LLVM  15.0.0git
PPCVSXSwapRemoval.cpp
Go to the documentation of this file.
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"
53 #include "llvm/Config/llvm-config.h"
54 #include "llvm/Support/Debug.h"
55 #include "llvm/Support/Format.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;
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.
109 
110  // Equivalence classes are used to gather webs of related computation.
111  // Swap entries are represented by their VSEId fields.
113 
114  PPCVSXSwapRemoval() : MachineFunctionPass(ID) {
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) {
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.
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  // All operands need to be checked because there are instructions that
258  // operate on a partial register and produce a full register (such as
259  // XXPERMDIs).
260  if (isAnyVecReg(Reg, Partial))
261  RelevantInstr = true;
262  }
263 
264  if (!RelevantInstr)
265  continue;
266 
267  RelevantFunction = true;
268 
269  // Create a SwapEntry initialized to zeros, then fill in the
270  // instruction and ID fields before pushing it to the back
271  // of the swap vector.
272  PPCVSXSwapEntry SwapEntry{};
273  int VecIdx = addSwapEntry(&MI, SwapEntry);
274 
275  switch(MI.getOpcode()) {
276  default:
277  // Unless noted otherwise, an instruction is considered
278  // safe for the optimization. There are a large number of
279  // such true-SIMD instructions (all vector math, logical,
280  // select, compare, etc.). However, if the instruction
281  // mentions a partial vector register and does not have
282  // special handling defined, it is not swappable.
283  if (Partial)
284  SwapVector[VecIdx].MentionsPartialVR = 1;
285  else
286  SwapVector[VecIdx].IsSwappable = 1;
287  break;
288  case PPC::XXPERMDI: {
289  // This is a swap if it is of the form XXPERMDI t, s, s, 2.
290  // Unfortunately, MachineCSE ignores COPY and SUBREG_TO_REG, so we
291  // can also see XXPERMDI t, SUBREG_TO_REG(s), SUBREG_TO_REG(s), 2,
292  // for example. We have to look through chains of COPY and
293  // SUBREG_TO_REG to find the real source value for comparison.
294  // If the real source value is a physical register, then mark the
295  // XXPERMDI as mentioning a physical register.
296  int immed = MI.getOperand(3).getImm();
297  if (immed == 2) {
298  unsigned trueReg1 = lookThruCopyLike(MI.getOperand(1).getReg(),
299  VecIdx);
300  unsigned trueReg2 = lookThruCopyLike(MI.getOperand(2).getReg(),
301  VecIdx);
302  if (trueReg1 == trueReg2)
303  SwapVector[VecIdx].IsSwap = 1;
304  else {
305  // We can still handle these if the two registers are not
306  // identical, by adjusting the form of the XXPERMDI.
307  SwapVector[VecIdx].IsSwappable = 1;
308  SwapVector[VecIdx].SpecialHandling = SHValues::SH_XXPERMDI;
309  }
310  // This is a doubleword splat if it is of the form
311  // XXPERMDI t, s, s, 0 or XXPERMDI t, s, s, 3. As above we
312  // must look through chains of copy-likes to find the source
313  // register. We turn off the marking for mention of a physical
314  // register, because splatting it is safe; the optimization
315  // will not swap the value in the physical register. Whether
316  // or not the two input registers are identical, we can handle
317  // these by adjusting the form of the XXPERMDI.
318  } else if (immed == 0 || immed == 3) {
319 
320  SwapVector[VecIdx].IsSwappable = 1;
321  SwapVector[VecIdx].SpecialHandling = SHValues::SH_XXPERMDI;
322 
323  unsigned trueReg1 = lookThruCopyLike(MI.getOperand(1).getReg(),
324  VecIdx);
325  unsigned trueReg2 = lookThruCopyLike(MI.getOperand(2).getReg(),
326  VecIdx);
327  if (trueReg1 == trueReg2)
328  SwapVector[VecIdx].MentionsPhysVR = 0;
329 
330  } else {
331  // We can still handle these by adjusting the form of the XXPERMDI.
332  SwapVector[VecIdx].IsSwappable = 1;
333  SwapVector[VecIdx].SpecialHandling = SHValues::SH_XXPERMDI;
334  }
335  break;
336  }
337  case PPC::LVX:
338  // Non-permuting loads are currently unsafe. We can use special
339  // handling for this in the future. By not marking these as
340  // IsSwap, we ensure computations containing them will be rejected
341  // for now.
342  SwapVector[VecIdx].IsLoad = 1;
343  break;
344  case PPC::LXVD2X:
345  case PPC::LXVW4X:
346  // Permuting loads are marked as both load and swap, and are
347  // safe for optimization.
348  SwapVector[VecIdx].IsLoad = 1;
349  SwapVector[VecIdx].IsSwap = 1;
350  break;
351  case PPC::LXSDX:
352  case PPC::LXSSPX:
353  case PPC::XFLOADf64:
354  case PPC::XFLOADf32:
355  // A load of a floating-point value into the high-order half of
356  // a vector register is safe, provided that we introduce a swap
357  // following the load, which will be done by the SUBREG_TO_REG
358  // support. So just mark these as safe.
359  SwapVector[VecIdx].IsLoad = 1;
360  SwapVector[VecIdx].IsSwappable = 1;
361  break;
362  case PPC::STVX:
363  // Non-permuting stores are currently unsafe. We can use special
364  // handling for this in the future. By not marking these as
365  // IsSwap, we ensure computations containing them will be rejected
366  // for now.
367  SwapVector[VecIdx].IsStore = 1;
368  break;
369  case PPC::STXVD2X:
370  case PPC::STXVW4X:
371  // Permuting stores are marked as both store and swap, and are
372  // safe for optimization.
373  SwapVector[VecIdx].IsStore = 1;
374  SwapVector[VecIdx].IsSwap = 1;
375  break;
376  case PPC::COPY:
377  // These are fine provided they are moving between full vector
378  // register classes.
379  if (isVecReg(MI.getOperand(0).getReg()) &&
380  isVecReg(MI.getOperand(1).getReg()))
381  SwapVector[VecIdx].IsSwappable = 1;
382  // If we have a copy from one scalar floating-point register
383  // to another, we can accept this even if it is a physical
384  // register. The only way this gets involved is if it feeds
385  // a SUBREG_TO_REG, which is handled by introducing a swap.
386  else if (isScalarVecReg(MI.getOperand(0).getReg()) &&
387  isScalarVecReg(MI.getOperand(1).getReg()))
388  SwapVector[VecIdx].IsSwappable = 1;
389  break;
390  case PPC::SUBREG_TO_REG: {
391  // These are fine provided they are moving between full vector
392  // register classes. If they are moving from a scalar
393  // floating-point class to a vector class, we can handle those
394  // as well, provided we introduce a swap. It is generally the
395  // case that we will introduce fewer swaps than we remove, but
396  // (FIXME) a cost model could be used. However, introduced
397  // swaps could potentially be CSEd, so this is not trivial.
398  if (isVecReg(MI.getOperand(0).getReg()) &&
399  isVecReg(MI.getOperand(2).getReg()))
400  SwapVector[VecIdx].IsSwappable = 1;
401  else if (isVecReg(MI.getOperand(0).getReg()) &&
402  isScalarVecReg(MI.getOperand(2).getReg())) {
403  SwapVector[VecIdx].IsSwappable = 1;
404  SwapVector[VecIdx].SpecialHandling = SHValues::SH_COPYWIDEN;
405  }
406  break;
407  }
408  case PPC::VSPLTB:
409  case PPC::VSPLTH:
410  case PPC::VSPLTW:
411  case PPC::XXSPLTW:
412  // Splats are lane-sensitive, but we can use special handling
413  // to adjust the source lane for the splat.
414  SwapVector[VecIdx].IsSwappable = 1;
415  SwapVector[VecIdx].SpecialHandling = SHValues::SH_SPLAT;
416  break;
417  // The presence of the following lane-sensitive operations in a
418  // web will kill the optimization, at least for now. For these
419  // we do nothing, causing the optimization to fail.
420  // FIXME: Some of these could be permitted with special handling,
421  // and will be phased in as time permits.
422  // FIXME: There is no simple and maintainable way to express a set
423  // of opcodes having a common attribute in TableGen. Should this
424  // change, this is a prime candidate to use such a mechanism.
425  case PPC::INLINEASM:
426  case PPC::INLINEASM_BR:
427  case PPC::EXTRACT_SUBREG:
428  case PPC::INSERT_SUBREG:
429  case PPC::COPY_TO_REGCLASS:
430  case PPC::LVEBX:
431  case PPC::LVEHX:
432  case PPC::LVEWX:
433  case PPC::LVSL:
434  case PPC::LVSR:
435  case PPC::LVXL:
436  case PPC::STVEBX:
437  case PPC::STVEHX:
438  case PPC::STVEWX:
439  case PPC::STVXL:
440  // We can handle STXSDX and STXSSPX similarly to LXSDX and LXSSPX,
441  // by adding special handling for narrowing copies as well as
442  // widening ones. However, I've experimented with this, and in
443  // practice we currently do not appear to use STXSDX fed by
444  // a narrowing copy from a full vector register. Since I can't
445  // generate any useful test cases, I've left this alone for now.
446  case PPC::STXSDX:
447  case PPC::STXSSPX:
448  case PPC::VCIPHER:
449  case PPC::VCIPHERLAST:
450  case PPC::VMRGHB:
451  case PPC::VMRGHH:
452  case PPC::VMRGHW:
453  case PPC::VMRGLB:
454  case PPC::VMRGLH:
455  case PPC::VMRGLW:
456  case PPC::VMULESB:
457  case PPC::VMULESH:
458  case PPC::VMULESW:
459  case PPC::VMULEUB:
460  case PPC::VMULEUH:
461  case PPC::VMULEUW:
462  case PPC::VMULOSB:
463  case PPC::VMULOSH:
464  case PPC::VMULOSW:
465  case PPC::VMULOUB:
466  case PPC::VMULOUH:
467  case PPC::VMULOUW:
468  case PPC::VNCIPHER:
469  case PPC::VNCIPHERLAST:
470  case PPC::VPERM:
471  case PPC::VPERMXOR:
472  case PPC::VPKPX:
473  case PPC::VPKSHSS:
474  case PPC::VPKSHUS:
475  case PPC::VPKSDSS:
476  case PPC::VPKSDUS:
477  case PPC::VPKSWSS:
478  case PPC::VPKSWUS:
479  case PPC::VPKUDUM:
480  case PPC::VPKUDUS:
481  case PPC::VPKUHUM:
482  case PPC::VPKUHUS:
483  case PPC::VPKUWUM:
484  case PPC::VPKUWUS:
485  case PPC::VPMSUMB:
486  case PPC::VPMSUMD:
487  case PPC::VPMSUMH:
488  case PPC::VPMSUMW:
489  case PPC::VRLB:
490  case PPC::VRLD:
491  case PPC::VRLH:
492  case PPC::VRLW:
493  case PPC::VSBOX:
494  case PPC::VSHASIGMAD:
495  case PPC::VSHASIGMAW:
496  case PPC::VSL:
497  case PPC::VSLDOI:
498  case PPC::VSLO:
499  case PPC::VSR:
500  case PPC::VSRO:
501  case PPC::VSUM2SWS:
502  case PPC::VSUM4SBS:
503  case PPC::VSUM4SHS:
504  case PPC::VSUM4UBS:
505  case PPC::VSUMSWS:
506  case PPC::VUPKHPX:
507  case PPC::VUPKHSB:
508  case PPC::VUPKHSH:
509  case PPC::VUPKHSW:
510  case PPC::VUPKLPX:
511  case PPC::VUPKLSB:
512  case PPC::VUPKLSH:
513  case PPC::VUPKLSW:
514  case PPC::XXMRGHW:
515  case PPC::XXMRGLW:
516  // XXSLDWI could be replaced by a general permute with one of three
517  // permute control vectors (for shift values 1, 2, 3). However,
518  // VPERM has a more restrictive register class.
519  case PPC::XXSLDWI:
520  case PPC::XSCVDPSPN:
521  case PPC::XSCVSPDPN:
522  break;
523  }
524  }
525  }
526 
527  if (RelevantFunction) {
528  LLVM_DEBUG(dbgs() << "Swap vector when first built\n\n");
529  LLVM_DEBUG(dumpSwapVector());
530  }
531 
532  return RelevantFunction;
533 }
534 
535 // Add an entry to the swap vector and swap map, and make a
536 // singleton equivalence class for the entry.
537 int PPCVSXSwapRemoval::addSwapEntry(MachineInstr *MI,
538  PPCVSXSwapEntry& SwapEntry) {
539  SwapEntry.VSEMI = MI;
540  SwapEntry.VSEId = SwapVector.size();
541  SwapVector.push_back(SwapEntry);
542  EC->insert(SwapEntry.VSEId);
543  SwapMap[MI] = SwapEntry.VSEId;
544  return SwapEntry.VSEId;
545 }
546 
547 // This is used to find the "true" source register for an
548 // XXPERMDI instruction, since MachineCSE does not handle the
549 // "copy-like" operations (Copy and SubregToReg). Returns
550 // the original SrcReg unless it is the target of a copy-like
551 // operation, in which case we chain backwards through all
552 // such operations to the ultimate source register. If a
553 // physical register is encountered, we stop the search and
554 // flag the swap entry indicated by VecIdx (the original
555 // XXPERMDI) as mentioning a physical register.
556 unsigned PPCVSXSwapRemoval::lookThruCopyLike(unsigned SrcReg,
557  unsigned VecIdx) {
558  MachineInstr *MI = MRI->getVRegDef(SrcReg);
559  if (!MI->isCopyLike())
560  return SrcReg;
561 
562  unsigned CopySrcReg;
563  if (MI->isCopy())
564  CopySrcReg = MI->getOperand(1).getReg();
565  else {
566  assert(MI->isSubregToReg() && "bad opcode for lookThruCopyLike");
567  CopySrcReg = MI->getOperand(2).getReg();
568  }
569 
570  if (!Register::isVirtualRegister(CopySrcReg)) {
571  if (!isScalarVecReg(CopySrcReg))
572  SwapVector[VecIdx].MentionsPhysVR = 1;
573  return CopySrcReg;
574  }
575 
576  return lookThruCopyLike(CopySrcReg, VecIdx);
577 }
578 
579 // Generate equivalence classes for related computations (webs) by
580 // def-use relationships of virtual registers. Mention of a physical
581 // register terminates the generation of equivalence classes as this
582 // indicates a use of a parameter, definition of a return value, use
583 // of a value returned from a call, or definition of a parameter to a
584 // call. Computations with physical register mentions are flagged
585 // as such so their containing webs will not be optimized.
586 void PPCVSXSwapRemoval::formWebs() {
587 
588  LLVM_DEBUG(dbgs() << "\n*** Forming webs for swap removal ***\n\n");
589 
590  for (unsigned EntryIdx = 0; EntryIdx < SwapVector.size(); ++EntryIdx) {
591 
592  MachineInstr *MI = SwapVector[EntryIdx].VSEMI;
593 
594  LLVM_DEBUG(dbgs() << "\n" << SwapVector[EntryIdx].VSEId << " ");
595  LLVM_DEBUG(MI->dump());
596 
597  // It's sufficient to walk vector uses and join them to their unique
598  // definitions. In addition, check full vector register operands
599  // for physical regs. We exclude partial-vector register operands
600  // because we can handle them if copied to a full vector.
601  for (const MachineOperand &MO : MI->operands()) {
602  if (!MO.isReg())
603  continue;
604 
605  Register Reg = MO.getReg();
606  if (!isVecReg(Reg) && !isScalarVecReg(Reg))
607  continue;
608 
610  if (!(MI->isCopy() && isScalarVecReg(Reg)))
611  SwapVector[EntryIdx].MentionsPhysVR = 1;
612  continue;
613  }
614 
615  if (!MO.isUse())
616  continue;
617 
619  assert(SwapMap.find(DefMI) != SwapMap.end() &&
620  "Inconsistency: def of vector reg not found in swap map!");
621  int DefIdx = SwapMap[DefMI];
622  (void)EC->unionSets(SwapVector[DefIdx].VSEId,
623  SwapVector[EntryIdx].VSEId);
624 
625  LLVM_DEBUG(dbgs() << format("Unioning %d with %d\n",
626  SwapVector[DefIdx].VSEId,
627  SwapVector[EntryIdx].VSEId));
628  LLVM_DEBUG(dbgs() << " Def: ");
629  LLVM_DEBUG(DefMI->dump());
630  }
631  }
632 }
633 
634 // Walk the swap vector entries looking for conditions that prevent their
635 // containing computations from being optimized. When such conditions are
636 // found, mark the representative of the computation's equivalence class
637 // as rejected.
638 void PPCVSXSwapRemoval::recordUnoptimizableWebs() {
639 
640  LLVM_DEBUG(dbgs() << "\n*** Rejecting webs for swap removal ***\n\n");
641 
642  for (unsigned EntryIdx = 0; EntryIdx < SwapVector.size(); ++EntryIdx) {
643  int Repr = EC->getLeaderValue(SwapVector[EntryIdx].VSEId);
644 
645  // If representative is already rejected, don't waste further time.
646  if (SwapVector[Repr].WebRejected)
647  continue;
648 
649  // Reject webs containing mentions of physical or partial registers, or
650  // containing operations that we don't know how to handle in a lane-
651  // permuted region.
652  if (SwapVector[EntryIdx].MentionsPhysVR ||
653  SwapVector[EntryIdx].MentionsPartialVR ||
654  !(SwapVector[EntryIdx].IsSwappable || SwapVector[EntryIdx].IsSwap)) {
655 
656  SwapVector[Repr].WebRejected = 1;
657 
658  LLVM_DEBUG(
659  dbgs() << format("Web %d rejected for physreg, partial reg, or not "
660  "swap[pable]\n",
661  Repr));
662  LLVM_DEBUG(dbgs() << " in " << EntryIdx << ": ");
663  LLVM_DEBUG(SwapVector[EntryIdx].VSEMI->dump());
664  LLVM_DEBUG(dbgs() << "\n");
665  }
666 
667  // Reject webs than contain swapping loads that feed something other
668  // than a swap instruction.
669  else if (SwapVector[EntryIdx].IsLoad && SwapVector[EntryIdx].IsSwap) {
670  MachineInstr *MI = SwapVector[EntryIdx].VSEMI;
671  Register DefReg = MI->getOperand(0).getReg();
672 
673  // We skip debug instructions in the analysis. (Note that debug
674  // location information is still maintained by this optimization
675  // because it remains on the LXVD2X and STXVD2X instructions after
676  // the XXPERMDIs are removed.)
677  for (MachineInstr &UseMI : MRI->use_nodbg_instructions(DefReg)) {
678  int UseIdx = SwapMap[&UseMI];
679 
680  if (!SwapVector[UseIdx].IsSwap || SwapVector[UseIdx].IsLoad ||
681  SwapVector[UseIdx].IsStore) {
682 
683  SwapVector[Repr].WebRejected = 1;
684 
685  LLVM_DEBUG(dbgs() << format(
686  "Web %d rejected for load not feeding swap\n", Repr));
687  LLVM_DEBUG(dbgs() << " def " << EntryIdx << ": ");
688  LLVM_DEBUG(MI->dump());
689  LLVM_DEBUG(dbgs() << " use " << UseIdx << ": ");
690  LLVM_DEBUG(UseMI.dump());
691  LLVM_DEBUG(dbgs() << "\n");
692  }
693 
694  // It is possible that the load feeds a swap and that swap feeds a
695  // store. In such a case, the code is actually trying to store a swapped
696  // vector. We must reject such webs.
697  if (SwapVector[UseIdx].IsSwap && !SwapVector[UseIdx].IsLoad &&
698  !SwapVector[UseIdx].IsStore) {
699  Register SwapDefReg = UseMI.getOperand(0).getReg();
700  for (MachineInstr &UseOfUseMI :
701  MRI->use_nodbg_instructions(SwapDefReg)) {
702  int UseOfUseIdx = SwapMap[&UseOfUseMI];
703  if (SwapVector[UseOfUseIdx].IsStore) {
704  SwapVector[Repr].WebRejected = 1;
705  LLVM_DEBUG(
706  dbgs() << format(
707  "Web %d rejected for load/swap feeding a store\n", Repr));
708  LLVM_DEBUG(dbgs() << " def " << EntryIdx << ": ");
709  LLVM_DEBUG(MI->dump());
710  LLVM_DEBUG(dbgs() << " use " << UseIdx << ": ");
711  LLVM_DEBUG(UseMI.dump());
712  LLVM_DEBUG(dbgs() << "\n");
713  }
714  }
715  }
716  }
717 
718  // Reject webs that contain swapping stores that are fed by something
719  // other than a swap instruction.
720  } else if (SwapVector[EntryIdx].IsStore && SwapVector[EntryIdx].IsSwap) {
721  MachineInstr *MI = SwapVector[EntryIdx].VSEMI;
722  Register UseReg = MI->getOperand(0).getReg();
724  Register DefReg = DefMI->getOperand(0).getReg();
725  int DefIdx = SwapMap[DefMI];
726 
727  if (!SwapVector[DefIdx].IsSwap || SwapVector[DefIdx].IsLoad ||
728  SwapVector[DefIdx].IsStore) {
729 
730  SwapVector[Repr].WebRejected = 1;
731 
732  LLVM_DEBUG(dbgs() << format(
733  "Web %d rejected for store not fed by swap\n", Repr));
734  LLVM_DEBUG(dbgs() << " def " << DefIdx << ": ");
735  LLVM_DEBUG(DefMI->dump());
736  LLVM_DEBUG(dbgs() << " use " << EntryIdx << ": ");
737  LLVM_DEBUG(MI->dump());
738  LLVM_DEBUG(dbgs() << "\n");
739  }
740 
741  // Ensure all uses of the register defined by DefMI feed store
742  // instructions
743  for (MachineInstr &UseMI : MRI->use_nodbg_instructions(DefReg)) {
744  int UseIdx = SwapMap[&UseMI];
745 
746  if (SwapVector[UseIdx].VSEMI->getOpcode() != MI->getOpcode()) {
747  SwapVector[Repr].WebRejected = 1;
748 
749  LLVM_DEBUG(
750  dbgs() << format(
751  "Web %d rejected for swap not feeding only stores\n", Repr));
752  LLVM_DEBUG(dbgs() << " def "
753  << " : ");
754  LLVM_DEBUG(DefMI->dump());
755  LLVM_DEBUG(dbgs() << " use " << UseIdx << ": ");
756  LLVM_DEBUG(SwapVector[UseIdx].VSEMI->dump());
757  LLVM_DEBUG(dbgs() << "\n");
758  }
759  }
760  }
761  }
762 
763  LLVM_DEBUG(dbgs() << "Swap vector after web analysis:\n\n");
764  LLVM_DEBUG(dumpSwapVector());
765 }
766 
767 // Walk the swap vector entries looking for swaps fed by permuting loads
768 // and swaps that feed permuting stores. If the containing computation
769 // has not been marked rejected, mark each such swap for removal.
770 // (Removal is delayed in case optimization has disturbed the pattern,
771 // such that multiple loads feed the same swap, etc.)
772 void PPCVSXSwapRemoval::markSwapsForRemoval() {
773 
774  LLVM_DEBUG(dbgs() << "\n*** Marking swaps for removal ***\n\n");
775 
776  for (unsigned EntryIdx = 0; EntryIdx < SwapVector.size(); ++EntryIdx) {
777 
778  if (SwapVector[EntryIdx].IsLoad && SwapVector[EntryIdx].IsSwap) {
779  int Repr = EC->getLeaderValue(SwapVector[EntryIdx].VSEId);
780 
781  if (!SwapVector[Repr].WebRejected) {
782  MachineInstr *MI = SwapVector[EntryIdx].VSEMI;
783  Register DefReg = MI->getOperand(0).getReg();
784 
785  for (MachineInstr &UseMI : MRI->use_nodbg_instructions(DefReg)) {
786  int UseIdx = SwapMap[&UseMI];
787  SwapVector[UseIdx].WillRemove = 1;
788 
789  LLVM_DEBUG(dbgs() << "Marking swap fed by load for removal: ");
790  LLVM_DEBUG(UseMI.dump());
791  }
792  }
793 
794  } else if (SwapVector[EntryIdx].IsStore && SwapVector[EntryIdx].IsSwap) {
795  int Repr = EC->getLeaderValue(SwapVector[EntryIdx].VSEId);
796 
797  if (!SwapVector[Repr].WebRejected) {
798  MachineInstr *MI = SwapVector[EntryIdx].VSEMI;
799  Register UseReg = MI->getOperand(0).getReg();
801  int DefIdx = SwapMap[DefMI];
802  SwapVector[DefIdx].WillRemove = 1;
803 
804  LLVM_DEBUG(dbgs() << "Marking swap feeding store for removal: ");
805  LLVM_DEBUG(DefMI->dump());
806  }
807 
808  } else if (SwapVector[EntryIdx].IsSwappable &&
809  SwapVector[EntryIdx].SpecialHandling != 0) {
810  int Repr = EC->getLeaderValue(SwapVector[EntryIdx].VSEId);
811 
812  if (!SwapVector[Repr].WebRejected)
813  handleSpecialSwappables(EntryIdx);
814  }
815  }
816 }
817 
818 // Create an xxswapd instruction and insert it prior to the given point.
819 // MI is used to determine basic block and debug loc information.
820 // FIXME: When inserting a swap, we should check whether SrcReg is
821 // defined by another swap: SrcReg = XXPERMDI Reg, Reg, 2; If so,
822 // then instead we should generate a copy from Reg to DstReg.
823 void PPCVSXSwapRemoval::insertSwap(MachineInstr *MI,
824  MachineBasicBlock::iterator InsertPoint,
825  unsigned DstReg, unsigned SrcReg) {
826  BuildMI(*MI->getParent(), InsertPoint, MI->getDebugLoc(),
827  TII->get(PPC::XXPERMDI), DstReg)
828  .addReg(SrcReg)
829  .addReg(SrcReg)
830  .addImm(2);
831 }
832 
833 // The identified swap entry requires special handling to allow its
834 // containing computation to be optimized. Perform that handling
835 // here.
836 // FIXME: Additional opportunities will be phased in with subsequent
837 // patches.
838 void PPCVSXSwapRemoval::handleSpecialSwappables(int EntryIdx) {
839  switch (SwapVector[EntryIdx].SpecialHandling) {
840 
841  default:
842  llvm_unreachable("Unexpected special handling type");
843 
844  // For splats based on an index into a vector, add N/2 modulo N
845  // to the index, where N is the number of vector elements.
846  case SHValues::SH_SPLAT: {
847  MachineInstr *MI = SwapVector[EntryIdx].VSEMI;
848  unsigned NElts;
849 
850  LLVM_DEBUG(dbgs() << "Changing splat: ");
851  LLVM_DEBUG(MI->dump());
852 
853  switch (MI->getOpcode()) {
854  default:
855  llvm_unreachable("Unexpected splat opcode");
856  case PPC::VSPLTB: NElts = 16; break;
857  case PPC::VSPLTH: NElts = 8; break;
858  case PPC::VSPLTW:
859  case PPC::XXSPLTW: NElts = 4; break;
860  }
861 
862  unsigned EltNo;
863  if (MI->getOpcode() == PPC::XXSPLTW)
864  EltNo = MI->getOperand(2).getImm();
865  else
866  EltNo = MI->getOperand(1).getImm();
867 
868  EltNo = (EltNo + NElts / 2) % NElts;
869  if (MI->getOpcode() == PPC::XXSPLTW)
870  MI->getOperand(2).setImm(EltNo);
871  else
872  MI->getOperand(1).setImm(EltNo);
873 
874  LLVM_DEBUG(dbgs() << " Into: ");
875  LLVM_DEBUG(MI->dump());
876  break;
877  }
878 
879  // For an XXPERMDI that isn't handled otherwise, we need to
880  // reverse the order of the operands. If the selector operand
881  // has a value of 0 or 3, we need to change it to 3 or 0,
882  // respectively. Otherwise we should leave it alone. (This
883  // is equivalent to reversing the two bits of the selector
884  // operand and complementing the result.)
885  case SHValues::SH_XXPERMDI: {
886  MachineInstr *MI = SwapVector[EntryIdx].VSEMI;
887 
888  LLVM_DEBUG(dbgs() << "Changing XXPERMDI: ");
889  LLVM_DEBUG(MI->dump());
890 
891  unsigned Selector = MI->getOperand(3).getImm();
892  if (Selector == 0 || Selector == 3)
893  Selector = 3 - Selector;
894  MI->getOperand(3).setImm(Selector);
895 
896  Register Reg1 = MI->getOperand(1).getReg();
897  Register Reg2 = MI->getOperand(2).getReg();
898  MI->getOperand(1).setReg(Reg2);
899  MI->getOperand(2).setReg(Reg1);
900 
901  // We also need to swap kill flag associated with the register.
902  bool IsKill1 = MI->getOperand(1).isKill();
903  bool IsKill2 = MI->getOperand(2).isKill();
904  MI->getOperand(1).setIsKill(IsKill2);
905  MI->getOperand(2).setIsKill(IsKill1);
906 
907  LLVM_DEBUG(dbgs() << " Into: ");
908  LLVM_DEBUG(MI->dump());
909  break;
910  }
911 
912  // For a copy from a scalar floating-point register to a vector
913  // register, removing swaps will leave the copied value in the
914  // wrong lane. Insert a swap following the copy to fix this.
915  case SHValues::SH_COPYWIDEN: {
916  MachineInstr *MI = SwapVector[EntryIdx].VSEMI;
917 
918  LLVM_DEBUG(dbgs() << "Changing SUBREG_TO_REG: ");
919  LLVM_DEBUG(MI->dump());
920 
921  Register DstReg = MI->getOperand(0).getReg();
922  const TargetRegisterClass *DstRC = MRI->getRegClass(DstReg);
923  Register NewVReg = MRI->createVirtualRegister(DstRC);
924 
925  MI->getOperand(0).setReg(NewVReg);
926  LLVM_DEBUG(dbgs() << " Into: ");
927  LLVM_DEBUG(MI->dump());
928 
929  auto InsertPoint = ++MachineBasicBlock::iterator(MI);
930 
931  // Note that an XXPERMDI requires a VSRC, so if the SUBREG_TO_REG
932  // is copying to a VRRC, we need to be careful to avoid a register
933  // assignment problem. In this case we must copy from VRRC to VSRC
934  // prior to the swap, and from VSRC to VRRC following the swap.
935  // Coalescing will usually remove all this mess.
936  if (DstRC == &PPC::VRRCRegClass) {
937  Register VSRCTmp1 = MRI->createVirtualRegister(&PPC::VSRCRegClass);
938  Register VSRCTmp2 = MRI->createVirtualRegister(&PPC::VSRCRegClass);
939 
940  BuildMI(*MI->getParent(), InsertPoint, MI->getDebugLoc(),
941  TII->get(PPC::COPY), VSRCTmp1)
942  .addReg(NewVReg);
943  LLVM_DEBUG(std::prev(InsertPoint)->dump());
944 
945  insertSwap(MI, InsertPoint, VSRCTmp2, VSRCTmp1);
946  LLVM_DEBUG(std::prev(InsertPoint)->dump());
947 
948  BuildMI(*MI->getParent(), InsertPoint, MI->getDebugLoc(),
949  TII->get(PPC::COPY), DstReg)
950  .addReg(VSRCTmp2);
951  LLVM_DEBUG(std::prev(InsertPoint)->dump());
952 
953  } else {
954  insertSwap(MI, InsertPoint, DstReg, NewVReg);
955  LLVM_DEBUG(std::prev(InsertPoint)->dump());
956  }
957  break;
958  }
959  }
960 }
961 
962 // Walk the swap vector and replace each entry marked for removal with
963 // a copy operation.
964 bool PPCVSXSwapRemoval::removeSwaps() {
965 
966  LLVM_DEBUG(dbgs() << "\n*** Removing swaps ***\n\n");
967 
968  bool Changed = false;
969 
970  for (unsigned EntryIdx = 0; EntryIdx < SwapVector.size(); ++EntryIdx) {
971  if (SwapVector[EntryIdx].WillRemove) {
972  Changed = true;
973  MachineInstr *MI = SwapVector[EntryIdx].VSEMI;
974  MachineBasicBlock *MBB = MI->getParent();
975  BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(TargetOpcode::COPY),
976  MI->getOperand(0).getReg())
977  .add(MI->getOperand(1));
978 
979  LLVM_DEBUG(dbgs() << format("Replaced %d with copy: ",
980  SwapVector[EntryIdx].VSEId));
981  LLVM_DEBUG(MI->dump());
982 
983  MI->eraseFromParent();
984  }
985  }
986 
987  return Changed;
988 }
989 
990 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
991 // For debug purposes, dump the contents of the swap vector.
992 LLVM_DUMP_METHOD void PPCVSXSwapRemoval::dumpSwapVector() {
993 
994  for (unsigned EntryIdx = 0; EntryIdx < SwapVector.size(); ++EntryIdx) {
995 
996  MachineInstr *MI = SwapVector[EntryIdx].VSEMI;
997  int ID = SwapVector[EntryIdx].VSEId;
998 
999  dbgs() << format("%6d", ID);
1000  dbgs() << format("%6d", EC->getLeaderValue(ID));
1001  dbgs() << format(" %bb.%3d", MI->getParent()->getNumber());
1002  dbgs() << format(" %14s ", TII->getName(MI->getOpcode()).str().c_str());
1003 
1004  if (SwapVector[EntryIdx].IsLoad)
1005  dbgs() << "load ";
1006  if (SwapVector[EntryIdx].IsStore)
1007  dbgs() << "store ";
1008  if (SwapVector[EntryIdx].IsSwap)
1009  dbgs() << "swap ";
1010  if (SwapVector[EntryIdx].MentionsPhysVR)
1011  dbgs() << "physreg ";
1012  if (SwapVector[EntryIdx].MentionsPartialVR)
1013  dbgs() << "partialreg ";
1014 
1015  if (SwapVector[EntryIdx].IsSwappable) {
1016  dbgs() << "swappable ";
1017  switch(SwapVector[EntryIdx].SpecialHandling) {
1018  default:
1019  dbgs() << "special:**unknown**";
1020  break;
1021  case SH_NONE:
1022  break;
1023  case SH_EXTRACT:
1024  dbgs() << "special:extract ";
1025  break;
1026  case SH_INSERT:
1027  dbgs() << "special:insert ";
1028  break;
1029  case SH_NOSWAP_LD:
1030  dbgs() << "special:load ";
1031  break;
1032  case SH_NOSWAP_ST:
1033  dbgs() << "special:store ";
1034  break;
1035  case SH_SPLAT:
1036  dbgs() << "special:splat ";
1037  break;
1038  case SH_XXPERMDI:
1039  dbgs() << "special:xxpermdi ";
1040  break;
1041  case SH_COPYWIDEN:
1042  dbgs() << "special:copywiden ";
1043  break;
1044  }
1045  }
1046 
1047  if (SwapVector[EntryIdx].WebRejected)
1048  dbgs() << "rejected ";
1049  if (SwapVector[EntryIdx].WillRemove)
1050  dbgs() << "remove ";
1051 
1052  dbgs() << "\n";
1053 
1054  // For no-asserts builds.
1055  (void)MI;
1056  (void)ID;
1057  }
1058 
1059  dbgs() << "\n";
1060 }
1061 #endif
1062 
1063 } // end default namespace
1064 
1065 INITIALIZE_PASS_BEGIN(PPCVSXSwapRemoval, DEBUG_TYPE,
1066  "PowerPC VSX Swap Removal", false, false)
1067 INITIALIZE_PASS_END(PPCVSXSwapRemoval, DEBUG_TYPE,
1068  "PowerPC VSX Swap Removal", false, false)
1069 
1070 char PPCVSXSwapRemoval::ID = 0;
1071 FunctionPass*
1072 llvm::createPPCVSXSwapRemovalPass() { return new PPCVSXSwapRemoval(); }
MI
IRTranslator LLVM IR MI
Definition: IRTranslator.cpp:104
LLVM_DUMP_METHOD
#define LLVM_DUMP_METHOD
Mark debug helper function definitions like dump() that should not be stripped from debug builds.
Definition: Compiler.h:494
llvm::MachineInstrBuilder::addImm
const MachineInstrBuilder & addImm(int64_t Val) const
Add a new immediate operand.
Definition: MachineInstrBuilder.h:131
llvm
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:17
UseMI
MachineInstrBuilder & UseMI
Definition: AArch64ExpandPseudoInsts.cpp:103
llvm::MachineRegisterInfo::createVirtualRegister
Register createVirtualRegister(const TargetRegisterClass *RegClass, StringRef Name="")
createVirtualRegister - Create and return a new virtual register in the function with the specified r...
Definition: MachineRegisterInfo.cpp:156
llvm::MachineRegisterInfo
MachineRegisterInfo - Keep track of information for virtual and physical registers,...
Definition: MachineRegisterInfo.h:50
llvm::MachineInstrBuilder::add
const MachineInstrBuilder & add(const MachineOperand &MO) const
Definition: MachineInstrBuilder.h:224
Removal
PowerPC VSX Swap Removal
Definition: PPCVSXSwapRemoval.cpp:1068
PPCInstrBuilder.h
llvm::EquivalenceClasses
EquivalenceClasses - This represents a collection of equivalence classes and supports three efficient...
Definition: EquivalenceClasses.h:60
llvm::X86Disassembler::Reg
Reg
All possible values of the reg field in the ModR/M byte.
Definition: X86DisassemblerDecoder.h:462
llvm::MachineFunctionPass
MachineFunctionPass - This class adapts the FunctionPass interface to allow convenient creation of pa...
Definition: MachineFunctionPass.h:30
llvm::PPCSubtarget::hasVSX
bool hasVSX() const
Definition: PPCSubtarget.h:280
llvm::PPCInstrInfo
Definition: PPCInstrInfo.h:191
llvm::PPCISD::VPERM
@ VPERM
VPERM - The PPC VPERM Instruction.
Definition: PPCISelLowering.h:100
llvm::MachineRegisterInfo::use_nodbg_instructions
iterator_range< use_instr_nodbg_iterator > use_nodbg_instructions(Register Reg) const
Definition: MachineRegisterInfo.h:551
DenseMap.h
llvm::dump
void dump(const SparseBitVector< ElementSize > &LHS, raw_ostream &out)
Definition: SparseBitVector.h:877
initialize
static void initialize(TargetLibraryInfoImpl &TLI, const Triple &T, ArrayRef< StringLiteral > StandardNames)
Initialize the set of available library functions based on the specified target triple.
Definition: TargetLibraryInfo.cpp:116
Format.h
LLVM_DEBUG
#define LLVM_DEBUG(X)
Definition: Debug.h:101
MachineRegisterInfo.h
llvm::ISD::INLINEASM
@ INLINEASM
INLINEASM - Represents an inline asm block.
Definition: ISDOpcodes.h:1018
llvm::dbgs
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:163
llvm::PPCISD::STXVD2X
@ STXVD2X
CHAIN = STXVD2X CHAIN, VSRC, Ptr - Occurs only for little endian.
Definition: PPCISelLowering.h:581
llvm::TargetRegisterClass::contains
bool contains(Register Reg) const
Return true if the specified register is included in this register class.
Definition: TargetRegisterInfo.h:94
llvm::PassRegistry::getPassRegistry
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
Definition: PassRegistry.cpp:31
llvm::PPCSubtarget
Definition: PPCSubtarget.h:71
llvm::MachineInstr::getOperand
const MachineOperand & getOperand(unsigned i) const
Definition: MachineInstr.h:501
llvm::TargetRegisterClass
Definition: TargetRegisterInfo.h:45
false
Definition: StackSlotColoring.cpp:141
TII
const HexagonInstrInfo * TII
Definition: HexagonCopyToCombine.cpp:127
llvm::MachineOperand
MachineOperand class - Representation of each machine instruction operand.
Definition: MachineOperand.h:48
PPC.h
llvm::MachineRegisterInfo::getVRegDef
MachineInstr * getVRegDef(Register Reg) const
getVRegDef - Return the machine instr that defines the specified virtual register or null if none is ...
Definition: MachineRegisterInfo.cpp:396
llvm::CallingConv::ID
unsigned ID
LLVM IR allows to use arbitrary numbers as calling convention identifiers.
Definition: CallingConv.h:24
llvm::MachineInstrBuilder::getReg
Register getReg(unsigned Idx) const
Get the register for the operand index.
Definition: MachineInstrBuilder.h:94
llvm::MachineBasicBlock
Definition: MachineBasicBlock.h:94
INITIALIZE_PASS_END
#define INITIALIZE_PASS_END(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:58
llvm::MachineRegisterInfo::getRegClass
const TargetRegisterClass * getRegClass(Register Reg) const
Return the register class of the specified virtual register.
Definition: MachineRegisterInfo.h:642
llvm::MachineFunction::getSubtarget
const TargetSubtargetInfo & getSubtarget() const
getSubtarget - Return the subtarget for which this machine code is being compiled.
Definition: MachineFunction.h:640
llvm::MachineInstr
Representation of each machine instruction.
Definition: MachineInstr.h:66
UseReg
static Register UseReg(const MachineOperand &MO)
Definition: HexagonCopyToCombine.cpp:254
llvm::PPCSubtarget::needsSwapsForVSXMemOps
bool needsSwapsForVSXMemOps() const
Definition: PPCSubtarget.h:355
INITIALIZE_PASS_BEGIN
INITIALIZE_PASS_BEGIN(PPCVSXSwapRemoval, DEBUG_TYPE, "PowerPC VSX Swap Removal", false, false) INITIALIZE_PASS_END(PPCVSXSwapRemoval
llvm::DenseMap
Definition: DenseMap.h:716
PPCInstrInfo.h
MachineFunctionPass.h
llvm::Register::isVirtualRegister
static bool isVirtualRegister(unsigned Reg)
Return true if the specified register number is in the virtual register namespace.
Definition: Register.h:71
assert
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
llvm::MachineInstrBuilder::addReg
const MachineInstrBuilder & addReg(Register RegNo, unsigned flags=0, unsigned SubReg=0) const
Add a new virtual register operand.
Definition: MachineInstrBuilder.h:97
llvm::MachineOperand::getReg
Register getReg() const
getReg - Returns the register number.
Definition: MachineOperand.h:359
isVecReg
static bool isVecReg(unsigned Reg)
Definition: HexagonVectorPrint.cpp:73
isRegInClass
static bool isRegInClass(const MachineOperand &MO, const TargetRegisterClass *Class)
Definition: ARMLowOverheadLoops.cpp:816
llvm::MachineFunction
Definition: MachineFunction.h:241
llvm::MachineInstr::dump
void dump() const
Definition: MachineInstr.cpp:1491
llvm::MachineBasicBlock::iterator
MachineInstrBundleIterator< MachineInstr > iterator
Definition: MachineBasicBlock.h:242
DEBUG_TYPE
#define DEBUG_TYPE
Definition: PPCVSXSwapRemoval.cpp:60
llvm_unreachable
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
Definition: ErrorHandling.h:143
llvm::format
format_object< Ts... > format(const char *Fmt, const Ts &... Vals)
These are helper functions used to produce formatted output.
Definition: Format.h:124
llvm::Pass::dump
void dump() const
Definition: Pass.cpp:135
MRI
unsigned const MachineRegisterInfo * MRI
Definition: AArch64AdvSIMDScalarPass.cpp:105
llvm::Register
Wrapper class representing virtual and physical registers.
Definition: Register.h:19
llvm::PPCISD::LXVD2X
@ LXVD2X
VSRC, CHAIN = LXVD2X_LE CHAIN, Ptr - Occurs only for little endian.
Definition: PPCISelLowering.h:549
MBB
MachineBasicBlock & MBB
Definition: AArch64SLSHardening.cpp:74
llvm::createPPCVSXSwapRemovalPass
FunctionPass * createPPCVSXSwapRemovalPass()
Definition: PPCVSXSwapRemoval.cpp:1072
llvm::MachineFunction::getFunction
Function & getFunction()
Return the LLVM function that this machine code represents.
Definition: MachineFunction.h:606
llvm::ISD::INLINEASM_BR
@ INLINEASM_BR
INLINEASM_BR - Branching version of inline asm. Used by asm-goto.
Definition: ISDOpcodes.h:1021
EquivalenceClasses.h
MachineInstrBuilder.h
llvm::BuildMI
MachineInstrBuilder BuildMI(MachineFunction &MF, const DebugLoc &DL, const MCInstrDesc &MCID)
Builder interface. Specify how to create the initial instruction itself.
Definition: MachineInstrBuilder.h:328
DefMI
MachineInstrBuilder MachineInstrBuilder & DefMI
Definition: AArch64ExpandPseudoInsts.cpp:104
llvm::PPCISD::XXPERMDI
@ XXPERMDI
XXPERMDI - The PPC XXPERMDI instruction.
Definition: PPCISelLowering.h:125
llvm::FunctionPass
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:308
llvm::TargetRegisterClass::hasSubClassEq
bool hasSubClassEq(const TargetRegisterClass *RC) const
Returns true if RC is a sub-class of or equal to this class.
Definition: TargetRegisterInfo.h:127
raw_ostream.h
llvm::MachineInstrBundleIterator< MachineInstr >
llvm::initializePPCVSXSwapRemovalPass
void initializePPCVSXSwapRemovalPass(PassRegistry &)
Debug.h
PPCTargetMachine.h
llvm::Intrinsic::ID
unsigned ID
Definition: TargetTransformInfo.h:37