Add X86 BZHI instruction as well as BMI2 feature detection.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@142122 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Craig Topper 2011-10-16 07:55:05 +00:00
parent dc479c4a89
commit b53fa8bf19
12 changed files with 101 additions and 65 deletions

View File

@ -387,20 +387,24 @@ namespace X86II {
/// and the additional register is encoded in VEX_VVVV prefix.
VEX_4V = 1U << 2,
/// VEX_4VOp3 - Similar to VEX_4V, but used on instructions that encode
/// operand 3 with VEX.vvvv.
VEX_4VOp3 = 1U << 3,
/// VEX_I8IMM - Specifies that the last register used in a AVX instruction,
/// must be encoded in the i8 immediate field. This usually happens in
/// instructions with 4 operands.
VEX_I8IMM = 1U << 3,
VEX_I8IMM = 1U << 4,
/// VEX_L - Stands for a bit in the VEX opcode prefix meaning the current
/// instruction uses 256-bit wide registers. This is usually auto detected
/// if a VR256 register is used, but some AVX instructions also have this
/// field marked when using a f256 memory references.
VEX_L = 1U << 4,
VEX_L = 1U << 5,
// VEX_LIG - Specifies that this instruction ignores the L-bit in the VEX
// prefix. Usually used for scalar instructions. Needed by disassembler.
VEX_LIG = 1U << 5,
VEX_LIG = 1U << 6,
/// Has3DNow0F0FOpcode - This flag indicates that the instruction uses the
/// wacky 0x0F 0x0F prefix for 3DNow! instructions. The manual documents
@ -408,7 +412,7 @@ namespace X86II {
/// storing a classifier in the imm8 field. To simplify our implementation,
/// we handle this by storeing the classifier in the opcode field and using
/// this flag to indicate that the encoder should do the wacky 3DNow! thing.
Has3DNow0F0FOpcode = 1U << 6
Has3DNow0F0FOpcode = 1U << 7
};
// getBaseOpcodeFor - This function returns the "base" X86 opcode for the
@ -477,12 +481,9 @@ namespace X86II {
case X86II::MRMDestMem:
return 0;
case X86II::MRMSrcMem: {
// FIXME: BEXTR uses VEX.vvvv for Operand 3
bool IsBEXTR = (Opcode == X86::BEXTR32rr || Opcode == X86::BEXTR32rm ||
Opcode == X86::BEXTR64rr || Opcode == X86::BEXTR64rm);
bool HasVEX_4V = (TSFlags >> X86II::VEXShift) & X86II::VEX_4V;
unsigned FirstMemOp = 1;
if (HasVEX_4V && !IsBEXTR)
if (HasVEX_4V)
++FirstMemOp;// Skip the register source (which is encoded in VEX_VVVV).
// FIXME: Maybe lea should have its own form? This is a horrible hack.

View File

@ -386,9 +386,8 @@ void X86MCCodeEmitter::EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
int MemOperand, const MCInst &MI,
const MCInstrDesc &Desc,
raw_ostream &OS) const {
bool HasVEX_4V = false;
if ((TSFlags >> X86II::VEXShift) & X86II::VEX_4V)
HasVEX_4V = true;
bool HasVEX_4V = (TSFlags >> X86II::VEXShift) & X86II::VEX_4V;
bool HasVEX_4VOp3 = (TSFlags >> X86II::VEXShift) & X86II::VEX_4VOp3;
// VEX_R: opcode externsion equivalent to REX.R in
// 1's complement (inverted) form
@ -447,11 +446,6 @@ void X86MCCodeEmitter::EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
//
unsigned char VEX_PP = 0;
// FIXME: BEXTR uses VEX.vvvv for Operand 3 instead of Operand 2
unsigned Opcode = MI.getOpcode();
bool IsBEXTR = (Opcode == X86::BEXTR32rr || Opcode == X86::BEXTR32rm ||
Opcode == X86::BEXTR64rr || Opcode == X86::BEXTR64rm);
// Encode the operand size opcode prefix as needed.
if (TSFlags & X86II::OpSize)
VEX_PP = 0x01;
@ -530,8 +524,7 @@ void X86MCCodeEmitter::EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
if (X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
VEX_R = 0x0;
// FIXME: BEXTR uses VEX.vvvv for Operand 3
if (HasVEX_4V && !IsBEXTR)
if (HasVEX_4V)
VEX_4V = getVEXRegisterEncoding(MI, 1);
if (X86II::isX86_64ExtendedReg(
@ -541,7 +534,7 @@ void X86MCCodeEmitter::EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
MI.getOperand(MemOperand+X86::AddrIndexReg).getReg()))
VEX_X = 0x0;
if (IsBEXTR)
if (HasVEX_4VOp3)
VEX_4V = getVEXRegisterEncoding(MI, X86::AddrNumOperands+1);
break;
case X86II::MRM0m: case X86II::MRM1m:
@ -572,13 +565,12 @@ void X86MCCodeEmitter::EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
VEX_R = 0x0;
CurOp++;
// FIXME: BEXTR uses VEX.vvvv for Operand 3
if (HasVEX_4V && !IsBEXTR)
if (HasVEX_4V)
VEX_4V = getVEXRegisterEncoding(MI, CurOp++);
if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
VEX_B = 0x0;
CurOp++;
if (IsBEXTR)
if (HasVEX_4VOp3)
VEX_4V = getVEXRegisterEncoding(MI, CurOp);
break;
case X86II::MRMDestReg:
@ -874,15 +866,11 @@ EncodeInstruction(const MCInst &MI, raw_ostream &OS,
unsigned CurByte = 0;
// Is this instruction encoded using the AVX VEX prefix?
bool HasVEXPrefix = false;
bool HasVEXPrefix = (TSFlags >> X86II::VEXShift) & X86II::VEX;
// It uses the VEX.VVVV field?
bool HasVEX_4V = false;
if ((TSFlags >> X86II::VEXShift) & X86II::VEX)
HasVEXPrefix = true;
if ((TSFlags >> X86II::VEXShift) & X86II::VEX_4V)
HasVEX_4V = true;
bool HasVEX_4V = (TSFlags >> X86II::VEXShift) & X86II::VEX_4V;
bool HasVEX_4VOp3 = (TSFlags >> X86II::VEXShift) & X86II::VEX_4VOp3;
// Determine where the memory operand starts, if present.
int MemoryOperand = X86II::getMemoryOperandNo(TSFlags, Opcode);
@ -898,10 +886,6 @@ EncodeInstruction(const MCInst &MI, raw_ostream &OS,
if ((TSFlags >> X86II::VEXShift) & X86II::Has3DNow0F0FOpcode)
BaseOpcode = 0x0F; // Weird 3DNow! encoding.
// FIXME: BEXTR uses VEX.vvvv for Operand 3 instead of Operand 2
bool IsBEXTR = (Opcode == X86::BEXTR32rr || Opcode == X86::BEXTR32rm ||
Opcode == X86::BEXTR64rr || Opcode == X86::BEXTR64rm);
unsigned SrcRegNum = 0;
switch (TSFlags & X86II::FormMask) {
case X86II::MRMInitReg:
@ -956,20 +940,20 @@ EncodeInstruction(const MCInst &MI, raw_ostream &OS,
EmitByte(BaseOpcode, CurByte, OS);
SrcRegNum = CurOp + 1;
if (HasVEX_4V && !IsBEXTR) // Skip 1st src (which is encoded in VEX_VVVV)
if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
SrcRegNum++;
EmitRegModRMByte(MI.getOperand(SrcRegNum),
GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
CurOp = SrcRegNum + 1;
if (IsBEXTR)
if (HasVEX_4VOp3)
++CurOp;
break;
case X86II::MRMSrcMem: {
int AddrOperands = X86::AddrNumOperands;
unsigned FirstMemOp = CurOp+1;
if (HasVEX_4V && !IsBEXTR) {
if (HasVEX_4V) {
++AddrOperands;
++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV).
}
@ -979,7 +963,7 @@ EncodeInstruction(const MCInst &MI, raw_ostream &OS,
EmitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)),
TSFlags, CurByte, OS, Fixups);
CurOp += AddrOperands + 1;
if (IsBEXTR)
if (HasVEX_4VOp3)
++CurOp;
break;
}

View File

@ -106,6 +106,8 @@ def FeatureLZCNT : SubtargetFeature<"lzcnt", "HasLZCNT", "true",
"Support LZCNT instruction">;
def FeatureBMI : SubtargetFeature<"bmi", "HasBMI", "true",
"Support BMI instructions">;
def FeatureBMI2 : SubtargetFeature<"bmi2", "HasBMI2", "true",
"Support BMI2 instructions">;
//===----------------------------------------------------------------------===//
// X86 processors supported.
@ -164,7 +166,7 @@ def : Proc<"core-avx-i", [FeatureSSE42, FeatureCMPXCHG16B,
def : Proc<"core-avx2", [FeatureSSE42, FeatureCMPXCHG16B, FeatureAES,
FeatureCLMUL, FeatureRDRAND, FeatureF16C,
FeatureFMA3, FeatureMOVBE, FeatureLZCNT,
FeatureBMI]>;
FeatureBMI, FeatureBMI2]>;
def : Proc<"k6", [FeatureMMX]>;
def : Proc<"k6-2", [Feature3DNow]>;

View File

@ -111,6 +111,7 @@ class TF { bits<5> Prefix = 17; }
class VEX { bit hasVEXPrefix = 1; }
class VEX_W { bit hasVEX_WPrefix = 1; }
class VEX_4V : VEX { bit hasVEX_4VPrefix = 1; }
class VEX_4VOp3 : VEX { bit hasVEX_4VOp3Prefix = 1; }
class VEX_I8IMM { bit hasVEX_i8ImmReg = 1; }
class VEX_L { bit hasVEX_L = 1; }
class VEX_LIG { bit ignoresVEX_L = 1; }
@ -148,6 +149,8 @@ class X86Inst<bits<8> opcod, Format f, ImmType i, dag outs, dag ins,
bit hasVEXPrefix = 0; // Does this inst require a VEX prefix?
bit hasVEX_WPrefix = 0; // Does this inst set the VEX_W field?
bit hasVEX_4VPrefix = 0; // Does this inst require the VEX.VVVV field?
bit hasVEX_4VOp3Prefix = 0; // Does this inst require the VEX.VVVV field to
// encode the third operand?
bit hasVEX_i8ImmReg = 0; // Does this inst require the last source register
// to be encoded in a immediate field?
bit hasVEX_L = 0; // Does this inst use large (256-bit) registers?
@ -169,10 +172,11 @@ class X86Inst<bits<8> opcod, Format f, ImmType i, dag outs, dag ins,
let TSFlags{33} = hasVEXPrefix;
let TSFlags{34} = hasVEX_WPrefix;
let TSFlags{35} = hasVEX_4VPrefix;
let TSFlags{36} = hasVEX_i8ImmReg;
let TSFlags{37} = hasVEX_L;
let TSFlags{38} = ignoresVEX_L;
let TSFlags{39} = has3DNow0F0FOpcode;
let TSFlags{36} = hasVEX_4VOp3Prefix;
let TSFlags{37} = hasVEX_i8ImmReg;
let TSFlags{38} = hasVEX_L;
let TSFlags{39} = ignoresVEX_L;
let TSFlags{40} = has3DNow0F0FOpcode;
}
class PseudoI<dag oops, dag iops, list<dag> pattern>

View File

@ -480,6 +480,7 @@ def HasRDRAND : Predicate<"Subtarget->hasRDRAND()">;
def HasF16C : Predicate<"Subtarget->hasF16C()">;
def HasLZCNT : Predicate<"Subtarget->hasLZCNT()">;
def HasBMI : Predicate<"Subtarget->hasBMI()">;
def HasBMI2 : Predicate<"Subtarget->hasBMI2()">;
def FPStackf32 : Predicate<"!Subtarget->hasXMM()">;
def FPStackf64 : Predicate<"!Subtarget->hasXMMInt()">;
def HasCmpxchg16b: Predicate<"Subtarget->hasCmpxchg16b()">;
@ -1422,19 +1423,24 @@ let Predicates = [HasBMI], Defs = [EFLAGS] in {
defm BLSI64 : bmi_bls<"blsi{q}", MRM3r, MRM3m, GR64, i64mem>, VEX_W;
}
multiclass bmi_bextr<string mnemonic, RegisterClass RC,
X86MemOperand x86memop> {
def rr : I<0xF7, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
multiclass bmi_bextr_bzhi<bits<8> opc, string mnemonic, RegisterClass RC,
X86MemOperand x86memop> {
def rr : I<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
!strconcat(mnemonic, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
[]>, T8, VEX_4V;
def rm : I<0xF7, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src1, RC:$src2),
[]>, T8, VEX_4VOp3;
def rm : I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src1, RC:$src2),
!strconcat(mnemonic, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
[]>, T8, VEX_4V;
[]>, T8, VEX_4VOp3;
}
let Predicates = [HasBMI], Defs = [EFLAGS] in {
defm BEXTR32 : bmi_bextr<"bextr{l}", GR32, i32mem>;
defm BEXTR64 : bmi_bextr<"bextr{q}", GR64, i64mem>, VEX_W;
defm BEXTR32 : bmi_bextr_bzhi<0xF7, "bextr{l}", GR32, i32mem>;
defm BEXTR64 : bmi_bextr_bzhi<0xF7, "bextr{q}", GR64, i64mem>, VEX_W;
}
let Predicates = [HasBMI2], Defs = [EFLAGS] in {
defm BZHI32 : bmi_bextr_bzhi<0xF5, "bzhi{l}", GR32, i32mem>;
defm BZHI64 : bmi_bextr_bzhi<0xF5, "bzhi{q}", GR64, i64mem>, VEX_W;
}
//===----------------------------------------------------------------------===//

View File

@ -283,6 +283,10 @@ void X86Subtarget::AutoDetectSubtargetFeatures() {
HasBMI = true;
ToggleFeature(X86::FeatureBMI);
}
if ((EBX >> 8) & 0x1) {
HasBMI2 = true;
ToggleFeature(X86::FeatureBMI2);
}
}
}
@ -307,6 +311,7 @@ X86Subtarget::X86Subtarget(const std::string &TT, const std::string &CPU,
, HasF16C(false)
, HasLZCNT(false)
, HasBMI(false)
, HasBMI2(false)
, IsBTMemSlow(false)
, IsUAMemFast(false)
, HasVectorUAMem(false)

View File

@ -105,6 +105,9 @@ protected:
/// HasBMI - Processor has BMI1 instructions.
bool HasBMI;
/// HasBMI2 - Processor has BMI2 instructions.
bool HasBMI2;
/// IsBTMemSlow - True if BT (bit test) of memory instructions are slow.
bool IsBTMemSlow;
@ -192,6 +195,7 @@ public:
bool hasF16C() const { return HasF16C; }
bool hasLZCNT() const { return HasLZCNT; }
bool hasBMI() const { return HasBMI; }
bool hasBMI2() const { return HasBMI2; }
bool isBTMemSlow() const { return IsBTMemSlow; }
bool isUnalignedMemAccessFast() const { return IsUAMemFast; }
bool hasVectorUAMem() const { return HasVectorUAMem; }

View File

@ -551,3 +551,15 @@
# CHECK: bextrq %r12, %r11, %r10
0xc4 0x42 0x98 0xf7 0xd3
# CHECK: bzhil %r12d, (%rax), %r10d
0xc4 0x62 0x18 0xf5 0x10
# CHECK: bzhil %r12d, %r11d, %r10d
0xc4 0x42 0x18 0xf5 0xd3
# CHECK: bzhiq %r12, (%rax), %r10
0xc4 0x62 0x98 0xf5 0x10
# CHECK: bzhiq %r12, %r11, %r10
0xc4 0x42 0x98 0xf5 0xd3

View File

@ -513,3 +513,9 @@
# CHECK: bextrl %esi, %ebx, %edx
0xc4 0xe2 0x08 0xf7 0xd3
# CHECK: bzhil %esi, (%eax), %edx
0xc4 0xe2 0x08 0xf5 0x10
# CHECK: bzhil %esi, %ebx, %edx
0xc4 0xe2 0x08 0xf5 0xd3

View File

@ -71,3 +71,19 @@
// CHECK: bextrq %r12, %r11, %r10
// CHECK: encoding: [0xc4,0x42,0x98,0xf7,0xd3]
bextrq %r12, %r11, %r10
// CHECK: bzhil %r12d, (%rax), %r10d
// CHECK: encoding: [0xc4,0x62,0x18,0xf5,0x10]
bzhil %r12d, (%rax), %r10d
// CHECK: bzhil %r12d, %r11d, %r10d
// CHECK: encoding: [0xc4,0x42,0x18,0xf5,0xd3]
bzhil %r12d, %r11d, %r10d
// CHECK: bzhiq %r12, (%rax), %r10
// CHECK: encoding: [0xc4,0x62,0x98,0xf5,0x10]
bzhiq %r12, (%rax), %r10
// CHECK: bzhiq %r12, %r11, %r10
// CHECK: encoding: [0xc4,0x42,0x98,0xf5,0xd3]
bzhiq %r12, %r11, %r10

View File

@ -219,6 +219,7 @@ RecognizableInstr::RecognizableInstr(DisassemblerTables &tables,
HasREX_WPrefix = Rec->getValueAsBit("hasREX_WPrefix");
HasVEXPrefix = Rec->getValueAsBit("hasVEXPrefix");
HasVEX_4VPrefix = Rec->getValueAsBit("hasVEX_4VPrefix");
HasVEX_4VOp3Prefix = Rec->getValueAsBit("hasVEX_4VOp3Prefix");
HasVEX_WPrefix = Rec->getValueAsBit("hasVEX_WPrefix");
IgnoresVEX_L = Rec->getValueAsBit("ignoresVEX_L");
HasLockPrefix = Rec->getValueAsBit("hasLockPrefix");
@ -261,9 +262,6 @@ RecognizableInstr::RecognizableInstr(DisassemblerTables &tables,
Rec->getName().find("PUSH64") != Name.npos ||
Rec->getName().find("POP64") != Name.npos;
// FIXME: BEXTR uses VEX.vvvv to encode its third operand
IsBEXTR = Rec->getName().find("BEXTR") != Name.npos;
ShouldBeEmitted = true;
}
@ -286,7 +284,7 @@ void RecognizableInstr::processInstr(DisassemblerTables &tables,
InstructionContext RecognizableInstr::insnContext() const {
InstructionContext insnContext;
if (HasVEX_4VPrefix || HasVEXPrefix) {
if (HasVEX_4VPrefix || HasVEX_4VOp3Prefix|| HasVEXPrefix) {
if (HasVEX_LPrefix && HasVEX_WPrefix)
llvm_unreachable("Don't support VEX.L and VEX.W together");
else if (HasOpSizePrefix && HasVEX_LPrefix)
@ -686,7 +684,7 @@ void RecognizableInstr::emitInstructionSpecifier(DisassemblerTables &tables) {
// - In AVX, there is a register operand in the VEX.vvvv field here -
// Operand 3 (optional) is an immediate.
if (HasVEX_4VPrefix)
if (HasVEX_4VPrefix || HasVEX_4VOp3Prefix)
assert(numPhysicalOperands >= 3 && numPhysicalOperands <= 4 &&
"Unexpected number of operands for MRMSrcRegFrm with VEX_4V");
else
@ -695,15 +693,14 @@ void RecognizableInstr::emitInstructionSpecifier(DisassemblerTables &tables) {
HANDLE_OPERAND(roRegister)
if (HasVEX_4VPrefix && !IsBEXTR)
if (HasVEX_4VPrefix)
// FIXME: In AVX, the register below becomes the one encoded
// in ModRMVEX and the one above the one in the VEX.VVVV field
HANDLE_OPERAND(vvvvRegister)
HANDLE_OPERAND(rmRegister)
// FIXME: BEXTR uses VEX.vvvv for Operand 3
if (IsBEXTR)
if (HasVEX_4VOp3Prefix)
HANDLE_OPERAND(vvvvRegister)
HANDLE_OPTIONAL(immediate)
@ -714,7 +711,7 @@ void RecognizableInstr::emitInstructionSpecifier(DisassemblerTables &tables) {
// - In AVX, there is a register operand in the VEX.vvvv field here -
// Operand 3 (optional) is an immediate.
if (HasVEX_4VPrefix)
if (HasVEX_4VPrefix || HasVEX_4VOp3Prefix)
assert(numPhysicalOperands >= 3 && numPhysicalOperands <= 4 &&
"Unexpected number of operands for MRMSrcMemFrm with VEX_4V");
else
@ -723,15 +720,14 @@ void RecognizableInstr::emitInstructionSpecifier(DisassemblerTables &tables) {
HANDLE_OPERAND(roRegister)
if (HasVEX_4VPrefix && !IsBEXTR)
if (HasVEX_4VPrefix)
// FIXME: In AVX, the register below becomes the one encoded
// in ModRMVEX and the one above the one in the VEX.VVVV field
HANDLE_OPERAND(vvvvRegister)
HANDLE_OPERAND(memory)
// FIXME: BEXTR uses VEX.vvvv for Operand 3
if (IsBEXTR)
if (HasVEX_4VOp3Prefix)
HANDLE_OPERAND(vvvvRegister)
HANDLE_OPTIONAL(immediate)

View File

@ -56,6 +56,8 @@ private:
bool HasVEXPrefix;
/// The hasVEX_4VPrefix field from the record
bool HasVEX_4VPrefix;
/// The hasVEX_4VOp3Prefix field from the record
bool HasVEX_4VOp3Prefix;
/// The hasVEX_WPrefix field from the record
bool HasVEX_WPrefix;
/// Inferred from the operands; indicates whether the L bit in the VEX prefix is set
@ -70,8 +72,6 @@ private:
bool Is64Bit;
// Whether the instruction has the predicate "In32BitMode"
bool Is32Bit;
// Whether the instruction is BEXTR
bool IsBEXTR;
/// The instruction name as listed in the tables
std::string Name;