xref: /freebsd/contrib/llvm-project/llvm/lib/Target/PowerPC/PPCVSXSwapRemoval.cpp (revision c66ec88fed842fbaad62c30d510644ceb7bd2d71)
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