dolphin/Source/Core/Common/Src/ArmEmitter.h

590 lines
16 KiB
C
Raw Normal View History

2013-02-26 12:49:00 -07:00
// Copyright (C) 2003 Dolphin Project.
// This program is free software: you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation, version 2.0.
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License 2.0 for more details.
// A copy of the GPL 2.0 should have been included with the program.
// If not, see http://www.gnu.org/licenses/
// Official SVN repository and contact information can be found at
// http://code.google.com/p/dolphin-emu/
// WARNING - THIS LIBRARY IS NOT THREAD SAFE!!!
#ifndef _DOLPHIN_ARM_CODEGEN_
#define _DOLPHIN_ARM_CODEGEN_
#include "Common.h"
#include "MemoryUtil.h"
#if defined(__SYMBIAN32__) || defined(PANDORA)
#include <signal.h>
#endif
#undef _IP
#undef R0
#undef _SP
#undef _LR
#undef _PC
namespace ArmGen
{
enum ARMReg
{
// GPRs
R0 = 0, R1, R2, R3, R4, R5,
R6, R7, R8, R9, R10, R11,
// SPRs
// R13 - R15 are SP, LR, and PC.
// Almost always referred to by name instead of register number
R12 = 12, R13 = 13, R14 = 14, R15 = 15,
_IP = 12, _SP = 13, _LR = 14, _PC = 15,
// VFP single precision registers
S0, S1, S2, S3, S4, S5, S6,
S7, S8, S9, S10, S11, S12, S13,
S14, S15, S16, S17, S18, S19, S20,
S21, S22, S23, S24, S25, S26, S27,
S28, S29, S30, S31,
// VFP Double Precision registers
D0, D1, D2, D3, D4, D5, D6, D7,
D8, D9, D10, D11, D12, D13, D14, D15,
D16, D17, D18, D19, D20, D21, D22, D23,
D24, D25, D26, D27, D28, D29, D30, D31,
// ASIMD Quad-Word registers
Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7,
Q8, Q9, Q10, Q11, Q12, Q13, Q14, Q15,
INVALID_REG = 0xFFFFFFFF
};
enum CCFlags
{
CC_EQ = 0, // Equal
CC_NEQ, // Not equal
CC_CS, // Carry Set
CC_CC, // Carry Clear
CC_MI, // Minus (Negative)
CC_PL, // Plus
CC_VS, // Overflow
CC_VC, // No Overflow
CC_HI, // Unsigned higher
CC_LS, // Unsigned lower or same
CC_GE, // Signed greater than or equal
CC_LT, // Signed less than
CC_GT, // Signed greater than
CC_LE, // Signed less than or equal
CC_AL, // Always (unconditional) 14
CC_HS = CC_CS, // Alias of CC_CS Unsigned higher or same
CC_LO = CC_CC, // Alias of CC_CC Unsigned lower
};
const u32 NO_COND = 0xE0000000;
enum ShiftType
{
ST_LSL = 0,
ST_ASL = 0,
ST_LSR = 1,
ST_ASR = 2,
ST_ROR = 3,
ST_RRX = 4
};
enum IntegerSize
{
I_I8 = 0,
I_I16,
I_I32,
I_I64
};
enum
{
NUMGPRs = 13,
};
class ARMXEmitter;
enum OpType
{
TYPE_IMM = 0,
TYPE_REG,
TYPE_IMMSREG,
TYPE_RSR,
TYPE_MEM
};
// This is no longer a proper operand2 class. Need to split up.
class Operand2
{
friend class ARMXEmitter;
protected:
u32 Value;
private:
OpType Type;
// IMM types
u8 Rotation; // Only for u8 values
// Register types
u8 IndexOrShift;
ShiftType Shift;
public:
OpType GetType()
{
return Type;
}
Operand2() {}
Operand2(u32 imm, OpType type = TYPE_IMM)
{
Type = type;
Value = imm;
Rotation = 0;
}
Operand2(ARMReg Reg)
{
Type = TYPE_REG;
Value = Reg;
Rotation = 0;
}
Operand2(u8 imm, u8 rotation)
{
Type = TYPE_IMM;
Value = imm;
Rotation = rotation;
}
Operand2(ARMReg base, ShiftType type, ARMReg shift) // RSR
{
Type = TYPE_RSR;
_assert_msg_(DYNA_REC, type != ST_RRX, "Invalid Operand2: RRX does not take a register shift amount");
IndexOrShift = shift;
Shift = type;
Value = base;
}
Operand2(u8 shift, ShiftType type, ARMReg base)// For IMM shifted register
{
if(shift == 32) shift = 0;
switch (type)
{
case ST_LSL:
_assert_msg_(DYNA_REC, shift < 32, "Invalid Operand2: LSL %u", shift);
break;
case ST_LSR:
_assert_msg_(DYNA_REC, shift <= 32, "Invalid Operand2: LSR %u", shift);
if (!shift)
type = ST_LSL;
if (shift == 32)
shift = 0;
break;
case ST_ASR:
_assert_msg_(DYNA_REC, shift < 32, "Invalid Operand2: LSR %u", shift);
if (!shift)
type = ST_LSL;
if (shift == 32)
shift = 0;
break;
case ST_ROR:
_assert_msg_(DYNA_REC, shift < 32, "Invalid Operand2: ROR %u", shift);
if (!shift)
type = ST_LSL;
break;
case ST_RRX:
_assert_msg_(DYNA_REC, shift == 0, "Invalid Operand2: RRX does not take an immediate shift amount");
type = ST_ROR;
break;
}
IndexOrShift = shift;
Shift = type;
Value = base;
Type = TYPE_IMMSREG;
}
const u32 GetData()
{
switch(Type)
{
case TYPE_IMM:
return Imm12Mod(); // This'll need to be changed later
case TYPE_REG:
return Rm();
case TYPE_IMMSREG:
return IMMSR();
case TYPE_RSR:
return RSR();
default:
_assert_msg_(DYNA_REC, false, "GetData with Invalid Type");
return 0;
}
}
const u32 IMMSR() // IMM shifted register
{
_assert_msg_(DYNA_REC, Type == TYPE_IMMSREG, "IMMSR must be imm shifted register");
2013-02-26 12:49:00 -07:00
return ((IndexOrShift & 0x1f) << 7 | (Shift << 5) | Value);
}
const 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()
{
_assert_msg_(DYNA_REC, Type == TYPE_REG, "Rm must be with Reg");
return Value;
}
const u32 Imm5()
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm5 not IMM value");
return ((Value & 0x0000001F) << 7);
}
const u32 Imm8()
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm8Rot not IMM value");
return Value & 0xFF;
}
const 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()
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm12 not IMM");
return (Value & 0x00000FFF);
}
const 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
// expand a 8bit IMM to a 32bit value and gives you some rotation as
// well.
// Each rotation rotates to the right by 2 bits
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm12Mod not IMM");
return ((Rotation & 0xF) << 8) | (Value & 0xFF);
}
const u32 Imm16()
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm16 not IMM");
return ( (Value & 0xF000) << 4) | (Value & 0x0FFF);
}
const u32 Imm16Low()
{
return Imm16();
}
const 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()
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm16 not IMM");
return (Value & 0x0FFFFFFF);
}
// NEON and ASIMD specific
const u32 Imm8ASIMD()
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm8ASIMD not IMM");
return ((Value & 0x80) << 17) | ((Value & 0x70) << 12) | (Value & 0xF);
}
const u32 Imm8VFP()
{
_assert_msg_(DYNA_REC, (Type == TYPE_IMM), "Imm8VFP not IMM");
return ((Value & 0xF0) << 12) | (Value & 0xF);
}
};
// Use these when you don't know if an imm can be represented as an operand2.
// This lets you generate both an optimal and a fallback solution by checking
// the return value, which will be false if these fail to find a Operand2 that
// represents your 32-bit imm value.
bool TryMakeOperand2(u32 imm, Operand2 &op2);
bool TryMakeOperand2_AllowInverse(u32 imm, Operand2 &op2, bool *inverse);
bool TryMakeOperand2_AllowNegation(s32 imm, Operand2 &op2, bool *negated);
inline Operand2 R(ARMReg Reg) { return Operand2(Reg, TYPE_REG); }
inline Operand2 IMM(u32 Imm) { return Operand2(Imm, TYPE_IMM); }
inline Operand2 Mem(void *ptr) { return Operand2((u32)ptr, TYPE_IMM); }
//usage: struct {int e;} s; STRUCT_OFFSET(s,e)
#define STRUCT_OFF(str,elem) ((u32)((u32)&(str).elem-(u32)&(str)))
struct FixupBranch
{
u8 *ptr;
u32 condition; // Remembers our codition at the time
int type; //0 = B 1 = BL
};
typedef const u8* JumpTarget;
class ARMXEmitter
{
friend struct OpArg; // for Write8 etc
private:
u8 *code, *startcode;
u8 *lastCacheFlushEnd;
u32 condition;
void WriteStoreOp(u32 op, ARMReg dest, ARMReg src, Operand2 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);
void WriteSignedMultiply(u32 Op, u32 Op2, u32 Op3, ARMReg dest, ARMReg r1, ARMReg r2);
2013-02-26 12:49:00 -07:00
void Write4OpMultiply(u32 op, ARMReg destLo, ARMReg destHi, ARMReg rn, ARMReg rm);
2013-02-26 12:49:00 -07:00
// New Ops
void WriteInstruction(u32 op, ARMReg Rd, ARMReg Rn, Operand2 Rm, bool SetFlags = false);
protected:
inline void Write32(u32 value) {*(u32*)code = value; code+=4;}
public:
ARMXEmitter() : code(0), startcode(0), lastCacheFlushEnd(0) {
condition = CC_AL << 28;
}
ARMXEmitter(u8 *code_ptr) {
code = code_ptr;
lastCacheFlushEnd = code_ptr;
startcode = code_ptr;
condition = CC_AL << 28;
}
virtual ~ARMXEmitter() {}
void SetCodePtr(u8 *ptr);
void ReserveCodeSpace(u32 bytes);
const u8 *AlignCode16();
const u8 *AlignCodePage();
const u8 *GetCodePtr() const;
void FlushIcache();
void FlushIcacheSection(u8 *start, u8 *end);
u8 *GetWritableCodePtr();
void SetCC(CCFlags cond = CC_AL);
// Special purpose instructions
// Dynamic Endian Switching
void SETEND(bool BE);
// Debug Breakpoint
void BKPT(u16 arg);
// Hint instruction
void YIELD();
// Do nothing
void NOP(int count = 1); //nop padding - TODO: fast nop slides, for amd and intel (check their manuals)
#ifdef CALL
#undef CALL
#endif
// Branching
FixupBranch B();
FixupBranch B_CC(CCFlags Cond);
void B_CC(CCFlags Cond, const void *fnptr);
FixupBranch BL();
FixupBranch BL_CC(CCFlags Cond);
void SetJumpTarget(FixupBranch const &branch);
void B (const void *fnptr);
void B (ARMReg src);
void BL(const void *fnptr);
void BL(ARMReg src);
void PUSH(const int num, ...);
void POP(const int num, ...);
// New Data Ops
void AND (ARMReg Rd, ARMReg Rn, Operand2 Rm);
void ANDS(ARMReg Rd, ARMReg Rn, Operand2 Rm);
void EOR (ARMReg dest, ARMReg src, Operand2 op2);
void EORS(ARMReg dest, ARMReg src, Operand2 op2);
void SUB (ARMReg dest, ARMReg src, Operand2 op2);
void SUBS(ARMReg dest, ARMReg src, Operand2 op2);
void RSB (ARMReg dest, ARMReg src, Operand2 op2);
void RSBS(ARMReg dest, ARMReg src, Operand2 op2);
void ADD (ARMReg dest, ARMReg src, Operand2 op2);
void ADDS(ARMReg dest, ARMReg src, Operand2 op2);
void ADC (ARMReg dest, ARMReg src, Operand2 op2);
void ADCS(ARMReg dest, ARMReg src, Operand2 op2);
void LSL (ARMReg dest, ARMReg src, Operand2 op2);
void LSL (ARMReg dest, ARMReg src, ARMReg op2);
void LSLS(ARMReg dest, ARMReg src, Operand2 op2);
void LSLS(ARMReg dest, ARMReg src, ARMReg op2);
void SBC (ARMReg dest, ARMReg src, Operand2 op2);
void SBCS(ARMReg dest, ARMReg src, Operand2 op2);
void REV (ARMReg dest, ARMReg src);
void REV16 (ARMReg dest, ARMReg src);
void RSC (ARMReg dest, ARMReg src, Operand2 op2);
void RSCS(ARMReg dest, ARMReg src, Operand2 op2);
void TST ( ARMReg src, Operand2 op2);
void TEQ ( ARMReg src, Operand2 op2);
void CMP ( ARMReg src, Operand2 op2);
void CMN ( ARMReg src, Operand2 op2);
void ORR (ARMReg dest, ARMReg src, Operand2 op2);
void ORRS(ARMReg dest, ARMReg src, Operand2 op2);
void MOV (ARMReg dest, Operand2 op2);
void MOVS(ARMReg dest, Operand2 op2);
void BIC (ARMReg dest, ARMReg src, Operand2 op2); // BIC = ANDN
void BICS(ARMReg dest, ARMReg src, Operand2 op2);
void MVN (ARMReg dest, Operand2 op2);
void MVNS(ARMReg dest, Operand2 op2);
void MOVW(ARMReg dest, Operand2 op2);
void MOVT(ARMReg dest, Operand2 op2, bool TopBits = false);
// UDIV and SDIV are only available on CPUs that have
// the idiva hardare capacity
void UDIV(ARMReg dest, ARMReg dividend, ARMReg divisor);
void SDIV(ARMReg dest, ARMReg dividend, ARMReg divisor);
void MUL (ARMReg dest, ARMReg src, ARMReg op2);
void MULS(ARMReg dest, ARMReg src, ARMReg op2);
2013-02-26 12:49:00 -07:00
void UMULL(ARMReg destLo, ARMReg destHi, ARMReg rn, ARMReg rm);
void SMULL(ARMReg destLo, ARMReg destHi, ARMReg rn, ARMReg rm);
void SXTB(ARMReg dest, ARMReg op2);
void SXTH(ARMReg dest, ARMReg op2, u8 rotation = 0);
void SXTAH(ARMReg dest, ARMReg src, ARMReg op2, u8 rotation = 0);
// Using just MSR here messes with our defines on the PPC side of stuff (when this code was in dolphin...)
// Just need to put an underscore here, bit annoying.
void _MSR (bool nzcvq, bool g, Operand2 op2);
void _MSR (bool nzcvq, bool g, ARMReg src );
void MRS (ARMReg dest);
// Memory load/store operations
void LDR (ARMReg dest, ARMReg src, Operand2 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);
// 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 STMFD(ARMReg dest, bool WriteBack, const int Regnum, ...);
void LDMFD(ARMReg dest, bool WriteBack, const int Regnum, ...);
// Exclusive Access operations
void LDREX(ARMReg dest, ARMReg base);
// dest contains the result if the instruction managed to store the value
void STREX(ARMReg dest, ARMReg base, ARMReg op);
void DMB ();
void SVC(Operand2 op);
// NEON and ASIMD instructions
// None of these will be created with conditional since ARM
// is deprecating conditional execution of ASIMD instructions.
// ASIMD instructions don't even have a conditional encoding.
// Subtracts the base from the register to give us the real one
ARMReg SubBase(ARMReg Reg);
// NEON Only
void VADD(IntegerSize Size, ARMReg Vd, ARMReg Vn, ARMReg Vm);
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);
2013-02-26 12:49:00 -07:00
void VCMP(ARMReg Vd, ARMReg Vm);
// Compares against zero
void VCMP(ARMReg Vd);
void VDIV(ARMReg Vd, ARMReg Vn, ARMReg Vm);
void VSQRT(ARMReg Vd, ARMReg Vm);
// NEON and VFP
void VADD(ARMReg Vd, ARMReg Vn, ARMReg Vm);
void VSUB(ARMReg Vd, ARMReg Vn, ARMReg Vm);
void VABS(ARMReg Vd, ARMReg Vm);
void VNEG(ARMReg Vd, ARMReg Vm);
void VMUL(ARMReg Vd, ARMReg Vn, ARMReg Vm);
2013-02-26 12:49:00 -07:00
void VMOV(ARMReg Dest, ARMReg Src, bool high);
void VMOV(ARMReg Dest, ARMReg Src);
void QuickCallFunction(ARMReg scratchreg, void *func);
// Utility functions
void MOVI2R(ARMReg reg, u32 val, bool optimize = true);
void ARMABI_MOVI2M(Operand2 op, Operand2 val);
}; // class ARMXEmitter
// Everything that needs to generate X86 code should inherit from this.
// You get memory management for free, plus, you can use all the MOV etc functions without
// having to prefix them with gen-> or something similar.
class ARMXCodeBlock : public ARMXEmitter
{
protected:
u8 *region;
size_t region_size;
public:
ARMXCodeBlock() : region(NULL), region_size(0) {}
virtual ~ARMXCodeBlock() { if (region) FreeCodeSpace(); }
// Call this before you generate any code.
void AllocCodeSpace(int size)
{
region_size = size;
region = (u8*)AllocateExecutableMemory(region_size);
SetCodePtr(region);
}
// Always clear code space with breakpoints, so that if someone accidentally executes
// uninitialized, it just breaks into the debugger.
void ClearCodeSpace()
{
// x86/64: 0xCC = breakpoint
memset(region, 0xCC, region_size);
ResetCodePtr();
}
// Call this when shutting down. Don't rely on the destructor, even though it'll do the job.
void FreeCodeSpace()
{
FreeMemoryPages(region, region_size);
region = NULL;
region_size = 0;
}
bool IsInSpace(u8 *ptr)
{
return ptr >= region && ptr < region + region_size;
}
// Cannot currently be undone. Will write protect the entire code region.
// Start over if you need to change the code (call FreeCodeSpace(), AllocCodeSpace()).
void WriteProtect()
{
WriteProtectMemory(region, region_size, true);
}
void ResetCodePtr()
{
SetCodePtr(region);
}
size_t GetSpaceLeft() const
{
return region_size - (GetCodePtr() - region);
}
};
} // namespace
#endif // _DOLPHIN_INTEL_CODEGEN_