1 //===----------- PPCVSXSwapRemoval.cpp - Remove VSX LE Swaps -------------===// 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 // This pass analyzes vector computations and removes unnecessary 10 // doubleword swaps (xxswapd instructions). This pass is performed 11 // only for little-endian VSX code generation. 12 // 13 // For this specific case, loads and stores of v4i32, v4f32, v2i64, 14 // and v2f64 vectors are inefficient. These are implemented using 15 // the lxvd2x and stxvd2x instructions, which invert the order of 16 // doublewords in a vector register. Thus code generation inserts 17 // an xxswapd after each such load, and prior to each such store. 18 // 19 // The extra xxswapd instructions reduce performance. The purpose 20 // of this pass is to reduce the number of xxswapd instructions 21 // required for correctness. 22 // 23 // The primary insight is that much code that operates on vectors 24 // does not care about the relative order of elements in a register, 25 // so long as the correct memory order is preserved. If we have a 26 // computation where all input values are provided by lxvd2x/xxswapd, 27 // all outputs are stored using xxswapd/lxvd2x, and all intermediate 28 // computations are lane-insensitive (independent of element order), 29 // then all the xxswapd instructions associated with the loads and 30 // stores may be removed without changing observable semantics. 31 // 32 // This pass uses standard equivalence class infrastructure to create 33 // maximal webs of computations fitting the above description. Each 34 // such web is then optimized by removing its unnecessary xxswapd 35 // instructions. 36 // 37 // There are some lane-sensitive operations for which we can still 38 // permit the optimization, provided we modify those operations 39 // accordingly. Such operations are identified as using "special 40 // handling" within this module. 41 // 42 //===---------------------------------------------------------------------===// 43 44 #include "PPC.h" 45 #include "PPCInstrBuilder.h" 46 #include "PPCInstrInfo.h" 47 #include "PPCTargetMachine.h" 48 #include "llvm/ADT/DenseMap.h" 49 #include "llvm/ADT/EquivalenceClasses.h" 50 #include "llvm/CodeGen/MachineFunctionPass.h" 51 #include "llvm/CodeGen/MachineInstrBuilder.h" 52 #include "llvm/CodeGen/MachineRegisterInfo.h" 53 #include "llvm/Config/llvm-config.h" 54 #include "llvm/Support/Debug.h" 55 #include "llvm/Support/Format.h" 56 #include "llvm/Support/raw_ostream.h" 57 58 using namespace llvm; 59 60 #define DEBUG_TYPE "ppc-vsx-swaps" 61 62 namespace { 63 64 // A PPCVSXSwapEntry is created for each machine instruction that 65 // is relevant to a vector computation. 66 struct PPCVSXSwapEntry { 67 // Pointer to the instruction. 68 MachineInstr *VSEMI; 69 70 // Unique ID (position in the swap vector). 71 int VSEId; 72 73 // Attributes of this node. 74 unsigned int IsLoad : 1; 75 unsigned int IsStore : 1; 76 unsigned int IsSwap : 1; 77 unsigned int MentionsPhysVR : 1; 78 unsigned int IsSwappable : 1; 79 unsigned int MentionsPartialVR : 1; 80 unsigned int SpecialHandling : 3; 81 unsigned int WebRejected : 1; 82 unsigned int WillRemove : 1; 83 }; 84 85 enum SHValues { 86 SH_NONE = 0, 87 SH_EXTRACT, 88 SH_INSERT, 89 SH_NOSWAP_LD, 90 SH_NOSWAP_ST, 91 SH_SPLAT, 92 SH_XXPERMDI, 93 SH_COPYWIDEN 94 }; 95 96 struct PPCVSXSwapRemoval : public MachineFunctionPass { 97 98 static char ID; 99 const PPCInstrInfo *TII; 100 MachineFunction *MF; 101 MachineRegisterInfo *MRI; 102 103 // Swap entries are allocated in a vector for better performance. 104 std::vector<PPCVSXSwapEntry> SwapVector; 105 106 // A mapping is maintained between machine instructions and 107 // their swap entries. The key is the address of the MI. 108 DenseMap<MachineInstr*, int> SwapMap; 109 110 // Equivalence classes are used to gather webs of related computation. 111 // Swap entries are represented by their VSEId fields. 112 EquivalenceClasses<int> *EC; 113 114 PPCVSXSwapRemoval() : MachineFunctionPass(ID) { 115 initializePPCVSXSwapRemovalPass(*PassRegistry::getPassRegistry()); 116 } 117 118 private: 119 // Initialize data structures. 120 void initialize(MachineFunction &MFParm); 121 122 // Walk the machine instructions to gather vector usage information. 123 // Return true iff vector mentions are present. 124 bool gatherVectorInstructions(); 125 126 // Add an entry to the swap vector and swap map. 127 int addSwapEntry(MachineInstr *MI, PPCVSXSwapEntry &SwapEntry); 128 129 // Hunt backwards through COPY and SUBREG_TO_REG chains for a 130 // source register. VecIdx indicates the swap vector entry to 131 // mark as mentioning a physical register if the search leads 132 // to one. 133 unsigned lookThruCopyLike(unsigned SrcReg, unsigned VecIdx); 134 135 // Generate equivalence classes for related computations (webs). 136 void formWebs(); 137 138 // Analyze webs and determine those that cannot be optimized. 139 void recordUnoptimizableWebs(); 140 141 // Record which swap instructions can be safely removed. 142 void markSwapsForRemoval(); 143 144 // Remove swaps and update other instructions requiring special 145 // handling. Return true iff any changes are made. 146 bool removeSwaps(); 147 148 // Insert a swap instruction from SrcReg to DstReg at the given 149 // InsertPoint. 150 void insertSwap(MachineInstr *MI, MachineBasicBlock::iterator InsertPoint, 151 unsigned DstReg, unsigned SrcReg); 152 153 // Update instructions requiring special handling. 154 void handleSpecialSwappables(int EntryIdx); 155 156 // Dump a description of the entries in the swap vector. 157 void dumpSwapVector(); 158 159 // Return true iff the given register is in the given class. 160 bool isRegInClass(unsigned Reg, const TargetRegisterClass *RC) { 161 if (Register::isVirtualRegister(Reg)) 162 return RC->hasSubClassEq(MRI->getRegClass(Reg)); 163 return RC->contains(Reg); 164 } 165 166 // Return true iff the given register is a full vector register. 167 bool isVecReg(unsigned Reg) { 168 return (isRegInClass(Reg, &PPC::VSRCRegClass) || 169 isRegInClass(Reg, &PPC::VRRCRegClass)); 170 } 171 172 // Return true iff the given register is a partial vector register. 173 bool isScalarVecReg(unsigned Reg) { 174 return (isRegInClass(Reg, &PPC::VSFRCRegClass) || 175 isRegInClass(Reg, &PPC::VSSRCRegClass)); 176 } 177 178 // Return true iff the given register mentions all or part of a 179 // vector register. Also sets Partial to true if the mention 180 // is for just the floating-point register overlap of the register. 181 bool isAnyVecReg(unsigned Reg, bool &Partial) { 182 if (isScalarVecReg(Reg)) 183 Partial = true; 184 return isScalarVecReg(Reg) || isVecReg(Reg); 185 } 186 187 public: 188 // Main entry point for this pass. 189 bool runOnMachineFunction(MachineFunction &MF) override { 190 if (skipFunction(MF.getFunction())) 191 return false; 192 193 // If we don't have VSX on the subtarget, don't do anything. 194 // Also, on Power 9 the load and store ops preserve element order and so 195 // the swaps are not required. 196 const PPCSubtarget &STI = MF.getSubtarget<PPCSubtarget>(); 197 if (!STI.hasVSX() || !STI.needsSwapsForVSXMemOps()) 198 return false; 199 200 bool Changed = false; 201 initialize(MF); 202 203 if (gatherVectorInstructions()) { 204 formWebs(); 205 recordUnoptimizableWebs(); 206 markSwapsForRemoval(); 207 Changed = removeSwaps(); 208 } 209 210 // FIXME: See the allocation of EC in initialize(). 211 delete EC; 212 return Changed; 213 } 214 }; 215 216 // Initialize data structures for this pass. In particular, clear the 217 // swap vector and allocate the equivalence class mapping before 218 // processing each function. 219 void PPCVSXSwapRemoval::initialize(MachineFunction &MFParm) { 220 MF = &MFParm; 221 MRI = &MF->getRegInfo(); 222 TII = MF->getSubtarget<PPCSubtarget>().getInstrInfo(); 223 224 // An initial vector size of 256 appears to work well in practice. 225 // Small/medium functions with vector content tend not to incur a 226 // reallocation at this size. Three of the vector tests in 227 // projects/test-suite reallocate, which seems like a reasonable rate. 228 const int InitialVectorSize(256); 229 SwapVector.clear(); 230 SwapVector.reserve(InitialVectorSize); 231 232 // FIXME: Currently we allocate EC each time because we don't have 233 // access to the set representation on which to call clear(). Should 234 // consider adding a clear() method to the EquivalenceClasses class. 235 EC = new EquivalenceClasses<int>; 236 } 237 238 // Create an entry in the swap vector for each instruction that mentions 239 // a full vector register, recording various characteristics of the 240 // instructions there. 241 bool PPCVSXSwapRemoval::gatherVectorInstructions() { 242 bool RelevantFunction = false; 243 244 for (MachineBasicBlock &MBB : *MF) { 245 for (MachineInstr &MI : MBB) { 246 247 if (MI.isDebugInstr()) 248 continue; 249 250 bool RelevantInstr = false; 251 bool Partial = false; 252 253 for (const MachineOperand &MO : MI.operands()) { 254 if (!MO.isReg()) 255 continue; 256 Register Reg = MO.getReg(); 257 if (isAnyVecReg(Reg, Partial)) { 258 RelevantInstr = true; 259 break; 260 } 261 } 262 263 if (!RelevantInstr) 264 continue; 265 266 RelevantFunction = true; 267 268 // Create a SwapEntry initialized to zeros, then fill in the 269 // instruction and ID fields before pushing it to the back 270 // of the swap vector. 271 PPCVSXSwapEntry SwapEntry{}; 272 int VecIdx = addSwapEntry(&MI, SwapEntry); 273 274 switch(MI.getOpcode()) { 275 default: 276 // Unless noted otherwise, an instruction is considered 277 // safe for the optimization. There are a large number of 278 // such true-SIMD instructions (all vector math, logical, 279 // select, compare, etc.). However, if the instruction 280 // mentions a partial vector register and does not have 281 // special handling defined, it is not swappable. 282 if (Partial) 283 SwapVector[VecIdx].MentionsPartialVR = 1; 284 else 285 SwapVector[VecIdx].IsSwappable = 1; 286 break; 287 case PPC::XXPERMDI: { 288 // This is a swap if it is of the form XXPERMDI t, s, s, 2. 289 // Unfortunately, MachineCSE ignores COPY and SUBREG_TO_REG, so we 290 // can also see XXPERMDI t, SUBREG_TO_REG(s), SUBREG_TO_REG(s), 2, 291 // for example. We have to look through chains of COPY and 292 // SUBREG_TO_REG to find the real source value for comparison. 293 // If the real source value is a physical register, then mark the 294 // XXPERMDI as mentioning a physical register. 295 int immed = MI.getOperand(3).getImm(); 296 if (immed == 2) { 297 unsigned trueReg1 = lookThruCopyLike(MI.getOperand(1).getReg(), 298 VecIdx); 299 unsigned trueReg2 = lookThruCopyLike(MI.getOperand(2).getReg(), 300 VecIdx); 301 if (trueReg1 == trueReg2) 302 SwapVector[VecIdx].IsSwap = 1; 303 else { 304 // We can still handle these if the two registers are not 305 // identical, by adjusting the form of the XXPERMDI. 306 SwapVector[VecIdx].IsSwappable = 1; 307 SwapVector[VecIdx].SpecialHandling = SHValues::SH_XXPERMDI; 308 } 309 // This is a doubleword splat if it is of the form 310 // XXPERMDI t, s, s, 0 or XXPERMDI t, s, s, 3. As above we 311 // must look through chains of copy-likes to find the source 312 // register. We turn off the marking for mention of a physical 313 // register, because splatting it is safe; the optimization 314 // will not swap the value in the physical register. Whether 315 // or not the two input registers are identical, we can handle 316 // these by adjusting the form of the XXPERMDI. 317 } else if (immed == 0 || immed == 3) { 318 319 SwapVector[VecIdx].IsSwappable = 1; 320 SwapVector[VecIdx].SpecialHandling = SHValues::SH_XXPERMDI; 321 322 unsigned trueReg1 = lookThruCopyLike(MI.getOperand(1).getReg(), 323 VecIdx); 324 unsigned trueReg2 = lookThruCopyLike(MI.getOperand(2).getReg(), 325 VecIdx); 326 if (trueReg1 == trueReg2) 327 SwapVector[VecIdx].MentionsPhysVR = 0; 328 329 } else { 330 // We can still handle these by adjusting the form of the XXPERMDI. 331 SwapVector[VecIdx].IsSwappable = 1; 332 SwapVector[VecIdx].SpecialHandling = SHValues::SH_XXPERMDI; 333 } 334 break; 335 } 336 case PPC::LVX: 337 // Non-permuting loads are currently unsafe. We can use special 338 // handling for this in the future. By not marking these as 339 // IsSwap, we ensure computations containing them will be rejected 340 // for now. 341 SwapVector[VecIdx].IsLoad = 1; 342 break; 343 case PPC::LXVD2X: 344 case PPC::LXVW4X: 345 // Permuting loads are marked as both load and swap, and are 346 // safe for optimization. 347 SwapVector[VecIdx].IsLoad = 1; 348 SwapVector[VecIdx].IsSwap = 1; 349 break; 350 case PPC::LXSDX: 351 case PPC::LXSSPX: 352 case PPC::XFLOADf64: 353 case PPC::XFLOADf32: 354 // A load of a floating-point value into the high-order half of 355 // a vector register is safe, provided that we introduce a swap 356 // following the load, which will be done by the SUBREG_TO_REG 357 // support. So just mark these as safe. 358 SwapVector[VecIdx].IsLoad = 1; 359 SwapVector[VecIdx].IsSwappable = 1; 360 break; 361 case PPC::STVX: 362 // Non-permuting stores are currently unsafe. We can use special 363 // handling for this in the future. By not marking these as 364 // IsSwap, we ensure computations containing them will be rejected 365 // for now. 366 SwapVector[VecIdx].IsStore = 1; 367 break; 368 case PPC::STXVD2X: 369 case PPC::STXVW4X: 370 // Permuting stores are marked as both store and swap, and are 371 // safe for optimization. 372 SwapVector[VecIdx].IsStore = 1; 373 SwapVector[VecIdx].IsSwap = 1; 374 break; 375 case PPC::COPY: 376 // These are fine provided they are moving between full vector 377 // register classes. 378 if (isVecReg(MI.getOperand(0).getReg()) && 379 isVecReg(MI.getOperand(1).getReg())) 380 SwapVector[VecIdx].IsSwappable = 1; 381 // If we have a copy from one scalar floating-point register 382 // to another, we can accept this even if it is a physical 383 // register. The only way this gets involved is if it feeds 384 // a SUBREG_TO_REG, which is handled by introducing a swap. 385 else if (isScalarVecReg(MI.getOperand(0).getReg()) && 386 isScalarVecReg(MI.getOperand(1).getReg())) 387 SwapVector[VecIdx].IsSwappable = 1; 388 break; 389 case PPC::SUBREG_TO_REG: { 390 // These are fine provided they are moving between full vector 391 // register classes. If they are moving from a scalar 392 // floating-point class to a vector class, we can handle those 393 // as well, provided we introduce a swap. It is generally the 394 // case that we will introduce fewer swaps than we remove, but 395 // (FIXME) a cost model could be used. However, introduced 396 // swaps could potentially be CSEd, so this is not trivial. 397 if (isVecReg(MI.getOperand(0).getReg()) && 398 isVecReg(MI.getOperand(2).getReg())) 399 SwapVector[VecIdx].IsSwappable = 1; 400 else if (isVecReg(MI.getOperand(0).getReg()) && 401 isScalarVecReg(MI.getOperand(2).getReg())) { 402 SwapVector[VecIdx].IsSwappable = 1; 403 SwapVector[VecIdx].SpecialHandling = SHValues::SH_COPYWIDEN; 404 } 405 break; 406 } 407 case PPC::VSPLTB: 408 case PPC::VSPLTH: 409 case PPC::VSPLTW: 410 case PPC::XXSPLTW: 411 // Splats are lane-sensitive, but we can use special handling 412 // to adjust the source lane for the splat. 413 SwapVector[VecIdx].IsSwappable = 1; 414 SwapVector[VecIdx].SpecialHandling = SHValues::SH_SPLAT; 415 break; 416 // The presence of the following lane-sensitive operations in a 417 // web will kill the optimization, at least for now. For these 418 // we do nothing, causing the optimization to fail. 419 // FIXME: Some of these could be permitted with special handling, 420 // and will be phased in as time permits. 421 // FIXME: There is no simple and maintainable way to express a set 422 // of opcodes having a common attribute in TableGen. Should this 423 // change, this is a prime candidate to use such a mechanism. 424 case PPC::INLINEASM: 425 case PPC::INLINEASM_BR: 426 case PPC::EXTRACT_SUBREG: 427 case PPC::INSERT_SUBREG: 428 case PPC::COPY_TO_REGCLASS: 429 case PPC::LVEBX: 430 case PPC::LVEHX: 431 case PPC::LVEWX: 432 case PPC::LVSL: 433 case PPC::LVSR: 434 case PPC::LVXL: 435 case PPC::STVEBX: 436 case PPC::STVEHX: 437 case PPC::STVEWX: 438 case PPC::STVXL: 439 // We can handle STXSDX and STXSSPX similarly to LXSDX and LXSSPX, 440 // by adding special handling for narrowing copies as well as 441 // widening ones. However, I've experimented with this, and in 442 // practice we currently do not appear to use STXSDX fed by 443 // a narrowing copy from a full vector register. Since I can't 444 // generate any useful test cases, I've left this alone for now. 445 case PPC::STXSDX: 446 case PPC::STXSSPX: 447 case PPC::VCIPHER: 448 case PPC::VCIPHERLAST: 449 case PPC::VMRGHB: 450 case PPC::VMRGHH: 451 case PPC::VMRGHW: 452 case PPC::VMRGLB: 453 case PPC::VMRGLH: 454 case PPC::VMRGLW: 455 case PPC::VMULESB: 456 case PPC::VMULESH: 457 case PPC::VMULESW: 458 case PPC::VMULEUB: 459 case PPC::VMULEUH: 460 case PPC::VMULEUW: 461 case PPC::VMULOSB: 462 case PPC::VMULOSH: 463 case PPC::VMULOSW: 464 case PPC::VMULOUB: 465 case PPC::VMULOUH: 466 case PPC::VMULOUW: 467 case PPC::VNCIPHER: 468 case PPC::VNCIPHERLAST: 469 case PPC::VPERM: 470 case PPC::VPERMXOR: 471 case PPC::VPKPX: 472 case PPC::VPKSHSS: 473 case PPC::VPKSHUS: 474 case PPC::VPKSDSS: 475 case PPC::VPKSDUS: 476 case PPC::VPKSWSS: 477 case PPC::VPKSWUS: 478 case PPC::VPKUDUM: 479 case PPC::VPKUDUS: 480 case PPC::VPKUHUM: 481 case PPC::VPKUHUS: 482 case PPC::VPKUWUM: 483 case PPC::VPKUWUS: 484 case PPC::VPMSUMB: 485 case PPC::VPMSUMD: 486 case PPC::VPMSUMH: 487 case PPC::VPMSUMW: 488 case PPC::VRLB: 489 case PPC::VRLD: 490 case PPC::VRLH: 491 case PPC::VRLW: 492 case PPC::VSBOX: 493 case PPC::VSHASIGMAD: 494 case PPC::VSHASIGMAW: 495 case PPC::VSL: 496 case PPC::VSLDOI: 497 case PPC::VSLO: 498 case PPC::VSR: 499 case PPC::VSRO: 500 case PPC::VSUM2SWS: 501 case PPC::VSUM4SBS: 502 case PPC::VSUM4SHS: 503 case PPC::VSUM4UBS: 504 case PPC::VSUMSWS: 505 case PPC::VUPKHPX: 506 case PPC::VUPKHSB: 507 case PPC::VUPKHSH: 508 case PPC::VUPKHSW: 509 case PPC::VUPKLPX: 510 case PPC::VUPKLSB: 511 case PPC::VUPKLSH: 512 case PPC::VUPKLSW: 513 case PPC::XXMRGHW: 514 case PPC::XXMRGLW: 515 // XXSLDWI could be replaced by a general permute with one of three 516 // permute control vectors (for shift values 1, 2, 3). However, 517 // VPERM has a more restrictive register class. 518 case PPC::XXSLDWI: 519 case PPC::XSCVDPSPN: 520 case PPC::XSCVSPDPN: 521 break; 522 } 523 } 524 } 525 526 if (RelevantFunction) { 527 LLVM_DEBUG(dbgs() << "Swap vector when first built\n\n"); 528 LLVM_DEBUG(dumpSwapVector()); 529 } 530 531 return RelevantFunction; 532 } 533 534 // Add an entry to the swap vector and swap map, and make a 535 // singleton equivalence class for the entry. 536 int PPCVSXSwapRemoval::addSwapEntry(MachineInstr *MI, 537 PPCVSXSwapEntry& SwapEntry) { 538 SwapEntry.VSEMI = MI; 539 SwapEntry.VSEId = SwapVector.size(); 540 SwapVector.push_back(SwapEntry); 541 EC->insert(SwapEntry.VSEId); 542 SwapMap[MI] = SwapEntry.VSEId; 543 return SwapEntry.VSEId; 544 } 545 546 // This is used to find the "true" source register for an 547 // XXPERMDI instruction, since MachineCSE does not handle the 548 // "copy-like" operations (Copy and SubregToReg). Returns 549 // the original SrcReg unless it is the target of a copy-like 550 // operation, in which case we chain backwards through all 551 // such operations to the ultimate source register. If a 552 // physical register is encountered, we stop the search and 553 // flag the swap entry indicated by VecIdx (the original 554 // XXPERMDI) as mentioning a physical register. 555 unsigned PPCVSXSwapRemoval::lookThruCopyLike(unsigned SrcReg, 556 unsigned VecIdx) { 557 MachineInstr *MI = MRI->getVRegDef(SrcReg); 558 if (!MI->isCopyLike()) 559 return SrcReg; 560 561 unsigned CopySrcReg; 562 if (MI->isCopy()) 563 CopySrcReg = MI->getOperand(1).getReg(); 564 else { 565 assert(MI->isSubregToReg() && "bad opcode for lookThruCopyLike"); 566 CopySrcReg = MI->getOperand(2).getReg(); 567 } 568 569 if (!Register::isVirtualRegister(CopySrcReg)) { 570 if (!isScalarVecReg(CopySrcReg)) 571 SwapVector[VecIdx].MentionsPhysVR = 1; 572 return CopySrcReg; 573 } 574 575 return lookThruCopyLike(CopySrcReg, VecIdx); 576 } 577 578 // Generate equivalence classes for related computations (webs) by 579 // def-use relationships of virtual registers. Mention of a physical 580 // register terminates the generation of equivalence classes as this 581 // indicates a use of a parameter, definition of a return value, use 582 // of a value returned from a call, or definition of a parameter to a 583 // call. Computations with physical register mentions are flagged 584 // as such so their containing webs will not be optimized. 585 void PPCVSXSwapRemoval::formWebs() { 586 587 LLVM_DEBUG(dbgs() << "\n*** Forming webs for swap removal ***\n\n"); 588 589 for (unsigned EntryIdx = 0; EntryIdx < SwapVector.size(); ++EntryIdx) { 590 591 MachineInstr *MI = SwapVector[EntryIdx].VSEMI; 592 593 LLVM_DEBUG(dbgs() << "\n" << SwapVector[EntryIdx].VSEId << " "); 594 LLVM_DEBUG(MI->dump()); 595 596 // It's sufficient to walk vector uses and join them to their unique 597 // definitions. In addition, check full vector register operands 598 // for physical regs. We exclude partial-vector register operands 599 // because we can handle them if copied to a full vector. 600 for (const MachineOperand &MO : MI->operands()) { 601 if (!MO.isReg()) 602 continue; 603 604 Register Reg = MO.getReg(); 605 if (!isVecReg(Reg) && !isScalarVecReg(Reg)) 606 continue; 607 608 if (!Register::isVirtualRegister(Reg)) { 609 if (!(MI->isCopy() && isScalarVecReg(Reg))) 610 SwapVector[EntryIdx].MentionsPhysVR = 1; 611 continue; 612 } 613 614 if (!MO.isUse()) 615 continue; 616 617 MachineInstr* DefMI = MRI->getVRegDef(Reg); 618 assert(SwapMap.find(DefMI) != SwapMap.end() && 619 "Inconsistency: def of vector reg not found in swap map!"); 620 int DefIdx = SwapMap[DefMI]; 621 (void)EC->unionSets(SwapVector[DefIdx].VSEId, 622 SwapVector[EntryIdx].VSEId); 623 624 LLVM_DEBUG(dbgs() << format("Unioning %d with %d\n", 625 SwapVector[DefIdx].VSEId, 626 SwapVector[EntryIdx].VSEId)); 627 LLVM_DEBUG(dbgs() << " Def: "); 628 LLVM_DEBUG(DefMI->dump()); 629 } 630 } 631 } 632 633 // Walk the swap vector entries looking for conditions that prevent their 634 // containing computations from being optimized. When such conditions are 635 // found, mark the representative of the computation's equivalence class 636 // as rejected. 637 void PPCVSXSwapRemoval::recordUnoptimizableWebs() { 638 639 LLVM_DEBUG(dbgs() << "\n*** Rejecting webs for swap removal ***\n\n"); 640 641 for (unsigned EntryIdx = 0; EntryIdx < SwapVector.size(); ++EntryIdx) { 642 int Repr = EC->getLeaderValue(SwapVector[EntryIdx].VSEId); 643 644 // If representative is already rejected, don't waste further time. 645 if (SwapVector[Repr].WebRejected) 646 continue; 647 648 // Reject webs containing mentions of physical or partial registers, or 649 // containing operations that we don't know how to handle in a lane- 650 // permuted region. 651 if (SwapVector[EntryIdx].MentionsPhysVR || 652 SwapVector[EntryIdx].MentionsPartialVR || 653 !(SwapVector[EntryIdx].IsSwappable || SwapVector[EntryIdx].IsSwap)) { 654 655 SwapVector[Repr].WebRejected = 1; 656 657 LLVM_DEBUG( 658 dbgs() << format("Web %d rejected for physreg, partial reg, or not " 659 "swap[pable]\n", 660 Repr)); 661 LLVM_DEBUG(dbgs() << " in " << EntryIdx << ": "); 662 LLVM_DEBUG(SwapVector[EntryIdx].VSEMI->dump()); 663 LLVM_DEBUG(dbgs() << "\n"); 664 } 665 666 // Reject webs than contain swapping loads that feed something other 667 // than a swap instruction. 668 else if (SwapVector[EntryIdx].IsLoad && SwapVector[EntryIdx].IsSwap) { 669 MachineInstr *MI = SwapVector[EntryIdx].VSEMI; 670 Register DefReg = MI->getOperand(0).getReg(); 671 672 // We skip debug instructions in the analysis. (Note that debug 673 // location information is still maintained by this optimization 674 // because it remains on the LXVD2X and STXVD2X instructions after 675 // the XXPERMDIs are removed.) 676 for (MachineInstr &UseMI : MRI->use_nodbg_instructions(DefReg)) { 677 int UseIdx = SwapMap[&UseMI]; 678 679 if (!SwapVector[UseIdx].IsSwap || SwapVector[UseIdx].IsLoad || 680 SwapVector[UseIdx].IsStore) { 681 682 SwapVector[Repr].WebRejected = 1; 683 684 LLVM_DEBUG(dbgs() << format( 685 "Web %d rejected for load not feeding swap\n", Repr)); 686 LLVM_DEBUG(dbgs() << " def " << EntryIdx << ": "); 687 LLVM_DEBUG(MI->dump()); 688 LLVM_DEBUG(dbgs() << " use " << UseIdx << ": "); 689 LLVM_DEBUG(UseMI.dump()); 690 LLVM_DEBUG(dbgs() << "\n"); 691 } 692 } 693 694 // Reject webs that contain swapping stores that are fed by something 695 // other than a swap instruction. 696 } else if (SwapVector[EntryIdx].IsStore && SwapVector[EntryIdx].IsSwap) { 697 MachineInstr *MI = SwapVector[EntryIdx].VSEMI; 698 Register UseReg = MI->getOperand(0).getReg(); 699 MachineInstr *DefMI = MRI->getVRegDef(UseReg); 700 Register DefReg = DefMI->getOperand(0).getReg(); 701 int DefIdx = SwapMap[DefMI]; 702 703 if (!SwapVector[DefIdx].IsSwap || SwapVector[DefIdx].IsLoad || 704 SwapVector[DefIdx].IsStore) { 705 706 SwapVector[Repr].WebRejected = 1; 707 708 LLVM_DEBUG(dbgs() << format( 709 "Web %d rejected for store not fed by swap\n", Repr)); 710 LLVM_DEBUG(dbgs() << " def " << DefIdx << ": "); 711 LLVM_DEBUG(DefMI->dump()); 712 LLVM_DEBUG(dbgs() << " use " << EntryIdx << ": "); 713 LLVM_DEBUG(MI->dump()); 714 LLVM_DEBUG(dbgs() << "\n"); 715 } 716 717 // Ensure all uses of the register defined by DefMI feed store 718 // instructions 719 for (MachineInstr &UseMI : MRI->use_nodbg_instructions(DefReg)) { 720 int UseIdx = SwapMap[&UseMI]; 721 722 if (SwapVector[UseIdx].VSEMI->getOpcode() != MI->getOpcode()) { 723 SwapVector[Repr].WebRejected = 1; 724 725 LLVM_DEBUG( 726 dbgs() << format( 727 "Web %d rejected for swap not feeding only stores\n", Repr)); 728 LLVM_DEBUG(dbgs() << " def " 729 << " : "); 730 LLVM_DEBUG(DefMI->dump()); 731 LLVM_DEBUG(dbgs() << " use " << UseIdx << ": "); 732 LLVM_DEBUG(SwapVector[UseIdx].VSEMI->dump()); 733 LLVM_DEBUG(dbgs() << "\n"); 734 } 735 } 736 } 737 } 738 739 LLVM_DEBUG(dbgs() << "Swap vector after web analysis:\n\n"); 740 LLVM_DEBUG(dumpSwapVector()); 741 } 742 743 // Walk the swap vector entries looking for swaps fed by permuting loads 744 // and swaps that feed permuting stores. If the containing computation 745 // has not been marked rejected, mark each such swap for removal. 746 // (Removal is delayed in case optimization has disturbed the pattern, 747 // such that multiple loads feed the same swap, etc.) 748 void PPCVSXSwapRemoval::markSwapsForRemoval() { 749 750 LLVM_DEBUG(dbgs() << "\n*** Marking swaps for removal ***\n\n"); 751 752 for (unsigned EntryIdx = 0; EntryIdx < SwapVector.size(); ++EntryIdx) { 753 754 if (SwapVector[EntryIdx].IsLoad && SwapVector[EntryIdx].IsSwap) { 755 int Repr = EC->getLeaderValue(SwapVector[EntryIdx].VSEId); 756 757 if (!SwapVector[Repr].WebRejected) { 758 MachineInstr *MI = SwapVector[EntryIdx].VSEMI; 759 Register DefReg = MI->getOperand(0).getReg(); 760 761 for (MachineInstr &UseMI : MRI->use_nodbg_instructions(DefReg)) { 762 int UseIdx = SwapMap[&UseMI]; 763 SwapVector[UseIdx].WillRemove = 1; 764 765 LLVM_DEBUG(dbgs() << "Marking swap fed by load for removal: "); 766 LLVM_DEBUG(UseMI.dump()); 767 } 768 } 769 770 } else if (SwapVector[EntryIdx].IsStore && SwapVector[EntryIdx].IsSwap) { 771 int Repr = EC->getLeaderValue(SwapVector[EntryIdx].VSEId); 772 773 if (!SwapVector[Repr].WebRejected) { 774 MachineInstr *MI = SwapVector[EntryIdx].VSEMI; 775 Register UseReg = MI->getOperand(0).getReg(); 776 MachineInstr *DefMI = MRI->getVRegDef(UseReg); 777 int DefIdx = SwapMap[DefMI]; 778 SwapVector[DefIdx].WillRemove = 1; 779 780 LLVM_DEBUG(dbgs() << "Marking swap feeding store for removal: "); 781 LLVM_DEBUG(DefMI->dump()); 782 } 783 784 } else if (SwapVector[EntryIdx].IsSwappable && 785 SwapVector[EntryIdx].SpecialHandling != 0) { 786 int Repr = EC->getLeaderValue(SwapVector[EntryIdx].VSEId); 787 788 if (!SwapVector[Repr].WebRejected) 789 handleSpecialSwappables(EntryIdx); 790 } 791 } 792 } 793 794 // Create an xxswapd instruction and insert it prior to the given point. 795 // MI is used to determine basic block and debug loc information. 796 // FIXME: When inserting a swap, we should check whether SrcReg is 797 // defined by another swap: SrcReg = XXPERMDI Reg, Reg, 2; If so, 798 // then instead we should generate a copy from Reg to DstReg. 799 void PPCVSXSwapRemoval::insertSwap(MachineInstr *MI, 800 MachineBasicBlock::iterator InsertPoint, 801 unsigned DstReg, unsigned SrcReg) { 802 BuildMI(*MI->getParent(), InsertPoint, MI->getDebugLoc(), 803 TII->get(PPC::XXPERMDI), DstReg) 804 .addReg(SrcReg) 805 .addReg(SrcReg) 806 .addImm(2); 807 } 808 809 // The identified swap entry requires special handling to allow its 810 // containing computation to be optimized. Perform that handling 811 // here. 812 // FIXME: Additional opportunities will be phased in with subsequent 813 // patches. 814 void PPCVSXSwapRemoval::handleSpecialSwappables(int EntryIdx) { 815 switch (SwapVector[EntryIdx].SpecialHandling) { 816 817 default: 818 llvm_unreachable("Unexpected special handling type"); 819 820 // For splats based on an index into a vector, add N/2 modulo N 821 // to the index, where N is the number of vector elements. 822 case SHValues::SH_SPLAT: { 823 MachineInstr *MI = SwapVector[EntryIdx].VSEMI; 824 unsigned NElts; 825 826 LLVM_DEBUG(dbgs() << "Changing splat: "); 827 LLVM_DEBUG(MI->dump()); 828 829 switch (MI->getOpcode()) { 830 default: 831 llvm_unreachable("Unexpected splat opcode"); 832 case PPC::VSPLTB: NElts = 16; break; 833 case PPC::VSPLTH: NElts = 8; break; 834 case PPC::VSPLTW: 835 case PPC::XXSPLTW: NElts = 4; break; 836 } 837 838 unsigned EltNo; 839 if (MI->getOpcode() == PPC::XXSPLTW) 840 EltNo = MI->getOperand(2).getImm(); 841 else 842 EltNo = MI->getOperand(1).getImm(); 843 844 EltNo = (EltNo + NElts / 2) % NElts; 845 if (MI->getOpcode() == PPC::XXSPLTW) 846 MI->getOperand(2).setImm(EltNo); 847 else 848 MI->getOperand(1).setImm(EltNo); 849 850 LLVM_DEBUG(dbgs() << " Into: "); 851 LLVM_DEBUG(MI->dump()); 852 break; 853 } 854 855 // For an XXPERMDI that isn't handled otherwise, we need to 856 // reverse the order of the operands. If the selector operand 857 // has a value of 0 or 3, we need to change it to 3 or 0, 858 // respectively. Otherwise we should leave it alone. (This 859 // is equivalent to reversing the two bits of the selector 860 // operand and complementing the result.) 861 case SHValues::SH_XXPERMDI: { 862 MachineInstr *MI = SwapVector[EntryIdx].VSEMI; 863 864 LLVM_DEBUG(dbgs() << "Changing XXPERMDI: "); 865 LLVM_DEBUG(MI->dump()); 866 867 unsigned Selector = MI->getOperand(3).getImm(); 868 if (Selector == 0 || Selector == 3) 869 Selector = 3 - Selector; 870 MI->getOperand(3).setImm(Selector); 871 872 Register Reg1 = MI->getOperand(1).getReg(); 873 Register Reg2 = MI->getOperand(2).getReg(); 874 MI->getOperand(1).setReg(Reg2); 875 MI->getOperand(2).setReg(Reg1); 876 877 // We also need to swap kill flag associated with the register. 878 bool IsKill1 = MI->getOperand(1).isKill(); 879 bool IsKill2 = MI->getOperand(2).isKill(); 880 MI->getOperand(1).setIsKill(IsKill2); 881 MI->getOperand(2).setIsKill(IsKill1); 882 883 LLVM_DEBUG(dbgs() << " Into: "); 884 LLVM_DEBUG(MI->dump()); 885 break; 886 } 887 888 // For a copy from a scalar floating-point register to a vector 889 // register, removing swaps will leave the copied value in the 890 // wrong lane. Insert a swap following the copy to fix this. 891 case SHValues::SH_COPYWIDEN: { 892 MachineInstr *MI = SwapVector[EntryIdx].VSEMI; 893 894 LLVM_DEBUG(dbgs() << "Changing SUBREG_TO_REG: "); 895 LLVM_DEBUG(MI->dump()); 896 897 Register DstReg = MI->getOperand(0).getReg(); 898 const TargetRegisterClass *DstRC = MRI->getRegClass(DstReg); 899 Register NewVReg = MRI->createVirtualRegister(DstRC); 900 901 MI->getOperand(0).setReg(NewVReg); 902 LLVM_DEBUG(dbgs() << " Into: "); 903 LLVM_DEBUG(MI->dump()); 904 905 auto InsertPoint = ++MachineBasicBlock::iterator(MI); 906 907 // Note that an XXPERMDI requires a VSRC, so if the SUBREG_TO_REG 908 // is copying to a VRRC, we need to be careful to avoid a register 909 // assignment problem. In this case we must copy from VRRC to VSRC 910 // prior to the swap, and from VSRC to VRRC following the swap. 911 // Coalescing will usually remove all this mess. 912 if (DstRC == &PPC::VRRCRegClass) { 913 Register VSRCTmp1 = MRI->createVirtualRegister(&PPC::VSRCRegClass); 914 Register VSRCTmp2 = MRI->createVirtualRegister(&PPC::VSRCRegClass); 915 916 BuildMI(*MI->getParent(), InsertPoint, MI->getDebugLoc(), 917 TII->get(PPC::COPY), VSRCTmp1) 918 .addReg(NewVReg); 919 LLVM_DEBUG(std::prev(InsertPoint)->dump()); 920 921 insertSwap(MI, InsertPoint, VSRCTmp2, VSRCTmp1); 922 LLVM_DEBUG(std::prev(InsertPoint)->dump()); 923 924 BuildMI(*MI->getParent(), InsertPoint, MI->getDebugLoc(), 925 TII->get(PPC::COPY), DstReg) 926 .addReg(VSRCTmp2); 927 LLVM_DEBUG(std::prev(InsertPoint)->dump()); 928 929 } else { 930 insertSwap(MI, InsertPoint, DstReg, NewVReg); 931 LLVM_DEBUG(std::prev(InsertPoint)->dump()); 932 } 933 break; 934 } 935 } 936 } 937 938 // Walk the swap vector and replace each entry marked for removal with 939 // a copy operation. 940 bool PPCVSXSwapRemoval::removeSwaps() { 941 942 LLVM_DEBUG(dbgs() << "\n*** Removing swaps ***\n\n"); 943 944 bool Changed = false; 945 946 for (unsigned EntryIdx = 0; EntryIdx < SwapVector.size(); ++EntryIdx) { 947 if (SwapVector[EntryIdx].WillRemove) { 948 Changed = true; 949 MachineInstr *MI = SwapVector[EntryIdx].VSEMI; 950 MachineBasicBlock *MBB = MI->getParent(); 951 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(TargetOpcode::COPY), 952 MI->getOperand(0).getReg()) 953 .add(MI->getOperand(1)); 954 955 LLVM_DEBUG(dbgs() << format("Replaced %d with copy: ", 956 SwapVector[EntryIdx].VSEId)); 957 LLVM_DEBUG(MI->dump()); 958 959 MI->eraseFromParent(); 960 } 961 } 962 963 return Changed; 964 } 965 966 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 967 // For debug purposes, dump the contents of the swap vector. 968 LLVM_DUMP_METHOD void PPCVSXSwapRemoval::dumpSwapVector() { 969 970 for (unsigned EntryIdx = 0; EntryIdx < SwapVector.size(); ++EntryIdx) { 971 972 MachineInstr *MI = SwapVector[EntryIdx].VSEMI; 973 int ID = SwapVector[EntryIdx].VSEId; 974 975 dbgs() << format("%6d", ID); 976 dbgs() << format("%6d", EC->getLeaderValue(ID)); 977 dbgs() << format(" %bb.%3d", MI->getParent()->getNumber()); 978 dbgs() << format(" %14s ", TII->getName(MI->getOpcode()).str().c_str()); 979 980 if (SwapVector[EntryIdx].IsLoad) 981 dbgs() << "load "; 982 if (SwapVector[EntryIdx].IsStore) 983 dbgs() << "store "; 984 if (SwapVector[EntryIdx].IsSwap) 985 dbgs() << "swap "; 986 if (SwapVector[EntryIdx].MentionsPhysVR) 987 dbgs() << "physreg "; 988 if (SwapVector[EntryIdx].MentionsPartialVR) 989 dbgs() << "partialreg "; 990 991 if (SwapVector[EntryIdx].IsSwappable) { 992 dbgs() << "swappable "; 993 switch(SwapVector[EntryIdx].SpecialHandling) { 994 default: 995 dbgs() << "special:**unknown**"; 996 break; 997 case SH_NONE: 998 break; 999 case SH_EXTRACT: 1000 dbgs() << "special:extract "; 1001 break; 1002 case SH_INSERT: 1003 dbgs() << "special:insert "; 1004 break; 1005 case SH_NOSWAP_LD: 1006 dbgs() << "special:load "; 1007 break; 1008 case SH_NOSWAP_ST: 1009 dbgs() << "special:store "; 1010 break; 1011 case SH_SPLAT: 1012 dbgs() << "special:splat "; 1013 break; 1014 case SH_XXPERMDI: 1015 dbgs() << "special:xxpermdi "; 1016 break; 1017 case SH_COPYWIDEN: 1018 dbgs() << "special:copywiden "; 1019 break; 1020 } 1021 } 1022 1023 if (SwapVector[EntryIdx].WebRejected) 1024 dbgs() << "rejected "; 1025 if (SwapVector[EntryIdx].WillRemove) 1026 dbgs() << "remove "; 1027 1028 dbgs() << "\n"; 1029 1030 // For no-asserts builds. 1031 (void)MI; 1032 (void)ID; 1033 } 1034 1035 dbgs() << "\n"; 1036 } 1037 #endif 1038 1039 } // end default namespace 1040 1041 INITIALIZE_PASS_BEGIN(PPCVSXSwapRemoval, DEBUG_TYPE, 1042 "PowerPC VSX Swap Removal", false, false) 1043 INITIALIZE_PASS_END(PPCVSXSwapRemoval, DEBUG_TYPE, 1044 "PowerPC VSX Swap Removal", false, false) 1045 1046 char PPCVSXSwapRemoval::ID = 0; 1047 FunctionPass* 1048 llvm::createPPCVSXSwapRemovalPass() { return new PPCVSXSwapRemoval(); } 1049