ARMRegisterInfo.td revision ac096808a3accc516ae7c193c9a2c1392bf3301a
1//===- ARMRegisterInfo.td - ARM Register defs -------------------*- C++ -*-===// 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//===----------------------------------------------------------------------===// 11// Declarations that describe the ARM register file 12//===----------------------------------------------------------------------===// 13 14// Registers are identified with 4-bit ID numbers. 15class ARMReg<bits<4> num, string n, list<Register> subregs = []> : Register<n> { 16 field bits<4> Num; 17 let Namespace = "ARM"; 18 let SubRegs = subregs; 19} 20 21class ARMFReg<bits<6> num, string n> : Register<n> { 22 field bits<6> Num; 23 let Namespace = "ARM"; 24} 25 26// Subregister indices. 27let Namespace = "ARM" in { 28// Note: Code depends on these having consecutive numbers. 29def ssub_0 : SubRegIndex; 30def ssub_1 : SubRegIndex; 31def ssub_2 : SubRegIndex; // In a Q reg. 32def ssub_3 : SubRegIndex; 33def ssub_4 : SubRegIndex; // In a QQ reg. 34def ssub_5 : SubRegIndex; 35def ssub_6 : SubRegIndex; 36def ssub_7 : SubRegIndex; 37def ssub_8 : SubRegIndex; // In a QQQQ reg. 38def ssub_9 : SubRegIndex; 39def ssub_10 : SubRegIndex; 40def ssub_11 : SubRegIndex; 41def ssub_12 : SubRegIndex; 42def ssub_13 : SubRegIndex; 43def ssub_14 : SubRegIndex; 44def ssub_15 : SubRegIndex; 45 46def dsub_0 : SubRegIndex; 47def dsub_1 : SubRegIndex; 48def dsub_2 : SubRegIndex; 49def dsub_3 : SubRegIndex; 50def dsub_4 : SubRegIndex; 51def dsub_5 : SubRegIndex; 52def dsub_6 : SubRegIndex; 53def dsub_7 : SubRegIndex; 54 55def qsub_0 : SubRegIndex; 56def qsub_1 : SubRegIndex; 57def qsub_2 : SubRegIndex; 58def qsub_3 : SubRegIndex; 59 60def qqsub_0 : SubRegIndex; 61def qqsub_1 : SubRegIndex; 62} 63 64// Integer registers 65def R0 : ARMReg< 0, "r0">, DwarfRegNum<[0]>; 66def R1 : ARMReg< 1, "r1">, DwarfRegNum<[1]>; 67def R2 : ARMReg< 2, "r2">, DwarfRegNum<[2]>; 68def R3 : ARMReg< 3, "r3">, DwarfRegNum<[3]>; 69def R4 : ARMReg< 4, "r4">, DwarfRegNum<[4]>; 70def R5 : ARMReg< 5, "r5">, DwarfRegNum<[5]>; 71def R6 : ARMReg< 6, "r6">, DwarfRegNum<[6]>; 72def R7 : ARMReg< 7, "r7">, DwarfRegNum<[7]>; 73def R8 : ARMReg< 8, "r8">, DwarfRegNum<[8]>; 74def R9 : ARMReg< 9, "r9">, DwarfRegNum<[9]>; 75def R10 : ARMReg<10, "r10">, DwarfRegNum<[10]>; 76def R11 : ARMReg<11, "r11">, DwarfRegNum<[11]>; 77def R12 : ARMReg<12, "r12">, DwarfRegNum<[12]>; 78def SP : ARMReg<13, "sp">, DwarfRegNum<[13]>; 79def LR : ARMReg<14, "lr">, DwarfRegNum<[14]>; 80def PC : ARMReg<15, "pc">, DwarfRegNum<[15]>; 81 82// Float registers 83def S0 : ARMFReg< 0, "s0">; def S1 : ARMFReg< 1, "s1">; 84def S2 : ARMFReg< 2, "s2">; def S3 : ARMFReg< 3, "s3">; 85def S4 : ARMFReg< 4, "s4">; def S5 : ARMFReg< 5, "s5">; 86def S6 : ARMFReg< 6, "s6">; def S7 : ARMFReg< 7, "s7">; 87def S8 : ARMFReg< 8, "s8">; def S9 : ARMFReg< 9, "s9">; 88def S10 : ARMFReg<10, "s10">; def S11 : ARMFReg<11, "s11">; 89def S12 : ARMFReg<12, "s12">; def S13 : ARMFReg<13, "s13">; 90def S14 : ARMFReg<14, "s14">; def S15 : ARMFReg<15, "s15">; 91def S16 : ARMFReg<16, "s16">; def S17 : ARMFReg<17, "s17">; 92def S18 : ARMFReg<18, "s18">; def S19 : ARMFReg<19, "s19">; 93def S20 : ARMFReg<20, "s20">; def S21 : ARMFReg<21, "s21">; 94def S22 : ARMFReg<22, "s22">; def S23 : ARMFReg<23, "s23">; 95def S24 : ARMFReg<24, "s24">; def S25 : ARMFReg<25, "s25">; 96def S26 : ARMFReg<26, "s26">; def S27 : ARMFReg<27, "s27">; 97def S28 : ARMFReg<28, "s28">; def S29 : ARMFReg<29, "s29">; 98def S30 : ARMFReg<30, "s30">; def S31 : ARMFReg<31, "s31">; 99 100// Aliases of the F* registers used to hold 64-bit fp values (doubles) 101let SubRegIndices = [ssub_0, ssub_1] in { 102def D0 : ARMReg< 0, "d0", [S0, S1]>; 103def D1 : ARMReg< 1, "d1", [S2, S3]>; 104def D2 : ARMReg< 2, "d2", [S4, S5]>; 105def D3 : ARMReg< 3, "d3", [S6, S7]>; 106def D4 : ARMReg< 4, "d4", [S8, S9]>; 107def D5 : ARMReg< 5, "d5", [S10, S11]>; 108def D6 : ARMReg< 6, "d6", [S12, S13]>; 109def D7 : ARMReg< 7, "d7", [S14, S15]>; 110def D8 : ARMReg< 8, "d8", [S16, S17]>; 111def D9 : ARMReg< 9, "d9", [S18, S19]>; 112def D10 : ARMReg<10, "d10", [S20, S21]>; 113def D11 : ARMReg<11, "d11", [S22, S23]>; 114def D12 : ARMReg<12, "d12", [S24, S25]>; 115def D13 : ARMReg<13, "d13", [S26, S27]>; 116def D14 : ARMReg<14, "d14", [S28, S29]>; 117def D15 : ARMReg<15, "d15", [S30, S31]>; 118} 119 120// VFP3 defines 16 additional double registers 121def D16 : ARMFReg<16, "d16">; def D17 : ARMFReg<17, "d17">; 122def D18 : ARMFReg<18, "d18">; def D19 : ARMFReg<19, "d19">; 123def D20 : ARMFReg<20, "d20">; def D21 : ARMFReg<21, "d21">; 124def D22 : ARMFReg<22, "d22">; def D23 : ARMFReg<23, "d23">; 125def D24 : ARMFReg<24, "d24">; def D25 : ARMFReg<25, "d25">; 126def D26 : ARMFReg<26, "d26">; def D27 : ARMFReg<27, "d27">; 127def D28 : ARMFReg<28, "d28">; def D29 : ARMFReg<29, "d29">; 128def D30 : ARMFReg<30, "d30">; def D31 : ARMFReg<31, "d31">; 129 130// Advanced SIMD (NEON) defines 16 quad-word aliases 131let SubRegIndices = [dsub_0, dsub_1], 132 CompositeIndices = [(ssub_2 dsub_1, ssub_0), 133 (ssub_3 dsub_1, ssub_1)] in { 134def Q0 : ARMReg< 0, "q0", [D0, D1]>; 135def Q1 : ARMReg< 1, "q1", [D2, D3]>; 136def Q2 : ARMReg< 2, "q2", [D4, D5]>; 137def Q3 : ARMReg< 3, "q3", [D6, D7]>; 138def Q4 : ARMReg< 4, "q4", [D8, D9]>; 139def Q5 : ARMReg< 5, "q5", [D10, D11]>; 140def Q6 : ARMReg< 6, "q6", [D12, D13]>; 141def Q7 : ARMReg< 7, "q7", [D14, D15]>; 142} 143let SubRegIndices = [dsub_0, dsub_1] in { 144def Q8 : ARMReg< 8, "q8", [D16, D17]>; 145def Q9 : ARMReg< 9, "q9", [D18, D19]>; 146def Q10 : ARMReg<10, "q10", [D20, D21]>; 147def Q11 : ARMReg<11, "q11", [D22, D23]>; 148def Q12 : ARMReg<12, "q12", [D24, D25]>; 149def Q13 : ARMReg<13, "q13", [D26, D27]>; 150def Q14 : ARMReg<14, "q14", [D28, D29]>; 151def Q15 : ARMReg<15, "q15", [D30, D31]>; 152} 153 154// Pseudo 256-bit registers to represent pairs of Q registers. These should 155// never be present in the emitted code. 156// These are used for NEON load / store instructions, e.g., vld4, vst3. 157// NOTE: It's possible to define more QQ registers since technically the 158// starting D register number doesn't have to be multiple of 4, e.g., 159// D1, D2, D3, D4 would be a legal quad, but that would make the subregister 160// stuff very messy. 161let SubRegIndices = [qsub_0, qsub_1] in { 162let CompositeIndices = [(dsub_2 qsub_1, dsub_0), (dsub_3 qsub_1, dsub_1), 163 (ssub_4 qsub_1, ssub_0), (ssub_5 qsub_1, ssub_1), 164 (ssub_6 qsub_1, ssub_2), (ssub_7 qsub_1, ssub_3)] in { 165def QQ0 : ARMReg<0, "qq0", [Q0, Q1]>; 166def QQ1 : ARMReg<1, "qq1", [Q2, Q3]>; 167def QQ2 : ARMReg<2, "qq2", [Q4, Q5]>; 168def QQ3 : ARMReg<3, "qq3", [Q6, Q7]>; 169} 170let CompositeIndices = [(dsub_2 qsub_1, dsub_0), (dsub_3 qsub_1, dsub_1)] in { 171def QQ4 : ARMReg<4, "qq4", [Q8, Q9]>; 172def QQ5 : ARMReg<5, "qq5", [Q10, Q11]>; 173def QQ6 : ARMReg<6, "qq6", [Q12, Q13]>; 174def QQ7 : ARMReg<7, "qq7", [Q14, Q15]>; 175} 176} 177 178// Pseudo 512-bit registers to represent four consecutive Q registers. 179let SubRegIndices = [qqsub_0, qqsub_1] in { 180let CompositeIndices = [(qsub_2 qqsub_1, qsub_0), (qsub_3 qqsub_1, qsub_1), 181 (dsub_4 qqsub_1, dsub_0), (dsub_5 qqsub_1, dsub_1), 182 (dsub_6 qqsub_1, dsub_2), (dsub_7 qqsub_1, dsub_3), 183 (ssub_8 qqsub_1, ssub_0), (ssub_9 qqsub_1, ssub_1), 184 (ssub_10 qqsub_1, ssub_2), (ssub_11 qqsub_1, ssub_3), 185 (ssub_12 qqsub_1, ssub_4), (ssub_13 qqsub_1, ssub_5), 186 (ssub_14 qqsub_1, ssub_6), (ssub_15 qqsub_1, ssub_7)] in 187{ 188def QQQQ0 : ARMReg<0, "qqqq0", [QQ0, QQ1]>; 189def QQQQ1 : ARMReg<1, "qqqq1", [QQ2, QQ3]>; 190} 191let CompositeIndices = [(qsub_2 qqsub_1, qsub_0), (qsub_3 qqsub_1, qsub_1), 192 (dsub_4 qqsub_1, dsub_0), (dsub_5 qqsub_1, dsub_1), 193 (dsub_6 qqsub_1, dsub_2), (dsub_7 qqsub_1, dsub_3)] in { 194def QQQQ2 : ARMReg<2, "qqqq2", [QQ4, QQ5]>; 195def QQQQ3 : ARMReg<3, "qqqq3", [QQ6, QQ7]>; 196} 197} 198 199// Current Program Status Register. 200def CPSR : ARMReg<0, "cpsr">; 201def FPSCR : ARMReg<1, "fpscr">; 202def ITSTATE : ARMReg<2, "itstate">; 203 204// Register classes. 205// 206// pc == Program Counter 207// lr == Link Register 208// sp == Stack Pointer 209// r12 == ip (scratch) 210// r7 == Frame Pointer (thumb-style backtraces) 211// r9 == May be reserved as Thread Register 212// r11 == Frame Pointer (arm-style backtraces) 213// r10 == Stack Limit 214// 215def GPR : RegisterClass<"ARM", [i32], 32, [R0, R1, R2, R3, R4, R5, R6, 216 R7, R8, R9, R10, R11, R12, 217 SP, LR, PC]> { 218 let MethodProtos = [{ 219 iterator allocation_order_begin(const MachineFunction &MF) const; 220 iterator allocation_order_end(const MachineFunction &MF) const; 221 }]; 222 let MethodBodies = [{ 223 // FP is R11, R9 is available. 224 static const unsigned ARM_GPR_AO_1[] = { 225 ARM::R0, ARM::R1, ARM::R2, ARM::R3, 226 ARM::R12,ARM::LR, 227 ARM::R4, ARM::R5, ARM::R6, ARM::R7, 228 ARM::R8, ARM::R9, ARM::R10, 229 ARM::R11 }; 230 // FP is R11, R9 is not available. 231 static const unsigned ARM_GPR_AO_2[] = { 232 ARM::R0, ARM::R1, ARM::R2, ARM::R3, 233 ARM::R12,ARM::LR, 234 ARM::R4, ARM::R5, ARM::R6, ARM::R7, 235 ARM::R8, ARM::R10, 236 ARM::R11 }; 237 // FP is R7, R9 is available as non-callee-saved register. 238 // This is used by Darwin. 239 static const unsigned ARM_GPR_AO_3[] = { 240 ARM::R0, ARM::R1, ARM::R2, ARM::R3, 241 ARM::R9, ARM::R12,ARM::LR, 242 ARM::R4, ARM::R5, ARM::R6, 243 ARM::R8, ARM::R10,ARM::R11,ARM::R7 }; 244 // FP is R7, R9 is not available. 245 static const unsigned ARM_GPR_AO_4[] = { 246 ARM::R0, ARM::R1, ARM::R2, ARM::R3, 247 ARM::R12,ARM::LR, 248 ARM::R4, ARM::R5, ARM::R6, 249 ARM::R8, ARM::R10,ARM::R11, 250 ARM::R7 }; 251 // FP is R7, R9 is available as callee-saved register. 252 // This is used by non-Darwin platform in Thumb mode. 253 static const unsigned ARM_GPR_AO_5[] = { 254 ARM::R0, ARM::R1, ARM::R2, ARM::R3, 255 ARM::R12,ARM::LR, 256 ARM::R4, ARM::R5, ARM::R6, 257 ARM::R8, ARM::R9, ARM::R10,ARM::R11,ARM::R7 }; 258 259 // For Thumb1 mode, we don't want to allocate hi regs at all, as we 260 // don't know how to spill them. If we make our prologue/epilogue code 261 // smarter at some point, we can go back to using the above allocation 262 // orders for the Thumb1 instructions that know how to use hi regs. 263 static const unsigned THUMB_GPR_AO[] = { 264 ARM::R0, ARM::R1, ARM::R2, ARM::R3, 265 ARM::R4, ARM::R5, ARM::R6, ARM::R7 }; 266 267 GPRClass::iterator 268 GPRClass::allocation_order_begin(const MachineFunction &MF) const { 269 const TargetMachine &TM = MF.getTarget(); 270 const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>(); 271 if (Subtarget.isThumb1Only()) 272 return THUMB_GPR_AO; 273 if (Subtarget.isTargetDarwin()) { 274 if (Subtarget.isR9Reserved()) 275 return ARM_GPR_AO_4; 276 else 277 return ARM_GPR_AO_3; 278 } else { 279 if (Subtarget.isR9Reserved()) 280 return ARM_GPR_AO_2; 281 else if (Subtarget.isThumb()) 282 return ARM_GPR_AO_5; 283 else 284 return ARM_GPR_AO_1; 285 } 286 } 287 288 GPRClass::iterator 289 GPRClass::allocation_order_end(const MachineFunction &MF) const { 290 const TargetMachine &TM = MF.getTarget(); 291 const TargetRegisterInfo *RI = TM.getRegisterInfo(); 292 const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>(); 293 GPRClass::iterator I; 294 295 if (Subtarget.isThumb1Only()) { 296 I = THUMB_GPR_AO + (sizeof(THUMB_GPR_AO)/sizeof(unsigned)); 297 return RI->hasFP(MF) ? I-1 : I; 298 } 299 300 if (Subtarget.isTargetDarwin()) { 301 if (Subtarget.isR9Reserved()) 302 I = ARM_GPR_AO_4 + (sizeof(ARM_GPR_AO_4)/sizeof(unsigned)); 303 else 304 I = ARM_GPR_AO_3 + (sizeof(ARM_GPR_AO_3)/sizeof(unsigned)); 305 } else { 306 if (Subtarget.isR9Reserved()) 307 I = ARM_GPR_AO_2 + (sizeof(ARM_GPR_AO_2)/sizeof(unsigned)); 308 else if (Subtarget.isThumb()) 309 I = ARM_GPR_AO_5 + (sizeof(ARM_GPR_AO_5)/sizeof(unsigned)); 310 else 311 I = ARM_GPR_AO_1 + (sizeof(ARM_GPR_AO_1)/sizeof(unsigned)); 312 } 313 314 return RI->hasFP(MF) ? I-1 : I; 315 } 316 }]; 317} 318 319// restricted GPR register class. Many Thumb2 instructions allow the full 320// register range for operands, but have undefined behaviours when PC 321// or SP (R13 or R15) are used. The ARM ARM refers to these operands 322// via the BadReg() pseudo-code description. 323def rGPR : RegisterClass<"ARM", [i32], 32, [R0, R1, R2, R3, R4, R5, R6, 324 R7, R8, R9, R10, R11, R12, LR]> { 325 let MethodProtos = [{ 326 iterator allocation_order_begin(const MachineFunction &MF) const; 327 iterator allocation_order_end(const MachineFunction &MF) const; 328 }]; 329 let MethodBodies = [{ 330 // FP is R11, R9 is available. 331 static const unsigned ARM_rGPRAO_1[] = { 332 ARM::R0, ARM::R1, ARM::R2, ARM::R3, 333 ARM::R12,ARM::LR, 334 ARM::R4, ARM::R5, ARM::R6, ARM::R7, 335 ARM::R8, ARM::R9, ARM::R10, 336 ARM::R11 }; 337 // FP is R11, R9 is not available. 338 static const unsigned ARM_rGPRAO_2[] = { 339 ARM::R0, ARM::R1, ARM::R2, ARM::R3, 340 ARM::R12,ARM::LR, 341 ARM::R4, ARM::R5, ARM::R6, ARM::R7, 342 ARM::R8, ARM::R10, 343 ARM::R11 }; 344 // FP is R7, R9 is available as non-callee-saved register. 345 // This is used by Darwin. 346 static const unsigned ARM_rGPRAO_3[] = { 347 ARM::R0, ARM::R1, ARM::R2, ARM::R3, 348 ARM::R9, ARM::R12,ARM::LR, 349 ARM::R4, ARM::R5, ARM::R6, 350 ARM::R8, ARM::R10,ARM::R11,ARM::R7 }; 351 // FP is R7, R9 is not available. 352 static const unsigned ARM_rGPRAO_4[] = { 353 ARM::R0, ARM::R1, ARM::R2, ARM::R3, 354 ARM::R12,ARM::LR, 355 ARM::R4, ARM::R5, ARM::R6, 356 ARM::R8, ARM::R10,ARM::R11, 357 ARM::R7 }; 358 // FP is R7, R9 is available as callee-saved register. 359 // This is used by non-Darwin platform in Thumb mode. 360 static const unsigned ARM_rGPRAO_5[] = { 361 ARM::R0, ARM::R1, ARM::R2, ARM::R3, 362 ARM::R12,ARM::LR, 363 ARM::R4, ARM::R5, ARM::R6, 364 ARM::R8, ARM::R9, ARM::R10,ARM::R11,ARM::R7 }; 365 366 // For Thumb1 mode, we don't want to allocate hi regs at all, as we 367 // don't know how to spill them. If we make our prologue/epilogue code 368 // smarter at some point, we can go back to using the above allocation 369 // orders for the Thumb1 instructions that know how to use hi regs. 370 static const unsigned THUMB_rGPRAO[] = { 371 ARM::R0, ARM::R1, ARM::R2, ARM::R3, 372 ARM::R4, ARM::R5, ARM::R6, ARM::R7 }; 373 374 rGPRClass::iterator 375 rGPRClass::allocation_order_begin(const MachineFunction &MF) const { 376 const TargetMachine &TM = MF.getTarget(); 377 const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>(); 378 if (Subtarget.isThumb1Only()) 379 return THUMB_rGPRAO; 380 if (Subtarget.isTargetDarwin()) { 381 if (Subtarget.isR9Reserved()) 382 return ARM_rGPRAO_4; 383 else 384 return ARM_rGPRAO_3; 385 } else { 386 if (Subtarget.isR9Reserved()) 387 return ARM_rGPRAO_2; 388 else if (Subtarget.isThumb()) 389 return ARM_rGPRAO_5; 390 else 391 return ARM_rGPRAO_1; 392 } 393 } 394 395 rGPRClass::iterator 396 rGPRClass::allocation_order_end(const MachineFunction &MF) const { 397 const TargetMachine &TM = MF.getTarget(); 398 const TargetRegisterInfo *RI = TM.getRegisterInfo(); 399 const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>(); 400 GPRClass::iterator I; 401 402 if (Subtarget.isThumb1Only()) { 403 I = THUMB_rGPRAO + (sizeof(THUMB_rGPRAO)/sizeof(unsigned)); 404 return RI->hasFP(MF) ? I-1 : I; 405 } 406 407 if (Subtarget.isTargetDarwin()) { 408 if (Subtarget.isR9Reserved()) 409 I = ARM_rGPRAO_4 + (sizeof(ARM_rGPRAO_4)/sizeof(unsigned)); 410 else 411 I = ARM_rGPRAO_3 + (sizeof(ARM_rGPRAO_3)/sizeof(unsigned)); 412 } else { 413 if (Subtarget.isR9Reserved()) 414 I = ARM_rGPRAO_2 + (sizeof(ARM_rGPRAO_2)/sizeof(unsigned)); 415 else if (Subtarget.isThumb()) 416 I = ARM_rGPRAO_5 + (sizeof(ARM_rGPRAO_5)/sizeof(unsigned)); 417 else 418 I = ARM_rGPRAO_1 + (sizeof(ARM_rGPRAO_1)/sizeof(unsigned)); 419 } 420 421 return RI->hasFP(MF) ? I-1 : I; 422 } 423 }]; 424} 425 426// Thumb registers are R0-R7 normally. Some instructions can still use 427// the general GPR register class above (MOV, e.g.) 428def tGPR : RegisterClass<"ARM", [i32], 32, [R0, R1, R2, R3, R4, R5, R6, R7]> { 429 let MethodProtos = [{ 430 iterator allocation_order_begin(const MachineFunction &MF) const; 431 iterator allocation_order_end(const MachineFunction &MF) const; 432 }]; 433 let MethodBodies = [{ 434 static const unsigned THUMB_tGPR_AO[] = { 435 ARM::R0, ARM::R1, ARM::R2, ARM::R3, 436 ARM::R4, ARM::R5, ARM::R6, ARM::R7 }; 437 438 // FP is R7, only low registers available. 439 tGPRClass::iterator 440 tGPRClass::allocation_order_begin(const MachineFunction &MF) const { 441 return THUMB_tGPR_AO; 442 } 443 444 tGPRClass::iterator 445 tGPRClass::allocation_order_end(const MachineFunction &MF) const { 446 const TargetMachine &TM = MF.getTarget(); 447 const TargetRegisterInfo *RI = TM.getRegisterInfo(); 448 tGPRClass::iterator I = 449 THUMB_tGPR_AO + (sizeof(THUMB_tGPR_AO)/sizeof(unsigned)); 450 return RI->hasFP(MF) ? I-1 : I; 451 } 452 }]; 453} 454 455// For tail calls, we can't use callee-saved registers, as they are restored 456// to the saved value before the tail call, which would clobber a call address. 457// Note, getMinimalPhysRegClass(R0) returns tGPR because of the names of 458// this class and the preceding one(!) This is what we want. 459def tcGPR : RegisterClass<"ARM", [i32], 32, [R0, R1, R2, R3, R9, R12]> { 460 let MethodProtos = [{ 461 iterator allocation_order_begin(const MachineFunction &MF) const; 462 iterator allocation_order_end(const MachineFunction &MF) const; 463 }]; 464 let MethodBodies = [{ 465 // R9 is available. 466 static const unsigned ARM_GPR_R9_TC[] = { 467 ARM::R0, ARM::R1, ARM::R2, ARM::R3, 468 ARM::R9, ARM::R12 }; 469 // R9 is not available. 470 static const unsigned ARM_GPR_NOR9_TC[] = { 471 ARM::R0, ARM::R1, ARM::R2, ARM::R3, 472 ARM::R12 }; 473 474 // For Thumb1 mode, we don't want to allocate hi regs at all, as we 475 // don't know how to spill them. If we make our prologue/epilogue code 476 // smarter at some point, we can go back to using the above allocation 477 // orders for the Thumb1 instructions that know how to use hi regs. 478 static const unsigned THUMB_GPR_AO_TC[] = { 479 ARM::R0, ARM::R1, ARM::R2, ARM::R3 }; 480 481 tcGPRClass::iterator 482 tcGPRClass::allocation_order_begin(const MachineFunction &MF) const { 483 const TargetMachine &TM = MF.getTarget(); 484 const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>(); 485 if (Subtarget.isThumb1Only()) 486 return THUMB_GPR_AO_TC; 487 if (Subtarget.isTargetDarwin()) { 488 if (Subtarget.isR9Reserved()) 489 return ARM_GPR_NOR9_TC; 490 else 491 return ARM_GPR_R9_TC; 492 } else 493 // R9 is either callee-saved or reserved; can't use it. 494 return ARM_GPR_NOR9_TC; 495 } 496 497 tcGPRClass::iterator 498 tcGPRClass::allocation_order_end(const MachineFunction &MF) const { 499 const TargetMachine &TM = MF.getTarget(); 500 const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>(); 501 GPRClass::iterator I; 502 503 if (Subtarget.isThumb1Only()) { 504 I = THUMB_GPR_AO_TC + (sizeof(THUMB_GPR_AO_TC)/sizeof(unsigned)); 505 return I; 506 } 507 508 if (Subtarget.isTargetDarwin()) { 509 if (Subtarget.isR9Reserved()) 510 I = ARM_GPR_NOR9_TC + (sizeof(ARM_GPR_NOR9_TC)/sizeof(unsigned)); 511 else 512 I = ARM_GPR_R9_TC + (sizeof(ARM_GPR_R9_TC)/sizeof(unsigned)); 513 } else 514 // R9 is either callee-saved or reserved; can't use it. 515 I = ARM_GPR_NOR9_TC + (sizeof(ARM_GPR_NOR9_TC)/sizeof(unsigned)); 516 return I; 517 } 518 }]; 519} 520 521 522// Scalar single precision floating point register class.. 523def SPR : RegisterClass<"ARM", [f32], 32, [S0, S1, S2, S3, S4, S5, S6, S7, S8, 524 S9, S10, S11, S12, S13, S14, S15, S16, S17, S18, S19, S20, S21, S22, 525 S23, S24, S25, S26, S27, S28, S29, S30, S31]>; 526 527// Subset of SPR which can be used as a source of NEON scalars for 16-bit 528// operations 529def SPR_8 : RegisterClass<"ARM", [f32], 32, 530 [S0, S1, S2, S3, S4, S5, S6, S7, 531 S8, S9, S10, S11, S12, S13, S14, S15]>; 532 533// Scalar double precision floating point / generic 64-bit vector register 534// class. 535// ARM requires only word alignment for double. It's more performant if it 536// is double-word alignment though. 537def DPR : RegisterClass<"ARM", [f64, v8i8, v4i16, v2i32, v1i64, v2f32], 64, 538 [D0, D1, D2, D3, D4, D5, D6, D7, 539 D8, D9, D10, D11, D12, D13, D14, D15, 540 D16, D17, D18, D19, D20, D21, D22, D23, 541 D24, D25, D26, D27, D28, D29, D30, D31]> { 542 let MethodProtos = [{ 543 iterator allocation_order_begin(const MachineFunction &MF) const; 544 iterator allocation_order_end(const MachineFunction &MF) const; 545 }]; 546 let MethodBodies = [{ 547 // VFP2 548 static const unsigned ARM_DPR_VFP2[] = { 549 ARM::D0, ARM::D1, ARM::D2, ARM::D3, 550 ARM::D4, ARM::D5, ARM::D6, ARM::D7, 551 ARM::D8, ARM::D9, ARM::D10, ARM::D11, 552 ARM::D12, ARM::D13, ARM::D14, ARM::D15 }; 553 // VFP3 554 static const unsigned ARM_DPR_VFP3[] = { 555 ARM::D0, ARM::D1, ARM::D2, ARM::D3, 556 ARM::D4, ARM::D5, ARM::D6, ARM::D7, 557 ARM::D8, ARM::D9, ARM::D10, ARM::D11, 558 ARM::D12, ARM::D13, ARM::D14, ARM::D15, 559 ARM::D16, ARM::D17, ARM::D18, ARM::D19, 560 ARM::D20, ARM::D21, ARM::D22, ARM::D23, 561 ARM::D24, ARM::D25, ARM::D26, ARM::D27, 562 ARM::D28, ARM::D29, ARM::D30, ARM::D31 }; 563 DPRClass::iterator 564 DPRClass::allocation_order_begin(const MachineFunction &MF) const { 565 const TargetMachine &TM = MF.getTarget(); 566 const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>(); 567 if (Subtarget.hasVFP3()) 568 return ARM_DPR_VFP3; 569 return ARM_DPR_VFP2; 570 } 571 572 DPRClass::iterator 573 DPRClass::allocation_order_end(const MachineFunction &MF) const { 574 const TargetMachine &TM = MF.getTarget(); 575 const ARMSubtarget &Subtarget = TM.getSubtarget<ARMSubtarget>(); 576 if (Subtarget.hasVFP3()) 577 return ARM_DPR_VFP3 + (sizeof(ARM_DPR_VFP3)/sizeof(unsigned)); 578 else 579 return ARM_DPR_VFP2 + (sizeof(ARM_DPR_VFP2)/sizeof(unsigned)); 580 } 581 }]; 582} 583 584// Subset of DPR that are accessible with VFP2 (and so that also have 585// 32-bit SPR subregs). 586def DPR_VFP2 : RegisterClass<"ARM", [f64, v8i8, v4i16, v2i32, v1i64, v2f32], 64, 587 [D0, D1, D2, D3, D4, D5, D6, D7, 588 D8, D9, D10, D11, D12, D13, D14, D15]> { 589 let SubRegClasses = [(SPR ssub_0, ssub_1)]; 590} 591 592// Subset of DPR which can be used as a source of NEON scalars for 16-bit 593// operations 594def DPR_8 : RegisterClass<"ARM", [f64, v8i8, v4i16, v2i32, v1i64, v2f32], 64, 595 [D0, D1, D2, D3, D4, D5, D6, D7]> { 596 let SubRegClasses = [(SPR_8 ssub_0, ssub_1)]; 597} 598 599// Generic 128-bit vector register class. 600def QPR : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64], 128, 601 [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7, 602 Q8, Q9, Q10, Q11, Q12, Q13, Q14, Q15]> { 603 let SubRegClasses = [(DPR dsub_0, dsub_1)]; 604} 605 606// Subset of QPR that have 32-bit SPR subregs. 607def QPR_VFP2 : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64], 608 128, 609 [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7]> { 610 let SubRegClasses = [(SPR ssub_0, ssub_1, ssub_2, ssub_3), 611 (DPR_VFP2 dsub_0, dsub_1)]; 612} 613 614// Subset of QPR that have DPR_8 and SPR_8 subregs. 615def QPR_8 : RegisterClass<"ARM", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64], 616 128, 617 [Q0, Q1, Q2, Q3]> { 618 let SubRegClasses = [(SPR_8 ssub_0, ssub_1, ssub_2, ssub_3), 619 (DPR_8 dsub_0, dsub_1)]; 620} 621 622// Pseudo 256-bit vector register class to model pairs of Q registers 623// (4 consecutive D registers). 624def QQPR : RegisterClass<"ARM", [v4i64], 625 256, 626 [QQ0, QQ1, QQ2, QQ3, QQ4, QQ5, QQ6, QQ7]> { 627 let SubRegClasses = [(DPR dsub_0, dsub_1, dsub_2, dsub_3), 628 (QPR qsub_0, qsub_1)]; 629} 630 631// Subset of QQPR that have 32-bit SPR subregs. 632def QQPR_VFP2 : RegisterClass<"ARM", [v4i64], 633 256, 634 [QQ0, QQ1, QQ2, QQ3]> { 635 let SubRegClasses = [(SPR ssub_0, ssub_1, ssub_2, ssub_3), 636 (DPR_VFP2 dsub_0, dsub_1, dsub_2, dsub_3), 637 (QPR_VFP2 qsub_0, qsub_1)]; 638 639} 640 641// Pseudo 512-bit vector register class to model 4 consecutive Q registers 642// (8 consecutive D registers). 643def QQQQPR : RegisterClass<"ARM", [v8i64], 644 256, 645 [QQQQ0, QQQQ1, QQQQ2, QQQQ3]> { 646 let SubRegClasses = [(DPR dsub_0, dsub_1, dsub_2, dsub_3, 647 dsub_4, dsub_5, dsub_6, dsub_7), 648 (QPR qsub_0, qsub_1, qsub_2, qsub_3)]; 649} 650 651// Condition code registers. 652def CCR : RegisterClass<"ARM", [i32], 32, [CPSR]>; 653