetCRRegisters(BitSet8 regs)
{
for (int i : regs)
{
OpArg& reg = m_guest_registers[GUEST_CR_OFFSET + i];
ARM64Reg host_reg = reg.GetReg();
ASSERT_MSG(DYNA_REC, host_reg == ARM64Reg::INVALID_REG,
"Attempted to reset a loaded register (did you mean to flush it?)");
reg.Flush();
}
}
void Arm64GPRCache::Flush(FlushMode mode, ARM64Reg tmp_reg) void Arm64GPRCache::Flush(FlushMode mode, ARM64Reg tmp_reg)
{ {
FlushRegisters(BitSet32(0xFFFFFFFF), mode == FlushMode::MaintainState, tmp_reg); FlushRegisters(BitSet32(0xFFFFFFFF), mode == FlushMode::MaintainState, tmp_reg);

3
Source/Core/Core/PowerPC/JitArm64/JitArm64_RegCache.h
View file

@ -335,6 +335,9 @@ public:
FlushCRRegisters(regs, false, tmp_reg); FlushCRRegisters(regs, false, tmp_reg);
} }
void DiscardCRRegisters(BitSet8 regs);
void ResetCRRegisters(BitSet8 regs);
protected: protected:
// Get the order of the host registers // Get the order of the host registers
void GetAllocationOrder() override; void GetAllocationOrder() override;

94
Source/Core/Core/PowerPC/PPCAnalyst.cpp
View file

@ -222,26 +222,13 @@ bool PPCAnalyzer::CanSwapAdjacentOps(const CodeOp& a, const CodeOp& b) const
// [1] https://bugs.dolphin-emu.org/issues/5864#note-7 // [1] https://bugs.dolphin-emu.org/issues/5864#note-7
if (a.canCauseException || b.canCauseException) if (a.canCauseException || b.canCauseException)
return false; return false;
if (a_flags & FL_ENDBLOCK) if (a_flags & (FL_ENDBLOCK | FL_TIMER | FL_NO_REORDER | FL_SET_OE))
return false; return false;
if (b_flags & (FL_SET_CRx | FL_ENDBLOCK | FL_TIMER | FL_EVIL | FL_SET_OE)) if (b_flags & (FL_ENDBLOCK | FL_TIMER | FL_NO_REORDER | FL_SET_OE))
return false;
if ((b_flags & (FL_RC_BIT | FL_RC_BIT_F)) && (b.inst.Rc))
return false; return false;
if ((a_flags & (FL_SET_CA | FL_READ_CA)) && (b_flags & (FL_SET_CA | FL_READ_CA))) if ((a_flags & (FL_SET_CA | FL_READ_CA)) && (b_flags & (FL_SET_CA | FL_READ_CA)))
return false; return false;
switch (b.inst.OPCD)
{
case 16:
case 18:
// branches. Do not swap.
case 17: // sc
case 46: // lmw
case 19: // table19 - lots of tricky stuff
return false;
}
// For now, only integer ops are acceptable. // For now, only integer ops are acceptable.
if (b_info->type != OpType::Integer) if (b_info->type != OpType::Integer)
return false; return false;
@ -249,16 +236,22 @@ bool PPCAnalyzer::CanSwapAdjacentOps(const CodeOp& a, const CodeOp& b) const
// Check that we have no register collisions. // Check that we have no register collisions.
// That is, check that none of b's outputs matches any of a's inputs, // That is, check that none of b's outputs matches any of a's inputs,
// and that none of a's outputs matches any of b's inputs. // and that none of a's outputs matches any of b's inputs.
// The latter does not apply if a is a cmp, of course, but doesn't hurt to check.
// register collision: b outputs to one of a's inputs // register collision: b outputs to one of a's inputs
if (b.regsOut & a.regsIn) if (b.regsOut & a.regsIn)
return false; return false;
if (b.crOut & a.crIn)
return false;
// register collision: a outputs to one of b's inputs // register collision: a outputs to one of b's inputs
if (a.regsOut & b.regsIn) if (a.regsOut & b.regsIn)
return false; return false;
if (a.crOut & b.crIn)
return false;
// register collision: b outputs to one of a's outputs (overwriting it) // register collision: b outputs to one of a's outputs (overwriting it)
if (b.regsOut & a.regsOut) if (b.regsOut & a.regsOut)
return false; return false;
if (b.crOut & a.crOut)
return false;
return true; return true;
} }
@ -451,12 +444,6 @@ void FindFunctions(const Core::CPUThreadGuard& guard, u32 startAddr, u32 endAddr
unniceSize); unniceSize);
} }
static bool isCmp(const CodeOp& a)
{
return (a.inst.OPCD == 10 || a.inst.OPCD == 11) ||
(a.inst.OPCD == 31 && (a.inst.SUBOP10 == 0 || a.inst.SUBOP10 == 32));
}
static bool isCarryOp(const CodeOp& a) static bool isCarryOp(const CodeOp& a)
{ {
return (a.opinfo->flags & FL_SET_CA) && !(a.opinfo->flags & FL_SET_OE) && return (a.opinfo->flags & FL_SET_CA) && !(a.opinfo->flags & FL_SET_OE) &&
@ -506,7 +493,7 @@ void PPCAnalyzer::ReorderInstructionsCore(u32 instructions, CodeOp* code, bool r
// Reorder integer compares, rlwinm., and carry-affecting ops // Reorder integer compares, rlwinm., and carry-affecting ops
// (if we add more merged branch instructions, add them here!) // (if we add more merged branch instructions, add them here!)
if ((type == ReorderType::CROR && isCror(a)) || (type == ReorderType::Carry && isCarryOp(a)) || if ((type == ReorderType::CROR && isCror(a)) || (type == ReorderType::Carry && isCarryOp(a)) ||
(type == ReorderType::CMP && (isCmp(a) || a.outputCR[0]))) (type == ReorderType::CMP && (type == ReorderType::CMP && a.crOut)))
{ {
// once we're next to a carry instruction, don't move away! // once we're next to a carry instruction, don't move away!
if (type == ReorderType::Carry && i != start) if (type == ReorderType::Carry && i != start)
@ -544,11 +531,6 @@ void PPCAnalyzer::ReorderInstructionsCore(u32 instructions, CodeOp* code, bool r
void PPCAnalyzer::ReorderInstructions(u32 instructions, CodeOp* code) const void PPCAnalyzer::ReorderInstructions(u32 instructions, CodeOp* code) const
{ {
// Reorder cror instructions upwards (e.g. towards an fcmp). Technically we should be more
// picky about this, but cror seems to almost solely be used for this purpose in real code.
// Additionally, the other boolean ops seem to almost never be used.
if (HasOption(OPTION_CROR_MERGE))
ReorderInstructionsCore(instructions, code, true, ReorderType::CROR);
// For carry, bubble instructions *towards* each other; one direction often isn't enough // For carry, bubble instructions *towards* each other; one direction often isn't enough
// to get pairs like addc/adde next to each other. // to get pairs like addc/adde next to each other.
if (HasOption(OPTION_CARRY_MERGE)) if (HasOption(OPTION_CARRY_MERGE))
@ -556,8 +538,16 @@ void PPCAnalyzer::ReorderInstructions(u32 instructions, CodeOp* code) const
ReorderInstructionsCore(instructions, code, false, ReorderType::Carry); ReorderInstructionsCore(instructions, code, false, ReorderType::Carry);
ReorderInstructionsCore(instructions, code, true, ReorderType::Carry); ReorderInstructionsCore(instructions, code, true, ReorderType::Carry);
} }
// Reorder instructions which write to CR (typically compare instructions) towards branches.
if (HasOption(OPTION_BRANCH_MERGE)) if (HasOption(OPTION_BRANCH_MERGE))
ReorderInstructionsCore(instructions, code, false, ReorderType::CMP); ReorderInstructionsCore(instructions, code, false, ReorderType::CMP);
// Reorder cror instructions upwards (e.g. towards an fcmp). Technically we should be more
// picky about this, but cror seems to almost solely be used for this purpose in real code.
// Additionally, the other boolean ops seem to almost never be used.
if (HasOption(OPTION_CROR_MERGE))
ReorderInstructionsCore(instructions, code, true, ReorderType::CROR);
} }
void PPCAnalyzer::SetInstructionStats(CodeBlock* block, CodeOp* code, void PPCAnalyzer::SetInstructionStats(CodeBlock* block, CodeOp* code,
@ -570,23 +560,40 @@ void PPCAnalyzer::SetInstructionStats(CodeBlock* block, CodeOp* code,
block->m_fpa->any = true; block->m_fpa->any = true;
} }
code->wantsCR = BitSet8(0); code->crIn = BitSet8(0);
if (opinfo->flags & FL_READ_ALL_CR) if (opinfo->flags & FL_READ_ALL_CR)
code->wantsCR = BitSet8(0xFF); {
code->crIn = BitSet8(0xFF);
}
else if (opinfo->flags & FL_READ_CRn) else if (opinfo->flags & FL_READ_CRn)
code->wantsCR[code->inst.CRFS] = true; {
code->crIn[code->inst.CRFS] = true;
}
else if (opinfo->flags & FL_READ_CR_BI) else if (opinfo->flags & FL_READ_CR_BI)
code->wantsCR[code->inst.BI] = true; {
code->crIn[code->inst.BI] = true;
}
else if (opinfo->type == OpType::CR)
{
code->crIn[code->inst.CRBA >> 2] = true;
code->crIn[code->inst.CRBB >> 2] = true;
code->outputCR = BitSet8(0); // CR instructions only write to one bit of the destination CR,
// so treat the other three bits of the destination as inputs
code->crIn[code->inst.CRBD >> 2] = true;
}
code->crOut = BitSet8(0);
if (opinfo->flags & FL_SET_ALL_CR) if (opinfo->flags & FL_SET_ALL_CR)
code->outputCR = BitSet8(0xFF); code->crOut = BitSet8(0xFF);
else if (opinfo->flags & FL_SET_CRn) else if (opinfo->flags & FL_SET_CRn)
code->outputCR[code->inst.CRFD] = true; code->crOut[code->inst.CRFD] = true;
else if ((opinfo->flags & FL_SET_CR0) || ((opinfo->flags & FL_RC_BIT) && code->inst.Rc)) else if ((opinfo->flags & FL_SET_CR0) || ((opinfo->flags & FL_RC_BIT) && code->inst.Rc))
code->outputCR[0] = true; code->crOut[0] = true;
else if ((opinfo->flags & FL_SET_CR1) || ((opinfo->flags & FL_RC_BIT_F) && code->inst.Rc)) else if ((opinfo->flags & FL_SET_CR1) || ((opinfo->flags & FL_RC_BIT_F) && code->inst.Rc))
code->outputCR[1] = true; code->crOut[1] = true;
else if (opinfo->type == OpType::CR)
code->crOut[code->inst.CRBD >> 2] = true;
code->wantsFPRF = (opinfo->flags & FL_READ_FPRF) != 0; code->wantsFPRF = (opinfo->flags & FL_READ_FPRF) != 0;
code->outputFPRF = (opinfo->flags & FL_SET_FPRF) != 0; code->outputFPRF = (opinfo->flags & FL_SET_FPRF) != 0;
@ -955,9 +962,9 @@ u32 PPCAnalyzer::Analyze(u32 address, CodeBlock* block, CodeBuffer* buffer,
// Scan for flag dependencies; assume the next block (or any branch that can leave the block) // Scan for flag dependencies; assume the next block (or any branch that can leave the block)
// wants flags, to be safe. // wants flags, to be safe.
BitSet8 wantsCR = BitSet8(0xFF);
bool wantsFPRF = true; bool wantsFPRF = true;
bool wantsCA = true; bool wantsCA = true;
BitSet8 crInUse, crDiscardable;
BitSet32 gprBlockInputs, gprInUse, fprInUse, gprDiscardable, fprDiscardable, fprInXmm; BitSet32 gprBlockInputs, gprInUse, fprInUse, gprDiscardable, fprDiscardable, fprInXmm;
for (int i = block->m_num_instructions - 1; i >= 0; i--) for (int i = block->m_num_instructions - 1; i >= 0; i--)
{ {
@ -974,27 +981,26 @@ u32 PPCAnalyzer::Analyze(u32 address, CodeBlock* block, CodeBuffer* buffer,
const bool hle = !!HLE::TryReplaceFunction(op.address); const bool hle = !!HLE::TryReplaceFunction(op.address);
const bool may_exit_block = hle || op.canEndBlock || op.canCauseException; const bool may_exit_block = hle || op.canEndBlock || op.canCauseException;
const BitSet8 opWantsCR = op.wantsCR;
const bool opWantsFPRF = op.wantsFPRF; const bool opWantsFPRF = op.wantsFPRF;
const bool opWantsCA = op.wantsCA; const bool opWantsCA = op.wantsCA;
op.wantsCR = wantsCR | BitSet8(may_exit_block ? 0xFF : 0);
op.wantsFPRF = wantsFPRF || may_exit_block; op.wantsFPRF = wantsFPRF || may_exit_block;
op.wantsCA = wantsCA || may_exit_block; op.wantsCA = wantsCA || may_exit_block;
wantsCR |= opWantsCR | BitSet8(may_exit_block ? 0xFF : 0);
wantsFPRF |= opWantsFPRF || may_exit_block; wantsFPRF |= opWantsFPRF || may_exit_block;
wantsCA |= opWantsCA || may_exit_block; wantsCA |= opWantsCA || may_exit_block;
wantsCR &= ~op.outputCR | opWantsCR;
wantsFPRF &= !op.outputFPRF || opWantsFPRF; wantsFPRF &= !op.outputFPRF || opWantsFPRF;
wantsCA &= !op.outputCA || opWantsCA; wantsCA &= !op.outputCA || opWantsCA;
op.gprInUse = gprInUse; op.gprInUse = gprInUse;
op.fprInUse = fprInUse; op.fprInUse = fprInUse;
op.crInUse = crInUse;
op.gprDiscardable = gprDiscardable; op.gprDiscardable = gprDiscardable;
op.fprDiscardable = fprDiscardable; op.fprDiscardable = fprDiscardable;
op.crDiscardable = crDiscardable;
op.fprInXmm = fprInXmm; op.fprInXmm = fprInXmm;
gprBlockInputs &= ~op.regsOut; gprBlockInputs &= ~op.regsOut;
gprBlockInputs |= op.regsIn; gprBlockInputs |= op.regsIn;
gprInUse |= op.regsIn | op.regsOut; gprInUse |= op.regsIn | op.regsOut;
fprInUse |= op.fregsIn | op.GetFregsOut(); fprInUse |= op.fregsIn | op.GetFregsOut();
crInUse |= op.crIn | op.crOut;
if (strncmp(op.opinfo->opname, "stfd", 4)) if (strncmp(op.opinfo->opname, "stfd", 4))
fprInXmm |= op.fregsIn; fprInXmm |= op.fregsIn;
@ -1006,11 +1012,13 @@ u32 PPCAnalyzer::Analyze(u32 address, CodeBlock* block, CodeBuffer* buffer,
fprInXmm = BitSet32{}; fprInXmm = BitSet32{};
gprDiscardable = BitSet32{}; gprDiscardable = BitSet32{};
fprDiscardable = BitSet32{}; fprDiscardable = BitSet32{};
crDiscardable = BitSet8{};
} }
else if (op.canEndBlock || op.canCauseException) else if (op.canEndBlock || op.canCauseException)
{ {
gprDiscardable = BitSet32{}; gprDiscardable = BitSet32{};
fprDiscardable = BitSet32{}; fprDiscardable = BitSet32{};
crDiscardable = BitSet8{};
} }
else else
{ {
@ -1018,6 +1026,8 @@ u32 PPCAnalyzer::Analyze(u32 address, CodeBlock* block, CodeBuffer* buffer,
gprDiscardable &= ~op.regsIn; gprDiscardable &= ~op.regsIn;
fprDiscardable |= op.GetFregsOut(); fprDiscardable |= op.GetFregsOut();
fprDiscardable &= ~op.fregsIn; fprDiscardable &= ~op.fregsIn;
crDiscardable |= op.crOut;
crDiscardable &= ~op.crIn;
} }
} }

6
Source/Core/Core/PowerPC/PPCAnalyst.h
View file

@ -32,10 +32,12 @@ struct CodeOp // 16B
const GekkoOPInfo* opinfo = nullptr; const GekkoOPInfo* opinfo = nullptr;
u32 address = 0; u32 address = 0;
u32 branchTo = 0; // if UINT32_MAX, not a branch u32 branchTo = 0; // if UINT32_MAX, not a branch
BitSet32 regsOut;
BitSet32 regsIn; BitSet32 regsIn;
BitSet32 regsOut;
BitSet32 fregsIn; BitSet32 fregsIn;
s8 fregOut = 0; s8 fregOut = 0;
BitSet8 crIn;
BitSet8 crOut;
bool isBranchTarget = false; bool isBranchTarget = false;
bool branchUsesCtr = false; bool branchUsesCtr = false;
bool branchIsIdleLoop = false; bool branchIsIdleLoop = false;
@ -50,6 +52,8 @@ struct CodeOp // 16B
bool canCauseException = false; bool canCauseException = false;
bool skipLRStack = false; bool skipLRStack = false;
bool skip = false; // followed BL-s for example bool skip = false; // followed BL-s for example
BitSet8 crInUse;
BitSet8 crDiscardable;
// which registers are still needed after this instruction in this block // which registers are still needed after this instruction in this block
BitSet32 fprInUse; BitSet32 fprInUse;
BitSet32 gprInUse; BitSet32 gprInUse;

36
Source/Core/Core/PowerPC/PPCTables.cpp
View file

@ -87,8 +87,8 @@ constexpr std::array<GekkoOPTemplate, 54> s_primary_table{{
{38, "stb", OpType::Store, 1, FL_IN_A0 | FL_IN_S | FL_LOADSTORE}, {38, "stb", OpType::Store, 1, FL_IN_A0 | FL_IN_S | FL_LOADSTORE},
{39, "stbu", OpType::Store, 1, FL_OUT_A | FL_IN_A | FL_IN_S | FL_LOADSTORE}, {39, "stbu", OpType::Store, 1, FL_OUT_A | FL_IN_A | FL_IN_S | FL_LOADSTORE},
{46, "lmw", OpType::System, 11, FL_EVIL | FL_IN_A0 | FL_LOADSTORE}, {46, "lmw", OpType::System, 11, FL_IN_A0 | FL_LOADSTORE},
{47, "stmw", OpType::System, 11, FL_EVIL | FL_IN_A0 | FL_LOADSTORE}, {47, "stmw", OpType::System, 11, FL_IN_A0 | FL_LOADSTORE},
{48, "lfs", OpType::LoadFP, 1, FL_OUT_FLOAT_D | FL_IN_A | FL_USE_FPU | FL_LOADSTORE}, {48, "lfs", OpType::LoadFP, 1, FL_OUT_FLOAT_D | FL_IN_A | FL_USE_FPU | FL_LOADSTORE},
{49, "lfsu", OpType::LoadFP, 1, {49, "lfsu", OpType::LoadFP, 1,
@ -224,17 +224,17 @@ constexpr std::array<GekkoOPTemplate, 4> s_table4_3{{
constexpr std::array<GekkoOPTemplate, 13> s_table19{{ constexpr std::array<GekkoOPTemplate, 13> s_table19{{
{528, "bcctrx", OpType::Branch, 1, FL_ENDBLOCK | FL_READ_CR_BI}, {528, "bcctrx", OpType::Branch, 1, FL_ENDBLOCK | FL_READ_CR_BI},
{16, "bclrx", OpType::Branch, 1, FL_ENDBLOCK | FL_READ_CR_BI}, {16, "bclrx", OpType::Branch, 1, FL_ENDBLOCK | FL_READ_CR_BI},
{257, "crand", OpType::CR, 1, FL_EVIL}, {257, "crand", OpType::CR, 1, 0},
{129, "crandc", OpType::CR, 1, FL_EVIL}, {129, "crandc", OpType::CR, 1, 0},
{289, "creqv", OpType::CR, 1, FL_EVIL}, {289, "creqv", OpType::CR, 1, 0},
{225, "crnand", OpType::CR, 1, FL_EVIL}, {225, "crnand", OpType::CR, 1, 0},
{33, "crnor", OpType::CR, 1, FL_EVIL}, {33, "crnor", OpType::CR, 1, 0},
{449, "cror", OpType::CR, 1, FL_EVIL}, {449, "cror", OpType::CR, 1, 0},
{417, "crorc", OpType::CR, 1, FL_EVIL}, {417, "crorc", OpType::CR, 1, 0},
{193, "crxor", OpType::CR, 1, FL_EVIL}, {193, "crxor", OpType::CR, 1, 0},
{150, "isync", OpType::InstructionCache, 1, FL_EVIL}, {150, "isync", OpType::InstructionCache, 1, FL_NO_REORDER},
{0, "mcrf", OpType::System, 1, FL_EVIL | FL_SET_CRn | FL_READ_CRn}, {0, "mcrf", OpType::System, 1, FL_SET_CRn | FL_READ_CRn},
{50, "rfi", OpType::System, 2, FL_ENDBLOCK | FL_CHECKEXCEPTIONS | FL_PROGRAMEXCEPTION}, {50, "rfi", OpType::System, 2, FL_ENDBLOCK | FL_CHECKEXCEPTIONS | FL_PROGRAMEXCEPTION},
}}; }};
@ -324,12 +324,12 @@ constexpr std::array<GekkoOPTemplate, 107> s_table31{{
{790, "lhbrx", OpType::Load, 1, FL_OUT_D | FL_IN_A0B | FL_LOADSTORE}, {790, "lhbrx", OpType::Load, 1, FL_OUT_D | FL_IN_A0B | FL_LOADSTORE},
// Conditional load/store (Wii SMP) // Conditional load/store (Wii SMP)
{150, "stwcxd", OpType::Store, 1, FL_EVIL | FL_IN_S | FL_IN_A0B | FL_SET_CR0 | FL_LOADSTORE}, {150, "stwcxd", OpType::Store, 1, FL_IN_S | FL_IN_A0B | FL_SET_CR0 | FL_LOADSTORE},
{20, "lwarx", OpType::Load, 1, FL_EVIL | FL_OUT_D | FL_IN_A0B | FL_SET_CR0 | FL_LOADSTORE}, {20, "lwarx", OpType::Load, 1, FL_OUT_D | FL_IN_A0B | FL_SET_CR0 | FL_LOADSTORE},
// load string (Inst these) // load string (Inst these)
{533, "lswx", OpType::Load, 1, FL_EVIL | FL_IN_A0B | FL_OUT_D | FL_LOADSTORE}, {533, "lswx", OpType::Load, 1, FL_IN_A0B | FL_OUT_D | FL_LOADSTORE},
{597, "lswi", OpType::Load, 1, FL_EVIL | FL_IN_A0 | FL_OUT_D | FL_LOADSTORE}, {597, "lswi", OpType::Load, 1, FL_IN_A0 | FL_OUT_D | FL_LOADSTORE},
// store word // store word
{151, "stwx", OpType::Store, 1, FL_IN_S | FL_IN_A0B | FL_LOADSTORE}, {151, "stwx", OpType::Store, 1, FL_IN_S | FL_IN_A0B | FL_LOADSTORE},
@ -347,8 +347,8 @@ constexpr std::array<GekkoOPTemplate, 107> s_table31{{
{662, "stwbrx", OpType::Store, 1, FL_IN_S | FL_IN_A0B | FL_LOADSTORE}, {662, "stwbrx", OpType::Store, 1, FL_IN_S | FL_IN_A0B | FL_LOADSTORE},
{918, "sthbrx", OpType::Store, 1, FL_IN_S | FL_IN_A0B | FL_LOADSTORE}, {918, "sthbrx", OpType::Store, 1, FL_IN_S | FL_IN_A0B | FL_LOADSTORE},
{661, "stswx", OpType::Store, 1, FL_EVIL | FL_IN_A0B | FL_LOADSTORE}, {661, "stswx", OpType::Store, 1, FL_IN_A0B | FL_LOADSTORE},
{725, "stswi", OpType::Store, 1, FL_EVIL | FL_IN_A0 | FL_LOADSTORE}, {725, "stswi", OpType::Store, 1, FL_IN_A0 | FL_LOADSTORE},
// fp load/store // fp load/store
{535, "lfsx", OpType::LoadFP, 1, FL_OUT_FLOAT_D | FL_IN_A0B | FL_USE_FPU | FL_LOADSTORE}, {535, "lfsx", OpType::LoadFP, 1, FL_OUT_FLOAT_D | FL_IN_A0B | FL_USE_FPU | FL_LOADSTORE},

23
Source/Core/Core/PowerPC/PPCTables.h
View file

@ -36,18 +36,17 @@ enum InstructionFlags : u64
FL_OUT_AD = FL_OUT_A | FL_OUT_D, FL_OUT_AD = FL_OUT_A | FL_OUT_D,
FL_TIMER = (1ull << 15), // Used only for mftb. FL_TIMER = (1ull << 15), // Used only for mftb.
FL_CHECKEXCEPTIONS = (1ull << 16), // Used with rfi/rfid. FL_CHECKEXCEPTIONS = (1ull << 16), // Used with rfi/rfid.
FL_EVIL = FL_NO_REORDER = (1ull << 17), // Instruction should not be reordered by our optimizations.
(1ull << 17), // Historically used to refer to instructions that messed up Super Monkey Ball. FL_USE_FPU = (1ull << 18), // Used to indicate a floating point instruction.
FL_USE_FPU = (1ull << 18), // Used to indicate a floating point instruction. FL_LOADSTORE = (1ull << 19), // Used to indicate a load/store instruction.
FL_LOADSTORE = (1ull << 19), // Used to indicate a load/store instruction. FL_SET_FPRF = (1ull << 20), // Sets bits in the FPRF.
FL_SET_FPRF = (1ull << 20), // Sets bits in the FPRF. FL_READ_FPRF = (1ull << 21), // Reads bits from the FPRF.
FL_READ_FPRF = (1ull << 21), // Reads bits from the FPRF. FL_SET_OE = (1ull << 22), // Sets the overflow flag.
FL_SET_OE = (1ull << 22), // Sets the overflow flag. FL_IN_FLOAT_A = (1ull << 23), // frA is used as an input.
FL_IN_FLOAT_A = (1ull << 23), // frA is used as an input. FL_IN_FLOAT_B = (1ull << 24), // frB is used as an input.
FL_IN_FLOAT_B = (1ull << 24), // frB is used as an input. FL_IN_FLOAT_C = (1ull << 25), // frC is used as an input.
FL_IN_FLOAT_C = (1ull << 25), // frC is used as an input. FL_IN_FLOAT_S = (1ull << 26), // frS is used as an input.
FL_IN_FLOAT_S = (1ull << 26), // frS is used as an input. FL_IN_FLOAT_D = (1ull << 27), // frD is used as an input.
FL_IN_FLOAT_D = (1ull << 27), // frD is used as an input.
FL_IN_FLOAT_AB = FL_IN_FLOAT_A | FL_IN_FLOAT_B, FL_IN_FLOAT_AB = FL_IN_FLOAT_A | FL_IN_FLOAT_B,
FL_IN_FLOAT_AC = FL_IN_FLOAT_A | FL_IN_FLOAT_C, FL_IN_FLOAT_AC = FL_IN_FLOAT_A | FL_IN_FLOAT_C,
FL_IN_FLOAT_ABC = FL_IN_FLOAT_A | FL_IN_FLOAT_B | FL_IN_FLOAT_C, FL_IN_FLOAT_ABC = FL_IN_FLOAT_A | FL_IN_FLOAT_B | FL_IN_FLOAT_C,