Chris Lattner | 0808414 | 2003-01-13 00:26:36 +0000 | [diff] [blame] | 1 | //===-- TargetInstrInfo.cpp - Target Instruction Information --------------===// |
Misha Brukman | f976c85 | 2005-04-21 22:55:34 +0000 | [diff] [blame] | 2 | // |
John Criswell | b576c94 | 2003-10-20 19:43:21 +0000 | [diff] [blame] | 3 | // The LLVM Compiler Infrastructure |
| 4 | // |
Chris Lattner | 4ee451d | 2007-12-29 20:36:04 +0000 | [diff] [blame] | 5 | // This file is distributed under the University of Illinois Open Source |
| 6 | // License. See LICENSE.TXT for details. |
Misha Brukman | f976c85 | 2005-04-21 22:55:34 +0000 | [diff] [blame] | 7 | // |
John Criswell | b576c94 | 2003-10-20 19:43:21 +0000 | [diff] [blame] | 8 | //===----------------------------------------------------------------------===// |
Chris Lattner | 93fa705 | 2002-10-28 23:55:33 +0000 | [diff] [blame] | 9 | // |
Chris Lattner | 167b10c | 2005-01-19 06:53:34 +0000 | [diff] [blame] | 10 | // This file implements the TargetInstrInfo class. |
Chris Lattner | 93fa705 | 2002-10-28 23:55:33 +0000 | [diff] [blame] | 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
Chris Lattner | 3501fea | 2003-01-14 22:00:31 +0000 | [diff] [blame] | 14 | #include "llvm/Target/TargetInstrInfo.h" |
Jakob Stoklund Olesen | fa2d986 | 2012-11-28 02:35:13 +0000 | [diff] [blame] | 15 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 16 | #include "llvm/CodeGen/MachineMemOperand.h" |
| 17 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 18 | #include "llvm/CodeGen/PseudoSourceValue.h" |
| 19 | #include "llvm/CodeGen/ScoreboardHazardRecognizer.h" |
Andrew Trick | bb756ca | 2013-11-17 01:36:23 +0000 | [diff] [blame^] | 20 | #include "llvm/IR/DataLayout.h" |
Evan Cheng | a0792de | 2010-10-06 06:27:31 +0000 | [diff] [blame] | 21 | #include "llvm/MC/MCAsmInfo.h" |
Evan Cheng | ab8be96 | 2011-06-29 01:14:12 +0000 | [diff] [blame] | 22 | #include "llvm/MC/MCInstrItineraries.h" |
Jakob Stoklund Olesen | fa2d986 | 2012-11-28 02:35:13 +0000 | [diff] [blame] | 23 | #include "llvm/Support/CommandLine.h" |
Chris Lattner | b6bbfebd | 2009-08-02 04:58:19 +0000 | [diff] [blame] | 24 | #include "llvm/Support/ErrorHandling.h" |
Jakob Stoklund Olesen | fa2d986 | 2012-11-28 02:35:13 +0000 | [diff] [blame] | 25 | #include "llvm/Support/raw_ostream.h" |
Jakob Stoklund Olesen | fa2d986 | 2012-11-28 02:35:13 +0000 | [diff] [blame] | 26 | #include "llvm/Target/TargetLowering.h" |
| 27 | #include "llvm/Target/TargetMachine.h" |
Chandler Carruth | d04a8d4 | 2012-12-03 16:50:05 +0000 | [diff] [blame] | 28 | #include "llvm/Target/TargetRegisterInfo.h" |
Nick Lewycky | 476b242 | 2010-12-19 20:43:38 +0000 | [diff] [blame] | 29 | #include <cctype> |
Chris Lattner | 167b10c | 2005-01-19 06:53:34 +0000 | [diff] [blame] | 30 | using namespace llvm; |
Chris Lattner | 93fa705 | 2002-10-28 23:55:33 +0000 | [diff] [blame] | 31 | |
Jakob Stoklund Olesen | fa2d986 | 2012-11-28 02:35:13 +0000 | [diff] [blame] | 32 | static cl::opt<bool> DisableHazardRecognizer( |
| 33 | "disable-sched-hazard", cl::Hidden, cl::init(false), |
| 34 | cl::desc("Disable hazard detection during preRA scheduling")); |
Chris Lattner | d90183d | 2009-08-02 05:20:37 +0000 | [diff] [blame] | 35 | |
Chris Lattner | 0808414 | 2003-01-13 00:26:36 +0000 | [diff] [blame] | 36 | TargetInstrInfo::~TargetInstrInfo() { |
Chris Lattner | 93fa705 | 2002-10-28 23:55:33 +0000 | [diff] [blame] | 37 | } |
| 38 | |
Evan Cheng | 15993f8 | 2011-06-27 21:26:13 +0000 | [diff] [blame] | 39 | const TargetRegisterClass* |
Evan Cheng | e837dea | 2011-06-28 19:10:37 +0000 | [diff] [blame] | 40 | TargetInstrInfo::getRegClass(const MCInstrDesc &MCID, unsigned OpNum, |
Jakob Stoklund Olesen | 397fc48 | 2012-05-07 22:10:26 +0000 | [diff] [blame] | 41 | const TargetRegisterInfo *TRI, |
| 42 | const MachineFunction &MF) const { |
Evan Cheng | e837dea | 2011-06-28 19:10:37 +0000 | [diff] [blame] | 43 | if (OpNum >= MCID.getNumOperands()) |
Evan Cheng | 15993f8 | 2011-06-27 21:26:13 +0000 | [diff] [blame] | 44 | return 0; |
| 45 | |
Evan Cheng | e837dea | 2011-06-28 19:10:37 +0000 | [diff] [blame] | 46 | short RegClass = MCID.OpInfo[OpNum].RegClass; |
| 47 | if (MCID.OpInfo[OpNum].isLookupPtrRegClass()) |
Jakob Stoklund Olesen | 397fc48 | 2012-05-07 22:10:26 +0000 | [diff] [blame] | 48 | return TRI->getPointerRegClass(MF, RegClass); |
Evan Cheng | 15993f8 | 2011-06-27 21:26:13 +0000 | [diff] [blame] | 49 | |
| 50 | // Instructions like INSERT_SUBREG do not have fixed register classes. |
| 51 | if (RegClass < 0) |
| 52 | return 0; |
| 53 | |
| 54 | // Otherwise just look it up normally. |
| 55 | return TRI->getRegClass(RegClass); |
| 56 | } |
| 57 | |
Chris Lattner | b6bbfebd | 2009-08-02 04:58:19 +0000 | [diff] [blame] | 58 | /// insertNoop - Insert a noop into the instruction stream at the specified |
| 59 | /// point. |
Andrew Trick | 6e8f4c4 | 2010-12-24 04:28:06 +0000 | [diff] [blame] | 60 | void TargetInstrInfo::insertNoop(MachineBasicBlock &MBB, |
Chris Lattner | b6bbfebd | 2009-08-02 04:58:19 +0000 | [diff] [blame] | 61 | MachineBasicBlock::iterator MI) const { |
| 62 | llvm_unreachable("Target didn't implement insertNoop!"); |
| 63 | } |
| 64 | |
Chris Lattner | d90183d | 2009-08-02 05:20:37 +0000 | [diff] [blame] | 65 | /// Measure the specified inline asm to determine an approximation of its |
| 66 | /// length. |
Jim Grosbach | d31d304 | 2011-03-24 18:46:34 +0000 | [diff] [blame] | 67 | /// Comments (which run till the next SeparatorString or newline) do not |
Chris Lattner | d90183d | 2009-08-02 05:20:37 +0000 | [diff] [blame] | 68 | /// count as an instruction. |
| 69 | /// Any other non-whitespace text is considered an instruction, with |
Jim Grosbach | d31d304 | 2011-03-24 18:46:34 +0000 | [diff] [blame] | 70 | /// multiple instructions separated by SeparatorString or newlines. |
Chris Lattner | d90183d | 2009-08-02 05:20:37 +0000 | [diff] [blame] | 71 | /// Variable-length instructions are not handled here; this function |
| 72 | /// may be overloaded in the target code to do that. |
| 73 | unsigned TargetInstrInfo::getInlineAsmLength(const char *Str, |
Chris Lattner | 33adcfb | 2009-08-22 21:43:10 +0000 | [diff] [blame] | 74 | const MCAsmInfo &MAI) const { |
Andrew Trick | 6e8f4c4 | 2010-12-24 04:28:06 +0000 | [diff] [blame] | 75 | |
| 76 | |
Chris Lattner | d90183d | 2009-08-02 05:20:37 +0000 | [diff] [blame] | 77 | // Count the number of instructions in the asm. |
| 78 | bool atInsnStart = true; |
| 79 | unsigned Length = 0; |
| 80 | for (; *Str; ++Str) { |
Jim Grosbach | d31d304 | 2011-03-24 18:46:34 +0000 | [diff] [blame] | 81 | if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(), |
| 82 | strlen(MAI.getSeparatorString())) == 0) |
Chris Lattner | d90183d | 2009-08-02 05:20:37 +0000 | [diff] [blame] | 83 | atInsnStart = true; |
Guy Benyei | 87d0b9e | 2013-02-12 21:21:59 +0000 | [diff] [blame] | 84 | if (atInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) { |
Chris Lattner | 33adcfb | 2009-08-22 21:43:10 +0000 | [diff] [blame] | 85 | Length += MAI.getMaxInstLength(); |
Chris Lattner | d90183d | 2009-08-02 05:20:37 +0000 | [diff] [blame] | 86 | atInsnStart = false; |
| 87 | } |
Chris Lattner | 33adcfb | 2009-08-22 21:43:10 +0000 | [diff] [blame] | 88 | if (atInsnStart && strncmp(Str, MAI.getCommentString(), |
| 89 | strlen(MAI.getCommentString())) == 0) |
Chris Lattner | d90183d | 2009-08-02 05:20:37 +0000 | [diff] [blame] | 90 | atInsnStart = false; |
| 91 | } |
Andrew Trick | 6e8f4c4 | 2010-12-24 04:28:06 +0000 | [diff] [blame] | 92 | |
Chris Lattner | d90183d | 2009-08-02 05:20:37 +0000 | [diff] [blame] | 93 | return Length; |
| 94 | } |
Jakob Stoklund Olesen | fa2d986 | 2012-11-28 02:35:13 +0000 | [diff] [blame] | 95 | |
| 96 | /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything |
| 97 | /// after it, replacing it with an unconditional branch to NewDest. |
| 98 | void |
| 99 | TargetInstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail, |
| 100 | MachineBasicBlock *NewDest) const { |
| 101 | MachineBasicBlock *MBB = Tail->getParent(); |
| 102 | |
| 103 | // Remove all the old successors of MBB from the CFG. |
| 104 | while (!MBB->succ_empty()) |
| 105 | MBB->removeSuccessor(MBB->succ_begin()); |
| 106 | |
| 107 | // Remove all the dead instructions from the end of MBB. |
| 108 | MBB->erase(Tail, MBB->end()); |
| 109 | |
| 110 | // If MBB isn't immediately before MBB, insert a branch to it. |
| 111 | if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest)) |
| 112 | InsertBranch(*MBB, NewDest, 0, SmallVector<MachineOperand, 0>(), |
| 113 | Tail->getDebugLoc()); |
| 114 | MBB->addSuccessor(NewDest); |
| 115 | } |
| 116 | |
| 117 | // commuteInstruction - The default implementation of this method just exchanges |
| 118 | // the two operands returned by findCommutedOpIndices. |
| 119 | MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr *MI, |
| 120 | bool NewMI) const { |
| 121 | const MCInstrDesc &MCID = MI->getDesc(); |
| 122 | bool HasDef = MCID.getNumDefs(); |
| 123 | if (HasDef && !MI->getOperand(0).isReg()) |
| 124 | // No idea how to commute this instruction. Target should implement its own. |
| 125 | return 0; |
| 126 | unsigned Idx1, Idx2; |
| 127 | if (!findCommutedOpIndices(MI, Idx1, Idx2)) { |
| 128 | std::string msg; |
| 129 | raw_string_ostream Msg(msg); |
| 130 | Msg << "Don't know how to commute: " << *MI; |
| 131 | report_fatal_error(Msg.str()); |
| 132 | } |
| 133 | |
| 134 | assert(MI->getOperand(Idx1).isReg() && MI->getOperand(Idx2).isReg() && |
| 135 | "This only knows how to commute register operands so far"); |
| 136 | unsigned Reg0 = HasDef ? MI->getOperand(0).getReg() : 0; |
| 137 | unsigned Reg1 = MI->getOperand(Idx1).getReg(); |
| 138 | unsigned Reg2 = MI->getOperand(Idx2).getReg(); |
| 139 | unsigned SubReg0 = HasDef ? MI->getOperand(0).getSubReg() : 0; |
| 140 | unsigned SubReg1 = MI->getOperand(Idx1).getSubReg(); |
| 141 | unsigned SubReg2 = MI->getOperand(Idx2).getSubReg(); |
| 142 | bool Reg1IsKill = MI->getOperand(Idx1).isKill(); |
| 143 | bool Reg2IsKill = MI->getOperand(Idx2).isKill(); |
| 144 | // If destination is tied to either of the commuted source register, then |
| 145 | // it must be updated. |
| 146 | if (HasDef && Reg0 == Reg1 && |
| 147 | MI->getDesc().getOperandConstraint(Idx1, MCOI::TIED_TO) == 0) { |
| 148 | Reg2IsKill = false; |
| 149 | Reg0 = Reg2; |
| 150 | SubReg0 = SubReg2; |
| 151 | } else if (HasDef && Reg0 == Reg2 && |
| 152 | MI->getDesc().getOperandConstraint(Idx2, MCOI::TIED_TO) == 0) { |
| 153 | Reg1IsKill = false; |
| 154 | Reg0 = Reg1; |
| 155 | SubReg0 = SubReg1; |
| 156 | } |
| 157 | |
| 158 | if (NewMI) { |
| 159 | // Create a new instruction. |
| 160 | MachineFunction &MF = *MI->getParent()->getParent(); |
| 161 | MI = MF.CloneMachineInstr(MI); |
| 162 | } |
| 163 | |
| 164 | if (HasDef) { |
| 165 | MI->getOperand(0).setReg(Reg0); |
| 166 | MI->getOperand(0).setSubReg(SubReg0); |
| 167 | } |
| 168 | MI->getOperand(Idx2).setReg(Reg1); |
| 169 | MI->getOperand(Idx1).setReg(Reg2); |
| 170 | MI->getOperand(Idx2).setSubReg(SubReg1); |
| 171 | MI->getOperand(Idx1).setSubReg(SubReg2); |
| 172 | MI->getOperand(Idx2).setIsKill(Reg1IsKill); |
| 173 | MI->getOperand(Idx1).setIsKill(Reg2IsKill); |
| 174 | return MI; |
| 175 | } |
| 176 | |
| 177 | /// findCommutedOpIndices - If specified MI is commutable, return the two |
| 178 | /// operand indices that would swap value. Return true if the instruction |
| 179 | /// is not in a form which this routine understands. |
| 180 | bool TargetInstrInfo::findCommutedOpIndices(MachineInstr *MI, |
| 181 | unsigned &SrcOpIdx1, |
| 182 | unsigned &SrcOpIdx2) const { |
| 183 | assert(!MI->isBundle() && |
| 184 | "TargetInstrInfo::findCommutedOpIndices() can't handle bundles"); |
| 185 | |
| 186 | const MCInstrDesc &MCID = MI->getDesc(); |
| 187 | if (!MCID.isCommutable()) |
| 188 | return false; |
| 189 | // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this |
| 190 | // is not true, then the target must implement this. |
| 191 | SrcOpIdx1 = MCID.getNumDefs(); |
| 192 | SrcOpIdx2 = SrcOpIdx1 + 1; |
| 193 | if (!MI->getOperand(SrcOpIdx1).isReg() || |
| 194 | !MI->getOperand(SrcOpIdx2).isReg()) |
| 195 | // No idea. |
| 196 | return false; |
| 197 | return true; |
| 198 | } |
| 199 | |
| 200 | |
| 201 | bool |
| 202 | TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const { |
| 203 | if (!MI->isTerminator()) return false; |
| 204 | |
| 205 | // Conditional branch is a special case. |
| 206 | if (MI->isBranch() && !MI->isBarrier()) |
| 207 | return true; |
| 208 | if (!MI->isPredicable()) |
| 209 | return true; |
| 210 | return !isPredicated(MI); |
| 211 | } |
| 212 | |
| 213 | |
| 214 | bool TargetInstrInfo::PredicateInstruction(MachineInstr *MI, |
| 215 | const SmallVectorImpl<MachineOperand> &Pred) const { |
| 216 | bool MadeChange = false; |
| 217 | |
| 218 | assert(!MI->isBundle() && |
| 219 | "TargetInstrInfo::PredicateInstruction() can't handle bundles"); |
| 220 | |
| 221 | const MCInstrDesc &MCID = MI->getDesc(); |
| 222 | if (!MI->isPredicable()) |
| 223 | return false; |
| 224 | |
| 225 | for (unsigned j = 0, i = 0, e = MI->getNumOperands(); i != e; ++i) { |
| 226 | if (MCID.OpInfo[i].isPredicate()) { |
| 227 | MachineOperand &MO = MI->getOperand(i); |
| 228 | if (MO.isReg()) { |
| 229 | MO.setReg(Pred[j].getReg()); |
| 230 | MadeChange = true; |
| 231 | } else if (MO.isImm()) { |
| 232 | MO.setImm(Pred[j].getImm()); |
| 233 | MadeChange = true; |
| 234 | } else if (MO.isMBB()) { |
| 235 | MO.setMBB(Pred[j].getMBB()); |
| 236 | MadeChange = true; |
| 237 | } |
| 238 | ++j; |
| 239 | } |
| 240 | } |
| 241 | return MadeChange; |
| 242 | } |
| 243 | |
| 244 | bool TargetInstrInfo::hasLoadFromStackSlot(const MachineInstr *MI, |
| 245 | const MachineMemOperand *&MMO, |
| 246 | int &FrameIndex) const { |
| 247 | for (MachineInstr::mmo_iterator o = MI->memoperands_begin(), |
| 248 | oe = MI->memoperands_end(); |
| 249 | o != oe; |
| 250 | ++o) { |
| 251 | if ((*o)->isLoad() && (*o)->getValue()) |
| 252 | if (const FixedStackPseudoSourceValue *Value = |
| 253 | dyn_cast<const FixedStackPseudoSourceValue>((*o)->getValue())) { |
| 254 | FrameIndex = Value->getFrameIndex(); |
| 255 | MMO = *o; |
| 256 | return true; |
| 257 | } |
| 258 | } |
| 259 | return false; |
| 260 | } |
| 261 | |
| 262 | bool TargetInstrInfo::hasStoreToStackSlot(const MachineInstr *MI, |
| 263 | const MachineMemOperand *&MMO, |
| 264 | int &FrameIndex) const { |
| 265 | for (MachineInstr::mmo_iterator o = MI->memoperands_begin(), |
| 266 | oe = MI->memoperands_end(); |
| 267 | o != oe; |
| 268 | ++o) { |
| 269 | if ((*o)->isStore() && (*o)->getValue()) |
| 270 | if (const FixedStackPseudoSourceValue *Value = |
| 271 | dyn_cast<const FixedStackPseudoSourceValue>((*o)->getValue())) { |
| 272 | FrameIndex = Value->getFrameIndex(); |
| 273 | MMO = *o; |
| 274 | return true; |
| 275 | } |
| 276 | } |
| 277 | return false; |
| 278 | } |
| 279 | |
Andrew Trick | bb756ca | 2013-11-17 01:36:23 +0000 | [diff] [blame^] | 280 | bool TargetInstrInfo::getStackSlotRange(const TargetRegisterClass *RC, |
| 281 | unsigned SubIdx, unsigned &Size, |
| 282 | unsigned &Offset, |
| 283 | const TargetMachine *TM) const { |
| 284 | if (!SubIdx) { |
| 285 | Size = RC->getSize(); |
| 286 | Offset = 0; |
| 287 | return true; |
| 288 | } |
| 289 | unsigned BitSize = TM->getRegisterInfo()->getSubRegIdxSize(SubIdx); |
| 290 | // Convert bit size to byte size to be consistent with |
| 291 | // MCRegisterClass::getSize(). |
| 292 | if (BitSize % 8) |
| 293 | return false; |
| 294 | |
| 295 | int BitOffset = TM->getRegisterInfo()->getSubRegIdxOffset(SubIdx); |
| 296 | if (BitOffset < 0 || BitOffset % 8) |
| 297 | return false; |
| 298 | |
| 299 | Size = BitSize /= 8; |
| 300 | Offset = (unsigned)BitOffset / 8; |
| 301 | |
| 302 | assert(RC->getSize() >= (Offset + Size) && "bad subregister range"); |
| 303 | |
| 304 | if (!TM->getDataLayout()->isLittleEndian()) { |
| 305 | Offset = RC->getSize() - (Offset + Size); |
| 306 | } |
| 307 | return true; |
| 308 | } |
| 309 | |
Jakob Stoklund Olesen | fa2d986 | 2012-11-28 02:35:13 +0000 | [diff] [blame] | 310 | void TargetInstrInfo::reMaterialize(MachineBasicBlock &MBB, |
| 311 | MachineBasicBlock::iterator I, |
| 312 | unsigned DestReg, |
| 313 | unsigned SubIdx, |
| 314 | const MachineInstr *Orig, |
| 315 | const TargetRegisterInfo &TRI) const { |
| 316 | MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig); |
| 317 | MI->substituteRegister(MI->getOperand(0).getReg(), DestReg, SubIdx, TRI); |
| 318 | MBB.insert(I, MI); |
| 319 | } |
| 320 | |
| 321 | bool |
| 322 | TargetInstrInfo::produceSameValue(const MachineInstr *MI0, |
| 323 | const MachineInstr *MI1, |
| 324 | const MachineRegisterInfo *MRI) const { |
| 325 | return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs); |
| 326 | } |
| 327 | |
| 328 | MachineInstr *TargetInstrInfo::duplicate(MachineInstr *Orig, |
| 329 | MachineFunction &MF) const { |
| 330 | assert(!Orig->isNotDuplicable() && |
| 331 | "Instruction cannot be duplicated"); |
| 332 | return MF.CloneMachineInstr(Orig); |
| 333 | } |
| 334 | |
| 335 | // If the COPY instruction in MI can be folded to a stack operation, return |
| 336 | // the register class to use. |
| 337 | static const TargetRegisterClass *canFoldCopy(const MachineInstr *MI, |
| 338 | unsigned FoldIdx) { |
| 339 | assert(MI->isCopy() && "MI must be a COPY instruction"); |
| 340 | if (MI->getNumOperands() != 2) |
| 341 | return 0; |
| 342 | assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand"); |
| 343 | |
| 344 | const MachineOperand &FoldOp = MI->getOperand(FoldIdx); |
| 345 | const MachineOperand &LiveOp = MI->getOperand(1-FoldIdx); |
| 346 | |
| 347 | if (FoldOp.getSubReg() || LiveOp.getSubReg()) |
| 348 | return 0; |
| 349 | |
| 350 | unsigned FoldReg = FoldOp.getReg(); |
| 351 | unsigned LiveReg = LiveOp.getReg(); |
| 352 | |
| 353 | assert(TargetRegisterInfo::isVirtualRegister(FoldReg) && |
| 354 | "Cannot fold physregs"); |
| 355 | |
| 356 | const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo(); |
| 357 | const TargetRegisterClass *RC = MRI.getRegClass(FoldReg); |
| 358 | |
| 359 | if (TargetRegisterInfo::isPhysicalRegister(LiveOp.getReg())) |
| 360 | return RC->contains(LiveOp.getReg()) ? RC : 0; |
| 361 | |
| 362 | if (RC->hasSubClassEq(MRI.getRegClass(LiveReg))) |
| 363 | return RC; |
| 364 | |
| 365 | // FIXME: Allow folding when register classes are memory compatible. |
| 366 | return 0; |
| 367 | } |
| 368 | |
| 369 | bool TargetInstrInfo:: |
| 370 | canFoldMemoryOperand(const MachineInstr *MI, |
| 371 | const SmallVectorImpl<unsigned> &Ops) const { |
| 372 | return MI->isCopy() && Ops.size() == 1 && canFoldCopy(MI, Ops[0]); |
| 373 | } |
| 374 | |
| 375 | /// foldMemoryOperand - Attempt to fold a load or store of the specified stack |
| 376 | /// slot into the specified machine instruction for the specified operand(s). |
| 377 | /// If this is possible, a new instruction is returned with the specified |
| 378 | /// operand folded, otherwise NULL is returned. The client is responsible for |
| 379 | /// removing the old instruction and adding the new one in the instruction |
| 380 | /// stream. |
| 381 | MachineInstr* |
| 382 | TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI, |
| 383 | const SmallVectorImpl<unsigned> &Ops, |
| 384 | int FI) const { |
| 385 | unsigned Flags = 0; |
| 386 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) |
| 387 | if (MI->getOperand(Ops[i]).isDef()) |
| 388 | Flags |= MachineMemOperand::MOStore; |
| 389 | else |
| 390 | Flags |= MachineMemOperand::MOLoad; |
| 391 | |
| 392 | MachineBasicBlock *MBB = MI->getParent(); |
| 393 | assert(MBB && "foldMemoryOperand needs an inserted instruction"); |
| 394 | MachineFunction &MF = *MBB->getParent(); |
| 395 | |
| 396 | // Ask the target to do the actual folding. |
| 397 | if (MachineInstr *NewMI = foldMemoryOperandImpl(MF, MI, Ops, FI)) { |
Andrew Trick | 849596c | 2013-11-14 23:45:04 +0000 | [diff] [blame] | 398 | NewMI->setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); |
Jakob Stoklund Olesen | fa2d986 | 2012-11-28 02:35:13 +0000 | [diff] [blame] | 399 | // Add a memory operand, foldMemoryOperandImpl doesn't do that. |
| 400 | assert((!(Flags & MachineMemOperand::MOStore) || |
| 401 | NewMI->mayStore()) && |
| 402 | "Folded a def to a non-store!"); |
| 403 | assert((!(Flags & MachineMemOperand::MOLoad) || |
| 404 | NewMI->mayLoad()) && |
| 405 | "Folded a use to a non-load!"); |
| 406 | const MachineFrameInfo &MFI = *MF.getFrameInfo(); |
| 407 | assert(MFI.getObjectOffset(FI) != -1); |
| 408 | MachineMemOperand *MMO = |
| 409 | MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FI), |
| 410 | Flags, MFI.getObjectSize(FI), |
| 411 | MFI.getObjectAlignment(FI)); |
| 412 | NewMI->addMemOperand(MF, MMO); |
| 413 | |
| 414 | // FIXME: change foldMemoryOperandImpl semantics to also insert NewMI. |
| 415 | return MBB->insert(MI, NewMI); |
| 416 | } |
| 417 | |
| 418 | // Straight COPY may fold as load/store. |
| 419 | if (!MI->isCopy() || Ops.size() != 1) |
| 420 | return 0; |
| 421 | |
| 422 | const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]); |
| 423 | if (!RC) |
| 424 | return 0; |
| 425 | |
| 426 | const MachineOperand &MO = MI->getOperand(1-Ops[0]); |
| 427 | MachineBasicBlock::iterator Pos = MI; |
| 428 | const TargetRegisterInfo *TRI = MF.getTarget().getRegisterInfo(); |
| 429 | |
| 430 | if (Flags == MachineMemOperand::MOStore) |
| 431 | storeRegToStackSlot(*MBB, Pos, MO.getReg(), MO.isKill(), FI, RC, TRI); |
| 432 | else |
| 433 | loadRegFromStackSlot(*MBB, Pos, MO.getReg(), FI, RC, TRI); |
| 434 | return --Pos; |
| 435 | } |
| 436 | |
| 437 | /// foldMemoryOperand - Same as the previous version except it allows folding |
| 438 | /// of any load and store from / to any address, not just from a specific |
| 439 | /// stack slot. |
| 440 | MachineInstr* |
| 441 | TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI, |
| 442 | const SmallVectorImpl<unsigned> &Ops, |
| 443 | MachineInstr* LoadMI) const { |
| 444 | assert(LoadMI->canFoldAsLoad() && "LoadMI isn't foldable!"); |
| 445 | #ifndef NDEBUG |
| 446 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) |
| 447 | assert(MI->getOperand(Ops[i]).isUse() && "Folding load into def!"); |
| 448 | #endif |
| 449 | MachineBasicBlock &MBB = *MI->getParent(); |
| 450 | MachineFunction &MF = *MBB.getParent(); |
| 451 | |
| 452 | // Ask the target to do the actual folding. |
| 453 | MachineInstr *NewMI = foldMemoryOperandImpl(MF, MI, Ops, LoadMI); |
| 454 | if (!NewMI) return 0; |
| 455 | |
| 456 | NewMI = MBB.insert(MI, NewMI); |
| 457 | |
| 458 | // Copy the memoperands from the load to the folded instruction. |
Andrew Trick | 849596c | 2013-11-14 23:45:04 +0000 | [diff] [blame] | 459 | if (MI->memoperands_empty()) { |
| 460 | NewMI->setMemRefs(LoadMI->memoperands_begin(), |
| 461 | LoadMI->memoperands_end()); |
| 462 | } |
| 463 | else { |
| 464 | // Handle the rare case of folding multiple loads. |
| 465 | NewMI->setMemRefs(MI->memoperands_begin(), |
| 466 | MI->memoperands_end()); |
| 467 | for (MachineInstr::mmo_iterator I = LoadMI->memoperands_begin(), |
| 468 | E = LoadMI->memoperands_end(); I != E; ++I) { |
| 469 | NewMI->addMemOperand(MF, *I); |
| 470 | } |
| 471 | } |
Jakob Stoklund Olesen | fa2d986 | 2012-11-28 02:35:13 +0000 | [diff] [blame] | 472 | return NewMI; |
| 473 | } |
| 474 | |
| 475 | bool TargetInstrInfo:: |
| 476 | isReallyTriviallyReMaterializableGeneric(const MachineInstr *MI, |
| 477 | AliasAnalysis *AA) const { |
| 478 | const MachineFunction &MF = *MI->getParent()->getParent(); |
| 479 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 480 | const TargetMachine &TM = MF.getTarget(); |
| 481 | const TargetInstrInfo &TII = *TM.getInstrInfo(); |
| 482 | |
| 483 | // Remat clients assume operand 0 is the defined register. |
| 484 | if (!MI->getNumOperands() || !MI->getOperand(0).isReg()) |
| 485 | return false; |
| 486 | unsigned DefReg = MI->getOperand(0).getReg(); |
| 487 | |
| 488 | // A sub-register definition can only be rematerialized if the instruction |
| 489 | // doesn't read the other parts of the register. Otherwise it is really a |
| 490 | // read-modify-write operation on the full virtual register which cannot be |
| 491 | // moved safely. |
| 492 | if (TargetRegisterInfo::isVirtualRegister(DefReg) && |
| 493 | MI->getOperand(0).getSubReg() && MI->readsVirtualRegister(DefReg)) |
| 494 | return false; |
| 495 | |
| 496 | // A load from a fixed stack slot can be rematerialized. This may be |
| 497 | // redundant with subsequent checks, but it's target-independent, |
| 498 | // simple, and a common case. |
| 499 | int FrameIdx = 0; |
| 500 | if (TII.isLoadFromStackSlot(MI, FrameIdx) && |
| 501 | MF.getFrameInfo()->isImmutableObjectIndex(FrameIdx)) |
| 502 | return true; |
| 503 | |
| 504 | // Avoid instructions obviously unsafe for remat. |
| 505 | if (MI->isNotDuplicable() || MI->mayStore() || |
| 506 | MI->hasUnmodeledSideEffects()) |
| 507 | return false; |
| 508 | |
| 509 | // Don't remat inline asm. We have no idea how expensive it is |
| 510 | // even if it's side effect free. |
| 511 | if (MI->isInlineAsm()) |
| 512 | return false; |
| 513 | |
| 514 | // Avoid instructions which load from potentially varying memory. |
| 515 | if (MI->mayLoad() && !MI->isInvariantLoad(AA)) |
| 516 | return false; |
| 517 | |
| 518 | // If any of the registers accessed are non-constant, conservatively assume |
| 519 | // the instruction is not rematerializable. |
| 520 | for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { |
| 521 | const MachineOperand &MO = MI->getOperand(i); |
| 522 | if (!MO.isReg()) continue; |
| 523 | unsigned Reg = MO.getReg(); |
| 524 | if (Reg == 0) |
| 525 | continue; |
| 526 | |
| 527 | // Check for a well-behaved physical register. |
| 528 | if (TargetRegisterInfo::isPhysicalRegister(Reg)) { |
| 529 | if (MO.isUse()) { |
| 530 | // If the physreg has no defs anywhere, it's just an ambient register |
| 531 | // and we can freely move its uses. Alternatively, if it's allocatable, |
| 532 | // it could get allocated to something with a def during allocation. |
| 533 | if (!MRI.isConstantPhysReg(Reg, MF)) |
| 534 | return false; |
| 535 | } else { |
| 536 | // A physreg def. We can't remat it. |
| 537 | return false; |
| 538 | } |
| 539 | continue; |
| 540 | } |
| 541 | |
| 542 | // Only allow one virtual-register def. There may be multiple defs of the |
| 543 | // same virtual register, though. |
| 544 | if (MO.isDef() && Reg != DefReg) |
| 545 | return false; |
| 546 | |
| 547 | // Don't allow any virtual-register uses. Rematting an instruction with |
| 548 | // virtual register uses would length the live ranges of the uses, which |
| 549 | // is not necessarily a good idea, certainly not "trivial". |
| 550 | if (MO.isUse()) |
| 551 | return false; |
| 552 | } |
| 553 | |
| 554 | // Everything checked out. |
| 555 | return true; |
| 556 | } |
| 557 | |
| 558 | /// isSchedulingBoundary - Test if the given instruction should be |
| 559 | /// considered a scheduling boundary. This primarily includes labels |
| 560 | /// and terminators. |
| 561 | bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr *MI, |
| 562 | const MachineBasicBlock *MBB, |
| 563 | const MachineFunction &MF) const { |
| 564 | // Terminators and labels can't be scheduled around. |
| 565 | if (MI->isTerminator() || MI->isLabel()) |
| 566 | return true; |
| 567 | |
| 568 | // Don't attempt to schedule around any instruction that defines |
| 569 | // a stack-oriented pointer, as it's unlikely to be profitable. This |
| 570 | // saves compile time, because it doesn't require every single |
| 571 | // stack slot reference to depend on the instruction that does the |
| 572 | // modification. |
| 573 | const TargetLowering &TLI = *MF.getTarget().getTargetLowering(); |
| 574 | const TargetRegisterInfo *TRI = MF.getTarget().getRegisterInfo(); |
| 575 | if (MI->modifiesRegister(TLI.getStackPointerRegisterToSaveRestore(), TRI)) |
| 576 | return true; |
| 577 | |
| 578 | return false; |
| 579 | } |
| 580 | |
| 581 | // Provide a global flag for disabling the PreRA hazard recognizer that targets |
| 582 | // may choose to honor. |
| 583 | bool TargetInstrInfo::usePreRAHazardRecognizer() const { |
| 584 | return !DisableHazardRecognizer; |
| 585 | } |
| 586 | |
| 587 | // Default implementation of CreateTargetRAHazardRecognizer. |
| 588 | ScheduleHazardRecognizer *TargetInstrInfo:: |
| 589 | CreateTargetHazardRecognizer(const TargetMachine *TM, |
| 590 | const ScheduleDAG *DAG) const { |
| 591 | // Dummy hazard recognizer allows all instructions to issue. |
| 592 | return new ScheduleHazardRecognizer(); |
| 593 | } |
| 594 | |
| 595 | // Default implementation of CreateTargetMIHazardRecognizer. |
| 596 | ScheduleHazardRecognizer *TargetInstrInfo:: |
| 597 | CreateTargetMIHazardRecognizer(const InstrItineraryData *II, |
| 598 | const ScheduleDAG *DAG) const { |
| 599 | return (ScheduleHazardRecognizer *) |
| 600 | new ScoreboardHazardRecognizer(II, DAG, "misched"); |
| 601 | } |
| 602 | |
| 603 | // Default implementation of CreateTargetPostRAHazardRecognizer. |
| 604 | ScheduleHazardRecognizer *TargetInstrInfo:: |
| 605 | CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, |
| 606 | const ScheduleDAG *DAG) const { |
| 607 | return (ScheduleHazardRecognizer *) |
| 608 | new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched"); |
| 609 | } |
| 610 | |
| 611 | //===----------------------------------------------------------------------===// |
| 612 | // SelectionDAG latency interface. |
| 613 | //===----------------------------------------------------------------------===// |
| 614 | |
| 615 | int |
| 616 | TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, |
| 617 | SDNode *DefNode, unsigned DefIdx, |
| 618 | SDNode *UseNode, unsigned UseIdx) const { |
| 619 | if (!ItinData || ItinData->isEmpty()) |
| 620 | return -1; |
| 621 | |
| 622 | if (!DefNode->isMachineOpcode()) |
| 623 | return -1; |
| 624 | |
| 625 | unsigned DefClass = get(DefNode->getMachineOpcode()).getSchedClass(); |
| 626 | if (!UseNode->isMachineOpcode()) |
| 627 | return ItinData->getOperandCycle(DefClass, DefIdx); |
| 628 | unsigned UseClass = get(UseNode->getMachineOpcode()).getSchedClass(); |
| 629 | return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx); |
| 630 | } |
| 631 | |
| 632 | int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, |
| 633 | SDNode *N) const { |
| 634 | if (!ItinData || ItinData->isEmpty()) |
| 635 | return 1; |
| 636 | |
| 637 | if (!N->isMachineOpcode()) |
| 638 | return 1; |
| 639 | |
| 640 | return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass()); |
| 641 | } |
| 642 | |
| 643 | //===----------------------------------------------------------------------===// |
| 644 | // MachineInstr latency interface. |
| 645 | //===----------------------------------------------------------------------===// |
| 646 | |
| 647 | unsigned |
| 648 | TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData, |
| 649 | const MachineInstr *MI) const { |
| 650 | if (!ItinData || ItinData->isEmpty()) |
| 651 | return 1; |
| 652 | |
| 653 | unsigned Class = MI->getDesc().getSchedClass(); |
| 654 | int UOps = ItinData->Itineraries[Class].NumMicroOps; |
| 655 | if (UOps >= 0) |
| 656 | return UOps; |
| 657 | |
| 658 | // The # of u-ops is dynamically determined. The specific target should |
| 659 | // override this function to return the right number. |
| 660 | return 1; |
| 661 | } |
| 662 | |
| 663 | /// Return the default expected latency for a def based on it's opcode. |
| 664 | unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel *SchedModel, |
| 665 | const MachineInstr *DefMI) const { |
| 666 | if (DefMI->isTransient()) |
| 667 | return 0; |
| 668 | if (DefMI->mayLoad()) |
| 669 | return SchedModel->LoadLatency; |
| 670 | if (isHighLatencyDef(DefMI->getOpcode())) |
| 671 | return SchedModel->HighLatency; |
| 672 | return 1; |
| 673 | } |
| 674 | |
Arnold Schwaighofer | d42730d | 2013-09-30 15:28:56 +0000 | [diff] [blame] | 675 | unsigned TargetInstrInfo::getPredicationCost(const MachineInstr *) const { |
| 676 | return 0; |
| 677 | } |
| 678 | |
Jakob Stoklund Olesen | fa2d986 | 2012-11-28 02:35:13 +0000 | [diff] [blame] | 679 | unsigned TargetInstrInfo:: |
| 680 | getInstrLatency(const InstrItineraryData *ItinData, |
| 681 | const MachineInstr *MI, |
| 682 | unsigned *PredCost) const { |
| 683 | // Default to one cycle for no itinerary. However, an "empty" itinerary may |
| 684 | // still have a MinLatency property, which getStageLatency checks. |
| 685 | if (!ItinData) |
| 686 | return MI->mayLoad() ? 2 : 1; |
| 687 | |
| 688 | return ItinData->getStageLatency(MI->getDesc().getSchedClass()); |
| 689 | } |
| 690 | |
| 691 | bool TargetInstrInfo::hasLowDefLatency(const InstrItineraryData *ItinData, |
| 692 | const MachineInstr *DefMI, |
| 693 | unsigned DefIdx) const { |
| 694 | if (!ItinData || ItinData->isEmpty()) |
| 695 | return false; |
| 696 | |
| 697 | unsigned DefClass = DefMI->getDesc().getSchedClass(); |
| 698 | int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx); |
| 699 | return (DefCycle != -1 && DefCycle <= 1); |
| 700 | } |
| 701 | |
| 702 | /// Both DefMI and UseMI must be valid. By default, call directly to the |
| 703 | /// itinerary. This may be overriden by the target. |
| 704 | int TargetInstrInfo:: |
| 705 | getOperandLatency(const InstrItineraryData *ItinData, |
| 706 | const MachineInstr *DefMI, unsigned DefIdx, |
| 707 | const MachineInstr *UseMI, unsigned UseIdx) const { |
| 708 | unsigned DefClass = DefMI->getDesc().getSchedClass(); |
| 709 | unsigned UseClass = UseMI->getDesc().getSchedClass(); |
| 710 | return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx); |
| 711 | } |
| 712 | |
| 713 | /// If we can determine the operand latency from the def only, without itinerary |
| 714 | /// lookup, do so. Otherwise return -1. |
| 715 | int TargetInstrInfo::computeDefOperandLatency( |
| 716 | const InstrItineraryData *ItinData, |
Andrew Trick | b86a0cd | 2013-06-15 04:49:57 +0000 | [diff] [blame] | 717 | const MachineInstr *DefMI) const { |
Jakob Stoklund Olesen | fa2d986 | 2012-11-28 02:35:13 +0000 | [diff] [blame] | 718 | |
| 719 | // Let the target hook getInstrLatency handle missing itineraries. |
| 720 | if (!ItinData) |
| 721 | return getInstrLatency(ItinData, DefMI); |
| 722 | |
Andrew Trick | b86a0cd | 2013-06-15 04:49:57 +0000 | [diff] [blame] | 723 | if(ItinData->isEmpty()) |
Jakob Stoklund Olesen | fa2d986 | 2012-11-28 02:35:13 +0000 | [diff] [blame] | 724 | return defaultDefLatency(ItinData->SchedModel, DefMI); |
| 725 | |
| 726 | // ...operand lookup required |
| 727 | return -1; |
| 728 | } |
| 729 | |
| 730 | /// computeOperandLatency - Compute and return the latency of the given data |
| 731 | /// dependent def and use when the operand indices are already known. UseMI may |
| 732 | /// be NULL for an unknown use. |
| 733 | /// |
| 734 | /// FindMin may be set to get the minimum vs. expected latency. Minimum |
| 735 | /// latency is used for scheduling groups, while expected latency is for |
| 736 | /// instruction cost and critical path. |
| 737 | /// |
| 738 | /// Depending on the subtarget's itinerary properties, this may or may not need |
| 739 | /// to call getOperandLatency(). For most subtargets, we don't need DefIdx or |
| 740 | /// UseIdx to compute min latency. |
| 741 | unsigned TargetInstrInfo:: |
| 742 | computeOperandLatency(const InstrItineraryData *ItinData, |
| 743 | const MachineInstr *DefMI, unsigned DefIdx, |
Andrew Trick | b86a0cd | 2013-06-15 04:49:57 +0000 | [diff] [blame] | 744 | const MachineInstr *UseMI, unsigned UseIdx) const { |
Jakob Stoklund Olesen | fa2d986 | 2012-11-28 02:35:13 +0000 | [diff] [blame] | 745 | |
Andrew Trick | b86a0cd | 2013-06-15 04:49:57 +0000 | [diff] [blame] | 746 | int DefLatency = computeDefOperandLatency(ItinData, DefMI); |
Jakob Stoklund Olesen | fa2d986 | 2012-11-28 02:35:13 +0000 | [diff] [blame] | 747 | if (DefLatency >= 0) |
| 748 | return DefLatency; |
| 749 | |
| 750 | assert(ItinData && !ItinData->isEmpty() && "computeDefOperandLatency fail"); |
| 751 | |
| 752 | int OperLatency = 0; |
| 753 | if (UseMI) |
| 754 | OperLatency = getOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx); |
| 755 | else { |
| 756 | unsigned DefClass = DefMI->getDesc().getSchedClass(); |
| 757 | OperLatency = ItinData->getOperandCycle(DefClass, DefIdx); |
| 758 | } |
| 759 | if (OperLatency >= 0) |
| 760 | return OperLatency; |
| 761 | |
| 762 | // No operand latency was found. |
| 763 | unsigned InstrLatency = getInstrLatency(ItinData, DefMI); |
| 764 | |
| 765 | // Expected latency is the max of the stage latency and itinerary props. |
Andrew Trick | b86a0cd | 2013-06-15 04:49:57 +0000 | [diff] [blame] | 766 | InstrLatency = std::max(InstrLatency, |
| 767 | defaultDefLatency(ItinData->SchedModel, DefMI)); |
Jakob Stoklund Olesen | fa2d986 | 2012-11-28 02:35:13 +0000 | [diff] [blame] | 768 | return InstrLatency; |
| 769 | } |