1 //===- PPCRegisterInfo.cpp - PowerPC Register Information -------*- 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 // This file contains the PowerPC implementation of the TargetRegisterInfo 11 // class. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #define DEBUG_TYPE "reginfo" 16 #include "PPC.h" 17 #include "PPCInstrBuilder.h" 18 #include "PPCMachineFunctionInfo.h" 19 #include "PPCRegisterInfo.h" 20 #include "PPCFrameInfo.h" 21 #include "PPCSubtarget.h" 22 #include "llvm/CallingConv.h" 23 #include "llvm/Constants.h" 24 #include "llvm/Function.h" 25 #include "llvm/Type.h" 26 #include "llvm/CodeGen/ValueTypes.h" 27 #include "llvm/CodeGen/MachineInstrBuilder.h" 28 #include "llvm/CodeGen/MachineModuleInfo.h" 29 #include "llvm/CodeGen/MachineFunction.h" 30 #include "llvm/CodeGen/MachineFrameInfo.h" 31 #include "llvm/CodeGen/MachineLocation.h" 32 #include "llvm/CodeGen/MachineRegisterInfo.h" 33 #include "llvm/CodeGen/RegisterScavenging.h" 34 #include "llvm/Target/TargetFrameInfo.h" 35 #include "llvm/Target/TargetInstrInfo.h" 36 #include "llvm/Target/TargetMachine.h" 37 #include "llvm/Target/TargetOptions.h" 38 #include "llvm/Support/CommandLine.h" 39 #include "llvm/Support/Debug.h" 40 #include "llvm/Support/ErrorHandling.h" 41 #include "llvm/Support/MathExtras.h" 42 #include "llvm/Support/raw_ostream.h" 43 #include "llvm/ADT/BitVector.h" 44 #include "llvm/ADT/STLExtras.h" 45 #include <cstdlib> 46 using namespace llvm; 47 48 // FIXME This disables some code that aligns the stack to a boundary 49 // bigger than the default (16 bytes on Darwin) when there is a stack local 50 // of greater alignment. This does not currently work, because the delta 51 // between old and new stack pointers is added to offsets that reference 52 // incoming parameters after the prolog is generated, and the code that 53 // does that doesn't handle a variable delta. You don't want to do that 54 // anyway; a better approach is to reserve another register that retains 55 // to the incoming stack pointer, and reference parameters relative to that. 56 #define ALIGN_STACK 0 57 58 // FIXME (64-bit): Eventually enable by default. 59 cl::opt<bool> EnablePPC32RS("enable-ppc32-regscavenger", 60 cl::init(false), 61 cl::desc("Enable PPC32 register scavenger"), 62 cl::Hidden); 63 cl::opt<bool> EnablePPC64RS("enable-ppc64-regscavenger", 64 cl::init(false), 65 cl::desc("Enable PPC64 register scavenger"), 66 cl::Hidden); 67 #define EnableRegisterScavenging \ 68 ((EnablePPC32RS && !Subtarget.isPPC64()) || \ 69 (EnablePPC64RS && Subtarget.isPPC64())) 70 71 // FIXME (64-bit): Should be inlined. 72 bool 73 PPCRegisterInfo::requiresRegisterScavenging(const MachineFunction &) const { 74 return EnableRegisterScavenging; 75 } 76 77 /// getRegisterNumbering - Given the enum value for some register, e.g. 78 /// PPC::F14, return the number that it corresponds to (e.g. 14). 79 unsigned PPCRegisterInfo::getRegisterNumbering(unsigned RegEnum) { 80 using namespace PPC; 81 switch (RegEnum) { 82 case 0: return 0; 83 case R0 : case X0 : case F0 : case V0 : case CR0: case CR0LT: return 0; 84 case R1 : case X1 : case F1 : case V1 : case CR1: case CR0GT: return 1; 85 case R2 : case X2 : case F2 : case V2 : case CR2: case CR0EQ: return 2; 86 case R3 : case X3 : case F3 : case V3 : case CR3: case CR0UN: return 3; 87 case R4 : case X4 : case F4 : case V4 : case CR4: case CR1LT: return 4; 88 case R5 : case X5 : case F5 : case V5 : case CR5: case CR1GT: return 5; 89 case R6 : case X6 : case F6 : case V6 : case CR6: case CR1EQ: return 6; 90 case R7 : case X7 : case F7 : case V7 : case CR7: case CR1UN: return 7; 91 case R8 : case X8 : case F8 : case V8 : case CR2LT: return 8; 92 case R9 : case X9 : case F9 : case V9 : case CR2GT: return 9; 93 case R10: case X10: case F10: case V10: case CR2EQ: return 10; 94 case R11: case X11: case F11: case V11: case CR2UN: return 11; 95 case R12: case X12: case F12: case V12: case CR3LT: return 12; 96 case R13: case X13: case F13: case V13: case CR3GT: return 13; 97 case R14: case X14: case F14: case V14: case CR3EQ: return 14; 98 case R15: case X15: case F15: case V15: case CR3UN: return 15; 99 case R16: case X16: case F16: case V16: case CR4LT: return 16; 100 case R17: case X17: case F17: case V17: case CR4GT: return 17; 101 case R18: case X18: case F18: case V18: case CR4EQ: return 18; 102 case R19: case X19: case F19: case V19: case CR4UN: return 19; 103 case R20: case X20: case F20: case V20: case CR5LT: return 20; 104 case R21: case X21: case F21: case V21: case CR5GT: return 21; 105 case R22: case X22: case F22: case V22: case CR5EQ: return 22; 106 case R23: case X23: case F23: case V23: case CR5UN: return 23; 107 case R24: case X24: case F24: case V24: case CR6LT: return 24; 108 case R25: case X25: case F25: case V25: case CR6GT: return 25; 109 case R26: case X26: case F26: case V26: case CR6EQ: return 26; 110 case R27: case X27: case F27: case V27: case CR6UN: return 27; 111 case R28: case X28: case F28: case V28: case CR7LT: return 28; 112 case R29: case X29: case F29: case V29: case CR7GT: return 29; 113 case R30: case X30: case F30: case V30: case CR7EQ: return 30; 114 case R31: case X31: case F31: case V31: case CR7UN: return 31; 115 default: 116 llvm_unreachable("Unhandled reg in PPCRegisterInfo::getRegisterNumbering!"); 117 } 118 } 119 120 PPCRegisterInfo::PPCRegisterInfo(const PPCSubtarget &ST, 121 const TargetInstrInfo &tii) 122 : PPCGenRegisterInfo(PPC::ADJCALLSTACKDOWN, PPC::ADJCALLSTACKUP), 123 Subtarget(ST), TII(tii) { 124 ImmToIdxMap[PPC::LD] = PPC::LDX; ImmToIdxMap[PPC::STD] = PPC::STDX; 125 ImmToIdxMap[PPC::LBZ] = PPC::LBZX; ImmToIdxMap[PPC::STB] = PPC::STBX; 126 ImmToIdxMap[PPC::LHZ] = PPC::LHZX; ImmToIdxMap[PPC::LHA] = PPC::LHAX; 127 ImmToIdxMap[PPC::LWZ] = PPC::LWZX; ImmToIdxMap[PPC::LWA] = PPC::LWAX; 128 ImmToIdxMap[PPC::LFS] = PPC::LFSX; ImmToIdxMap[PPC::LFD] = PPC::LFDX; 129 ImmToIdxMap[PPC::STH] = PPC::STHX; ImmToIdxMap[PPC::STW] = PPC::STWX; 130 ImmToIdxMap[PPC::STFS] = PPC::STFSX; ImmToIdxMap[PPC::STFD] = PPC::STFDX; 131 ImmToIdxMap[PPC::ADDI] = PPC::ADD4; 132 133 // 64-bit 134 ImmToIdxMap[PPC::LHA8] = PPC::LHAX8; ImmToIdxMap[PPC::LBZ8] = PPC::LBZX8; 135 ImmToIdxMap[PPC::LHZ8] = PPC::LHZX8; ImmToIdxMap[PPC::LWZ8] = PPC::LWZX8; 136 ImmToIdxMap[PPC::STB8] = PPC::STBX8; ImmToIdxMap[PPC::STH8] = PPC::STHX8; 137 ImmToIdxMap[PPC::STW8] = PPC::STWX8; ImmToIdxMap[PPC::STDU] = PPC::STDUX; 138 ImmToIdxMap[PPC::ADDI8] = PPC::ADD8; ImmToIdxMap[PPC::STD_32] = PPC::STDX_32; 139 } 140 141 /// getPointerRegClass - Return the register class to use to hold pointers. 142 /// This is used for addressing modes. 143 const TargetRegisterClass * 144 PPCRegisterInfo::getPointerRegClass(unsigned Kind) const { 145 if (Subtarget.isPPC64()) 146 return &PPC::G8RCRegClass; 147 return &PPC::GPRCRegClass; 148 } 149 150 const unsigned* 151 PPCRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const { 152 // 32-bit Darwin calling convention. 153 static const unsigned Darwin32_CalleeSavedRegs[] = { 154 PPC::R13, PPC::R14, PPC::R15, 155 PPC::R16, PPC::R17, PPC::R18, PPC::R19, 156 PPC::R20, PPC::R21, PPC::R22, PPC::R23, 157 PPC::R24, PPC::R25, PPC::R26, PPC::R27, 158 PPC::R28, PPC::R29, PPC::R30, PPC::R31, 159 160 PPC::F14, PPC::F15, PPC::F16, PPC::F17, 161 PPC::F18, PPC::F19, PPC::F20, PPC::F21, 162 PPC::F22, PPC::F23, PPC::F24, PPC::F25, 163 PPC::F26, PPC::F27, PPC::F28, PPC::F29, 164 PPC::F30, PPC::F31, 165 166 PPC::CR2, PPC::CR3, PPC::CR4, 167 PPC::V20, PPC::V21, PPC::V22, PPC::V23, 168 PPC::V24, PPC::V25, PPC::V26, PPC::V27, 169 PPC::V28, PPC::V29, PPC::V30, PPC::V31, 170 171 PPC::CR2LT, PPC::CR2GT, PPC::CR2EQ, PPC::CR2UN, 172 PPC::CR3LT, PPC::CR3GT, PPC::CR3EQ, PPC::CR3UN, 173 PPC::CR4LT, PPC::CR4GT, PPC::CR4EQ, PPC::CR4UN, 174 175 PPC::LR, 0 176 }; 177 178 // 32-bit SVR4 calling convention. 179 static const unsigned SVR4_CalleeSavedRegs[] = { 180 PPC::R14, PPC::R15, 181 PPC::R16, PPC::R17, PPC::R18, PPC::R19, 182 PPC::R20, PPC::R21, PPC::R22, PPC::R23, 183 PPC::R24, PPC::R25, PPC::R26, PPC::R27, 184 PPC::R28, PPC::R29, PPC::R30, PPC::R31, 185 186 PPC::F14, PPC::F15, PPC::F16, PPC::F17, 187 PPC::F18, PPC::F19, PPC::F20, PPC::F21, 188 PPC::F22, PPC::F23, PPC::F24, PPC::F25, 189 PPC::F26, PPC::F27, PPC::F28, PPC::F29, 190 PPC::F30, PPC::F31, 191 192 PPC::CR2, PPC::CR3, PPC::CR4, 193 194 PPC::VRSAVE, 195 196 PPC::V20, PPC::V21, PPC::V22, PPC::V23, 197 PPC::V24, PPC::V25, PPC::V26, PPC::V27, 198 PPC::V28, PPC::V29, PPC::V30, PPC::V31, 199 200 PPC::CR2LT, PPC::CR2GT, PPC::CR2EQ, PPC::CR2UN, 201 PPC::CR3LT, PPC::CR3GT, PPC::CR3EQ, PPC::CR3UN, 202 PPC::CR4LT, PPC::CR4GT, PPC::CR4EQ, PPC::CR4UN, 203 204 0 205 }; 206 // 64-bit Darwin calling convention. 207 static const unsigned Darwin64_CalleeSavedRegs[] = { 208 PPC::X14, PPC::X15, 209 PPC::X16, PPC::X17, PPC::X18, PPC::X19, 210 PPC::X20, PPC::X21, PPC::X22, PPC::X23, 211 PPC::X24, PPC::X25, PPC::X26, PPC::X27, 212 PPC::X28, PPC::X29, PPC::X30, PPC::X31, 213 214 PPC::F14, PPC::F15, PPC::F16, PPC::F17, 215 PPC::F18, PPC::F19, PPC::F20, PPC::F21, 216 PPC::F22, PPC::F23, PPC::F24, PPC::F25, 217 PPC::F26, PPC::F27, PPC::F28, PPC::F29, 218 PPC::F30, PPC::F31, 219 220 PPC::CR2, PPC::CR3, PPC::CR4, 221 PPC::V20, PPC::V21, PPC::V22, PPC::V23, 222 PPC::V24, PPC::V25, PPC::V26, PPC::V27, 223 PPC::V28, PPC::V29, PPC::V30, PPC::V31, 224 225 PPC::CR2LT, PPC::CR2GT, PPC::CR2EQ, PPC::CR2UN, 226 PPC::CR3LT, PPC::CR3GT, PPC::CR3EQ, PPC::CR3UN, 227 PPC::CR4LT, PPC::CR4GT, PPC::CR4EQ, PPC::CR4UN, 228 229 PPC::LR8, 0 230 }; 231 232 // 64-bit SVR4 calling convention. 233 static const unsigned SVR4_64_CalleeSavedRegs[] = { 234 PPC::X14, PPC::X15, 235 PPC::X16, PPC::X17, PPC::X18, PPC::X19, 236 PPC::X20, PPC::X21, PPC::X22, PPC::X23, 237 PPC::X24, PPC::X25, PPC::X26, PPC::X27, 238 PPC::X28, PPC::X29, PPC::X30, PPC::X31, 239 240 PPC::F14, PPC::F15, PPC::F16, PPC::F17, 241 PPC::F18, PPC::F19, PPC::F20, PPC::F21, 242 PPC::F22, PPC::F23, PPC::F24, PPC::F25, 243 PPC::F26, PPC::F27, PPC::F28, PPC::F29, 244 PPC::F30, PPC::F31, 245 246 PPC::CR2, PPC::CR3, PPC::CR4, 247 248 PPC::VRSAVE, 249 250 PPC::V20, PPC::V21, PPC::V22, PPC::V23, 251 PPC::V24, PPC::V25, PPC::V26, PPC::V27, 252 PPC::V28, PPC::V29, PPC::V30, PPC::V31, 253 254 PPC::CR2LT, PPC::CR2GT, PPC::CR2EQ, PPC::CR2UN, 255 PPC::CR3LT, PPC::CR3GT, PPC::CR3EQ, PPC::CR3UN, 256 PPC::CR4LT, PPC::CR4GT, PPC::CR4EQ, PPC::CR4UN, 257 258 0 259 }; 260 261 if (Subtarget.isDarwinABI()) 262 return Subtarget.isPPC64() ? Darwin64_CalleeSavedRegs : 263 Darwin32_CalleeSavedRegs; 264 265 return Subtarget.isPPC64() ? SVR4_64_CalleeSavedRegs : SVR4_CalleeSavedRegs; 266 } 267 268 const TargetRegisterClass* const* 269 PPCRegisterInfo::getCalleeSavedRegClasses(const MachineFunction *MF) const { 270 // 32-bit Darwin calling convention. 271 static const TargetRegisterClass * const Darwin32_CalleeSavedRegClasses[] = { 272 &PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass, 273 &PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass, 274 &PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass, 275 &PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass, 276 &PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass, 277 278 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 279 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 280 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 281 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 282 &PPC::F8RCRegClass,&PPC::F8RCRegClass, 283 284 &PPC::CRRCRegClass,&PPC::CRRCRegClass,&PPC::CRRCRegClass, 285 286 &PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass, 287 &PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass, 288 &PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass, 289 290 &PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass, 291 &PPC::CRBITRCRegClass, 292 &PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass, 293 &PPC::CRBITRCRegClass, 294 &PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass, 295 &PPC::CRBITRCRegClass, 296 297 &PPC::GPRCRegClass, 0 298 }; 299 300 // 32-bit SVR4 calling convention. 301 static const TargetRegisterClass * const SVR4_CalleeSavedRegClasses[] = { 302 &PPC::GPRCRegClass,&PPC::GPRCRegClass, 303 &PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass, 304 &PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass, 305 &PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass, 306 &PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass,&PPC::GPRCRegClass, 307 308 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 309 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 310 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 311 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 312 &PPC::F8RCRegClass,&PPC::F8RCRegClass, 313 314 &PPC::CRRCRegClass,&PPC::CRRCRegClass,&PPC::CRRCRegClass, 315 316 &PPC::VRSAVERCRegClass, 317 318 &PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass, 319 &PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass, 320 &PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass, 321 322 &PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass, 323 &PPC::CRBITRCRegClass, 324 &PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass, 325 &PPC::CRBITRCRegClass, 326 &PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass, 327 &PPC::CRBITRCRegClass, 328 329 0 330 }; 331 332 // 64-bit Darwin calling convention. 333 static const TargetRegisterClass * const Darwin64_CalleeSavedRegClasses[] = { 334 &PPC::G8RCRegClass,&PPC::G8RCRegClass, 335 &PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass, 336 &PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass, 337 &PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass, 338 &PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass, 339 340 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 341 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 342 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 343 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 344 &PPC::F8RCRegClass,&PPC::F8RCRegClass, 345 346 &PPC::CRRCRegClass,&PPC::CRRCRegClass,&PPC::CRRCRegClass, 347 348 &PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass, 349 &PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass, 350 &PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass, 351 352 &PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass, 353 &PPC::CRBITRCRegClass, 354 &PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass, 355 &PPC::CRBITRCRegClass, 356 &PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass, 357 &PPC::CRBITRCRegClass, 358 359 &PPC::G8RCRegClass, 0 360 }; 361 362 // 64-bit SVR4 calling convention. 363 static const TargetRegisterClass * const SVR4_64_CalleeSavedRegClasses[] = { 364 &PPC::G8RCRegClass,&PPC::G8RCRegClass, 365 &PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass, 366 &PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass, 367 &PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass, 368 &PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass,&PPC::G8RCRegClass, 369 370 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 371 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 372 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 373 &PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass,&PPC::F8RCRegClass, 374 &PPC::F8RCRegClass,&PPC::F8RCRegClass, 375 376 &PPC::CRRCRegClass,&PPC::CRRCRegClass,&PPC::CRRCRegClass, 377 378 &PPC::VRSAVERCRegClass, 379 380 &PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass, 381 &PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass, 382 &PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass,&PPC::VRRCRegClass, 383 384 &PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass, 385 &PPC::CRBITRCRegClass, 386 &PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass, 387 &PPC::CRBITRCRegClass, 388 &PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass,&PPC::CRBITRCRegClass, 389 &PPC::CRBITRCRegClass, 390 391 0 392 }; 393 394 if (Subtarget.isDarwinABI()) 395 return Subtarget.isPPC64() ? Darwin64_CalleeSavedRegClasses : 396 Darwin32_CalleeSavedRegClasses; 397 398 return Subtarget.isPPC64() ? SVR4_64_CalleeSavedRegClasses 399 : SVR4_CalleeSavedRegClasses; 400 } 401 402 // needsFP - Return true if the specified function should have a dedicated frame 403 // pointer register. This is true if the function has variable sized allocas or 404 // if frame pointer elimination is disabled. 405 // 406 static bool needsFP(const MachineFunction &MF) { 407 const MachineFrameInfo *MFI = MF.getFrameInfo(); 408 return NoFramePointerElim || MFI->hasVarSizedObjects() || 409 (GuaranteedTailCallOpt && MF.getInfo<PPCFunctionInfo>()->hasFastCall()); 410 } 411 412 static bool spillsCR(const MachineFunction &MF) { 413 const PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); 414 return FuncInfo->isCRSpilled(); 415 } 416 417 BitVector PPCRegisterInfo::getReservedRegs(const MachineFunction &MF) const { 418 BitVector Reserved(getNumRegs()); 419 Reserved.set(PPC::R0); 420 Reserved.set(PPC::R1); 421 Reserved.set(PPC::LR); 422 Reserved.set(PPC::LR8); 423 Reserved.set(PPC::RM); 424 425 // The SVR4 ABI reserves r2 and r13 426 if (Subtarget.isSVR4ABI()) { 427 Reserved.set(PPC::R2); // System-reserved register 428 Reserved.set(PPC::R13); // Small Data Area pointer register 429 } 430 // Reserve R2 on Darwin to hack around the problem of save/restore of CR 431 // when the stack frame is too big to address directly; we need two regs. 432 // This is a hack. 433 if (Subtarget.isDarwinABI()) { 434 Reserved.set(PPC::R2); 435 } 436 437 // On PPC64, r13 is the thread pointer. Never allocate this register. 438 // Note that this is over conservative, as it also prevents allocation of R31 439 // when the FP is not needed. 440 if (Subtarget.isPPC64()) { 441 Reserved.set(PPC::R13); 442 Reserved.set(PPC::R31); 443 444 if (!EnableRegisterScavenging) 445 Reserved.set(PPC::R0); // FIXME (64-bit): Remove 446 447 Reserved.set(PPC::X0); 448 Reserved.set(PPC::X1); 449 Reserved.set(PPC::X13); 450 Reserved.set(PPC::X31); 451 452 // The 64-bit SVR4 ABI reserves r2 for the TOC pointer. 453 if (Subtarget.isSVR4ABI()) { 454 Reserved.set(PPC::X2); 455 } 456 // Reserve R2 on Darwin to hack around the problem of save/restore of CR 457 // when the stack frame is too big to address directly; we need two regs. 458 // This is a hack. 459 if (Subtarget.isDarwinABI()) { 460 Reserved.set(PPC::X2); 461 } 462 } 463 464 if (needsFP(MF)) 465 Reserved.set(PPC::R31); 466 467 return Reserved; 468 } 469 470 //===----------------------------------------------------------------------===// 471 // Stack Frame Processing methods 472 //===----------------------------------------------------------------------===// 473 474 // hasFP - Return true if the specified function actually has a dedicated frame 475 // pointer register. This is true if the function needs a frame pointer and has 476 // a non-zero stack size. 477 bool PPCRegisterInfo::hasFP(const MachineFunction &MF) const { 478 const MachineFrameInfo *MFI = MF.getFrameInfo(); 479 return MFI->getStackSize() && needsFP(MF); 480 } 481 482 /// MustSaveLR - Return true if this function requires that we save the LR 483 /// register onto the stack in the prolog and restore it in the epilog of the 484 /// function. 485 static bool MustSaveLR(const MachineFunction &MF, unsigned LR) { 486 const PPCFunctionInfo *MFI = MF.getInfo<PPCFunctionInfo>(); 487 488 // We need a save/restore of LR if there is any def of LR (which is 489 // defined by calls, including the PIC setup sequence), or if there is 490 // some use of the LR stack slot (e.g. for builtin_return_address). 491 // (LR comes in 32 and 64 bit versions.) 492 MachineRegisterInfo::def_iterator RI = MF.getRegInfo().def_begin(LR); 493 return RI !=MF.getRegInfo().def_end() || MFI->isLRStoreRequired(); 494 } 495 496 497 498 void PPCRegisterInfo:: 499 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, 500 MachineBasicBlock::iterator I) const { 501 if (GuaranteedTailCallOpt && I->getOpcode() == PPC::ADJCALLSTACKUP) { 502 // Add (actually subtract) back the amount the callee popped on return. 503 if (int CalleeAmt = I->getOperand(1).getImm()) { 504 bool is64Bit = Subtarget.isPPC64(); 505 CalleeAmt *= -1; 506 unsigned StackReg = is64Bit ? PPC::X1 : PPC::R1; 507 unsigned TmpReg = is64Bit ? PPC::X0 : PPC::R0; 508 unsigned ADDIInstr = is64Bit ? PPC::ADDI8 : PPC::ADDI; 509 unsigned ADDInstr = is64Bit ? PPC::ADD8 : PPC::ADD4; 510 unsigned LISInstr = is64Bit ? PPC::LIS8 : PPC::LIS; 511 unsigned ORIInstr = is64Bit ? PPC::ORI8 : PPC::ORI; 512 MachineInstr *MI = I; 513 DebugLoc dl = MI->getDebugLoc(); 514 515 if (isInt<16>(CalleeAmt)) { 516 BuildMI(MBB, I, dl, TII.get(ADDIInstr), StackReg).addReg(StackReg). 517 addImm(CalleeAmt); 518 } else { 519 MachineBasicBlock::iterator MBBI = I; 520 BuildMI(MBB, MBBI, dl, TII.get(LISInstr), TmpReg) 521 .addImm(CalleeAmt >> 16); 522 BuildMI(MBB, MBBI, dl, TII.get(ORIInstr), TmpReg) 523 .addReg(TmpReg, RegState::Kill) 524 .addImm(CalleeAmt & 0xFFFF); 525 BuildMI(MBB, MBBI, dl, TII.get(ADDInstr)) 526 .addReg(StackReg) 527 .addReg(StackReg) 528 .addReg(TmpReg); 529 } 530 } 531 } 532 // Simply discard ADJCALLSTACKDOWN, ADJCALLSTACKUP instructions. 533 MBB.erase(I); 534 } 535 536 /// findScratchRegister - Find a 'free' PPC register. Try for a call-clobbered 537 /// register first and then a spilled callee-saved register if that fails. 538 static 539 unsigned findScratchRegister(MachineBasicBlock::iterator II, RegScavenger *RS, 540 const TargetRegisterClass *RC, int SPAdj) { 541 assert(RS && "Register scavenging must be on"); 542 unsigned Reg = RS->FindUnusedReg(RC); 543 // FIXME: move ARM callee-saved reg scan to target independent code, then 544 // search for already spilled CS register here. 545 if (Reg == 0) 546 Reg = RS->scavengeRegister(RC, II, SPAdj); 547 return Reg; 548 } 549 550 /// lowerDynamicAlloc - Generate the code for allocating an object in the 551 /// current frame. The sequence of code with be in the general form 552 /// 553 /// addi R0, SP, \#frameSize ; get the address of the previous frame 554 /// stwxu R0, SP, Rnegsize ; add and update the SP with the negated size 555 /// addi Rnew, SP, \#maxCalFrameSize ; get the top of the allocation 556 /// 557 void PPCRegisterInfo::lowerDynamicAlloc(MachineBasicBlock::iterator II, 558 int SPAdj, RegScavenger *RS) const { 559 // Get the instruction. 560 MachineInstr &MI = *II; 561 // Get the instruction's basic block. 562 MachineBasicBlock &MBB = *MI.getParent(); 563 // Get the basic block's function. 564 MachineFunction &MF = *MBB.getParent(); 565 // Get the frame info. 566 MachineFrameInfo *MFI = MF.getFrameInfo(); 567 // Determine whether 64-bit pointers are used. 568 bool LP64 = Subtarget.isPPC64(); 569 DebugLoc dl = MI.getDebugLoc(); 570 571 // Get the maximum call stack size. 572 unsigned maxCallFrameSize = MFI->getMaxCallFrameSize(); 573 // Get the total frame size. 574 unsigned FrameSize = MFI->getStackSize(); 575 576 // Get stack alignments. 577 unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment(); 578 unsigned MaxAlign = MFI->getMaxAlignment(); 579 assert(MaxAlign <= TargetAlign && 580 "Dynamic alloca with large aligns not supported"); 581 582 // Determine the previous frame's address. If FrameSize can't be 583 // represented as 16 bits or we need special alignment, then we load the 584 // previous frame's address from 0(SP). Why not do an addis of the hi? 585 // Because R0 is our only safe tmp register and addi/addis treat R0 as zero. 586 // Constructing the constant and adding would take 3 instructions. 587 // Fortunately, a frame greater than 32K is rare. 588 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 589 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 590 const TargetRegisterClass *RC = LP64 ? G8RC : GPRC; 591 592 // FIXME (64-bit): Use "findScratchRegister" 593 unsigned Reg; 594 if (EnableRegisterScavenging) 595 Reg = findScratchRegister(II, RS, RC, SPAdj); 596 else 597 Reg = PPC::R0; 598 599 if (MaxAlign < TargetAlign && isInt<16>(FrameSize)) { 600 BuildMI(MBB, II, dl, TII.get(PPC::ADDI), Reg) 601 .addReg(PPC::R31) 602 .addImm(FrameSize); 603 } else if (LP64) { 604 if (EnableRegisterScavenging) // FIXME (64-bit): Use "true" part. 605 BuildMI(MBB, II, dl, TII.get(PPC::LD), Reg) 606 .addImm(0) 607 .addReg(PPC::X1); 608 else 609 BuildMI(MBB, II, dl, TII.get(PPC::LD), PPC::X0) 610 .addImm(0) 611 .addReg(PPC::X1); 612 } else { 613 BuildMI(MBB, II, dl, TII.get(PPC::LWZ), Reg) 614 .addImm(0) 615 .addReg(PPC::R1); 616 } 617 618 // Grow the stack and update the stack pointer link, then determine the 619 // address of new allocated space. 620 if (LP64) { 621 if (EnableRegisterScavenging) // FIXME (64-bit): Use "true" part. 622 BuildMI(MBB, II, dl, TII.get(PPC::STDUX)) 623 .addReg(Reg, RegState::Kill) 624 .addReg(PPC::X1) 625 .addReg(MI.getOperand(1).getReg()); 626 else 627 BuildMI(MBB, II, dl, TII.get(PPC::STDUX)) 628 .addReg(PPC::X0, RegState::Kill) 629 .addReg(PPC::X1) 630 .addReg(MI.getOperand(1).getReg()); 631 632 if (!MI.getOperand(1).isKill()) 633 BuildMI(MBB, II, dl, TII.get(PPC::ADDI8), MI.getOperand(0).getReg()) 634 .addReg(PPC::X1) 635 .addImm(maxCallFrameSize); 636 else 637 // Implicitly kill the register. 638 BuildMI(MBB, II, dl, TII.get(PPC::ADDI8), MI.getOperand(0).getReg()) 639 .addReg(PPC::X1) 640 .addImm(maxCallFrameSize) 641 .addReg(MI.getOperand(1).getReg(), RegState::ImplicitKill); 642 } else { 643 BuildMI(MBB, II, dl, TII.get(PPC::STWUX)) 644 .addReg(Reg, RegState::Kill) 645 .addReg(PPC::R1) 646 .addReg(MI.getOperand(1).getReg()); 647 648 if (!MI.getOperand(1).isKill()) 649 BuildMI(MBB, II, dl, TII.get(PPC::ADDI), MI.getOperand(0).getReg()) 650 .addReg(PPC::R1) 651 .addImm(maxCallFrameSize); 652 else 653 // Implicitly kill the register. 654 BuildMI(MBB, II, dl, TII.get(PPC::ADDI), MI.getOperand(0).getReg()) 655 .addReg(PPC::R1) 656 .addImm(maxCallFrameSize) 657 .addReg(MI.getOperand(1).getReg(), RegState::ImplicitKill); 658 } 659 660 // Discard the DYNALLOC instruction. 661 MBB.erase(II); 662 } 663 664 /// lowerCRSpilling - Generate the code for spilling a CR register. Instead of 665 /// reserving a whole register (R0), we scrounge for one here. This generates 666 /// code like this: 667 /// 668 /// mfcr rA ; Move the conditional register into GPR rA. 669 /// rlwinm rA, rA, SB, 0, 31 ; Shift the bits left so they are in CR0's slot. 670 /// stw rA, FI ; Store rA to the frame. 671 /// 672 void PPCRegisterInfo::lowerCRSpilling(MachineBasicBlock::iterator II, 673 unsigned FrameIndex, int SPAdj, 674 RegScavenger *RS) const { 675 // Get the instruction. 676 MachineInstr &MI = *II; // ; SPILL_CR <SrcReg>, <offset>, <FI> 677 // Get the instruction's basic block. 678 MachineBasicBlock &MBB = *MI.getParent(); 679 DebugLoc dl = MI.getDebugLoc(); 680 681 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 682 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 683 const TargetRegisterClass *RC = Subtarget.isPPC64() ? G8RC : GPRC; 684 unsigned Reg = findScratchRegister(II, RS, RC, SPAdj); 685 686 // We need to store the CR in the low 4-bits of the saved value. First, issue 687 // an MFCR to save all of the CRBits. Add an implicit kill of the CR. 688 if (!MI.getOperand(0).isKill()) 689 BuildMI(MBB, II, dl, TII.get(PPC::MFCR), Reg); 690 else 691 // Implicitly kill the CR register. 692 BuildMI(MBB, II, dl, TII.get(PPC::MFCR), Reg) 693 .addReg(MI.getOperand(0).getReg(), RegState::ImplicitKill); 694 695 // If the saved register wasn't CR0, shift the bits left so that they are in 696 // CR0's slot. 697 unsigned SrcReg = MI.getOperand(0).getReg(); 698 if (SrcReg != PPC::CR0) 699 // rlwinm rA, rA, ShiftBits, 0, 31. 700 BuildMI(MBB, II, dl, TII.get(PPC::RLWINM), Reg) 701 .addReg(Reg, RegState::Kill) 702 .addImm(PPCRegisterInfo::getRegisterNumbering(SrcReg) * 4) 703 .addImm(0) 704 .addImm(31); 705 706 addFrameReference(BuildMI(MBB, II, dl, TII.get(PPC::STW)) 707 .addReg(Reg, getKillRegState(MI.getOperand(1).getImm())), 708 FrameIndex); 709 710 // Discard the pseudo instruction. 711 MBB.erase(II); 712 } 713 714 unsigned 715 PPCRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, 716 int SPAdj, FrameIndexValue *Value, 717 RegScavenger *RS) const { 718 assert(SPAdj == 0 && "Unexpected"); 719 720 // Get the instruction. 721 MachineInstr &MI = *II; 722 // Get the instruction's basic block. 723 MachineBasicBlock &MBB = *MI.getParent(); 724 // Get the basic block's function. 725 MachineFunction &MF = *MBB.getParent(); 726 // Get the frame info. 727 MachineFrameInfo *MFI = MF.getFrameInfo(); 728 DebugLoc dl = MI.getDebugLoc(); 729 730 // Find out which operand is the frame index. 731 unsigned FIOperandNo = 0; 732 while (!MI.getOperand(FIOperandNo).isFI()) { 733 ++FIOperandNo; 734 assert(FIOperandNo != MI.getNumOperands() && 735 "Instr doesn't have FrameIndex operand!"); 736 } 737 // Take into account whether it's an add or mem instruction 738 unsigned OffsetOperandNo = (FIOperandNo == 2) ? 1 : 2; 739 if (MI.isInlineAsm()) 740 OffsetOperandNo = FIOperandNo-1; 741 742 // Get the frame index. 743 int FrameIndex = MI.getOperand(FIOperandNo).getIndex(); 744 745 // Get the frame pointer save index. Users of this index are primarily 746 // DYNALLOC instructions. 747 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>(); 748 int FPSI = FI->getFramePointerSaveIndex(); 749 // Get the instruction opcode. 750 unsigned OpC = MI.getOpcode(); 751 752 // Special case for dynamic alloca. 753 if (FPSI && FrameIndex == FPSI && 754 (OpC == PPC::DYNALLOC || OpC == PPC::DYNALLOC8)) { 755 lowerDynamicAlloc(II, SPAdj, RS); 756 return 0; 757 } 758 759 // Special case for pseudo-op SPILL_CR. 760 if (EnableRegisterScavenging) // FIXME (64-bit): Enable by default. 761 if (OpC == PPC::SPILL_CR) { 762 lowerCRSpilling(II, FrameIndex, SPAdj, RS); 763 return 0; 764 } 765 766 // Replace the FrameIndex with base register with GPR1 (SP) or GPR31 (FP). 767 MI.getOperand(FIOperandNo).ChangeToRegister(hasFP(MF) ? PPC::R31 : PPC::R1, 768 false); 769 770 // Figure out if the offset in the instruction is shifted right two bits. This 771 // is true for instructions like "STD", which the machine implicitly adds two 772 // low zeros to. 773 bool isIXAddr = false; 774 switch (OpC) { 775 case PPC::LWA: 776 case PPC::LD: 777 case PPC::STD: 778 case PPC::STD_32: 779 isIXAddr = true; 780 break; 781 } 782 783 // Now add the frame object offset to the offset from r1. 784 int Offset = MFI->getObjectOffset(FrameIndex); 785 if (!isIXAddr) 786 Offset += MI.getOperand(OffsetOperandNo).getImm(); 787 else 788 Offset += MI.getOperand(OffsetOperandNo).getImm() << 2; 789 790 // If we're not using a Frame Pointer that has been set to the value of the 791 // SP before having the stack size subtracted from it, then add the stack size 792 // to Offset to get the correct offset. 793 Offset += MFI->getStackSize(); 794 795 // If we can, encode the offset directly into the instruction. If this is a 796 // normal PPC "ri" instruction, any 16-bit value can be safely encoded. If 797 // this is a PPC64 "ix" instruction, only a 16-bit value with the low two bits 798 // clear can be encoded. This is extremely uncommon, because normally you 799 // only "std" to a stack slot that is at least 4-byte aligned, but it can 800 // happen in invalid code. 801 if (isInt<16>(Offset) && (!isIXAddr || (Offset & 3) == 0)) { 802 if (isIXAddr) 803 Offset >>= 2; // The actual encoded value has the low two bits zero. 804 MI.getOperand(OffsetOperandNo).ChangeToImmediate(Offset); 805 return 0; 806 } 807 808 // The offset doesn't fit into a single register, scavenge one to build the 809 // offset in. 810 // FIXME: figure out what SPAdj is doing here. 811 812 // FIXME (64-bit): Use "findScratchRegister". 813 unsigned SReg; 814 if (EnableRegisterScavenging) 815 SReg = findScratchRegister(II, RS, &PPC::GPRCRegClass, SPAdj); 816 else 817 SReg = PPC::R0; 818 819 // Insert a set of rA with the full offset value before the ld, st, or add 820 BuildMI(MBB, II, dl, TII.get(PPC::LIS), SReg) 821 .addImm(Offset >> 16); 822 BuildMI(MBB, II, dl, TII.get(PPC::ORI), SReg) 823 .addReg(SReg, RegState::Kill) 824 .addImm(Offset); 825 826 // Convert into indexed form of the instruction: 827 // 828 // sth 0:rA, 1:imm 2:(rB) ==> sthx 0:rA, 2:rB, 1:r0 829 // addi 0:rA 1:rB, 2, imm ==> add 0:rA, 1:rB, 2:r0 830 unsigned OperandBase; 831 832 if (OpC != TargetOpcode::INLINEASM) { 833 assert(ImmToIdxMap.count(OpC) && 834 "No indexed form of load or store available!"); 835 unsigned NewOpcode = ImmToIdxMap.find(OpC)->second; 836 MI.setDesc(TII.get(NewOpcode)); 837 OperandBase = 1; 838 } else { 839 OperandBase = OffsetOperandNo; 840 } 841 842 unsigned StackReg = MI.getOperand(FIOperandNo).getReg(); 843 MI.getOperand(OperandBase).ChangeToRegister(StackReg, false); 844 MI.getOperand(OperandBase + 1).ChangeToRegister(SReg, false); 845 return 0; 846 } 847 848 /// VRRegNo - Map from a numbered VR register to its enum value. 849 /// 850 static const unsigned short VRRegNo[] = { 851 PPC::V0 , PPC::V1 , PPC::V2 , PPC::V3 , PPC::V4 , PPC::V5 , PPC::V6 , PPC::V7 , 852 PPC::V8 , PPC::V9 , PPC::V10, PPC::V11, PPC::V12, PPC::V13, PPC::V14, PPC::V15, 853 PPC::V16, PPC::V17, PPC::V18, PPC::V19, PPC::V20, PPC::V21, PPC::V22, PPC::V23, 854 PPC::V24, PPC::V25, PPC::V26, PPC::V27, PPC::V28, PPC::V29, PPC::V30, PPC::V31 855 }; 856 857 /// RemoveVRSaveCode - We have found that this function does not need any code 858 /// to manipulate the VRSAVE register, even though it uses vector registers. 859 /// This can happen when the only registers used are known to be live in or out 860 /// of the function. Remove all of the VRSAVE related code from the function. 861 static void RemoveVRSaveCode(MachineInstr *MI) { 862 MachineBasicBlock *Entry = MI->getParent(); 863 MachineFunction *MF = Entry->getParent(); 864 865 // We know that the MTVRSAVE instruction immediately follows MI. Remove it. 866 MachineBasicBlock::iterator MBBI = MI; 867 ++MBBI; 868 assert(MBBI != Entry->end() && MBBI->getOpcode() == PPC::MTVRSAVE); 869 MBBI->eraseFromParent(); 870 871 bool RemovedAllMTVRSAVEs = true; 872 // See if we can find and remove the MTVRSAVE instruction from all of the 873 // epilog blocks. 874 for (MachineFunction::iterator I = MF->begin(), E = MF->end(); I != E; ++I) { 875 // If last instruction is a return instruction, add an epilogue 876 if (!I->empty() && I->back().getDesc().isReturn()) { 877 bool FoundIt = false; 878 for (MBBI = I->end(); MBBI != I->begin(); ) { 879 --MBBI; 880 if (MBBI->getOpcode() == PPC::MTVRSAVE) { 881 MBBI->eraseFromParent(); // remove it. 882 FoundIt = true; 883 break; 884 } 885 } 886 RemovedAllMTVRSAVEs &= FoundIt; 887 } 888 } 889 890 // If we found and removed all MTVRSAVE instructions, remove the read of 891 // VRSAVE as well. 892 if (RemovedAllMTVRSAVEs) { 893 MBBI = MI; 894 assert(MBBI != Entry->begin() && "UPDATE_VRSAVE is first instr in block?"); 895 --MBBI; 896 assert(MBBI->getOpcode() == PPC::MFVRSAVE && "VRSAVE instrs wandered?"); 897 MBBI->eraseFromParent(); 898 } 899 900 // Finally, nuke the UPDATE_VRSAVE. 901 MI->eraseFromParent(); 902 } 903 904 // HandleVRSaveUpdate - MI is the UPDATE_VRSAVE instruction introduced by the 905 // instruction selector. Based on the vector registers that have been used, 906 // transform this into the appropriate ORI instruction. 907 static void HandleVRSaveUpdate(MachineInstr *MI, const TargetInstrInfo &TII) { 908 MachineFunction *MF = MI->getParent()->getParent(); 909 DebugLoc dl = MI->getDebugLoc(); 910 911 unsigned UsedRegMask = 0; 912 for (unsigned i = 0; i != 32; ++i) 913 if (MF->getRegInfo().isPhysRegUsed(VRRegNo[i])) 914 UsedRegMask |= 1 << (31-i); 915 916 // Live in and live out values already must be in the mask, so don't bother 917 // marking them. 918 for (MachineRegisterInfo::livein_iterator 919 I = MF->getRegInfo().livein_begin(), 920 E = MF->getRegInfo().livein_end(); I != E; ++I) { 921 unsigned RegNo = PPCRegisterInfo::getRegisterNumbering(I->first); 922 if (VRRegNo[RegNo] == I->first) // If this really is a vector reg. 923 UsedRegMask &= ~(1 << (31-RegNo)); // Doesn't need to be marked. 924 } 925 for (MachineRegisterInfo::liveout_iterator 926 I = MF->getRegInfo().liveout_begin(), 927 E = MF->getRegInfo().liveout_end(); I != E; ++I) { 928 unsigned RegNo = PPCRegisterInfo::getRegisterNumbering(*I); 929 if (VRRegNo[RegNo] == *I) // If this really is a vector reg. 930 UsedRegMask &= ~(1 << (31-RegNo)); // Doesn't need to be marked. 931 } 932 933 // If no registers are used, turn this into a copy. 934 if (UsedRegMask == 0) { 935 // Remove all VRSAVE code. 936 RemoveVRSaveCode(MI); 937 return; 938 } 939 940 unsigned SrcReg = MI->getOperand(1).getReg(); 941 unsigned DstReg = MI->getOperand(0).getReg(); 942 943 if ((UsedRegMask & 0xFFFF) == UsedRegMask) { 944 if (DstReg != SrcReg) 945 BuildMI(*MI->getParent(), MI, dl, TII.get(PPC::ORI), DstReg) 946 .addReg(SrcReg) 947 .addImm(UsedRegMask); 948 else 949 BuildMI(*MI->getParent(), MI, dl, TII.get(PPC::ORI), DstReg) 950 .addReg(SrcReg, RegState::Kill) 951 .addImm(UsedRegMask); 952 } else if ((UsedRegMask & 0xFFFF0000) == UsedRegMask) { 953 if (DstReg != SrcReg) 954 BuildMI(*MI->getParent(), MI, dl, TII.get(PPC::ORIS), DstReg) 955 .addReg(SrcReg) 956 .addImm(UsedRegMask >> 16); 957 else 958 BuildMI(*MI->getParent(), MI, dl, TII.get(PPC::ORIS), DstReg) 959 .addReg(SrcReg, RegState::Kill) 960 .addImm(UsedRegMask >> 16); 961 } else { 962 if (DstReg != SrcReg) 963 BuildMI(*MI->getParent(), MI, dl, TII.get(PPC::ORIS), DstReg) 964 .addReg(SrcReg) 965 .addImm(UsedRegMask >> 16); 966 else 967 BuildMI(*MI->getParent(), MI, dl, TII.get(PPC::ORIS), DstReg) 968 .addReg(SrcReg, RegState::Kill) 969 .addImm(UsedRegMask >> 16); 970 971 BuildMI(*MI->getParent(), MI, dl, TII.get(PPC::ORI), DstReg) 972 .addReg(DstReg, RegState::Kill) 973 .addImm(UsedRegMask & 0xFFFF); 974 } 975 976 // Remove the old UPDATE_VRSAVE instruction. 977 MI->eraseFromParent(); 978 } 979 980 /// determineFrameLayout - Determine the size of the frame and maximum call 981 /// frame size. 982 void PPCRegisterInfo::determineFrameLayout(MachineFunction &MF) const { 983 MachineFrameInfo *MFI = MF.getFrameInfo(); 984 985 // Get the number of bytes to allocate from the FrameInfo 986 unsigned FrameSize = MFI->getStackSize(); 987 988 // Get the alignments provided by the target, and the maximum alignment 989 // (if any) of the fixed frame objects. 990 unsigned MaxAlign = MFI->getMaxAlignment(); 991 unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment(); 992 unsigned AlignMask = TargetAlign - 1; // 993 994 // If we are a leaf function, and use up to 224 bytes of stack space, 995 // don't have a frame pointer, calls, or dynamic alloca then we do not need 996 // to adjust the stack pointer (we fit in the Red Zone). 997 bool DisableRedZone = MF.getFunction()->hasFnAttr(Attribute::NoRedZone); 998 // FIXME SVR4 The 32-bit SVR4 ABI has no red zone. 999 if (!DisableRedZone && 1000 FrameSize <= 224 && // Fits in red zone. 1001 !MFI->hasVarSizedObjects() && // No dynamic alloca. 1002 !MFI->hasCalls() && // No calls. 1003 (!ALIGN_STACK || MaxAlign <= TargetAlign)) { // No special alignment. 1004 // No need for frame 1005 MFI->setStackSize(0); 1006 return; 1007 } 1008 1009 // Get the maximum call frame size of all the calls. 1010 unsigned maxCallFrameSize = MFI->getMaxCallFrameSize(); 1011 1012 // Maximum call frame needs to be at least big enough for linkage and 8 args. 1013 unsigned minCallFrameSize = 1014 PPCFrameInfo::getMinCallFrameSize(Subtarget.isPPC64(), 1015 Subtarget.isDarwinABI()); 1016 maxCallFrameSize = std::max(maxCallFrameSize, minCallFrameSize); 1017 1018 // If we have dynamic alloca then maxCallFrameSize needs to be aligned so 1019 // that allocations will be aligned. 1020 if (MFI->hasVarSizedObjects()) 1021 maxCallFrameSize = (maxCallFrameSize + AlignMask) & ~AlignMask; 1022 1023 // Update maximum call frame size. 1024 MFI->setMaxCallFrameSize(maxCallFrameSize); 1025 1026 // Include call frame size in total. 1027 FrameSize += maxCallFrameSize; 1028 1029 // Make sure the frame is aligned. 1030 FrameSize = (FrameSize + AlignMask) & ~AlignMask; 1031 1032 // Update frame info. 1033 MFI->setStackSize(FrameSize); 1034 } 1035 1036 void 1037 PPCRegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF, 1038 RegScavenger *RS) const { 1039 // Save and clear the LR state. 1040 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>(); 1041 unsigned LR = getRARegister(); 1042 FI->setMustSaveLR(MustSaveLR(MF, LR)); 1043 MF.getRegInfo().setPhysRegUnused(LR); 1044 1045 // Save R31 if necessary 1046 int FPSI = FI->getFramePointerSaveIndex(); 1047 bool isPPC64 = Subtarget.isPPC64(); 1048 bool isDarwinABI = Subtarget.isDarwinABI(); 1049 MachineFrameInfo *MFI = MF.getFrameInfo(); 1050 1051 // If the frame pointer save index hasn't been defined yet. 1052 if (!FPSI && needsFP(MF)) { 1053 // Find out what the fix offset of the frame pointer save area. 1054 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(isPPC64, 1055 isDarwinABI); 1056 // Allocate the frame index for frame pointer save area. 1057 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, 1058 true, false); 1059 // Save the result. 1060 FI->setFramePointerSaveIndex(FPSI); 1061 } 1062 1063 // Reserve stack space to move the linkage area to in case of a tail call. 1064 int TCSPDelta = 0; 1065 if (GuaranteedTailCallOpt && (TCSPDelta = FI->getTailCallSPDelta()) < 0) { 1066 MF.getFrameInfo()->CreateFixedObject(-1 * TCSPDelta, TCSPDelta, 1067 true, false); 1068 } 1069 1070 // Reserve a slot closest to SP or frame pointer if we have a dynalloc or 1071 // a large stack, which will require scavenging a register to materialize a 1072 // large offset. 1073 // FIXME: this doesn't actually check stack size, so is a bit pessimistic 1074 // FIXME: doesn't detect whether or not we need to spill vXX, which requires 1075 // r0 for now. 1076 1077 if (EnableRegisterScavenging) // FIXME (64-bit): Enable. 1078 if (needsFP(MF) || spillsCR(MF)) { 1079 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 1080 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 1081 const TargetRegisterClass *RC = isPPC64 ? G8RC : GPRC; 1082 RS->setScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(), 1083 RC->getAlignment(), 1084 false)); 1085 } 1086 } 1087 1088 void 1089 PPCRegisterInfo::processFunctionBeforeFrameFinalized(MachineFunction &MF) 1090 const { 1091 // Early exit if not using the SVR4 ABI. 1092 if (!Subtarget.isSVR4ABI()) { 1093 return; 1094 } 1095 1096 // Get callee saved register information. 1097 MachineFrameInfo *FFI = MF.getFrameInfo(); 1098 const std::vector<CalleeSavedInfo> &CSI = FFI->getCalleeSavedInfo(); 1099 1100 // Early exit if no callee saved registers are modified! 1101 if (CSI.empty() && !needsFP(MF)) { 1102 return; 1103 } 1104 1105 unsigned MinGPR = PPC::R31; 1106 unsigned MinG8R = PPC::X31; 1107 unsigned MinFPR = PPC::F31; 1108 unsigned MinVR = PPC::V31; 1109 1110 bool HasGPSaveArea = false; 1111 bool HasG8SaveArea = false; 1112 bool HasFPSaveArea = false; 1113 bool HasCRSaveArea = false; 1114 bool HasVRSAVESaveArea = false; 1115 bool HasVRSaveArea = false; 1116 1117 SmallVector<CalleeSavedInfo, 18> GPRegs; 1118 SmallVector<CalleeSavedInfo, 18> G8Regs; 1119 SmallVector<CalleeSavedInfo, 18> FPRegs; 1120 SmallVector<CalleeSavedInfo, 18> VRegs; 1121 1122 for (unsigned i = 0, e = CSI.size(); i != e; ++i) { 1123 unsigned Reg = CSI[i].getReg(); 1124 const TargetRegisterClass *RC = CSI[i].getRegClass(); 1125 1126 if (RC == PPC::GPRCRegisterClass) { 1127 HasGPSaveArea = true; 1128 1129 GPRegs.push_back(CSI[i]); 1130 1131 if (Reg < MinGPR) { 1132 MinGPR = Reg; 1133 } 1134 } else if (RC == PPC::G8RCRegisterClass) { 1135 HasG8SaveArea = true; 1136 1137 G8Regs.push_back(CSI[i]); 1138 1139 if (Reg < MinG8R) { 1140 MinG8R = Reg; 1141 } 1142 } else if (RC == PPC::F8RCRegisterClass) { 1143 HasFPSaveArea = true; 1144 1145 FPRegs.push_back(CSI[i]); 1146 1147 if (Reg < MinFPR) { 1148 MinFPR = Reg; 1149 } 1150 // FIXME SVR4: Disable CR save area for now. 1151 } else if ( RC == PPC::CRBITRCRegisterClass 1152 || RC == PPC::CRRCRegisterClass) { 1153 // HasCRSaveArea = true; 1154 } else if (RC == PPC::VRSAVERCRegisterClass) { 1155 HasVRSAVESaveArea = true; 1156 } else if (RC == PPC::VRRCRegisterClass) { 1157 HasVRSaveArea = true; 1158 1159 VRegs.push_back(CSI[i]); 1160 1161 if (Reg < MinVR) { 1162 MinVR = Reg; 1163 } 1164 } else { 1165 llvm_unreachable("Unknown RegisterClass!"); 1166 } 1167 } 1168 1169 PPCFunctionInfo *PFI = MF.getInfo<PPCFunctionInfo>(); 1170 1171 int64_t LowerBound = 0; 1172 1173 // Take into account stack space reserved for tail calls. 1174 int TCSPDelta = 0; 1175 if (GuaranteedTailCallOpt && (TCSPDelta = PFI->getTailCallSPDelta()) < 0) { 1176 LowerBound = TCSPDelta; 1177 } 1178 1179 // The Floating-point register save area is right below the back chain word 1180 // of the previous stack frame. 1181 if (HasFPSaveArea) { 1182 for (unsigned i = 0, e = FPRegs.size(); i != e; ++i) { 1183 int FI = FPRegs[i].getFrameIdx(); 1184 1185 FFI->setObjectOffset(FI, LowerBound + FFI->getObjectOffset(FI)); 1186 } 1187 1188 LowerBound -= (31 - getRegisterNumbering(MinFPR) + 1) * 8; 1189 } 1190 1191 // Check whether the frame pointer register is allocated. If so, make sure it 1192 // is spilled to the correct offset. 1193 if (needsFP(MF)) { 1194 HasGPSaveArea = true; 1195 1196 int FI = PFI->getFramePointerSaveIndex(); 1197 assert(FI && "No Frame Pointer Save Slot!"); 1198 1199 FFI->setObjectOffset(FI, LowerBound + FFI->getObjectOffset(FI)); 1200 } 1201 1202 // General register save area starts right below the Floating-point 1203 // register save area. 1204 if (HasGPSaveArea || HasG8SaveArea) { 1205 // Move general register save area spill slots down, taking into account 1206 // the size of the Floating-point register save area. 1207 for (unsigned i = 0, e = GPRegs.size(); i != e; ++i) { 1208 int FI = GPRegs[i].getFrameIdx(); 1209 1210 FFI->setObjectOffset(FI, LowerBound + FFI->getObjectOffset(FI)); 1211 } 1212 1213 // Move general register save area spill slots down, taking into account 1214 // the size of the Floating-point register save area. 1215 for (unsigned i = 0, e = G8Regs.size(); i != e; ++i) { 1216 int FI = G8Regs[i].getFrameIdx(); 1217 1218 FFI->setObjectOffset(FI, LowerBound + FFI->getObjectOffset(FI)); 1219 } 1220 1221 unsigned MinReg = std::min<unsigned>(getRegisterNumbering(MinGPR), 1222 getRegisterNumbering(MinG8R)); 1223 1224 if (Subtarget.isPPC64()) { 1225 LowerBound -= (31 - MinReg + 1) * 8; 1226 } else { 1227 LowerBound -= (31 - MinReg + 1) * 4; 1228 } 1229 } 1230 1231 // The CR save area is below the general register save area. 1232 if (HasCRSaveArea) { 1233 // FIXME SVR4: Is it actually possible to have multiple elements in CSI 1234 // which have the CR/CRBIT register class? 1235 // Adjust the frame index of the CR spill slot. 1236 for (unsigned i = 0, e = CSI.size(); i != e; ++i) { 1237 const TargetRegisterClass *RC = CSI[i].getRegClass(); 1238 1239 if (RC == PPC::CRBITRCRegisterClass || RC == PPC::CRRCRegisterClass) { 1240 int FI = CSI[i].getFrameIdx(); 1241 1242 FFI->setObjectOffset(FI, LowerBound + FFI->getObjectOffset(FI)); 1243 } 1244 } 1245 1246 LowerBound -= 4; // The CR save area is always 4 bytes long. 1247 } 1248 1249 if (HasVRSAVESaveArea) { 1250 // FIXME SVR4: Is it actually possible to have multiple elements in CSI 1251 // which have the VRSAVE register class? 1252 // Adjust the frame index of the VRSAVE spill slot. 1253 for (unsigned i = 0, e = CSI.size(); i != e; ++i) { 1254 const TargetRegisterClass *RC = CSI[i].getRegClass(); 1255 1256 if (RC == PPC::VRSAVERCRegisterClass) { 1257 int FI = CSI[i].getFrameIdx(); 1258 1259 FFI->setObjectOffset(FI, LowerBound + FFI->getObjectOffset(FI)); 1260 } 1261 } 1262 1263 LowerBound -= 4; // The VRSAVE save area is always 4 bytes long. 1264 } 1265 1266 if (HasVRSaveArea) { 1267 // Insert alignment padding, we need 16-byte alignment. 1268 LowerBound = (LowerBound - 15) & ~(15); 1269 1270 for (unsigned i = 0, e = VRegs.size(); i != e; ++i) { 1271 int FI = VRegs[i].getFrameIdx(); 1272 1273 FFI->setObjectOffset(FI, LowerBound + FFI->getObjectOffset(FI)); 1274 } 1275 } 1276 } 1277 1278 void 1279 PPCRegisterInfo::emitPrologue(MachineFunction &MF) const { 1280 MachineBasicBlock &MBB = MF.front(); // Prolog goes in entry BB 1281 MachineBasicBlock::iterator MBBI = MBB.begin(); 1282 MachineFrameInfo *MFI = MF.getFrameInfo(); 1283 MachineModuleInfo &MMI = MF.getMMI(); 1284 DebugLoc dl; 1285 bool needsFrameMoves = MMI.hasDebugInfo() || 1286 !MF.getFunction()->doesNotThrow() || 1287 UnwindTablesMandatory; 1288 1289 // Prepare for frame info. 1290 MCSymbol *FrameLabel = 0; 1291 1292 // Scan the prolog, looking for an UPDATE_VRSAVE instruction. If we find it, 1293 // process it. 1294 for (unsigned i = 0; MBBI != MBB.end(); ++i, ++MBBI) { 1295 if (MBBI->getOpcode() == PPC::UPDATE_VRSAVE) { 1296 HandleVRSaveUpdate(MBBI, TII); 1297 break; 1298 } 1299 } 1300 1301 // Move MBBI back to the beginning of the function. 1302 MBBI = MBB.begin(); 1303 1304 // Work out frame sizes. 1305 determineFrameLayout(MF); 1306 unsigned FrameSize = MFI->getStackSize(); 1307 1308 int NegFrameSize = -FrameSize; 1309 1310 // Get processor type. 1311 bool isPPC64 = Subtarget.isPPC64(); 1312 // Get operating system 1313 bool isDarwinABI = Subtarget.isDarwinABI(); 1314 // Check if the link register (LR) must be saved. 1315 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>(); 1316 bool MustSaveLR = FI->mustSaveLR(); 1317 // Do we have a frame pointer for this function? 1318 bool HasFP = hasFP(MF) && FrameSize; 1319 1320 int LROffset = PPCFrameInfo::getReturnSaveOffset(isPPC64, isDarwinABI); 1321 1322 int FPOffset = 0; 1323 if (HasFP) { 1324 if (Subtarget.isSVR4ABI()) { 1325 MachineFrameInfo *FFI = MF.getFrameInfo(); 1326 int FPIndex = FI->getFramePointerSaveIndex(); 1327 assert(FPIndex && "No Frame Pointer Save Slot!"); 1328 FPOffset = FFI->getObjectOffset(FPIndex); 1329 } else { 1330 FPOffset = PPCFrameInfo::getFramePointerSaveOffset(isPPC64, isDarwinABI); 1331 } 1332 } 1333 1334 if (isPPC64) { 1335 if (MustSaveLR) 1336 BuildMI(MBB, MBBI, dl, TII.get(PPC::MFLR8), PPC::X0); 1337 1338 if (HasFP) 1339 BuildMI(MBB, MBBI, dl, TII.get(PPC::STD)) 1340 .addReg(PPC::X31) 1341 .addImm(FPOffset/4) 1342 .addReg(PPC::X1); 1343 1344 if (MustSaveLR) 1345 BuildMI(MBB, MBBI, dl, TII.get(PPC::STD)) 1346 .addReg(PPC::X0) 1347 .addImm(LROffset / 4) 1348 .addReg(PPC::X1); 1349 } else { 1350 if (MustSaveLR) 1351 BuildMI(MBB, MBBI, dl, TII.get(PPC::MFLR), PPC::R0); 1352 1353 if (HasFP) 1354 BuildMI(MBB, MBBI, dl, TII.get(PPC::STW)) 1355 .addReg(PPC::R31) 1356 .addImm(FPOffset) 1357 .addReg(PPC::R1); 1358 1359 if (MustSaveLR) 1360 BuildMI(MBB, MBBI, dl, TII.get(PPC::STW)) 1361 .addReg(PPC::R0) 1362 .addImm(LROffset) 1363 .addReg(PPC::R1); 1364 } 1365 1366 // Skip if a leaf routine. 1367 if (!FrameSize) return; 1368 1369 // Get stack alignments. 1370 unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment(); 1371 unsigned MaxAlign = MFI->getMaxAlignment(); 1372 1373 // Adjust stack pointer: r1 += NegFrameSize. 1374 // If there is a preferred stack alignment, align R1 now 1375 if (!isPPC64) { 1376 // PPC32. 1377 if (ALIGN_STACK && MaxAlign > TargetAlign) { 1378 assert(isPowerOf2_32(MaxAlign) && isInt<16>(MaxAlign) && 1379 "Invalid alignment!"); 1380 assert(isInt<16>(NegFrameSize) && "Unhandled stack size and alignment!"); 1381 1382 BuildMI(MBB, MBBI, dl, TII.get(PPC::RLWINM), PPC::R0) 1383 .addReg(PPC::R1) 1384 .addImm(0) 1385 .addImm(32 - Log2_32(MaxAlign)) 1386 .addImm(31); 1387 BuildMI(MBB, MBBI, dl, TII.get(PPC::SUBFIC) ,PPC::R0) 1388 .addReg(PPC::R0, RegState::Kill) 1389 .addImm(NegFrameSize); 1390 BuildMI(MBB, MBBI, dl, TII.get(PPC::STWUX)) 1391 .addReg(PPC::R1) 1392 .addReg(PPC::R1) 1393 .addReg(PPC::R0); 1394 } else if (isInt<16>(NegFrameSize)) { 1395 BuildMI(MBB, MBBI, dl, TII.get(PPC::STWU), PPC::R1) 1396 .addReg(PPC::R1) 1397 .addImm(NegFrameSize) 1398 .addReg(PPC::R1); 1399 } else { 1400 BuildMI(MBB, MBBI, dl, TII.get(PPC::LIS), PPC::R0) 1401 .addImm(NegFrameSize >> 16); 1402 BuildMI(MBB, MBBI, dl, TII.get(PPC::ORI), PPC::R0) 1403 .addReg(PPC::R0, RegState::Kill) 1404 .addImm(NegFrameSize & 0xFFFF); 1405 BuildMI(MBB, MBBI, dl, TII.get(PPC::STWUX)) 1406 .addReg(PPC::R1) 1407 .addReg(PPC::R1) 1408 .addReg(PPC::R0); 1409 } 1410 } else { // PPC64. 1411 if (ALIGN_STACK && MaxAlign > TargetAlign) { 1412 assert(isPowerOf2_32(MaxAlign) && isInt<16>(MaxAlign) && 1413 "Invalid alignment!"); 1414 assert(isInt<16>(NegFrameSize) && "Unhandled stack size and alignment!"); 1415 1416 BuildMI(MBB, MBBI, dl, TII.get(PPC::RLDICL), PPC::X0) 1417 .addReg(PPC::X1) 1418 .addImm(0) 1419 .addImm(64 - Log2_32(MaxAlign)); 1420 BuildMI(MBB, MBBI, dl, TII.get(PPC::SUBFIC8), PPC::X0) 1421 .addReg(PPC::X0) 1422 .addImm(NegFrameSize); 1423 BuildMI(MBB, MBBI, dl, TII.get(PPC::STDUX)) 1424 .addReg(PPC::X1) 1425 .addReg(PPC::X1) 1426 .addReg(PPC::X0); 1427 } else if (isInt<16>(NegFrameSize)) { 1428 BuildMI(MBB, MBBI, dl, TII.get(PPC::STDU), PPC::X1) 1429 .addReg(PPC::X1) 1430 .addImm(NegFrameSize / 4) 1431 .addReg(PPC::X1); 1432 } else { 1433 BuildMI(MBB, MBBI, dl, TII.get(PPC::LIS8), PPC::X0) 1434 .addImm(NegFrameSize >> 16); 1435 BuildMI(MBB, MBBI, dl, TII.get(PPC::ORI8), PPC::X0) 1436 .addReg(PPC::X0, RegState::Kill) 1437 .addImm(NegFrameSize & 0xFFFF); 1438 BuildMI(MBB, MBBI, dl, TII.get(PPC::STDUX)) 1439 .addReg(PPC::X1) 1440 .addReg(PPC::X1) 1441 .addReg(PPC::X0); 1442 } 1443 } 1444 1445 std::vector<MachineMove> &Moves = MMI.getFrameMoves(); 1446 1447 // Add the "machine moves" for the instructions we generated above, but in 1448 // reverse order. 1449 if (needsFrameMoves) { 1450 // Mark effective beginning of when frame pointer becomes valid. 1451 FrameLabel = MMI.getContext().CreateTempSymbol(); 1452 BuildMI(MBB, MBBI, dl, TII.get(PPC::DBG_LABEL)).addSym(FrameLabel); 1453 1454 // Show update of SP. 1455 if (NegFrameSize) { 1456 MachineLocation SPDst(MachineLocation::VirtualFP); 1457 MachineLocation SPSrc(MachineLocation::VirtualFP, NegFrameSize); 1458 Moves.push_back(MachineMove(FrameLabel, SPDst, SPSrc)); 1459 } else { 1460 MachineLocation SP(isPPC64 ? PPC::X31 : PPC::R31); 1461 Moves.push_back(MachineMove(FrameLabel, SP, SP)); 1462 } 1463 1464 if (HasFP) { 1465 MachineLocation FPDst(MachineLocation::VirtualFP, FPOffset); 1466 MachineLocation FPSrc(isPPC64 ? PPC::X31 : PPC::R31); 1467 Moves.push_back(MachineMove(FrameLabel, FPDst, FPSrc)); 1468 } 1469 1470 if (MustSaveLR) { 1471 MachineLocation LRDst(MachineLocation::VirtualFP, LROffset); 1472 MachineLocation LRSrc(isPPC64 ? PPC::LR8 : PPC::LR); 1473 Moves.push_back(MachineMove(FrameLabel, LRDst, LRSrc)); 1474 } 1475 } 1476 1477 MCSymbol *ReadyLabel = 0; 1478 1479 // If there is a frame pointer, copy R1 into R31 1480 if (HasFP) { 1481 if (!isPPC64) { 1482 BuildMI(MBB, MBBI, dl, TII.get(PPC::OR), PPC::R31) 1483 .addReg(PPC::R1) 1484 .addReg(PPC::R1); 1485 } else { 1486 BuildMI(MBB, MBBI, dl, TII.get(PPC::OR8), PPC::X31) 1487 .addReg(PPC::X1) 1488 .addReg(PPC::X1); 1489 } 1490 1491 if (needsFrameMoves) { 1492 ReadyLabel = MMI.getContext().CreateTempSymbol(); 1493 1494 // Mark effective beginning of when frame pointer is ready. 1495 BuildMI(MBB, MBBI, dl, TII.get(PPC::DBG_LABEL)).addSym(ReadyLabel); 1496 1497 MachineLocation FPDst(HasFP ? (isPPC64 ? PPC::X31 : PPC::R31) : 1498 (isPPC64 ? PPC::X1 : PPC::R1)); 1499 MachineLocation FPSrc(MachineLocation::VirtualFP); 1500 Moves.push_back(MachineMove(ReadyLabel, FPDst, FPSrc)); 1501 } 1502 } 1503 1504 if (needsFrameMoves) { 1505 MCSymbol *Label = HasFP ? ReadyLabel : FrameLabel; 1506 1507 // Add callee saved registers to move list. 1508 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo(); 1509 for (unsigned I = 0, E = CSI.size(); I != E; ++I) { 1510 int Offset = MFI->getObjectOffset(CSI[I].getFrameIdx()); 1511 unsigned Reg = CSI[I].getReg(); 1512 if (Reg == PPC::LR || Reg == PPC::LR8 || Reg == PPC::RM) continue; 1513 MachineLocation CSDst(MachineLocation::VirtualFP, Offset); 1514 MachineLocation CSSrc(Reg); 1515 Moves.push_back(MachineMove(Label, CSDst, CSSrc)); 1516 } 1517 } 1518 } 1519 1520 void PPCRegisterInfo::emitEpilogue(MachineFunction &MF, 1521 MachineBasicBlock &MBB) const { 1522 MachineBasicBlock::iterator MBBI = prior(MBB.end()); 1523 unsigned RetOpcode = MBBI->getOpcode(); 1524 DebugLoc dl; 1525 1526 assert( (RetOpcode == PPC::BLR || 1527 RetOpcode == PPC::TCRETURNri || 1528 RetOpcode == PPC::TCRETURNdi || 1529 RetOpcode == PPC::TCRETURNai || 1530 RetOpcode == PPC::TCRETURNri8 || 1531 RetOpcode == PPC::TCRETURNdi8 || 1532 RetOpcode == PPC::TCRETURNai8) && 1533 "Can only insert epilog into returning blocks"); 1534 1535 // Get alignment info so we know how to restore r1 1536 const MachineFrameInfo *MFI = MF.getFrameInfo(); 1537 unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment(); 1538 unsigned MaxAlign = MFI->getMaxAlignment(); 1539 1540 // Get the number of bytes allocated from the FrameInfo. 1541 int FrameSize = MFI->getStackSize(); 1542 1543 // Get processor type. 1544 bool isPPC64 = Subtarget.isPPC64(); 1545 // Get operating system 1546 bool isDarwinABI = Subtarget.isDarwinABI(); 1547 // Check if the link register (LR) has been saved. 1548 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>(); 1549 bool MustSaveLR = FI->mustSaveLR(); 1550 // Do we have a frame pointer for this function? 1551 bool HasFP = hasFP(MF) && FrameSize; 1552 1553 int LROffset = PPCFrameInfo::getReturnSaveOffset(isPPC64, isDarwinABI); 1554 1555 int FPOffset = 0; 1556 if (HasFP) { 1557 if (Subtarget.isSVR4ABI()) { 1558 MachineFrameInfo *FFI = MF.getFrameInfo(); 1559 int FPIndex = FI->getFramePointerSaveIndex(); 1560 assert(FPIndex && "No Frame Pointer Save Slot!"); 1561 FPOffset = FFI->getObjectOffset(FPIndex); 1562 } else { 1563 FPOffset = PPCFrameInfo::getFramePointerSaveOffset(isPPC64, isDarwinABI); 1564 } 1565 } 1566 1567 bool UsesTCRet = RetOpcode == PPC::TCRETURNri || 1568 RetOpcode == PPC::TCRETURNdi || 1569 RetOpcode == PPC::TCRETURNai || 1570 RetOpcode == PPC::TCRETURNri8 || 1571 RetOpcode == PPC::TCRETURNdi8 || 1572 RetOpcode == PPC::TCRETURNai8; 1573 1574 if (UsesTCRet) { 1575 int MaxTCRetDelta = FI->getTailCallSPDelta(); 1576 MachineOperand &StackAdjust = MBBI->getOperand(1); 1577 assert(StackAdjust.isImm() && "Expecting immediate value."); 1578 // Adjust stack pointer. 1579 int StackAdj = StackAdjust.getImm(); 1580 int Delta = StackAdj - MaxTCRetDelta; 1581 assert((Delta >= 0) && "Delta must be positive"); 1582 if (MaxTCRetDelta>0) 1583 FrameSize += (StackAdj +Delta); 1584 else 1585 FrameSize += StackAdj; 1586 } 1587 1588 if (FrameSize) { 1589 // The loaded (or persistent) stack pointer value is offset by the 'stwu' 1590 // on entry to the function. Add this offset back now. 1591 if (!isPPC64) { 1592 // If this function contained a fastcc call and GuaranteedTailCallOpt is 1593 // enabled (=> hasFastCall()==true) the fastcc call might contain a tail 1594 // call which invalidates the stack pointer value in SP(0). So we use the 1595 // value of R31 in this case. 1596 if (FI->hasFastCall() && isInt<16>(FrameSize)) { 1597 assert(hasFP(MF) && "Expecting a valid the frame pointer."); 1598 BuildMI(MBB, MBBI, dl, TII.get(PPC::ADDI), PPC::R1) 1599 .addReg(PPC::R31).addImm(FrameSize); 1600 } else if(FI->hasFastCall()) { 1601 BuildMI(MBB, MBBI, dl, TII.get(PPC::LIS), PPC::R0) 1602 .addImm(FrameSize >> 16); 1603 BuildMI(MBB, MBBI, dl, TII.get(PPC::ORI), PPC::R0) 1604 .addReg(PPC::R0, RegState::Kill) 1605 .addImm(FrameSize & 0xFFFF); 1606 BuildMI(MBB, MBBI, dl, TII.get(PPC::ADD4)) 1607 .addReg(PPC::R1) 1608 .addReg(PPC::R31) 1609 .addReg(PPC::R0); 1610 } else if (isInt<16>(FrameSize) && 1611 (!ALIGN_STACK || TargetAlign >= MaxAlign) && 1612 !MFI->hasVarSizedObjects()) { 1613 BuildMI(MBB, MBBI, dl, TII.get(PPC::ADDI), PPC::R1) 1614 .addReg(PPC::R1).addImm(FrameSize); 1615 } else { 1616 BuildMI(MBB, MBBI, dl, TII.get(PPC::LWZ),PPC::R1) 1617 .addImm(0).addReg(PPC::R1); 1618 } 1619 } else { 1620 if (FI->hasFastCall() && isInt<16>(FrameSize)) { 1621 assert(hasFP(MF) && "Expecting a valid the frame pointer."); 1622 BuildMI(MBB, MBBI, dl, TII.get(PPC::ADDI8), PPC::X1) 1623 .addReg(PPC::X31).addImm(FrameSize); 1624 } else if(FI->hasFastCall()) { 1625 BuildMI(MBB, MBBI, dl, TII.get(PPC::LIS8), PPC::X0) 1626 .addImm(FrameSize >> 16); 1627 BuildMI(MBB, MBBI, dl, TII.get(PPC::ORI8), PPC::X0) 1628 .addReg(PPC::X0, RegState::Kill) 1629 .addImm(FrameSize & 0xFFFF); 1630 BuildMI(MBB, MBBI, dl, TII.get(PPC::ADD8)) 1631 .addReg(PPC::X1) 1632 .addReg(PPC::X31) 1633 .addReg(PPC::X0); 1634 } else if (isInt<16>(FrameSize) && TargetAlign >= MaxAlign && 1635 !MFI->hasVarSizedObjects()) { 1636 BuildMI(MBB, MBBI, dl, TII.get(PPC::ADDI8), PPC::X1) 1637 .addReg(PPC::X1).addImm(FrameSize); 1638 } else { 1639 BuildMI(MBB, MBBI, dl, TII.get(PPC::LD), PPC::X1) 1640 .addImm(0).addReg(PPC::X1); 1641 } 1642 } 1643 } 1644 1645 if (isPPC64) { 1646 if (MustSaveLR) 1647 BuildMI(MBB, MBBI, dl, TII.get(PPC::LD), PPC::X0) 1648 .addImm(LROffset/4).addReg(PPC::X1); 1649 1650 if (HasFP) 1651 BuildMI(MBB, MBBI, dl, TII.get(PPC::LD), PPC::X31) 1652 .addImm(FPOffset/4).addReg(PPC::X1); 1653 1654 if (MustSaveLR) 1655 BuildMI(MBB, MBBI, dl, TII.get(PPC::MTLR8)).addReg(PPC::X0); 1656 } else { 1657 if (MustSaveLR) 1658 BuildMI(MBB, MBBI, dl, TII.get(PPC::LWZ), PPC::R0) 1659 .addImm(LROffset).addReg(PPC::R1); 1660 1661 if (HasFP) 1662 BuildMI(MBB, MBBI, dl, TII.get(PPC::LWZ), PPC::R31) 1663 .addImm(FPOffset).addReg(PPC::R1); 1664 1665 if (MustSaveLR) 1666 BuildMI(MBB, MBBI, dl, TII.get(PPC::MTLR)).addReg(PPC::R0); 1667 } 1668 1669 // Callee pop calling convention. Pop parameter/linkage area. Used for tail 1670 // call optimization 1671 if (GuaranteedTailCallOpt && RetOpcode == PPC::BLR && 1672 MF.getFunction()->getCallingConv() == CallingConv::Fast) { 1673 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>(); 1674 unsigned CallerAllocatedAmt = FI->getMinReservedArea(); 1675 unsigned StackReg = isPPC64 ? PPC::X1 : PPC::R1; 1676 unsigned FPReg = isPPC64 ? PPC::X31 : PPC::R31; 1677 unsigned TmpReg = isPPC64 ? PPC::X0 : PPC::R0; 1678 unsigned ADDIInstr = isPPC64 ? PPC::ADDI8 : PPC::ADDI; 1679 unsigned ADDInstr = isPPC64 ? PPC::ADD8 : PPC::ADD4; 1680 unsigned LISInstr = isPPC64 ? PPC::LIS8 : PPC::LIS; 1681 unsigned ORIInstr = isPPC64 ? PPC::ORI8 : PPC::ORI; 1682 1683 if (CallerAllocatedAmt && isInt<16>(CallerAllocatedAmt)) { 1684 BuildMI(MBB, MBBI, dl, TII.get(ADDIInstr), StackReg) 1685 .addReg(StackReg).addImm(CallerAllocatedAmt); 1686 } else { 1687 BuildMI(MBB, MBBI, dl, TII.get(LISInstr), TmpReg) 1688 .addImm(CallerAllocatedAmt >> 16); 1689 BuildMI(MBB, MBBI, dl, TII.get(ORIInstr), TmpReg) 1690 .addReg(TmpReg, RegState::Kill) 1691 .addImm(CallerAllocatedAmt & 0xFFFF); 1692 BuildMI(MBB, MBBI, dl, TII.get(ADDInstr)) 1693 .addReg(StackReg) 1694 .addReg(FPReg) 1695 .addReg(TmpReg); 1696 } 1697 } else if (RetOpcode == PPC::TCRETURNdi) { 1698 MBBI = prior(MBB.end()); 1699 MachineOperand &JumpTarget = MBBI->getOperand(0); 1700 BuildMI(MBB, MBBI, dl, TII.get(PPC::TAILB)). 1701 addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset()); 1702 } else if (RetOpcode == PPC::TCRETURNri) { 1703 MBBI = prior(MBB.end()); 1704 assert(MBBI->getOperand(0).isReg() && "Expecting register operand."); 1705 BuildMI(MBB, MBBI, dl, TII.get(PPC::TAILBCTR)); 1706 } else if (RetOpcode == PPC::TCRETURNai) { 1707 MBBI = prior(MBB.end()); 1708 MachineOperand &JumpTarget = MBBI->getOperand(0); 1709 BuildMI(MBB, MBBI, dl, TII.get(PPC::TAILBA)).addImm(JumpTarget.getImm()); 1710 } else if (RetOpcode == PPC::TCRETURNdi8) { 1711 MBBI = prior(MBB.end()); 1712 MachineOperand &JumpTarget = MBBI->getOperand(0); 1713 BuildMI(MBB, MBBI, dl, TII.get(PPC::TAILB8)). 1714 addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset()); 1715 } else if (RetOpcode == PPC::TCRETURNri8) { 1716 MBBI = prior(MBB.end()); 1717 assert(MBBI->getOperand(0).isReg() && "Expecting register operand."); 1718 BuildMI(MBB, MBBI, dl, TII.get(PPC::TAILBCTR8)); 1719 } else if (RetOpcode == PPC::TCRETURNai8) { 1720 MBBI = prior(MBB.end()); 1721 MachineOperand &JumpTarget = MBBI->getOperand(0); 1722 BuildMI(MBB, MBBI, dl, TII.get(PPC::TAILBA8)).addImm(JumpTarget.getImm()); 1723 } 1724 } 1725 1726 unsigned PPCRegisterInfo::getRARegister() const { 1727 return !Subtarget.isPPC64() ? PPC::LR : PPC::LR8; 1728 } 1729 1730 unsigned PPCRegisterInfo::getFrameRegister(const MachineFunction &MF) const { 1731 if (!Subtarget.isPPC64()) 1732 return hasFP(MF) ? PPC::R31 : PPC::R1; 1733 else 1734 return hasFP(MF) ? PPC::X31 : PPC::X1; 1735 } 1736 1737 void PPCRegisterInfo::getInitialFrameState(std::vector<MachineMove> &Moves) 1738 const { 1739 // Initial state of the frame pointer is R1. 1740 MachineLocation Dst(MachineLocation::VirtualFP); 1741 MachineLocation Src(PPC::R1, 0); 1742 Moves.push_back(MachineMove(0, Dst, Src)); 1743 } 1744 1745 unsigned PPCRegisterInfo::getEHExceptionRegister() const { 1746 return !Subtarget.isPPC64() ? PPC::R3 : PPC::X3; 1747 } 1748 1749 unsigned PPCRegisterInfo::getEHHandlerRegister() const { 1750 return !Subtarget.isPPC64() ? PPC::R4 : PPC::X4; 1751 } 1752 1753 int PPCRegisterInfo::getDwarfRegNum(unsigned RegNum, bool isEH) const { 1754 // FIXME: Most probably dwarf numbers differs for Linux and Darwin 1755 return PPCGenRegisterInfo::getDwarfRegNumFull(RegNum, 0); 1756 } 1757 1758 #include "PPCGenRegisterInfo.inc" 1759 1760