1 //===- BitTracker.cpp -----------------------------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 // SSA-based bit propagation. 10 // 11 // The purpose of this code is, for a given virtual register, to provide 12 // information about the value of each bit in the register. The values 13 // of bits are represented by the class BitValue, and take one of four 14 // cases: 0, 1, "ref" and "bottom". The 0 and 1 are rather clear, the 15 // "ref" value means that the bit is a copy of another bit (which itself 16 // cannot be a copy of yet another bit---such chains are not allowed). 17 // A "ref" value is associated with a BitRef structure, which indicates 18 // which virtual register, and which bit in that register is the origin 19 // of the value. For example, given an instruction 20 // %2 = ASL %1, 1 21 // assuming that nothing is known about bits of %1, bit 1 of %2 22 // will be a "ref" to (%1, 0). If there is a subsequent instruction 23 // %3 = ASL %2, 2 24 // then bit 3 of %3 will be a "ref" to (%1, 0) as well. 25 // The "bottom" case means that the bit's value cannot be determined, 26 // and that this virtual register actually defines it. The "bottom" case 27 // is discussed in detail in BitTracker.h. In fact, "bottom" is a "ref 28 // to self", so for the %1 above, the bit 0 of it will be a "ref" to 29 // (%1, 0), bit 1 will be a "ref" to (%1, 1), etc. 30 // 31 // The tracker implements the Wegman-Zadeck algorithm, originally developed 32 // for SSA-based constant propagation. Each register is represented as 33 // a sequence of bits, with the convention that bit 0 is the least signi- 34 // ficant bit. Each bit is propagated individually. The class RegisterCell 35 // implements the register's representation, and is also the subject of 36 // the lattice operations in the tracker. 37 // 38 // The intended usage of the bit tracker is to create a target-specific 39 // machine instruction evaluator, pass the evaluator to the BitTracker 40 // object, and run the tracker. The tracker will then collect the bit 41 // value information for a given machine function. After that, it can be 42 // queried for the cells for each virtual register. 43 // Sample code: 44 // const TargetSpecificEvaluator TSE(TRI, MRI); 45 // BitTracker BT(TSE, MF); 46 // BT.run(); 47 // ... 48 // unsigned Reg = interestingRegister(); 49 // RegisterCell RC = BT.get(Reg); 50 // if (RC[3].is(1)) 51 // Reg0bit3 = 1; 52 // 53 // The code below is intended to be fully target-independent. 54 55 #include "BitTracker.h" 56 #include "llvm/ADT/APInt.h" 57 #include "llvm/ADT/BitVector.h" 58 #include "llvm/CodeGen/MachineBasicBlock.h" 59 #include "llvm/CodeGen/MachineFunction.h" 60 #include "llvm/CodeGen/MachineInstr.h" 61 #include "llvm/CodeGen/MachineOperand.h" 62 #include "llvm/CodeGen/MachineRegisterInfo.h" 63 #include "llvm/IR/Constants.h" 64 #include "llvm/Support/Debug.h" 65 66 using namespace llvm; 67 68 using BT = BitTracker; 69 70 namespace { 71 72 // Local trickery to pretty print a register (without the whole "%number" 73 // business). 74 struct printv { 75 printv(unsigned r) : R(r) {} 76 77 unsigned R; 78 }; 79 80 raw_ostream &operator<< (raw_ostream &OS, const printv &PV) { 81 if (PV.R) 82 OS << 'v' << Register(PV.R).virtRegIndex(); 83 else 84 OS << 's'; 85 return OS; 86 } 87 88 } // end anonymous namespace 89 90 namespace llvm { 91 92 raw_ostream &operator<<(raw_ostream &OS, const BT::BitValue &BV) { 93 switch (BV.Type) { 94 case BT::BitValue::Top: 95 OS << 'T'; 96 break; 97 case BT::BitValue::Zero: 98 OS << '0'; 99 break; 100 case BT::BitValue::One: 101 OS << '1'; 102 break; 103 case BT::BitValue::Ref: 104 OS << printv(BV.RefI.Reg) << '[' << BV.RefI.Pos << ']'; 105 break; 106 } 107 return OS; 108 } 109 110 raw_ostream &operator<<(raw_ostream &OS, const BT::RegisterCell &RC) { 111 unsigned n = RC.Bits.size(); 112 OS << "{ w:" << n; 113 // Instead of printing each bit value individually, try to group them 114 // into logical segments, such as sequences of 0 or 1 bits or references 115 // to consecutive bits (e.g. "bits 3-5 are same as bits 7-9 of reg xyz"). 116 // "Start" will be the index of the beginning of the most recent segment. 117 unsigned Start = 0; 118 bool SeqRef = false; // A sequence of refs to consecutive bits. 119 bool ConstRef = false; // A sequence of refs to the same bit. 120 121 for (unsigned i = 1, n = RC.Bits.size(); i < n; ++i) { 122 const BT::BitValue &V = RC[i]; 123 const BT::BitValue &SV = RC[Start]; 124 bool IsRef = (V.Type == BT::BitValue::Ref); 125 // If the current value is the same as Start, skip to the next one. 126 if (!IsRef && V == SV) 127 continue; 128 if (IsRef && SV.Type == BT::BitValue::Ref && V.RefI.Reg == SV.RefI.Reg) { 129 if (Start+1 == i) { 130 SeqRef = (V.RefI.Pos == SV.RefI.Pos+1); 131 ConstRef = (V.RefI.Pos == SV.RefI.Pos); 132 } 133 if (SeqRef && V.RefI.Pos == SV.RefI.Pos+(i-Start)) 134 continue; 135 if (ConstRef && V.RefI.Pos == SV.RefI.Pos) 136 continue; 137 } 138 139 // The current value is different. Print the previous one and reset 140 // the Start. 141 OS << " [" << Start; 142 unsigned Count = i - Start; 143 if (Count == 1) { 144 OS << "]:" << SV; 145 } else { 146 OS << '-' << i-1 << "]:"; 147 if (SV.Type == BT::BitValue::Ref && SeqRef) 148 OS << printv(SV.RefI.Reg) << '[' << SV.RefI.Pos << '-' 149 << SV.RefI.Pos+(Count-1) << ']'; 150 else 151 OS << SV; 152 } 153 Start = i; 154 SeqRef = ConstRef = false; 155 } 156 157 OS << " [" << Start; 158 unsigned Count = n - Start; 159 if (n-Start == 1) { 160 OS << "]:" << RC[Start]; 161 } else { 162 OS << '-' << n-1 << "]:"; 163 const BT::BitValue &SV = RC[Start]; 164 if (SV.Type == BT::BitValue::Ref && SeqRef) 165 OS << printv(SV.RefI.Reg) << '[' << SV.RefI.Pos << '-' 166 << SV.RefI.Pos+(Count-1) << ']'; 167 else 168 OS << SV; 169 } 170 OS << " }"; 171 172 return OS; 173 } 174 175 } // end namespace llvm 176 177 void BitTracker::print_cells(raw_ostream &OS) const { 178 for (const std::pair<unsigned, RegisterCell> P : Map) 179 dbgs() << printReg(P.first, &ME.TRI) << " -> " << P.second << "\n"; 180 } 181 182 BitTracker::BitTracker(const MachineEvaluator &E, MachineFunction &F) 183 : ME(E), MF(F), MRI(F.getRegInfo()), Map(*new CellMapType), Trace(false) { 184 } 185 186 BitTracker::~BitTracker() { 187 delete ⤅ 188 } 189 190 // If we were allowed to update a cell for a part of a register, the meet 191 // operation would need to be parametrized by the register number and the 192 // exact part of the register, so that the computer BitRefs correspond to 193 // the actual bits of the "self" register. 194 // While this cannot happen in the current implementation, I'm not sure 195 // if this should be ruled out in the future. 196 bool BT::RegisterCell::meet(const RegisterCell &RC, Register SelfR) { 197 // An example when "meet" can be invoked with SelfR == 0 is a phi node 198 // with a physical register as an operand. 199 assert(SelfR == 0 || SelfR.isVirtual()); 200 bool Changed = false; 201 for (uint16_t i = 0, n = Bits.size(); i < n; ++i) { 202 const BitValue &RCV = RC[i]; 203 Changed |= Bits[i].meet(RCV, BitRef(SelfR, i)); 204 } 205 return Changed; 206 } 207 208 // Insert the entire cell RC into the current cell at position given by M. 209 BT::RegisterCell &BT::RegisterCell::insert(const BT::RegisterCell &RC, 210 const BitMask &M) { 211 uint16_t B = M.first(), E = M.last(), W = width(); 212 // M must be a valid mask for *this. 213 assert(B < W && E < W); 214 // The masked part of *this must have the same number of bits 215 // as the source. 216 assert(B > E || E-B+1 == RC.width()); // B <= E => E-B+1 = |RC|. 217 assert(B <= E || E+(W-B)+1 == RC.width()); // E < B => E+(W-B)+1 = |RC|. 218 if (B <= E) { 219 for (uint16_t i = 0; i <= E-B; ++i) 220 Bits[i+B] = RC[i]; 221 } else { 222 for (uint16_t i = 0; i < W-B; ++i) 223 Bits[i+B] = RC[i]; 224 for (uint16_t i = 0; i <= E; ++i) 225 Bits[i] = RC[i+(W-B)]; 226 } 227 return *this; 228 } 229 230 BT::RegisterCell BT::RegisterCell::extract(const BitMask &M) const { 231 uint16_t B = M.first(), E = M.last(), W = width(); 232 assert(B < W && E < W); 233 if (B <= E) { 234 RegisterCell RC(E-B+1); 235 for (uint16_t i = B; i <= E; ++i) 236 RC.Bits[i-B] = Bits[i]; 237 return RC; 238 } 239 240 RegisterCell RC(E+(W-B)+1); 241 for (uint16_t i = 0; i < W-B; ++i) 242 RC.Bits[i] = Bits[i+B]; 243 for (uint16_t i = 0; i <= E; ++i) 244 RC.Bits[i+(W-B)] = Bits[i]; 245 return RC; 246 } 247 248 BT::RegisterCell &BT::RegisterCell::rol(uint16_t Sh) { 249 // Rotate left (i.e. towards increasing bit indices). 250 // Swap the two parts: [0..W-Sh-1] [W-Sh..W-1] 251 uint16_t W = width(); 252 Sh = Sh % W; 253 if (Sh == 0) 254 return *this; 255 256 RegisterCell Tmp(W-Sh); 257 // Tmp = [0..W-Sh-1]. 258 for (uint16_t i = 0; i < W-Sh; ++i) 259 Tmp[i] = Bits[i]; 260 // Shift [W-Sh..W-1] to [0..Sh-1]. 261 for (uint16_t i = 0; i < Sh; ++i) 262 Bits[i] = Bits[W-Sh+i]; 263 // Copy Tmp to [Sh..W-1]. 264 for (uint16_t i = 0; i < W-Sh; ++i) 265 Bits[i+Sh] = Tmp.Bits[i]; 266 return *this; 267 } 268 269 BT::RegisterCell &BT::RegisterCell::fill(uint16_t B, uint16_t E, 270 const BitValue &V) { 271 assert(B <= E); 272 while (B < E) 273 Bits[B++] = V; 274 return *this; 275 } 276 277 BT::RegisterCell &BT::RegisterCell::cat(const RegisterCell &RC) { 278 // Append the cell given as the argument to the "this" cell. 279 // Bit 0 of RC becomes bit W of the result, where W is this->width(). 280 uint16_t W = width(), WRC = RC.width(); 281 Bits.resize(W+WRC); 282 for (uint16_t i = 0; i < WRC; ++i) 283 Bits[i+W] = RC.Bits[i]; 284 return *this; 285 } 286 287 uint16_t BT::RegisterCell::ct(bool B) const { 288 uint16_t W = width(); 289 uint16_t C = 0; 290 BitValue V = B; 291 while (C < W && Bits[C] == V) 292 C++; 293 return C; 294 } 295 296 uint16_t BT::RegisterCell::cl(bool B) const { 297 uint16_t W = width(); 298 uint16_t C = 0; 299 BitValue V = B; 300 while (C < W && Bits[W-(C+1)] == V) 301 C++; 302 return C; 303 } 304 305 bool BT::RegisterCell::operator== (const RegisterCell &RC) const { 306 uint16_t W = Bits.size(); 307 if (RC.Bits.size() != W) 308 return false; 309 for (uint16_t i = 0; i < W; ++i) 310 if (Bits[i] != RC[i]) 311 return false; 312 return true; 313 } 314 315 BT::RegisterCell &BT::RegisterCell::regify(unsigned R) { 316 for (unsigned i = 0, n = width(); i < n; ++i) { 317 const BitValue &V = Bits[i]; 318 if (V.Type == BitValue::Ref && V.RefI.Reg == 0) 319 Bits[i].RefI = BitRef(R, i); 320 } 321 return *this; 322 } 323 324 uint16_t BT::MachineEvaluator::getRegBitWidth(const RegisterRef &RR) const { 325 // The general problem is with finding a register class that corresponds 326 // to a given reference reg:sub. There can be several such classes, and 327 // since we only care about the register size, it does not matter which 328 // such class we would find. 329 // The easiest way to accomplish what we want is to 330 // 1. find a physical register PhysR from the same class as RR.Reg, 331 // 2. find a physical register PhysS that corresponds to PhysR:RR.Sub, 332 // 3. find a register class that contains PhysS. 333 if (RR.Reg.isVirtual()) { 334 const auto &VC = composeWithSubRegIndex(*MRI.getRegClass(RR.Reg), RR.Sub); 335 return TRI.getRegSizeInBits(VC); 336 } 337 assert(RR.Reg.isPhysical()); 338 MCRegister PhysR = 339 (RR.Sub == 0) ? RR.Reg.asMCReg() : TRI.getSubReg(RR.Reg, RR.Sub); 340 return getPhysRegBitWidth(PhysR); 341 } 342 343 BT::RegisterCell BT::MachineEvaluator::getCell(const RegisterRef &RR, 344 const CellMapType &M) const { 345 uint16_t BW = getRegBitWidth(RR); 346 347 // Physical registers are assumed to be present in the map with an unknown 348 // value. Don't actually insert anything in the map, just return the cell. 349 if (RR.Reg.isPhysical()) 350 return RegisterCell::self(0, BW); 351 352 assert(RR.Reg.isVirtual()); 353 // For virtual registers that belong to a class that is not tracked, 354 // generate an "unknown" value as well. 355 const TargetRegisterClass *C = MRI.getRegClass(RR.Reg); 356 if (!track(C)) 357 return RegisterCell::self(0, BW); 358 359 CellMapType::const_iterator F = M.find(RR.Reg); 360 if (F != M.end()) { 361 if (!RR.Sub) 362 return F->second; 363 BitMask M = mask(RR.Reg, RR.Sub); 364 return F->second.extract(M); 365 } 366 // If not found, create a "top" entry, but do not insert it in the map. 367 return RegisterCell::top(BW); 368 } 369 370 void BT::MachineEvaluator::putCell(const RegisterRef &RR, RegisterCell RC, 371 CellMapType &M) const { 372 // While updating the cell map can be done in a meaningful way for 373 // a part of a register, it makes little sense to implement it as the 374 // SSA representation would never contain such "partial definitions". 375 if (!RR.Reg.isVirtual()) 376 return; 377 assert(RR.Sub == 0 && "Unexpected sub-register in definition"); 378 // Eliminate all ref-to-reg-0 bit values: replace them with "self". 379 M[RR.Reg] = RC.regify(RR.Reg); 380 } 381 382 // Check if the cell represents a compile-time integer value. 383 bool BT::MachineEvaluator::isInt(const RegisterCell &A) const { 384 uint16_t W = A.width(); 385 for (uint16_t i = 0; i < W; ++i) 386 if (!A[i].is(0) && !A[i].is(1)) 387 return false; 388 return true; 389 } 390 391 // Convert a cell to the integer value. The result must fit in uint64_t. 392 uint64_t BT::MachineEvaluator::toInt(const RegisterCell &A) const { 393 assert(isInt(A)); 394 uint64_t Val = 0; 395 uint16_t W = A.width(); 396 for (uint16_t i = 0; i < W; ++i) { 397 Val <<= 1; 398 Val |= A[i].is(1); 399 } 400 return Val; 401 } 402 403 // Evaluator helper functions. These implement some common operation on 404 // register cells that can be used to implement target-specific instructions 405 // in a target-specific evaluator. 406 407 BT::RegisterCell BT::MachineEvaluator::eIMM(int64_t V, uint16_t W) const { 408 RegisterCell Res(W); 409 // For bits beyond the 63rd, this will generate the sign bit of V. 410 for (uint16_t i = 0; i < W; ++i) { 411 Res[i] = BitValue(V & 1); 412 V >>= 1; 413 } 414 return Res; 415 } 416 417 BT::RegisterCell BT::MachineEvaluator::eIMM(const ConstantInt *CI) const { 418 const APInt &A = CI->getValue(); 419 uint16_t BW = A.getBitWidth(); 420 assert((unsigned)BW == A.getBitWidth() && "BitWidth overflow"); 421 RegisterCell Res(BW); 422 for (uint16_t i = 0; i < BW; ++i) 423 Res[i] = A[i]; 424 return Res; 425 } 426 427 BT::RegisterCell BT::MachineEvaluator::eADD(const RegisterCell &A1, 428 const RegisterCell &A2) const { 429 uint16_t W = A1.width(); 430 assert(W == A2.width()); 431 RegisterCell Res(W); 432 bool Carry = false; 433 uint16_t I; 434 for (I = 0; I < W; ++I) { 435 const BitValue &V1 = A1[I]; 436 const BitValue &V2 = A2[I]; 437 if (!V1.num() || !V2.num()) 438 break; 439 unsigned S = bool(V1) + bool(V2) + Carry; 440 Res[I] = BitValue(S & 1); 441 Carry = (S > 1); 442 } 443 for (; I < W; ++I) { 444 const BitValue &V1 = A1[I]; 445 const BitValue &V2 = A2[I]; 446 // If the next bit is same as Carry, the result will be 0 plus the 447 // other bit. The Carry bit will remain unchanged. 448 if (V1.is(Carry)) 449 Res[I] = BitValue::ref(V2); 450 else if (V2.is(Carry)) 451 Res[I] = BitValue::ref(V1); 452 else 453 break; 454 } 455 for (; I < W; ++I) 456 Res[I] = BitValue::self(); 457 return Res; 458 } 459 460 BT::RegisterCell BT::MachineEvaluator::eSUB(const RegisterCell &A1, 461 const RegisterCell &A2) const { 462 uint16_t W = A1.width(); 463 assert(W == A2.width()); 464 RegisterCell Res(W); 465 bool Borrow = false; 466 uint16_t I; 467 for (I = 0; I < W; ++I) { 468 const BitValue &V1 = A1[I]; 469 const BitValue &V2 = A2[I]; 470 if (!V1.num() || !V2.num()) 471 break; 472 unsigned S = bool(V1) - bool(V2) - Borrow; 473 Res[I] = BitValue(S & 1); 474 Borrow = (S > 1); 475 } 476 for (; I < W; ++I) { 477 const BitValue &V1 = A1[I]; 478 const BitValue &V2 = A2[I]; 479 if (V1.is(Borrow)) { 480 Res[I] = BitValue::ref(V2); 481 break; 482 } 483 if (V2.is(Borrow)) 484 Res[I] = BitValue::ref(V1); 485 else 486 break; 487 } 488 for (; I < W; ++I) 489 Res[I] = BitValue::self(); 490 return Res; 491 } 492 493 BT::RegisterCell BT::MachineEvaluator::eMLS(const RegisterCell &A1, 494 const RegisterCell &A2) const { 495 uint16_t W = A1.width() + A2.width(); 496 uint16_t Z = A1.ct(false) + A2.ct(false); 497 RegisterCell Res(W); 498 Res.fill(0, Z, BitValue::Zero); 499 Res.fill(Z, W, BitValue::self()); 500 return Res; 501 } 502 503 BT::RegisterCell BT::MachineEvaluator::eMLU(const RegisterCell &A1, 504 const RegisterCell &A2) const { 505 uint16_t W = A1.width() + A2.width(); 506 uint16_t Z = A1.ct(false) + A2.ct(false); 507 RegisterCell Res(W); 508 Res.fill(0, Z, BitValue::Zero); 509 Res.fill(Z, W, BitValue::self()); 510 return Res; 511 } 512 513 BT::RegisterCell BT::MachineEvaluator::eASL(const RegisterCell &A1, 514 uint16_t Sh) const { 515 assert(Sh <= A1.width()); 516 RegisterCell Res = RegisterCell::ref(A1); 517 Res.rol(Sh); 518 Res.fill(0, Sh, BitValue::Zero); 519 return Res; 520 } 521 522 BT::RegisterCell BT::MachineEvaluator::eLSR(const RegisterCell &A1, 523 uint16_t Sh) const { 524 uint16_t W = A1.width(); 525 assert(Sh <= W); 526 RegisterCell Res = RegisterCell::ref(A1); 527 Res.rol(W-Sh); 528 Res.fill(W-Sh, W, BitValue::Zero); 529 return Res; 530 } 531 532 BT::RegisterCell BT::MachineEvaluator::eASR(const RegisterCell &A1, 533 uint16_t Sh) const { 534 uint16_t W = A1.width(); 535 assert(Sh <= W); 536 RegisterCell Res = RegisterCell::ref(A1); 537 BitValue Sign = Res[W-1]; 538 Res.rol(W-Sh); 539 Res.fill(W-Sh, W, Sign); 540 return Res; 541 } 542 543 BT::RegisterCell BT::MachineEvaluator::eAND(const RegisterCell &A1, 544 const RegisterCell &A2) const { 545 uint16_t W = A1.width(); 546 assert(W == A2.width()); 547 RegisterCell Res(W); 548 for (uint16_t i = 0; i < W; ++i) { 549 const BitValue &V1 = A1[i]; 550 const BitValue &V2 = A2[i]; 551 if (V1.is(1)) 552 Res[i] = BitValue::ref(V2); 553 else if (V2.is(1)) 554 Res[i] = BitValue::ref(V1); 555 else if (V1.is(0) || V2.is(0)) 556 Res[i] = BitValue::Zero; 557 else if (V1 == V2) 558 Res[i] = V1; 559 else 560 Res[i] = BitValue::self(); 561 } 562 return Res; 563 } 564 565 BT::RegisterCell BT::MachineEvaluator::eORL(const RegisterCell &A1, 566 const RegisterCell &A2) const { 567 uint16_t W = A1.width(); 568 assert(W == A2.width()); 569 RegisterCell Res(W); 570 for (uint16_t i = 0; i < W; ++i) { 571 const BitValue &V1 = A1[i]; 572 const BitValue &V2 = A2[i]; 573 if (V1.is(1) || V2.is(1)) 574 Res[i] = BitValue::One; 575 else if (V1.is(0)) 576 Res[i] = BitValue::ref(V2); 577 else if (V2.is(0)) 578 Res[i] = BitValue::ref(V1); 579 else if (V1 == V2) 580 Res[i] = V1; 581 else 582 Res[i] = BitValue::self(); 583 } 584 return Res; 585 } 586 587 BT::RegisterCell BT::MachineEvaluator::eXOR(const RegisterCell &A1, 588 const RegisterCell &A2) const { 589 uint16_t W = A1.width(); 590 assert(W == A2.width()); 591 RegisterCell Res(W); 592 for (uint16_t i = 0; i < W; ++i) { 593 const BitValue &V1 = A1[i]; 594 const BitValue &V2 = A2[i]; 595 if (V1.is(0)) 596 Res[i] = BitValue::ref(V2); 597 else if (V2.is(0)) 598 Res[i] = BitValue::ref(V1); 599 else if (V1 == V2) 600 Res[i] = BitValue::Zero; 601 else 602 Res[i] = BitValue::self(); 603 } 604 return Res; 605 } 606 607 BT::RegisterCell BT::MachineEvaluator::eNOT(const RegisterCell &A1) const { 608 uint16_t W = A1.width(); 609 RegisterCell Res(W); 610 for (uint16_t i = 0; i < W; ++i) { 611 const BitValue &V = A1[i]; 612 if (V.is(0)) 613 Res[i] = BitValue::One; 614 else if (V.is(1)) 615 Res[i] = BitValue::Zero; 616 else 617 Res[i] = BitValue::self(); 618 } 619 return Res; 620 } 621 622 BT::RegisterCell BT::MachineEvaluator::eSET(const RegisterCell &A1, 623 uint16_t BitN) const { 624 assert(BitN < A1.width()); 625 RegisterCell Res = RegisterCell::ref(A1); 626 Res[BitN] = BitValue::One; 627 return Res; 628 } 629 630 BT::RegisterCell BT::MachineEvaluator::eCLR(const RegisterCell &A1, 631 uint16_t BitN) const { 632 assert(BitN < A1.width()); 633 RegisterCell Res = RegisterCell::ref(A1); 634 Res[BitN] = BitValue::Zero; 635 return Res; 636 } 637 638 BT::RegisterCell BT::MachineEvaluator::eCLB(const RegisterCell &A1, bool B, 639 uint16_t W) const { 640 uint16_t C = A1.cl(B), AW = A1.width(); 641 // If the last leading non-B bit is not a constant, then we don't know 642 // the real count. 643 if ((C < AW && A1[AW-1-C].num()) || C == AW) 644 return eIMM(C, W); 645 return RegisterCell::self(0, W); 646 } 647 648 BT::RegisterCell BT::MachineEvaluator::eCTB(const RegisterCell &A1, bool B, 649 uint16_t W) const { 650 uint16_t C = A1.ct(B), AW = A1.width(); 651 // If the last trailing non-B bit is not a constant, then we don't know 652 // the real count. 653 if ((C < AW && A1[C].num()) || C == AW) 654 return eIMM(C, W); 655 return RegisterCell::self(0, W); 656 } 657 658 BT::RegisterCell BT::MachineEvaluator::eSXT(const RegisterCell &A1, 659 uint16_t FromN) const { 660 uint16_t W = A1.width(); 661 assert(FromN <= W); 662 RegisterCell Res = RegisterCell::ref(A1); 663 BitValue Sign = Res[FromN-1]; 664 // Sign-extend "inreg". 665 Res.fill(FromN, W, Sign); 666 return Res; 667 } 668 669 BT::RegisterCell BT::MachineEvaluator::eZXT(const RegisterCell &A1, 670 uint16_t FromN) const { 671 uint16_t W = A1.width(); 672 assert(FromN <= W); 673 RegisterCell Res = RegisterCell::ref(A1); 674 Res.fill(FromN, W, BitValue::Zero); 675 return Res; 676 } 677 678 BT::RegisterCell BT::MachineEvaluator::eXTR(const RegisterCell &A1, 679 uint16_t B, uint16_t E) const { 680 uint16_t W = A1.width(); 681 assert(B < W && E <= W); 682 if (B == E) 683 return RegisterCell(0); 684 uint16_t Last = (E > 0) ? E-1 : W-1; 685 RegisterCell Res = RegisterCell::ref(A1).extract(BT::BitMask(B, Last)); 686 // Return shorter cell. 687 return Res; 688 } 689 690 BT::RegisterCell BT::MachineEvaluator::eINS(const RegisterCell &A1, 691 const RegisterCell &A2, uint16_t AtN) const { 692 uint16_t W1 = A1.width(), W2 = A2.width(); 693 (void)W1; 694 assert(AtN < W1 && AtN+W2 <= W1); 695 // Copy bits from A1, insert A2 at position AtN. 696 RegisterCell Res = RegisterCell::ref(A1); 697 if (W2 > 0) 698 Res.insert(RegisterCell::ref(A2), BT::BitMask(AtN, AtN+W2-1)); 699 return Res; 700 } 701 702 BT::BitMask BT::MachineEvaluator::mask(Register Reg, unsigned Sub) const { 703 assert(Sub == 0 && "Generic BitTracker::mask called for Sub != 0"); 704 uint16_t W = getRegBitWidth(Reg); 705 assert(W > 0 && "Cannot generate mask for empty register"); 706 return BitMask(0, W-1); 707 } 708 709 uint16_t BT::MachineEvaluator::getPhysRegBitWidth(MCRegister Reg) const { 710 const TargetRegisterClass &PC = *TRI.getMinimalPhysRegClass(Reg); 711 return TRI.getRegSizeInBits(PC); 712 } 713 714 bool BT::MachineEvaluator::evaluate(const MachineInstr &MI, 715 const CellMapType &Inputs, 716 CellMapType &Outputs) const { 717 unsigned Opc = MI.getOpcode(); 718 switch (Opc) { 719 case TargetOpcode::REG_SEQUENCE: { 720 RegisterRef RD = MI.getOperand(0); 721 assert(RD.Sub == 0); 722 RegisterRef RS = MI.getOperand(1); 723 unsigned SS = MI.getOperand(2).getImm(); 724 RegisterRef RT = MI.getOperand(3); 725 unsigned ST = MI.getOperand(4).getImm(); 726 assert(SS != ST); 727 728 uint16_t W = getRegBitWidth(RD); 729 RegisterCell Res(W); 730 Res.insert(RegisterCell::ref(getCell(RS, Inputs)), mask(RD.Reg, SS)); 731 Res.insert(RegisterCell::ref(getCell(RT, Inputs)), mask(RD.Reg, ST)); 732 putCell(RD, Res, Outputs); 733 break; 734 } 735 736 case TargetOpcode::COPY: { 737 // COPY can transfer a smaller register into a wider one. 738 // If that is the case, fill the remaining high bits with 0. 739 RegisterRef RD = MI.getOperand(0); 740 RegisterRef RS = MI.getOperand(1); 741 assert(RD.Sub == 0); 742 uint16_t WD = getRegBitWidth(RD); 743 uint16_t WS = getRegBitWidth(RS); 744 assert(WD >= WS); 745 RegisterCell Src = getCell(RS, Inputs); 746 RegisterCell Res(WD); 747 Res.insert(Src, BitMask(0, WS-1)); 748 Res.fill(WS, WD, BitValue::Zero); 749 putCell(RD, Res, Outputs); 750 break; 751 } 752 753 default: 754 return false; 755 } 756 757 return true; 758 } 759 760 bool BT::UseQueueType::Cmp::operator()(const MachineInstr *InstA, 761 const MachineInstr *InstB) const { 762 // This is a comparison function for a priority queue: give higher priority 763 // to earlier instructions. 764 // This operator is used as "less", so returning "true" gives InstB higher 765 // priority (because then InstA < InstB). 766 if (InstA == InstB) 767 return false; 768 const MachineBasicBlock *BA = InstA->getParent(); 769 const MachineBasicBlock *BB = InstB->getParent(); 770 if (BA != BB) { 771 // If the blocks are different, ideally the dominating block would 772 // have a higher priority, but it may be too expensive to check. 773 return BA->getNumber() > BB->getNumber(); 774 } 775 776 auto getDist = [this] (const MachineInstr *MI) { 777 auto F = Dist.find(MI); 778 if (F != Dist.end()) 779 return F->second; 780 MachineBasicBlock::const_iterator I = MI->getParent()->begin(); 781 MachineBasicBlock::const_iterator E = MI->getIterator(); 782 unsigned D = std::distance(I, E); 783 Dist.insert(std::make_pair(MI, D)); 784 return D; 785 }; 786 787 return getDist(InstA) > getDist(InstB); 788 } 789 790 // Main W-Z implementation. 791 792 void BT::visitPHI(const MachineInstr &PI) { 793 int ThisN = PI.getParent()->getNumber(); 794 if (Trace) 795 dbgs() << "Visit FI(" << printMBBReference(*PI.getParent()) << "): " << PI; 796 797 const MachineOperand &MD = PI.getOperand(0); 798 assert(MD.getSubReg() == 0 && "Unexpected sub-register in definition"); 799 RegisterRef DefRR(MD); 800 uint16_t DefBW = ME.getRegBitWidth(DefRR); 801 802 RegisterCell DefC = ME.getCell(DefRR, Map); 803 if (DefC == RegisterCell::self(DefRR.Reg, DefBW)) // XXX slow 804 return; 805 806 bool Changed = false; 807 808 for (unsigned i = 1, n = PI.getNumOperands(); i < n; i += 2) { 809 const MachineBasicBlock *PB = PI.getOperand(i + 1).getMBB(); 810 int PredN = PB->getNumber(); 811 if (Trace) 812 dbgs() << " edge " << printMBBReference(*PB) << "->" 813 << printMBBReference(*PI.getParent()); 814 if (!EdgeExec.count(CFGEdge(PredN, ThisN))) { 815 if (Trace) 816 dbgs() << " not executable\n"; 817 continue; 818 } 819 820 RegisterRef RU = PI.getOperand(i); 821 RegisterCell ResC = ME.getCell(RU, Map); 822 if (Trace) 823 dbgs() << " input reg: " << printReg(RU.Reg, &ME.TRI, RU.Sub) 824 << " cell: " << ResC << "\n"; 825 Changed |= DefC.meet(ResC, DefRR.Reg); 826 } 827 828 if (Changed) { 829 if (Trace) 830 dbgs() << "Output: " << printReg(DefRR.Reg, &ME.TRI, DefRR.Sub) 831 << " cell: " << DefC << "\n"; 832 ME.putCell(DefRR, DefC, Map); 833 visitUsesOf(DefRR.Reg); 834 } 835 } 836 837 void BT::visitNonBranch(const MachineInstr &MI) { 838 if (Trace) 839 dbgs() << "Visit MI(" << printMBBReference(*MI.getParent()) << "): " << MI; 840 if (MI.isDebugInstr()) 841 return; 842 assert(!MI.isBranch() && "Unexpected branch instruction"); 843 844 CellMapType ResMap; 845 bool Eval = ME.evaluate(MI, Map, ResMap); 846 847 if (Trace && Eval) { 848 for (const MachineOperand &MO : MI.operands()) { 849 if (!MO.isReg() || !MO.isUse()) 850 continue; 851 RegisterRef RU(MO); 852 dbgs() << " input reg: " << printReg(RU.Reg, &ME.TRI, RU.Sub) 853 << " cell: " << ME.getCell(RU, Map) << "\n"; 854 } 855 dbgs() << "Outputs:\n"; 856 for (const std::pair<const unsigned, RegisterCell> &P : ResMap) { 857 RegisterRef RD(P.first); 858 dbgs() << " " << printReg(P.first, &ME.TRI) << " cell: " 859 << ME.getCell(RD, ResMap) << "\n"; 860 } 861 } 862 863 // Iterate over all definitions of the instruction, and update the 864 // cells accordingly. 865 for (const MachineOperand &MO : MI.operands()) { 866 // Visit register defs only. 867 if (!MO.isReg() || !MO.isDef()) 868 continue; 869 RegisterRef RD(MO); 870 assert(RD.Sub == 0 && "Unexpected sub-register in definition"); 871 if (!RD.Reg.isVirtual()) 872 continue; 873 874 bool Changed = false; 875 if (!Eval || ResMap.count(RD.Reg) == 0) { 876 // Set to "ref" (aka "bottom"). 877 uint16_t DefBW = ME.getRegBitWidth(RD); 878 RegisterCell RefC = RegisterCell::self(RD.Reg, DefBW); 879 if (RefC != ME.getCell(RD, Map)) { 880 ME.putCell(RD, RefC, Map); 881 Changed = true; 882 } 883 } else { 884 RegisterCell DefC = ME.getCell(RD, Map); 885 RegisterCell ResC = ME.getCell(RD, ResMap); 886 // This is a non-phi instruction, so the values of the inputs come 887 // from the same registers each time this instruction is evaluated. 888 // During the propagation, the values of the inputs can become lowered 889 // in the sense of the lattice operation, which may cause different 890 // results to be calculated in subsequent evaluations. This should 891 // not cause the bottoming of the result in the map, since the new 892 // result is already reflecting the lowered inputs. 893 for (uint16_t i = 0, w = DefC.width(); i < w; ++i) { 894 BitValue &V = DefC[i]; 895 // Bits that are already "bottom" should not be updated. 896 if (V.Type == BitValue::Ref && V.RefI.Reg == RD.Reg) 897 continue; 898 // Same for those that are identical in DefC and ResC. 899 if (V == ResC[i]) 900 continue; 901 V = ResC[i]; 902 Changed = true; 903 } 904 if (Changed) 905 ME.putCell(RD, DefC, Map); 906 } 907 if (Changed) 908 visitUsesOf(RD.Reg); 909 } 910 } 911 912 void BT::visitBranchesFrom(const MachineInstr &BI) { 913 const MachineBasicBlock &B = *BI.getParent(); 914 MachineBasicBlock::const_iterator It = BI, End = B.end(); 915 BranchTargetList Targets, BTs; 916 bool FallsThrough = true, DefaultToAll = false; 917 int ThisN = B.getNumber(); 918 919 do { 920 BTs.clear(); 921 const MachineInstr &MI = *It; 922 if (Trace) 923 dbgs() << "Visit BR(" << printMBBReference(B) << "): " << MI; 924 assert(MI.isBranch() && "Expecting branch instruction"); 925 InstrExec.insert(&MI); 926 bool Eval = ME.evaluate(MI, Map, BTs, FallsThrough); 927 if (!Eval) { 928 // If the evaluation failed, we will add all targets. Keep going in 929 // the loop to mark all executable branches as such. 930 DefaultToAll = true; 931 FallsThrough = true; 932 if (Trace) 933 dbgs() << " failed to evaluate: will add all CFG successors\n"; 934 } else if (!DefaultToAll) { 935 // If evaluated successfully add the targets to the cumulative list. 936 if (Trace) { 937 dbgs() << " adding targets:"; 938 for (const MachineBasicBlock *BT : BTs) 939 dbgs() << " " << printMBBReference(*BT); 940 if (FallsThrough) 941 dbgs() << "\n falls through\n"; 942 else 943 dbgs() << "\n does not fall through\n"; 944 } 945 Targets.insert_range(BTs); 946 } 947 ++It; 948 } while (FallsThrough && It != End); 949 950 if (B.mayHaveInlineAsmBr()) 951 DefaultToAll = true; 952 953 if (!DefaultToAll) { 954 // Need to add all CFG successors that lead to EH landing pads. 955 // There won't be explicit branches to these blocks, but they must 956 // be processed. 957 for (const MachineBasicBlock *SB : B.successors()) { 958 if (SB->isEHPad()) 959 Targets.insert(SB); 960 } 961 if (FallsThrough) { 962 MachineFunction::const_iterator BIt = B.getIterator(); 963 MachineFunction::const_iterator Next = std::next(BIt); 964 if (Next != MF.end()) 965 Targets.insert(&*Next); 966 } 967 } else { 968 Targets.insert_range(B.successors()); 969 } 970 971 for (const MachineBasicBlock *TB : Targets) 972 FlowQ.push(CFGEdge(ThisN, TB->getNumber())); 973 } 974 975 void BT::visitUsesOf(Register Reg) { 976 if (Trace) 977 dbgs() << "queuing uses of modified reg " << printReg(Reg, &ME.TRI) 978 << " cell: " << ME.getCell(Reg, Map) << '\n'; 979 980 for (MachineInstr &UseI : MRI.use_nodbg_instructions(Reg)) 981 UseQ.push(&UseI); 982 } 983 984 BT::RegisterCell BT::get(RegisterRef RR) const { 985 return ME.getCell(RR, Map); 986 } 987 988 void BT::put(RegisterRef RR, const RegisterCell &RC) { 989 ME.putCell(RR, RC, Map); 990 } 991 992 // Replace all references to bits from OldRR with the corresponding bits 993 // in NewRR. 994 void BT::subst(RegisterRef OldRR, RegisterRef NewRR) { 995 assert(Map.count(OldRR.Reg) > 0 && "OldRR not present in map"); 996 BitMask OM = ME.mask(OldRR.Reg, OldRR.Sub); 997 BitMask NM = ME.mask(NewRR.Reg, NewRR.Sub); 998 uint16_t OMB = OM.first(), OME = OM.last(); 999 uint16_t NMB = NM.first(), NME = NM.last(); 1000 (void)NME; 1001 assert((OME-OMB == NME-NMB) && 1002 "Substituting registers of different lengths"); 1003 for (std::pair<const unsigned, RegisterCell> &P : Map) { 1004 RegisterCell &RC = P.second; 1005 for (uint16_t i = 0, w = RC.width(); i < w; ++i) { 1006 BitValue &V = RC[i]; 1007 if (V.Type != BitValue::Ref || V.RefI.Reg != OldRR.Reg) 1008 continue; 1009 if (V.RefI.Pos < OMB || V.RefI.Pos > OME) 1010 continue; 1011 V.RefI.Reg = NewRR.Reg; 1012 V.RefI.Pos += NMB-OMB; 1013 } 1014 } 1015 } 1016 1017 // Check if the block has been "executed" during propagation. (If not, the 1018 // block is dead, but it may still appear to be reachable.) 1019 bool BT::reached(const MachineBasicBlock *B) const { 1020 int BN = B->getNumber(); 1021 assert(BN >= 0); 1022 return ReachedBB.count(BN); 1023 } 1024 1025 // Visit an individual instruction. This could be a newly added instruction, 1026 // or one that has been modified by an optimization. 1027 void BT::visit(const MachineInstr &MI) { 1028 assert(!MI.isBranch() && "Only non-branches are allowed"); 1029 InstrExec.insert(&MI); 1030 visitNonBranch(MI); 1031 // Make sure to flush all the pending use updates. 1032 runUseQueue(); 1033 // The call to visitNonBranch could propagate the changes until a branch 1034 // is actually visited. This could result in adding CFG edges to the flow 1035 // queue. Since the queue won't be processed, clear it. 1036 while (!FlowQ.empty()) 1037 FlowQ.pop(); 1038 } 1039 1040 void BT::reset() { 1041 EdgeExec.clear(); 1042 InstrExec.clear(); 1043 Map.clear(); 1044 ReachedBB.clear(); 1045 ReachedBB.reserve(MF.size()); 1046 } 1047 1048 void BT::runEdgeQueue(BitVector &BlockScanned) { 1049 while (!FlowQ.empty()) { 1050 CFGEdge Edge = FlowQ.front(); 1051 FlowQ.pop(); 1052 1053 if (!EdgeExec.insert(Edge).second) 1054 return; 1055 ReachedBB.insert(Edge.second); 1056 1057 const MachineBasicBlock &B = *MF.getBlockNumbered(Edge.second); 1058 MachineBasicBlock::const_iterator It = B.begin(), End = B.end(); 1059 // Visit PHI nodes first. 1060 while (It != End && It->isPHI()) { 1061 const MachineInstr &PI = *It++; 1062 InstrExec.insert(&PI); 1063 visitPHI(PI); 1064 } 1065 1066 // If this block has already been visited through a flow graph edge, 1067 // then the instructions have already been processed. Any updates to 1068 // the cells would now only happen through visitUsesOf... 1069 if (BlockScanned[Edge.second]) 1070 return; 1071 BlockScanned[Edge.second] = true; 1072 1073 // Visit non-branch instructions. 1074 while (It != End && !It->isBranch()) { 1075 const MachineInstr &MI = *It++; 1076 InstrExec.insert(&MI); 1077 visitNonBranch(MI); 1078 } 1079 // If block end has been reached, add the fall-through edge to the queue. 1080 if (It == End) { 1081 MachineFunction::const_iterator BIt = B.getIterator(); 1082 MachineFunction::const_iterator Next = std::next(BIt); 1083 if (Next != MF.end() && B.isSuccessor(&*Next)) { 1084 int ThisN = B.getNumber(); 1085 int NextN = Next->getNumber(); 1086 FlowQ.push(CFGEdge(ThisN, NextN)); 1087 } 1088 } else { 1089 // Handle the remaining sequence of branches. This function will update 1090 // the work queue. 1091 visitBranchesFrom(*It); 1092 } 1093 } // while (!FlowQ->empty()) 1094 } 1095 1096 void BT::runUseQueue() { 1097 while (!UseQ.empty()) { 1098 MachineInstr &UseI = *UseQ.front(); 1099 UseQ.pop(); 1100 1101 if (!InstrExec.count(&UseI)) 1102 continue; 1103 if (UseI.isPHI()) 1104 visitPHI(UseI); 1105 else if (!UseI.isBranch()) 1106 visitNonBranch(UseI); 1107 else 1108 visitBranchesFrom(UseI); 1109 } 1110 } 1111 1112 void BT::run() { 1113 reset(); 1114 assert(FlowQ.empty()); 1115 1116 using MachineFlowGraphTraits = GraphTraits<const MachineFunction*>; 1117 const MachineBasicBlock *Entry = MachineFlowGraphTraits::getEntryNode(&MF); 1118 1119 unsigned MaxBN = 0; 1120 for (const MachineBasicBlock &B : MF) { 1121 assert(B.getNumber() >= 0 && "Disconnected block"); 1122 unsigned BN = B.getNumber(); 1123 if (BN > MaxBN) 1124 MaxBN = BN; 1125 } 1126 1127 // Keep track of visited blocks. 1128 BitVector BlockScanned(MaxBN+1); 1129 1130 int EntryN = Entry->getNumber(); 1131 // Generate a fake edge to get something to start with. 1132 FlowQ.push(CFGEdge(-1, EntryN)); 1133 1134 while (!FlowQ.empty() || !UseQ.empty()) { 1135 runEdgeQueue(BlockScanned); 1136 runUseQueue(); 1137 } 1138 UseQ.reset(); 1139 1140 if (Trace) 1141 print_cells(dbgs() << "Cells after propagation:\n"); 1142 } 1143