1//===-- X86InstrExtension.td - Sign and Zero Extensions ----*- tablegen -*-===// 2// 3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4// See https://llvm.org/LICENSE.txt for license information. 5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6// 7//===----------------------------------------------------------------------===// 8// 9// This file describes the sign and zero extension operations. 10// 11//===----------------------------------------------------------------------===// 12 13let hasSideEffects = 0 in { 14 let Defs = [AX], Uses = [AL] in // AX = signext(AL) 15 def CBW : I<0x98, RawFrm, (outs), (ins), 16 "{cbtw|cbw}", []>, OpSize16, Sched<[WriteALU]>; 17 let Defs = [EAX], Uses = [AX] in // EAX = signext(AX) 18 def CWDE : I<0x98, RawFrm, (outs), (ins), 19 "{cwtl|cwde}", []>, OpSize32, Sched<[WriteALU]>; 20 let Defs = [RAX], Uses = [EAX] in // RAX = signext(EAX) 21 def CDQE : RI<0x98, RawFrm, (outs), (ins), 22 "{cltq|cdqe}", []>, Sched<[WriteALU]>, Requires<[In64BitMode]>; 23 24 // FIXME: CWD/CDQ/CQO shouldn't Def the A register, but the fast register 25 // allocator crashes if you remove it. 26 let Defs = [AX,DX], Uses = [AX] in // DX:AX = signext(AX) 27 def CWD : I<0x99, RawFrm, (outs), (ins), 28 "{cwtd|cwd}", []>, OpSize16, Sched<[WriteALU]>; 29 let Defs = [EAX,EDX], Uses = [EAX] in // EDX:EAX = signext(EAX) 30 def CDQ : I<0x99, RawFrm, (outs), (ins), 31 "{cltd|cdq}", []>, OpSize32, Sched<[WriteALU]>; 32 let Defs = [RAX,RDX], Uses = [RAX] in // RDX:RAX = signext(RAX) 33 def CQO : RI<0x99, RawFrm, (outs), (ins), 34 "{cqto|cqo}", []>, Sched<[WriteALU]>, Requires<[In64BitMode]>; 35} 36 37// Sign/Zero extenders 38let hasSideEffects = 0 in { 39def MOVSX16rr8 : I<0xBE, MRMSrcReg, (outs GR16:$dst), (ins GR8:$src), 40 "movs{bw|x}\t{$src, $dst|$dst, $src}", []>, 41 TB, OpSize16, Sched<[WriteALU]>; 42let mayLoad = 1 in 43def MOVSX16rm8 : I<0xBE, MRMSrcMem, (outs GR16:$dst), (ins i8mem:$src), 44 "movs{bw|x}\t{$src, $dst|$dst, $src}", []>, 45 TB, OpSize16, Sched<[WriteLoad]>; 46} // hasSideEffects = 0 47def MOVSX32rr8 : I<0xBE, MRMSrcReg, (outs GR32:$dst), (ins GR8:$src), 48 "movs{bl|x}\t{$src, $dst|$dst, $src}", 49 [(set GR32:$dst, (sext GR8:$src))]>, TB, 50 OpSize32, Sched<[WriteALU]>; 51def MOVSX32rm8 : I<0xBE, MRMSrcMem, (outs GR32:$dst), (ins i8mem :$src), 52 "movs{bl|x}\t{$src, $dst|$dst, $src}", 53 [(set GR32:$dst, (sextloadi32i8 addr:$src))]>, TB, 54 OpSize32, Sched<[WriteLoad]>; 55def MOVSX32rr16: I<0xBF, MRMSrcReg, (outs GR32:$dst), (ins GR16:$src), 56 "movs{wl|x}\t{$src, $dst|$dst, $src}", 57 [(set GR32:$dst, (sext GR16:$src))]>, TB, 58 OpSize32, Sched<[WriteALU]>; 59def MOVSX32rm16: I<0xBF, MRMSrcMem, (outs GR32:$dst), (ins i16mem:$src), 60 "movs{wl|x}\t{$src, $dst|$dst, $src}", 61 [(set GR32:$dst, (sextloadi32i16 addr:$src))]>, 62 OpSize32, TB, Sched<[WriteLoad]>; 63 64let hasSideEffects = 0 in { 65def MOVZX16rr8 : I<0xB6, MRMSrcReg, (outs GR16:$dst), (ins GR8:$src), 66 "movz{bw|x}\t{$src, $dst|$dst, $src}", []>, 67 TB, OpSize16, Sched<[WriteALU]>; 68let mayLoad = 1 in 69def MOVZX16rm8 : I<0xB6, MRMSrcMem, (outs GR16:$dst), (ins i8mem:$src), 70 "movz{bw|x}\t{$src, $dst|$dst, $src}", []>, 71 TB, OpSize16, Sched<[WriteLoad]>; 72} // hasSideEffects = 0 73def MOVZX32rr8 : I<0xB6, MRMSrcReg, (outs GR32:$dst), (ins GR8 :$src), 74 "movz{bl|x}\t{$src, $dst|$dst, $src}", 75 [(set GR32:$dst, (zext GR8:$src))]>, TB, 76 OpSize32, Sched<[WriteALU]>; 77def MOVZX32rm8 : I<0xB6, MRMSrcMem, (outs GR32:$dst), (ins i8mem :$src), 78 "movz{bl|x}\t{$src, $dst|$dst, $src}", 79 [(set GR32:$dst, (zextloadi32i8 addr:$src))]>, TB, 80 OpSize32, Sched<[WriteLoad]>; 81def MOVZX32rr16: I<0xB7, MRMSrcReg, (outs GR32:$dst), (ins GR16:$src), 82 "movz{wl|x}\t{$src, $dst|$dst, $src}", 83 [(set GR32:$dst, (zext GR16:$src))]>, TB, 84 OpSize32, Sched<[WriteALU]>; 85def MOVZX32rm16: I<0xB7, MRMSrcMem, (outs GR32:$dst), (ins i16mem:$src), 86 "movz{wl|x}\t{$src, $dst|$dst, $src}", 87 [(set GR32:$dst, (zextloadi32i16 addr:$src))]>, 88 TB, OpSize32, Sched<[WriteLoad]>; 89 90// These instructions exist as a consequence of operand size prefix having 91// control of the destination size, but not the input size. Only support them 92// for the disassembler. 93let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in { 94def MOVSX16rr16: I<0xBF, MRMSrcReg, (outs GR16:$dst), (ins GR16:$src), 95 "movs{ww|x}\t{$src, $dst|$dst, $src}", 96 []>, TB, OpSize16, Sched<[WriteALU]>, NotMemoryFoldable; 97def MOVZX16rr16: I<0xB7, MRMSrcReg, (outs GR16:$dst), (ins GR16:$src), 98 "movz{ww|x}\t{$src, $dst|$dst, $src}", 99 []>, TB, OpSize16, Sched<[WriteALU]>, NotMemoryFoldable; 100let mayLoad = 1 in { 101def MOVSX16rm16: I<0xBF, MRMSrcMem, (outs GR16:$dst), (ins i16mem:$src), 102 "movs{ww|x}\t{$src, $dst|$dst, $src}", 103 []>, OpSize16, TB, Sched<[WriteLoad]>, NotMemoryFoldable; 104def MOVZX16rm16: I<0xB7, MRMSrcMem, (outs GR16:$dst), (ins i16mem:$src), 105 "movz{ww|x}\t{$src, $dst|$dst, $src}", 106 []>, TB, OpSize16, Sched<[WriteLoad]>, NotMemoryFoldable; 107} // mayLoad = 1 108} // isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 109 110// These are the same as the regular MOVZX32rr8 and MOVZX32rm8 111// except that they use GR32_NOREX for the output operand register class 112// instead of GR32. This allows them to operate on h registers on x86-64. 113let hasSideEffects = 0, isCodeGenOnly = 1 in { 114def MOVZX32rr8_NOREX : I<0xB6, MRMSrcReg, 115 (outs GR32_NOREX:$dst), (ins GR8_NOREX:$src), 116 "movz{bl|x}\t{$src, $dst|$dst, $src}", 117 []>, TB, OpSize32, Sched<[WriteALU]>; 118let mayLoad = 1 in 119def MOVZX32rm8_NOREX : I<0xB6, MRMSrcMem, 120 (outs GR32_NOREX:$dst), (ins i8mem_NOREX:$src), 121 "movz{bl|x}\t{$src, $dst|$dst, $src}", 122 []>, TB, OpSize32, Sched<[WriteLoad]>; 123 124def MOVSX32rr8_NOREX : I<0xBE, MRMSrcReg, 125 (outs GR32_NOREX:$dst), (ins GR8_NOREX:$src), 126 "movs{bl|x}\t{$src, $dst|$dst, $src}", 127 []>, TB, OpSize32, Sched<[WriteALU]>; 128let mayLoad = 1 in 129def MOVSX32rm8_NOREX : I<0xBE, MRMSrcMem, 130 (outs GR32_NOREX:$dst), (ins i8mem_NOREX:$src), 131 "movs{bl|x}\t{$src, $dst|$dst, $src}", 132 []>, TB, OpSize32, Sched<[WriteLoad]>; 133} 134 135// MOVSX64rr8 always has a REX prefix and it has an 8-bit register 136// operand, which makes it a rare instruction with an 8-bit register 137// operand that can never access an h register. If support for h registers 138// were generalized, this would require a special register class. 139def MOVSX64rr8 : RI<0xBE, MRMSrcReg, (outs GR64:$dst), (ins GR8 :$src), 140 "movs{bq|x}\t{$src, $dst|$dst, $src}", 141 [(set GR64:$dst, (sext GR8:$src))]>, TB, 142 Sched<[WriteALU]>; 143def MOVSX64rm8 : RI<0xBE, MRMSrcMem, (outs GR64:$dst), (ins i8mem :$src), 144 "movs{bq|x}\t{$src, $dst|$dst, $src}", 145 [(set GR64:$dst, (sextloadi64i8 addr:$src))]>, 146 TB, Sched<[WriteLoad]>; 147def MOVSX64rr16: RI<0xBF, MRMSrcReg, (outs GR64:$dst), (ins GR16:$src), 148 "movs{wq|x}\t{$src, $dst|$dst, $src}", 149 [(set GR64:$dst, (sext GR16:$src))]>, TB, 150 Sched<[WriteALU]>; 151def MOVSX64rm16: RI<0xBF, MRMSrcMem, (outs GR64:$dst), (ins i16mem:$src), 152 "movs{wq|x}\t{$src, $dst|$dst, $src}", 153 [(set GR64:$dst, (sextloadi64i16 addr:$src))]>, 154 TB, Sched<[WriteLoad]>; 155def MOVSX64rr32: RI<0x63, MRMSrcReg, (outs GR64:$dst), (ins GR32:$src), 156 "movs{lq|xd}\t{$src, $dst|$dst, $src}", 157 [(set GR64:$dst, (sext GR32:$src))]>, 158 Sched<[WriteALU]>, Requires<[In64BitMode]>; 159def MOVSX64rm32: RI<0x63, MRMSrcMem, (outs GR64:$dst), (ins i32mem:$src), 160 "movs{lq|xd}\t{$src, $dst|$dst, $src}", 161 [(set GR64:$dst, (sextloadi64i32 addr:$src))]>, 162 Sched<[WriteLoad]>, Requires<[In64BitMode]>; 163 164// These instructions exist as a consequence of operand size prefix having 165// control of the destination size, but not the input size. Only support them 166// for the disassembler. 167let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in { 168def MOVSX16rr32: I<0x63, MRMSrcReg, (outs GR16:$dst), (ins GR32:$src), 169 "movs{lq|xd}\t{$src, $dst|$dst, $src}", []>, 170 Sched<[WriteALU]>, OpSize16, Requires<[In64BitMode]>; 171def MOVSX32rr32: I<0x63, MRMSrcReg, (outs GR32:$dst), (ins GR32:$src), 172 "movs{lq|xd}\t{$src, $dst|$dst, $src}", []>, 173 Sched<[WriteALU]>, OpSize32, Requires<[In64BitMode]>; 174let mayLoad = 1 in { 175def MOVSX16rm32: I<0x63, MRMSrcMem, (outs GR16:$dst), (ins i32mem:$src), 176 "movs{lq|xd}\t{$src, $dst|$dst, $src}", []>, 177 Sched<[WriteLoad]>, OpSize16, Requires<[In64BitMode]>; 178def MOVSX32rm32: I<0x63, MRMSrcMem, (outs GR32:$dst), (ins i32mem:$src), 179 "movs{lq|xd}\t{$src, $dst|$dst, $src}", []>, 180 Sched<[WriteLoad]>, OpSize32, Requires<[In64BitMode]>; 181} // mayLoad = 1 182} // isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 183 184// movzbq and movzwq encodings for the disassembler 185let hasSideEffects = 0 in { 186def MOVZX64rr8 : RI<0xB6, MRMSrcReg, (outs GR64:$dst), (ins GR8:$src), 187 "movz{bq|x}\t{$src, $dst|$dst, $src}", []>, 188 TB, Sched<[WriteALU]>; 189let mayLoad = 1 in 190def MOVZX64rm8 : RI<0xB6, MRMSrcMem, (outs GR64:$dst), (ins i8mem:$src), 191 "movz{bq|x}\t{$src, $dst|$dst, $src}", []>, 192 TB, Sched<[WriteLoad]>; 193def MOVZX64rr16 : RI<0xB7, MRMSrcReg, (outs GR64:$dst), (ins GR16:$src), 194 "movz{wq|x}\t{$src, $dst|$dst, $src}", []>, 195 TB, Sched<[WriteALU]>; 196let mayLoad = 1 in 197def MOVZX64rm16 : RI<0xB7, MRMSrcMem, (outs GR64:$dst), (ins i16mem:$src), 198 "movz{wq|x}\t{$src, $dst|$dst, $src}", []>, 199 TB, Sched<[WriteLoad]>; 200} 201 202// 64-bit zero-extension patterns use SUBREG_TO_REG and an operation writing a 203// 32-bit register. 204def : Pat<(i64 (zext GR8:$src)), 205 (SUBREG_TO_REG (i64 0), (MOVZX32rr8 GR8:$src), sub_32bit)>; 206def : Pat<(zextloadi64i8 addr:$src), 207 (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>; 208 209def : Pat<(i64 (zext GR16:$src)), 210 (SUBREG_TO_REG (i64 0), (MOVZX32rr16 GR16:$src), sub_32bit)>; 211def : Pat<(zextloadi64i16 addr:$src), 212 (SUBREG_TO_REG (i64 0), (MOVZX32rm16 addr:$src), sub_32bit)>; 213 214// The preferred way to do 32-bit-to-64-bit zero extension on x86-64 is to use a 215// SUBREG_TO_REG to utilize implicit zero-extension, however this isn't possible 216// when the 32-bit value is defined by a truncate or is copied from something 217// where the high bits aren't necessarily all zero. In such cases, we fall back 218// to these explicit zext instructions. 219def : Pat<(i64 (zext GR32:$src)), 220 (SUBREG_TO_REG (i64 0), (MOV32rr GR32:$src), sub_32bit)>; 221def : Pat<(i64 (zextloadi64i32 addr:$src)), 222 (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src), sub_32bit)>; 223