ShuriZma
/
dolphin
mirror of https://github.com/dolphin-emu/dolphin.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Make the (V)LDR/(V)STR instructions support negative offsets. This fixes a bug where Arm Jit couldn't load the top 33 FPRs. Also makes it so the core can access all GPRs, FPRs, and SPRs in ppcState. This increases VPS 15-20 on SSBM intro movie on ODROIDX

This commit is contained in:
Ryan Houdek 2013-03-06 01:51:24 +00:00
parent 2095641af0
commit d6558e1c31
13 changed files with 145 additions and 136 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

58
Source/Core/Common/Src/ArmEmitter.cpp
View File

@ -533,15 +533,15 @@ void ARMXEmitter::MRS (ARMReg dest)
Write32(condition | (16 << 20) | (15 << 16) | (dest << 12));
}
void ARMXEmitter::WriteStoreOp(u32 op, ARMReg dest, ARMReg src, Operand2 op2)
void ARMXEmitter::WriteStoreOp(u32 op, ARMReg dest, ARMReg src, s16 op2)
{
if (op2.GetData() == 0) // set the preindex bit, but not the W bit!
Write32(condition | 0x01800000 | (op << 20) | (dest << 16) | (src << 12) | op2.Imm12());
else
Write32(condition | (op << 20) | (3 << 23) | (dest << 16) | (src << 12) | op2.Imm12());
bool Index = op2 != 0 ? true : false;
bool Add = op2 >= 0 ? true : false;
u32 imm = abs(op2);
Write32(condition | (op << 20) | (Index << 24) | (Add << 23) | (dest << 16) | (src << 12) | imm);
}
void ARMXEmitter::STR (ARMReg dest, ARMReg src, Operand2 op) { WriteStoreOp(0x40, dest, src, op);}
void ARMXEmitter::STRB(ARMReg dest, ARMReg src, Operand2 op) { WriteStoreOp(0x44, dest, src, op);}
void ARMXEmitter::STR (ARMReg dest, ARMReg src, s16 op) { WriteStoreOp(0x40, dest, src, op);}
void ARMXEmitter::STRB(ARMReg dest, ARMReg src, s16 op) { WriteStoreOp(0x44, dest, src, op);}
void ARMXEmitter::STR (ARMReg dest, ARMReg base, ARMReg offset, bool Index, bool Add)
{
Write32(condition | (0x60 << 20) | (Index << 24) | (Add << 23) | (dest << 16) | (base << 12) | offset);
@ -564,13 +564,13 @@ void ARMXEmitter::SVC(Operand2 op)
Write32(condition | (0x0F << 24) | op.Imm24());
}
void ARMXEmitter::LDR (ARMReg dest, ARMReg src, Operand2 op) { WriteStoreOp(0x41, src, dest, op);}
void ARMXEmitter::LDR (ARMReg dest, ARMReg src, s16 op) { WriteStoreOp(0x41, src, dest, op);}
void ARMXEmitter::LDRH(ARMReg dest, ARMReg src, Operand2 op)
{
u8 Imm = op.Imm8();
Write32(condition | (0x05 << 20) | (src << 16) | (dest << 12) | ((Imm >> 4) << 8) | (0xB << 4) | (Imm & 0x0F));
}
void ARMXEmitter::LDRB(ARMReg dest, ARMReg src, Operand2 op) { WriteStoreOp(0x45, src, dest, op);}
void ARMXEmitter::LDRB(ARMReg dest, ARMReg src, s16 op) { WriteStoreOp(0x45, src, dest, op);}
void ARMXEmitter::LDR (ARMReg dest, ARMReg base, ARMReg offset, bool Index, bool Add)
{
@ -661,14 +661,18 @@ void ARMXEmitter::VSUB(IntegerSize Size, ARMReg Vd, ARMReg Vn, ARMReg Vm)
// VFP Specific
void ARMXEmitter::VLDR(ARMReg Dest, ARMReg Base, u16 offset)
void ARMXEmitter::VLDR(ARMReg Dest, ARMReg Base, s16 offset)
{
_assert_msg_(DYNA_REC, Dest >= S0 && Dest <= D31, "Passed Invalid dest register to VLDR");
_assert_msg_(DYNA_REC, Base <= R15, "Passed invalid Base register to VLDR");
_assert_msg_(DYNA_REC, (offset & 0xC03) == 0, "VLDR: Offset needs to be word aligned and small enough");
if (offset & 0xC03) {
ERROR_LOG(DYNA_REC, "VLDR: Bad offset %08x", offset);
bool Add = offset >= 0 ? true : false;
u32 imm = abs(offset);
_assert_msg_(DYNA_REC, (imm & 0xC03) == 0, "VLDR: Offset needs to be word aligned and small enough");
if (imm & 0xC03) {
ERROR_LOG(DYNA_REC, "VLDR: Bad offset %08x", imm);
}
bool single_reg = Dest < D0;
@ -677,24 +681,28 @@ void ARMXEmitter::VLDR(ARMReg Dest, ARMReg Base, u16 offset)
if (single_reg)
{
Write32(NO_COND | (0x1B << 23) | ((Dest & 0x1) << 22) | (1 << 20) | (Base << 16) \
| ((Dest & 0x1E) << 11) | (10 << 8) | (offset >> 2));
Write32(NO_COND | (0xD << 24) | (Add << 23) | ((Dest & 0x1) << 22) | (1 << 20) | (Base << 16) \
| ((Dest & 0x1E) << 11) | (10 << 8) | (imm >> 2));
}
else
{
Write32(NO_COND | (0x1B << 23) | ((Dest & 0x10) << 18) | (1 << 20) | (Base << 16) \
| ((Dest & 0xF) << 12) | (11 << 8) | (offset >> 2));
Write32(NO_COND | (0xD << 24) | (Add << 23) | ((Dest & 0x10) << 18) | (1 << 20) | (Base << 16) \
| ((Dest & 0xF) << 12) | (11 << 8) | (imm >> 2));
}
}
void ARMXEmitter::VSTR(ARMReg Src, ARMReg Base, u16 offset)
void ARMXEmitter::VSTR(ARMReg Src, ARMReg Base, s16 offset)
{
_assert_msg_(DYNA_REC, Src >= S0 && Src <= D31, "Passed invalid src register to VSTR");
_assert_msg_(DYNA_REC, Base <= R15, "Passed invalid base register to VSTR");
_assert_msg_(DYNA_REC, (offset & 0xC03) == 0, "VSTR: Offset needs to be word aligned");
if (offset & 0xC03) {
ERROR_LOG(DYNA_REC, "VSTR: Bad offset %08x", offset);
bool Add = offset >= 0 ? true : false;
u32 imm = abs(offset);
_assert_msg_(DYNA_REC, (imm & 0xC03) == 0, "VSTR: Offset needs to be word aligned and small enough");
if (imm & 0xC03) {
ERROR_LOG(DYNA_REC, "VSTR: Bad offset %08x", imm);
}
bool single_reg = Src < D0;
@ -703,14 +711,14 @@ void ARMXEmitter::VSTR(ARMReg Src, ARMReg Base, u16 offset)
if (single_reg)
{
Write32(NO_COND | (0x1B << 23) | ((Src & 0x1) << 22) | (Base << 16) \
| ((Src & 0x1E) << 11) | (10 << 8) | (offset >> 2));
Write32(NO_COND | (0xD << 24) | (Add << 23) | ((Src & 0x1) << 22) | (Base << 16) \
| ((Src & 0x1E) << 11) | (10 << 8) | (imm >> 2));
}
else
{
Write32(NO_COND | (0x1B << 23) | ((Src & 0x10) << 18) | (Base << 16) \
| ((Src & 0xF) << 12) | (11 << 8) | (offset >> 2));
Write32(NO_COND | (0xD << 24) | (Add << 23) | ((Src & 0x10) << 18) | (Base << 16) \
| ((Src & 0xF) << 12) | (11 << 8) | (imm >> 2));
}
}
void ARMXEmitter::VCMP(ARMReg Vd, ARMReg Vm)

44
Source/Core/Common/Src/ArmEmitter.h
View File

@ -208,7 +208,7 @@ public:
Value = base;
Type = TYPE_IMMSREG;
}
const u32 GetData()
u32 GetData()
{
switch(Type)
{
@ -225,45 +225,45 @@ public:
return 0;
}
}
const u32 IMMSR() // IMM shifted register
u32 IMMSR() // IMM shifted register
{
_assert_msg_(DYNA_REC, Type == TYPE_IMMSREG, "IMMSR must be imm shifted register");
return ((IndexOrShift & 0x1f) << 7 | (Shift << 5) | Value);
}
const u32 RSR() // Register shifted register
u32 RSR() // Register shifted register
{
_assert_msg_(DYNA_REC, Type == TYPE_RSR, "RSR must be RSR Of Course");
return (IndexOrShift << 8) | (Shift << 5) | 0x10 | Value;
}
const u32 Rm()
u32 Rm()
{
_assert_msg_(DYNA_REC, Type == TYPE_REG, "Rm must be with Reg");
return Value;
}
const u32 Imm5()
u32 Imm5()
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm5 not IMM value");
return ((Value & 0x0000001F) << 7);
}
const u32 Imm8()
u32 Imm8()
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm8Rot not IMM value");
return Value & 0xFF;
}
const u32 Imm8Rot() // IMM8 with Rotation
u32 Imm8Rot() // IMM8 with Rotation
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm8Rot not IMM value");
_assert_msg_(DYNA_REC, (Rotation & 0xE1) != 0, "Invalid Operand2: immediate rotation %u", Rotation);
return (1 << 25) | (Rotation << 7) | (Value & 0x000000FF);
}
const u32 Imm12()
u32 Imm12()
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm12 not IMM");
return (Value & 0x00000FFF);
}
const u32 Imm12Mod()
u32 Imm12Mod()
{
// This is a IMM12 with the top four bits being rotation and the
// bottom eight being a IMM. This is for instructions that need to
@ -273,32 +273,32 @@ public:
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm12Mod not IMM");
return ((Rotation & 0xF) << 8) | (Value & 0xFF);
}
const u32 Imm16()
u32 Imm16()
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm16 not IMM");
return ( (Value & 0xF000) << 4) | (Value & 0x0FFF);
}
const u32 Imm16Low()
u32 Imm16Low()
{
return Imm16();
}
const u32 Imm16High() // Returns high 16bits
u32 Imm16High() // Returns high 16bits
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm16 not IMM");
return ( ((Value >> 16) & 0xF000) << 4) | ((Value >> 16) & 0x0FFF);
}
const u32 Imm24()
u32 Imm24()
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm16 not IMM");
return (Value & 0x0FFFFFFF);
}
// NEON and ASIMD specific
const u32 Imm8ASIMD()
u32 Imm8ASIMD()
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm8ASIMD not IMM");
return ((Value & 0x80) << 17) | ((Value & 0x70) << 12) | (Value & 0xF);
}
const u32 Imm8VFP()
u32 Imm8VFP()
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm8VFP not IMM");
return ((Value & 0xF0) << 12) | (Value & 0xF);
@ -337,7 +337,7 @@ private:
u8 *lastCacheFlushEnd;
u32 condition;
void WriteStoreOp(u32 op, ARMReg dest, ARMReg src, Operand2 op2);
void WriteStoreOp(u32 op, ARMReg dest, ARMReg src, s16 op2);
void WriteRegStoreOp(u32 op, ARMReg dest, bool WriteBack, u16 RegList);
void WriteShiftedDataOp(u32 op, bool SetFlags, ARMReg dest, ARMReg src, ARMReg op2);
void WriteShiftedDataOp(u32 op, bool SetFlags, ARMReg dest, ARMReg src, Operand2 op2);
@ -467,16 +467,16 @@ public:
void MRS (ARMReg dest);
// Memory load/store operations
void LDR (ARMReg dest, ARMReg src, Operand2 op2 = 0);
void LDR (ARMReg dest, ARMReg src, s16 op2 = 0);
// Offset adds to the base register in LDR
void LDR (ARMReg dest, ARMReg base, ARMReg offset, bool Index, bool Add);
void LDRH(ARMReg dest, ARMReg src, Operand2 op = 0);
void LDRB(ARMReg dest, ARMReg src, Operand2 op2 = 0);
void STR (ARMReg dest, ARMReg src, Operand2 op2 = 0);
void LDRB(ARMReg dest, ARMReg src, s16 op2 = 0);
void STR (ARMReg dest, ARMReg src, s16 op2 = 0);
// Offset adds on to the destination register in STR
void STR (ARMReg dest, ARMReg base, ARMReg offset, bool Index, bool Add);
void STRB(ARMReg dest, ARMReg src, Operand2 op2 = 0);
void STRB(ARMReg dest, ARMReg src, s16 op2 = 0);
void STMFD(ARMReg dest, bool WriteBack, const int Regnum, ...);
void LDMFD(ARMReg dest, bool WriteBack, const int Regnum, ...);
@ -499,8 +499,8 @@ public:
void VSUB(IntegerSize Size, ARMReg Vd, ARMReg Vn, ARMReg Vm);
// VFP Only
void VLDR(ARMReg Dest, ARMReg Base, u16 offset);
void VSTR(ARMReg Src, ARMReg Base, u16 offset);
void VLDR(ARMReg Dest, ARMReg Base, s16 offset);
void VSTR(ARMReg Src, ARMReg Base, s16 offset);
void VCMP(ARMReg Vd, ARMReg Vm);
// Compares against zero
void VCMP(ARMReg Vd);

13
Source/Core/Core/Src/PowerPC/JitArm32/Jit.cpp
View File

@ -100,7 +100,7 @@ void JitArm::HLEFunction(UGeckoInstruction _inst)
MOVI2R(R1, _inst.hex);
QuickCallFunction(R14, (void*)&HLE::Execute);
ARMReg rA = gpr.GetReg();
LDR(rA, R9, STRUCT_OFF(PowerPC::ppcState, npc));
LDR(rA, R9, PPCSTATE_OFF(PowerPC::ppcState, npc));
WriteExitDestInR(rA);
}
@ -161,7 +161,7 @@ void JitArm::DoDownCount()
}
void JitArm::WriteExitDestInR(ARMReg Reg)
{
STR(R9, Reg, STRUCT_OFF(PowerPC::ppcState, pc));
STR(R9, Reg, PPCSTATE_OFF(PowerPC::ppcState, pc));
Cleanup();
DoDownCount();
MOVI2R(Reg, (u32)asm_routines.dispatcher);
@ -170,7 +170,7 @@ void JitArm::WriteExitDestInR(ARMReg Reg)
}
void JitArm::WriteRfiExitDestInR(ARMReg Reg)
{
STR(R9, Reg, STRUCT_OFF(PowerPC::ppcState, pc));
STR(R9, Reg, PPCSTATE_OFF(PowerPC::ppcState, pc));
Cleanup();
DoDownCount();
@ -209,7 +209,7 @@ void JitArm::WriteExit(u32 destination, int exit_num)
{
ARMReg A = gpr.GetReg(false);
MOVI2R(A, destination);
STR(R9, A, STRUCT_OFF(PowerPC::ppcState, pc));
STR(R9, A, PPCSTATE_OFF(PowerPC::ppcState, pc));
MOVI2R(A, (u32)asm_routines.dispatcher);
B(A);
}
@ -303,7 +303,6 @@ void JitArm::Break(UGeckoInstruction inst)
const u8* JitArm::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBlock *b)
{
int blockSize = code_buf->GetSize();
// Memory exception on instruction fetch
bool memory_exception = false;
@ -382,10 +381,10 @@ const u8* JitArm::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBlo
ARMReg C = gpr.GetReg();
Operand2 Shift(2, 10); // 1 << 13
MOVI2R(C, js.blockStart); // R3
LDR(A, R9, STRUCT_OFF(PowerPC::ppcState, msr));
LDR(A, R9, PPCSTATE_OFF(PowerPC::ppcState, msr));
TST(A, Shift);
FixupBranch b1 = B_CC(CC_NEQ);
STR(R9, C, STRUCT_OFF(PowerPC::ppcState, pc));
STR(R9, C, PPCSTATE_OFF(PowerPC::ppcState, pc));
MOVI2R(A, (u32)asm_routines.fpException);
B(A);
SetJumpTarget(b1);

2
Source/Core/Core/Src/PowerPC/JitArm32/Jit.h
View File

@ -47,7 +47,7 @@
if (Core::g_CoreStartupParameter.bJITOff || \
Core::g_CoreStartupParameter.bJIT##type##Off) \
{Default(inst); return;}
#define PPCSTATE_OFF(str, elem) ((s32)STRUCT_OFF(PowerPC::ppcState, elem) - (s32)STRUCT_OFF(PowerPC::ppcState, spr[0]))
class JitArm : public JitBase, public ArmGen::ARMXCodeBlock
{
private:

3
Source/Core/Core/Src/PowerPC/JitArm32/JitArmCache.cpp
View File

@ -36,10 +36,9 @@ using namespace ArmGen;
void JitArmBlockCache::WriteDestroyBlock(const u8* location, u32 address)
{
ARMXEmitter emit((u8 *)location);
emit.MOVI2R(R10, (u32)&PC);
emit.MOVI2R(R11, address);
emit.MOVI2R(R12, (u32)jit->GetAsmRoutines()->dispatcher);
emit.STR(R10, R11);
emit.STR(R9, R11, PPCSTATE_OFF(PowerPC::ppcState, pc));
emit.B(R12);
}

90
Source/Core/Core/Src/PowerPC/JitArm32/JitArm_Branch.cpp
View File

@ -50,9 +50,9 @@ void JitArm::sc(UGeckoInstruction inst)
ARMABI_MOVI2M((u32)&PC, js.compilerPC + 4); // Destroys R12 and R14
ARMReg rA = gpr.GetReg();
LDR(rA, R9, STRUCT_OFF(PowerPC::ppcState, Exceptions));
LDR(rA, R9, PPCSTATE_OFF(PowerPC::ppcState, Exceptions));
ORR(rA, rA, EXCEPTION_SYSCALL);
STR(R9, rA, STRUCT_OFF(PowerPC::ppcState, Exceptions));
STR(R9, rA, PPCSTATE_OFF(PowerPC::ppcState, Exceptions));
gpr.Unlock(rA);
WriteExceptionExit();
@ -78,25 +78,22 @@ void JitArm::rfi(UGeckoInstruction inst)
ARMReg rB = gpr.GetReg();
ARMReg rC = gpr.GetReg();
ARMReg rD = gpr.GetReg();
MOVI2R(rA, (u32)&MSR);
MOVI2R(rB, (~mask) & clearMSR13);
MOVI2R(rC, mask & clearMSR13);
LDR(rD, rA);
LDR(rD, R9, PPCSTATE_OFF(PowerPC::ppcState, msr));
AND(rD, rD, rB); // rD = Masked MSR
STR(rA, rD);
STR(R9, rD, PPCSTATE_OFF(PowerPC::ppcState, msr));
MOVI2R(rB, (u32)&SRR1);
LDR(rB, rB); // rB contains SRR1 here
LDR(rB, R9, PPCSTATE_OFF(PowerPC::ppcState, spr[SPR_SRR1])); // rB contains SRR1 here
AND(rB, rB, rC); // rB contains masked SRR1 here
ORR(rB, rD, rB); // rB = Masked MSR OR masked SRR1
STR(rA, rB); // STR rB in to rA
STR(R9, rB, PPCSTATE_OFF(PowerPC::ppcState, msr)); // STR rB in to rA
MOVI2R(rA, (u32)&SRR0);
LDR(rA, rA);
LDR(rA, R9, PPCSTATE_OFF(PowerPC::ppcState, spr[SPR_SRR0]));
gpr.Unlock(rB, rC, rD);
WriteRfiExitDestInR(rA); // rA gets unlocked here
@ -116,8 +113,13 @@ void JitArm::bx(UGeckoInstruction inst)
// We must always process the following sentence
// even if the blocks are merged by PPCAnalyst::Flatten().
if (inst.LK)
ARMABI_MOVI2M((u32)&LR, js.compilerPC + 4);
{
ARMReg rA = gpr.GetReg(false);
u32 Jumpto = js.compilerPC + 4;
MOVI2R(rA, Jumpto);
STR(R9, rA, PPCSTATE_OFF(PowerPC::ppcState, spr[SPR_LR]));
//ARMABI_MOVI2M((u32)&LR, js.compilerPC + 4);
}
// If this is not the last instruction of a block,
// we will skip the rest process.
// Because PPCAnalyst::Flatten() merged the blocks.
@ -164,10 +166,9 @@ void JitArm::bcx(UGeckoInstruction inst)
FixupBranch pCTRDontBranch;
if ((inst.BO & BO_DONT_DECREMENT_FLAG) == 0) // Decrement and test CTR
{
MOVI2R(rA, (u32)&CTR);
LDR(rB, rA);
LDR(rB, R9, PPCSTATE_OFF(PowerPC::ppcState, spr[SPR_CTR]));
SUBS(rB, rB, 1);
STR(rA, rB);
STR(R9, rB, PPCSTATE_OFF(PowerPC::ppcState, spr[SPR_CTR]));
//SUB(32, M(&CTR), Imm8(1));
if (inst.BO & BO_BRANCH_IF_CTR_0)
@ -179,7 +180,7 @@ void JitArm::bcx(UGeckoInstruction inst)
FixupBranch pConditionDontBranch;
if ((inst.BO & BO_DONT_CHECK_CONDITION) == 0) // Test a CR bit
{
LDRB(rA, R9, STRUCT_OFF(PowerPC::ppcState, cr_fast) + (inst.BI >> 2));
LDRB(rA, R9, PPCSTATE_OFF(PowerPC::ppcState, cr_fast) + (inst.BI >> 2));
TST(rA, 8 >> (inst.BI & 3));
//TEST(8, M(&PowerPC::ppcState.cr_fast[inst.BI >> 2]), Imm8(8 >> (inst.BI & 3)));
@ -188,10 +189,15 @@ void JitArm::bcx(UGeckoInstruction inst)
else
pConditionDontBranch = B_CC(CC_NEQ); // Not Zero
}
gpr.Unlock(rA, rB);
if (inst.LK)
ARMABI_MOVI2M((u32)&LR, js.compilerPC + 4); // Careful, destroys R14, R12
{
u32 Jumpto = js.compilerPC + 4;
MOVI2R(rB, Jumpto);
STR(R9, rB, PPCSTATE_OFF(PowerPC::ppcState, spr[SPR_LR]));
//ARMABI_MOVI2M((u32)&LR, js.compilerPC + 4); // Careful, destroys R14, R12
}
gpr.Unlock(rA, rB);
u32 destination;
if(inst.AA)
destination = SignExt16(inst.BD << 2);
@ -222,13 +228,18 @@ void JitArm::bcctrx(UGeckoInstruction inst)
// BO_2 == 1z1zz -> b always
//NPC = CTR & 0xfffffffc;
if(inst.LK_3)
ARMABI_MOVI2M((u32)&LR, js.compilerPC + 4);
ARMReg rA = gpr.GetReg();
ARMReg rB = gpr.GetReg();
MOVI2R(rA, (u32)&CTR);
if(inst.LK_3)
{
u32 Jumpto = js.compilerPC + 4;
MOVI2R(rA, Jumpto);
STR(R9, rA, PPCSTATE_OFF(PowerPC::ppcState, spr[SPR_LR]));
// ARMABI_MOVI2M((u32)&LR, js.compilerPC + 4);
}
MVN(rB, 0x3); // 0xFFFFFFFC
LDR(rA, rA);
LDR(rA, R9, PPCSTATE_OFF(PowerPC::ppcState, spr[SPR_CTR]));
AND(rA, rA, rB);
gpr.Unlock(rB);
WriteExitDestInR(rA);
@ -242,7 +253,7 @@ void JitArm::bcctrx(UGeckoInstruction inst)
ARMReg rA = gpr.GetReg();
ARMReg rB = gpr.GetReg();
LDRB(rA, R9, STRUCT_OFF(PowerPC::ppcState, cr_fast) + (inst.BI >> 2));
LDRB(rA, R9, PPCSTATE_OFF(PowerPC::ppcState, cr_fast) + (inst.BI >> 2));
TST(rA, 8 >> (inst.BI & 3));
CCFlags branch;
if (inst.BO_2 & BO_BRANCH_IF_TRUE)
@ -251,17 +262,14 @@ void JitArm::bcctrx(UGeckoInstruction inst)
branch = CC_NEQ;
FixupBranch b = B_CC(branch);
MOVI2R(rA, (u32)&CTR);
LDR(rA, rA);
LDR(rA, R9, PPCSTATE_OFF(PowerPC::ppcState, spr[SPR_CTR]));
MVN(rB, 0x3); // 0xFFFFFFFC
AND(rA, rA, rB);
if (inst.LK_3){
ARMReg rC = gpr.GetReg(false);
u32 Jumpto = js.compilerPC + 4;
MOVI2R(rB, (u32)&LR);
MOVI2R(rC, Jumpto);
STR(rB, rC);
MOVI2R(rB, Jumpto);
STR(R9, rB, PPCSTATE_OFF(PowerPC::ppcState, spr[SPR_LR]));
//ARMABI_MOVI2M((u32)&LR, js.compilerPC + 4);
}
gpr.Unlock(rB); // rA gets unlocked in WriteExitDestInR
@ -279,7 +287,11 @@ void JitArm::bclrx(UGeckoInstruction inst)
(inst.BO & (1 << 4)) && (inst.BO & (1 << 2))) {
if (inst.LK)
{
ARMABI_MOVI2M((u32)&LR, js.compilerPC + 4);
ARMReg rA = gpr.GetReg(false);
u32 Jumpto = js.compilerPC + 4;
MOVI2R(rA, Jumpto);
STR(R9, rA, PPCSTATE_OFF(PowerPC::ppcState, spr[SPR_LR]));
// ARMABI_MOVI2M((u32)&LR, js.compilerPC + 4);
}
return;
}
@ -291,10 +303,9 @@ void JitArm::bclrx(UGeckoInstruction inst)
FixupBranch pCTRDontBranch;
if ((inst.BO & BO_DONT_DECREMENT_FLAG) == 0) // Decrement and test CTR
{
MOVI2R(rA, (u32)&CTR);
LDR(rB, rA);
LDR(rB, R9, PPCSTATE_OFF(PowerPC::ppcState, spr[SPR_CTR]));
SUBS(rB, rB, 1);
STR(rA, rB);
STR(R9, rB, PPCSTATE_OFF(PowerPC::ppcState, spr[SPR_CTR]));
//SUB(32, M(&CTR), Imm8(1));
if (inst.BO & BO_BRANCH_IF_CTR_0)
@ -306,7 +317,7 @@ void JitArm::bclrx(UGeckoInstruction inst)
FixupBranch pConditionDontBranch;
if ((inst.BO & BO_DONT_CHECK_CONDITION) == 0) // Test a CR bit
{
LDRB(rA, R9, STRUCT_OFF(PowerPC::ppcState, cr_fast) + (inst.BI >> 2));
LDRB(rA, R9, PPCSTATE_OFF(PowerPC::ppcState, cr_fast) + (inst.BI >> 2));
TST(rA, 8 >> (inst.BI & 3));
//TEST(8, M(&PowerPC::ppcState.cr_fast[inst.BI >> 2]), Imm8(8 >> (inst.BI & 3)));
if (inst.BO & BO_BRANCH_IF_TRUE) // Conditional branch
@ -324,16 +335,13 @@ void JitArm::bclrx(UGeckoInstruction inst)
//MOV(32, R(EAX), M(&LR));
//AND(32, R(EAX), Imm32(0xFFFFFFFC));
MOVI2R(rA, (u32)&LR);
MVN(rB, 0x3); // 0xFFFFFFFC
LDR(rA, rA);
LDR(rA, R9, PPCSTATE_OFF(PowerPC::ppcState, spr[SPR_LR]));
AND(rA, rA, rB);
if (inst.LK){
ARMReg rC = gpr.GetReg(false);
u32 Jumpto = js.compilerPC + 4;
MOVI2R(rB, (u32)&LR);
MOVI2R(rC, Jumpto);
STR(rB, rC);
MOVI2R(rB, Jumpto);
STR(R9, rB, PPCSTATE_OFF(PowerPC::ppcState, spr[SPR_LR]));
//ARMABI_MOVI2M((u32)&LR, js.compilerPC + 4);
}
gpr.Unlock(rB); // rA gets unlocked in WriteExitDestInR

6
Source/Core/Core/Src/PowerPC/JitArm32/JitArm_Integer.cpp
View File

@ -40,7 +40,7 @@ void JitArm::GenerateRC(int cr) {
SetCC(CC_MI); MOV(rB, 0x8); // Result < 0
SetCC();
STRB(R9, rB, STRUCT_OFF(PowerPC::ppcState, cr_fast) + cr);
STRB(R9, rB, PPCSTATE_OFF(PowerPC::ppcState, cr_fast) + cr);
gpr.Unlock(rB);
}
void JitArm::ComputeRC(int cr) {
@ -51,7 +51,7 @@ void JitArm::ComputeRC(int cr) {
SetCC(CC_GT); MOV(rB, 0x4); // Result > 0
SetCC();
STRB(R9, rB, STRUCT_OFF(PowerPC::ppcState, cr_fast) + cr);
STRB(R9, rB, PPCSTATE_OFF(PowerPC::ppcState, cr_fast) + cr);
gpr.Unlock(rB);
}
@ -232,7 +232,7 @@ void JitArm::cmpli(UGeckoInstruction inst)
SetCC(CC_HI); MOV(rA, 0x4); // Result > 0
SetCC();
STRB(R9, rA, STRUCT_OFF(PowerPC::ppcState, cr_fast) + crf);
STRB(R9, rA, PPCSTATE_OFF(PowerPC::ppcState, cr_fast) + crf);
gpr.Unlock(rA);
}

10
Source/Core/Core/Src/PowerPC/JitArm32/JitArm_LoadStore.cpp
View File

@ -100,7 +100,7 @@ void JitArm::stwu(UGeckoInstruction inst)
// Check and set the update before writing since calling a function can
// mess with the "special registers R11+ which may cause some issues.
LDR(Function, R9, STRUCT_OFF(PowerPC::ppcState, Exceptions));
LDR(Function, R9, PPCSTATE_OFF(PowerPC::ppcState, Exceptions));
CMP(Function, EXCEPTION_DSI);
FixupBranch DoNotWrite = B_CC(CC_EQ);
MOV(RA, Addr);
@ -198,7 +198,7 @@ void JitArm::lbz(UGeckoInstruction inst)
ARMReg rA = gpr.GetReg();
ARMReg rB = gpr.GetReg();
ARMReg RD = gpr.R(inst.RD);
LDR(rA, R9, STRUCT_OFF(PowerPC::ppcState, Exceptions));
LDR(rA, R9, PPCSTATE_OFF(PowerPC::ppcState, Exceptions));
CMP(rA, EXCEPTION_DSI);
FixupBranch DoNotLoad = B_CC(CC_EQ);
#if FASTMEM
@ -245,7 +245,7 @@ void JitArm::lhz(UGeckoInstruction inst)
ARMReg rA = gpr.GetReg();
ARMReg rB = gpr.GetReg();
ARMReg RD = gpr.R(inst.RD);
LDR(rA, R9, STRUCT_OFF(PowerPC::ppcState, Exceptions));
LDR(rA, R9, PPCSTATE_OFF(PowerPC::ppcState, Exceptions));
CMP(rA, EXCEPTION_DSI);
FixupBranch DoNotLoad = B_CC(CC_EQ);
#if 0 // FASTMEM
@ -295,7 +295,7 @@ void JitArm::lwz(UGeckoInstruction inst)
ARMReg rA = gpr.GetReg();
ARMReg rB = gpr.GetReg();
ARMReg RD = gpr.R(inst.RD);
LDR(rA, R9, STRUCT_OFF(PowerPC::ppcState, Exceptions));
LDR(rA, R9, PPCSTATE_OFF(PowerPC::ppcState, Exceptions));
CMP(rA, EXCEPTION_DSI);
FixupBranch DoNotLoad = B_CC(CC_EQ);
@ -369,7 +369,7 @@ void JitArm::lwzx(UGeckoInstruction inst)
ARMReg RB = gpr.R(inst.RB);
ARMReg RD = gpr.R(inst.RD);
LDR(rA, R9, STRUCT_OFF(PowerPC::ppcState, Exceptions));
LDR(rA, R9, PPCSTATE_OFF(PowerPC::ppcState, Exceptions));
CMP(rA, EXCEPTION_DSI);
FixupBranch DoNotLoad = B_CC(CC_EQ);
#if FASTMEM

2
Source/Core/Core/Src/PowerPC/JitArm32/JitArm_LoadStoreFloating.cpp
View File

@ -40,7 +40,7 @@ void JitArm::lfs(UGeckoInstruction inst)
ARMReg rA = gpr.GetReg();
ARMReg rB = gpr.GetReg();
LDR(rA, R9, STRUCT_OFF(PowerPC::ppcState, Exceptions));
LDR(rA, R9, PPCSTATE_OFF(PowerPC::ppcState, Exceptions));
CMP(rA, EXCEPTION_DSI);
FixupBranch DoNotLoad = B_CC(CC_EQ);

13
Source/Core/Core/Src/PowerPC/JitArm32/JitArm_SystemRegisters.cpp
View File

@ -70,9 +70,7 @@ void JitArm::mtspr(UGeckoInstruction inst)
}
// OK, this is easy.
ARMReg rA = gpr.GetReg(false);
MOVI2R(rA, (u32)&PowerPC::ppcState.spr);
STR(rA, RD, iIndex * 4);
STR(R9, RD, PPCSTATE_OFF(PowerPC::ppcState, spr) + iIndex * 4);
}
void JitArm::mfspr(UGeckoInstruction inst)
@ -91,9 +89,7 @@ void JitArm::mfspr(UGeckoInstruction inst)
Default(inst);
return;
default:
ARMReg rA = gpr.GetReg(false);
MOVI2R(rA, (u32)&PowerPC::ppcState.spr);
LDR(RD, rA, iIndex * 4);
LDR(RD, R9, PPCSTATE_OFF(PowerPC::ppcState, spr) + iIndex * 4);
break;
}
}
@ -103,9 +99,6 @@ void JitArm::mtmsr(UGeckoInstruction inst)
// Don't interpret this, if we do we get thrown out
//JITDISABLE(SystemRegisters)
ARMReg rA = gpr.GetReg();
MOVI2R(rA, (u32)&MSR);
STR(rA, gpr.R(inst.RS));
gpr.Unlock(rA);
STR(R9, gpr.R(inst.RS), PPCSTATE_OFF(PowerPC::ppcState, msr));
WriteExit(js.compilerPC + 4, 0);
}

22
Source/Core/Core/Src/PowerPC/JitArm32/JitAsm.cpp
View File

@ -47,7 +47,7 @@ void JitArmAsmRoutineManager::Generate()
PUSH(2, R11, _LR); // R11 is frame pointer in Debug.
MOVI2R(R0, (u32)&CoreTiming::downcount);
MOVI2R(R9, (u32)&PowerPC::ppcState);
MOVI2R(R9, (u32)&PowerPC::ppcState.spr[0]);
FixupBranch skipToRealDispatcher = B();
dispatcher = GetCodePtr();
@ -63,7 +63,7 @@ void JitArmAsmRoutineManager::Generate()
// This block of code gets the address of the compiled block of code
// It runs though to the compiling portion if it isn't found
LDR(R12, R9, STRUCT_OFF(PowerPC::ppcState, pc));// Load the current PC into R12
LDR(R12, R9, PPCSTATE_OFF(PowerPC::ppcState, pc));// Load the current PC into R12
MOVI2R(R14, JIT_ICACHE_MASK); // Potential for optimization
AND(R12, R12, R14); // R12 contains PC & JIT_ICACHE_MASK here.
@ -92,7 +92,7 @@ void JitArmAsmRoutineManager::Generate()
// If we get to this point, that means that we don't have the block cached to execute
// So call ArmJit to compile the block and then execute it.
MOVI2R(R14, (u32)&Jit);
LDR(R0, R9, STRUCT_OFF(PowerPC::ppcState, pc));
LDR(R0, R9, PPCSTATE_OFF(PowerPC::ppcState, pc));
BL(R14);
B(dispatcherNoCheck);
@ -100,12 +100,12 @@ void JitArmAsmRoutineManager::Generate()
// fpException()
// Floating Point Exception Check, Jumped to if false
fpException = GetCodePtr();
LDR(R0, R9, STRUCT_OFF(PowerPC::ppcState, Exceptions));
LDR(R0, R9, PPCSTATE_OFF(PowerPC::ppcState, Exceptions));
ORR(R0, R0, EXCEPTION_FPU_UNAVAILABLE);
STR(R9, R0, STRUCT_OFF(PowerPC::ppcState, Exceptions));
STR(R9, R0, PPCSTATE_OFF(PowerPC::ppcState, Exceptions));
QuickCallFunction(R14, (void*)&PowerPC::CheckExceptions);
LDR(R0, R9, STRUCT_OFF(PowerPC::ppcState, npc));
STR(R9, R0, STRUCT_OFF(PowerPC::ppcState, pc));
LDR(R0, R9, PPCSTATE_OFF(PowerPC::ppcState, npc));
STR(R9, R0, PPCSTATE_OFF(PowerPC::ppcState, pc));
B(dispatcher);
SetJumpTarget(bail);
@ -116,11 +116,11 @@ void JitArmAsmRoutineManager::Generate()
// Does exception checking
testExceptions = GetCodePtr();
LDR(R0, R9, STRUCT_OFF(PowerPC::ppcState, pc));
STR(R9, R0, STRUCT_OFF(PowerPC::ppcState, npc));
LDR(R0, R9, PPCSTATE_OFF(PowerPC::ppcState, pc));
STR(R9, R0, PPCSTATE_OFF(PowerPC::ppcState, npc));
QuickCallFunction(R14, (void*)&PowerPC::CheckExceptions);
LDR(R0, R9, STRUCT_OFF(PowerPC::ppcState, npc));
STR(R9, R0, STRUCT_OFF(PowerPC::ppcState, pc));
LDR(R0, R9, PPCSTATE_OFF(PowerPC::ppcState, npc));
STR(R9, R0, PPCSTATE_OFF(PowerPC::ppcState, pc));
// Check the state pointer to see if we are exiting
// Gets checked on every exception check
MOVI2R(R0, (u32)PowerPC::GetStatePtr());

9
Source/Core/Core/Src/PowerPC/JitArm32/JitFPRCache.cpp
View File

@ -15,6 +15,7 @@
// Official SVN repository and contact information can be found at
// http://code.google.com/p/dolphin-emu/
#include "Jit.h"
#include "JitFPRCache.h"
ArmFPRCache::ArmFPRCache()
@ -122,7 +123,7 @@ ARMReg ArmFPRCache::GetPPCReg(u32 preg, bool PS1, bool preLoad)
for (u8 a = 0; a < NUMPPCREG; ++a)
if (ArmCRegs[a].PPCReg == 33)
{
u16 offset = STRUCT_OFF(PowerPC::ppcState, ps) + (preg * 16) + (PS1 ? 8 : 0);
u16 offset = PPCSTATE_OFF(PowerPC::ppcState, ps) + (preg * 16) + (PS1 ? 8 : 0);
if (preLoad)
emit->VLDR(ArmCRegs[a].Reg, R9, offset);
ArmCRegs[a].PPCReg = preg;
@ -131,10 +132,10 @@ ARMReg ArmFPRCache::GetPPCReg(u32 preg, bool PS1, bool preLoad)
return ArmCRegs[a].Reg;
}
// Alright, we couldn't get a free space, dump that least used register
u16 offsetOld = STRUCT_OFF(PowerPC::ppcState, ps) + (ArmCRegs[Num].PPCReg * 16) + (ArmCRegs[Num].PS1 ? 8 : 0);
u16 offsetOld = PPCSTATE_OFF(PowerPC::ppcState, ps) + (ArmCRegs[Num].PPCReg * 16) + (ArmCRegs[Num].PS1 ? 8 : 0);
emit->VSTR(ArmCRegs[Num].Reg, R9, offsetOld);
u16 offsetNew = STRUCT_OFF(PowerPC::ppcState, ps) + (preg * 16) + (PS1 ? 8 : 0);
u16 offsetNew = PPCSTATE_OFF(PowerPC::ppcState, ps) + (preg * 16) + (PS1 ? 8 : 0);
if (preLoad)
emit->VLDR(ArmCRegs[Num].Reg, R9, offsetNew);
ArmCRegs[Num].PPCReg = preg;
@ -159,7 +160,7 @@ void ArmFPRCache::Flush()
for(u8 a = 0; a < NUMPPCREG; ++a)
if (ArmCRegs[a].PPCReg != 33)
{
u16 offset = STRUCT_OFF(PowerPC::ppcState, ps) + (ArmCRegs[a].PPCReg * 16) + (ArmCRegs[a].PS1 ? 8 : 0);
u16 offset = PPCSTATE_OFF(PowerPC::ppcState, ps) + (ArmCRegs[a].PPCReg * 16) + (ArmCRegs[a].PS1 ? 8 : 0);
emit->VSTR(ArmCRegs[a].Reg, R9, offset);
ArmCRegs[a].PPCReg = 33;
ArmCRegs[a].LastLoad = 0;

9
Source/Core/Core/Src/PowerPC/JitArm32/JitRegCache.cpp
View File

@ -15,6 +15,7 @@
// Official SVN repository and contact information can be found at
// http://code.google.com/p/dolphin-emu/
#include "Jit.h"
#include "JitRegCache.h"
ArmRegCache::ArmRegCache()
@ -141,14 +142,14 @@ ARMReg ArmRegCache::R(u32 preg)
for (u8 a = 0; a < NUMPPCREG; ++a)
if (ArmCRegs[a].PPCReg == 33)
{
emit->LDR(ArmCRegs[a].Reg, R9, STRUCT_OFF(PowerPC::ppcState, gpr) + preg * 4);
emit->LDR(ArmCRegs[a].Reg, R9, PPCSTATE_OFF(PowerPC::ppcState, gpr) + preg * 4);
ArmCRegs[a].PPCReg = preg;
ArmCRegs[a].LastLoad = 0;
return ArmCRegs[a].Reg;
}
// Alright, we couldn't get a free space, dump that least used register
emit->STR(R9, ArmCRegs[Num].Reg, STRUCT_OFF(PowerPC::ppcState, gpr) + ArmCRegs[Num].PPCReg * 4);
emit->LDR(ArmCRegs[Num].Reg, R9, STRUCT_OFF(PowerPC::ppcState, gpr) + preg * 4);
emit->STR(R9, ArmCRegs[Num].Reg, PPCSTATE_OFF(PowerPC::ppcState, gpr) + ArmCRegs[Num].PPCReg * 4);
emit->LDR(ArmCRegs[Num].Reg, R9, PPCSTATE_OFF(PowerPC::ppcState, gpr) + preg * 4);
ArmCRegs[Num].PPCReg = preg;
ArmCRegs[Num].LastLoad = 0;
return ArmCRegs[Num].Reg;
@ -159,7 +160,7 @@ void ArmRegCache::Flush()
for(u8 a = 0; a < NUMPPCREG; ++a)
if (ArmCRegs[a].PPCReg != 33)
{
emit->STR(R9, ArmCRegs[a].Reg, STRUCT_OFF(PowerPC::ppcState, gpr) + ArmCRegs[a].PPCReg * 4);
emit->STR(R9, ArmCRegs[a].Reg, PPCSTATE_OFF(PowerPC::ppcState, gpr) + ArmCRegs[a].PPCReg * 4);
ArmCRegs[a].PPCReg = 33;
ArmCRegs[a].LastLoad = 0;
}