1 //===----- AArch64InstrInfo.td - AArch64 Instruction Info ----*- tablegen -*-=// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file describes the AArch64 scalar instructions in TableGen format. 11 // 12 //===----------------------------------------------------------------------===// 13 14 include "AArch64InstrFormats.td" 15 16 //===----------------------------------------------------------------------===// 17 // Target-specific ISD nodes and profiles 18 //===----------------------------------------------------------------------===// 19 20 def SDT_A64ret : SDTypeProfile<0, 0, []>; 21 def A64ret : SDNode<"AArch64ISD::Ret", SDT_A64ret, [SDNPHasChain, 22 SDNPOptInGlue, 23 SDNPVariadic]>; 24 25 // (ins NZCV, Condition, Dest) 26 def SDT_A64br_cc : SDTypeProfile<0, 3, [SDTCisVT<0, i32>]>; 27 def A64br_cc : SDNode<"AArch64ISD::BR_CC", SDT_A64br_cc, [SDNPHasChain]>; 28 29 // (outs Result), (ins NZCV, IfTrue, IfFalse, Condition) 30 def SDT_A64select_cc : SDTypeProfile<1, 4, [SDTCisVT<1, i32>, 31 SDTCisSameAs<0, 2>, 32 SDTCisSameAs<2, 3>]>; 33 def A64select_cc : SDNode<"AArch64ISD::SELECT_CC", SDT_A64select_cc>; 34 35 // (outs NZCV), (ins LHS, RHS, Condition) 36 def SDT_A64setcc : SDTypeProfile<1, 3, [SDTCisVT<0, i32>, 37 SDTCisSameAs<1, 2>]>; 38 def A64setcc : SDNode<"AArch64ISD::SETCC", SDT_A64setcc>; 39 40 41 // (outs GPR64), (ins) 42 def A64threadpointer : SDNode<"AArch64ISD::THREAD_POINTER", SDTPtrLeaf>; 43 44 // A64 compares don't care about the cond really (they set all flags) so a 45 // simple binary operator is useful. 46 def A64cmp : PatFrag<(ops node:$lhs, node:$rhs), 47 (A64setcc node:$lhs, node:$rhs, cond)>; 48 49 50 // When matching a notional (CMP op1, (sub 0, op2)), we'd like to use a CMN 51 // instruction on the grounds that "op1 - (-op2) == op1 + op2". However, the C 52 // and V flags can be set differently by this operation. It comes down to 53 // whether "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are 54 // then everything is fine. If not then the optimization is wrong. Thus general 55 // comparisons are only valid if op2 != 0. 56 57 // So, finally, the only LLVM-native comparisons that don't mention C and V are 58 // SETEQ and SETNE. They're the only ones we can safely use CMN for in the 59 // absence of information about op2. 60 def equality_cond : PatLeaf<(cond), [{ 61 return N->get() == ISD::SETEQ || N->get() == ISD::SETNE; 62 }]>; 63 64 def A64cmn : PatFrag<(ops node:$lhs, node:$rhs), 65 (A64setcc node:$lhs, (sub 0, node:$rhs), equality_cond)>; 66 67 // There are two layers of indirection here, driven by the following 68 // considerations. 69 // + TableGen does not know CodeModel or Reloc so that decision should be 70 // made for a variable/address at ISelLowering. 71 // + The output of ISelLowering should be selectable (hence the Wrapper, 72 // rather than a bare target opcode) 73 def SDTAArch64Wrapper : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>, 74 SDTCisSameAs<1, 2>, 75 SDTCisVT<3, i32>, 76 SDTCisPtrTy<0>]>; 77 78 def A64WrapperSmall : SDNode<"AArch64ISD::WrapperSmall", SDTAArch64Wrapper>; 79 80 81 def SDTAArch64GOTLoad : SDTypeProfile<1, 1, [SDTCisPtrTy<0>, SDTCisPtrTy<1>]>; 82 def A64GOTLoad : SDNode<"AArch64ISD::GOTLoad", SDTAArch64GOTLoad, 83 [SDNPHasChain]>; 84 85 86 // (A64BFI LHS, RHS, LSB, Width) 87 def SDTA64BFI : SDTypeProfile<1, 4, [SDTCisSameAs<0, 1>, 88 SDTCisSameAs<1, 2>, 89 SDTCisVT<3, i64>, 90 SDTCisVT<4, i64>]>; 91 92 def A64Bfi : SDNode<"AArch64ISD::BFI", SDTA64BFI>; 93 94 // (A64EXTR HiReg, LoReg, LSB) 95 def SDTA64EXTR : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>, SDTCisSameAs<1, 2>, 96 SDTCisVT<3, i64>]>; 97 def A64Extr : SDNode<"AArch64ISD::EXTR", SDTA64EXTR>; 98 99 // (A64[SU]BFX Field, ImmR, ImmS). 100 // 101 // Note that ImmR and ImmS are already encoded for the actual instructions. The 102 // more natural LSB and Width mix together to form ImmR and ImmS, something 103 // which TableGen can't handle. 104 def SDTA64BFX : SDTypeProfile<1, 3, [SDTCisVT<2, i64>, SDTCisVT<3, i64>]>; 105 def A64Sbfx : SDNode<"AArch64ISD::SBFX", SDTA64BFX>; 106 107 def A64Ubfx : SDNode<"AArch64ISD::UBFX", SDTA64BFX>; 108 109 //===----------------------------------------------------------------------===// 110 // Call sequence pseudo-instructions 111 //===----------------------------------------------------------------------===// 112 113 114 def SDT_AArch64Call : SDTypeProfile<0, -1, [SDTCisPtrTy<0>]>; 115 def AArch64Call : SDNode<"AArch64ISD::Call", SDT_AArch64Call, 116 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>; 117 118 def AArch64tcret : SDNode<"AArch64ISD::TC_RETURN", SDT_AArch64Call, 119 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>; 120 121 // The TLSDESCCALL node is a variant call which goes to an indirectly calculated 122 // destination but needs a relocation against a fixed symbol. As such it has two 123 // certain operands: the callee and the relocated variable. 124 // 125 // The TLS ABI only allows it to be selected to a BLR instructin (with 126 // appropriate relocation). 127 def SDTTLSDescCall : SDTypeProfile<0, -2, [SDTCisPtrTy<0>, SDTCisPtrTy<1>]>; 128 129 def A64tlsdesc_blr : SDNode<"AArch64ISD::TLSDESCCALL", SDTTLSDescCall, 130 [SDNPInGlue, SDNPOutGlue, SDNPHasChain, 131 SDNPVariadic]>; 132 133 134 def SDT_AArch64CallSeqStart : SDCallSeqStart<[ SDTCisPtrTy<0> ]>; 135 def AArch64callseq_start : SDNode<"ISD::CALLSEQ_START", SDT_AArch64CallSeqStart, 136 [SDNPHasChain, SDNPOutGlue]>; 137 138 def SDT_AArch64CallSeqEnd : SDCallSeqEnd<[ SDTCisPtrTy<0>, SDTCisPtrTy<1> ]>; 139 def AArch64callseq_end : SDNode<"ISD::CALLSEQ_END", SDT_AArch64CallSeqEnd, 140 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>; 141 142 143 144 // These pseudo-instructions have special semantics by virtue of being passed to 145 // the InstrInfo constructor. CALLSEQ_START/CALLSEQ_END are produced by 146 // LowerCall to (in our case) tell the back-end about stack adjustments for 147 // arguments passed on the stack. Here we select those markers to 148 // pseudo-instructions which explicitly set the stack, and finally in the 149 // RegisterInfo we convert them to a true stack adjustment. 150 let Defs = [XSP], Uses = [XSP] in { 151 def ADJCALLSTACKDOWN : PseudoInst<(outs), (ins i64imm:$amt), 152 [(AArch64callseq_start timm:$amt)]>; 153 154 def ADJCALLSTACKUP : PseudoInst<(outs), (ins i64imm:$amt1, i64imm:$amt2), 155 [(AArch64callseq_end timm:$amt1, timm:$amt2)]>; 156 } 157 158 //===----------------------------------------------------------------------===// 159 // Atomic operation pseudo-instructions 160 //===----------------------------------------------------------------------===// 161 162 let usesCustomInserter = 1 in { 163 multiclass AtomicSizes<string opname> { 164 def _I8 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$incr), 165 [(set GPR32:$dst, (!cast<SDNode>(opname # "_8") GPR64:$ptr, GPR32:$incr))]>; 166 def _I16 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$incr), 167 [(set GPR32:$dst, (!cast<SDNode>(opname # "_16") GPR64:$ptr, GPR32:$incr))]>; 168 def _I32 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$incr), 169 [(set GPR32:$dst, (!cast<SDNode>(opname # "_32") GPR64:$ptr, GPR32:$incr))]>; 170 def _I64 : PseudoInst<(outs GPR64:$dst), (ins GPR64:$ptr, GPR64:$incr), 171 [(set GPR64:$dst, (!cast<SDNode>(opname # "_64") GPR64:$ptr, GPR64:$incr))]>; 172 } 173 } 174 175 defm ATOMIC_LOAD_ADD : AtomicSizes<"atomic_load_add">; 176 defm ATOMIC_LOAD_SUB : AtomicSizes<"atomic_load_sub">; 177 defm ATOMIC_LOAD_AND : AtomicSizes<"atomic_load_and">; 178 defm ATOMIC_LOAD_OR : AtomicSizes<"atomic_load_or">; 179 defm ATOMIC_LOAD_XOR : AtomicSizes<"atomic_load_xor">; 180 defm ATOMIC_LOAD_NAND : AtomicSizes<"atomic_load_nand">; 181 defm ATOMIC_SWAP : AtomicSizes<"atomic_swap">; 182 let Defs = [NZCV] in { 183 // These operations need a CMP to calculate the correct value 184 defm ATOMIC_LOAD_MIN : AtomicSizes<"atomic_load_min">; 185 defm ATOMIC_LOAD_MAX : AtomicSizes<"atomic_load_max">; 186 defm ATOMIC_LOAD_UMIN : AtomicSizes<"atomic_load_umin">; 187 defm ATOMIC_LOAD_UMAX : AtomicSizes<"atomic_load_umax">; 188 } 189 190 let usesCustomInserter = 1, Defs = [NZCV] in { 191 def ATOMIC_CMP_SWAP_I8 192 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$old, GPR32:$new), 193 [(set GPR32:$dst, 194 (atomic_cmp_swap_8 GPR64:$ptr, GPR32:$old, GPR32:$new))]>; 195 def ATOMIC_CMP_SWAP_I16 196 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$old, GPR32:$new), 197 [(set GPR32:$dst, 198 (atomic_cmp_swap_16 GPR64:$ptr, GPR32:$old, GPR32:$new))]>; 199 def ATOMIC_CMP_SWAP_I32 200 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$old, GPR32:$new), 201 [(set GPR32:$dst, 202 (atomic_cmp_swap_32 GPR64:$ptr, GPR32:$old, GPR32:$new))]>; 203 def ATOMIC_CMP_SWAP_I64 204 : PseudoInst<(outs GPR64:$dst), (ins GPR64:$ptr, GPR64:$old, GPR64:$new), 205 [(set GPR64:$dst, 206 (atomic_cmp_swap_64 GPR64:$ptr, GPR64:$old, GPR64:$new))]>; 207 } 208 209 //===----------------------------------------------------------------------===// 210 // Add-subtract (extended register) instructions 211 //===----------------------------------------------------------------------===// 212 // Contains: ADD, ADDS, SUB, SUBS + aliases CMN, CMP 213 214 // The RHS of these operations is conceptually a sign/zero-extended 215 // register, optionally shifted left by 1-4. The extension can be a 216 // NOP (e.g. "sxtx" sign-extending a 64-bit register to 64-bits) but 217 // must be specified with one exception: 218 219 // If one of the registers is sp/wsp then LSL is an alias for UXTW in 220 // 32-bit instructions and UXTX in 64-bit versions, the shift amount 221 // is not optional in that case (but can explicitly be 0), and the 222 // entire suffix can be skipped (e.g. "add sp, x3, x2"). 223 224 multiclass extend_operands<string PREFIX, string Diag> { 225 def _asmoperand : AsmOperandClass { 226 let Name = PREFIX; 227 let RenderMethod = "addRegExtendOperands"; 228 let PredicateMethod = "isRegExtend<A64SE::" # PREFIX # ">"; 229 let DiagnosticType = "AddSubRegExtend" # Diag; 230 } 231 232 def _operand : Operand<i64>, 233 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 4; }]> { 234 let PrintMethod = "printRegExtendOperand<A64SE::" # PREFIX # ">"; 235 let DecoderMethod = "DecodeRegExtendOperand"; 236 let ParserMatchClass = !cast<AsmOperandClass>(PREFIX # "_asmoperand"); 237 } 238 } 239 240 defm UXTB : extend_operands<"UXTB", "Small">; 241 defm UXTH : extend_operands<"UXTH", "Small">; 242 defm UXTW : extend_operands<"UXTW", "Small">; 243 defm UXTX : extend_operands<"UXTX", "Large">; 244 defm SXTB : extend_operands<"SXTB", "Small">; 245 defm SXTH : extend_operands<"SXTH", "Small">; 246 defm SXTW : extend_operands<"SXTW", "Small">; 247 defm SXTX : extend_operands<"SXTX", "Large">; 248 249 def LSL_extasmoperand : AsmOperandClass { 250 let Name = "RegExtendLSL"; 251 let RenderMethod = "addRegExtendOperands"; 252 let DiagnosticType = "AddSubRegExtendLarge"; 253 } 254 255 def LSL_extoperand : Operand<i64> { 256 let ParserMatchClass = LSL_extasmoperand; 257 } 258 259 260 // The patterns for various sign-extensions are a little ugly and 261 // non-uniform because everything has already been promoted to the 262 // legal i64 and i32 types. We'll wrap the various variants up in a 263 // class for use later. 264 class extend_types { 265 dag uxtb; dag uxth; dag uxtw; dag uxtx; 266 dag sxtb; dag sxth; dag sxtw; dag sxtx; 267 } 268 269 def extends_to_i64 : extend_types { 270 let uxtb = (and (anyext GPR32:$Rm), 255); 271 let uxth = (and (anyext GPR32:$Rm), 65535); 272 let uxtw = (zext GPR32:$Rm); 273 let uxtx = (i64 GPR64:$Rm); 274 275 let sxtb = (sext_inreg (anyext GPR32:$Rm), i8); 276 let sxth = (sext_inreg (anyext GPR32:$Rm), i16); 277 let sxtw = (sext GPR32:$Rm); 278 let sxtx = (i64 GPR64:$Rm); 279 } 280 281 282 def extends_to_i32 : extend_types { 283 let uxtb = (and GPR32:$Rm, 255); 284 let uxth = (and GPR32:$Rm, 65535); 285 let uxtw = (i32 GPR32:$Rm); 286 let uxtx = (i32 GPR32:$Rm); 287 288 let sxtb = (sext_inreg GPR32:$Rm, i8); 289 let sxth = (sext_inreg GPR32:$Rm, i16); 290 let sxtw = (i32 GPR32:$Rm); 291 let sxtx = (i32 GPR32:$Rm); 292 } 293 294 // Now, six of the extensions supported are easy and uniform: if the source size 295 // is 32-bits or less, then Rm is always a 32-bit register. We'll instantiate 296 // those instructions in one block. 297 298 // The uxtx/sxtx could potentially be merged in, but three facts dissuaded me: 299 // + It would break the naming scheme: either ADDxx_uxtx or ADDww_uxtx would 300 // be impossible. 301 // + Patterns are very different as well. 302 // + Passing different registers would be ugly (more fields in extend_types 303 // would probably be the best option). 304 multiclass addsub_exts<bit sf, bit op, bit S, string asmop, 305 SDPatternOperator opfrag, 306 dag outs, extend_types exts, RegisterClass GPRsp> { 307 def w_uxtb : A64I_addsubext<sf, op, S, 0b00, 0b000, 308 outs, 309 (ins GPRsp:$Rn, GPR32:$Rm, UXTB_operand:$Imm3), 310 !strconcat(asmop, "$Rn, $Rm, $Imm3"), 311 [(opfrag GPRsp:$Rn, (shl exts.uxtb, UXTB_operand:$Imm3))], 312 NoItinerary>; 313 def w_uxth : A64I_addsubext<sf, op, S, 0b00, 0b001, 314 outs, 315 (ins GPRsp:$Rn, GPR32:$Rm, UXTH_operand:$Imm3), 316 !strconcat(asmop, "$Rn, $Rm, $Imm3"), 317 [(opfrag GPRsp:$Rn, (shl exts.uxth, UXTH_operand:$Imm3))], 318 NoItinerary>; 319 def w_uxtw : A64I_addsubext<sf, op, S, 0b00, 0b010, 320 outs, 321 (ins GPRsp:$Rn, GPR32:$Rm, UXTW_operand:$Imm3), 322 !strconcat(asmop, "$Rn, $Rm, $Imm3"), 323 [(opfrag GPRsp:$Rn, (shl exts.uxtw, UXTW_operand:$Imm3))], 324 NoItinerary>; 325 326 def w_sxtb : A64I_addsubext<sf, op, S, 0b00, 0b100, 327 outs, 328 (ins GPRsp:$Rn, GPR32:$Rm, SXTB_operand:$Imm3), 329 !strconcat(asmop, "$Rn, $Rm, $Imm3"), 330 [(opfrag GPRsp:$Rn, (shl exts.sxtb, SXTB_operand:$Imm3))], 331 NoItinerary>; 332 def w_sxth : A64I_addsubext<sf, op, S, 0b00, 0b101, 333 outs, 334 (ins GPRsp:$Rn, GPR32:$Rm, SXTH_operand:$Imm3), 335 !strconcat(asmop, "$Rn, $Rm, $Imm3"), 336 [(opfrag GPRsp:$Rn, (shl exts.sxth, SXTH_operand:$Imm3))], 337 NoItinerary>; 338 def w_sxtw : A64I_addsubext<sf, op, S, 0b00, 0b110, 339 outs, 340 (ins GPRsp:$Rn, GPR32:$Rm, SXTW_operand:$Imm3), 341 !strconcat(asmop, "$Rn, $Rm, $Imm3"), 342 [(opfrag GPRsp:$Rn, (shl exts.sxtw, SXTW_operand:$Imm3))], 343 NoItinerary>; 344 } 345 346 // These two could be merge in with the above, but their patterns aren't really 347 // necessary and the naming-scheme would necessarily break: 348 multiclass addsub_xxtx<bit op, bit S, string asmop, SDPatternOperator opfrag, 349 dag outs> { 350 def x_uxtx : A64I_addsubext<0b1, op, S, 0b00, 0b011, 351 outs, 352 (ins GPR64xsp:$Rn, GPR64:$Rm, UXTX_operand:$Imm3), 353 !strconcat(asmop, "$Rn, $Rm, $Imm3"), 354 [(opfrag GPR64xsp:$Rn, (shl GPR64:$Rm, UXTX_operand:$Imm3))], 355 NoItinerary>; 356 357 def x_sxtx : A64I_addsubext<0b1, op, S, 0b00, 0b111, 358 outs, 359 (ins GPR64xsp:$Rn, GPR64:$Rm, SXTX_operand:$Imm3), 360 !strconcat(asmop, "$Rn, $Rm, $Imm3"), 361 [/* No Pattern: same as uxtx */], 362 NoItinerary>; 363 } 364 365 multiclass addsub_wxtx<bit op, bit S, string asmop, dag outs> { 366 def w_uxtx : A64I_addsubext<0b0, op, S, 0b00, 0b011, 367 outs, 368 (ins GPR32wsp:$Rn, GPR32:$Rm, UXTX_operand:$Imm3), 369 !strconcat(asmop, "$Rn, $Rm, $Imm3"), 370 [/* No pattern: probably same as uxtw */], 371 NoItinerary>; 372 373 def w_sxtx : A64I_addsubext<0b0, op, S, 0b00, 0b111, 374 outs, 375 (ins GPR32wsp:$Rn, GPR32:$Rm, SXTX_operand:$Imm3), 376 !strconcat(asmop, "$Rn, $Rm, $Imm3"), 377 [/* No Pattern: probably same as uxtw */], 378 NoItinerary>; 379 } 380 381 class SetRD<RegisterClass RC, SDPatternOperator op> 382 : PatFrag<(ops node:$lhs, node:$rhs), (set RC:$Rd, (op node:$lhs, node:$rhs))>; 383 class SetNZCV<SDPatternOperator op> 384 : PatFrag<(ops node:$lhs, node:$rhs), (set NZCV, (op node:$lhs, node:$rhs))>; 385 386 defm ADDxx :addsub_exts<0b1, 0b0, 0b0, "add\t$Rd, ", SetRD<GPR64xsp, add>, 387 (outs GPR64xsp:$Rd), extends_to_i64, GPR64xsp>, 388 addsub_xxtx< 0b0, 0b0, "add\t$Rd, ", SetRD<GPR64xsp, add>, 389 (outs GPR64xsp:$Rd)>; 390 defm ADDww :addsub_exts<0b0, 0b0, 0b0, "add\t$Rd, ", SetRD<GPR32wsp, add>, 391 (outs GPR32wsp:$Rd), extends_to_i32, GPR32wsp>, 392 addsub_wxtx< 0b0, 0b0, "add\t$Rd, ", 393 (outs GPR32wsp:$Rd)>; 394 defm SUBxx :addsub_exts<0b1, 0b1, 0b0, "sub\t$Rd, ", SetRD<GPR64xsp, sub>, 395 (outs GPR64xsp:$Rd), extends_to_i64, GPR64xsp>, 396 addsub_xxtx< 0b1, 0b0, "sub\t$Rd, ", SetRD<GPR64xsp, sub>, 397 (outs GPR64xsp:$Rd)>; 398 defm SUBww :addsub_exts<0b0, 0b1, 0b0, "sub\t$Rd, ", SetRD<GPR32wsp, sub>, 399 (outs GPR32wsp:$Rd), extends_to_i32, GPR32wsp>, 400 addsub_wxtx< 0b1, 0b0, "sub\t$Rd, ", 401 (outs GPR32wsp:$Rd)>; 402 403 let Defs = [NZCV] in { 404 defm ADDSxx :addsub_exts<0b1, 0b0, 0b1, "adds\t$Rd, ", SetRD<GPR64, addc>, 405 (outs GPR64:$Rd), extends_to_i64, GPR64xsp>, 406 addsub_xxtx< 0b0, 0b1, "adds\t$Rd, ", SetRD<GPR64, addc>, 407 (outs GPR64:$Rd)>; 408 defm ADDSww :addsub_exts<0b0, 0b0, 0b1, "adds\t$Rd, ", SetRD<GPR32, addc>, 409 (outs GPR32:$Rd), extends_to_i32, GPR32wsp>, 410 addsub_wxtx< 0b0, 0b1, "adds\t$Rd, ", 411 (outs GPR32:$Rd)>; 412 defm SUBSxx :addsub_exts<0b1, 0b1, 0b1, "subs\t$Rd, ", SetRD<GPR64, subc>, 413 (outs GPR64:$Rd), extends_to_i64, GPR64xsp>, 414 addsub_xxtx< 0b1, 0b1, "subs\t$Rd, ", SetRD<GPR64, subc>, 415 (outs GPR64:$Rd)>; 416 defm SUBSww :addsub_exts<0b0, 0b1, 0b1, "subs\t$Rd, ", SetRD<GPR32, subc>, 417 (outs GPR32:$Rd), extends_to_i32, GPR32wsp>, 418 addsub_wxtx< 0b1, 0b1, "subs\t$Rd, ", 419 (outs GPR32:$Rd)>; 420 421 422 let Rd = 0b11111, isCompare = 1 in { 423 defm CMNx : addsub_exts<0b1, 0b0, 0b1, "cmn\t", SetNZCV<A64cmn>, 424 (outs), extends_to_i64, GPR64xsp>, 425 addsub_xxtx< 0b0, 0b1, "cmn\t", SetNZCV<A64cmn>, (outs)>; 426 defm CMNw : addsub_exts<0b0, 0b0, 0b1, "cmn\t", SetNZCV<A64cmn>, 427 (outs), extends_to_i32, GPR32wsp>, 428 addsub_wxtx< 0b0, 0b1, "cmn\t", (outs)>; 429 defm CMPx : addsub_exts<0b1, 0b1, 0b1, "cmp\t", SetNZCV<A64cmp>, 430 (outs), extends_to_i64, GPR64xsp>, 431 addsub_xxtx< 0b1, 0b1, "cmp\t", SetNZCV<A64cmp>, (outs)>; 432 defm CMPw : addsub_exts<0b0, 0b1, 0b1, "cmp\t", SetNZCV<A64cmp>, 433 (outs), extends_to_i32, GPR32wsp>, 434 addsub_wxtx< 0b1, 0b1, "cmp\t", (outs)>; 435 } 436 } 437 438 // Now patterns for the operation without a shift being needed. No patterns are 439 // created for uxtx/sxtx since they're non-uniform and it's expected that 440 // add/sub (shifted register) will handle those cases anyway. 441 multiclass addsubext_noshift_patterns<string prefix, SDPatternOperator nodeop, 442 RegisterClass GPRsp, extend_types exts> { 443 def : Pat<(nodeop GPRsp:$Rn, exts.uxtb), 444 (!cast<Instruction>(prefix # "w_uxtb") GPRsp:$Rn, GPR32:$Rm, 0)>; 445 def : Pat<(nodeop GPRsp:$Rn, exts.uxth), 446 (!cast<Instruction>(prefix # "w_uxth") GPRsp:$Rn, GPR32:$Rm, 0)>; 447 def : Pat<(nodeop GPRsp:$Rn, exts.uxtw), 448 (!cast<Instruction>(prefix # "w_uxtw") GPRsp:$Rn, GPR32:$Rm, 0)>; 449 450 def : Pat<(nodeop GPRsp:$Rn, exts.sxtb), 451 (!cast<Instruction>(prefix # "w_sxtb") GPRsp:$Rn, GPR32:$Rm, 0)>; 452 def : Pat<(nodeop GPRsp:$Rn, exts.sxth), 453 (!cast<Instruction>(prefix # "w_sxth") GPRsp:$Rn, GPR32:$Rm, 0)>; 454 def : Pat<(nodeop GPRsp:$Rn, exts.sxtw), 455 (!cast<Instruction>(prefix # "w_sxtw") GPRsp:$Rn, GPR32:$Rm, 0)>; 456 } 457 458 defm : addsubext_noshift_patterns<"ADDxx", add, GPR64xsp, extends_to_i64>; 459 defm : addsubext_noshift_patterns<"ADDww", add, GPR32wsp, extends_to_i32>; 460 defm : addsubext_noshift_patterns<"SUBxx", sub, GPR64xsp, extends_to_i64>; 461 defm : addsubext_noshift_patterns<"SUBww", sub, GPR32wsp, extends_to_i32>; 462 463 defm : addsubext_noshift_patterns<"CMNx", A64cmn, GPR64xsp, extends_to_i64>; 464 defm : addsubext_noshift_patterns<"CMNw", A64cmn, GPR32wsp, extends_to_i32>; 465 defm : addsubext_noshift_patterns<"CMPx", A64cmp, GPR64xsp, extends_to_i64>; 466 defm : addsubext_noshift_patterns<"CMPw", A64cmp, GPR32wsp, extends_to_i32>; 467 468 // An extend of "lsl #imm" is valid if and only if one of Rn and Rd is 469 // sp/wsp. It is synonymous with uxtx/uxtw depending on the size of the 470 // operation. Also permitted in this case is complete omission of the argument, 471 // which implies "lsl #0". 472 multiclass lsl_aliases<string asmop, Instruction inst, RegisterClass GPR_Rd, 473 RegisterClass GPR_Rn, RegisterClass GPR_Rm> { 474 def : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm"), 475 (inst GPR_Rd:$Rd, GPR_Rn:$Rn, GPR_Rm:$Rm, 0)>; 476 477 def : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm, $LSL"), 478 (inst GPR_Rd:$Rd, GPR_Rn:$Rn, GPR_Rm:$Rm, LSL_extoperand:$LSL)>; 479 480 } 481 482 defm : lsl_aliases<"add", ADDxxx_uxtx, Rxsp, GPR64xsp, GPR64>; 483 defm : lsl_aliases<"add", ADDxxx_uxtx, GPR64xsp, Rxsp, GPR64>; 484 defm : lsl_aliases<"add", ADDwww_uxtw, Rwsp, GPR32wsp, GPR32>; 485 defm : lsl_aliases<"add", ADDwww_uxtw, GPR32wsp, Rwsp, GPR32>; 486 defm : lsl_aliases<"sub", SUBxxx_uxtx, Rxsp, GPR64xsp, GPR64>; 487 defm : lsl_aliases<"sub", SUBxxx_uxtx, GPR64xsp, Rxsp, GPR64>; 488 defm : lsl_aliases<"sub", SUBwww_uxtw, Rwsp, GPR32wsp, GPR32>; 489 defm : lsl_aliases<"sub", SUBwww_uxtw, GPR32wsp, Rwsp, GPR32>; 490 491 // Rd cannot be sp for flag-setting variants so only half of the aliases are 492 // needed. 493 defm : lsl_aliases<"adds", ADDSxxx_uxtx, GPR64, Rxsp, GPR64>; 494 defm : lsl_aliases<"adds", ADDSwww_uxtw, GPR32, Rwsp, GPR32>; 495 defm : lsl_aliases<"subs", SUBSxxx_uxtx, GPR64, Rxsp, GPR64>; 496 defm : lsl_aliases<"subs", SUBSwww_uxtw, GPR32, Rwsp, GPR32>; 497 498 // CMP unfortunately has to be different because the instruction doesn't have a 499 // dest register. 500 multiclass cmp_lsl_aliases<string asmop, Instruction inst, 501 RegisterClass GPR_Rn, RegisterClass GPR_Rm> { 502 def : InstAlias<!strconcat(asmop, " $Rn, $Rm"), 503 (inst GPR_Rn:$Rn, GPR_Rm:$Rm, 0)>; 504 505 def : InstAlias<!strconcat(asmop, " $Rn, $Rm, $LSL"), 506 (inst GPR_Rn:$Rn, GPR_Rm:$Rm, LSL_extoperand:$LSL)>; 507 } 508 509 defm : cmp_lsl_aliases<"cmp", CMPxx_uxtx, Rxsp, GPR64>; 510 defm : cmp_lsl_aliases<"cmp", CMPww_uxtw, Rwsp, GPR32>; 511 defm : cmp_lsl_aliases<"cmn", CMNxx_uxtx, Rxsp, GPR64>; 512 defm : cmp_lsl_aliases<"cmn", CMNww_uxtw, Rwsp, GPR32>; 513 514 //===----------------------------------------------------------------------===// 515 // Add-subtract (immediate) instructions 516 //===----------------------------------------------------------------------===// 517 // Contains: ADD, ADDS, SUB, SUBS + aliases CMN, CMP, MOV 518 519 // These instructions accept a 12-bit unsigned immediate, optionally shifted 520 // left by 12 bits. Official assembly format specifies a 12 bit immediate with 521 // one of "", "LSL #0", "LSL #12" supplementary operands. 522 523 // There are surprisingly few ways to make this work with TableGen, so this 524 // implementation has separate instructions for the "LSL #0" and "LSL #12" 525 // variants. 526 527 // If the MCInst retained a single combined immediate (which could be 0x123000, 528 // for example) then both components (imm & shift) would have to be delegated to 529 // a single assembly operand. This would entail a separate operand parser 530 // (because the LSL would have to live in the same AArch64Operand as the 531 // immediate to be accessible); assembly parsing is rather complex and 532 // error-prone C++ code. 533 // 534 // By splitting the immediate, we can delegate handling this optional operand to 535 // an InstAlias. Supporting functions to generate the correct MCInst are still 536 // required, but these are essentially trivial and parsing can remain generic. 537 // 538 // Rejected plans with rationale: 539 // ------------------------------ 540 // 541 // In an ideal world you'de have two first class immediate operands (in 542 // InOperandList, specifying imm12 and shift). Unfortunately this is not 543 // selectable by any means I could discover. 544 // 545 // An Instruction with two MCOperands hidden behind a single entry in 546 // InOperandList (expanded by ComplexPatterns and MIOperandInfo) was functional, 547 // but required more C++ code to handle encoding/decoding. Parsing (the intended 548 // main beneficiary) ended up equally complex because of the optional nature of 549 // "LSL #0". 550 // 551 // Attempting to circumvent the need for a custom OperandParser above by giving 552 // InstAliases without the "lsl #0" failed. add/sub could be accommodated but 553 // the cmp/cmn aliases didn't use the MIOperandInfo to determine how operands 554 // should be parsed: there was no way to accommodate an "lsl #12". 555 556 let ParserMethod = "ParseImmWithLSLOperand", 557 RenderMethod = "addImmWithLSLOperands" in { 558 // Derived PredicateMethod fields are different for each 559 def addsubimm_lsl0_asmoperand : AsmOperandClass { 560 let Name = "AddSubImmLSL0"; 561 // If an error is reported against this operand, instruction could also be a 562 // register variant. 563 let DiagnosticType = "AddSubSecondSource"; 564 } 565 566 def addsubimm_lsl12_asmoperand : AsmOperandClass { 567 let Name = "AddSubImmLSL12"; 568 let DiagnosticType = "AddSubSecondSource"; 569 } 570 } 571 572 def shr_12_XFORM : SDNodeXForm<imm, [{ 573 return CurDAG->getTargetConstant(N->getSExtValue() >> 12, MVT::i32); 574 }]>; 575 576 def shr_12_neg_XFORM : SDNodeXForm<imm, [{ 577 return CurDAG->getTargetConstant((-N->getSExtValue()) >> 12, MVT::i32); 578 }]>; 579 580 def neg_XFORM : SDNodeXForm<imm, [{ 581 return CurDAG->getTargetConstant(-N->getSExtValue(), MVT::i32); 582 }]>; 583 584 585 multiclass addsub_imm_operands<ValueType ty> { 586 let PrintMethod = "printAddSubImmLSL0Operand", 587 EncoderMethod = "getAddSubImmOpValue", 588 ParserMatchClass = addsubimm_lsl0_asmoperand in { 589 def _posimm_lsl0 : Operand<ty>, 590 ImmLeaf<ty, [{ return Imm >= 0 && (Imm & ~0xfff) == 0; }]>; 591 def _negimm_lsl0 : Operand<ty>, 592 ImmLeaf<ty, [{ return Imm < 0 && (-Imm & ~0xfff) == 0; }], 593 neg_XFORM>; 594 } 595 596 let PrintMethod = "printAddSubImmLSL12Operand", 597 EncoderMethod = "getAddSubImmOpValue", 598 ParserMatchClass = addsubimm_lsl12_asmoperand in { 599 def _posimm_lsl12 : Operand<ty>, 600 ImmLeaf<ty, [{ return Imm >= 0 && (Imm & ~0xfff000) == 0; }], 601 shr_12_XFORM>; 602 603 def _negimm_lsl12 : Operand<ty>, 604 ImmLeaf<ty, [{ return Imm < 0 && (-Imm & ~0xfff000) == 0; }], 605 shr_12_neg_XFORM>; 606 } 607 } 608 609 // The add operands don't need any transformation 610 defm addsubimm_operand_i32 : addsub_imm_operands<i32>; 611 defm addsubimm_operand_i64 : addsub_imm_operands<i64>; 612 613 multiclass addsubimm_varieties<string prefix, bit sf, bit op, bits<2> shift, 614 string asmop, string cmpasmop, 615 Operand imm_operand, Operand cmp_imm_operand, 616 RegisterClass GPR, RegisterClass GPRsp, 617 AArch64Reg ZR> { 618 // All registers for non-S variants allow SP 619 def _s : A64I_addsubimm<sf, op, 0b0, shift, 620 (outs GPRsp:$Rd), 621 (ins GPRsp:$Rn, imm_operand:$Imm12), 622 !strconcat(asmop, "\t$Rd, $Rn, $Imm12"), 623 [(set GPRsp:$Rd, 624 (add GPRsp:$Rn, imm_operand:$Imm12))], 625 NoItinerary>; 626 627 628 // S variants can read SP but would write to ZR 629 def _S : A64I_addsubimm<sf, op, 0b1, shift, 630 (outs GPR:$Rd), 631 (ins GPRsp:$Rn, imm_operand:$Imm12), 632 !strconcat(asmop, "s\t$Rd, $Rn, $Imm12"), 633 [(set GPR:$Rd, (addc GPRsp:$Rn, imm_operand:$Imm12))], 634 NoItinerary> { 635 let Defs = [NZCV]; 636 } 637 638 // Note that the pattern here for ADDS is subtle. Canonically CMP 639 // a, b becomes SUBS a, b. If b < 0 then this is equivalent to 640 // ADDS a, (-b). This is not true in general. 641 def _cmp : A64I_addsubimm<sf, op, 0b1, shift, 642 (outs), (ins GPRsp:$Rn, imm_operand:$Imm12), 643 !strconcat(cmpasmop, " $Rn, $Imm12"), 644 [(set NZCV, 645 (A64cmp GPRsp:$Rn, cmp_imm_operand:$Imm12))], 646 NoItinerary> { 647 let Rd = 0b11111; 648 let Defs = [NZCV]; 649 let isCompare = 1; 650 } 651 } 652 653 654 multiclass addsubimm_shifts<string prefix, bit sf, bit op, 655 string asmop, string cmpasmop, string operand, string cmpoperand, 656 RegisterClass GPR, RegisterClass GPRsp, AArch64Reg ZR> { 657 defm _lsl0 : addsubimm_varieties<prefix # "_lsl0", sf, op, 0b00, 658 asmop, cmpasmop, 659 !cast<Operand>(operand # "_lsl0"), 660 !cast<Operand>(cmpoperand # "_lsl0"), 661 GPR, GPRsp, ZR>; 662 663 defm _lsl12 : addsubimm_varieties<prefix # "_lsl12", sf, op, 0b01, 664 asmop, cmpasmop, 665 !cast<Operand>(operand # "_lsl12"), 666 !cast<Operand>(cmpoperand # "_lsl12"), 667 GPR, GPRsp, ZR>; 668 } 669 670 defm ADDwwi : addsubimm_shifts<"ADDwi", 0b0, 0b0, "add", "cmn", 671 "addsubimm_operand_i32_posimm", 672 "addsubimm_operand_i32_negimm", 673 GPR32, GPR32wsp, WZR>; 674 defm ADDxxi : addsubimm_shifts<"ADDxi", 0b1, 0b0, "add", "cmn", 675 "addsubimm_operand_i64_posimm", 676 "addsubimm_operand_i64_negimm", 677 GPR64, GPR64xsp, XZR>; 678 defm SUBwwi : addsubimm_shifts<"SUBwi", 0b0, 0b1, "sub", "cmp", 679 "addsubimm_operand_i32_negimm", 680 "addsubimm_operand_i32_posimm", 681 GPR32, GPR32wsp, WZR>; 682 defm SUBxxi : addsubimm_shifts<"SUBxi", 0b1, 0b1, "sub", "cmp", 683 "addsubimm_operand_i64_negimm", 684 "addsubimm_operand_i64_posimm", 685 GPR64, GPR64xsp, XZR>; 686 687 multiclass MOVsp<RegisterClass GPRsp, RegisterClass SP, Instruction addop> { 688 def _fromsp : InstAlias<"mov $Rd, $Rn", 689 (addop GPRsp:$Rd, SP:$Rn, 0), 690 0b1>; 691 692 def _tosp : InstAlias<"mov $Rd, $Rn", 693 (addop SP:$Rd, GPRsp:$Rn, 0), 694 0b1>; 695 } 696 697 // Recall Rxsp is a RegisterClass containing *just* xsp. 698 defm MOVxx : MOVsp<GPR64xsp, Rxsp, ADDxxi_lsl0_s>; 699 defm MOVww : MOVsp<GPR32wsp, Rwsp, ADDwwi_lsl0_s>; 700 701 //===----------------------------------------------------------------------===// 702 // Add-subtract (shifted register) instructions 703 //===----------------------------------------------------------------------===// 704 // Contains: ADD, ADDS, SUB, SUBS + aliases CMN, CMP, NEG, NEGS 705 706 //===------------------------------- 707 // 1. The "shifed register" operands. Shared with logical insts. 708 //===------------------------------- 709 710 multiclass shift_operands<string prefix, string form> { 711 def _asmoperand_i32 : AsmOperandClass { 712 let Name = "Shift" # form # "i32"; 713 let RenderMethod = "addShiftOperands"; 714 let PredicateMethod = "isShift<A64SE::" # form # ", false>"; 715 let DiagnosticType = "AddSubRegShift32"; 716 } 717 718 // Note that the operand type is intentionally i64 because the DAGCombiner 719 // puts these into a canonical form. 720 def _i32 : Operand<i64>, ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 31; }]> { 721 let ParserMatchClass 722 = !cast<AsmOperandClass>(prefix # "_asmoperand_i32"); 723 let PrintMethod = "printShiftOperand<A64SE::" # form # ">"; 724 let DecoderMethod = "Decode32BitShiftOperand"; 725 } 726 727 def _asmoperand_i64 : AsmOperandClass { 728 let Name = "Shift" # form # "i64"; 729 let RenderMethod = "addShiftOperands"; 730 let PredicateMethod = "isShift<A64SE::" # form # ", true>"; 731 let DiagnosticType = "AddSubRegShift64"; 732 } 733 734 def _i64 : Operand<i64>, ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 63; }]> { 735 let ParserMatchClass 736 = !cast<AsmOperandClass>(prefix # "_asmoperand_i64"); 737 let PrintMethod = "printShiftOperand<A64SE::" # form # ">"; 738 } 739 } 740 741 defm lsl_operand : shift_operands<"lsl_operand", "LSL">; 742 defm lsr_operand : shift_operands<"lsr_operand", "LSR">; 743 defm asr_operand : shift_operands<"asr_operand", "ASR">; 744 745 // Not used for add/sub, but defined here for completeness. The "logical 746 // (shifted register)" instructions *do* have an ROR variant. 747 defm ror_operand : shift_operands<"ror_operand", "ROR">; 748 749 //===------------------------------- 750 // 2. The basic 3.5-operand ADD/SUB/ADDS/SUBS instructions. 751 //===------------------------------- 752 753 // N.b. the commutable parameter is just !N. It will be first against the wall 754 // when the revolution comes. 755 multiclass addsub_shifts<string prefix, bit sf, bit op, bit s, bit commutable, 756 string asmop, SDPatternOperator opfrag, string sty, 757 RegisterClass GPR, list<Register> defs> { 758 let isCommutable = commutable, Defs = defs in { 759 def _lsl : A64I_addsubshift<sf, op, s, 0b00, 760 (outs GPR:$Rd), 761 (ins GPR:$Rn, GPR:$Rm, 762 !cast<Operand>("lsl_operand_" # sty):$Imm6), 763 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"), 764 [(set GPR:$Rd, (opfrag GPR:$Rn, (shl GPR:$Rm, 765 !cast<Operand>("lsl_operand_" # sty):$Imm6)) 766 )], 767 NoItinerary>; 768 769 def _lsr : A64I_addsubshift<sf, op, s, 0b01, 770 (outs GPR:$Rd), 771 (ins GPR:$Rn, GPR:$Rm, 772 !cast<Operand>("lsr_operand_" # sty):$Imm6), 773 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"), 774 [(set GPR:$Rd, (opfrag GPR:$Rn, (srl GPR:$Rm, 775 !cast<Operand>("lsr_operand_" # sty):$Imm6)) 776 )], 777 NoItinerary>; 778 779 def _asr : A64I_addsubshift<sf, op, s, 0b10, 780 (outs GPR:$Rd), 781 (ins GPR:$Rn, GPR:$Rm, 782 !cast<Operand>("asr_operand_" # sty):$Imm6), 783 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"), 784 [(set GPR:$Rd, (opfrag GPR:$Rn, (sra GPR:$Rm, 785 !cast<Operand>("asr_operand_" # sty):$Imm6)) 786 )], 787 NoItinerary>; 788 } 789 790 def _noshift 791 : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm"), 792 (!cast<Instruction>(prefix # "_lsl") GPR:$Rd, GPR:$Rn, 793 GPR:$Rm, 0)>; 794 795 def : Pat<(opfrag GPR:$Rn, GPR:$Rm), 796 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>; 797 } 798 799 multiclass addsub_sizes<string prefix, bit op, bit s, bit commutable, 800 string asmop, SDPatternOperator opfrag, 801 list<Register> defs> { 802 defm xxx : addsub_shifts<prefix # "xxx", 0b1, op, s, 803 commutable, asmop, opfrag, "i64", GPR64, defs>; 804 defm www : addsub_shifts<prefix # "www", 0b0, op, s, 805 commutable, asmop, opfrag, "i32", GPR32, defs>; 806 } 807 808 809 defm ADD : addsub_sizes<"ADD", 0b0, 0b0, 0b1, "add", add, []>; 810 defm SUB : addsub_sizes<"SUB", 0b1, 0b0, 0b0, "sub", sub, []>; 811 812 defm ADDS : addsub_sizes<"ADDS", 0b0, 0b1, 0b1, "adds", addc, [NZCV]>; 813 defm SUBS : addsub_sizes<"SUBS", 0b1, 0b1, 0b0, "subs", subc, [NZCV]>; 814 815 //===------------------------------- 816 // 1. The NEG/NEGS aliases 817 //===------------------------------- 818 819 multiclass neg_alias<Instruction INST, RegisterClass GPR, 820 Register ZR, Operand shift_operand, SDNode shiftop> { 821 def : InstAlias<"neg $Rd, $Rm, $Imm6", 822 (INST GPR:$Rd, ZR, GPR:$Rm, shift_operand:$Imm6)>; 823 824 def : Pat<(sub 0, (shiftop GPR:$Rm, shift_operand:$Imm6)), 825 (INST ZR, GPR:$Rm, shift_operand:$Imm6)>; 826 } 827 828 defm : neg_alias<SUBwww_lsl, GPR32, WZR, lsl_operand_i32, shl>; 829 defm : neg_alias<SUBwww_lsr, GPR32, WZR, lsr_operand_i32, srl>; 830 defm : neg_alias<SUBwww_asr, GPR32, WZR, asr_operand_i32, sra>; 831 def : InstAlias<"neg $Rd, $Rm", (SUBwww_lsl GPR32:$Rd, WZR, GPR32:$Rm, 0)>; 832 def : Pat<(sub 0, GPR32:$Rm), (SUBwww_lsl WZR, GPR32:$Rm, 0)>; 833 834 defm : neg_alias<SUBxxx_lsl, GPR64, XZR, lsl_operand_i64, shl>; 835 defm : neg_alias<SUBxxx_lsr, GPR64, XZR, lsr_operand_i64, srl>; 836 defm : neg_alias<SUBxxx_asr, GPR64, XZR, asr_operand_i64, sra>; 837 def : InstAlias<"neg $Rd, $Rm", (SUBxxx_lsl GPR64:$Rd, XZR, GPR64:$Rm, 0)>; 838 def : Pat<(sub 0, GPR64:$Rm), (SUBxxx_lsl XZR, GPR64:$Rm, 0)>; 839 840 // NEGS doesn't get any patterns yet: defining multiple outputs means C++ has to 841 // be involved. 842 class negs_alias<Instruction INST, RegisterClass GPR, 843 Register ZR, Operand shift_operand, SDNode shiftop> 844 : InstAlias<"negs $Rd, $Rm, $Imm6", 845 (INST GPR:$Rd, ZR, GPR:$Rm, shift_operand:$Imm6)>; 846 847 def : negs_alias<SUBSwww_lsl, GPR32, WZR, lsl_operand_i32, shl>; 848 def : negs_alias<SUBSwww_lsr, GPR32, WZR, lsr_operand_i32, srl>; 849 def : negs_alias<SUBSwww_asr, GPR32, WZR, asr_operand_i32, sra>; 850 def : InstAlias<"negs $Rd, $Rm", (SUBSwww_lsl GPR32:$Rd, WZR, GPR32:$Rm, 0)>; 851 852 def : negs_alias<SUBSxxx_lsl, GPR64, XZR, lsl_operand_i64, shl>; 853 def : negs_alias<SUBSxxx_lsr, GPR64, XZR, lsr_operand_i64, srl>; 854 def : negs_alias<SUBSxxx_asr, GPR64, XZR, asr_operand_i64, sra>; 855 def : InstAlias<"negs $Rd, $Rm", (SUBSxxx_lsl GPR64:$Rd, XZR, GPR64:$Rm, 0)>; 856 857 //===------------------------------- 858 // 1. The CMP/CMN aliases 859 //===------------------------------- 860 861 multiclass cmp_shifts<string prefix, bit sf, bit op, bit commutable, 862 string asmop, SDPatternOperator opfrag, string sty, 863 RegisterClass GPR> { 864 let isCommutable = commutable, Rd = 0b11111, Defs = [NZCV] in { 865 def _lsl : A64I_addsubshift<sf, op, 0b1, 0b00, 866 (outs), 867 (ins GPR:$Rn, GPR:$Rm, 868 !cast<Operand>("lsl_operand_" # sty):$Imm6), 869 !strconcat(asmop, "\t$Rn, $Rm, $Imm6"), 870 [(set NZCV, (opfrag GPR:$Rn, (shl GPR:$Rm, 871 !cast<Operand>("lsl_operand_" # sty):$Imm6)) 872 )], 873 NoItinerary>; 874 875 def _lsr : A64I_addsubshift<sf, op, 0b1, 0b01, 876 (outs), 877 (ins GPR:$Rn, GPR:$Rm, 878 !cast<Operand>("lsr_operand_" # sty):$Imm6), 879 !strconcat(asmop, "\t$Rn, $Rm, $Imm6"), 880 [(set NZCV, (opfrag GPR:$Rn, (srl GPR:$Rm, 881 !cast<Operand>("lsr_operand_" # sty):$Imm6)) 882 )], 883 NoItinerary>; 884 885 def _asr : A64I_addsubshift<sf, op, 0b1, 0b10, 886 (outs), 887 (ins GPR:$Rn, GPR:$Rm, 888 !cast<Operand>("asr_operand_" # sty):$Imm6), 889 !strconcat(asmop, "\t$Rn, $Rm, $Imm6"), 890 [(set NZCV, (opfrag GPR:$Rn, (sra GPR:$Rm, 891 !cast<Operand>("asr_operand_" # sty):$Imm6)) 892 )], 893 NoItinerary>; 894 } 895 896 def _noshift 897 : InstAlias<!strconcat(asmop, " $Rn, $Rm"), 898 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>; 899 900 def : Pat<(opfrag GPR:$Rn, GPR:$Rm), 901 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>; 902 } 903 904 defm CMPww : cmp_shifts<"CMPww", 0b0, 0b1, 0b0, "cmp", A64cmp, "i32", GPR32>; 905 defm CMPxx : cmp_shifts<"CMPxx", 0b1, 0b1, 0b0, "cmp", A64cmp, "i64", GPR64>; 906 907 defm CMNww : cmp_shifts<"CMNww", 0b0, 0b0, 0b1, "cmn", A64cmn, "i32", GPR32>; 908 defm CMNxx : cmp_shifts<"CMNxx", 0b1, 0b0, 0b1, "cmn", A64cmn, "i64", GPR64>; 909 910 //===----------------------------------------------------------------------===// 911 // Add-subtract (with carry) instructions 912 //===----------------------------------------------------------------------===// 913 // Contains: ADC, ADCS, SBC, SBCS + aliases NGC, NGCS 914 915 multiclass A64I_addsubcarrySizes<bit op, bit s, string asmop> { 916 let Uses = [NZCV] in { 917 def www : A64I_addsubcarry<0b0, op, s, 0b000000, 918 (outs GPR32:$Rd), (ins GPR32:$Rn, GPR32:$Rm), 919 !strconcat(asmop, "\t$Rd, $Rn, $Rm"), 920 [], NoItinerary>; 921 922 def xxx : A64I_addsubcarry<0b1, op, s, 0b000000, 923 (outs GPR64:$Rd), (ins GPR64:$Rn, GPR64:$Rm), 924 !strconcat(asmop, "\t$Rd, $Rn, $Rm"), 925 [], NoItinerary>; 926 } 927 } 928 929 let isCommutable = 1 in { 930 defm ADC : A64I_addsubcarrySizes<0b0, 0b0, "adc">; 931 } 932 933 defm SBC : A64I_addsubcarrySizes<0b1, 0b0, "sbc">; 934 935 let Defs = [NZCV] in { 936 let isCommutable = 1 in { 937 defm ADCS : A64I_addsubcarrySizes<0b0, 0b1, "adcs">; 938 } 939 940 defm SBCS : A64I_addsubcarrySizes<0b1, 0b1, "sbcs">; 941 } 942 943 def : InstAlias<"ngc $Rd, $Rm", (SBCwww GPR32:$Rd, WZR, GPR32:$Rm)>; 944 def : InstAlias<"ngc $Rd, $Rm", (SBCxxx GPR64:$Rd, XZR, GPR64:$Rm)>; 945 def : InstAlias<"ngcs $Rd, $Rm", (SBCSwww GPR32:$Rd, WZR, GPR32:$Rm)>; 946 def : InstAlias<"ngcs $Rd, $Rm", (SBCSxxx GPR64:$Rd, XZR, GPR64:$Rm)>; 947 948 // Note that adde and sube can form a chain longer than two (e.g. for 256-bit 949 // addition). So the flag-setting instructions are appropriate. 950 def : Pat<(adde GPR32:$Rn, GPR32:$Rm), (ADCSwww GPR32:$Rn, GPR32:$Rm)>; 951 def : Pat<(adde GPR64:$Rn, GPR64:$Rm), (ADCSxxx GPR64:$Rn, GPR64:$Rm)>; 952 def : Pat<(sube GPR32:$Rn, GPR32:$Rm), (SBCSwww GPR32:$Rn, GPR32:$Rm)>; 953 def : Pat<(sube GPR64:$Rn, GPR64:$Rm), (SBCSxxx GPR64:$Rn, GPR64:$Rm)>; 954 955 //===----------------------------------------------------------------------===// 956 // Bitfield 957 //===----------------------------------------------------------------------===// 958 // Contains: SBFM, BFM, UBFM, [SU]XT[BHW], ASR, LSR, LSL, SBFI[ZX], BFI, BFXIL, 959 // UBFIZ, UBFX 960 961 // Because of the rather complicated nearly-overlapping aliases, the decoding of 962 // this range of instructions is handled manually. The architectural 963 // instructions are BFM, SBFM and UBFM but a disassembler should never produce 964 // these. 965 // 966 // In the end, the best option was to use BFM instructions for decoding under 967 // almost all circumstances, but to create aliasing *Instructions* for each of 968 // the canonical forms and specify a completely custom decoder which would 969 // substitute the correct MCInst as needed. 970 // 971 // This also simplifies instruction selection, parsing etc because the MCInsts 972 // have a shape that's closer to their use in code. 973 974 //===------------------------------- 975 // 1. The architectural BFM instructions 976 //===------------------------------- 977 978 def uimm5_asmoperand : AsmOperandClass { 979 let Name = "UImm5"; 980 let PredicateMethod = "isUImm<5>"; 981 let RenderMethod = "addImmOperands"; 982 let DiagnosticType = "UImm5"; 983 } 984 985 def uimm6_asmoperand : AsmOperandClass { 986 let Name = "UImm6"; 987 let PredicateMethod = "isUImm<6>"; 988 let RenderMethod = "addImmOperands"; 989 let DiagnosticType = "UImm6"; 990 } 991 992 def bitfield32_imm : Operand<i64>, 993 ImmLeaf<i64, [{ return Imm >= 0 && Imm < 32; }]> { 994 let ParserMatchClass = uimm5_asmoperand; 995 996 let DecoderMethod = "DecodeBitfield32ImmOperand"; 997 } 998 999 1000 def bitfield64_imm : Operand<i64>, 1001 ImmLeaf<i64, [{ return Imm >= 0 && Imm < 64; }]> { 1002 let ParserMatchClass = uimm6_asmoperand; 1003 1004 // Default decoder works in 64-bit case: the 6-bit field can take any value. 1005 } 1006 1007 multiclass A64I_bitfieldSizes<bits<2> opc, string asmop> { 1008 def wwii : A64I_bitfield<0b0, opc, 0b0, (outs GPR32:$Rd), 1009 (ins GPR32:$Rn, bitfield32_imm:$ImmR, bitfield32_imm:$ImmS), 1010 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"), 1011 [], NoItinerary> { 1012 let DecoderMethod = "DecodeBitfieldInstruction"; 1013 } 1014 1015 def xxii : A64I_bitfield<0b1, opc, 0b1, (outs GPR64:$Rd), 1016 (ins GPR64:$Rn, bitfield64_imm:$ImmR, bitfield64_imm:$ImmS), 1017 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"), 1018 [], NoItinerary> { 1019 let DecoderMethod = "DecodeBitfieldInstruction"; 1020 } 1021 } 1022 1023 defm SBFM : A64I_bitfieldSizes<0b00, "sbfm">; 1024 defm UBFM : A64I_bitfieldSizes<0b10, "ubfm">; 1025 1026 // BFM instructions modify the destination register rather than defining it 1027 // completely. 1028 def BFMwwii : 1029 A64I_bitfield<0b0, 0b01, 0b0, (outs GPR32:$Rd), 1030 (ins GPR32:$src, GPR32:$Rn, bitfield32_imm:$ImmR, bitfield32_imm:$ImmS), 1031 "bfm\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> { 1032 let DecoderMethod = "DecodeBitfieldInstruction"; 1033 let Constraints = "$src = $Rd"; 1034 } 1035 1036 def BFMxxii : 1037 A64I_bitfield<0b1, 0b01, 0b1, (outs GPR64:$Rd), 1038 (ins GPR64:$src, GPR64:$Rn, bitfield64_imm:$ImmR, bitfield64_imm:$ImmS), 1039 "bfm\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> { 1040 let DecoderMethod = "DecodeBitfieldInstruction"; 1041 let Constraints = "$src = $Rd"; 1042 } 1043 1044 1045 //===------------------------------- 1046 // 2. Extend aliases to 64-bit dest 1047 //===------------------------------- 1048 1049 // Unfortunately the extensions that end up as 64-bits cannot be handled by an 1050 // instruction alias: their syntax is (for example) "SXTB x0, w0", which needs 1051 // to be mapped to "SBFM x0, x0, #0, 7" (changing the class of Rn). InstAlias is 1052 // not capable of such a map as far as I'm aware 1053 1054 // Note that these instructions are strictly more specific than the 1055 // BFM ones (in ImmR) so they can handle their own decoding. 1056 class A64I_bf_ext<bit sf, bits<2> opc, RegisterClass GPRDest, string asmop, 1057 bits<6> imms, dag pattern> 1058 : A64I_bitfield<sf, opc, sf, 1059 (outs GPRDest:$Rd), (ins GPR32:$Rn), 1060 !strconcat(asmop, "\t$Rd, $Rn"), 1061 [(set GPRDest:$Rd, pattern)], NoItinerary> { 1062 let ImmR = 0b000000; 1063 let ImmS = imms; 1064 } 1065 1066 // Signed extensions 1067 def SXTBxw : A64I_bf_ext<0b1, 0b00, GPR64, "sxtb", 7, 1068 (sext_inreg (anyext GPR32:$Rn), i8)>; 1069 def SXTBww : A64I_bf_ext<0b0, 0b00, GPR32, "sxtb", 7, 1070 (sext_inreg GPR32:$Rn, i8)>; 1071 def SXTHxw : A64I_bf_ext<0b1, 0b00, GPR64, "sxth", 15, 1072 (sext_inreg (anyext GPR32:$Rn), i16)>; 1073 def SXTHww : A64I_bf_ext<0b0, 0b00, GPR32, "sxth", 15, 1074 (sext_inreg GPR32:$Rn, i16)>; 1075 def SXTWxw : A64I_bf_ext<0b1, 0b00, GPR64, "sxtw", 31, (sext GPR32:$Rn)>; 1076 1077 // Unsigned extensions 1078 def UXTBww : A64I_bf_ext<0b0, 0b10, GPR32, "uxtb", 7, 1079 (and GPR32:$Rn, 255)>; 1080 def UXTHww : A64I_bf_ext<0b0, 0b10, GPR32, "uxth", 15, 1081 (and GPR32:$Rn, 65535)>; 1082 1083 // The 64-bit unsigned variants are not strictly architectural but recommended 1084 // for consistency. 1085 let isAsmParserOnly = 1 in { 1086 def UXTBxw : A64I_bf_ext<0b0, 0b10, GPR64, "uxtb", 7, 1087 (and (anyext GPR32:$Rn), 255)>; 1088 def UXTHxw : A64I_bf_ext<0b0, 0b10, GPR64, "uxth", 15, 1089 (and (anyext GPR32:$Rn), 65535)>; 1090 } 1091 1092 // Extra patterns for when the source register is actually 64-bits 1093 // too. There's no architectural difference here, it's just LLVM 1094 // shinanigans. There's no need for equivalent zero-extension patterns 1095 // because they'll already be caught by logical (immediate) matching. 1096 def : Pat<(sext_inreg GPR64:$Rn, i8), 1097 (SXTBxw (EXTRACT_SUBREG GPR64:$Rn, sub_32))>; 1098 def : Pat<(sext_inreg GPR64:$Rn, i16), 1099 (SXTHxw (EXTRACT_SUBREG GPR64:$Rn, sub_32))>; 1100 def : Pat<(sext_inreg GPR64:$Rn, i32), 1101 (SXTWxw (EXTRACT_SUBREG GPR64:$Rn, sub_32))>; 1102 1103 1104 //===------------------------------- 1105 // 3. Aliases for ASR and LSR (the simple shifts) 1106 //===------------------------------- 1107 1108 // These also handle their own decoding because ImmS being set makes 1109 // them take precedence over BFM. 1110 multiclass A64I_shift<bits<2> opc, string asmop, SDNode opnode> { 1111 def wwi : A64I_bitfield<0b0, opc, 0b0, 1112 (outs GPR32:$Rd), (ins GPR32:$Rn, bitfield32_imm:$ImmR), 1113 !strconcat(asmop, "\t$Rd, $Rn, $ImmR"), 1114 [(set GPR32:$Rd, (opnode GPR32:$Rn, bitfield32_imm:$ImmR))], 1115 NoItinerary> { 1116 let ImmS = 31; 1117 } 1118 1119 def xxi : A64I_bitfield<0b1, opc, 0b1, 1120 (outs GPR64:$Rd), (ins GPR64:$Rn, bitfield64_imm:$ImmR), 1121 !strconcat(asmop, "\t$Rd, $Rn, $ImmR"), 1122 [(set GPR64:$Rd, (opnode GPR64:$Rn, bitfield64_imm:$ImmR))], 1123 NoItinerary> { 1124 let ImmS = 63; 1125 } 1126 1127 } 1128 1129 defm ASR : A64I_shift<0b00, "asr", sra>; 1130 defm LSR : A64I_shift<0b10, "lsr", srl>; 1131 1132 //===------------------------------- 1133 // 4. Aliases for LSL 1134 //===------------------------------- 1135 1136 // Unfortunately LSL and subsequent aliases are much more complicated. We need 1137 // to be able to say certain output instruction fields depend in a complex 1138 // manner on combinations of input assembly fields). 1139 // 1140 // MIOperandInfo *might* have been able to do it, but at the cost of 1141 // significantly more C++ code. 1142 1143 // N.b. contrary to usual practice these operands store the shift rather than 1144 // the machine bits in an MCInst. The complexity overhead of consistency 1145 // outweighed the benefits in this case (custom asmparser, printer and selection 1146 // vs custom encoder). 1147 def bitfield32_lsl_imm : Operand<i64>, 1148 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 31; }]> { 1149 let ParserMatchClass = uimm5_asmoperand; 1150 let EncoderMethod = "getBitfield32LSLOpValue"; 1151 } 1152 1153 def bitfield64_lsl_imm : Operand<i64>, 1154 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 63; }]> { 1155 let ParserMatchClass = uimm6_asmoperand; 1156 let EncoderMethod = "getBitfield64LSLOpValue"; 1157 } 1158 1159 class A64I_bitfield_lsl<bit sf, RegisterClass GPR, Operand operand> 1160 : A64I_bitfield<sf, 0b10, sf, (outs GPR:$Rd), (ins GPR:$Rn, operand:$FullImm), 1161 "lsl\t$Rd, $Rn, $FullImm", 1162 [(set GPR:$Rd, (shl GPR:$Rn, operand:$FullImm))], 1163 NoItinerary> { 1164 bits<12> FullImm; 1165 let ImmR = FullImm{5-0}; 1166 let ImmS = FullImm{11-6}; 1167 1168 // No disassembler allowed because it would overlap with BFM which does the 1169 // actual work. 1170 let isAsmParserOnly = 1; 1171 } 1172 1173 def LSLwwi : A64I_bitfield_lsl<0b0, GPR32, bitfield32_lsl_imm>; 1174 def LSLxxi : A64I_bitfield_lsl<0b1, GPR64, bitfield64_lsl_imm>; 1175 1176 //===------------------------------- 1177 // 5. Aliases for bitfield extract instructions 1178 //===------------------------------- 1179 1180 def bfx32_width_asmoperand : AsmOperandClass { 1181 let Name = "BFX32Width"; 1182 let PredicateMethod = "isBitfieldWidth<32>"; 1183 let RenderMethod = "addBFXWidthOperands"; 1184 let DiagnosticType = "Width32"; 1185 } 1186 1187 def bfx32_width : Operand<i64>, ImmLeaf<i64, [{ return true; }]> { 1188 let PrintMethod = "printBFXWidthOperand"; 1189 let ParserMatchClass = bfx32_width_asmoperand; 1190 } 1191 1192 def bfx64_width_asmoperand : AsmOperandClass { 1193 let Name = "BFX64Width"; 1194 let PredicateMethod = "isBitfieldWidth<64>"; 1195 let RenderMethod = "addBFXWidthOperands"; 1196 let DiagnosticType = "Width64"; 1197 } 1198 1199 def bfx64_width : Operand<i64> { 1200 let PrintMethod = "printBFXWidthOperand"; 1201 let ParserMatchClass = bfx64_width_asmoperand; 1202 } 1203 1204 1205 multiclass A64I_bitfield_extract<bits<2> opc, string asmop, SDNode op> { 1206 def wwii : A64I_bitfield<0b0, opc, 0b0, (outs GPR32:$Rd), 1207 (ins GPR32:$Rn, bitfield32_imm:$ImmR, bfx32_width:$ImmS), 1208 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"), 1209 [(set GPR32:$Rd, (op GPR32:$Rn, imm:$ImmR, imm:$ImmS))], 1210 NoItinerary> { 1211 // As above, no disassembler allowed. 1212 let isAsmParserOnly = 1; 1213 } 1214 1215 def xxii : A64I_bitfield<0b1, opc, 0b1, (outs GPR64:$Rd), 1216 (ins GPR64:$Rn, bitfield64_imm:$ImmR, bfx64_width:$ImmS), 1217 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"), 1218 [(set GPR64:$Rd, (op GPR64:$Rn, imm:$ImmR, imm:$ImmS))], 1219 NoItinerary> { 1220 // As above, no disassembler allowed. 1221 let isAsmParserOnly = 1; 1222 } 1223 } 1224 1225 defm SBFX : A64I_bitfield_extract<0b00, "sbfx", A64Sbfx>; 1226 defm UBFX : A64I_bitfield_extract<0b10, "ubfx", A64Ubfx>; 1227 1228 // Again, variants based on BFM modify Rd so need it as an input too. 1229 def BFXILwwii : A64I_bitfield<0b0, 0b01, 0b0, (outs GPR32:$Rd), 1230 (ins GPR32:$src, GPR32:$Rn, bitfield32_imm:$ImmR, bfx32_width:$ImmS), 1231 "bfxil\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> { 1232 // As above, no disassembler allowed. 1233 let isAsmParserOnly = 1; 1234 let Constraints = "$src = $Rd"; 1235 } 1236 1237 def BFXILxxii : A64I_bitfield<0b1, 0b01, 0b1, (outs GPR64:$Rd), 1238 (ins GPR64:$src, GPR64:$Rn, bitfield64_imm:$ImmR, bfx64_width:$ImmS), 1239 "bfxil\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> { 1240 // As above, no disassembler allowed. 1241 let isAsmParserOnly = 1; 1242 let Constraints = "$src = $Rd"; 1243 } 1244 1245 // SBFX instructions can do a 1-instruction sign-extension of boolean values. 1246 def : Pat<(sext_inreg GPR64:$Rn, i1), (SBFXxxii GPR64:$Rn, 0, 0)>; 1247 def : Pat<(sext_inreg GPR32:$Rn, i1), (SBFXwwii GPR32:$Rn, 0, 0)>; 1248 def : Pat<(i64 (sext_inreg (anyext GPR32:$Rn), i1)), 1249 (SBFXxxii (SUBREG_TO_REG (i64 0), GPR32:$Rn, sub_32), 0, 0)>; 1250 1251 // UBFX makes sense as an implementation of a 64-bit zero-extension too. Could 1252 // use either 64-bit or 32-bit variant, but 32-bit might be more efficient. 1253 def : Pat<(zext GPR32:$Rn), (SUBREG_TO_REG (i64 0), (UBFXwwii GPR32:$Rn, 0, 31), 1254 sub_32)>; 1255 1256 //===------------------------------- 1257 // 6. Aliases for bitfield insert instructions 1258 //===------------------------------- 1259 1260 def bfi32_lsb_asmoperand : AsmOperandClass { 1261 let Name = "BFI32LSB"; 1262 let PredicateMethod = "isUImm<5>"; 1263 let RenderMethod = "addBFILSBOperands<32>"; 1264 let DiagnosticType = "UImm5"; 1265 } 1266 1267 def bfi32_lsb : Operand<i64>, 1268 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 31; }]> { 1269 let PrintMethod = "printBFILSBOperand<32>"; 1270 let ParserMatchClass = bfi32_lsb_asmoperand; 1271 } 1272 1273 def bfi64_lsb_asmoperand : AsmOperandClass { 1274 let Name = "BFI64LSB"; 1275 let PredicateMethod = "isUImm<6>"; 1276 let RenderMethod = "addBFILSBOperands<64>"; 1277 let DiagnosticType = "UImm6"; 1278 } 1279 1280 def bfi64_lsb : Operand<i64>, 1281 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 63; }]> { 1282 let PrintMethod = "printBFILSBOperand<64>"; 1283 let ParserMatchClass = bfi64_lsb_asmoperand; 1284 } 1285 1286 // Width verification is performed during conversion so width operand can be 1287 // shared between 32/64-bit cases. Still needed for the print method though 1288 // because ImmR encodes "width - 1". 1289 def bfi32_width_asmoperand : AsmOperandClass { 1290 let Name = "BFI32Width"; 1291 let PredicateMethod = "isBitfieldWidth<32>"; 1292 let RenderMethod = "addBFIWidthOperands"; 1293 let DiagnosticType = "Width32"; 1294 } 1295 1296 def bfi32_width : Operand<i64>, 1297 ImmLeaf<i64, [{ return Imm >= 1 && Imm <= 32; }]> { 1298 let PrintMethod = "printBFIWidthOperand"; 1299 let ParserMatchClass = bfi32_width_asmoperand; 1300 } 1301 1302 def bfi64_width_asmoperand : AsmOperandClass { 1303 let Name = "BFI64Width"; 1304 let PredicateMethod = "isBitfieldWidth<64>"; 1305 let RenderMethod = "addBFIWidthOperands"; 1306 let DiagnosticType = "Width64"; 1307 } 1308 1309 def bfi64_width : Operand<i64>, 1310 ImmLeaf<i64, [{ return Imm >= 1 && Imm <= 64; }]> { 1311 let PrintMethod = "printBFIWidthOperand"; 1312 let ParserMatchClass = bfi64_width_asmoperand; 1313 } 1314 1315 multiclass A64I_bitfield_insert<bits<2> opc, string asmop> { 1316 def wwii : A64I_bitfield<0b0, opc, 0b0, (outs GPR32:$Rd), 1317 (ins GPR32:$Rn, bfi32_lsb:$ImmR, bfi32_width:$ImmS), 1318 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"), 1319 [], NoItinerary> { 1320 // As above, no disassembler allowed. 1321 let isAsmParserOnly = 1; 1322 } 1323 1324 def xxii : A64I_bitfield<0b1, opc, 0b1, (outs GPR64:$Rd), 1325 (ins GPR64:$Rn, bfi64_lsb:$ImmR, bfi64_width:$ImmS), 1326 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"), 1327 [], NoItinerary> { 1328 // As above, no disassembler allowed. 1329 let isAsmParserOnly = 1; 1330 } 1331 } 1332 1333 defm SBFIZ : A64I_bitfield_insert<0b00, "sbfiz">; 1334 defm UBFIZ : A64I_bitfield_insert<0b10, "ubfiz">; 1335 1336 1337 def BFIwwii : A64I_bitfield<0b0, 0b01, 0b0, (outs GPR32:$Rd), 1338 (ins GPR32:$src, GPR32:$Rn, bfi32_lsb:$ImmR, bfi32_width:$ImmS), 1339 "bfi\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> { 1340 // As above, no disassembler allowed. 1341 let isAsmParserOnly = 1; 1342 let Constraints = "$src = $Rd"; 1343 } 1344 1345 def BFIxxii : A64I_bitfield<0b1, 0b01, 0b1, (outs GPR64:$Rd), 1346 (ins GPR64:$src, GPR64:$Rn, bfi64_lsb:$ImmR, bfi64_width:$ImmS), 1347 "bfi\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> { 1348 // As above, no disassembler allowed. 1349 let isAsmParserOnly = 1; 1350 let Constraints = "$src = $Rd"; 1351 } 1352 1353 //===----------------------------------------------------------------------===// 1354 // Compare and branch (immediate) 1355 //===----------------------------------------------------------------------===// 1356 // Contains: CBZ, CBNZ 1357 1358 class label_asmoperand<int width, int scale> : AsmOperandClass { 1359 let Name = "Label" # width # "_" # scale; 1360 let PredicateMethod = "isLabel<" # width # "," # scale # ">"; 1361 let RenderMethod = "addLabelOperands<" # width # ", " # scale # ">"; 1362 let DiagnosticType = "Label"; 1363 } 1364 1365 def label_wid19_scal4_asmoperand : label_asmoperand<19, 4>; 1366 1367 // All conditional immediate branches are the same really: 19 signed bits scaled 1368 // by the instruction-size (4). 1369 def bcc_target : Operand<OtherVT> { 1370 // This label is a 19-bit offset from PC, scaled by the instruction-width: 4. 1371 let ParserMatchClass = label_wid19_scal4_asmoperand; 1372 let PrintMethod = "printLabelOperand<19, 4>"; 1373 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_condbr>"; 1374 let OperandType = "OPERAND_PCREL"; 1375 } 1376 1377 multiclass cmpbr_sizes<bit op, string asmop, ImmLeaf SETOP> { 1378 let isBranch = 1, isTerminator = 1 in { 1379 def x : A64I_cmpbr<0b1, op, 1380 (outs), 1381 (ins GPR64:$Rt, bcc_target:$Label), 1382 !strconcat(asmop,"\t$Rt, $Label"), 1383 [(A64br_cc (A64cmp GPR64:$Rt, 0), SETOP, bb:$Label)], 1384 NoItinerary>; 1385 1386 def w : A64I_cmpbr<0b0, op, 1387 (outs), 1388 (ins GPR32:$Rt, bcc_target:$Label), 1389 !strconcat(asmop,"\t$Rt, $Label"), 1390 [(A64br_cc (A64cmp GPR32:$Rt, 0), SETOP, bb:$Label)], 1391 NoItinerary>; 1392 } 1393 } 1394 1395 defm CBZ : cmpbr_sizes<0b0, "cbz", ImmLeaf<i32, [{ 1396 return Imm == A64CC::EQ; 1397 }]> >; 1398 defm CBNZ : cmpbr_sizes<0b1, "cbnz", ImmLeaf<i32, [{ 1399 return Imm == A64CC::NE; 1400 }]> >; 1401 1402 //===----------------------------------------------------------------------===// 1403 // Conditional branch (immediate) instructions 1404 //===----------------------------------------------------------------------===// 1405 // Contains: B.cc 1406 1407 def cond_code_asmoperand : AsmOperandClass { 1408 let Name = "CondCode"; 1409 let DiagnosticType = "CondCode"; 1410 } 1411 1412 def cond_code : Operand<i32>, ImmLeaf<i32, [{ 1413 return Imm >= 0 && Imm <= 15; 1414 }]> { 1415 let PrintMethod = "printCondCodeOperand"; 1416 let ParserMatchClass = cond_code_asmoperand; 1417 } 1418 1419 def Bcc : A64I_condbr<0b0, 0b0, (outs), 1420 (ins cond_code:$Cond, bcc_target:$Label), 1421 "b.$Cond $Label", [(A64br_cc NZCV, (i32 imm:$Cond), bb:$Label)], 1422 NoItinerary> { 1423 let Uses = [NZCV]; 1424 let isBranch = 1; 1425 let isTerminator = 1; 1426 } 1427 1428 //===----------------------------------------------------------------------===// 1429 // Conditional compare (immediate) instructions 1430 //===----------------------------------------------------------------------===// 1431 // Contains: CCMN, CCMP 1432 1433 def uimm4_asmoperand : AsmOperandClass { 1434 let Name = "UImm4"; 1435 let PredicateMethod = "isUImm<4>"; 1436 let RenderMethod = "addImmOperands"; 1437 let DiagnosticType = "UImm4"; 1438 } 1439 1440 def uimm4 : Operand<i32> { 1441 let ParserMatchClass = uimm4_asmoperand; 1442 } 1443 1444 def uimm5 : Operand<i32> { 1445 let ParserMatchClass = uimm5_asmoperand; 1446 } 1447 1448 // The only difference between this operand and the one for instructions like 1449 // B.cc is that it's parsed manually. The other get parsed implicitly as part of 1450 // the mnemonic handling. 1451 def cond_code_op_asmoperand : AsmOperandClass { 1452 let Name = "CondCodeOp"; 1453 let RenderMethod = "addCondCodeOperands"; 1454 let PredicateMethod = "isCondCode"; 1455 let ParserMethod = "ParseCondCodeOperand"; 1456 let DiagnosticType = "CondCode"; 1457 } 1458 1459 def cond_code_op : Operand<i32> { 1460 let PrintMethod = "printCondCodeOperand"; 1461 let ParserMatchClass = cond_code_op_asmoperand; 1462 } 1463 1464 class A64I_condcmpimmImpl<bit sf, bit op, RegisterClass GPR, string asmop> 1465 : A64I_condcmpimm<sf, op, 0b0, 0b0, 0b1, (outs), 1466 (ins GPR:$Rn, uimm5:$UImm5, uimm4:$NZCVImm, cond_code_op:$Cond), 1467 !strconcat(asmop, "\t$Rn, $UImm5, $NZCVImm, $Cond"), 1468 [], NoItinerary> { 1469 let Defs = [NZCV]; 1470 } 1471 1472 def CCMNwi : A64I_condcmpimmImpl<0b0, 0b0, GPR32, "ccmn">; 1473 def CCMNxi : A64I_condcmpimmImpl<0b1, 0b0, GPR64, "ccmn">; 1474 def CCMPwi : A64I_condcmpimmImpl<0b0, 0b1, GPR32, "ccmp">; 1475 def CCMPxi : A64I_condcmpimmImpl<0b1, 0b1, GPR64, "ccmp">; 1476 1477 //===----------------------------------------------------------------------===// 1478 // Conditional compare (register) instructions 1479 //===----------------------------------------------------------------------===// 1480 // Contains: CCMN, CCMP 1481 1482 class A64I_condcmpregImpl<bit sf, bit op, RegisterClass GPR, string asmop> 1483 : A64I_condcmpreg<sf, op, 0b0, 0b0, 0b1, 1484 (outs), 1485 (ins GPR:$Rn, GPR:$Rm, uimm4:$NZCVImm, cond_code_op:$Cond), 1486 !strconcat(asmop, "\t$Rn, $Rm, $NZCVImm, $Cond"), 1487 [], NoItinerary> { 1488 let Defs = [NZCV]; 1489 } 1490 1491 def CCMNww : A64I_condcmpregImpl<0b0, 0b0, GPR32, "ccmn">; 1492 def CCMNxx : A64I_condcmpregImpl<0b1, 0b0, GPR64, "ccmn">; 1493 def CCMPww : A64I_condcmpregImpl<0b0, 0b1, GPR32, "ccmp">; 1494 def CCMPxx : A64I_condcmpregImpl<0b1, 0b1, GPR64, "ccmp">; 1495 1496 //===----------------------------------------------------------------------===// 1497 // Conditional select instructions 1498 //===----------------------------------------------------------------------===// 1499 // Contains: CSEL, CSINC, CSINV, CSNEG + aliases CSET, CSETM, CINC, CINV, CNEG 1500 1501 // Condition code which is encoded as the inversion (semantically rather than 1502 // bitwise) in the instruction. 1503 def inv_cond_code_op_asmoperand : AsmOperandClass { 1504 let Name = "InvCondCodeOp"; 1505 let RenderMethod = "addInvCondCodeOperands"; 1506 let PredicateMethod = "isCondCode"; 1507 let ParserMethod = "ParseCondCodeOperand"; 1508 let DiagnosticType = "CondCode"; 1509 } 1510 1511 def inv_cond_code_op : Operand<i32> { 1512 let ParserMatchClass = inv_cond_code_op_asmoperand; 1513 } 1514 1515 // Having a separate operand for the selectable use-case is debatable, but gives 1516 // consistency with cond_code. 1517 def inv_cond_XFORM : SDNodeXForm<imm, [{ 1518 A64CC::CondCodes CC = static_cast<A64CC::CondCodes>(N->getZExtValue()); 1519 return CurDAG->getTargetConstant(A64InvertCondCode(CC), MVT::i32); 1520 }]>; 1521 1522 def inv_cond_code 1523 : ImmLeaf<i32, [{ return Imm >= 0 && Imm <= 15; }], inv_cond_XFORM>; 1524 1525 1526 multiclass A64I_condselSizes<bit op, bits<2> op2, string asmop, 1527 SDPatternOperator select> { 1528 let Uses = [NZCV] in { 1529 def wwwc : A64I_condsel<0b0, op, 0b0, op2, 1530 (outs GPR32:$Rd), 1531 (ins GPR32:$Rn, GPR32:$Rm, cond_code_op:$Cond), 1532 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Cond"), 1533 [(set GPR32:$Rd, (select GPR32:$Rn, GPR32:$Rm))], 1534 NoItinerary>; 1535 1536 1537 def xxxc : A64I_condsel<0b1, op, 0b0, op2, 1538 (outs GPR64:$Rd), 1539 (ins GPR64:$Rn, GPR64:$Rm, cond_code_op:$Cond), 1540 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Cond"), 1541 [(set GPR64:$Rd, (select GPR64:$Rn, GPR64:$Rm))], 1542 NoItinerary>; 1543 } 1544 } 1545 1546 def simple_select 1547 : PatFrag<(ops node:$lhs, node:$rhs), 1548 (A64select_cc NZCV, node:$lhs, node:$rhs, (i32 imm:$Cond))>; 1549 1550 class complex_select<SDPatternOperator opnode> 1551 : PatFrag<(ops node:$lhs, node:$rhs), 1552 (A64select_cc NZCV, node:$lhs, (opnode node:$rhs), (i32 imm:$Cond))>; 1553 1554 1555 defm CSEL : A64I_condselSizes<0b0, 0b00, "csel", simple_select>; 1556 defm CSINC : A64I_condselSizes<0b0, 0b01, "csinc", 1557 complex_select<PatFrag<(ops node:$val), 1558 (add node:$val, 1)>>>; 1559 defm CSINV : A64I_condselSizes<0b1, 0b00, "csinv", complex_select<not>>; 1560 defm CSNEG : A64I_condselSizes<0b1, 0b01, "csneg", complex_select<ineg>>; 1561 1562 // Now the instruction aliases, which fit nicely into LLVM's model: 1563 1564 def : InstAlias<"cset $Rd, $Cond", 1565 (CSINCwwwc GPR32:$Rd, WZR, WZR, inv_cond_code_op:$Cond)>; 1566 def : InstAlias<"cset $Rd, $Cond", 1567 (CSINCxxxc GPR64:$Rd, XZR, XZR, inv_cond_code_op:$Cond)>; 1568 def : InstAlias<"csetm $Rd, $Cond", 1569 (CSINVwwwc GPR32:$Rd, WZR, WZR, inv_cond_code_op:$Cond)>; 1570 def : InstAlias<"csetm $Rd, $Cond", 1571 (CSINVxxxc GPR64:$Rd, XZR, XZR, inv_cond_code_op:$Cond)>; 1572 def : InstAlias<"cinc $Rd, $Rn, $Cond", 1573 (CSINCwwwc GPR32:$Rd, GPR32:$Rn, GPR32:$Rn, inv_cond_code_op:$Cond)>; 1574 def : InstAlias<"cinc $Rd, $Rn, $Cond", 1575 (CSINCxxxc GPR64:$Rd, GPR64:$Rn, GPR64:$Rn, inv_cond_code_op:$Cond)>; 1576 def : InstAlias<"cinv $Rd, $Rn, $Cond", 1577 (CSINVwwwc GPR32:$Rd, GPR32:$Rn, GPR32:$Rn, inv_cond_code_op:$Cond)>; 1578 def : InstAlias<"cinv $Rd, $Rn, $Cond", 1579 (CSINVxxxc GPR64:$Rd, GPR64:$Rn, GPR64:$Rn, inv_cond_code_op:$Cond)>; 1580 def : InstAlias<"cneg $Rd, $Rn, $Cond", 1581 (CSNEGwwwc GPR32:$Rd, GPR32:$Rn, GPR32:$Rn, inv_cond_code_op:$Cond)>; 1582 def : InstAlias<"cneg $Rd, $Rn, $Cond", 1583 (CSNEGxxxc GPR64:$Rd, GPR64:$Rn, GPR64:$Rn, inv_cond_code_op:$Cond)>; 1584 1585 // Finally some helper patterns. 1586 1587 // For CSET (a.k.a. zero-extension of icmp) 1588 def : Pat<(A64select_cc NZCV, 0, 1, cond_code:$Cond), 1589 (CSINCwwwc WZR, WZR, cond_code:$Cond)>; 1590 def : Pat<(A64select_cc NZCV, 1, 0, inv_cond_code:$Cond), 1591 (CSINCwwwc WZR, WZR, inv_cond_code:$Cond)>; 1592 1593 def : Pat<(A64select_cc NZCV, 0, 1, cond_code:$Cond), 1594 (CSINCxxxc XZR, XZR, cond_code:$Cond)>; 1595 def : Pat<(A64select_cc NZCV, 1, 0, inv_cond_code:$Cond), 1596 (CSINCxxxc XZR, XZR, inv_cond_code:$Cond)>; 1597 1598 // For CSETM (a.k.a. sign-extension of icmp) 1599 def : Pat<(A64select_cc NZCV, 0, -1, cond_code:$Cond), 1600 (CSINVwwwc WZR, WZR, cond_code:$Cond)>; 1601 def : Pat<(A64select_cc NZCV, -1, 0, inv_cond_code:$Cond), 1602 (CSINVwwwc WZR, WZR, inv_cond_code:$Cond)>; 1603 1604 def : Pat<(A64select_cc NZCV, 0, -1, cond_code:$Cond), 1605 (CSINVxxxc XZR, XZR, cond_code:$Cond)>; 1606 def : Pat<(A64select_cc NZCV, -1, 0, inv_cond_code:$Cond), 1607 (CSINVxxxc XZR, XZR, inv_cond_code:$Cond)>; 1608 1609 // CINC, CINV and CNEG get dealt with automatically, which leaves the issue of 1610 // commutativity. The instructions are to complex for isCommutable to be used, 1611 // so we have to create the patterns manually: 1612 1613 // No commutable pattern for CSEL since the commuted version is isomorphic. 1614 1615 // CSINC 1616 def :Pat<(A64select_cc NZCV, (add GPR32:$Rm, 1), GPR32:$Rn, 1617 inv_cond_code:$Cond), 1618 (CSINCwwwc GPR32:$Rn, GPR32:$Rm, inv_cond_code:$Cond)>; 1619 def :Pat<(A64select_cc NZCV, (add GPR64:$Rm, 1), GPR64:$Rn, 1620 inv_cond_code:$Cond), 1621 (CSINCxxxc GPR64:$Rn, GPR64:$Rm, inv_cond_code:$Cond)>; 1622 1623 // CSINV 1624 def :Pat<(A64select_cc NZCV, (not GPR32:$Rm), GPR32:$Rn, inv_cond_code:$Cond), 1625 (CSINVwwwc GPR32:$Rn, GPR32:$Rm, inv_cond_code:$Cond)>; 1626 def :Pat<(A64select_cc NZCV, (not GPR64:$Rm), GPR64:$Rn, inv_cond_code:$Cond), 1627 (CSINVxxxc GPR64:$Rn, GPR64:$Rm, inv_cond_code:$Cond)>; 1628 1629 // CSNEG 1630 def :Pat<(A64select_cc NZCV, (ineg GPR32:$Rm), GPR32:$Rn, inv_cond_code:$Cond), 1631 (CSNEGwwwc GPR32:$Rn, GPR32:$Rm, inv_cond_code:$Cond)>; 1632 def :Pat<(A64select_cc NZCV, (ineg GPR64:$Rm), GPR64:$Rn, inv_cond_code:$Cond), 1633 (CSNEGxxxc GPR64:$Rn, GPR64:$Rm, inv_cond_code:$Cond)>; 1634 1635 //===----------------------------------------------------------------------===// 1636 // Data Processing (1 source) instructions 1637 //===----------------------------------------------------------------------===// 1638 // Contains: RBIT, REV16, REV, REV32, CLZ, CLS. 1639 1640 // We define an unary operator which always fails. We will use this to 1641 // define unary operators that cannot be matched. 1642 1643 class A64I_dp_1src_impl<bit sf, bits<6> opcode, string asmop, 1644 list<dag> patterns, RegisterClass GPRrc, 1645 InstrItinClass itin>: 1646 A64I_dp_1src<sf, 1647 0, 1648 0b00000, 1649 opcode, 1650 !strconcat(asmop, "\t$Rd, $Rn"), 1651 (outs GPRrc:$Rd), 1652 (ins GPRrc:$Rn), 1653 patterns, 1654 itin>; 1655 1656 multiclass A64I_dp_1src <bits<6> opcode, string asmop> { 1657 let hasSideEffects = 0 in { 1658 def ww : A64I_dp_1src_impl<0b0, opcode, asmop, [], GPR32, NoItinerary>; 1659 def xx : A64I_dp_1src_impl<0b1, opcode, asmop, [], GPR64, NoItinerary>; 1660 } 1661 } 1662 1663 defm RBIT : A64I_dp_1src<0b000000, "rbit">; 1664 defm CLS : A64I_dp_1src<0b000101, "cls">; 1665 defm CLZ : A64I_dp_1src<0b000100, "clz">; 1666 1667 def : Pat<(ctlz GPR32:$Rn), (CLZww GPR32:$Rn)>; 1668 def : Pat<(ctlz GPR64:$Rn), (CLZxx GPR64:$Rn)>; 1669 def : Pat<(ctlz_zero_undef GPR32:$Rn), (CLZww GPR32:$Rn)>; 1670 def : Pat<(ctlz_zero_undef GPR64:$Rn), (CLZxx GPR64:$Rn)>; 1671 1672 def : Pat<(cttz GPR32:$Rn), (CLZww (RBITww GPR32:$Rn))>; 1673 def : Pat<(cttz GPR64:$Rn), (CLZxx (RBITxx GPR64:$Rn))>; 1674 def : Pat<(cttz_zero_undef GPR32:$Rn), (CLZww (RBITww GPR32:$Rn))>; 1675 def : Pat<(cttz_zero_undef GPR64:$Rn), (CLZxx (RBITxx GPR64:$Rn))>; 1676 1677 1678 def REVww : A64I_dp_1src_impl<0b0, 0b000010, "rev", 1679 [(set GPR32:$Rd, (bswap GPR32:$Rn))], 1680 GPR32, NoItinerary>; 1681 def REVxx : A64I_dp_1src_impl<0b1, 0b000011, "rev", 1682 [(set GPR64:$Rd, (bswap GPR64:$Rn))], 1683 GPR64, NoItinerary>; 1684 def REV32xx : A64I_dp_1src_impl<0b1, 0b000010, "rev32", 1685 [(set GPR64:$Rd, (bswap (rotr GPR64:$Rn, (i64 32))))], 1686 GPR64, NoItinerary>; 1687 def REV16ww : A64I_dp_1src_impl<0b0, 0b000001, "rev16", 1688 [(set GPR32:$Rd, (bswap (rotr GPR32:$Rn, (i64 16))))], 1689 GPR32, 1690 NoItinerary>; 1691 def REV16xx : A64I_dp_1src_impl<0b1, 0b000001, "rev16", [], GPR64, NoItinerary>; 1692 1693 //===----------------------------------------------------------------------===// 1694 // Data Processing (2 sources) instructions 1695 //===----------------------------------------------------------------------===// 1696 // Contains: CRC32C?[BHWX], UDIV, SDIV, LSLV, LSRV, ASRV, RORV + aliases LSL, 1697 // LSR, ASR, ROR 1698 1699 1700 class dp_2src_impl<bit sf, bits<6> opcode, string asmop, list<dag> patterns, 1701 RegisterClass GPRsp, 1702 InstrItinClass itin>: 1703 A64I_dp_2src<sf, 1704 opcode, 1705 0, 1706 !strconcat(asmop, "\t$Rd, $Rn, $Rm"), 1707 (outs GPRsp:$Rd), 1708 (ins GPRsp:$Rn, GPRsp:$Rm), 1709 patterns, 1710 itin>; 1711 1712 multiclass dp_2src_crc<bit c, string asmop> { 1713 def B_www : dp_2src_impl<0b0, {0, 1, 0, c, 0, 0}, 1714 !strconcat(asmop, "b"), [], GPR32, NoItinerary>; 1715 def H_www : dp_2src_impl<0b0, {0, 1, 0, c, 0, 1}, 1716 !strconcat(asmop, "h"), [], GPR32, NoItinerary>; 1717 def W_www : dp_2src_impl<0b0, {0, 1, 0, c, 1, 0}, 1718 !strconcat(asmop, "w"), [], GPR32, NoItinerary>; 1719 def X_wwx : A64I_dp_2src<0b1, {0, 1, 0, c, 1, 1}, 0b0, 1720 !strconcat(asmop, "x\t$Rd, $Rn, $Rm"), 1721 (outs GPR32:$Rd), (ins GPR32:$Rn, GPR64:$Rm), [], 1722 NoItinerary>; 1723 } 1724 1725 multiclass dp_2src_zext <bits<6> opcode, string asmop, SDPatternOperator op> { 1726 def www : dp_2src_impl<0b0, 1727 opcode, 1728 asmop, 1729 [(set GPR32:$Rd, 1730 (op GPR32:$Rn, (i64 (zext GPR32:$Rm))))], 1731 GPR32, 1732 NoItinerary>; 1733 def xxx : dp_2src_impl<0b1, 1734 opcode, 1735 asmop, 1736 [(set GPR64:$Rd, (op GPR64:$Rn, GPR64:$Rm))], 1737 GPR64, 1738 NoItinerary>; 1739 } 1740 1741 1742 multiclass dp_2src <bits<6> opcode, string asmop, SDPatternOperator op> { 1743 def www : dp_2src_impl<0b0, 1744 opcode, 1745 asmop, 1746 [(set GPR32:$Rd, (op GPR32:$Rn, GPR32:$Rm))], 1747 GPR32, 1748 NoItinerary>; 1749 def xxx : dp_2src_impl<0b1, 1750 opcode, 1751 asmop, 1752 [(set GPR64:$Rd, (op GPR64:$Rn, GPR64:$Rm))], 1753 GPR64, 1754 NoItinerary>; 1755 } 1756 1757 // Here we define the data processing 2 source instructions. 1758 defm CRC32 : dp_2src_crc<0b0, "crc32">; 1759 defm CRC32C : dp_2src_crc<0b1, "crc32c">; 1760 1761 defm UDIV : dp_2src<0b000010, "udiv", udiv>; 1762 defm SDIV : dp_2src<0b000011, "sdiv", sdiv>; 1763 1764 defm LSLV : dp_2src_zext<0b001000, "lsl", shl>; 1765 defm LSRV : dp_2src_zext<0b001001, "lsr", srl>; 1766 defm ASRV : dp_2src_zext<0b001010, "asr", sra>; 1767 defm RORV : dp_2src_zext<0b001011, "ror", rotr>; 1768 1769 // Extra patterns for an incoming 64-bit value for a 32-bit 1770 // operation. Since the LLVM operations are undefined (as in C) if the 1771 // RHS is out of range, it's perfectly permissible to discard the high 1772 // bits of the GPR64. 1773 def : Pat<(shl GPR32:$Rn, GPR64:$Rm), 1774 (LSLVwww GPR32:$Rn, (EXTRACT_SUBREG GPR64:$Rm, sub_32))>; 1775 def : Pat<(srl GPR32:$Rn, GPR64:$Rm), 1776 (LSRVwww GPR32:$Rn, (EXTRACT_SUBREG GPR64:$Rm, sub_32))>; 1777 def : Pat<(sra GPR32:$Rn, GPR64:$Rm), 1778 (ASRVwww GPR32:$Rn, (EXTRACT_SUBREG GPR64:$Rm, sub_32))>; 1779 def : Pat<(rotr GPR32:$Rn, GPR64:$Rm), 1780 (RORVwww GPR32:$Rn, (EXTRACT_SUBREG GPR64:$Rm, sub_32))>; 1781 1782 // Here we define the aliases for the data processing 2 source instructions. 1783 def LSL_mnemonic : MnemonicAlias<"lslv", "lsl">; 1784 def LSR_mnemonic : MnemonicAlias<"lsrv", "lsr">; 1785 def ASR_menmonic : MnemonicAlias<"asrv", "asr">; 1786 def ROR_menmonic : MnemonicAlias<"rorv", "ror">; 1787 1788 //===----------------------------------------------------------------------===// 1789 // Data Processing (3 sources) instructions 1790 //===----------------------------------------------------------------------===// 1791 // Contains: MADD, MSUB, SMADDL, SMSUBL, SMULH, UMADDL, UMSUBL, UMULH 1792 // + aliases MUL, MNEG, SMULL, SMNEGL, UMULL, UMNEGL 1793 1794 class A64I_dp3_4operand<bit sf, bits<6> opcode, RegisterClass AccReg, 1795 RegisterClass SrcReg, string asmop, dag pattern> 1796 : A64I_dp3<sf, opcode, 1797 (outs AccReg:$Rd), (ins SrcReg:$Rn, SrcReg:$Rm, AccReg:$Ra), 1798 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Ra"), 1799 [(set AccReg:$Rd, pattern)], NoItinerary> { 1800 RegisterClass AccGPR = AccReg; 1801 RegisterClass SrcGPR = SrcReg; 1802 } 1803 1804 def MADDwwww : A64I_dp3_4operand<0b0, 0b000000, GPR32, GPR32, "madd", 1805 (add GPR32:$Ra, (mul GPR32:$Rn, GPR32:$Rm))>; 1806 def MADDxxxx : A64I_dp3_4operand<0b1, 0b000000, GPR64, GPR64, "madd", 1807 (add GPR64:$Ra, (mul GPR64:$Rn, GPR64:$Rm))>; 1808 1809 def MSUBwwww : A64I_dp3_4operand<0b0, 0b000001, GPR32, GPR32, "msub", 1810 (sub GPR32:$Ra, (mul GPR32:$Rn, GPR32:$Rm))>; 1811 def MSUBxxxx : A64I_dp3_4operand<0b1, 0b000001, GPR64, GPR64, "msub", 1812 (sub GPR64:$Ra, (mul GPR64:$Rn, GPR64:$Rm))>; 1813 1814 def SMADDLxwwx : A64I_dp3_4operand<0b1, 0b000010, GPR64, GPR32, "smaddl", 1815 (add GPR64:$Ra, (mul (i64 (sext GPR32:$Rn)), (sext GPR32:$Rm)))>; 1816 def SMSUBLxwwx : A64I_dp3_4operand<0b1, 0b000011, GPR64, GPR32, "smsubl", 1817 (sub GPR64:$Ra, (mul (i64 (sext GPR32:$Rn)), (sext GPR32:$Rm)))>; 1818 1819 def UMADDLxwwx : A64I_dp3_4operand<0b1, 0b001010, GPR64, GPR32, "umaddl", 1820 (add GPR64:$Ra, (mul (i64 (zext GPR32:$Rn)), (zext GPR32:$Rm)))>; 1821 def UMSUBLxwwx : A64I_dp3_4operand<0b1, 0b001011, GPR64, GPR32, "umsubl", 1822 (sub GPR64:$Ra, (mul (i64 (zext GPR32:$Rn)), (zext GPR32:$Rm)))>; 1823 1824 let isCommutable = 1, PostEncoderMethod = "fixMulHigh" in { 1825 def UMULHxxx : A64I_dp3<0b1, 0b001100, (outs GPR64:$Rd), 1826 (ins GPR64:$Rn, GPR64:$Rm), 1827 "umulh\t$Rd, $Rn, $Rm", 1828 [(set GPR64:$Rd, (mulhu GPR64:$Rn, GPR64:$Rm))], 1829 NoItinerary>; 1830 1831 def SMULHxxx : A64I_dp3<0b1, 0b000100, (outs GPR64:$Rd), 1832 (ins GPR64:$Rn, GPR64:$Rm), 1833 "smulh\t$Rd, $Rn, $Rm", 1834 [(set GPR64:$Rd, (mulhs GPR64:$Rn, GPR64:$Rm))], 1835 NoItinerary>; 1836 } 1837 1838 multiclass A64I_dp3_3operand<string asmop, A64I_dp3_4operand INST, 1839 Register ZR, dag pattern> { 1840 def : InstAlias<asmop # " $Rd, $Rn, $Rm", 1841 (INST INST.AccGPR:$Rd, INST.SrcGPR:$Rn, INST.SrcGPR:$Rm, ZR)>; 1842 1843 def : Pat<pattern, (INST INST.SrcGPR:$Rn, INST.SrcGPR:$Rm, ZR)>; 1844 } 1845 1846 defm : A64I_dp3_3operand<"mul", MADDwwww, WZR, (mul GPR32:$Rn, GPR32:$Rm)>; 1847 defm : A64I_dp3_3operand<"mul", MADDxxxx, XZR, (mul GPR64:$Rn, GPR64:$Rm)>; 1848 1849 defm : A64I_dp3_3operand<"mneg", MSUBwwww, WZR, 1850 (sub 0, (mul GPR32:$Rn, GPR32:$Rm))>; 1851 defm : A64I_dp3_3operand<"mneg", MSUBxxxx, XZR, 1852 (sub 0, (mul GPR64:$Rn, GPR64:$Rm))>; 1853 1854 defm : A64I_dp3_3operand<"smull", SMADDLxwwx, XZR, 1855 (mul (i64 (sext GPR32:$Rn)), (sext GPR32:$Rm))>; 1856 defm : A64I_dp3_3operand<"smnegl", SMSUBLxwwx, XZR, 1857 (sub 0, (mul (i64 (sext GPR32:$Rn)), (sext GPR32:$Rm)))>; 1858 1859 defm : A64I_dp3_3operand<"umull", UMADDLxwwx, XZR, 1860 (mul (i64 (zext GPR32:$Rn)), (zext GPR32:$Rm))>; 1861 defm : A64I_dp3_3operand<"umnegl", UMSUBLxwwx, XZR, 1862 (sub 0, (mul (i64 (zext GPR32:$Rn)), (zext GPR32:$Rm)))>; 1863 1864 1865 //===----------------------------------------------------------------------===// 1866 // Exception generation 1867 //===----------------------------------------------------------------------===// 1868 // Contains: SVC, HVC, SMC, BRK, HLT, DCPS1, DCPS2, DCPS3 1869 1870 def uimm16_asmoperand : AsmOperandClass { 1871 let Name = "UImm16"; 1872 let PredicateMethod = "isUImm<16>"; 1873 let RenderMethod = "addImmOperands"; 1874 let DiagnosticType = "UImm16"; 1875 } 1876 1877 def uimm16 : Operand<i32> { 1878 let ParserMatchClass = uimm16_asmoperand; 1879 } 1880 1881 class A64I_exceptImpl<bits<3> opc, bits<2> ll, string asmop> 1882 : A64I_exception<opc, 0b000, ll, (outs), (ins uimm16:$UImm16), 1883 !strconcat(asmop, "\t$UImm16"), [], NoItinerary> { 1884 let isBranch = 1; 1885 let isTerminator = 1; 1886 } 1887 1888 def SVCi : A64I_exceptImpl<0b000, 0b01, "svc">; 1889 def HVCi : A64I_exceptImpl<0b000, 0b10, "hvc">; 1890 def SMCi : A64I_exceptImpl<0b000, 0b11, "smc">; 1891 def BRKi : A64I_exceptImpl<0b001, 0b00, "brk">; 1892 def HLTi : A64I_exceptImpl<0b010, 0b00, "hlt">; 1893 1894 def DCPS1i : A64I_exceptImpl<0b101, 0b01, "dcps1">; 1895 def DCPS2i : A64I_exceptImpl<0b101, 0b10, "dcps2">; 1896 def DCPS3i : A64I_exceptImpl<0b101, 0b11, "dcps3">; 1897 1898 // The immediate is optional for the DCPS instructions, defaulting to 0. 1899 def : InstAlias<"dcps1", (DCPS1i 0)>; 1900 def : InstAlias<"dcps2", (DCPS2i 0)>; 1901 def : InstAlias<"dcps3", (DCPS3i 0)>; 1902 1903 //===----------------------------------------------------------------------===// 1904 // Extract (immediate) 1905 //===----------------------------------------------------------------------===// 1906 // Contains: EXTR + alias ROR 1907 1908 def EXTRwwwi : A64I_extract<0b0, 0b000, 0b0, 1909 (outs GPR32:$Rd), 1910 (ins GPR32:$Rn, GPR32:$Rm, bitfield32_imm:$LSB), 1911 "extr\t$Rd, $Rn, $Rm, $LSB", 1912 [(set GPR32:$Rd, 1913 (A64Extr GPR32:$Rn, GPR32:$Rm, imm:$LSB))], 1914 NoItinerary>; 1915 def EXTRxxxi : A64I_extract<0b1, 0b000, 0b1, 1916 (outs GPR64:$Rd), 1917 (ins GPR64:$Rn, GPR64:$Rm, bitfield64_imm:$LSB), 1918 "extr\t$Rd, $Rn, $Rm, $LSB", 1919 [(set GPR64:$Rd, 1920 (A64Extr GPR64:$Rn, GPR64:$Rm, imm:$LSB))], 1921 NoItinerary>; 1922 1923 def : InstAlias<"ror $Rd, $Rs, $LSB", 1924 (EXTRwwwi GPR32:$Rd, GPR32:$Rs, GPR32:$Rs, bitfield32_imm:$LSB)>; 1925 def : InstAlias<"ror $Rd, $Rs, $LSB", 1926 (EXTRxxxi GPR64:$Rd, GPR64:$Rs, GPR64:$Rs, bitfield64_imm:$LSB)>; 1927 1928 def : Pat<(rotr GPR32:$Rn, bitfield32_imm:$LSB), 1929 (EXTRwwwi GPR32:$Rn, GPR32:$Rn, bitfield32_imm:$LSB)>; 1930 def : Pat<(rotr GPR64:$Rn, bitfield64_imm:$LSB), 1931 (EXTRxxxi GPR64:$Rn, GPR64:$Rn, bitfield64_imm:$LSB)>; 1932 1933 //===----------------------------------------------------------------------===// 1934 // Floating-point compare instructions 1935 //===----------------------------------------------------------------------===// 1936 // Contains: FCMP, FCMPE 1937 1938 def fpzero_asmoperand : AsmOperandClass { 1939 let Name = "FPZero"; 1940 let ParserMethod = "ParseFPImmOperand"; 1941 let DiagnosticType = "FPZero"; 1942 } 1943 1944 def fpz32 : Operand<f32>, 1945 ComplexPattern<f32, 1, "SelectFPZeroOperand", [fpimm]> { 1946 let ParserMatchClass = fpzero_asmoperand; 1947 let PrintMethod = "printFPZeroOperand"; 1948 let DecoderMethod = "DecodeFPZeroOperand"; 1949 } 1950 1951 def fpz64 : Operand<f64>, 1952 ComplexPattern<f64, 1, "SelectFPZeroOperand", [fpimm]> { 1953 let ParserMatchClass = fpzero_asmoperand; 1954 let PrintMethod = "printFPZeroOperand"; 1955 let DecoderMethod = "DecodeFPZeroOperand"; 1956 } 1957 1958 multiclass A64I_fpcmpSignal<bits<2> type, bit imm, dag ins, dag pattern> { 1959 def _quiet : A64I_fpcmp<0b0, 0b0, type, 0b00, {0b0, imm, 0b0, 0b0, 0b0}, 1960 (outs), ins, "fcmp\t$Rn, $Rm", [pattern], 1961 NoItinerary> { 1962 let Defs = [NZCV]; 1963 } 1964 1965 def _sig : A64I_fpcmp<0b0, 0b0, type, 0b00, {0b1, imm, 0b0, 0b0, 0b0}, 1966 (outs), ins, "fcmpe\t$Rn, $Rm", [], NoItinerary> { 1967 let Defs = [NZCV]; 1968 } 1969 } 1970 1971 defm FCMPss : A64I_fpcmpSignal<0b00, 0b0, (ins FPR32:$Rn, FPR32:$Rm), 1972 (set NZCV, (A64cmp (f32 FPR32:$Rn), FPR32:$Rm))>; 1973 defm FCMPdd : A64I_fpcmpSignal<0b01, 0b0, (ins FPR64:$Rn, FPR64:$Rm), 1974 (set NZCV, (A64cmp (f64 FPR64:$Rn), FPR64:$Rm))>; 1975 1976 // What would be Rm should be written as 0; note that even though it's called 1977 // "$Rm" here to fit in with the InstrFormats, it's actually an immediate. 1978 defm FCMPsi : A64I_fpcmpSignal<0b00, 0b1, (ins FPR32:$Rn, fpz32:$Rm), 1979 (set NZCV, (A64cmp (f32 FPR32:$Rn), fpz32:$Rm))>; 1980 1981 defm FCMPdi : A64I_fpcmpSignal<0b01, 0b1, (ins FPR64:$Rn, fpz64:$Rm), 1982 (set NZCV, (A64cmp (f64 FPR64:$Rn), fpz64:$Rm))>; 1983 1984 1985 //===----------------------------------------------------------------------===// 1986 // Floating-point conditional compare instructions 1987 //===----------------------------------------------------------------------===// 1988 // Contains: FCCMP, FCCMPE 1989 1990 class A64I_fpccmpImpl<bits<2> type, bit op, RegisterClass FPR, string asmop> 1991 : A64I_fpccmp<0b0, 0b0, type, op, 1992 (outs), 1993 (ins FPR:$Rn, FPR:$Rm, uimm4:$NZCVImm, cond_code_op:$Cond), 1994 !strconcat(asmop, "\t$Rn, $Rm, $NZCVImm, $Cond"), 1995 [], NoItinerary> { 1996 let Defs = [NZCV]; 1997 } 1998 1999 def FCCMPss : A64I_fpccmpImpl<0b00, 0b0, FPR32, "fccmp">; 2000 def FCCMPEss : A64I_fpccmpImpl<0b00, 0b1, FPR32, "fccmpe">; 2001 def FCCMPdd : A64I_fpccmpImpl<0b01, 0b0, FPR64, "fccmp">; 2002 def FCCMPEdd : A64I_fpccmpImpl<0b01, 0b1, FPR64, "fccmpe">; 2003 2004 //===----------------------------------------------------------------------===// 2005 // Floating-point conditional select instructions 2006 //===----------------------------------------------------------------------===// 2007 // Contains: FCSEL 2008 2009 let Uses = [NZCV] in { 2010 def FCSELsssc : A64I_fpcondsel<0b0, 0b0, 0b00, (outs FPR32:$Rd), 2011 (ins FPR32:$Rn, FPR32:$Rm, cond_code_op:$Cond), 2012 "fcsel\t$Rd, $Rn, $Rm, $Cond", 2013 [(set FPR32:$Rd, 2014 (simple_select (f32 FPR32:$Rn), 2015 FPR32:$Rm))], 2016 NoItinerary>; 2017 2018 2019 def FCSELdddc : A64I_fpcondsel<0b0, 0b0, 0b01, (outs FPR64:$Rd), 2020 (ins FPR64:$Rn, FPR64:$Rm, cond_code_op:$Cond), 2021 "fcsel\t$Rd, $Rn, $Rm, $Cond", 2022 [(set FPR64:$Rd, 2023 (simple_select (f64 FPR64:$Rn), 2024 FPR64:$Rm))], 2025 NoItinerary>; 2026 } 2027 2028 //===----------------------------------------------------------------------===// 2029 // Floating-point data-processing (1 source) 2030 //===----------------------------------------------------------------------===// 2031 // Contains: FMOV, FABS, FNEG, FSQRT, FCVT, FRINT[NPMZAXI]. 2032 2033 def FPNoUnop : PatFrag<(ops node:$val), (fneg node:$val), 2034 [{ (void)N; return false; }]>; 2035 2036 // First we do the fairly trivial bunch with uniform "OP s, s" and "OP d, d" 2037 // syntax. Default to no pattern because most are odd enough not to have one. 2038 multiclass A64I_fpdp1sizes<bits<6> opcode, string asmstr, 2039 SDPatternOperator opnode = FPNoUnop> { 2040 def ss : A64I_fpdp1<0b0, 0b0, 0b00, opcode, (outs FPR32:$Rd), (ins FPR32:$Rn), 2041 !strconcat(asmstr, "\t$Rd, $Rn"), 2042 [(set (f32 FPR32:$Rd), (opnode FPR32:$Rn))], 2043 NoItinerary>; 2044 2045 def dd : A64I_fpdp1<0b0, 0b0, 0b01, opcode, (outs FPR64:$Rd), (ins FPR64:$Rn), 2046 !strconcat(asmstr, "\t$Rd, $Rn"), 2047 [(set (f64 FPR64:$Rd), (opnode FPR64:$Rn))], 2048 NoItinerary>; 2049 } 2050 2051 defm FMOV : A64I_fpdp1sizes<0b000000, "fmov">; 2052 defm FABS : A64I_fpdp1sizes<0b000001, "fabs", fabs>; 2053 defm FNEG : A64I_fpdp1sizes<0b000010, "fneg", fneg>; 2054 defm FSQRT : A64I_fpdp1sizes<0b000011, "fsqrt", fsqrt>; 2055 2056 defm FRINTN : A64I_fpdp1sizes<0b001000, "frintn">; 2057 defm FRINTP : A64I_fpdp1sizes<0b001001, "frintp", fceil>; 2058 defm FRINTM : A64I_fpdp1sizes<0b001010, "frintm", ffloor>; 2059 defm FRINTZ : A64I_fpdp1sizes<0b001011, "frintz", ftrunc>; 2060 defm FRINTA : A64I_fpdp1sizes<0b001100, "frinta">; 2061 defm FRINTX : A64I_fpdp1sizes<0b001110, "frintx", frint>; 2062 defm FRINTI : A64I_fpdp1sizes<0b001111, "frinti", fnearbyint>; 2063 2064 // The FCVT instrucitons have different source and destination register-types, 2065 // but the fields are uniform everywhere a D-register (say) crops up. Package 2066 // this information in a Record. 2067 class FCVTRegType<RegisterClass rc, bits<2> fld, ValueType vt> { 2068 RegisterClass Class = rc; 2069 ValueType VT = vt; 2070 bit t1 = fld{1}; 2071 bit t0 = fld{0}; 2072 } 2073 2074 def FCVT16 : FCVTRegType<FPR16, 0b11, f16>; 2075 def FCVT32 : FCVTRegType<FPR32, 0b00, f32>; 2076 def FCVT64 : FCVTRegType<FPR64, 0b01, f64>; 2077 2078 class A64I_fpdp1_fcvt<FCVTRegType DestReg, FCVTRegType SrcReg, SDNode opnode> 2079 : A64I_fpdp1<0b0, 0b0, {SrcReg.t1, SrcReg.t0}, 2080 {0,0,0,1, DestReg.t1, DestReg.t0}, 2081 (outs DestReg.Class:$Rd), (ins SrcReg.Class:$Rn), 2082 "fcvt\t$Rd, $Rn", 2083 [(set (DestReg.VT DestReg.Class:$Rd), 2084 (opnode (SrcReg.VT SrcReg.Class:$Rn)))], NoItinerary>; 2085 2086 def FCVTds : A64I_fpdp1_fcvt<FCVT64, FCVT32, fextend>; 2087 def FCVThs : A64I_fpdp1_fcvt<FCVT16, FCVT32, fround>; 2088 def FCVTsd : A64I_fpdp1_fcvt<FCVT32, FCVT64, fround>; 2089 def FCVThd : A64I_fpdp1_fcvt<FCVT16, FCVT64, fround>; 2090 def FCVTsh : A64I_fpdp1_fcvt<FCVT32, FCVT16, fextend>; 2091 def FCVTdh : A64I_fpdp1_fcvt<FCVT64, FCVT16, fextend>; 2092 2093 2094 //===----------------------------------------------------------------------===// 2095 // Floating-point data-processing (2 sources) instructions 2096 //===----------------------------------------------------------------------===// 2097 // Contains: FMUL, FDIV, FADD, FSUB, FMAX, FMIN, FMAXNM, FMINNM, FNMUL 2098 2099 def FPNoBinop : PatFrag<(ops node:$lhs, node:$rhs), (fadd node:$lhs, node:$rhs), 2100 [{ (void)N; return false; }]>; 2101 2102 multiclass A64I_fpdp2sizes<bits<4> opcode, string asmstr, 2103 SDPatternOperator opnode> { 2104 def sss : A64I_fpdp2<0b0, 0b0, 0b00, opcode, 2105 (outs FPR32:$Rd), 2106 (ins FPR32:$Rn, FPR32:$Rm), 2107 !strconcat(asmstr, "\t$Rd, $Rn, $Rm"), 2108 [(set (f32 FPR32:$Rd), (opnode FPR32:$Rn, FPR32:$Rm))], 2109 NoItinerary>; 2110 2111 def ddd : A64I_fpdp2<0b0, 0b0, 0b01, opcode, 2112 (outs FPR64:$Rd), 2113 (ins FPR64:$Rn, FPR64:$Rm), 2114 !strconcat(asmstr, "\t$Rd, $Rn, $Rm"), 2115 [(set (f64 FPR64:$Rd), (opnode FPR64:$Rn, FPR64:$Rm))], 2116 NoItinerary>; 2117 } 2118 2119 let isCommutable = 1 in { 2120 defm FMUL : A64I_fpdp2sizes<0b0000, "fmul", fmul>; 2121 defm FADD : A64I_fpdp2sizes<0b0010, "fadd", fadd>; 2122 2123 // No patterns for these. 2124 defm FMAX : A64I_fpdp2sizes<0b0100, "fmax", FPNoBinop>; 2125 defm FMIN : A64I_fpdp2sizes<0b0101, "fmin", FPNoBinop>; 2126 defm FMAXNM : A64I_fpdp2sizes<0b0110, "fmaxnm", FPNoBinop>; 2127 defm FMINNM : A64I_fpdp2sizes<0b0111, "fminnm", FPNoBinop>; 2128 2129 defm FNMUL : A64I_fpdp2sizes<0b1000, "fnmul", 2130 PatFrag<(ops node:$lhs, node:$rhs), 2131 (fneg (fmul node:$lhs, node:$rhs))> >; 2132 } 2133 2134 defm FDIV : A64I_fpdp2sizes<0b0001, "fdiv", fdiv>; 2135 defm FSUB : A64I_fpdp2sizes<0b0011, "fsub", fsub>; 2136 2137 //===----------------------------------------------------------------------===// 2138 // Floating-point data-processing (3 sources) instructions 2139 //===----------------------------------------------------------------------===// 2140 // Contains: FMADD, FMSUB, FNMADD, FNMSUB 2141 2142 def fmsub : PatFrag<(ops node:$Rn, node:$Rm, node:$Ra), 2143 (fma (fneg node:$Rn), node:$Rm, node:$Ra)>; 2144 def fnmadd : PatFrag<(ops node:$Rn, node:$Rm, node:$Ra), 2145 (fma node:$Rn, node:$Rm, (fneg node:$Ra))>; 2146 def fnmsub : PatFrag<(ops node:$Rn, node:$Rm, node:$Ra), 2147 (fma (fneg node:$Rn), node:$Rm, (fneg node:$Ra))>; 2148 2149 class A64I_fpdp3Impl<string asmop, RegisterClass FPR, ValueType VT, 2150 bits<2> type, bit o1, bit o0, SDPatternOperator fmakind> 2151 : A64I_fpdp3<0b0, 0b0, type, o1, o0, (outs FPR:$Rd), 2152 (ins FPR:$Rn, FPR:$Rm, FPR:$Ra), 2153 !strconcat(asmop,"\t$Rd, $Rn, $Rm, $Ra"), 2154 [(set FPR:$Rd, (fmakind (VT FPR:$Rn), FPR:$Rm, FPR:$Ra))], 2155 NoItinerary>; 2156 2157 def FMADDssss : A64I_fpdp3Impl<"fmadd", FPR32, f32, 0b00, 0b0, 0b0, fma>; 2158 def FMSUBssss : A64I_fpdp3Impl<"fmsub", FPR32, f32, 0b00, 0b0, 0b1, fmsub>; 2159 def FNMADDssss : A64I_fpdp3Impl<"fnmadd", FPR32, f32, 0b00, 0b1, 0b0, fnmadd>; 2160 def FNMSUBssss : A64I_fpdp3Impl<"fnmsub", FPR32, f32, 0b00, 0b1, 0b1, fnmsub>; 2161 2162 def FMADDdddd : A64I_fpdp3Impl<"fmadd", FPR64, f64, 0b01, 0b0, 0b0, fma>; 2163 def FMSUBdddd : A64I_fpdp3Impl<"fmsub", FPR64, f64, 0b01, 0b0, 0b1, fmsub>; 2164 def FNMADDdddd : A64I_fpdp3Impl<"fnmadd", FPR64, f64, 0b01, 0b1, 0b0, fnmadd>; 2165 def FNMSUBdddd : A64I_fpdp3Impl<"fnmsub", FPR64, f64, 0b01, 0b1, 0b1, fnmsub>; 2166 2167 //===----------------------------------------------------------------------===// 2168 // Floating-point <-> fixed-point conversion instructions 2169 //===----------------------------------------------------------------------===// 2170 // Contains: FCVTZS, FCVTZU, SCVTF, UCVTF 2171 2172 // #1-#32 allowed, encoded as "64 - <specified imm> 2173 def fixedpos_asmoperand_i32 : AsmOperandClass { 2174 let Name = "CVTFixedPos32"; 2175 let RenderMethod = "addCVTFixedPosOperands"; 2176 let PredicateMethod = "isCVTFixedPos<32>"; 2177 let DiagnosticType = "CVTFixedPos32"; 2178 } 2179 2180 // Also encoded as "64 - <specified imm>" but #1-#64 allowed. 2181 def fixedpos_asmoperand_i64 : AsmOperandClass { 2182 let Name = "CVTFixedPos64"; 2183 let RenderMethod = "addCVTFixedPosOperands"; 2184 let PredicateMethod = "isCVTFixedPos<64>"; 2185 let DiagnosticType = "CVTFixedPos64"; 2186 } 2187 2188 // We need the cartesian product of f32/f64 i32/i64 operands for 2189 // conversions: 2190 // + Selection needs to use operands of correct floating type 2191 // + Assembly parsing and decoding depend on integer width 2192 class cvtfix_i32_op<ValueType FloatVT> 2193 : Operand<FloatVT>, 2194 ComplexPattern<FloatVT, 1, "SelectCVTFixedPosOperand<32>", [fpimm]> { 2195 let ParserMatchClass = fixedpos_asmoperand_i32; 2196 let DecoderMethod = "DecodeCVT32FixedPosOperand"; 2197 let PrintMethod = "printCVTFixedPosOperand"; 2198 } 2199 2200 class cvtfix_i64_op<ValueType FloatVT> 2201 : Operand<FloatVT>, 2202 ComplexPattern<FloatVT, 1, "SelectCVTFixedPosOperand<64>", [fpimm]> { 2203 let ParserMatchClass = fixedpos_asmoperand_i64; 2204 let PrintMethod = "printCVTFixedPosOperand"; 2205 } 2206 2207 // Because of the proliferation of weird operands, it's not really 2208 // worth going for a multiclass here. Oh well. 2209 2210 class A64I_fptofix<bit sf, bits<2> type, bits<3> opcode, 2211 RegisterClass GPR, RegisterClass FPR, Operand scale_op, 2212 string asmop, SDNode cvtop> 2213 : A64I_fpfixed<sf, 0b0, type, 0b11, opcode, 2214 (outs GPR:$Rd), (ins FPR:$Rn, scale_op:$Scale), 2215 !strconcat(asmop, "\t$Rd, $Rn, $Scale"), 2216 [(set GPR:$Rd, (cvtop (fmul FPR:$Rn, scale_op:$Scale)))], 2217 NoItinerary>; 2218 2219 def FCVTZSwsi : A64I_fptofix<0b0, 0b00, 0b000, GPR32, FPR32, 2220 cvtfix_i32_op<f32>, "fcvtzs", fp_to_sint>; 2221 def FCVTZSxsi : A64I_fptofix<0b1, 0b00, 0b000, GPR64, FPR32, 2222 cvtfix_i64_op<f32>, "fcvtzs", fp_to_sint>; 2223 def FCVTZUwsi : A64I_fptofix<0b0, 0b00, 0b001, GPR32, FPR32, 2224 cvtfix_i32_op<f32>, "fcvtzu", fp_to_uint>; 2225 def FCVTZUxsi : A64I_fptofix<0b1, 0b00, 0b001, GPR64, FPR32, 2226 cvtfix_i64_op<f32>, "fcvtzu", fp_to_uint>; 2227 2228 def FCVTZSwdi : A64I_fptofix<0b0, 0b01, 0b000, GPR32, FPR64, 2229 cvtfix_i32_op<f64>, "fcvtzs", fp_to_sint>; 2230 def FCVTZSxdi : A64I_fptofix<0b1, 0b01, 0b000, GPR64, FPR64, 2231 cvtfix_i64_op<f64>, "fcvtzs", fp_to_sint>; 2232 def FCVTZUwdi : A64I_fptofix<0b0, 0b01, 0b001, GPR32, FPR64, 2233 cvtfix_i32_op<f64>, "fcvtzu", fp_to_uint>; 2234 def FCVTZUxdi : A64I_fptofix<0b1, 0b01, 0b001, GPR64, FPR64, 2235 cvtfix_i64_op<f64>, "fcvtzu", fp_to_uint>; 2236 2237 2238 class A64I_fixtofp<bit sf, bits<2> type, bits<3> opcode, 2239 RegisterClass FPR, RegisterClass GPR, Operand scale_op, 2240 string asmop, SDNode cvtop> 2241 : A64I_fpfixed<sf, 0b0, type, 0b00, opcode, 2242 (outs FPR:$Rd), (ins GPR:$Rn, scale_op:$Scale), 2243 !strconcat(asmop, "\t$Rd, $Rn, $Scale"), 2244 [(set FPR:$Rd, (fdiv (cvtop GPR:$Rn), scale_op:$Scale))], 2245 NoItinerary>; 2246 2247 def SCVTFswi : A64I_fixtofp<0b0, 0b00, 0b010, FPR32, GPR32, 2248 cvtfix_i32_op<f32>, "scvtf", sint_to_fp>; 2249 def SCVTFsxi : A64I_fixtofp<0b1, 0b00, 0b010, FPR32, GPR64, 2250 cvtfix_i64_op<f32>, "scvtf", sint_to_fp>; 2251 def UCVTFswi : A64I_fixtofp<0b0, 0b00, 0b011, FPR32, GPR32, 2252 cvtfix_i32_op<f32>, "ucvtf", uint_to_fp>; 2253 def UCVTFsxi : A64I_fixtofp<0b1, 0b00, 0b011, FPR32, GPR64, 2254 cvtfix_i64_op<f32>, "ucvtf", uint_to_fp>; 2255 def SCVTFdwi : A64I_fixtofp<0b0, 0b01, 0b010, FPR64, GPR32, 2256 cvtfix_i32_op<f64>, "scvtf", sint_to_fp>; 2257 def SCVTFdxi : A64I_fixtofp<0b1, 0b01, 0b010, FPR64, GPR64, 2258 cvtfix_i64_op<f64>, "scvtf", sint_to_fp>; 2259 def UCVTFdwi : A64I_fixtofp<0b0, 0b01, 0b011, FPR64, GPR32, 2260 cvtfix_i32_op<f64>, "ucvtf", uint_to_fp>; 2261 def UCVTFdxi : A64I_fixtofp<0b1, 0b01, 0b011, FPR64, GPR64, 2262 cvtfix_i64_op<f64>, "ucvtf", uint_to_fp>; 2263 2264 //===----------------------------------------------------------------------===// 2265 // Floating-point <-> integer conversion instructions 2266 //===----------------------------------------------------------------------===// 2267 // Contains: FCVTZS, FCVTZU, SCVTF, UCVTF 2268 2269 class A64I_fpintI<bit sf, bits<2> type, bits<2> rmode, bits<3> opcode, 2270 RegisterClass DestPR, RegisterClass SrcPR, string asmop> 2271 : A64I_fpint<sf, 0b0, type, rmode, opcode, (outs DestPR:$Rd), (ins SrcPR:$Rn), 2272 !strconcat(asmop, "\t$Rd, $Rn"), [], NoItinerary>; 2273 2274 multiclass A64I_fptointRM<bits<2> rmode, bit o2, string asmop> { 2275 def Sws : A64I_fpintI<0b0, 0b00, rmode, {o2, 0, 0}, 2276 GPR32, FPR32, asmop # "s">; 2277 def Sxs : A64I_fpintI<0b1, 0b00, rmode, {o2, 0, 0}, 2278 GPR64, FPR32, asmop # "s">; 2279 def Uws : A64I_fpintI<0b0, 0b00, rmode, {o2, 0, 1}, 2280 GPR32, FPR32, asmop # "u">; 2281 def Uxs : A64I_fpintI<0b1, 0b00, rmode, {o2, 0, 1}, 2282 GPR64, FPR32, asmop # "u">; 2283 2284 def Swd : A64I_fpintI<0b0, 0b01, rmode, {o2, 0, 0}, 2285 GPR32, FPR64, asmop # "s">; 2286 def Sxd : A64I_fpintI<0b1, 0b01, rmode, {o2, 0, 0}, 2287 GPR64, FPR64, asmop # "s">; 2288 def Uwd : A64I_fpintI<0b0, 0b01, rmode, {o2, 0, 1}, 2289 GPR32, FPR64, asmop # "u">; 2290 def Uxd : A64I_fpintI<0b1, 0b01, rmode, {o2, 0, 1}, 2291 GPR64, FPR64, asmop # "u">; 2292 } 2293 2294 defm FCVTN : A64I_fptointRM<0b00, 0b0, "fcvtn">; 2295 defm FCVTP : A64I_fptointRM<0b01, 0b0, "fcvtp">; 2296 defm FCVTM : A64I_fptointRM<0b10, 0b0, "fcvtm">; 2297 defm FCVTZ : A64I_fptointRM<0b11, 0b0, "fcvtz">; 2298 defm FCVTA : A64I_fptointRM<0b00, 0b1, "fcvta">; 2299 2300 def : Pat<(i32 (fp_to_sint FPR32:$Rn)), (FCVTZSws FPR32:$Rn)>; 2301 def : Pat<(i64 (fp_to_sint FPR32:$Rn)), (FCVTZSxs FPR32:$Rn)>; 2302 def : Pat<(i32 (fp_to_uint FPR32:$Rn)), (FCVTZUws FPR32:$Rn)>; 2303 def : Pat<(i64 (fp_to_uint FPR32:$Rn)), (FCVTZUxs FPR32:$Rn)>; 2304 def : Pat<(i32 (fp_to_sint (f64 FPR64:$Rn))), (FCVTZSwd FPR64:$Rn)>; 2305 def : Pat<(i64 (fp_to_sint (f64 FPR64:$Rn))), (FCVTZSxd FPR64:$Rn)>; 2306 def : Pat<(i32 (fp_to_uint (f64 FPR64:$Rn))), (FCVTZUwd FPR64:$Rn)>; 2307 def : Pat<(i64 (fp_to_uint (f64 FPR64:$Rn))), (FCVTZUxd FPR64:$Rn)>; 2308 2309 multiclass A64I_inttofp<bit o0, string asmop> { 2310 def CVTFsw : A64I_fpintI<0b0, 0b00, 0b00, {0, 1, o0}, FPR32, GPR32, asmop>; 2311 def CVTFsx : A64I_fpintI<0b1, 0b00, 0b00, {0, 1, o0}, FPR32, GPR64, asmop>; 2312 def CVTFdw : A64I_fpintI<0b0, 0b01, 0b00, {0, 1, o0}, FPR64, GPR32, asmop>; 2313 def CVTFdx : A64I_fpintI<0b1, 0b01, 0b00, {0, 1, o0}, FPR64, GPR64, asmop>; 2314 } 2315 2316 defm S : A64I_inttofp<0b0, "scvtf">; 2317 defm U : A64I_inttofp<0b1, "ucvtf">; 2318 2319 def : Pat<(f32 (sint_to_fp GPR32:$Rn)), (SCVTFsw GPR32:$Rn)>; 2320 def : Pat<(f32 (sint_to_fp GPR64:$Rn)), (SCVTFsx GPR64:$Rn)>; 2321 def : Pat<(f64 (sint_to_fp GPR32:$Rn)), (SCVTFdw GPR32:$Rn)>; 2322 def : Pat<(f64 (sint_to_fp GPR64:$Rn)), (SCVTFdx GPR64:$Rn)>; 2323 def : Pat<(f32 (uint_to_fp GPR32:$Rn)), (UCVTFsw GPR32:$Rn)>; 2324 def : Pat<(f32 (uint_to_fp GPR64:$Rn)), (UCVTFsx GPR64:$Rn)>; 2325 def : Pat<(f64 (uint_to_fp GPR32:$Rn)), (UCVTFdw GPR32:$Rn)>; 2326 def : Pat<(f64 (uint_to_fp GPR64:$Rn)), (UCVTFdx GPR64:$Rn)>; 2327 2328 def FMOVws : A64I_fpintI<0b0, 0b00, 0b00, 0b110, GPR32, FPR32, "fmov">; 2329 def FMOVsw : A64I_fpintI<0b0, 0b00, 0b00, 0b111, FPR32, GPR32, "fmov">; 2330 def FMOVxd : A64I_fpintI<0b1, 0b01, 0b00, 0b110, GPR64, FPR64, "fmov">; 2331 def FMOVdx : A64I_fpintI<0b1, 0b01, 0b00, 0b111, FPR64, GPR64, "fmov">; 2332 2333 def : Pat<(i32 (bitconvert (f32 FPR32:$Rn))), (FMOVws FPR32:$Rn)>; 2334 def : Pat<(f32 (bitconvert (i32 GPR32:$Rn))), (FMOVsw GPR32:$Rn)>; 2335 def : Pat<(i64 (bitconvert (f64 FPR64:$Rn))), (FMOVxd FPR64:$Rn)>; 2336 def : Pat<(f64 (bitconvert (i64 GPR64:$Rn))), (FMOVdx GPR64:$Rn)>; 2337 2338 def lane1_asmoperand : AsmOperandClass { 2339 let Name = "Lane1"; 2340 let RenderMethod = "addImmOperands"; 2341 let DiagnosticType = "Lane1"; 2342 } 2343 2344 def lane1 : Operand<i32> { 2345 let ParserMatchClass = lane1_asmoperand; 2346 let PrintMethod = "printBareImmOperand"; 2347 } 2348 2349 let DecoderMethod = "DecodeFMOVLaneInstruction" in { 2350 def FMOVxv : A64I_fpint<0b1, 0b0, 0b10, 0b01, 0b110, 2351 (outs GPR64:$Rd), (ins VPR128:$Rn, lane1:$Lane), 2352 "fmov\t$Rd, $Rn.d[$Lane]", [], NoItinerary>; 2353 2354 def FMOVvx : A64I_fpint<0b1, 0b0, 0b10, 0b01, 0b111, 2355 (outs VPR128:$Rd), (ins GPR64:$Rn, lane1:$Lane), 2356 "fmov\t$Rd.d[$Lane], $Rn", [], NoItinerary>; 2357 } 2358 2359 def : InstAlias<"fmov $Rd, $Rn.2d[$Lane]", 2360 (FMOVxv GPR64:$Rd, VPR128:$Rn, lane1:$Lane), 0b0>; 2361 2362 def : InstAlias<"fmov $Rd.2d[$Lane], $Rn", 2363 (FMOVvx VPR128:$Rd, GPR64:$Rn, lane1:$Lane), 0b0>; 2364 2365 //===----------------------------------------------------------------------===// 2366 // Floating-point immediate instructions 2367 //===----------------------------------------------------------------------===// 2368 // Contains: FMOV 2369 2370 def fpimm_asmoperand : AsmOperandClass { 2371 let Name = "FMOVImm"; 2372 let ParserMethod = "ParseFPImmOperand"; 2373 let DiagnosticType = "FPImm"; 2374 } 2375 2376 // The MCOperand for these instructions are the encoded 8-bit values. 2377 def SDXF_fpimm : SDNodeXForm<fpimm, [{ 2378 uint32_t Imm8; 2379 A64Imms::isFPImm(N->getValueAPF(), Imm8); 2380 return CurDAG->getTargetConstant(Imm8, MVT::i32); 2381 }]>; 2382 2383 class fmov_operand<ValueType FT> 2384 : Operand<i32>, 2385 PatLeaf<(FT fpimm), [{ return A64Imms::isFPImm(N->getValueAPF()); }], 2386 SDXF_fpimm> { 2387 let PrintMethod = "printFPImmOperand"; 2388 let ParserMatchClass = fpimm_asmoperand; 2389 } 2390 2391 def fmov32_operand : fmov_operand<f32>; 2392 def fmov64_operand : fmov_operand<f64>; 2393 2394 class A64I_fpimm_impl<bits<2> type, RegisterClass Reg, ValueType VT, 2395 Operand fmov_operand> 2396 : A64I_fpimm<0b0, 0b0, type, 0b00000, 2397 (outs Reg:$Rd), 2398 (ins fmov_operand:$Imm8), 2399 "fmov\t$Rd, $Imm8", 2400 [(set (VT Reg:$Rd), fmov_operand:$Imm8)], 2401 NoItinerary>; 2402 2403 def FMOVsi : A64I_fpimm_impl<0b00, FPR32, f32, fmov32_operand>; 2404 def FMOVdi : A64I_fpimm_impl<0b01, FPR64, f64, fmov64_operand>; 2405 2406 //===----------------------------------------------------------------------===// 2407 // Load-register (literal) instructions 2408 //===----------------------------------------------------------------------===// 2409 // Contains: LDR, LDRSW, PRFM 2410 2411 def ldrlit_label_asmoperand : AsmOperandClass { 2412 let Name = "LoadLitLabel"; 2413 let RenderMethod = "addLabelOperands<19, 4>"; 2414 let DiagnosticType = "Label"; 2415 } 2416 2417 def ldrlit_label : Operand<i64> { 2418 let EncoderMethod = "getLoadLitLabelOpValue"; 2419 2420 // This label is a 19-bit offset from PC, scaled by the instruction-width: 4. 2421 let PrintMethod = "printLabelOperand<19, 4>"; 2422 let ParserMatchClass = ldrlit_label_asmoperand; 2423 let OperandType = "OPERAND_PCREL"; 2424 } 2425 2426 // Various instructions take an immediate value (which can always be used), 2427 // where some numbers have a symbolic name to make things easier. These operands 2428 // and the associated functions abstract away the differences. 2429 multiclass namedimm<string prefix, string mapper> { 2430 def _asmoperand : AsmOperandClass { 2431 let Name = "NamedImm" # prefix; 2432 let PredicateMethod = "isUImm"; 2433 let RenderMethod = "addImmOperands"; 2434 let ParserMethod = "ParseNamedImmOperand<" # mapper # ">"; 2435 let DiagnosticType = "NamedImm_" # prefix; 2436 } 2437 2438 def _op : Operand<i32> { 2439 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand"); 2440 let PrintMethod = "printNamedImmOperand<" # mapper # ">"; 2441 let DecoderMethod = "DecodeNamedImmOperand<" # mapper # ">"; 2442 } 2443 } 2444 2445 defm prefetch : namedimm<"prefetch", "A64PRFM::PRFMMapper">; 2446 2447 class A64I_LDRlitSimple<bits<2> opc, bit v, RegisterClass OutReg, 2448 list<dag> patterns = []> 2449 : A64I_LDRlit<opc, v, (outs OutReg:$Rt), (ins ldrlit_label:$Imm19), 2450 "ldr\t$Rt, $Imm19", patterns, NoItinerary>; 2451 2452 let mayLoad = 1 in { 2453 def LDRw_lit : A64I_LDRlitSimple<0b00, 0b0, GPR32>; 2454 def LDRx_lit : A64I_LDRlitSimple<0b01, 0b0, GPR64>; 2455 } 2456 2457 def LDRs_lit : A64I_LDRlitSimple<0b00, 0b1, FPR32>; 2458 def LDRd_lit : A64I_LDRlitSimple<0b01, 0b1, FPR64>; 2459 2460 let mayLoad = 1 in { 2461 def LDRq_lit : A64I_LDRlitSimple<0b10, 0b1, FPR128>; 2462 2463 2464 def LDRSWx_lit : A64I_LDRlit<0b10, 0b0, 2465 (outs GPR64:$Rt), 2466 (ins ldrlit_label:$Imm19), 2467 "ldrsw\t$Rt, $Imm19", 2468 [], NoItinerary>; 2469 2470 def PRFM_lit : A64I_LDRlit<0b11, 0b0, 2471 (outs), (ins prefetch_op:$Rt, ldrlit_label:$Imm19), 2472 "prfm\t$Rt, $Imm19", 2473 [], NoItinerary>; 2474 } 2475 2476 //===----------------------------------------------------------------------===// 2477 // Load-store exclusive instructions 2478 //===----------------------------------------------------------------------===// 2479 // Contains: STXRB, STXRH, STXR, LDXRB, LDXRH, LDXR. STXP, LDXP, STLXRB, 2480 // STLXRH, STLXR, LDAXRB, LDAXRH, LDAXR, STLXP, LDAXP, STLRB, 2481 // STLRH, STLR, LDARB, LDARH, LDAR 2482 2483 // Since these instructions have the undefined register bits set to 1 in 2484 // their canonical form, we need a post encoder method to set those bits 2485 // to 1 when encoding these instructions. We do this using the 2486 // fixLoadStoreExclusive function. This function has template parameters: 2487 // 2488 // fixLoadStoreExclusive<int hasRs, int hasRt2> 2489 // 2490 // hasRs indicates that the instruction uses the Rs field, so we won't set 2491 // it to 1 (and the same for Rt2). We don't need template parameters for 2492 // the other register fiels since Rt and Rn are always used. 2493 2494 // This operand parses a GPR64xsp register, followed by an optional immediate 2495 // #0. 2496 def GPR64xsp0_asmoperand : AsmOperandClass { 2497 let Name = "GPR64xsp0"; 2498 let PredicateMethod = "isWrappedReg"; 2499 let RenderMethod = "addRegOperands"; 2500 let ParserMethod = "ParseLSXAddressOperand"; 2501 // Diagnostics are provided by ParserMethod 2502 } 2503 2504 def GPR64xsp0 : RegisterOperand<GPR64xsp> { 2505 let ParserMatchClass = GPR64xsp0_asmoperand; 2506 } 2507 2508 //===---------------------------------- 2509 // Store-exclusive (releasing & normal) 2510 //===---------------------------------- 2511 2512 class A64I_SRexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs, 2513 dag ins, list<dag> pat, 2514 InstrItinClass itin> : 2515 A64I_LDSTex_stn <size, 2516 opcode{2}, 0, opcode{1}, opcode{0}, 2517 outs, ins, 2518 !strconcat(asm, "\t$Rs, $Rt, [$Rn]"), 2519 pat, itin> { 2520 let mayStore = 1; 2521 let PostEncoderMethod = "fixLoadStoreExclusive<1,0>"; 2522 } 2523 2524 multiclass A64I_SRex<string asmstr, bits<3> opcode, string prefix> { 2525 def _byte: A64I_SRexs_impl<0b00, opcode, !strconcat(asmstr, "b"), 2526 (outs GPR32:$Rs), (ins GPR32:$Rt, GPR64xsp0:$Rn), 2527 [], NoItinerary>; 2528 2529 def _hword: A64I_SRexs_impl<0b01, opcode, !strconcat(asmstr, "h"), 2530 (outs GPR32:$Rs), (ins GPR32:$Rt, GPR64xsp0:$Rn), 2531 [],NoItinerary>; 2532 2533 def _word: A64I_SRexs_impl<0b10, opcode, asmstr, 2534 (outs GPR32:$Rs), (ins GPR32:$Rt, GPR64xsp0:$Rn), 2535 [], NoItinerary>; 2536 2537 def _dword: A64I_SRexs_impl<0b11, opcode, asmstr, 2538 (outs GPR32:$Rs), (ins GPR64:$Rt, GPR64xsp0:$Rn), 2539 [], NoItinerary>; 2540 } 2541 2542 defm STXR : A64I_SRex<"stxr", 0b000, "STXR">; 2543 defm STLXR : A64I_SRex<"stlxr", 0b001, "STLXR">; 2544 2545 //===---------------------------------- 2546 // Loads 2547 //===---------------------------------- 2548 2549 class A64I_LRexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs, 2550 dag ins, list<dag> pat, 2551 InstrItinClass itin> : 2552 A64I_LDSTex_tn <size, 2553 opcode{2}, 1, opcode{1}, opcode{0}, 2554 outs, ins, 2555 !strconcat(asm, "\t$Rt, [$Rn]"), 2556 pat, itin> { 2557 let mayLoad = 1; 2558 let PostEncoderMethod = "fixLoadStoreExclusive<0,0>"; 2559 } 2560 2561 multiclass A64I_LRex<string asmstr, bits<3> opcode> { 2562 def _byte: A64I_LRexs_impl<0b00, opcode, !strconcat(asmstr, "b"), 2563 (outs GPR32:$Rt), (ins GPR64xsp0:$Rn), 2564 [], NoItinerary>; 2565 2566 def _hword: A64I_LRexs_impl<0b01, opcode, !strconcat(asmstr, "h"), 2567 (outs GPR32:$Rt), (ins GPR64xsp0:$Rn), 2568 [], NoItinerary>; 2569 2570 def _word: A64I_LRexs_impl<0b10, opcode, asmstr, 2571 (outs GPR32:$Rt), (ins GPR64xsp0:$Rn), 2572 [], NoItinerary>; 2573 2574 def _dword: A64I_LRexs_impl<0b11, opcode, asmstr, 2575 (outs GPR64:$Rt), (ins GPR64xsp0:$Rn), 2576 [], NoItinerary>; 2577 } 2578 2579 defm LDXR : A64I_LRex<"ldxr", 0b000>; 2580 defm LDAXR : A64I_LRex<"ldaxr", 0b001>; 2581 defm LDAR : A64I_LRex<"ldar", 0b101>; 2582 2583 class acquiring_load<PatFrag base> 2584 : PatFrag<(ops node:$ptr), (base node:$ptr), [{ 2585 return cast<AtomicSDNode>(N)->getOrdering() == Acquire; 2586 }]>; 2587 2588 def atomic_load_acquire_8 : acquiring_load<atomic_load_8>; 2589 def atomic_load_acquire_16 : acquiring_load<atomic_load_16>; 2590 def atomic_load_acquire_32 : acquiring_load<atomic_load_32>; 2591 def atomic_load_acquire_64 : acquiring_load<atomic_load_64>; 2592 2593 def : Pat<(atomic_load_acquire_8 GPR64xsp:$Rn), (LDAR_byte GPR64xsp0:$Rn)>; 2594 def : Pat<(atomic_load_acquire_16 GPR64xsp:$Rn), (LDAR_hword GPR64xsp0:$Rn)>; 2595 def : Pat<(atomic_load_acquire_32 GPR64xsp:$Rn), (LDAR_word GPR64xsp0:$Rn)>; 2596 def : Pat<(atomic_load_acquire_64 GPR64xsp:$Rn), (LDAR_dword GPR64xsp0:$Rn)>; 2597 2598 //===---------------------------------- 2599 // Store-release (no exclusivity) 2600 //===---------------------------------- 2601 2602 class A64I_SLexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs, 2603 dag ins, list<dag> pat, 2604 InstrItinClass itin> : 2605 A64I_LDSTex_tn <size, 2606 opcode{2}, 0, opcode{1}, opcode{0}, 2607 outs, ins, 2608 !strconcat(asm, "\t$Rt, [$Rn]"), 2609 pat, itin> { 2610 let mayStore = 1; 2611 let PostEncoderMethod = "fixLoadStoreExclusive<0,0>"; 2612 } 2613 2614 class releasing_store<PatFrag base> 2615 : PatFrag<(ops node:$ptr, node:$val), (base node:$ptr, node:$val), [{ 2616 return cast<AtomicSDNode>(N)->getOrdering() == Release; 2617 }]>; 2618 2619 def atomic_store_release_8 : releasing_store<atomic_store_8>; 2620 def atomic_store_release_16 : releasing_store<atomic_store_16>; 2621 def atomic_store_release_32 : releasing_store<atomic_store_32>; 2622 def atomic_store_release_64 : releasing_store<atomic_store_64>; 2623 2624 multiclass A64I_SLex<string asmstr, bits<3> opcode, string prefix> { 2625 def _byte: A64I_SLexs_impl<0b00, opcode, !strconcat(asmstr, "b"), 2626 (outs), (ins GPR32:$Rt, GPR64xsp0:$Rn), 2627 [(atomic_store_release_8 GPR64xsp0:$Rn, GPR32:$Rt)], 2628 NoItinerary>; 2629 2630 def _hword: A64I_SLexs_impl<0b01, opcode, !strconcat(asmstr, "h"), 2631 (outs), (ins GPR32:$Rt, GPR64xsp0:$Rn), 2632 [(atomic_store_release_16 GPR64xsp0:$Rn, GPR32:$Rt)], 2633 NoItinerary>; 2634 2635 def _word: A64I_SLexs_impl<0b10, opcode, asmstr, 2636 (outs), (ins GPR32:$Rt, GPR64xsp0:$Rn), 2637 [(atomic_store_release_32 GPR64xsp0:$Rn, GPR32:$Rt)], 2638 NoItinerary>; 2639 2640 def _dword: A64I_SLexs_impl<0b11, opcode, asmstr, 2641 (outs), (ins GPR64:$Rt, GPR64xsp0:$Rn), 2642 [(atomic_store_release_64 GPR64xsp0:$Rn, GPR64:$Rt)], 2643 NoItinerary>; 2644 } 2645 2646 defm STLR : A64I_SLex<"stlr", 0b101, "STLR">; 2647 2648 //===---------------------------------- 2649 // Store-exclusive pair (releasing & normal) 2650 //===---------------------------------- 2651 2652 class A64I_SPexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs, 2653 dag ins, list<dag> pat, 2654 InstrItinClass itin> : 2655 A64I_LDSTex_stt2n <size, 2656 opcode{2}, 0, opcode{1}, opcode{0}, 2657 outs, ins, 2658 !strconcat(asm, "\t$Rs, $Rt, $Rt2, [$Rn]"), 2659 pat, itin> { 2660 let mayStore = 1; 2661 } 2662 2663 2664 multiclass A64I_SPex<string asmstr, bits<3> opcode> { 2665 def _word: A64I_SPexs_impl<0b10, opcode, asmstr, (outs), 2666 (ins GPR32:$Rs, GPR32:$Rt, GPR32:$Rt2, 2667 GPR64xsp0:$Rn), 2668 [], NoItinerary>; 2669 2670 def _dword: A64I_SPexs_impl<0b11, opcode, asmstr, (outs), 2671 (ins GPR32:$Rs, GPR64:$Rt, GPR64:$Rt2, 2672 GPR64xsp0:$Rn), 2673 [], NoItinerary>; 2674 } 2675 2676 defm STXP : A64I_SPex<"stxp", 0b010>; 2677 defm STLXP : A64I_SPex<"stlxp", 0b011>; 2678 2679 //===---------------------------------- 2680 // Load-exclusive pair (acquiring & normal) 2681 //===---------------------------------- 2682 2683 class A64I_LPexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs, 2684 dag ins, list<dag> pat, 2685 InstrItinClass itin> : 2686 A64I_LDSTex_tt2n <size, 2687 opcode{2}, 1, opcode{1}, opcode{0}, 2688 outs, ins, 2689 !strconcat(asm, "\t$Rt, $Rt2, [$Rn]"), 2690 pat, itin>{ 2691 let mayLoad = 1; 2692 let DecoderMethod = "DecodeLoadPairExclusiveInstruction"; 2693 let PostEncoderMethod = "fixLoadStoreExclusive<0,1>"; 2694 } 2695 2696 multiclass A64I_LPex<string asmstr, bits<3> opcode> { 2697 def _word: A64I_LPexs_impl<0b10, opcode, asmstr, 2698 (outs GPR32:$Rt, GPR32:$Rt2), 2699 (ins GPR64xsp0:$Rn), 2700 [], NoItinerary>; 2701 2702 def _dword: A64I_LPexs_impl<0b11, opcode, asmstr, 2703 (outs GPR64:$Rt, GPR64:$Rt2), 2704 (ins GPR64xsp0:$Rn), 2705 [], NoItinerary>; 2706 } 2707 2708 defm LDXP : A64I_LPex<"ldxp", 0b010>; 2709 defm LDAXP : A64I_LPex<"ldaxp", 0b011>; 2710 2711 //===----------------------------------------------------------------------===// 2712 // Load-store register (unscaled immediate) instructions 2713 //===----------------------------------------------------------------------===// 2714 // Contains: LDURB, LDURH, LDRUSB, LDRUSH, LDRUSW, STUR, STURB, STURH and PRFUM 2715 // 2716 // and 2717 // 2718 //===----------------------------------------------------------------------===// 2719 // Load-store register (register offset) instructions 2720 //===----------------------------------------------------------------------===// 2721 // Contains: LDRB, LDRH, LDRSB, LDRSH, LDRSW, STR, STRB, STRH and PRFM 2722 // 2723 // and 2724 // 2725 //===----------------------------------------------------------------------===// 2726 // Load-store register (unsigned immediate) instructions 2727 //===----------------------------------------------------------------------===// 2728 // Contains: LDRB, LDRH, LDRSB, LDRSH, LDRSW, STR, STRB, STRH and PRFM 2729 // 2730 // and 2731 // 2732 //===----------------------------------------------------------------------===// 2733 // Load-store register (immediate post-indexed) instructions 2734 //===----------------------------------------------------------------------===// 2735 // Contains: STRB, STRH, STR, LDRB, LDRH, LDR, LDRSB, LDRSH, LDRSW 2736 // 2737 // and 2738 // 2739 //===----------------------------------------------------------------------===// 2740 // Load-store register (immediate pre-indexed) instructions 2741 //===----------------------------------------------------------------------===// 2742 // Contains: STRB, STRH, STR, LDRB, LDRH, LDR, LDRSB, LDRSH, LDRSW 2743 2744 // Note that patterns are much later on in a completely separate section (they 2745 // need ADRPxi to be defined). 2746 2747 //===------------------------------- 2748 // 1. Various operands needed 2749 //===------------------------------- 2750 2751 //===------------------------------- 2752 // 1.1 Unsigned 12-bit immediate operands 2753 //===------------------------------- 2754 // The addressing mode for these instructions consists of an unsigned 12-bit 2755 // immediate which is scaled by the size of the memory access. 2756 // 2757 // We represent this in the MC layer by two operands: 2758 // 1. A base register. 2759 // 2. A 12-bit immediate: not multiplied by access size, so "LDR x0,[x0,#8]" 2760 // would have '1' in this field. 2761 // This means that separate functions are needed for converting representations 2762 // which *are* aware of the intended access size. 2763 2764 // Anything that creates an MCInst (Decoding, selection and AsmParsing) has to 2765 // know the access size via some means. An isolated operand does not have this 2766 // information unless told from here, which means we need separate tablegen 2767 // Operands for each access size. This multiclass takes care of instantiating 2768 // the correct template functions in the rest of the backend. 2769 2770 //===------------------------------- 2771 // 1.1 Unsigned 12-bit immediate operands 2772 //===------------------------------- 2773 2774 multiclass offsets_uimm12<int MemSize, string prefix> { 2775 def uimm12_asmoperand : AsmOperandClass { 2776 let Name = "OffsetUImm12_" # MemSize; 2777 let PredicateMethod = "isOffsetUImm12<" # MemSize # ">"; 2778 let RenderMethod = "addOffsetUImm12Operands<" # MemSize # ">"; 2779 let DiagnosticType = "LoadStoreUImm12_" # MemSize; 2780 } 2781 2782 // Pattern is really no more than an ImmLeaf, but predicated on MemSize which 2783 // complicates things beyond TableGen's ken. 2784 def uimm12 : Operand<i64>, 2785 ComplexPattern<i64, 1, "SelectOffsetUImm12<" # MemSize # ">"> { 2786 let ParserMatchClass 2787 = !cast<AsmOperandClass>(prefix # uimm12_asmoperand); 2788 2789 let PrintMethod = "printOffsetUImm12Operand<" # MemSize # ">"; 2790 let EncoderMethod = "getOffsetUImm12OpValue<" # MemSize # ">"; 2791 } 2792 } 2793 2794 defm byte_ : offsets_uimm12<1, "byte_">; 2795 defm hword_ : offsets_uimm12<2, "hword_">; 2796 defm word_ : offsets_uimm12<4, "word_">; 2797 defm dword_ : offsets_uimm12<8, "dword_">; 2798 defm qword_ : offsets_uimm12<16, "qword_">; 2799 2800 //===------------------------------- 2801 // 1.1 Signed 9-bit immediate operands 2802 //===------------------------------- 2803 2804 // The MCInst is expected to store the bit-wise encoding of the value, 2805 // which amounts to lopping off the extended sign bits. 2806 def SDXF_simm9 : SDNodeXForm<imm, [{ 2807 return CurDAG->getTargetConstant(N->getZExtValue() & 0x1ff, MVT::i32); 2808 }]>; 2809 2810 def simm9_asmoperand : AsmOperandClass { 2811 let Name = "SImm9"; 2812 let PredicateMethod = "isSImm<9>"; 2813 let RenderMethod = "addSImmOperands<9>"; 2814 let DiagnosticType = "LoadStoreSImm9"; 2815 } 2816 2817 def simm9 : Operand<i64>, 2818 ImmLeaf<i64, [{ return Imm >= -0x100 && Imm <= 0xff; }], 2819 SDXF_simm9> { 2820 let PrintMethod = "printOffsetSImm9Operand"; 2821 let ParserMatchClass = simm9_asmoperand; 2822 } 2823 2824 2825 //===------------------------------- 2826 // 1.3 Register offset extensions 2827 //===------------------------------- 2828 2829 // The assembly-syntax for these addressing-modes is: 2830 // [<Xn|SP>, <R><m> {, <extend> {<amount>}}] 2831 // 2832 // The essential semantics are: 2833 // + <amount> is a shift: #<log(transfer size)> or #0 2834 // + <R> can be W or X. 2835 // + If <R> is W, <extend> can be UXTW or SXTW 2836 // + If <R> is X, <extend> can be LSL or SXTX 2837 // 2838 // The trickiest of those constraints is that Rm can be either GPR32 or GPR64, 2839 // which will need separate instructions for LLVM type-consistency. We'll also 2840 // need separate operands, of course. 2841 multiclass regexts<int MemSize, int RmSize, RegisterClass GPR, 2842 string Rm, string prefix> { 2843 def regext_asmoperand : AsmOperandClass { 2844 let Name = "AddrRegExtend_" # MemSize # "_" # Rm; 2845 let PredicateMethod = "isAddrRegExtend<" # MemSize # "," # RmSize # ">"; 2846 let RenderMethod = "addAddrRegExtendOperands<" # MemSize # ">"; 2847 let DiagnosticType = "LoadStoreExtend" # RmSize # "_" # MemSize; 2848 } 2849 2850 def regext : Operand<i64> { 2851 let PrintMethod 2852 = "printAddrRegExtendOperand<" # MemSize # ", " # RmSize # ">"; 2853 2854 let DecoderMethod = "DecodeAddrRegExtendOperand"; 2855 let ParserMatchClass 2856 = !cast<AsmOperandClass>(prefix # regext_asmoperand); 2857 } 2858 } 2859 2860 multiclass regexts_wx<int MemSize, string prefix> { 2861 // Rm is an X-register if LSL or SXTX are specified as the shift. 2862 defm Xm_ : regexts<MemSize, 64, GPR64, "Xm", prefix # "Xm_">; 2863 2864 // Rm is a W-register if UXTW or SXTW are specified as the shift. 2865 defm Wm_ : regexts<MemSize, 32, GPR32, "Wm", prefix # "Wm_">; 2866 } 2867 2868 defm byte_ : regexts_wx<1, "byte_">; 2869 defm hword_ : regexts_wx<2, "hword_">; 2870 defm word_ : regexts_wx<4, "word_">; 2871 defm dword_ : regexts_wx<8, "dword_">; 2872 defm qword_ : regexts_wx<16, "qword_">; 2873 2874 2875 //===------------------------------ 2876 // 2. The instructions themselves. 2877 //===------------------------------ 2878 2879 // We have the following instructions to implement: 2880 // | | B | H | W | X | 2881 // |-----------------+-------+-------+-------+--------| 2882 // | unsigned str | STRB | STRH | STR | STR | 2883 // | unsigned ldr | LDRB | LDRH | LDR | LDR | 2884 // | signed ldr to W | LDRSB | LDRSH | - | - | 2885 // | signed ldr to X | LDRSB | LDRSH | LDRSW | (PRFM) | 2886 2887 // This will instantiate the LDR/STR instructions you'd expect to use for an 2888 // unsigned datatype (first two rows above) or floating-point register, which is 2889 // reasonably uniform across all access sizes. 2890 2891 2892 //===------------------------------ 2893 // 2.1 Regular instructions 2894 //===------------------------------ 2895 2896 // This class covers the basic unsigned or irrelevantly-signed loads and stores, 2897 // to general-purpose and floating-point registers. 2898 2899 class AddrParams<string prefix> { 2900 Operand uimm12 = !cast<Operand>(prefix # "_uimm12"); 2901 2902 Operand regextWm = !cast<Operand>(prefix # "_Wm_regext"); 2903 Operand regextXm = !cast<Operand>(prefix # "_Xm_regext"); 2904 } 2905 2906 def byte_addrparams : AddrParams<"byte">; 2907 def hword_addrparams : AddrParams<"hword">; 2908 def word_addrparams : AddrParams<"word">; 2909 def dword_addrparams : AddrParams<"dword">; 2910 def qword_addrparams : AddrParams<"qword">; 2911 2912 multiclass A64I_LDRSTR_unsigned<string prefix, bits<2> size, bit v, 2913 bit high_opc, string asmsuffix, 2914 RegisterClass GPR, AddrParams params> { 2915 // Unsigned immediate 2916 def _STR : A64I_LSunsigimm<size, v, {high_opc, 0b0}, 2917 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, params.uimm12:$UImm12), 2918 "str" # asmsuffix # "\t$Rt, [$Rn, $UImm12]", 2919 [], NoItinerary> { 2920 let mayStore = 1; 2921 } 2922 def : InstAlias<"str" # asmsuffix # " $Rt, [$Rn]", 2923 (!cast<Instruction>(prefix # "_STR") GPR:$Rt, GPR64xsp:$Rn, 0)>; 2924 2925 def _LDR : A64I_LSunsigimm<size, v, {high_opc, 0b1}, 2926 (outs GPR:$Rt), (ins GPR64xsp:$Rn, params.uimm12:$UImm12), 2927 "ldr" # asmsuffix # "\t$Rt, [$Rn, $UImm12]", 2928 [], NoItinerary> { 2929 let mayLoad = 1; 2930 } 2931 def : InstAlias<"ldr" # asmsuffix # " $Rt, [$Rn]", 2932 (!cast<Instruction>(prefix # "_LDR") GPR:$Rt, GPR64xsp:$Rn, 0)>; 2933 2934 // Register offset (four of these: load/store and Wm/Xm). 2935 let mayLoad = 1 in { 2936 def _Wm_RegOffset_LDR : A64I_LSregoff<size, v, {high_opc, 0b1}, 0b0, 2937 (outs GPR:$Rt), 2938 (ins GPR64xsp:$Rn, GPR32:$Rm, params.regextWm:$Ext), 2939 "ldr" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]", 2940 [], NoItinerary>; 2941 2942 def _Xm_RegOffset_LDR : A64I_LSregoff<size, v, {high_opc, 0b1}, 0b1, 2943 (outs GPR:$Rt), 2944 (ins GPR64xsp:$Rn, GPR64:$Rm, params.regextXm:$Ext), 2945 "ldr" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]", 2946 [], NoItinerary>; 2947 } 2948 def : InstAlias<"ldr" # asmsuffix # " $Rt, [$Rn, $Rm]", 2949 (!cast<Instruction>(prefix # "_Xm_RegOffset_LDR") GPR:$Rt, GPR64xsp:$Rn, 2950 GPR64:$Rm, 2)>; 2951 2952 let mayStore = 1 in { 2953 def _Wm_RegOffset_STR : A64I_LSregoff<size, v, {high_opc, 0b0}, 0b0, 2954 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, GPR32:$Rm, 2955 params.regextWm:$Ext), 2956 "str" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]", 2957 [], NoItinerary>; 2958 2959 def _Xm_RegOffset_STR : A64I_LSregoff<size, v, {high_opc, 0b0}, 0b1, 2960 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, GPR64:$Rm, 2961 params.regextXm:$Ext), 2962 "str" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]", 2963 [], NoItinerary>; 2964 } 2965 def : InstAlias<"str" # asmsuffix # " $Rt, [$Rn, $Rm]", 2966 (!cast<Instruction>(prefix # "_Xm_RegOffset_STR") GPR:$Rt, GPR64xsp:$Rn, 2967 GPR64:$Rm, 2)>; 2968 2969 // Unaligned immediate 2970 def _STUR : A64I_LSunalimm<size, v, {high_opc, 0b0}, 2971 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9), 2972 "stur" # asmsuffix # "\t$Rt, [$Rn, $SImm9]", 2973 [], NoItinerary> { 2974 let mayStore = 1; 2975 } 2976 def : InstAlias<"stur" # asmsuffix # " $Rt, [$Rn]", 2977 (!cast<Instruction>(prefix # "_STUR") GPR:$Rt, GPR64xsp:$Rn, 0)>; 2978 2979 def _LDUR : A64I_LSunalimm<size, v, {high_opc, 0b1}, 2980 (outs GPR:$Rt), (ins GPR64xsp:$Rn, simm9:$SImm9), 2981 "ldur" # asmsuffix # "\t$Rt, [$Rn, $SImm9]", 2982 [], NoItinerary> { 2983 let mayLoad = 1; 2984 } 2985 def : InstAlias<"ldur" # asmsuffix # " $Rt, [$Rn]", 2986 (!cast<Instruction>(prefix # "_LDUR") GPR:$Rt, GPR64xsp:$Rn, 0)>; 2987 2988 // Post-indexed 2989 def _PostInd_STR : A64I_LSpostind<size, v, {high_opc, 0b0}, 2990 (outs GPR64xsp:$Rn_wb), 2991 (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9), 2992 "str" # asmsuffix # "\t$Rt, [$Rn], $SImm9", 2993 [], NoItinerary> { 2994 let Constraints = "$Rn = $Rn_wb"; 2995 let mayStore = 1; 2996 2997 // Decoder only needed for unpredictability checking (FIXME). 2998 let DecoderMethod = "DecodeSingleIndexedInstruction"; 2999 } 3000 3001 def _PostInd_LDR : A64I_LSpostind<size, v, {high_opc, 0b1}, 3002 (outs GPR:$Rt, GPR64xsp:$Rn_wb), 3003 (ins GPR64xsp:$Rn, simm9:$SImm9), 3004 "ldr" # asmsuffix # "\t$Rt, [$Rn], $SImm9", 3005 [], NoItinerary> { 3006 let mayLoad = 1; 3007 let Constraints = "$Rn = $Rn_wb"; 3008 let DecoderMethod = "DecodeSingleIndexedInstruction"; 3009 } 3010 3011 // Pre-indexed 3012 def _PreInd_STR : A64I_LSpreind<size, v, {high_opc, 0b0}, 3013 (outs GPR64xsp:$Rn_wb), 3014 (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9), 3015 "str" # asmsuffix # "\t$Rt, [$Rn, $SImm9]!", 3016 [], NoItinerary> { 3017 let Constraints = "$Rn = $Rn_wb"; 3018 let mayStore = 1; 3019 3020 // Decoder only needed for unpredictability checking (FIXME). 3021 let DecoderMethod = "DecodeSingleIndexedInstruction"; 3022 } 3023 3024 def _PreInd_LDR : A64I_LSpreind<size, v, {high_opc, 0b1}, 3025 (outs GPR:$Rt, GPR64xsp:$Rn_wb), 3026 (ins GPR64xsp:$Rn, simm9:$SImm9), 3027 "ldr" # asmsuffix # "\t$Rt, [$Rn, $SImm9]!", 3028 [], NoItinerary> { 3029 let mayLoad = 1; 3030 let Constraints = "$Rn = $Rn_wb"; 3031 let DecoderMethod = "DecodeSingleIndexedInstruction"; 3032 } 3033 3034 } 3035 3036 // STRB/LDRB: First define the instructions 3037 defm LS8 3038 : A64I_LDRSTR_unsigned<"LS8", 0b00, 0b0, 0b0, "b", GPR32, byte_addrparams>; 3039 3040 // STRH/LDRH 3041 defm LS16 3042 : A64I_LDRSTR_unsigned<"LS16", 0b01, 0b0, 0b0, "h", GPR32, hword_addrparams>; 3043 3044 3045 // STR/LDR to/from a W register 3046 defm LS32 3047 : A64I_LDRSTR_unsigned<"LS32", 0b10, 0b0, 0b0, "", GPR32, word_addrparams>; 3048 3049 // STR/LDR to/from an X register 3050 defm LS64 3051 : A64I_LDRSTR_unsigned<"LS64", 0b11, 0b0, 0b0, "", GPR64, dword_addrparams>; 3052 3053 // STR/LDR to/from a B register 3054 defm LSFP8 3055 : A64I_LDRSTR_unsigned<"LSFP8", 0b00, 0b1, 0b0, "", FPR8, byte_addrparams>; 3056 3057 // STR/LDR to/from an H register 3058 defm LSFP16 3059 : A64I_LDRSTR_unsigned<"LSFP16", 0b01, 0b1, 0b0, "", FPR16, hword_addrparams>; 3060 3061 // STR/LDR to/from an S register 3062 defm LSFP32 3063 : A64I_LDRSTR_unsigned<"LSFP32", 0b10, 0b1, 0b0, "", FPR32, word_addrparams>; 3064 // STR/LDR to/from a D register 3065 defm LSFP64 3066 : A64I_LDRSTR_unsigned<"LSFP64", 0b11, 0b1, 0b0, "", FPR64, dword_addrparams>; 3067 // STR/LDR to/from a Q register 3068 defm LSFP128 3069 : A64I_LDRSTR_unsigned<"LSFP128", 0b00, 0b1, 0b1, "", FPR128, 3070 qword_addrparams>; 3071 3072 //===------------------------------ 3073 // 2.3 Signed loads 3074 //===------------------------------ 3075 3076 // Byte and half-word signed loads can both go into either an X or a W register, 3077 // so it's worth factoring out. Signed word loads don't fit because there is no 3078 // W version. 3079 multiclass A64I_LDR_signed<bits<2> size, string asmopcode, AddrParams params, 3080 string prefix> { 3081 // Unsigned offset 3082 def w : A64I_LSunsigimm<size, 0b0, 0b11, 3083 (outs GPR32:$Rt), 3084 (ins GPR64xsp:$Rn, params.uimm12:$UImm12), 3085 "ldrs" # asmopcode # "\t$Rt, [$Rn, $UImm12]", 3086 [], NoItinerary> { 3087 let mayLoad = 1; 3088 } 3089 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn]", 3090 (!cast<Instruction>(prefix # w) GPR32:$Rt, GPR64xsp:$Rn, 0)>; 3091 3092 def x : A64I_LSunsigimm<size, 0b0, 0b10, 3093 (outs GPR64:$Rt), 3094 (ins GPR64xsp:$Rn, params.uimm12:$UImm12), 3095 "ldrs" # asmopcode # "\t$Rt, [$Rn, $UImm12]", 3096 [], NoItinerary> { 3097 let mayLoad = 1; 3098 } 3099 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn]", 3100 (!cast<Instruction>(prefix # x) GPR64:$Rt, GPR64xsp:$Rn, 0)>; 3101 3102 // Register offset 3103 let mayLoad = 1 in { 3104 def w_Wm_RegOffset : A64I_LSregoff<size, 0b0, 0b11, 0b0, 3105 (outs GPR32:$Rt), 3106 (ins GPR64xsp:$Rn, GPR32:$Rm, params.regextWm:$Ext), 3107 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]", 3108 [], NoItinerary>; 3109 3110 def w_Xm_RegOffset : A64I_LSregoff<size, 0b0, 0b11, 0b1, 3111 (outs GPR32:$Rt), 3112 (ins GPR64xsp:$Rn, GPR64:$Rm, params.regextXm:$Ext), 3113 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]", 3114 [], NoItinerary>; 3115 3116 def x_Wm_RegOffset : A64I_LSregoff<size, 0b0, 0b10, 0b0, 3117 (outs GPR64:$Rt), 3118 (ins GPR64xsp:$Rn, GPR32:$Rm, params.regextWm:$Ext), 3119 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]", 3120 [], NoItinerary>; 3121 3122 def x_Xm_RegOffset : A64I_LSregoff<size, 0b0, 0b10, 0b1, 3123 (outs GPR64:$Rt), 3124 (ins GPR64xsp:$Rn, GPR64:$Rm, params.regextXm:$Ext), 3125 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]", 3126 [], NoItinerary>; 3127 } 3128 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn, $Rm]", 3129 (!cast<Instruction>(prefix # "w_Xm_RegOffset") GPR32:$Rt, GPR64xsp:$Rn, 3130 GPR64:$Rm, 2)>; 3131 3132 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn, $Rm]", 3133 (!cast<Instruction>(prefix # "x_Xm_RegOffset") GPR64:$Rt, GPR64xsp:$Rn, 3134 GPR64:$Rm, 2)>; 3135 3136 3137 let mayLoad = 1 in { 3138 // Unaligned offset 3139 def w_U : A64I_LSunalimm<size, 0b0, 0b11, 3140 (outs GPR32:$Rt), 3141 (ins GPR64xsp:$Rn, simm9:$SImm9), 3142 "ldurs" # asmopcode # "\t$Rt, [$Rn, $SImm9]", 3143 [], NoItinerary>; 3144 3145 def x_U : A64I_LSunalimm<size, 0b0, 0b10, 3146 (outs GPR64:$Rt), 3147 (ins GPR64xsp:$Rn, simm9:$SImm9), 3148 "ldurs" # asmopcode # "\t$Rt, [$Rn, $SImm9]", 3149 [], NoItinerary>; 3150 3151 3152 // Post-indexed 3153 def w_PostInd : A64I_LSpostind<size, 0b0, 0b11, 3154 (outs GPR32:$Rt, GPR64xsp:$Rn_wb), 3155 (ins GPR64xsp:$Rn, simm9:$SImm9), 3156 "ldrs" # asmopcode # "\t$Rt, [$Rn], $SImm9", 3157 [], NoItinerary> { 3158 let Constraints = "$Rn = $Rn_wb"; 3159 let DecoderMethod = "DecodeSingleIndexedInstruction"; 3160 } 3161 3162 def x_PostInd : A64I_LSpostind<size, 0b0, 0b10, 3163 (outs GPR64:$Rt, GPR64xsp:$Rn_wb), 3164 (ins GPR64xsp:$Rn, simm9:$SImm9), 3165 "ldrs" # asmopcode # "\t$Rt, [$Rn], $SImm9", 3166 [], NoItinerary> { 3167 let Constraints = "$Rn = $Rn_wb"; 3168 let DecoderMethod = "DecodeSingleIndexedInstruction"; 3169 } 3170 3171 // Pre-indexed 3172 def w_PreInd : A64I_LSpreind<size, 0b0, 0b11, 3173 (outs GPR32:$Rt, GPR64xsp:$Rn_wb), 3174 (ins GPR64xsp:$Rn, simm9:$SImm9), 3175 "ldrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]!", 3176 [], NoItinerary> { 3177 let Constraints = "$Rn = $Rn_wb"; 3178 let DecoderMethod = "DecodeSingleIndexedInstruction"; 3179 } 3180 3181 def x_PreInd : A64I_LSpreind<size, 0b0, 0b10, 3182 (outs GPR64:$Rt, GPR64xsp:$Rn_wb), 3183 (ins GPR64xsp:$Rn, simm9:$SImm9), 3184 "ldrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]!", 3185 [], NoItinerary> { 3186 let Constraints = "$Rn = $Rn_wb"; 3187 let DecoderMethod = "DecodeSingleIndexedInstruction"; 3188 } 3189 } // let mayLoad = 1 3190 } 3191 3192 // LDRSB 3193 defm LDRSB : A64I_LDR_signed<0b00, "b", byte_addrparams, "LDRSB">; 3194 // LDRSH 3195 defm LDRSH : A64I_LDR_signed<0b01, "h", hword_addrparams, "LDRSH">; 3196 3197 // LDRSW: load a 32-bit register, sign-extending to 64-bits. 3198 def LDRSWx 3199 : A64I_LSunsigimm<0b10, 0b0, 0b10, 3200 (outs GPR64:$Rt), 3201 (ins GPR64xsp:$Rn, word_uimm12:$UImm12), 3202 "ldrsw\t$Rt, [$Rn, $UImm12]", 3203 [], NoItinerary> { 3204 let mayLoad = 1; 3205 } 3206 def : InstAlias<"ldrsw $Rt, [$Rn]", (LDRSWx GPR64:$Rt, GPR64xsp:$Rn, 0)>; 3207 3208 let mayLoad = 1 in { 3209 def LDRSWx_Wm_RegOffset : A64I_LSregoff<0b10, 0b0, 0b10, 0b0, 3210 (outs GPR64:$Rt), 3211 (ins GPR64xsp:$Rn, GPR32:$Rm, word_Wm_regext:$Ext), 3212 "ldrsw\t$Rt, [$Rn, $Rm, $Ext]", 3213 [], NoItinerary>; 3214 3215 def LDRSWx_Xm_RegOffset : A64I_LSregoff<0b10, 0b0, 0b10, 0b1, 3216 (outs GPR64:$Rt), 3217 (ins GPR64xsp:$Rn, GPR64:$Rm, word_Xm_regext:$Ext), 3218 "ldrsw\t$Rt, [$Rn, $Rm, $Ext]", 3219 [], NoItinerary>; 3220 } 3221 def : InstAlias<"ldrsw $Rt, [$Rn, $Rm]", 3222 (LDRSWx_Xm_RegOffset GPR64:$Rt, GPR64xsp:$Rn, GPR64:$Rm, 2)>; 3223 3224 3225 def LDURSWx 3226 : A64I_LSunalimm<0b10, 0b0, 0b10, 3227 (outs GPR64:$Rt), 3228 (ins GPR64xsp:$Rn, simm9:$SImm9), 3229 "ldursw\t$Rt, [$Rn, $SImm9]", 3230 [], NoItinerary> { 3231 let mayLoad = 1; 3232 } 3233 def : InstAlias<"ldursw $Rt, [$Rn]", (LDURSWx GPR64:$Rt, GPR64xsp:$Rn, 0)>; 3234 3235 def LDRSWx_PostInd 3236 : A64I_LSpostind<0b10, 0b0, 0b10, 3237 (outs GPR64:$Rt, GPR64xsp:$Rn_wb), 3238 (ins GPR64xsp:$Rn, simm9:$SImm9), 3239 "ldrsw\t$Rt, [$Rn], $SImm9", 3240 [], NoItinerary> { 3241 let mayLoad = 1; 3242 let Constraints = "$Rn = $Rn_wb"; 3243 let DecoderMethod = "DecodeSingleIndexedInstruction"; 3244 } 3245 3246 def LDRSWx_PreInd : A64I_LSpreind<0b10, 0b0, 0b10, 3247 (outs GPR64:$Rt, GPR64xsp:$Rn_wb), 3248 (ins GPR64xsp:$Rn, simm9:$SImm9), 3249 "ldrsw\t$Rt, [$Rn, $SImm9]!", 3250 [], NoItinerary> { 3251 let mayLoad = 1; 3252 let Constraints = "$Rn = $Rn_wb"; 3253 let DecoderMethod = "DecodeSingleIndexedInstruction"; 3254 } 3255 3256 //===------------------------------ 3257 // 2.4 Prefetch operations 3258 //===------------------------------ 3259 3260 def PRFM : A64I_LSunsigimm<0b11, 0b0, 0b10, (outs), 3261 (ins prefetch_op:$Rt, GPR64xsp:$Rn, dword_uimm12:$UImm12), 3262 "prfm\t$Rt, [$Rn, $UImm12]", 3263 [], NoItinerary> { 3264 let mayLoad = 1; 3265 } 3266 def : InstAlias<"prfm $Rt, [$Rn]", 3267 (PRFM prefetch_op:$Rt, GPR64xsp:$Rn, 0)>; 3268 3269 let mayLoad = 1 in { 3270 def PRFM_Wm_RegOffset : A64I_LSregoff<0b11, 0b0, 0b10, 0b0, (outs), 3271 (ins prefetch_op:$Rt, GPR64xsp:$Rn, 3272 GPR32:$Rm, dword_Wm_regext:$Ext), 3273 "prfm\t$Rt, [$Rn, $Rm, $Ext]", 3274 [], NoItinerary>; 3275 def PRFM_Xm_RegOffset : A64I_LSregoff<0b11, 0b0, 0b10, 0b1, (outs), 3276 (ins prefetch_op:$Rt, GPR64xsp:$Rn, 3277 GPR64:$Rm, dword_Xm_regext:$Ext), 3278 "prfm\t$Rt, [$Rn, $Rm, $Ext]", 3279 [], NoItinerary>; 3280 } 3281 3282 def : InstAlias<"prfm $Rt, [$Rn, $Rm]", 3283 (PRFM_Xm_RegOffset prefetch_op:$Rt, GPR64xsp:$Rn, 3284 GPR64:$Rm, 2)>; 3285 3286 3287 def PRFUM : A64I_LSunalimm<0b11, 0b0, 0b10, (outs), 3288 (ins prefetch_op:$Rt, GPR64xsp:$Rn, simm9:$SImm9), 3289 "prfum\t$Rt, [$Rn, $SImm9]", 3290 [], NoItinerary> { 3291 let mayLoad = 1; 3292 } 3293 def : InstAlias<"prfum $Rt, [$Rn]", 3294 (PRFUM prefetch_op:$Rt, GPR64xsp:$Rn, 0)>; 3295 3296 //===----------------------------------------------------------------------===// 3297 // Load-store register (unprivileged) instructions 3298 //===----------------------------------------------------------------------===// 3299 // Contains: LDTRB, LDTRH, LDTRSB, LDTRSH, LDTRSW, STTR, STTRB and STTRH 3300 3301 // These instructions very much mirror the "unscaled immediate" loads, but since 3302 // there are no floating-point variants we need to split them out into their own 3303 // section to avoid instantiation of "ldtr d0, [sp]" etc. 3304 3305 multiclass A64I_LDTRSTTR<bits<2> size, string asmsuffix, RegisterClass GPR, 3306 string prefix> { 3307 def _UnPriv_STR : A64I_LSunpriv<size, 0b0, 0b00, 3308 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9), 3309 "sttr" # asmsuffix # "\t$Rt, [$Rn, $SImm9]", 3310 [], NoItinerary> { 3311 let mayStore = 1; 3312 } 3313 3314 def : InstAlias<"sttr" # asmsuffix # " $Rt, [$Rn]", 3315 (!cast<Instruction>(prefix # "_UnPriv_STR") GPR:$Rt, GPR64xsp:$Rn, 0)>; 3316 3317 def _UnPriv_LDR : A64I_LSunpriv<size, 0b0, 0b01, 3318 (outs GPR:$Rt), (ins GPR64xsp:$Rn, simm9:$SImm9), 3319 "ldtr" # asmsuffix # "\t$Rt, [$Rn, $SImm9]", 3320 [], NoItinerary> { 3321 let mayLoad = 1; 3322 } 3323 3324 def : InstAlias<"ldtr" # asmsuffix # " $Rt, [$Rn]", 3325 (!cast<Instruction>(prefix # "_UnPriv_LDR") GPR:$Rt, GPR64xsp:$Rn, 0)>; 3326 3327 } 3328 3329 // STTRB/LDTRB: First define the instructions 3330 defm LS8 : A64I_LDTRSTTR<0b00, "b", GPR32, "LS8">; 3331 3332 // STTRH/LDTRH 3333 defm LS16 : A64I_LDTRSTTR<0b01, "h", GPR32, "LS16">; 3334 3335 // STTR/LDTR to/from a W register 3336 defm LS32 : A64I_LDTRSTTR<0b10, "", GPR32, "LS32">; 3337 3338 // STTR/LDTR to/from an X register 3339 defm LS64 : A64I_LDTRSTTR<0b11, "", GPR64, "LS64">; 3340 3341 // Now a class for the signed instructions that can go to either 32 or 64 3342 // bits... 3343 multiclass A64I_LDTR_signed<bits<2> size, string asmopcode, string prefix> { 3344 let mayLoad = 1 in { 3345 def w : A64I_LSunpriv<size, 0b0, 0b11, 3346 (outs GPR32:$Rt), 3347 (ins GPR64xsp:$Rn, simm9:$SImm9), 3348 "ldtrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]", 3349 [], NoItinerary>; 3350 3351 def x : A64I_LSunpriv<size, 0b0, 0b10, 3352 (outs GPR64:$Rt), 3353 (ins GPR64xsp:$Rn, simm9:$SImm9), 3354 "ldtrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]", 3355 [], NoItinerary>; 3356 } 3357 3358 def : InstAlias<"ldtrs" # asmopcode # " $Rt, [$Rn]", 3359 (!cast<Instruction>(prefix # "w") GPR32:$Rt, GPR64xsp:$Rn, 0)>; 3360 3361 def : InstAlias<"ldtrs" # asmopcode # " $Rt, [$Rn]", 3362 (!cast<Instruction>(prefix # "x") GPR64:$Rt, GPR64xsp:$Rn, 0)>; 3363 3364 } 3365 3366 // LDTRSB 3367 defm LDTRSB : A64I_LDTR_signed<0b00, "b", "LDTRSB">; 3368 // LDTRSH 3369 defm LDTRSH : A64I_LDTR_signed<0b01, "h", "LDTRSH">; 3370 3371 // And finally LDTRSW which only goes to 64 bits. 3372 def LDTRSWx : A64I_LSunpriv<0b10, 0b0, 0b10, 3373 (outs GPR64:$Rt), 3374 (ins GPR64xsp:$Rn, simm9:$SImm9), 3375 "ldtrsw\t$Rt, [$Rn, $SImm9]", 3376 [], NoItinerary> { 3377 let mayLoad = 1; 3378 } 3379 def : InstAlias<"ldtrsw $Rt, [$Rn]", (LDTRSWx GPR64:$Rt, GPR64xsp:$Rn, 0)>; 3380 3381 //===----------------------------------------------------------------------===// 3382 // Load-store register pair (offset) instructions 3383 //===----------------------------------------------------------------------===// 3384 // 3385 // and 3386 // 3387 //===----------------------------------------------------------------------===// 3388 // Load-store register pair (post-indexed) instructions 3389 //===----------------------------------------------------------------------===// 3390 // Contains: STP, LDP, LDPSW 3391 // 3392 // and 3393 // 3394 //===----------------------------------------------------------------------===// 3395 // Load-store register pair (pre-indexed) instructions 3396 //===----------------------------------------------------------------------===// 3397 // Contains: STP, LDP, LDPSW 3398 // 3399 // and 3400 // 3401 //===----------------------------------------------------------------------===// 3402 // Load-store non-temporal register pair (offset) instructions 3403 //===----------------------------------------------------------------------===// 3404 // Contains: STNP, LDNP 3405 3406 3407 // Anything that creates an MCInst (Decoding, selection and AsmParsing) has to 3408 // know the access size via some means. An isolated operand does not have this 3409 // information unless told from here, which means we need separate tablegen 3410 // Operands for each access size. This multiclass takes care of instantiating 3411 // the correct template functions in the rest of the backend. 3412 3413 multiclass offsets_simm7<string MemSize, string prefix> { 3414 // The bare signed 7-bit immediate is used in post-indexed instructions, but 3415 // because of the scaling performed a generic "simm7" operand isn't 3416 // appropriate here either. 3417 def simm7_asmoperand : AsmOperandClass { 3418 let Name = "SImm7_Scaled" # MemSize; 3419 let PredicateMethod = "isSImm7Scaled<" # MemSize # ">"; 3420 let RenderMethod = "addSImm7ScaledOperands<" # MemSize # ">"; 3421 let DiagnosticType = "LoadStoreSImm7_" # MemSize; 3422 } 3423 3424 def simm7 : Operand<i64> { 3425 let PrintMethod = "printSImm7ScaledOperand<" # MemSize # ">"; 3426 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "simm7_asmoperand"); 3427 } 3428 } 3429 3430 defm word_ : offsets_simm7<"4", "word_">; 3431 defm dword_ : offsets_simm7<"8", "dword_">; 3432 defm qword_ : offsets_simm7<"16", "qword_">; 3433 3434 multiclass A64I_LSPsimple<bits<2> opc, bit v, RegisterClass SomeReg, 3435 Operand simm7, string prefix> { 3436 def _STR : A64I_LSPoffset<opc, v, 0b0, (outs), 3437 (ins SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn, simm7:$SImm7), 3438 "stp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> { 3439 let mayStore = 1; 3440 let DecoderMethod = "DecodeLDSTPairInstruction"; 3441 } 3442 def : InstAlias<"stp $Rt, $Rt2, [$Rn]", 3443 (!cast<Instruction>(prefix # "_STR") SomeReg:$Rt, 3444 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>; 3445 3446 def _LDR : A64I_LSPoffset<opc, v, 0b1, 3447 (outs SomeReg:$Rt, SomeReg:$Rt2), 3448 (ins GPR64xsp:$Rn, simm7:$SImm7), 3449 "ldp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> { 3450 let mayLoad = 1; 3451 let DecoderMethod = "DecodeLDSTPairInstruction"; 3452 } 3453 def : InstAlias<"ldp $Rt, $Rt2, [$Rn]", 3454 (!cast<Instruction>(prefix # "_LDR") SomeReg:$Rt, 3455 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>; 3456 3457 def _PostInd_STR : A64I_LSPpostind<opc, v, 0b0, 3458 (outs GPR64xsp:$Rn_wb), 3459 (ins SomeReg:$Rt, SomeReg:$Rt2, 3460 GPR64xsp:$Rn, 3461 simm7:$SImm7), 3462 "stp\t$Rt, $Rt2, [$Rn], $SImm7", 3463 [], NoItinerary> { 3464 let mayStore = 1; 3465 let Constraints = "$Rn = $Rn_wb"; 3466 3467 // Decoder only needed for unpredictability checking (FIXME). 3468 let DecoderMethod = "DecodeLDSTPairInstruction"; 3469 } 3470 3471 def _PostInd_LDR : A64I_LSPpostind<opc, v, 0b1, 3472 (outs SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn_wb), 3473 (ins GPR64xsp:$Rn, simm7:$SImm7), 3474 "ldp\t$Rt, $Rt2, [$Rn], $SImm7", 3475 [], NoItinerary> { 3476 let mayLoad = 1; 3477 let Constraints = "$Rn = $Rn_wb"; 3478 let DecoderMethod = "DecodeLDSTPairInstruction"; 3479 } 3480 3481 def _PreInd_STR : A64I_LSPpreind<opc, v, 0b0, (outs GPR64xsp:$Rn_wb), 3482 (ins SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn, simm7:$SImm7), 3483 "stp\t$Rt, $Rt2, [$Rn, $SImm7]!", 3484 [], NoItinerary> { 3485 let mayStore = 1; 3486 let Constraints = "$Rn = $Rn_wb"; 3487 let DecoderMethod = "DecodeLDSTPairInstruction"; 3488 } 3489 3490 def _PreInd_LDR : A64I_LSPpreind<opc, v, 0b1, 3491 (outs SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn_wb), 3492 (ins GPR64xsp:$Rn, simm7:$SImm7), 3493 "ldp\t$Rt, $Rt2, [$Rn, $SImm7]!", 3494 [], NoItinerary> { 3495 let mayLoad = 1; 3496 let Constraints = "$Rn = $Rn_wb"; 3497 let DecoderMethod = "DecodeLDSTPairInstruction"; 3498 } 3499 3500 def _NonTemp_STR : A64I_LSPnontemp<opc, v, 0b0, (outs), 3501 (ins SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn, simm7:$SImm7), 3502 "stnp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> { 3503 let mayStore = 1; 3504 let DecoderMethod = "DecodeLDSTPairInstruction"; 3505 } 3506 def : InstAlias<"stnp $Rt, $Rt2, [$Rn]", 3507 (!cast<Instruction>(prefix # "_NonTemp_STR") SomeReg:$Rt, 3508 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>; 3509 3510 def _NonTemp_LDR : A64I_LSPnontemp<opc, v, 0b1, 3511 (outs SomeReg:$Rt, SomeReg:$Rt2), 3512 (ins GPR64xsp:$Rn, simm7:$SImm7), 3513 "ldnp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> { 3514 let mayLoad = 1; 3515 let DecoderMethod = "DecodeLDSTPairInstruction"; 3516 } 3517 def : InstAlias<"ldnp $Rt, $Rt2, [$Rn]", 3518 (!cast<Instruction>(prefix # "_NonTemp_LDR") SomeReg:$Rt, 3519 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>; 3520 3521 } 3522 3523 3524 defm LSPair32 : A64I_LSPsimple<0b00, 0b0, GPR32, word_simm7, "LSPair32">; 3525 defm LSPair64 : A64I_LSPsimple<0b10, 0b0, GPR64, dword_simm7, "LSPair64">; 3526 defm LSFPPair32 : A64I_LSPsimple<0b00, 0b1, FPR32, word_simm7, "LSFPPair32">; 3527 defm LSFPPair64 : A64I_LSPsimple<0b01, 0b1, FPR64, dword_simm7, "LSFPPair64">; 3528 defm LSFPPair128 : A64I_LSPsimple<0b10, 0b1, FPR128, qword_simm7, 3529 "LSFPPair128">; 3530 3531 3532 def LDPSWx : A64I_LSPoffset<0b01, 0b0, 0b1, 3533 (outs GPR64:$Rt, GPR64:$Rt2), 3534 (ins GPR64xsp:$Rn, word_simm7:$SImm7), 3535 "ldpsw\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> { 3536 let mayLoad = 1; 3537 let DecoderMethod = "DecodeLDSTPairInstruction"; 3538 } 3539 def : InstAlias<"ldpsw $Rt, $Rt2, [$Rn]", 3540 (LDPSWx GPR64:$Rt, GPR64:$Rt2, GPR64xsp:$Rn, 0)>; 3541 3542 def LDPSWx_PostInd : A64I_LSPpostind<0b01, 0b0, 0b1, 3543 (outs GPR64:$Rt, GPR64:$Rt2, GPR64:$Rn_wb), 3544 (ins GPR64xsp:$Rn, word_simm7:$SImm7), 3545 "ldpsw\t$Rt, $Rt2, [$Rn], $SImm7", 3546 [], NoItinerary> { 3547 let mayLoad = 1; 3548 let Constraints = "$Rn = $Rn_wb"; 3549 let DecoderMethod = "DecodeLDSTPairInstruction"; 3550 } 3551 3552 def LDPSWx_PreInd : A64I_LSPpreind<0b01, 0b0, 0b1, 3553 (outs GPR64:$Rt, GPR64:$Rt2, GPR64:$Rn_wb), 3554 (ins GPR64xsp:$Rn, word_simm7:$SImm7), 3555 "ldpsw\t$Rt, $Rt2, [$Rn, $SImm7]!", 3556 [], NoItinerary> { 3557 let mayLoad = 1; 3558 let Constraints = "$Rn = $Rn_wb"; 3559 let DecoderMethod = "DecodeLDSTPairInstruction"; 3560 } 3561 3562 //===----------------------------------------------------------------------===// 3563 // Logical (immediate) instructions 3564 //===----------------------------------------------------------------------===// 3565 // Contains: AND, ORR, EOR, ANDS, + aliases TST, MOV 3566 3567 multiclass logical_imm_operands<string prefix, string note, 3568 int size, ValueType VT> { 3569 def _asmoperand : AsmOperandClass { 3570 let Name = "LogicalImm" # note # size; 3571 let PredicateMethod = "isLogicalImm" # note # "<" # size # ">"; 3572 let RenderMethod = "addLogicalImmOperands<" # size # ">"; 3573 let DiagnosticType = "LogicalSecondSource"; 3574 } 3575 3576 def _operand 3577 : Operand<VT>, ComplexPattern<VT, 1, "SelectLogicalImm", [imm]> { 3578 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand"); 3579 let PrintMethod = "printLogicalImmOperand<" # size # ">"; 3580 let DecoderMethod = "DecodeLogicalImmOperand<" # size # ">"; 3581 } 3582 } 3583 3584 defm logical_imm32 : logical_imm_operands<"logical_imm32", "", 32, i32>; 3585 defm logical_imm64 : logical_imm_operands<"logical_imm64", "", 64, i64>; 3586 3587 // The mov versions only differ in assembly parsing, where they 3588 // exclude values representable with either MOVZ or MOVN. 3589 defm logical_imm32_mov 3590 : logical_imm_operands<"logical_imm32_mov", "MOV", 32, i32>; 3591 defm logical_imm64_mov 3592 : logical_imm_operands<"logical_imm64_mov", "MOV", 64, i64>; 3593 3594 3595 multiclass A64I_logimmSizes<bits<2> opc, string asmop, SDNode opnode> { 3596 def wwi : A64I_logicalimm<0b0, opc, (outs GPR32wsp:$Rd), 3597 (ins GPR32:$Rn, logical_imm32_operand:$Imm), 3598 !strconcat(asmop, "\t$Rd, $Rn, $Imm"), 3599 [(set GPR32wsp:$Rd, 3600 (opnode GPR32:$Rn, logical_imm32_operand:$Imm))], 3601 NoItinerary>; 3602 3603 def xxi : A64I_logicalimm<0b1, opc, (outs GPR64xsp:$Rd), 3604 (ins GPR64:$Rn, logical_imm64_operand:$Imm), 3605 !strconcat(asmop, "\t$Rd, $Rn, $Imm"), 3606 [(set GPR64xsp:$Rd, 3607 (opnode GPR64:$Rn, logical_imm64_operand:$Imm))], 3608 NoItinerary>; 3609 } 3610 3611 defm AND : A64I_logimmSizes<0b00, "and", and>; 3612 defm ORR : A64I_logimmSizes<0b01, "orr", or>; 3613 defm EOR : A64I_logimmSizes<0b10, "eor", xor>; 3614 3615 let Defs = [NZCV] in { 3616 def ANDSwwi : A64I_logicalimm<0b0, 0b11, (outs GPR32:$Rd), 3617 (ins GPR32:$Rn, logical_imm32_operand:$Imm), 3618 "ands\t$Rd, $Rn, $Imm", 3619 [], NoItinerary>; 3620 3621 def ANDSxxi : A64I_logicalimm<0b1, 0b11, (outs GPR64:$Rd), 3622 (ins GPR64:$Rn, logical_imm64_operand:$Imm), 3623 "ands\t$Rd, $Rn, $Imm", 3624 [], NoItinerary>; 3625 } 3626 3627 3628 def : InstAlias<"tst $Rn, $Imm", 3629 (ANDSwwi WZR, GPR32:$Rn, logical_imm32_operand:$Imm)>; 3630 def : InstAlias<"tst $Rn, $Imm", 3631 (ANDSxxi XZR, GPR64:$Rn, logical_imm64_operand:$Imm)>; 3632 def : InstAlias<"mov $Rd, $Imm", 3633 (ORRwwi GPR32wsp:$Rd, WZR, logical_imm32_mov_operand:$Imm)>; 3634 def : InstAlias<"mov $Rd, $Imm", 3635 (ORRxxi GPR64xsp:$Rd, XZR, logical_imm64_mov_operand:$Imm)>; 3636 3637 //===----------------------------------------------------------------------===// 3638 // Logical (shifted register) instructions 3639 //===----------------------------------------------------------------------===// 3640 // Contains: AND, BIC, ORR, ORN, EOR, EON, ANDS, BICS + aliases TST, MVN, MOV 3641 3642 // Operand for optimizing (icmp (and LHS, RHS), 0, SomeCode). In theory "ANDS" 3643 // behaves differently for unsigned comparisons, so we defensively only allow 3644 // signed or n/a as the operand. In practice "unsigned greater than 0" is "not 3645 // equal to 0" and LLVM gives us this. 3646 def signed_cond : PatLeaf<(cond), [{ 3647 return !isUnsignedIntSetCC(N->get()); 3648 }]>; 3649 3650 3651 // These instructions share their "shift" operands with add/sub (shifted 3652 // register instructions). They are defined there. 3653 3654 // N.b. the commutable parameter is just !N. It will be first against the wall 3655 // when the revolution comes. 3656 multiclass logical_shifts<string prefix, bit sf, bits<2> opc, 3657 bit N, bit commutable, 3658 string asmop, SDPatternOperator opfrag, string sty, 3659 RegisterClass GPR, list<Register> defs> { 3660 let isCommutable = commutable, Defs = defs in { 3661 def _lsl : A64I_logicalshift<sf, opc, 0b00, N, 3662 (outs GPR:$Rd), 3663 (ins GPR:$Rn, GPR:$Rm, 3664 !cast<Operand>("lsl_operand_" # sty):$Imm6), 3665 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"), 3666 [(set GPR:$Rd, (opfrag GPR:$Rn, (shl GPR:$Rm, 3667 !cast<Operand>("lsl_operand_" # sty):$Imm6)) 3668 )], 3669 NoItinerary>; 3670 3671 def _lsr : A64I_logicalshift<sf, opc, 0b01, N, 3672 (outs GPR:$Rd), 3673 (ins GPR:$Rn, GPR:$Rm, 3674 !cast<Operand>("lsr_operand_" # sty):$Imm6), 3675 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"), 3676 [(set GPR:$Rd, (opfrag GPR:$Rn, (srl GPR:$Rm, 3677 !cast<Operand>("lsr_operand_" # sty):$Imm6)) 3678 )], 3679 NoItinerary>; 3680 3681 def _asr : A64I_logicalshift<sf, opc, 0b10, N, 3682 (outs GPR:$Rd), 3683 (ins GPR:$Rn, GPR:$Rm, 3684 !cast<Operand>("asr_operand_" # sty):$Imm6), 3685 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"), 3686 [(set GPR:$Rd, (opfrag GPR:$Rn, (sra GPR:$Rm, 3687 !cast<Operand>("asr_operand_" # sty):$Imm6)) 3688 )], 3689 NoItinerary>; 3690 3691 def _ror : A64I_logicalshift<sf, opc, 0b11, N, 3692 (outs GPR:$Rd), 3693 (ins GPR:$Rn, GPR:$Rm, 3694 !cast<Operand>("ror_operand_" # sty):$Imm6), 3695 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"), 3696 [(set GPR:$Rd, (opfrag GPR:$Rn, (rotr GPR:$Rm, 3697 !cast<Operand>("ror_operand_" # sty):$Imm6)) 3698 )], 3699 NoItinerary>; 3700 } 3701 3702 def _noshift 3703 : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm"), 3704 (!cast<Instruction>(prefix # "_lsl") GPR:$Rd, GPR:$Rn, 3705 GPR:$Rm, 0)>; 3706 3707 def : Pat<(opfrag GPR:$Rn, GPR:$Rm), 3708 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>; 3709 } 3710 3711 multiclass logical_sizes<string prefix, bits<2> opc, bit N, bit commutable, 3712 string asmop, SDPatternOperator opfrag, 3713 list<Register> defs> { 3714 defm xxx : logical_shifts<prefix # "xxx", 0b1, opc, N, 3715 commutable, asmop, opfrag, "i64", GPR64, defs>; 3716 defm www : logical_shifts<prefix # "www", 0b0, opc, N, 3717 commutable, asmop, opfrag, "i32", GPR32, defs>; 3718 } 3719 3720 3721 defm AND : logical_sizes<"AND", 0b00, 0b0, 0b1, "and", and, []>; 3722 defm ORR : logical_sizes<"ORR", 0b01, 0b0, 0b1, "orr", or, []>; 3723 defm EOR : logical_sizes<"EOR", 0b10, 0b0, 0b1, "eor", xor, []>; 3724 defm ANDS : logical_sizes<"ANDS", 0b11, 0b0, 0b1, "ands", 3725 PatFrag<(ops node:$lhs, node:$rhs), (and node:$lhs, node:$rhs), 3726 [{ (void)N; return false; }]>, 3727 [NZCV]>; 3728 3729 defm BIC : logical_sizes<"BIC", 0b00, 0b1, 0b0, "bic", 3730 PatFrag<(ops node:$lhs, node:$rhs), 3731 (and node:$lhs, (not node:$rhs))>, []>; 3732 defm ORN : logical_sizes<"ORN", 0b01, 0b1, 0b0, "orn", 3733 PatFrag<(ops node:$lhs, node:$rhs), 3734 (or node:$lhs, (not node:$rhs))>, []>; 3735 defm EON : logical_sizes<"EON", 0b10, 0b1, 0b0, "eon", 3736 PatFrag<(ops node:$lhs, node:$rhs), 3737 (xor node:$lhs, (not node:$rhs))>, []>; 3738 defm BICS : logical_sizes<"BICS", 0b11, 0b1, 0b0, "bics", 3739 PatFrag<(ops node:$lhs, node:$rhs), 3740 (and node:$lhs, (not node:$rhs)), 3741 [{ (void)N; return false; }]>, 3742 [NZCV]>; 3743 3744 multiclass tst_shifts<string prefix, bit sf, string sty, RegisterClass GPR> { 3745 let isCommutable = 1, Rd = 0b11111, Defs = [NZCV] in { 3746 def _lsl : A64I_logicalshift<sf, 0b11, 0b00, 0b0, 3747 (outs), 3748 (ins GPR:$Rn, GPR:$Rm, 3749 !cast<Operand>("lsl_operand_" # sty):$Imm6), 3750 "tst\t$Rn, $Rm, $Imm6", 3751 [(set NZCV, (A64setcc (and GPR:$Rn, (shl GPR:$Rm, 3752 !cast<Operand>("lsl_operand_" # sty):$Imm6)), 3753 0, signed_cond))], 3754 NoItinerary>; 3755 3756 3757 def _lsr : A64I_logicalshift<sf, 0b11, 0b01, 0b0, 3758 (outs), 3759 (ins GPR:$Rn, GPR:$Rm, 3760 !cast<Operand>("lsr_operand_" # sty):$Imm6), 3761 "tst\t$Rn, $Rm, $Imm6", 3762 [(set NZCV, (A64setcc (and GPR:$Rn, (srl GPR:$Rm, 3763 !cast<Operand>("lsr_operand_" # sty):$Imm6)), 3764 0, signed_cond))], 3765 NoItinerary>; 3766 3767 def _asr : A64I_logicalshift<sf, 0b11, 0b10, 0b0, 3768 (outs), 3769 (ins GPR:$Rn, GPR:$Rm, 3770 !cast<Operand>("asr_operand_" # sty):$Imm6), 3771 "tst\t$Rn, $Rm, $Imm6", 3772 [(set NZCV, (A64setcc (and GPR:$Rn, (sra GPR:$Rm, 3773 !cast<Operand>("asr_operand_" # sty):$Imm6)), 3774 0, signed_cond))], 3775 NoItinerary>; 3776 3777 def _ror : A64I_logicalshift<sf, 0b11, 0b11, 0b0, 3778 (outs), 3779 (ins GPR:$Rn, GPR:$Rm, 3780 !cast<Operand>("ror_operand_" # sty):$Imm6), 3781 "tst\t$Rn, $Rm, $Imm6", 3782 [(set NZCV, (A64setcc (and GPR:$Rn, (rotr GPR:$Rm, 3783 !cast<Operand>("ror_operand_" # sty):$Imm6)), 3784 0, signed_cond))], 3785 NoItinerary>; 3786 } 3787 3788 def _noshift : InstAlias<"tst $Rn, $Rm", 3789 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>; 3790 3791 def : Pat<(A64setcc (and GPR:$Rn, GPR:$Rm), 0, signed_cond), 3792 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>; 3793 } 3794 3795 defm TSTxx : tst_shifts<"TSTxx", 0b1, "i64", GPR64>; 3796 defm TSTww : tst_shifts<"TSTww", 0b0, "i32", GPR32>; 3797 3798 3799 multiclass mvn_shifts<string prefix, bit sf, string sty, RegisterClass GPR> { 3800 let isCommutable = 0, Rn = 0b11111 in { 3801 def _lsl : A64I_logicalshift<sf, 0b01, 0b00, 0b1, 3802 (outs GPR:$Rd), 3803 (ins GPR:$Rm, 3804 !cast<Operand>("lsl_operand_" # sty):$Imm6), 3805 "mvn\t$Rd, $Rm, $Imm6", 3806 [(set GPR:$Rd, (not (shl GPR:$Rm, 3807 !cast<Operand>("lsl_operand_" # sty):$Imm6)))], 3808 NoItinerary>; 3809 3810 3811 def _lsr : A64I_logicalshift<sf, 0b01, 0b01, 0b1, 3812 (outs GPR:$Rd), 3813 (ins GPR:$Rm, 3814 !cast<Operand>("lsr_operand_" # sty):$Imm6), 3815 "mvn\t$Rd, $Rm, $Imm6", 3816 [(set GPR:$Rd, (not (srl GPR:$Rm, 3817 !cast<Operand>("lsr_operand_" # sty):$Imm6)))], 3818 NoItinerary>; 3819 3820 def _asr : A64I_logicalshift<sf, 0b01, 0b10, 0b1, 3821 (outs GPR:$Rd), 3822 (ins GPR:$Rm, 3823 !cast<Operand>("asr_operand_" # sty):$Imm6), 3824 "mvn\t$Rd, $Rm, $Imm6", 3825 [(set GPR:$Rd, (not (sra GPR:$Rm, 3826 !cast<Operand>("asr_operand_" # sty):$Imm6)))], 3827 NoItinerary>; 3828 3829 def _ror : A64I_logicalshift<sf, 0b01, 0b11, 0b1, 3830 (outs GPR:$Rd), 3831 (ins GPR:$Rm, 3832 !cast<Operand>("ror_operand_" # sty):$Imm6), 3833 "mvn\t$Rd, $Rm, $Imm6", 3834 [(set GPR:$Rd, (not (rotr GPR:$Rm, 3835 !cast<Operand>("lsl_operand_" # sty):$Imm6)))], 3836 NoItinerary>; 3837 } 3838 3839 def _noshift : InstAlias<"mvn $Rn, $Rm", 3840 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>; 3841 3842 def : Pat<(not GPR:$Rm), 3843 (!cast<Instruction>(prefix # "_lsl") GPR:$Rm, 0)>; 3844 } 3845 3846 defm MVNxx : mvn_shifts<"MVNxx", 0b1, "i64", GPR64>; 3847 defm MVNww : mvn_shifts<"MVNww", 0b0, "i32", GPR32>; 3848 3849 def MOVxx :InstAlias<"mov $Rd, $Rm", (ORRxxx_lsl GPR64:$Rd, XZR, GPR64:$Rm, 0)>; 3850 def MOVww :InstAlias<"mov $Rd, $Rm", (ORRwww_lsl GPR32:$Rd, WZR, GPR32:$Rm, 0)>; 3851 3852 //===----------------------------------------------------------------------===// 3853 // Move wide (immediate) instructions 3854 //===----------------------------------------------------------------------===// 3855 // Contains: MOVN, MOVZ, MOVK + MOV aliases 3856 3857 // A wide variety of different relocations are needed for variants of these 3858 // instructions, so it turns out that we need a different operand for all of 3859 // them. 3860 multiclass movw_operands<string prefix, string instname, int width> { 3861 def _imm_asmoperand : AsmOperandClass { 3862 let Name = instname # width # "Shifted" # shift; 3863 let PredicateMethod = "is" # instname # width # "Imm"; 3864 let RenderMethod = "addMoveWideImmOperands"; 3865 let ParserMethod = "ParseImmWithLSLOperand"; 3866 let DiagnosticType = "MOVWUImm16"; 3867 } 3868 3869 def _imm : Operand<i32> { 3870 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_imm_asmoperand"); 3871 let PrintMethod = "printMoveWideImmOperand"; 3872 let EncoderMethod = "getMoveWideImmOpValue"; 3873 let DecoderMethod = "DecodeMoveWideImmOperand<" # width # ">"; 3874 3875 let MIOperandInfo = (ops uimm16:$UImm16, imm:$Shift); 3876 } 3877 } 3878 3879 defm movn32 : movw_operands<"movn32", "MOVN", 32>; 3880 defm movn64 : movw_operands<"movn64", "MOVN", 64>; 3881 defm movz32 : movw_operands<"movz32", "MOVZ", 32>; 3882 defm movz64 : movw_operands<"movz64", "MOVZ", 64>; 3883 defm movk32 : movw_operands<"movk32", "MOVK", 32>; 3884 defm movk64 : movw_operands<"movk64", "MOVK", 64>; 3885 3886 multiclass A64I_movwSizes<bits<2> opc, string asmop, dag ins32bit, 3887 dag ins64bit> { 3888 3889 def wii : A64I_movw<0b0, opc, (outs GPR32:$Rd), ins32bit, 3890 !strconcat(asmop, "\t$Rd, $FullImm"), 3891 [], NoItinerary> { 3892 bits<18> FullImm; 3893 let UImm16 = FullImm{15-0}; 3894 let Shift = FullImm{17-16}; 3895 } 3896 3897 def xii : A64I_movw<0b1, opc, (outs GPR64:$Rd), ins64bit, 3898 !strconcat(asmop, "\t$Rd, $FullImm"), 3899 [], NoItinerary> { 3900 bits<18> FullImm; 3901 let UImm16 = FullImm{15-0}; 3902 let Shift = FullImm{17-16}; 3903 } 3904 } 3905 3906 let isMoveImm = 1, isReMaterializable = 1, 3907 isAsCheapAsAMove = 1, hasSideEffects = 0 in { 3908 defm MOVN : A64I_movwSizes<0b00, "movn", 3909 (ins movn32_imm:$FullImm), 3910 (ins movn64_imm:$FullImm)>; 3911 3912 // Some relocations are able to convert between a MOVZ and a MOVN. If these 3913 // are applied the instruction must be emitted with the corresponding bits as 3914 // 0, which means a MOVZ needs to override that bit from the default. 3915 let PostEncoderMethod = "fixMOVZ" in 3916 defm MOVZ : A64I_movwSizes<0b10, "movz", 3917 (ins movz32_imm:$FullImm), 3918 (ins movz64_imm:$FullImm)>; 3919 } 3920 3921 let Constraints = "$src = $Rd" in 3922 defm MOVK : A64I_movwSizes<0b11, "movk", 3923 (ins GPR32:$src, movk32_imm:$FullImm), 3924 (ins GPR64:$src, movk64_imm:$FullImm)>; 3925 3926 3927 // And now the "MOV" aliases. These also need their own operands because what 3928 // they accept is completely different to what the base instructions accept. 3929 multiclass movalias_operand<string prefix, string basename, 3930 string immpredicate, int width> { 3931 def _asmoperand : AsmOperandClass { 3932 let Name = basename # width # "MovAlias"; 3933 let PredicateMethod 3934 = "isMoveWideMovAlias<" # width # ", A64Imms::" # immpredicate # ">"; 3935 let RenderMethod 3936 = "addMoveWideMovAliasOperands<" # width # ", " 3937 # "A64Imms::" # immpredicate # ">"; 3938 } 3939 3940 def _movimm : Operand<i32> { 3941 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand"); 3942 3943 let MIOperandInfo = (ops uimm16:$UImm16, imm:$Shift); 3944 } 3945 } 3946 3947 defm movz32 : movalias_operand<"movz32", "MOVZ", "isMOVZImm", 32>; 3948 defm movz64 : movalias_operand<"movz64", "MOVZ", "isMOVZImm", 64>; 3949 defm movn32 : movalias_operand<"movn32", "MOVN", "isOnlyMOVNImm", 32>; 3950 defm movn64 : movalias_operand<"movn64", "MOVN", "isOnlyMOVNImm", 64>; 3951 3952 // FIXME: these are officially canonical aliases, but TableGen is too limited to 3953 // print them at the moment. I believe in this case an "AliasPredicate" method 3954 // will need to be implemented. to allow it, as well as the more generally 3955 // useful handling of non-register, non-constant operands. 3956 class movalias<Instruction INST, RegisterClass GPR, Operand operand> 3957 : InstAlias<"mov $Rd, $FullImm", (INST GPR:$Rd, operand:$FullImm)>; 3958 3959 def : movalias<MOVZwii, GPR32, movz32_movimm>; 3960 def : movalias<MOVZxii, GPR64, movz64_movimm>; 3961 def : movalias<MOVNwii, GPR32, movn32_movimm>; 3962 def : movalias<MOVNxii, GPR64, movn64_movimm>; 3963 3964 //===----------------------------------------------------------------------===// 3965 // PC-relative addressing instructions 3966 //===----------------------------------------------------------------------===// 3967 // Contains: ADR, ADRP 3968 3969 def adr_label : Operand<i64> { 3970 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_adr_prel>"; 3971 3972 // This label is a 21-bit offset from PC, unscaled 3973 let PrintMethod = "printLabelOperand<21, 1>"; 3974 let ParserMatchClass = label_asmoperand<21, 1>; 3975 let OperandType = "OPERAND_PCREL"; 3976 } 3977 3978 def adrp_label_asmoperand : AsmOperandClass { 3979 let Name = "AdrpLabel"; 3980 let RenderMethod = "addLabelOperands<21, 4096>"; 3981 let DiagnosticType = "Label"; 3982 } 3983 3984 def adrp_label : Operand<i64> { 3985 let EncoderMethod = "getAdrpLabelOpValue"; 3986 3987 // This label is a 21-bit offset from PC, scaled by the page-size: 4096. 3988 let PrintMethod = "printLabelOperand<21, 4096>"; 3989 let ParserMatchClass = adrp_label_asmoperand; 3990 let OperandType = "OPERAND_PCREL"; 3991 } 3992 3993 let hasSideEffects = 0 in { 3994 def ADRxi : A64I_PCADR<0b0, (outs GPR64:$Rd), (ins adr_label:$Label), 3995 "adr\t$Rd, $Label", [], NoItinerary>; 3996 3997 def ADRPxi : A64I_PCADR<0b1, (outs GPR64:$Rd), (ins adrp_label:$Label), 3998 "adrp\t$Rd, $Label", [], NoItinerary>; 3999 } 4000 4001 //===----------------------------------------------------------------------===// 4002 // System instructions 4003 //===----------------------------------------------------------------------===// 4004 // Contains: HINT, CLREX, DSB, DMB, ISB, MSR, SYS, SYSL, MRS 4005 // + aliases IC, DC, AT, TLBI, NOP, YIELD, WFE, WFI, SEV, SEVL 4006 4007 // Op1 and Op2 fields are sometimes simple 3-bit unsigned immediate values. 4008 def uimm3_asmoperand : AsmOperandClass { 4009 let Name = "UImm3"; 4010 let PredicateMethod = "isUImm<3>"; 4011 let RenderMethod = "addImmOperands"; 4012 let DiagnosticType = "UImm3"; 4013 } 4014 4015 def uimm3 : Operand<i32> { 4016 let ParserMatchClass = uimm3_asmoperand; 4017 } 4018 4019 // The HINT alias can accept a simple unsigned 7-bit immediate. 4020 def uimm7_asmoperand : AsmOperandClass { 4021 let Name = "UImm7"; 4022 let PredicateMethod = "isUImm<7>"; 4023 let RenderMethod = "addImmOperands"; 4024 let DiagnosticType = "UImm7"; 4025 } 4026 4027 def uimm7 : Operand<i32> { 4028 let ParserMatchClass = uimm7_asmoperand; 4029 } 4030 4031 // Multiclass namedimm is defined with the prefetch operands. Most of these fit 4032 // into the NamedImmMapper scheme well: they either accept a named operand or 4033 // any immediate under a particular value (which may be 0, implying no immediate 4034 // is allowed). 4035 defm dbarrier : namedimm<"dbarrier", "A64DB::DBarrierMapper">; 4036 defm isb : namedimm<"isb", "A64ISB::ISBMapper">; 4037 defm ic : namedimm<"ic", "A64IC::ICMapper">; 4038 defm dc : namedimm<"dc", "A64DC::DCMapper">; 4039 defm at : namedimm<"at", "A64AT::ATMapper">; 4040 defm tlbi : namedimm<"tlbi", "A64TLBI::TLBIMapper">; 4041 4042 // However, MRS and MSR are more complicated for a few reasons: 4043 // * There are ~1000 generic names S3_<op1>_<CRn>_<CRm>_<Op2> which have an 4044 // implementation-defined effect 4045 // * Most registers are shared, but some are read-only or write-only. 4046 // * There is a variant of MSR which accepts the same register name (SPSel), 4047 // but which would have a different encoding. 4048 4049 // In principle these could be resolved in with more complicated subclasses of 4050 // NamedImmMapper, however that imposes an overhead on other "named 4051 // immediates". Both in concrete terms with virtual tables and in unnecessary 4052 // abstraction. 4053 4054 // The solution adopted here is to take the MRS/MSR Mappers out of the usual 4055 // hierarchy (they're not derived from NamedImmMapper) and to add logic for 4056 // their special situation. 4057 def mrs_asmoperand : AsmOperandClass { 4058 let Name = "MRS"; 4059 let ParserMethod = "ParseSysRegOperand"; 4060 let DiagnosticType = "MRS"; 4061 } 4062 4063 def mrs_op : Operand<i32> { 4064 let ParserMatchClass = mrs_asmoperand; 4065 let PrintMethod = "printMRSOperand"; 4066 let DecoderMethod = "DecodeMRSOperand"; 4067 } 4068 4069 def msr_asmoperand : AsmOperandClass { 4070 let Name = "MSRWithReg"; 4071 4072 // Note that SPSel is valid for both this and the pstate operands, but with 4073 // different immediate encodings. This is why these operands provide a string 4074 // AArch64Operand rather than an immediate. The overlap is small enough that 4075 // it could be resolved with hackery now, but who can say in future? 4076 let ParserMethod = "ParseSysRegOperand"; 4077 let DiagnosticType = "MSR"; 4078 } 4079 4080 def msr_op : Operand<i32> { 4081 let ParserMatchClass = msr_asmoperand; 4082 let PrintMethod = "printMSROperand"; 4083 let DecoderMethod = "DecodeMSROperand"; 4084 } 4085 4086 def pstate_asmoperand : AsmOperandClass { 4087 let Name = "MSRPState"; 4088 // See comment above about parser. 4089 let ParserMethod = "ParseSysRegOperand"; 4090 let DiagnosticType = "MSR"; 4091 } 4092 4093 def pstate_op : Operand<i32> { 4094 let ParserMatchClass = pstate_asmoperand; 4095 let PrintMethod = "printNamedImmOperand<A64PState::PStateMapper>"; 4096 let DecoderMethod = "DecodeNamedImmOperand<A64PState::PStateMapper>"; 4097 } 4098 4099 // When <CRn> is specified, an assembler should accept something like "C4", not 4100 // the usual "#4" immediate. 4101 def CRx_asmoperand : AsmOperandClass { 4102 let Name = "CRx"; 4103 let PredicateMethod = "isUImm<4>"; 4104 let RenderMethod = "addImmOperands"; 4105 let ParserMethod = "ParseCRxOperand"; 4106 // Diagnostics are handled in all cases by ParseCRxOperand. 4107 } 4108 4109 def CRx : Operand<i32> { 4110 let ParserMatchClass = CRx_asmoperand; 4111 let PrintMethod = "printCRxOperand"; 4112 } 4113 4114 4115 // Finally, we can start defining the instructions. 4116 4117 // HINT is straightforward, with a few aliases. 4118 def HINTi : A64I_system<0b0, (outs), (ins uimm7:$UImm7), "hint\t$UImm7", 4119 [], NoItinerary> { 4120 bits<7> UImm7; 4121 let CRm = UImm7{6-3}; 4122 let Op2 = UImm7{2-0}; 4123 4124 let Op0 = 0b00; 4125 let Op1 = 0b011; 4126 let CRn = 0b0010; 4127 let Rt = 0b11111; 4128 } 4129 4130 def : InstAlias<"nop", (HINTi 0)>; 4131 def : InstAlias<"yield", (HINTi 1)>; 4132 def : InstAlias<"wfe", (HINTi 2)>; 4133 def : InstAlias<"wfi", (HINTi 3)>; 4134 def : InstAlias<"sev", (HINTi 4)>; 4135 def : InstAlias<"sevl", (HINTi 5)>; 4136 4137 // Quite a few instructions then follow a similar pattern of fixing common 4138 // fields in the bitpattern, we'll define a helper-class for them. 4139 class simple_sys<bits<2> op0, bits<3> op1, bits<4> crn, bits<3> op2, 4140 Operand operand, string asmop> 4141 : A64I_system<0b0, (outs), (ins operand:$CRm), !strconcat(asmop, "\t$CRm"), 4142 [], NoItinerary> { 4143 let Op0 = op0; 4144 let Op1 = op1; 4145 let CRn = crn; 4146 let Op2 = op2; 4147 let Rt = 0b11111; 4148 } 4149 4150 4151 def CLREXi : simple_sys<0b00, 0b011, 0b0011, 0b010, uimm4, "clrex">; 4152 def DSBi : simple_sys<0b00, 0b011, 0b0011, 0b100, dbarrier_op, "dsb">; 4153 def DMBi : simple_sys<0b00, 0b011, 0b0011, 0b101, dbarrier_op, "dmb">; 4154 def ISBi : simple_sys<0b00, 0b011, 0b0011, 0b110, isb_op, "isb">; 4155 4156 def : InstAlias<"clrex", (CLREXi 0b1111)>; 4157 def : InstAlias<"isb", (ISBi 0b1111)>; 4158 4159 // (DMBi 0xb) is a "DMB ISH" instruciton, appropriate for Linux SMP 4160 // configurations at least. 4161 def : Pat<(atomic_fence imm, imm), (DMBi 0xb)>; 4162 4163 // Any SYS bitpattern can be represented with a complex and opaque "SYS" 4164 // instruction. 4165 def SYSiccix : A64I_system<0b0, (outs), 4166 (ins uimm3:$Op1, CRx:$CRn, CRx:$CRm, 4167 uimm3:$Op2, GPR64:$Rt), 4168 "sys\t$Op1, $CRn, $CRm, $Op2, $Rt", 4169 [], NoItinerary> { 4170 let Op0 = 0b01; 4171 } 4172 4173 // You can skip the Xt argument whether it makes sense or not for the generic 4174 // SYS instruction. 4175 def : InstAlias<"sys $Op1, $CRn, $CRm, $Op2", 4176 (SYSiccix uimm3:$Op1, CRx:$CRn, CRx:$CRm, uimm3:$Op2, XZR)>; 4177 4178 4179 // But many have aliases, which obviously don't fit into 4180 class SYSalias<dag ins, string asmstring> 4181 : A64I_system<0b0, (outs), ins, asmstring, [], NoItinerary> { 4182 let isAsmParserOnly = 1; 4183 4184 bits<14> SysOp; 4185 let Op0 = 0b01; 4186 let Op1 = SysOp{13-11}; 4187 let CRn = SysOp{10-7}; 4188 let CRm = SysOp{6-3}; 4189 let Op2 = SysOp{2-0}; 4190 } 4191 4192 def ICix : SYSalias<(ins ic_op:$SysOp, GPR64:$Rt), "ic\t$SysOp, $Rt">; 4193 4194 def ICi : SYSalias<(ins ic_op:$SysOp), "ic\t$SysOp"> { 4195 let Rt = 0b11111; 4196 } 4197 4198 def DCix : SYSalias<(ins dc_op:$SysOp, GPR64:$Rt), "dc\t$SysOp, $Rt">; 4199 def ATix : SYSalias<(ins at_op:$SysOp, GPR64:$Rt), "at\t$SysOp, $Rt">; 4200 4201 def TLBIix : SYSalias<(ins tlbi_op:$SysOp, GPR64:$Rt), "tlbi\t$SysOp, $Rt">; 4202 4203 def TLBIi : SYSalias<(ins tlbi_op:$SysOp), "tlbi\t$SysOp"> { 4204 let Rt = 0b11111; 4205 } 4206 4207 4208 def SYSLxicci : A64I_system<0b1, (outs GPR64:$Rt), 4209 (ins uimm3:$Op1, CRx:$CRn, CRx:$CRm, uimm3:$Op2), 4210 "sysl\t$Rt, $Op1, $CRn, $CRm, $Op2", 4211 [], NoItinerary> { 4212 let Op0 = 0b01; 4213 } 4214 4215 // The instructions themselves are rather simple for MSR and MRS. 4216 def MSRix : A64I_system<0b0, (outs), (ins msr_op:$SysReg, GPR64:$Rt), 4217 "msr\t$SysReg, $Rt", [], NoItinerary> { 4218 bits<16> SysReg; 4219 let Op0 = SysReg{15-14}; 4220 let Op1 = SysReg{13-11}; 4221 let CRn = SysReg{10-7}; 4222 let CRm = SysReg{6-3}; 4223 let Op2 = SysReg{2-0}; 4224 } 4225 4226 def MRSxi : A64I_system<0b1, (outs GPR64:$Rt), (ins mrs_op:$SysReg), 4227 "mrs\t$Rt, $SysReg", [], NoItinerary> { 4228 bits<16> SysReg; 4229 let Op0 = SysReg{15-14}; 4230 let Op1 = SysReg{13-11}; 4231 let CRn = SysReg{10-7}; 4232 let CRm = SysReg{6-3}; 4233 let Op2 = SysReg{2-0}; 4234 } 4235 4236 def MSRii : A64I_system<0b0, (outs), (ins pstate_op:$PState, uimm4:$CRm), 4237 "msr\t$PState, $CRm", [], NoItinerary> { 4238 bits<6> PState; 4239 4240 let Op0 = 0b00; 4241 let Op1 = PState{5-3}; 4242 let CRn = 0b0100; 4243 let Op2 = PState{2-0}; 4244 let Rt = 0b11111; 4245 } 4246 4247 //===----------------------------------------------------------------------===// 4248 // Test & branch (immediate) instructions 4249 //===----------------------------------------------------------------------===// 4250 // Contains: TBZ, TBNZ 4251 4252 // The bit to test is a simple unsigned 6-bit immediate in the X-register 4253 // versions. 4254 def uimm6 : Operand<i64> { 4255 let ParserMatchClass = uimm6_asmoperand; 4256 } 4257 4258 def label_wid14_scal4_asmoperand : label_asmoperand<14, 4>; 4259 4260 def tbimm_target : Operand<OtherVT> { 4261 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_tstbr>"; 4262 4263 // This label is a 14-bit offset from PC, scaled by the instruction-width: 4. 4264 let PrintMethod = "printLabelOperand<14, 4>"; 4265 let ParserMatchClass = label_wid14_scal4_asmoperand; 4266 4267 let OperandType = "OPERAND_PCREL"; 4268 } 4269 4270 def A64eq : ImmLeaf<i32, [{ return Imm == A64CC::EQ; }]>; 4271 def A64ne : ImmLeaf<i32, [{ return Imm == A64CC::NE; }]>; 4272 4273 // These instructions correspond to patterns involving "and" with a power of 4274 // two, which we need to be able to select. 4275 def tstb64_pat : ComplexPattern<i64, 1, "SelectTSTBOperand<64>">; 4276 def tstb32_pat : ComplexPattern<i32, 1, "SelectTSTBOperand<32>">; 4277 4278 let isBranch = 1, isTerminator = 1 in { 4279 def TBZxii : A64I_TBimm<0b0, (outs), 4280 (ins GPR64:$Rt, uimm6:$Imm, tbimm_target:$Label), 4281 "tbz\t$Rt, $Imm, $Label", 4282 [(A64br_cc (A64cmp (and GPR64:$Rt, tstb64_pat:$Imm), 0), 4283 A64eq, bb:$Label)], 4284 NoItinerary>; 4285 4286 def TBNZxii : A64I_TBimm<0b1, (outs), 4287 (ins GPR64:$Rt, uimm6:$Imm, tbimm_target:$Label), 4288 "tbnz\t$Rt, $Imm, $Label", 4289 [(A64br_cc (A64cmp (and GPR64:$Rt, tstb64_pat:$Imm), 0), 4290 A64ne, bb:$Label)], 4291 NoItinerary>; 4292 4293 4294 // Note, these instructions overlap with the above 64-bit patterns. This is 4295 // intentional, "tbz x3, #1, somewhere" and "tbz w3, #1, somewhere" would both 4296 // do the same thing and are both permitted assembly. They also both have 4297 // sensible DAG patterns. 4298 def TBZwii : A64I_TBimm<0b0, (outs), 4299 (ins GPR32:$Rt, uimm5:$Imm, tbimm_target:$Label), 4300 "tbz\t$Rt, $Imm, $Label", 4301 [(A64br_cc (A64cmp (and GPR32:$Rt, tstb32_pat:$Imm), 0), 4302 A64eq, bb:$Label)], 4303 NoItinerary> { 4304 let Imm{5} = 0b0; 4305 } 4306 4307 def TBNZwii : A64I_TBimm<0b1, (outs), 4308 (ins GPR32:$Rt, uimm5:$Imm, tbimm_target:$Label), 4309 "tbnz\t$Rt, $Imm, $Label", 4310 [(A64br_cc (A64cmp (and GPR32:$Rt, tstb32_pat:$Imm), 0), 4311 A64ne, bb:$Label)], 4312 NoItinerary> { 4313 let Imm{5} = 0b0; 4314 } 4315 } 4316 4317 //===----------------------------------------------------------------------===// 4318 // Unconditional branch (immediate) instructions 4319 //===----------------------------------------------------------------------===// 4320 // Contains: B, BL 4321 4322 def label_wid26_scal4_asmoperand : label_asmoperand<26, 4>; 4323 4324 def bimm_target : Operand<OtherVT> { 4325 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_uncondbr>"; 4326 4327 // This label is a 26-bit offset from PC, scaled by the instruction-width: 4. 4328 let PrintMethod = "printLabelOperand<26, 4>"; 4329 let ParserMatchClass = label_wid26_scal4_asmoperand; 4330 4331 let OperandType = "OPERAND_PCREL"; 4332 } 4333 4334 def blimm_target : Operand<i64> { 4335 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_call>"; 4336 4337 // This label is a 26-bit offset from PC, scaled by the instruction-width: 4. 4338 let PrintMethod = "printLabelOperand<26, 4>"; 4339 let ParserMatchClass = label_wid26_scal4_asmoperand; 4340 4341 let OperandType = "OPERAND_PCREL"; 4342 } 4343 4344 class A64I_BimmImpl<bit op, string asmop, list<dag> patterns, Operand lbl_type> 4345 : A64I_Bimm<op, (outs), (ins lbl_type:$Label), 4346 !strconcat(asmop, "\t$Label"), patterns, 4347 NoItinerary>; 4348 4349 let isBranch = 1 in { 4350 def Bimm : A64I_BimmImpl<0b0, "b", [(br bb:$Label)], bimm_target> { 4351 let isTerminator = 1; 4352 let isBarrier = 1; 4353 } 4354 4355 def BLimm : A64I_BimmImpl<0b1, "bl", 4356 [(AArch64Call tglobaladdr:$Label)], blimm_target> { 4357 let isCall = 1; 4358 let Defs = [X30]; 4359 } 4360 } 4361 4362 def : Pat<(AArch64Call texternalsym:$Label), (BLimm texternalsym:$Label)>; 4363 4364 //===----------------------------------------------------------------------===// 4365 // Unconditional branch (register) instructions 4366 //===----------------------------------------------------------------------===// 4367 // Contains: BR, BLR, RET, ERET, DRP. 4368 4369 // Most of the notional opcode fields in the A64I_Breg format are fixed in A64 4370 // at the moment. 4371 class A64I_BregImpl<bits<4> opc, 4372 dag outs, dag ins, string asmstr, list<dag> patterns, 4373 InstrItinClass itin = NoItinerary> 4374 : A64I_Breg<opc, 0b11111, 0b000000, 0b00000, 4375 outs, ins, asmstr, patterns, itin> { 4376 let isBranch = 1; 4377 let isIndirectBranch = 1; 4378 } 4379 4380 // Note that these are not marked isCall or isReturn because as far as LLVM is 4381 // concerned they're not. "ret" is just another jump unless it has been selected 4382 // by LLVM as the function's return. 4383 4384 let isBranch = 1 in { 4385 def BRx : A64I_BregImpl<0b0000,(outs), (ins GPR64:$Rn), 4386 "br\t$Rn", [(brind GPR64:$Rn)]> { 4387 let isBarrier = 1; 4388 let isTerminator = 1; 4389 } 4390 4391 def BLRx : A64I_BregImpl<0b0001, (outs), (ins GPR64:$Rn), 4392 "blr\t$Rn", [(AArch64Call GPR64:$Rn)]> { 4393 let isBarrier = 0; 4394 let isCall = 1; 4395 let Defs = [X30]; 4396 } 4397 4398 def RETx : A64I_BregImpl<0b0010, (outs), (ins GPR64:$Rn), 4399 "ret\t$Rn", []> { 4400 let isBarrier = 1; 4401 let isTerminator = 1; 4402 let isReturn = 1; 4403 } 4404 4405 // Create a separate pseudo-instruction for codegen to use so that we don't 4406 // flag x30 as used in every function. It'll be restored before the RET by the 4407 // epilogue if it's legitimately used. 4408 def RET : A64PseudoExpand<(outs), (ins), [(A64ret)], (RETx (ops X30))> { 4409 let isTerminator = 1; 4410 let isBarrier = 1; 4411 let isReturn = 1; 4412 } 4413 4414 def ERET : A64I_BregImpl<0b0100, (outs), (ins), "eret", []> { 4415 let Rn = 0b11111; 4416 let isBarrier = 1; 4417 let isTerminator = 1; 4418 let isReturn = 1; 4419 } 4420 4421 def DRPS : A64I_BregImpl<0b0101, (outs), (ins), "drps", []> { 4422 let Rn = 0b11111; 4423 let isBarrier = 1; 4424 } 4425 } 4426 4427 def RETAlias : InstAlias<"ret", (RETx X30)>; 4428 4429 4430 //===----------------------------------------------------------------------===// 4431 // Address generation patterns 4432 //===----------------------------------------------------------------------===// 4433 4434 // Primary method of address generation for the small/absolute memory model is 4435 // an ADRP/ADR pair: 4436 // ADRP x0, some_variable 4437 // ADD x0, x0, #:lo12:some_variable 4438 // 4439 // The load/store elision of the ADD is accomplished when selecting 4440 // addressing-modes. This just mops up the cases where that doesn't work and we 4441 // really need an address in some register. 4442 4443 // This wrapper applies a LO12 modifier to the address. Otherwise we could just 4444 // use the same address. 4445 4446 class ADRP_ADD<SDNode Wrapper, SDNode addrop> 4447 : Pat<(Wrapper addrop:$Hi, addrop:$Lo12, (i32 imm)), 4448 (ADDxxi_lsl0_s (ADRPxi addrop:$Hi), addrop:$Lo12)>; 4449 4450 def : ADRP_ADD<A64WrapperSmall, tblockaddress>; 4451 def : ADRP_ADD<A64WrapperSmall, texternalsym>; 4452 def : ADRP_ADD<A64WrapperSmall, tglobaladdr>; 4453 def : ADRP_ADD<A64WrapperSmall, tglobaltlsaddr>; 4454 def : ADRP_ADD<A64WrapperSmall, tjumptable>; 4455 4456 //===----------------------------------------------------------------------===// 4457 // GOT access patterns 4458 //===----------------------------------------------------------------------===// 4459 4460 // FIXME: Wibble 4461 4462 class GOTLoadSmall<SDNode addrfrag> 4463 : Pat<(A64GOTLoad (A64WrapperSmall addrfrag:$Hi, addrfrag:$Lo12, 8)), 4464 (LS64_LDR (ADRPxi addrfrag:$Hi), addrfrag:$Lo12)>; 4465 4466 def : GOTLoadSmall<texternalsym>; 4467 def : GOTLoadSmall<tglobaladdr>; 4468 def : GOTLoadSmall<tglobaltlsaddr>; 4469 4470 //===----------------------------------------------------------------------===// 4471 // Tail call handling 4472 //===----------------------------------------------------------------------===// 4473 4474 let isCall = 1, isTerminator = 1, isReturn = 1, isBarrier = 1, Uses = [XSP] in { 4475 def TC_RETURNdi 4476 : PseudoInst<(outs), (ins i64imm:$dst, i32imm:$FPDiff), 4477 [(AArch64tcret tglobaladdr:$dst, (i32 timm:$FPDiff))]>; 4478 4479 def TC_RETURNxi 4480 : PseudoInst<(outs), (ins tcGPR64:$dst, i32imm:$FPDiff), 4481 [(AArch64tcret tcGPR64:$dst, (i32 timm:$FPDiff))]>; 4482 } 4483 4484 let isCall = 1, isTerminator = 1, isReturn = 1, isBarrier = 1, 4485 Uses = [XSP] in { 4486 def TAIL_Bimm : A64PseudoExpand<(outs), (ins bimm_target:$Label), [], 4487 (Bimm bimm_target:$Label)>; 4488 4489 def TAIL_BRx : A64PseudoExpand<(outs), (ins tcGPR64:$Rd), [], 4490 (BRx GPR64:$Rd)>; 4491 } 4492 4493 4494 def : Pat<(AArch64tcret texternalsym:$dst, (i32 timm:$FPDiff)), 4495 (TC_RETURNdi texternalsym:$dst, imm:$FPDiff)>; 4496 4497 //===----------------------------------------------------------------------===// 4498 // Thread local storage 4499 //===----------------------------------------------------------------------===// 4500 4501 // This is a pseudo-instruction representing the ".tlsdesccall" directive in 4502 // assembly. Its effect is to insert an R_AARCH64_TLSDESC_CALL relocation at the 4503 // current location. It should always be immediately followed by a BLR 4504 // instruction, and is intended solely for relaxation by the linker. 4505 4506 def : Pat<(A64threadpointer), (MRSxi 0xde82)>; 4507 4508 def TLSDESCCALL : PseudoInst<(outs), (ins i64imm:$Lbl), []> { 4509 let hasSideEffects = 1; 4510 } 4511 4512 def TLSDESC_BLRx : PseudoInst<(outs), (ins GPR64:$Rn, i64imm:$Var), 4513 [(A64tlsdesc_blr GPR64:$Rn, tglobaltlsaddr:$Var)]> { 4514 let isCall = 1; 4515 let Defs = [X30]; 4516 } 4517 4518 def : Pat<(A64tlsdesc_blr GPR64:$Rn, texternalsym:$Var), 4519 (TLSDESC_BLRx GPR64:$Rn, texternalsym:$Var)>; 4520 4521 //===----------------------------------------------------------------------===// 4522 // Bitfield patterns 4523 //===----------------------------------------------------------------------===// 4524 4525 def bfi32_lsb_to_immr : SDNodeXForm<imm, [{ 4526 return CurDAG->getTargetConstant((32 - N->getZExtValue()) % 32, MVT::i64); 4527 }]>; 4528 4529 def bfi64_lsb_to_immr : SDNodeXForm<imm, [{ 4530 return CurDAG->getTargetConstant((64 - N->getZExtValue()) % 64, MVT::i64); 4531 }]>; 4532 4533 def bfi_width_to_imms : SDNodeXForm<imm, [{ 4534 return CurDAG->getTargetConstant(N->getZExtValue() - 1, MVT::i64); 4535 }]>; 4536 4537 4538 // The simpler patterns deal with cases where no AND mask is actually needed 4539 // (either all bits are used or the low 32 bits are used). 4540 let AddedComplexity = 10 in { 4541 4542 def : Pat<(A64Bfi GPR64:$src, GPR64:$Rn, imm:$ImmR, imm:$ImmS), 4543 (BFIxxii GPR64:$src, GPR64:$Rn, 4544 (bfi64_lsb_to_immr (i64 imm:$ImmR)), 4545 (bfi_width_to_imms (i64 imm:$ImmS)))>; 4546 4547 def : Pat<(A64Bfi GPR32:$src, GPR32:$Rn, imm:$ImmR, imm:$ImmS), 4548 (BFIwwii GPR32:$src, GPR32:$Rn, 4549 (bfi32_lsb_to_immr (i64 imm:$ImmR)), 4550 (bfi_width_to_imms (i64 imm:$ImmS)))>; 4551 4552 4553 def : Pat<(and (A64Bfi GPR64:$src, GPR64:$Rn, imm:$ImmR, imm:$ImmS), 4554 (i64 4294967295)), 4555 (SUBREG_TO_REG (i64 0), 4556 (BFIwwii (EXTRACT_SUBREG GPR64:$src, sub_32), 4557 (EXTRACT_SUBREG GPR64:$Rn, sub_32), 4558 (bfi32_lsb_to_immr (i64 imm:$ImmR)), 4559 (bfi_width_to_imms (i64 imm:$ImmS))), 4560 sub_32)>; 4561 4562 } 4563 4564 //===----------------------------------------------------------------------===// 4565 // Miscellaneous patterns 4566 //===----------------------------------------------------------------------===// 4567 4568 // Truncation from 64 to 32-bits just involves renaming your register. 4569 def : Pat<(i32 (trunc (i64 GPR64:$val))), (EXTRACT_SUBREG GPR64:$val, sub_32)>; 4570 4571 // Similarly, extension where we don't care about the high bits is 4572 // just a rename. 4573 def : Pat<(i64 (anyext (i32 GPR32:$val))), 4574 (INSERT_SUBREG (IMPLICIT_DEF), GPR32:$val, sub_32)>; 4575 4576 // SELECT instructions providing f128 types need to be handled by a 4577 // pseudo-instruction since the eventual code will need to introduce basic 4578 // blocks and control flow. 4579 def F128CSEL : PseudoInst<(outs FPR128:$Rd), 4580 (ins FPR128:$Rn, FPR128:$Rm, cond_code_op:$Cond), 4581 [(set FPR128:$Rd, (simple_select (f128 FPR128:$Rn), 4582 FPR128:$Rm))]> { 4583 let Uses = [NZCV]; 4584 let usesCustomInserter = 1; 4585 } 4586 4587 //===----------------------------------------------------------------------===// 4588 // Load/store patterns 4589 //===----------------------------------------------------------------------===// 4590 4591 // There are lots of patterns here, because we need to allow at least three 4592 // parameters to vary independently. 4593 // 1. Instruction: "ldrb w9, [sp]", "ldrh w9, [sp]", ... 4594 // 2. LLVM source: zextloadi8, anyextloadi8, ... 4595 // 3. Address-generation: A64Wrapper, (add BASE, OFFSET), ... 4596 // 4597 // The biggest problem turns out to be the address-generation variable. At the 4598 // point of instantiation we need to produce two DAGs, one for the pattern and 4599 // one for the instruction. Doing this at the lowest level of classes doesn't 4600 // work. 4601 // 4602 // Consider the simple uimm12 addressing mode, and the desire to match both (add 4603 // GPR64xsp:$Rn, uimm12:$Offset) and GPR64xsp:$Rn, particularly on the 4604 // instruction side. We'd need to insert either "GPR64xsp" and "uimm12" or 4605 // "GPR64xsp" and "0" into an unknown dag. !subst is not capable of this 4606 // operation, and PatFrags are for selection not output. 4607 // 4608 // As a result, the address-generation patterns are the final 4609 // instantiations. However, we do still need to vary the operand for the address 4610 // further down (At the point we're deciding A64WrapperSmall, we don't know 4611 // the memory width of the operation). 4612 4613 //===------------------------------ 4614 // 1. Basic infrastructural defs 4615 //===------------------------------ 4616 4617 // First, some simple classes for !foreach and !subst to use: 4618 class Decls { 4619 dag pattern; 4620 } 4621 4622 def decls : Decls; 4623 def ALIGN; 4624 def INST; 4625 def OFFSET; 4626 def SHIFT; 4627 4628 // You can't use !subst on an actual immediate, but you *can* use it on an 4629 // operand record that happens to match a single immediate. So we do. 4630 def imm_eq0 : ImmLeaf<i64, [{ return Imm == 0; }]>; 4631 def imm_eq1 : ImmLeaf<i64, [{ return Imm == 1; }]>; 4632 def imm_eq2 : ImmLeaf<i64, [{ return Imm == 2; }]>; 4633 def imm_eq3 : ImmLeaf<i64, [{ return Imm == 3; }]>; 4634 def imm_eq4 : ImmLeaf<i64, [{ return Imm == 4; }]>; 4635 4636 // If the low bits of a pointer are known to be 0 then an "or" is just as good 4637 // as addition for computing an offset. This fragment forwards that check for 4638 // TableGen's use. 4639 def add_like_or : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs), 4640 [{ 4641 return CurDAG->isBaseWithConstantOffset(SDValue(N, 0)); 4642 }]>; 4643 4644 // Load/store (unsigned immediate) operations with relocations against global 4645 // symbols (for lo12) are only valid if those symbols have correct alignment 4646 // (since the immediate offset is divided by the access scale, it can't have a 4647 // remainder). 4648 // 4649 // The guaranteed alignment is provided as part of the WrapperSmall 4650 // operation, and checked against one of these. 4651 def any_align : ImmLeaf<i32, [{ (void)Imm; return true; }]>; 4652 def min_align2 : ImmLeaf<i32, [{ return Imm >= 2; }]>; 4653 def min_align4 : ImmLeaf<i32, [{ return Imm >= 4; }]>; 4654 def min_align8 : ImmLeaf<i32, [{ return Imm >= 8; }]>; 4655 def min_align16 : ImmLeaf<i32, [{ return Imm >= 16; }]>; 4656 4657 // "Normal" load/store instructions can be used on atomic operations, provided 4658 // the ordering parameter is at most "monotonic". Anything above that needs 4659 // special handling with acquire/release instructions. 4660 class simple_load<PatFrag base> 4661 : PatFrag<(ops node:$ptr), (base node:$ptr), [{ 4662 return cast<AtomicSDNode>(N)->getOrdering() <= Monotonic; 4663 }]>; 4664 4665 def atomic_load_simple_i8 : simple_load<atomic_load_8>; 4666 def atomic_load_simple_i16 : simple_load<atomic_load_16>; 4667 def atomic_load_simple_i32 : simple_load<atomic_load_32>; 4668 def atomic_load_simple_i64 : simple_load<atomic_load_64>; 4669 4670 class simple_store<PatFrag base> 4671 : PatFrag<(ops node:$ptr, node:$val), (base node:$ptr, node:$val), [{ 4672 return cast<AtomicSDNode>(N)->getOrdering() <= Monotonic; 4673 }]>; 4674 4675 def atomic_store_simple_i8 : simple_store<atomic_store_8>; 4676 def atomic_store_simple_i16 : simple_store<atomic_store_16>; 4677 def atomic_store_simple_i32 : simple_store<atomic_store_32>; 4678 def atomic_store_simple_i64 : simple_store<atomic_store_64>; 4679 4680 //===------------------------------ 4681 // 2. UImm12 and SImm9 4682 //===------------------------------ 4683 4684 // These instructions have two operands providing the address so they can be 4685 // treated similarly for most purposes. 4686 4687 //===------------------------------ 4688 // 2.1 Base patterns covering extend/truncate semantics 4689 //===------------------------------ 4690 4691 // Atomic patterns can be shared between integer operations of all sizes, a 4692 // quick multiclass here allows reuse. 4693 multiclass ls_atomic_pats<Instruction LOAD, Instruction STORE, dag Base, 4694 dag Offset, dag address, RegisterClass TPR, 4695 ValueType sty> { 4696 def : Pat<(!cast<PatFrag>("atomic_load_simple_" # sty) address), 4697 (LOAD Base, Offset)>; 4698 4699 def : Pat<(!cast<PatFrag>("atomic_store_simple_" # sty) address, TPR:$Rt), 4700 (STORE TPR:$Rt, Base, Offset)>; 4701 } 4702 4703 // Instructions accessing a memory chunk smaller than a register (or, in a 4704 // pinch, the same size) have a characteristic set of patterns they want to 4705 // match: extending loads and truncating stores. This class deals with the 4706 // sign-neutral version of those patterns. 4707 // 4708 // It will be instantiated across multiple addressing-modes. 4709 multiclass ls_small_pats<Instruction LOAD, Instruction STORE, 4710 dag Base, dag Offset, 4711 dag address, ValueType sty> 4712 : ls_atomic_pats<LOAD, STORE, Base, Offset, address, GPR32, sty> { 4713 def : Pat<(!cast<SDNode>(zextload # sty) address), (LOAD Base, Offset)>; 4714 4715 def : Pat<(!cast<SDNode>(extload # sty) address), (LOAD Base, Offset)>; 4716 4717 // For zero-extension to 64-bits we have to tell LLVM that the whole 64-bit 4718 // register was actually set. 4719 def : Pat<(i64 (!cast<SDNode>(zextload # sty) address)), 4720 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset), sub_32)>; 4721 4722 def : Pat<(i64 (!cast<SDNode>(extload # sty) address)), 4723 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset), sub_32)>; 4724 4725 def : Pat<(!cast<SDNode>(truncstore # sty) GPR32:$Rt, address), 4726 (STORE GPR32:$Rt, Base, Offset)>; 4727 4728 // For truncating store from 64-bits, we have to manually tell LLVM to 4729 // ignore the high bits of the x register. 4730 def : Pat<(!cast<SDNode>(truncstore # sty) GPR64:$Rt, address), 4731 (STORE (EXTRACT_SUBREG GPR64:$Rt, sub_32), Base, Offset)>; 4732 } 4733 4734 // Next come patterns for sign-extending loads. 4735 multiclass load_signed_pats<string T, string U, dag Base, dag Offset, 4736 dag address, ValueType sty> { 4737 def : Pat<(i32 (!cast<SDNode>("sextload" # sty) address)), 4738 (!cast<Instruction>("LDRS" # T # "w" # U) Base, Offset)>; 4739 4740 def : Pat<(i64 (!cast<SDNode>("sextload" # sty) address)), 4741 (!cast<Instruction>("LDRS" # T # "x" # U) Base, Offset)>; 4742 4743 } 4744 4745 // and finally "natural-width" loads and stores come next. 4746 multiclass ls_neutral_pats<Instruction LOAD, Instruction STORE, dag Base, 4747 dag Offset, dag address, RegisterClass TPR, 4748 ValueType sty> { 4749 def : Pat<(sty (load address)), (LOAD Base, Offset)>; 4750 def : Pat<(store (sty TPR:$Rt), address), (STORE TPR:$Rt, Base, Offset)>; 4751 } 4752 4753 // Integer operations also get atomic instructions to select for. 4754 multiclass ls_int_neutral_pats<Instruction LOAD, Instruction STORE, dag Base, 4755 dag Offset, dag address, RegisterClass TPR, 4756 ValueType sty> 4757 : ls_neutral_pats<LOAD, STORE, Base, Offset, address, TPR, sty>, 4758 ls_atomic_pats<LOAD, STORE, Base, Offset, address, TPR, sty>; 4759 4760 //===------------------------------ 4761 // 2.2. Addressing-mode instantiations 4762 //===------------------------------ 4763 4764 multiclass uimm12_pats<dag address, dag Base, dag Offset> { 4765 defm : ls_small_pats<LS8_LDR, LS8_STR, Base, 4766 !foreach(decls.pattern, Offset, 4767 !subst(OFFSET, byte_uimm12, decls.pattern)), 4768 !foreach(decls.pattern, address, 4769 !subst(OFFSET, byte_uimm12, 4770 !subst(ALIGN, any_align, decls.pattern))), 4771 i8>; 4772 defm : ls_small_pats<LS16_LDR, LS16_STR, Base, 4773 !foreach(decls.pattern, Offset, 4774 !subst(OFFSET, hword_uimm12, decls.pattern)), 4775 !foreach(decls.pattern, address, 4776 !subst(OFFSET, hword_uimm12, 4777 !subst(ALIGN, min_align2, decls.pattern))), 4778 i16>; 4779 defm : ls_small_pats<LS32_LDR, LS32_STR, Base, 4780 !foreach(decls.pattern, Offset, 4781 !subst(OFFSET, word_uimm12, decls.pattern)), 4782 !foreach(decls.pattern, address, 4783 !subst(OFFSET, word_uimm12, 4784 !subst(ALIGN, min_align4, decls.pattern))), 4785 i32>; 4786 4787 defm : ls_int_neutral_pats<LS32_LDR, LS32_STR, Base, 4788 !foreach(decls.pattern, Offset, 4789 !subst(OFFSET, word_uimm12, decls.pattern)), 4790 !foreach(decls.pattern, address, 4791 !subst(OFFSET, word_uimm12, 4792 !subst(ALIGN, min_align4, decls.pattern))), 4793 GPR32, i32>; 4794 4795 defm : ls_int_neutral_pats<LS64_LDR, LS64_STR, Base, 4796 !foreach(decls.pattern, Offset, 4797 !subst(OFFSET, dword_uimm12, decls.pattern)), 4798 !foreach(decls.pattern, address, 4799 !subst(OFFSET, dword_uimm12, 4800 !subst(ALIGN, min_align8, decls.pattern))), 4801 GPR64, i64>; 4802 4803 defm : ls_neutral_pats<LSFP16_LDR, LSFP16_STR, Base, 4804 !foreach(decls.pattern, Offset, 4805 !subst(OFFSET, hword_uimm12, decls.pattern)), 4806 !foreach(decls.pattern, address, 4807 !subst(OFFSET, hword_uimm12, 4808 !subst(ALIGN, min_align2, decls.pattern))), 4809 FPR16, f16>; 4810 4811 defm : ls_neutral_pats<LSFP32_LDR, LSFP32_STR, Base, 4812 !foreach(decls.pattern, Offset, 4813 !subst(OFFSET, word_uimm12, decls.pattern)), 4814 !foreach(decls.pattern, address, 4815 !subst(OFFSET, word_uimm12, 4816 !subst(ALIGN, min_align4, decls.pattern))), 4817 FPR32, f32>; 4818 4819 defm : ls_neutral_pats<LSFP64_LDR, LSFP64_STR, Base, 4820 !foreach(decls.pattern, Offset, 4821 !subst(OFFSET, dword_uimm12, decls.pattern)), 4822 !foreach(decls.pattern, address, 4823 !subst(OFFSET, dword_uimm12, 4824 !subst(ALIGN, min_align8, decls.pattern))), 4825 FPR64, f64>; 4826 4827 defm : ls_neutral_pats<LSFP128_LDR, LSFP128_STR, Base, 4828 !foreach(decls.pattern, Offset, 4829 !subst(OFFSET, qword_uimm12, decls.pattern)), 4830 !foreach(decls.pattern, address, 4831 !subst(OFFSET, qword_uimm12, 4832 !subst(ALIGN, min_align16, decls.pattern))), 4833 FPR128, f128>; 4834 4835 defm : load_signed_pats<"B", "", Base, 4836 !foreach(decls.pattern, Offset, 4837 !subst(OFFSET, byte_uimm12, decls.pattern)), 4838 !foreach(decls.pattern, address, 4839 !subst(OFFSET, byte_uimm12, 4840 !subst(ALIGN, any_align, decls.pattern))), 4841 i8>; 4842 4843 defm : load_signed_pats<"H", "", Base, 4844 !foreach(decls.pattern, Offset, 4845 !subst(OFFSET, hword_uimm12, decls.pattern)), 4846 !foreach(decls.pattern, address, 4847 !subst(OFFSET, hword_uimm12, 4848 !subst(ALIGN, min_align2, decls.pattern))), 4849 i16>; 4850 4851 def : Pat<(sextloadi32 !foreach(decls.pattern, address, 4852 !subst(OFFSET, word_uimm12, 4853 !subst(ALIGN, min_align4, decls.pattern)))), 4854 (LDRSWx Base, !foreach(decls.pattern, Offset, 4855 !subst(OFFSET, word_uimm12, decls.pattern)))>; 4856 } 4857 4858 // Straightforward patterns of last resort: a pointer with or without an 4859 // appropriate offset. 4860 defm : uimm12_pats<(i64 GPR64xsp:$Rn), (i64 GPR64xsp:$Rn), (i64 0)>; 4861 defm : uimm12_pats<(add GPR64xsp:$Rn, OFFSET:$UImm12), 4862 (i64 GPR64xsp:$Rn), (i64 OFFSET:$UImm12)>; 4863 4864 // The offset could be hidden behind an "or", of course: 4865 defm : uimm12_pats<(add_like_or GPR64xsp:$Rn, OFFSET:$UImm12), 4866 (i64 GPR64xsp:$Rn), (i64 OFFSET:$UImm12)>; 4867 4868 // Global addresses under the small-absolute model should use these 4869 // instructions. There are ELF relocations specifically for it. 4870 defm : uimm12_pats<(A64WrapperSmall tglobaladdr:$Hi, tglobaladdr:$Lo12, ALIGN), 4871 (ADRPxi tglobaladdr:$Hi), (i64 tglobaladdr:$Lo12)>; 4872 4873 defm : uimm12_pats<(A64WrapperSmall tglobaltlsaddr:$Hi, tglobaltlsaddr:$Lo12, 4874 ALIGN), 4875 (ADRPxi tglobaltlsaddr:$Hi), (i64 tglobaltlsaddr:$Lo12)>; 4876 4877 // External symbols that make it this far should also get standard relocations. 4878 defm : uimm12_pats<(A64WrapperSmall texternalsym:$Hi, texternalsym:$Lo12, 4879 ALIGN), 4880 (ADRPxi texternalsym:$Hi), (i64 texternalsym:$Lo12)>; 4881 4882 defm : uimm12_pats<(A64WrapperSmall tconstpool:$Hi, tconstpool:$Lo12, ALIGN), 4883 (ADRPxi tconstpool:$Hi), (i64 tconstpool:$Lo12)>; 4884 4885 // We also want to use uimm12 instructions for local variables at the moment. 4886 def tframeindex_XFORM : SDNodeXForm<frameindex, [{ 4887 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 4888 return CurDAG->getTargetFrameIndex(FI, MVT::i64); 4889 }]>; 4890 4891 defm : uimm12_pats<(i64 frameindex:$Rn), 4892 (tframeindex_XFORM tframeindex:$Rn), (i64 0)>; 4893 4894 // These can be much simpler than uimm12 because we don't to change the operand 4895 // type (e.g. LDURB and LDURH take the same operands). 4896 multiclass simm9_pats<dag address, dag Base, dag Offset> { 4897 defm : ls_small_pats<LS8_LDUR, LS8_STUR, Base, Offset, address, i8>; 4898 defm : ls_small_pats<LS16_LDUR, LS16_STUR, Base, Offset, address, i16>; 4899 4900 defm : ls_int_neutral_pats<LS32_LDUR, LS32_STUR, Base, Offset, address, 4901 GPR32, i32>; 4902 defm : ls_int_neutral_pats<LS64_LDUR, LS64_STUR, Base, Offset, address, 4903 GPR64, i64>; 4904 4905 defm : ls_neutral_pats<LSFP16_LDUR, LSFP16_STUR, Base, Offset, address, 4906 FPR16, f16>; 4907 defm : ls_neutral_pats<LSFP32_LDUR, LSFP32_STUR, Base, Offset, address, 4908 FPR32, f32>; 4909 defm : ls_neutral_pats<LSFP64_LDUR, LSFP64_STUR, Base, Offset, address, 4910 FPR64, f64>; 4911 defm : ls_neutral_pats<LSFP128_LDUR, LSFP128_STUR, Base, Offset, address, 4912 FPR128, f128>; 4913 4914 def : Pat<(i64 (zextloadi32 address)), 4915 (SUBREG_TO_REG (i64 0), (LS32_LDUR Base, Offset), sub_32)>; 4916 4917 def : Pat<(truncstorei32 GPR64:$Rt, address), 4918 (LS32_STUR (EXTRACT_SUBREG GPR64:$Rt, sub_32), Base, Offset)>; 4919 4920 defm : load_signed_pats<"B", "_U", Base, Offset, address, i8>; 4921 defm : load_signed_pats<"H", "_U", Base, Offset, address, i16>; 4922 def : Pat<(sextloadi32 address), (LDURSWx Base, Offset)>; 4923 } 4924 4925 defm : simm9_pats<(add GPR64xsp:$Rn, simm9:$SImm9), 4926 (i64 GPR64xsp:$Rn), (SDXF_simm9 simm9:$SImm9)>; 4927 4928 defm : simm9_pats<(add_like_or GPR64xsp:$Rn, simm9:$SImm9), 4929 (i64 GPR64xsp:$Rn), (SDXF_simm9 simm9:$SImm9)>; 4930 4931 4932 //===------------------------------ 4933 // 3. Register offset patterns 4934 //===------------------------------ 4935 4936 // Atomic patterns can be shared between integer operations of all sizes, a 4937 // quick multiclass here allows reuse. 4938 multiclass ro_atomic_pats<Instruction LOAD, Instruction STORE, dag Base, 4939 dag Offset, dag Extend, dag address, 4940 RegisterClass TPR, ValueType sty> { 4941 def : Pat<(!cast<PatFrag>("atomic_load_simple_" # sty) address), 4942 (LOAD Base, Offset, Extend)>; 4943 4944 def : Pat<(!cast<PatFrag>("atomic_store_simple_" # sty) address, TPR:$Rt), 4945 (STORE TPR:$Rt, Base, Offset, Extend)>; 4946 } 4947 4948 // The register offset instructions take three operands giving the instruction, 4949 // and have an annoying split between instructions where Rm is 32-bit and 4950 // 64-bit. So we need a special hierarchy to describe them. Other than that the 4951 // same operations should be supported as for simm9 and uimm12 addressing. 4952 4953 multiclass ro_small_pats<Instruction LOAD, Instruction STORE, 4954 dag Base, dag Offset, dag Extend, 4955 dag address, ValueType sty> 4956 : ro_atomic_pats<LOAD, STORE, Base, Offset, Extend, address, GPR32, sty> { 4957 def : Pat<(!cast<SDNode>(zextload # sty) address), 4958 (LOAD Base, Offset, Extend)>; 4959 4960 def : Pat<(!cast<SDNode>(extload # sty) address), 4961 (LOAD Base, Offset, Extend)>; 4962 4963 // For zero-extension to 64-bits we have to tell LLVM that the whole 64-bit 4964 // register was actually set. 4965 def : Pat<(i64 (!cast<SDNode>(zextload # sty) address)), 4966 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset, Extend), sub_32)>; 4967 4968 def : Pat<(i64 (!cast<SDNode>(extload # sty) address)), 4969 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset, Extend), sub_32)>; 4970 4971 def : Pat<(!cast<SDNode>(truncstore # sty) GPR32:$Rt, address), 4972 (STORE GPR32:$Rt, Base, Offset, Extend)>; 4973 4974 // For truncating store from 64-bits, we have to manually tell LLVM to 4975 // ignore the high bits of the x register. 4976 def : Pat<(!cast<SDNode>(truncstore # sty) GPR64:$Rt, address), 4977 (STORE (EXTRACT_SUBREG GPR64:$Rt, sub_32), Base, Offset, Extend)>; 4978 4979 } 4980 4981 // Next come patterns for sign-extending loads. 4982 multiclass ro_signed_pats<string T, string Rm, dag Base, dag Offset, dag Extend, 4983 dag address, ValueType sty> { 4984 def : Pat<(i32 (!cast<SDNode>("sextload" # sty) address)), 4985 (!cast<Instruction>("LDRS" # T # "w_" # Rm # "_RegOffset") 4986 Base, Offset, Extend)>; 4987 4988 def : Pat<(i64 (!cast<SDNode>("sextload" # sty) address)), 4989 (!cast<Instruction>("LDRS" # T # "x_" # Rm # "_RegOffset") 4990 Base, Offset, Extend)>; 4991 } 4992 4993 // and finally "natural-width" loads and stores come next. 4994 multiclass ro_neutral_pats<Instruction LOAD, Instruction STORE, 4995 dag Base, dag Offset, dag Extend, dag address, 4996 RegisterClass TPR, ValueType sty> { 4997 def : Pat<(sty (load address)), (LOAD Base, Offset, Extend)>; 4998 def : Pat<(store (sty TPR:$Rt), address), 4999 (STORE TPR:$Rt, Base, Offset, Extend)>; 5000 } 5001 5002 multiclass ro_int_neutral_pats<Instruction LOAD, Instruction STORE, 5003 dag Base, dag Offset, dag Extend, dag address, 5004 RegisterClass TPR, ValueType sty> 5005 : ro_neutral_pats<LOAD, STORE, Base, Offset, Extend, address, TPR, sty>, 5006 ro_atomic_pats<LOAD, STORE, Base, Offset, Extend, address, TPR, sty>; 5007 5008 multiclass regoff_pats<string Rm, dag address, dag Base, dag Offset, 5009 dag Extend> { 5010 defm : ro_small_pats<!cast<Instruction>("LS8_" # Rm # "_RegOffset_LDR"), 5011 !cast<Instruction>("LS8_" # Rm # "_RegOffset_STR"), 5012 Base, Offset, Extend, 5013 !foreach(decls.pattern, address, 5014 !subst(SHIFT, imm_eq0, decls.pattern)), 5015 i8>; 5016 defm : ro_small_pats<!cast<Instruction>("LS16_" # Rm # "_RegOffset_LDR"), 5017 !cast<Instruction>("LS16_" # Rm # "_RegOffset_STR"), 5018 Base, Offset, Extend, 5019 !foreach(decls.pattern, address, 5020 !subst(SHIFT, imm_eq1, decls.pattern)), 5021 i16>; 5022 defm : ro_small_pats<!cast<Instruction>("LS32_" # Rm # "_RegOffset_LDR"), 5023 !cast<Instruction>("LS32_" # Rm # "_RegOffset_STR"), 5024 Base, Offset, Extend, 5025 !foreach(decls.pattern, address, 5026 !subst(SHIFT, imm_eq2, decls.pattern)), 5027 i32>; 5028 5029 defm : ro_int_neutral_pats< 5030 !cast<Instruction>("LS32_" # Rm # "_RegOffset_LDR"), 5031 !cast<Instruction>("LS32_" # Rm # "_RegOffset_STR"), 5032 Base, Offset, Extend, 5033 !foreach(decls.pattern, address, 5034 !subst(SHIFT, imm_eq2, decls.pattern)), 5035 GPR32, i32>; 5036 5037 defm : ro_int_neutral_pats< 5038 !cast<Instruction>("LS64_" # Rm # "_RegOffset_LDR"), 5039 !cast<Instruction>("LS64_" # Rm # "_RegOffset_STR"), 5040 Base, Offset, Extend, 5041 !foreach(decls.pattern, address, 5042 !subst(SHIFT, imm_eq3, decls.pattern)), 5043 GPR64, i64>; 5044 5045 defm : ro_neutral_pats<!cast<Instruction>("LSFP16_" # Rm # "_RegOffset_LDR"), 5046 !cast<Instruction>("LSFP16_" # Rm # "_RegOffset_STR"), 5047 Base, Offset, Extend, 5048 !foreach(decls.pattern, address, 5049 !subst(SHIFT, imm_eq1, decls.pattern)), 5050 FPR16, f16>; 5051 5052 defm : ro_neutral_pats<!cast<Instruction>("LSFP32_" # Rm # "_RegOffset_LDR"), 5053 !cast<Instruction>("LSFP32_" # Rm # "_RegOffset_STR"), 5054 Base, Offset, Extend, 5055 !foreach(decls.pattern, address, 5056 !subst(SHIFT, imm_eq2, decls.pattern)), 5057 FPR32, f32>; 5058 5059 defm : ro_neutral_pats<!cast<Instruction>("LSFP64_" # Rm # "_RegOffset_LDR"), 5060 !cast<Instruction>("LSFP64_" # Rm # "_RegOffset_STR"), 5061 Base, Offset, Extend, 5062 !foreach(decls.pattern, address, 5063 !subst(SHIFT, imm_eq3, decls.pattern)), 5064 FPR64, f64>; 5065 5066 defm : ro_neutral_pats<!cast<Instruction>("LSFP128_" # Rm # "_RegOffset_LDR"), 5067 !cast<Instruction>("LSFP128_" # Rm # "_RegOffset_STR"), 5068 Base, Offset, Extend, 5069 !foreach(decls.pattern, address, 5070 !subst(SHIFT, imm_eq4, decls.pattern)), 5071 FPR128, f128>; 5072 5073 defm : ro_signed_pats<"B", Rm, Base, Offset, Extend, 5074 !foreach(decls.pattern, address, 5075 !subst(SHIFT, imm_eq0, decls.pattern)), 5076 i8>; 5077 5078 defm : ro_signed_pats<"H", Rm, Base, Offset, Extend, 5079 !foreach(decls.pattern, address, 5080 !subst(SHIFT, imm_eq1, decls.pattern)), 5081 i16>; 5082 5083 def : Pat<(sextloadi32 !foreach(decls.pattern, address, 5084 !subst(SHIFT, imm_eq2, decls.pattern))), 5085 (!cast<Instruction>("LDRSWx_" # Rm # "_RegOffset") 5086 Base, Offset, Extend)>; 5087 } 5088 5089 5090 // Finally we're in a position to tell LLVM exactly what addresses are reachable 5091 // using register-offset instructions. Essentially a base plus a possibly 5092 // extended, possibly shifted (by access size) offset. 5093 5094 defm : regoff_pats<"Wm", (add GPR64xsp:$Rn, (sext GPR32:$Rm)), 5095 (i64 GPR64xsp:$Rn), (i32 GPR32:$Rm), (i64 6)>; 5096 5097 defm : regoff_pats<"Wm", (add GPR64xsp:$Rn, (shl (sext GPR32:$Rm), SHIFT)), 5098 (i64 GPR64xsp:$Rn), (i32 GPR32:$Rm), (i64 7)>; 5099 5100 defm : regoff_pats<"Wm", (add GPR64xsp:$Rn, (zext GPR32:$Rm)), 5101 (i64 GPR64xsp:$Rn), (i32 GPR32:$Rm), (i64 2)>; 5102 5103 defm : regoff_pats<"Wm", (add GPR64xsp:$Rn, (shl (zext GPR32:$Rm), SHIFT)), 5104 (i64 GPR64xsp:$Rn), (i32 GPR32:$Rm), (i64 3)>; 5105 5106 defm : regoff_pats<"Xm", (add GPR64xsp:$Rn, GPR64:$Rm), 5107 (i64 GPR64xsp:$Rn), (i64 GPR64:$Rm), (i64 2)>; 5108 5109 defm : regoff_pats<"Xm", (add GPR64xsp:$Rn, (shl GPR64:$Rm, SHIFT)), 5110 (i64 GPR64xsp:$Rn), (i64 GPR64:$Rm), (i64 3)>; 5111