bool is32BitMode() const {
// FIXME: Can tablegen auto-generate this?
- return (STI.getFeatureBits() & X86::Mode64Bit) == 0;
+ return (STI.getFeatureBits() & X86::Mode32Bit) != 0;
+ }
+
+ bool is16BitMode() const {
+ // FIXME: Can tablegen auto-generate this?
+ return (STI.getFeatureBits() & X86::Mode16Bit) != 0;
+ }
+
+ /// Is16BitMemOperand - Return true if the specified instruction has
+ /// a 16-bit memory operand. Op specifies the operand # of the memoperand.
+ bool Is16BitMemOperand(const MCInst &MI, unsigned Op) const {
+ const MCOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
+ const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
+ const MCOperand &Disp = MI.getOperand(Op+X86::AddrDisp);
+
+ if (is16BitMode() && BaseReg.getReg() == 0 &&
+ Disp.isImm() && Disp.getImm() < 0x10000)
+ return true;
+ if ((BaseReg.getReg() != 0 &&
+ X86MCRegisterClasses[X86::GR16RegClassID].contains(BaseReg.getReg())) ||
+ (IndexReg.getReg() != 0 &&
+ X86MCRegisterClasses[X86::GR16RegClassID].contains(IndexReg.getReg())))
+ return true;
+ return false;
}
unsigned GetX86RegNum(const MCOperand &MO) const {
}
#endif
-/// Is16BitMemOperand - Return true if the specified instruction has
-/// a 16-bit memory operand. Op specifies the operand # of the memoperand.
-static bool Is16BitMemOperand(const MCInst &MI, unsigned Op) {
- const MCOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
- const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
-
- if ((BaseReg.getReg() != 0 &&
- X86MCRegisterClasses[X86::GR16RegClassID].contains(BaseReg.getReg())) ||
- (IndexReg.getReg() != 0 &&
- X86MCRegisterClasses[X86::GR16RegClassID].contains(IndexReg.getReg())))
- return true;
- return false;
-}
-
/// StartsWithGlobalOffsetTable - Check if this expression starts with
/// _GLOBAL_OFFSET_TABLE_ and if it is of the form
/// _GLOBAL_OFFSET_TABLE_-symbol. This is needed to support PIC on ELF
unsigned BaseRegNo = BaseReg ? GetX86RegNum(Base) : -1U;
+ // 16-bit addressing forms of the ModR/M byte have a different encoding for
+ // the R/M field and are far more limited in which registers can be used.
+ if (Is16BitMemOperand(MI, Op)) {
+ if (BaseReg) {
+ // For 32-bit addressing, the row and column values in Table 2-2 are
+ // basically the same. It's AX/CX/DX/BX/SP/BP/SI/DI in that order, with
+ // some special cases. And GetX86RegNum reflects that numbering.
+ // For 16-bit addressing it's more fun, as shown in the SDM Vol 2A,
+ // Table 2-1 "16-Bit Addressing Forms with the ModR/M byte". We can only
+ // use SI/DI/BP/BX, which have "row" values 4-7 in no particular order,
+ // while values 0-3 indicate the allowed combinations (base+index) of
+ // those: 0 for BX+SI, 1 for BX+DI, 2 for BP+SI, 3 for BP+DI.
+ //
+ // R16Table[] is a lookup from the normal RegNo, to the row values from
+ // Table 2-1 for 16-bit addressing modes. Where zero means disallowed.
+ static const unsigned R16Table[] = { 0, 0, 0, 7, 0, 6, 4, 5 };
+ unsigned RMfield = R16Table[BaseRegNo];
+
+ assert(RMfield && "invalid 16-bit base register");
+
+ if (IndexReg.getReg()) {
+ unsigned IndexReg16 = R16Table[GetX86RegNum(IndexReg)];
+
+ assert(IndexReg16 && "invalid 16-bit index register");
+ // We must have one of SI/DI (4,5), and one of BP/BX (6,7).
+ assert(((IndexReg16 ^ RMfield) & 2) &&
+ "invalid 16-bit base/index register combination");
+ assert(Scale.getImm() == 1 &&
+ "invalid scale for 16-bit memory reference");
+
+ // Allow base/index to appear in either order (although GAS doesn't).
+ if (IndexReg16 & 2)
+ RMfield = (RMfield & 1) | ((7 - IndexReg16) << 1);
+ else
+ RMfield = (IndexReg16 & 1) | ((7 - RMfield) << 1);
+ }
+
+ if (Disp.isImm() && isDisp8(Disp.getImm())) {
+ if (Disp.getImm() == 0 && BaseRegNo != N86::EBP) {
+ // There is no displacement; just the register.
+ EmitByte(ModRMByte(0, RegOpcodeField, RMfield), CurByte, OS);
+ return;
+ }
+ // Use the [REG]+disp8 form, including for [BP] which cannot be encoded.
+ EmitByte(ModRMByte(1, RegOpcodeField, RMfield), CurByte, OS);
+ EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups);
+ return;
+ }
+ // This is the [REG]+disp16 case.
+ EmitByte(ModRMByte(2, RegOpcodeField, RMfield), CurByte, OS);
+ } else {
+ // There is no BaseReg; this is the plain [disp16] case.
+ EmitByte(ModRMByte(0, RegOpcodeField, 6), CurByte, OS);
+ }
+
+ // Emit 16-bit displacement for plain disp16 or [REG]+disp16 cases.
+ EmitImmediate(Disp, MI.getLoc(), 2, FK_Data_2, CurByte, OS, Fixups);
+ return;
+ }
+
// Determine whether a SIB byte is needed.
// If no BaseReg, issue a RIP relative instruction only if the MCE can
// resolve addresses on-the-fly, otherwise use SIB (Intel Manual 2A, table
bool HasVEX_4V = (TSFlags >> X86II::VEXShift) & X86II::VEX_4V;
bool HasVEX_4VOp3 = (TSFlags >> X86II::VEXShift) & X86II::VEX_4VOp3;
bool HasMemOp4 = (TSFlags >> X86II::VEXShift) & X86II::MemOp4;
+ bool HasEVEX_RC = (TSFlags >> X86II::VEXShift) & X86II::EVEX_RC;
// VEX_R: opcode externsion equivalent to REX.R in
// 1's complement (inverted) form
unsigned char VEX_W = 0;
// XOP: Use XOP prefix byte 0x8f instead of VEX.
- unsigned char XOP = 0;
+ bool XOP = false;
// VEX_5M (VEX m-mmmmm field):
//
// 0b00011: implied 0F 3A leading opcode bytes
// 0b00100-0b11111: Reserved for future use
// 0b01000: XOP map select - 08h instructions with imm byte
- // 0b10001: XOP map select - 09h instructions with no imm byte
+ // 0b01001: XOP map select - 09h instructions with no imm byte
+ // 0b01010: XOP map select - 0Ah instructions with imm dword
unsigned char VEX_5M = 0x1;
// VEX_4V (VEX vvvv field): a register specifier
// EVEX_b
unsigned char EVEX_b = 0;
+ // EVEX_rc
+ unsigned char EVEX_rc = 0;
+
// EVEX_aaa
unsigned char EVEX_aaa = 0;
+ bool EncodeRC = false;
+
// Encode the operand size opcode prefix as needed.
if (TSFlags & X86II::OpSize)
VEX_PP = 0x01;
VEX_W = 1;
if ((TSFlags >> X86II::VEXShift) & X86II::XOP)
- XOP = 1;
+ XOP = true;
if ((TSFlags >> X86II::VEXShift) & X86II::VEX_L)
VEX_L = 1;
case X86II::TA: // 0F 3A
VEX_5M = 0x3;
break;
+ case X86II::T8PD: // 66 0F 38
+ VEX_PP = 0x1;
+ VEX_5M = 0x2;
+ break;
case X86II::T8XS: // F3 0F 38
VEX_PP = 0x2;
VEX_5M = 0x2;
VEX_PP = 0x3;
VEX_5M = 0x2;
break;
+ case X86II::TAPD: // 66 0F 3A
+ VEX_PP = 0x1;
+ VEX_5M = 0x3;
+ break;
case X86II::TAXD: // F2 0F 3A
VEX_PP = 0x3;
VEX_5M = 0x3;
break;
+ case X86II::PD: // 66 0F
+ VEX_PP = 0x1;
+ break;
case X86II::XS: // F3 0F
VEX_PP = 0x2;
break;
case X86II::XOPA:
VEX_5M = 0xA;
break;
- case X86II::A6: // Bypass: Not used by VEX
- case X86II::A7: // Bypass: Not used by VEX
- case X86II::TB: // Bypass: Not used by VEX
- case 0:
- break; // No prefix!
+ case X86II::TB: // VEX_5M/VEX_PP already correct
+ break;
}
++CurOp;
switch (TSFlags & X86II::FormMask) {
- case X86II::MRMInitReg: llvm_unreachable("FIXME: Remove this!");
case X86II::MRMDestMem: {
// MRMDestMem instructions forms:
// MemAddr, src1(ModR/M)
VEX_X = 0x0;
CurOp++;
if (HasVEX_4VOp3)
- VEX_4V = getVEXRegisterEncoding(MI, CurOp);
+ VEX_4V = getVEXRegisterEncoding(MI, CurOp++);
+ if (EVEX_b) {
+ if (HasEVEX_RC) {
+ unsigned RcOperand = NumOps-1;
+ assert(RcOperand >= CurOp);
+ EVEX_rc = MI.getOperand(RcOperand).getImm() & 0x3;
+ }
+ EncodeRC = true;
+ }
break;
case X86II::MRMDestReg:
// MRMDestReg instructions forms:
VEX_R = 0x0;
if (HasEVEX && X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
EVEX_R2 = 0x0;
+ if (EVEX_b)
+ EncodeRC = true;
break;
case X86II::MRM0r: case X86II::MRM1r:
case X86II::MRM2r: case X86II::MRM3r:
(VEX_4V << 3) |
(EVEX_U << 2) |
VEX_PP, CurByte, OS);
- EmitByte((EVEX_z << 7) |
- (EVEX_L2 << 6) |
- (VEX_L << 5) |
- (EVEX_b << 4) |
- (EVEX_V2 << 3) |
- EVEX_aaa, CurByte, OS);
+ if (EncodeRC)
+ EmitByte((EVEX_z << 7) |
+ (EVEX_rc << 5) |
+ (EVEX_b << 4) |
+ (EVEX_V2 << 3) |
+ EVEX_aaa, CurByte, OS);
+ else
+ EmitByte((EVEX_z << 7) |
+ (EVEX_L2 << 6) |
+ (VEX_L << 5) |
+ (EVEX_b << 4) |
+ (EVEX_V2 << 3) |
+ EVEX_aaa, CurByte, OS);
}
}
}
switch (TSFlags & X86II::FormMask) {
- case X86II::MRMInitReg: llvm_unreachable("FIXME: Remove this!");
case X86II::MRMSrcReg:
if (MI.getOperand(0).isReg() &&
X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
unsigned &CurByte, int MemOperand,
const MCInst &MI,
raw_ostream &OS) const {
- switch (TSFlags & X86II::SegOvrMask) {
- default: llvm_unreachable("Invalid segment!");
- case 0:
- // No segment override, check for explicit one on memory operand.
- if (MemOperand != -1) { // If the instruction has a memory operand.
- switch (MI.getOperand(MemOperand+X86::AddrSegmentReg).getReg()) {
- default: llvm_unreachable("Unknown segment register!");
- case 0: break;
- case X86::CS: EmitByte(0x2E, CurByte, OS); break;
- case X86::SS: EmitByte(0x36, CurByte, OS); break;
- case X86::DS: EmitByte(0x3E, CurByte, OS); break;
- case X86::ES: EmitByte(0x26, CurByte, OS); break;
- case X86::FS: EmitByte(0x64, CurByte, OS); break;
- case X86::GS: EmitByte(0x65, CurByte, OS); break;
- }
- }
- break;
- case X86II::FS:
- EmitByte(0x64, CurByte, OS);
- break;
- case X86II::GS:
- EmitByte(0x65, CurByte, OS);
- break;
+ if (MemOperand < 0)
+ return; // No memory operand
+
+ // Check for explicit segment override on memory operand.
+ switch (MI.getOperand(MemOperand+X86::AddrSegmentReg).getReg()) {
+ default: llvm_unreachable("Unknown segment register!");
+ case 0: break;
+ case X86::CS: EmitByte(0x2E, CurByte, OS); break;
+ case X86::SS: EmitByte(0x36, CurByte, OS); break;
+ case X86::DS: EmitByte(0x3E, CurByte, OS); break;
+ case X86::ES: EmitByte(0x26, CurByte, OS); break;
+ case X86::FS: EmitByte(0x64, CurByte, OS); break;
+ case X86::GS: EmitByte(0x65, CurByte, OS); break;
}
}
// Emit the address size opcode prefix as needed.
bool need_address_override;
- if (TSFlags & X86II::AdSize) {
+ // The AdSize prefix is only for 32-bit and 64-bit modes. Hm, perhaps we
+ // should introduce an AdSize16 bit instead of having seven special cases?
+ if ((!is16BitMode() && TSFlags & X86II::AdSize) ||
+ (is16BitMode() && (MI.getOpcode() == X86::JECXZ_32 ||
+ MI.getOpcode() == X86::MOV8o8a ||
+ MI.getOpcode() == X86::MOV16o16a ||
+ MI.getOpcode() == X86::MOV32o32a ||
+ MI.getOpcode() == X86::MOV8ao8 ||
+ MI.getOpcode() == X86::MOV16ao16 ||
+ MI.getOpcode() == X86::MOV32ao32))) {
need_address_override = true;
} else if (MemOperand == -1) {
need_address_override = false;
assert(!Is64BitMemOperand(MI, MemOperand));
need_address_override = Is16BitMemOperand(MI, MemOperand);
} else {
- need_address_override = false;
+ assert(is16BitMode());
+ assert(!Is64BitMemOperand(MI, MemOperand));
+ need_address_override = !Is16BitMemOperand(MI, MemOperand);
}
if (need_address_override)
EmitByte(0x67, CurByte, OS);
// Emit the operand size opcode prefix as needed.
- if (TSFlags & X86II::OpSize)
+ if (TSFlags & (is16BitMode() ? X86II::OpSize16 : X86II::OpSize))
EmitByte(0x66, CurByte, OS);
bool Need0FPrefix = false;
case X86II::A7: // 0F A7
Need0FPrefix = true;
break;
- case X86II::T8XS: // F3 0F 38
- EmitByte(0xF3, CurByte, OS);
- Need0FPrefix = true;
- break;
- case X86II::T8XD: // F2 0F 38
- EmitByte(0xF2, CurByte, OS);
- Need0FPrefix = true;
- break;
- case X86II::TAXD: // F2 0F 3A
- EmitByte(0xF2, CurByte, OS);
+ case X86II::PD: // 66 0F
+ case X86II::T8PD: // 66 0F 38
+ case X86II::TAPD: // 66 0F 3A
+ EmitByte(0x66, CurByte, OS);
Need0FPrefix = true;
break;
case X86II::XS: // F3 0F
+ case X86II::T8XS: // F3 0F 38
EmitByte(0xF3, CurByte, OS);
Need0FPrefix = true;
break;
case X86II::XD: // F2 0F
+ case X86II::T8XD: // F2 0F 38
+ case X86II::TAXD: // F2 0F 3A
EmitByte(0xF2, CurByte, OS);
Need0FPrefix = true;
break;
- case X86II::D8: EmitByte(0xD8, CurByte, OS); break;
- case X86II::D9: EmitByte(0xD9, CurByte, OS); break;
- case X86II::DA: EmitByte(0xDA, CurByte, OS); break;
- case X86II::DB: EmitByte(0xDB, CurByte, OS); break;
- case X86II::DC: EmitByte(0xDC, CurByte, OS); break;
- case X86II::DD: EmitByte(0xDD, CurByte, OS); break;
- case X86II::DE: EmitByte(0xDE, CurByte, OS); break;
- case X86II::DF: EmitByte(0xDF, CurByte, OS); break;
+ case X86II::D8:
+ case X86II::D9:
+ case X86II::DA:
+ case X86II::DB:
+ case X86II::DC:
+ case X86II::DD:
+ case X86II::DE:
+ case X86II::DF:
+ EmitByte(0xD8+(((TSFlags & X86II::Op0Mask) - X86II::D8) >> X86II::Op0Shift),
+ CurByte, OS);
+ break;
}
// Handle REX prefix.
// FIXME: Pull this up into previous switch if REX can be moved earlier.
switch (TSFlags & X86II::Op0Mask) {
+ case X86II::T8PD: // 66 0F 38
case X86II::T8XS: // F3 0F 38
case X86II::T8XD: // F2 0F 38
case X86II::T8: // 0F 38
EmitByte(0x38, CurByte, OS);
break;
+ case X86II::TAPD: // 66 0F 3A
case X86II::TAXD: // F2 0F 3A
case X86II::TA: // 0F 3A
EmitByte(0x3A, CurByte, OS);
// It uses the EVEX.aaa field?
bool HasEVEX = (TSFlags >> X86II::VEXShift) & X86II::EVEX;
bool HasEVEX_K = HasEVEX && ((TSFlags >> X86II::VEXShift) & X86II::EVEX_K);
-
+ bool HasEVEX_RC = HasEVEX && ((TSFlags >> X86II::VEXShift) & X86II::EVEX_RC);
+
// Determine where the memory operand starts, if present.
int MemoryOperand = X86II::getMemoryOperandNo(TSFlags, Opcode);
if (MemoryOperand != -1) MemoryOperand += CurOp;
unsigned SrcRegNum = 0;
switch (TSFlags & X86II::FormMask) {
- case X86II::MRMInitReg:
- llvm_unreachable("FIXME: Remove this form when the JIT moves to MCCodeEmitter!");
default: errs() << "FORM: " << (TSFlags & X86II::FormMask) << "\n";
llvm_unreachable("Unknown FormMask value in X86MCCodeEmitter!");
case X86II::Pseudo:
CurOp = HasMemOp4 ? SrcRegNum : SrcRegNum + 1;
if (HasVEX_4VOp3)
++CurOp;
+ // do not count the rounding control operand
+ if (HasEVEX_RC)
+ NumOps--;
break;
case X86II::MRMSrcMem: {
projects / oota-llvm.git / blobdiff