xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp (revision 43a5ec4eb41567cc92586503212743d89686d78f)
1 //===-- SILowerI1Copies.cpp - Lower I1 Copies -----------------------------===//
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 lowers all occurrences of i1 values (with a vreg_1 register class)
10 // to lane masks (32 / 64-bit scalar registers). The pass assumes machine SSA
11 // form and a wave-level control flow graph.
12 //
13 // Before this pass, values that are semantically i1 and are defined and used
14 // within the same basic block are already represented as lane masks in scalar
15 // registers. However, values that cross basic blocks are always transferred
16 // between basic blocks in vreg_1 virtual registers and are lowered by this
17 // pass.
18 //
19 // The only instructions that use or define vreg_1 virtual registers are COPY,
20 // PHI, and IMPLICIT_DEF.
21 //
22 //===----------------------------------------------------------------------===//
23 
24 #include "AMDGPU.h"
25 #include "GCNSubtarget.h"
26 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
27 #include "llvm/CodeGen/MachineDominators.h"
28 #include "llvm/CodeGen/MachineFunctionPass.h"
29 #include "llvm/CodeGen/MachinePostDominators.h"
30 #include "llvm/CodeGen/MachineSSAUpdater.h"
31 #include "llvm/InitializePasses.h"
32 
33 #define DEBUG_TYPE "si-i1-copies"
34 
35 using namespace llvm;
36 
37 static unsigned createLaneMaskReg(MachineFunction &MF);
38 static unsigned insertUndefLaneMask(MachineBasicBlock &MBB);
39 
40 namespace {
41 
42 class SILowerI1Copies : public MachineFunctionPass {
43 public:
44   static char ID;
45 
46 private:
47   bool IsWave32 = false;
48   MachineFunction *MF = nullptr;
49   MachineDominatorTree *DT = nullptr;
50   MachinePostDominatorTree *PDT = nullptr;
51   MachineRegisterInfo *MRI = nullptr;
52   const GCNSubtarget *ST = nullptr;
53   const SIInstrInfo *TII = nullptr;
54 
55   unsigned ExecReg;
56   unsigned MovOp;
57   unsigned AndOp;
58   unsigned OrOp;
59   unsigned XorOp;
60   unsigned AndN2Op;
61   unsigned OrN2Op;
62 
63   DenseSet<unsigned> ConstrainRegs;
64 
65 public:
66   SILowerI1Copies() : MachineFunctionPass(ID) {
67     initializeSILowerI1CopiesPass(*PassRegistry::getPassRegistry());
68   }
69 
70   bool runOnMachineFunction(MachineFunction &MF) override;
71 
72   StringRef getPassName() const override { return "SI Lower i1 Copies"; }
73 
74   void getAnalysisUsage(AnalysisUsage &AU) const override {
75     AU.setPreservesCFG();
76     AU.addRequired<MachineDominatorTree>();
77     AU.addRequired<MachinePostDominatorTree>();
78     MachineFunctionPass::getAnalysisUsage(AU);
79   }
80 
81 private:
82   void lowerCopiesFromI1();
83   void lowerPhis();
84   void lowerCopiesToI1();
85   bool isConstantLaneMask(Register Reg, bool &Val) const;
86   void buildMergeLaneMasks(MachineBasicBlock &MBB,
87                            MachineBasicBlock::iterator I, const DebugLoc &DL,
88                            unsigned DstReg, unsigned PrevReg, unsigned CurReg);
89   MachineBasicBlock::iterator
90   getSaluInsertionAtEnd(MachineBasicBlock &MBB) const;
91 
92   bool isVreg1(Register Reg) const {
93     return Reg.isVirtual() && MRI->getRegClass(Reg) == &AMDGPU::VReg_1RegClass;
94   }
95 
96   bool isLaneMaskReg(unsigned Reg) const {
97     return TII->getRegisterInfo().isSGPRReg(*MRI, Reg) &&
98            TII->getRegisterInfo().getRegSizeInBits(Reg, *MRI) ==
99                ST->getWavefrontSize();
100   }
101 };
102 
103 /// Helper class that determines the relationship between incoming values of a
104 /// phi in the control flow graph to determine where an incoming value can
105 /// simply be taken as a scalar lane mask as-is, and where it needs to be
106 /// merged with another, previously defined lane mask.
107 ///
108 /// The approach is as follows:
109 ///  - Determine all basic blocks which, starting from the incoming blocks,
110 ///    a wave may reach before entering the def block (the block containing the
111 ///    phi).
112 ///  - If an incoming block has no predecessors in this set, we can take the
113 ///    incoming value as a scalar lane mask as-is.
114 ///  -- A special case of this is when the def block has a self-loop.
115 ///  - Otherwise, the incoming value needs to be merged with a previously
116 ///    defined lane mask.
117 ///  - If there is a path into the set of reachable blocks that does _not_ go
118 ///    through an incoming block where we can take the scalar lane mask as-is,
119 ///    we need to invent an available value for the SSAUpdater. Choices are
120 ///    0 and undef, with differing consequences for how to merge values etc.
121 ///
122 /// TODO: We could use region analysis to quickly skip over SESE regions during
123 ///       the traversal.
124 ///
125 class PhiIncomingAnalysis {
126   MachinePostDominatorTree &PDT;
127 
128   // For each reachable basic block, whether it is a source in the induced
129   // subgraph of the CFG.
130   DenseMap<MachineBasicBlock *, bool> ReachableMap;
131   SmallVector<MachineBasicBlock *, 4> ReachableOrdered;
132   SmallVector<MachineBasicBlock *, 4> Stack;
133   SmallVector<MachineBasicBlock *, 4> Predecessors;
134 
135 public:
136   PhiIncomingAnalysis(MachinePostDominatorTree &PDT) : PDT(PDT) {}
137 
138   /// Returns whether \p MBB is a source in the induced subgraph of reachable
139   /// blocks.
140   bool isSource(MachineBasicBlock &MBB) const {
141     return ReachableMap.find(&MBB)->second;
142   }
143 
144   ArrayRef<MachineBasicBlock *> predecessors() const { return Predecessors; }
145 
146   void analyze(MachineBasicBlock &DefBlock,
147                ArrayRef<MachineBasicBlock *> IncomingBlocks) {
148     assert(Stack.empty());
149     ReachableMap.clear();
150     ReachableOrdered.clear();
151     Predecessors.clear();
152 
153     // Insert the def block first, so that it acts as an end point for the
154     // traversal.
155     ReachableMap.try_emplace(&DefBlock, false);
156     ReachableOrdered.push_back(&DefBlock);
157 
158     for (MachineBasicBlock *MBB : IncomingBlocks) {
159       if (MBB == &DefBlock) {
160         ReachableMap[&DefBlock] = true; // self-loop on DefBlock
161         continue;
162       }
163 
164       ReachableMap.try_emplace(MBB, false);
165       ReachableOrdered.push_back(MBB);
166 
167       // If this block has a divergent terminator and the def block is its
168       // post-dominator, the wave may first visit the other successors.
169       bool Divergent = false;
170       for (MachineInstr &MI : MBB->terminators()) {
171         if (MI.getOpcode() == AMDGPU::SI_NON_UNIFORM_BRCOND_PSEUDO ||
172             MI.getOpcode() == AMDGPU::SI_IF ||
173             MI.getOpcode() == AMDGPU::SI_ELSE ||
174             MI.getOpcode() == AMDGPU::SI_LOOP) {
175           Divergent = true;
176           break;
177         }
178       }
179 
180       if (Divergent && PDT.dominates(&DefBlock, MBB))
181         append_range(Stack, MBB->successors());
182     }
183 
184     while (!Stack.empty()) {
185       MachineBasicBlock *MBB = Stack.pop_back_val();
186       if (!ReachableMap.try_emplace(MBB, false).second)
187         continue;
188       ReachableOrdered.push_back(MBB);
189 
190       append_range(Stack, MBB->successors());
191     }
192 
193     for (MachineBasicBlock *MBB : ReachableOrdered) {
194       bool HaveReachablePred = false;
195       for (MachineBasicBlock *Pred : MBB->predecessors()) {
196         if (ReachableMap.count(Pred)) {
197           HaveReachablePred = true;
198         } else {
199           Stack.push_back(Pred);
200         }
201       }
202       if (!HaveReachablePred)
203         ReachableMap[MBB] = true;
204       if (HaveReachablePred) {
205         for (MachineBasicBlock *UnreachablePred : Stack) {
206           if (!llvm::is_contained(Predecessors, UnreachablePred))
207             Predecessors.push_back(UnreachablePred);
208         }
209       }
210       Stack.clear();
211     }
212   }
213 };
214 
215 /// Helper class that detects loops which require us to lower an i1 COPY into
216 /// bitwise manipulation.
217 ///
218 /// Unfortunately, we cannot use LoopInfo because LoopInfo does not distinguish
219 /// between loops with the same header. Consider this example:
220 ///
221 ///  A-+-+
222 ///  | | |
223 ///  B-+ |
224 ///  |   |
225 ///  C---+
226 ///
227 /// A is the header of a loop containing A, B, and C as far as LoopInfo is
228 /// concerned. However, an i1 COPY in B that is used in C must be lowered to
229 /// bitwise operations to combine results from different loop iterations when
230 /// B has a divergent branch (since by default we will compile this code such
231 /// that threads in a wave are merged at the entry of C).
232 ///
233 /// The following rule is implemented to determine whether bitwise operations
234 /// are required: use the bitwise lowering for a def in block B if a backward
235 /// edge to B is reachable without going through the nearest common
236 /// post-dominator of B and all uses of the def.
237 ///
238 /// TODO: This rule is conservative because it does not check whether the
239 ///       relevant branches are actually divergent.
240 ///
241 /// The class is designed to cache the CFG traversal so that it can be re-used
242 /// for multiple defs within the same basic block.
243 ///
244 /// TODO: We could use region analysis to quickly skip over SESE regions during
245 ///       the traversal.
246 ///
247 class LoopFinder {
248   MachineDominatorTree &DT;
249   MachinePostDominatorTree &PDT;
250 
251   // All visited / reachable block, tagged by level (level 0 is the def block,
252   // level 1 are all blocks reachable including but not going through the def
253   // block's IPDOM, etc.).
254   DenseMap<MachineBasicBlock *, unsigned> Visited;
255 
256   // Nearest common dominator of all visited blocks by level (level 0 is the
257   // def block). Used for seeding the SSAUpdater.
258   SmallVector<MachineBasicBlock *, 4> CommonDominators;
259 
260   // Post-dominator of all visited blocks.
261   MachineBasicBlock *VisitedPostDom = nullptr;
262 
263   // Level at which a loop was found: 0 is not possible; 1 = a backward edge is
264   // reachable without going through the IPDOM of the def block (if the IPDOM
265   // itself has an edge to the def block, the loop level is 2), etc.
266   unsigned FoundLoopLevel = ~0u;
267 
268   MachineBasicBlock *DefBlock = nullptr;
269   SmallVector<MachineBasicBlock *, 4> Stack;
270   SmallVector<MachineBasicBlock *, 4> NextLevel;
271 
272 public:
273   LoopFinder(MachineDominatorTree &DT, MachinePostDominatorTree &PDT)
274       : DT(DT), PDT(PDT) {}
275 
276   void initialize(MachineBasicBlock &MBB) {
277     Visited.clear();
278     CommonDominators.clear();
279     Stack.clear();
280     NextLevel.clear();
281     VisitedPostDom = nullptr;
282     FoundLoopLevel = ~0u;
283 
284     DefBlock = &MBB;
285   }
286 
287   /// Check whether a backward edge can be reached without going through the
288   /// given \p PostDom of the def block.
289   ///
290   /// Return the level of \p PostDom if a loop was found, or 0 otherwise.
291   unsigned findLoop(MachineBasicBlock *PostDom) {
292     MachineDomTreeNode *PDNode = PDT.getNode(DefBlock);
293 
294     if (!VisitedPostDom)
295       advanceLevel();
296 
297     unsigned Level = 0;
298     while (PDNode->getBlock() != PostDom) {
299       if (PDNode->getBlock() == VisitedPostDom)
300         advanceLevel();
301       PDNode = PDNode->getIDom();
302       Level++;
303       if (FoundLoopLevel == Level)
304         return Level;
305     }
306 
307     return 0;
308   }
309 
310   /// Add undef values dominating the loop and the optionally given additional
311   /// blocks, so that the SSA updater doesn't have to search all the way to the
312   /// function entry.
313   void addLoopEntries(unsigned LoopLevel, MachineSSAUpdater &SSAUpdater,
314                       ArrayRef<MachineBasicBlock *> Blocks = {}) {
315     assert(LoopLevel < CommonDominators.size());
316 
317     MachineBasicBlock *Dom = CommonDominators[LoopLevel];
318     for (MachineBasicBlock *MBB : Blocks)
319       Dom = DT.findNearestCommonDominator(Dom, MBB);
320 
321     if (!inLoopLevel(*Dom, LoopLevel, Blocks)) {
322       SSAUpdater.AddAvailableValue(Dom, insertUndefLaneMask(*Dom));
323     } else {
324       // The dominator is part of the loop or the given blocks, so add the
325       // undef value to unreachable predecessors instead.
326       for (MachineBasicBlock *Pred : Dom->predecessors()) {
327         if (!inLoopLevel(*Pred, LoopLevel, Blocks))
328           SSAUpdater.AddAvailableValue(Pred, insertUndefLaneMask(*Pred));
329       }
330     }
331   }
332 
333 private:
334   bool inLoopLevel(MachineBasicBlock &MBB, unsigned LoopLevel,
335                    ArrayRef<MachineBasicBlock *> Blocks) const {
336     auto DomIt = Visited.find(&MBB);
337     if (DomIt != Visited.end() && DomIt->second <= LoopLevel)
338       return true;
339 
340     if (llvm::is_contained(Blocks, &MBB))
341       return true;
342 
343     return false;
344   }
345 
346   void advanceLevel() {
347     MachineBasicBlock *VisitedDom;
348 
349     if (!VisitedPostDom) {
350       VisitedPostDom = DefBlock;
351       VisitedDom = DefBlock;
352       Stack.push_back(DefBlock);
353     } else {
354       VisitedPostDom = PDT.getNode(VisitedPostDom)->getIDom()->getBlock();
355       VisitedDom = CommonDominators.back();
356 
357       for (unsigned i = 0; i < NextLevel.size();) {
358         if (PDT.dominates(VisitedPostDom, NextLevel[i])) {
359           Stack.push_back(NextLevel[i]);
360 
361           NextLevel[i] = NextLevel.back();
362           NextLevel.pop_back();
363         } else {
364           i++;
365         }
366       }
367     }
368 
369     unsigned Level = CommonDominators.size();
370     while (!Stack.empty()) {
371       MachineBasicBlock *MBB = Stack.pop_back_val();
372       if (!PDT.dominates(VisitedPostDom, MBB))
373         NextLevel.push_back(MBB);
374 
375       Visited[MBB] = Level;
376       VisitedDom = DT.findNearestCommonDominator(VisitedDom, MBB);
377 
378       for (MachineBasicBlock *Succ : MBB->successors()) {
379         if (Succ == DefBlock) {
380           if (MBB == VisitedPostDom)
381             FoundLoopLevel = std::min(FoundLoopLevel, Level + 1);
382           else
383             FoundLoopLevel = std::min(FoundLoopLevel, Level);
384           continue;
385         }
386 
387         if (Visited.try_emplace(Succ, ~0u).second) {
388           if (MBB == VisitedPostDom)
389             NextLevel.push_back(Succ);
390           else
391             Stack.push_back(Succ);
392         }
393       }
394     }
395 
396     CommonDominators.push_back(VisitedDom);
397   }
398 };
399 
400 } // End anonymous namespace.
401 
402 INITIALIZE_PASS_BEGIN(SILowerI1Copies, DEBUG_TYPE, "SI Lower i1 Copies", false,
403                       false)
404 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
405 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)
406 INITIALIZE_PASS_END(SILowerI1Copies, DEBUG_TYPE, "SI Lower i1 Copies", false,
407                     false)
408 
409 char SILowerI1Copies::ID = 0;
410 
411 char &llvm::SILowerI1CopiesID = SILowerI1Copies::ID;
412 
413 FunctionPass *llvm::createSILowerI1CopiesPass() {
414   return new SILowerI1Copies();
415 }
416 
417 static unsigned createLaneMaskReg(MachineFunction &MF) {
418   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
419   MachineRegisterInfo &MRI = MF.getRegInfo();
420   return MRI.createVirtualRegister(ST.isWave32() ? &AMDGPU::SReg_32RegClass
421                                                  : &AMDGPU::SReg_64RegClass);
422 }
423 
424 static unsigned insertUndefLaneMask(MachineBasicBlock &MBB) {
425   MachineFunction &MF = *MBB.getParent();
426   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
427   const SIInstrInfo *TII = ST.getInstrInfo();
428   unsigned UndefReg = createLaneMaskReg(MF);
429   BuildMI(MBB, MBB.getFirstTerminator(), {}, TII->get(AMDGPU::IMPLICIT_DEF),
430           UndefReg);
431   return UndefReg;
432 }
433 
434 /// Lower all instructions that def or use vreg_1 registers.
435 ///
436 /// In a first pass, we lower COPYs from vreg_1 to vector registers, as can
437 /// occur around inline assembly. We do this first, before vreg_1 registers
438 /// are changed to scalar mask registers.
439 ///
440 /// Then we lower all defs of vreg_1 registers. Phi nodes are lowered before
441 /// all others, because phi lowering looks through copies and can therefore
442 /// often make copy lowering unnecessary.
443 bool SILowerI1Copies::runOnMachineFunction(MachineFunction &TheMF) {
444   // Only need to run this in SelectionDAG path.
445   if (TheMF.getProperties().hasProperty(
446         MachineFunctionProperties::Property::Selected))
447     return false;
448 
449   MF = &TheMF;
450   MRI = &MF->getRegInfo();
451   DT = &getAnalysis<MachineDominatorTree>();
452   PDT = &getAnalysis<MachinePostDominatorTree>();
453 
454   ST = &MF->getSubtarget<GCNSubtarget>();
455   TII = ST->getInstrInfo();
456   IsWave32 = ST->isWave32();
457 
458   if (IsWave32) {
459     ExecReg = AMDGPU::EXEC_LO;
460     MovOp = AMDGPU::S_MOV_B32;
461     AndOp = AMDGPU::S_AND_B32;
462     OrOp = AMDGPU::S_OR_B32;
463     XorOp = AMDGPU::S_XOR_B32;
464     AndN2Op = AMDGPU::S_ANDN2_B32;
465     OrN2Op = AMDGPU::S_ORN2_B32;
466   } else {
467     ExecReg = AMDGPU::EXEC;
468     MovOp = AMDGPU::S_MOV_B64;
469     AndOp = AMDGPU::S_AND_B64;
470     OrOp = AMDGPU::S_OR_B64;
471     XorOp = AMDGPU::S_XOR_B64;
472     AndN2Op = AMDGPU::S_ANDN2_B64;
473     OrN2Op = AMDGPU::S_ORN2_B64;
474   }
475 
476   lowerCopiesFromI1();
477   lowerPhis();
478   lowerCopiesToI1();
479 
480   for (unsigned Reg : ConstrainRegs)
481     MRI->constrainRegClass(Reg, &AMDGPU::SReg_1_XEXECRegClass);
482   ConstrainRegs.clear();
483 
484   return true;
485 }
486 
487 #ifndef NDEBUG
488 static bool isVRegCompatibleReg(const SIRegisterInfo &TRI,
489                                 const MachineRegisterInfo &MRI,
490                                 Register Reg) {
491   unsigned Size = TRI.getRegSizeInBits(Reg, MRI);
492   return Size == 1 || Size == 32;
493 }
494 #endif
495 
496 void SILowerI1Copies::lowerCopiesFromI1() {
497   SmallVector<MachineInstr *, 4> DeadCopies;
498 
499   for (MachineBasicBlock &MBB : *MF) {
500     for (MachineInstr &MI : MBB) {
501       if (MI.getOpcode() != AMDGPU::COPY)
502         continue;
503 
504       Register DstReg = MI.getOperand(0).getReg();
505       Register SrcReg = MI.getOperand(1).getReg();
506       if (!isVreg1(SrcReg))
507         continue;
508 
509       if (isLaneMaskReg(DstReg) || isVreg1(DstReg))
510         continue;
511 
512       // Copy into a 32-bit vector register.
513       LLVM_DEBUG(dbgs() << "Lower copy from i1: " << MI);
514       DebugLoc DL = MI.getDebugLoc();
515 
516       assert(isVRegCompatibleReg(TII->getRegisterInfo(), *MRI, DstReg));
517       assert(!MI.getOperand(0).getSubReg());
518 
519       ConstrainRegs.insert(SrcReg);
520       BuildMI(MBB, MI, DL, TII->get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
521           .addImm(0)
522           .addImm(0)
523           .addImm(0)
524           .addImm(-1)
525           .addReg(SrcReg);
526       DeadCopies.push_back(&MI);
527     }
528 
529     for (MachineInstr *MI : DeadCopies)
530       MI->eraseFromParent();
531     DeadCopies.clear();
532   }
533 }
534 
535 void SILowerI1Copies::lowerPhis() {
536   MachineSSAUpdater SSAUpdater(*MF);
537   LoopFinder LF(*DT, *PDT);
538   PhiIncomingAnalysis PIA(*PDT);
539   SmallVector<MachineInstr *, 4> Vreg1Phis;
540   SmallVector<MachineBasicBlock *, 4> IncomingBlocks;
541   SmallVector<unsigned, 4> IncomingRegs;
542   SmallVector<unsigned, 4> IncomingUpdated;
543 #ifndef NDEBUG
544   DenseSet<unsigned> PhiRegisters;
545 #endif
546 
547   for (MachineBasicBlock &MBB : *MF) {
548     for (MachineInstr &MI : MBB.phis()) {
549       if (isVreg1(MI.getOperand(0).getReg()))
550         Vreg1Phis.push_back(&MI);
551     }
552   }
553 
554   MachineBasicBlock *PrevMBB = nullptr;
555   for (MachineInstr *MI : Vreg1Phis) {
556     MachineBasicBlock &MBB = *MI->getParent();
557     if (&MBB != PrevMBB) {
558       LF.initialize(MBB);
559       PrevMBB = &MBB;
560     }
561 
562     LLVM_DEBUG(dbgs() << "Lower PHI: " << *MI);
563 
564     Register DstReg = MI->getOperand(0).getReg();
565     MRI->setRegClass(DstReg, IsWave32 ? &AMDGPU::SReg_32RegClass
566                                       : &AMDGPU::SReg_64RegClass);
567 
568     // Collect incoming values.
569     for (unsigned i = 1; i < MI->getNumOperands(); i += 2) {
570       assert(i + 1 < MI->getNumOperands());
571       Register IncomingReg = MI->getOperand(i).getReg();
572       MachineBasicBlock *IncomingMBB = MI->getOperand(i + 1).getMBB();
573       MachineInstr *IncomingDef = MRI->getUniqueVRegDef(IncomingReg);
574 
575       if (IncomingDef->getOpcode() == AMDGPU::COPY) {
576         IncomingReg = IncomingDef->getOperand(1).getReg();
577         assert(isLaneMaskReg(IncomingReg) || isVreg1(IncomingReg));
578         assert(!IncomingDef->getOperand(1).getSubReg());
579       } else if (IncomingDef->getOpcode() == AMDGPU::IMPLICIT_DEF) {
580         continue;
581       } else {
582         assert(IncomingDef->isPHI() || PhiRegisters.count(IncomingReg));
583       }
584 
585       IncomingBlocks.push_back(IncomingMBB);
586       IncomingRegs.push_back(IncomingReg);
587     }
588 
589 #ifndef NDEBUG
590     PhiRegisters.insert(DstReg);
591 #endif
592 
593     // Phis in a loop that are observed outside the loop receive a simple but
594     // conservatively correct treatment.
595     std::vector<MachineBasicBlock *> DomBlocks = {&MBB};
596     for (MachineInstr &Use : MRI->use_instructions(DstReg))
597       DomBlocks.push_back(Use.getParent());
598 
599     MachineBasicBlock *PostDomBound =
600         PDT->findNearestCommonDominator(DomBlocks);
601 
602     // FIXME: This fails to find irreducible cycles. If we have a def (other
603     // than a constant) in a pair of blocks that end up looping back to each
604     // other, it will be mishandle. Due to structurization this shouldn't occur
605     // in practice.
606     unsigned FoundLoopLevel = LF.findLoop(PostDomBound);
607 
608     SSAUpdater.Initialize(DstReg);
609 
610     if (FoundLoopLevel) {
611       LF.addLoopEntries(FoundLoopLevel, SSAUpdater, IncomingBlocks);
612 
613       for (unsigned i = 0; i < IncomingRegs.size(); ++i) {
614         IncomingUpdated.push_back(createLaneMaskReg(*MF));
615         SSAUpdater.AddAvailableValue(IncomingBlocks[i],
616                                      IncomingUpdated.back());
617       }
618 
619       for (unsigned i = 0; i < IncomingRegs.size(); ++i) {
620         MachineBasicBlock &IMBB = *IncomingBlocks[i];
621         buildMergeLaneMasks(
622             IMBB, getSaluInsertionAtEnd(IMBB), {}, IncomingUpdated[i],
623             SSAUpdater.GetValueInMiddleOfBlock(&IMBB), IncomingRegs[i]);
624       }
625     } else {
626       // The phi is not observed from outside a loop. Use a more accurate
627       // lowering.
628       PIA.analyze(MBB, IncomingBlocks);
629 
630       for (MachineBasicBlock *MBB : PIA.predecessors())
631         SSAUpdater.AddAvailableValue(MBB, insertUndefLaneMask(*MBB));
632 
633       for (unsigned i = 0; i < IncomingRegs.size(); ++i) {
634         MachineBasicBlock &IMBB = *IncomingBlocks[i];
635         if (PIA.isSource(IMBB)) {
636           IncomingUpdated.push_back(0);
637           SSAUpdater.AddAvailableValue(&IMBB, IncomingRegs[i]);
638         } else {
639           IncomingUpdated.push_back(createLaneMaskReg(*MF));
640           SSAUpdater.AddAvailableValue(&IMBB, IncomingUpdated.back());
641         }
642       }
643 
644       for (unsigned i = 0; i < IncomingRegs.size(); ++i) {
645         if (!IncomingUpdated[i])
646           continue;
647 
648         MachineBasicBlock &IMBB = *IncomingBlocks[i];
649         buildMergeLaneMasks(
650             IMBB, getSaluInsertionAtEnd(IMBB), {}, IncomingUpdated[i],
651             SSAUpdater.GetValueInMiddleOfBlock(&IMBB), IncomingRegs[i]);
652       }
653     }
654 
655     Register NewReg = SSAUpdater.GetValueInMiddleOfBlock(&MBB);
656     if (NewReg != DstReg) {
657       MRI->replaceRegWith(NewReg, DstReg);
658       MI->eraseFromParent();
659     }
660 
661     IncomingBlocks.clear();
662     IncomingRegs.clear();
663     IncomingUpdated.clear();
664   }
665 }
666 
667 void SILowerI1Copies::lowerCopiesToI1() {
668   MachineSSAUpdater SSAUpdater(*MF);
669   LoopFinder LF(*DT, *PDT);
670   SmallVector<MachineInstr *, 4> DeadCopies;
671 
672   for (MachineBasicBlock &MBB : *MF) {
673     LF.initialize(MBB);
674 
675     for (MachineInstr &MI : MBB) {
676       if (MI.getOpcode() != AMDGPU::IMPLICIT_DEF &&
677           MI.getOpcode() != AMDGPU::COPY)
678         continue;
679 
680       Register DstReg = MI.getOperand(0).getReg();
681       if (!isVreg1(DstReg))
682         continue;
683 
684       if (MRI->use_empty(DstReg)) {
685         DeadCopies.push_back(&MI);
686         continue;
687       }
688 
689       LLVM_DEBUG(dbgs() << "Lower Other: " << MI);
690 
691       MRI->setRegClass(DstReg, IsWave32 ? &AMDGPU::SReg_32RegClass
692                                         : &AMDGPU::SReg_64RegClass);
693       if (MI.getOpcode() == AMDGPU::IMPLICIT_DEF)
694         continue;
695 
696       DebugLoc DL = MI.getDebugLoc();
697       Register SrcReg = MI.getOperand(1).getReg();
698       assert(!MI.getOperand(1).getSubReg());
699 
700       if (!SrcReg.isVirtual() || (!isLaneMaskReg(SrcReg) && !isVreg1(SrcReg))) {
701         assert(TII->getRegisterInfo().getRegSizeInBits(SrcReg, *MRI) == 32);
702         unsigned TmpReg = createLaneMaskReg(*MF);
703         BuildMI(MBB, MI, DL, TII->get(AMDGPU::V_CMP_NE_U32_e64), TmpReg)
704             .addReg(SrcReg)
705             .addImm(0);
706         MI.getOperand(1).setReg(TmpReg);
707         SrcReg = TmpReg;
708       }
709 
710       // Defs in a loop that are observed outside the loop must be transformed
711       // into appropriate bit manipulation.
712       std::vector<MachineBasicBlock *> DomBlocks = {&MBB};
713       for (MachineInstr &Use : MRI->use_instructions(DstReg))
714         DomBlocks.push_back(Use.getParent());
715 
716       MachineBasicBlock *PostDomBound =
717           PDT->findNearestCommonDominator(DomBlocks);
718       unsigned FoundLoopLevel = LF.findLoop(PostDomBound);
719       if (FoundLoopLevel) {
720         SSAUpdater.Initialize(DstReg);
721         SSAUpdater.AddAvailableValue(&MBB, DstReg);
722         LF.addLoopEntries(FoundLoopLevel, SSAUpdater);
723 
724         buildMergeLaneMasks(MBB, MI, DL, DstReg,
725                             SSAUpdater.GetValueInMiddleOfBlock(&MBB), SrcReg);
726         DeadCopies.push_back(&MI);
727       }
728     }
729 
730     for (MachineInstr *MI : DeadCopies)
731       MI->eraseFromParent();
732     DeadCopies.clear();
733   }
734 }
735 
736 bool SILowerI1Copies::isConstantLaneMask(Register Reg, bool &Val) const {
737   const MachineInstr *MI;
738   for (;;) {
739     MI = MRI->getUniqueVRegDef(Reg);
740     if (MI->getOpcode() == AMDGPU::IMPLICIT_DEF)
741       return true;
742 
743     if (MI->getOpcode() != AMDGPU::COPY)
744       break;
745 
746     Reg = MI->getOperand(1).getReg();
747     if (!Reg.isVirtual())
748       return false;
749     if (!isLaneMaskReg(Reg))
750       return false;
751   }
752 
753   if (MI->getOpcode() != MovOp)
754     return false;
755 
756   if (!MI->getOperand(1).isImm())
757     return false;
758 
759   int64_t Imm = MI->getOperand(1).getImm();
760   if (Imm == 0) {
761     Val = false;
762     return true;
763   }
764   if (Imm == -1) {
765     Val = true;
766     return true;
767   }
768 
769   return false;
770 }
771 
772 static void instrDefsUsesSCC(const MachineInstr &MI, bool &Def, bool &Use) {
773   Def = false;
774   Use = false;
775 
776   for (const MachineOperand &MO : MI.operands()) {
777     if (MO.isReg() && MO.getReg() == AMDGPU::SCC) {
778       if (MO.isUse())
779         Use = true;
780       else
781         Def = true;
782     }
783   }
784 }
785 
786 /// Return a point at the end of the given \p MBB to insert SALU instructions
787 /// for lane mask calculation. Take terminators and SCC into account.
788 MachineBasicBlock::iterator
789 SILowerI1Copies::getSaluInsertionAtEnd(MachineBasicBlock &MBB) const {
790   auto InsertionPt = MBB.getFirstTerminator();
791   bool TerminatorsUseSCC = false;
792   for (auto I = InsertionPt, E = MBB.end(); I != E; ++I) {
793     bool DefsSCC;
794     instrDefsUsesSCC(*I, DefsSCC, TerminatorsUseSCC);
795     if (TerminatorsUseSCC || DefsSCC)
796       break;
797   }
798 
799   if (!TerminatorsUseSCC)
800     return InsertionPt;
801 
802   while (InsertionPt != MBB.begin()) {
803     InsertionPt--;
804 
805     bool DefSCC, UseSCC;
806     instrDefsUsesSCC(*InsertionPt, DefSCC, UseSCC);
807     if (DefSCC)
808       return InsertionPt;
809   }
810 
811   // We should have at least seen an IMPLICIT_DEF or COPY
812   llvm_unreachable("SCC used by terminator but no def in block");
813 }
814 
815 void SILowerI1Copies::buildMergeLaneMasks(MachineBasicBlock &MBB,
816                                           MachineBasicBlock::iterator I,
817                                           const DebugLoc &DL, unsigned DstReg,
818                                           unsigned PrevReg, unsigned CurReg) {
819   bool PrevVal = false;
820   bool PrevConstant = isConstantLaneMask(PrevReg, PrevVal);
821   bool CurVal = false;
822   bool CurConstant = isConstantLaneMask(CurReg, CurVal);
823 
824   if (PrevConstant && CurConstant) {
825     if (PrevVal == CurVal) {
826       BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), DstReg).addReg(CurReg);
827     } else if (CurVal) {
828       BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), DstReg).addReg(ExecReg);
829     } else {
830       BuildMI(MBB, I, DL, TII->get(XorOp), DstReg)
831           .addReg(ExecReg)
832           .addImm(-1);
833     }
834     return;
835   }
836 
837   unsigned PrevMaskedReg = 0;
838   unsigned CurMaskedReg = 0;
839   if (!PrevConstant) {
840     if (CurConstant && CurVal) {
841       PrevMaskedReg = PrevReg;
842     } else {
843       PrevMaskedReg = createLaneMaskReg(*MF);
844       BuildMI(MBB, I, DL, TII->get(AndN2Op), PrevMaskedReg)
845           .addReg(PrevReg)
846           .addReg(ExecReg);
847     }
848   }
849   if (!CurConstant) {
850     // TODO: check whether CurReg is already masked by EXEC
851     if (PrevConstant && PrevVal) {
852       CurMaskedReg = CurReg;
853     } else {
854       CurMaskedReg = createLaneMaskReg(*MF);
855       BuildMI(MBB, I, DL, TII->get(AndOp), CurMaskedReg)
856           .addReg(CurReg)
857           .addReg(ExecReg);
858     }
859   }
860 
861   if (PrevConstant && !PrevVal) {
862     BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), DstReg)
863         .addReg(CurMaskedReg);
864   } else if (CurConstant && !CurVal) {
865     BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), DstReg)
866         .addReg(PrevMaskedReg);
867   } else if (PrevConstant && PrevVal) {
868     BuildMI(MBB, I, DL, TII->get(OrN2Op), DstReg)
869         .addReg(CurMaskedReg)
870         .addReg(ExecReg);
871   } else {
872     BuildMI(MBB, I, DL, TII->get(OrOp), DstReg)
873         .addReg(PrevMaskedReg)
874         .addReg(CurMaskedReg ? CurMaskedReg : ExecReg);
875   }
876 }
877