xref: /freebsd/contrib/llvm-project/llvm/lib/Target/ARM/ARMParallelDSP.cpp (revision e53a21abdf2953714e44e3c54b4bb78557cb096c)
1 //===- ARMParallelDSP.cpp - Parallel DSP Pass -----------------------------===//
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 /// \file
10 /// Armv6 introduced instructions to perform 32-bit SIMD operations. The
11 /// purpose of this pass is do some IR pattern matching to create ACLE
12 /// DSP intrinsics, which map on these 32-bit SIMD operations.
13 /// This pass runs only when unaligned accesses is supported/enabled.
14 //
15 //===----------------------------------------------------------------------===//
16 
17 #include "ARM.h"
18 #include "ARMSubtarget.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/Analysis/AssumptionCache.h"
23 #include "llvm/Analysis/GlobalsModRef.h"
24 #include "llvm/Analysis/LoopAccessAnalysis.h"
25 #include "llvm/Analysis/TargetLibraryInfo.h"
26 #include "llvm/CodeGen/TargetPassConfig.h"
27 #include "llvm/IR/IRBuilder.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/IntrinsicsARM.h"
30 #include "llvm/IR/Module.h"
31 #include "llvm/IR/NoFolder.h"
32 #include "llvm/IR/PatternMatch.h"
33 #include "llvm/Pass.h"
34 #include "llvm/PassRegistry.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Transforms/Scalar.h"
37 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
38 
39 using namespace llvm;
40 using namespace PatternMatch;
41 
42 #define DEBUG_TYPE "arm-parallel-dsp"
43 
44 STATISTIC(NumSMLAD , "Number of smlad instructions generated");
45 
46 static cl::opt<bool>
47 DisableParallelDSP("disable-arm-parallel-dsp", cl::Hidden, cl::init(false),
48                    cl::desc("Disable the ARM Parallel DSP pass"));
49 
50 static cl::opt<unsigned>
51 NumLoadLimit("arm-parallel-dsp-load-limit", cl::Hidden, cl::init(16),
52              cl::desc("Limit the number of loads analysed"));
53 
54 namespace {
55   struct MulCandidate;
56   class Reduction;
57 
58   using MulCandList = SmallVector<std::unique_ptr<MulCandidate>, 8>;
59   using MemInstList = SmallVectorImpl<LoadInst*>;
60   using MulPairList = SmallVector<std::pair<MulCandidate*, MulCandidate*>, 8>;
61 
62   // 'MulCandidate' holds the multiplication instructions that are candidates
63   // for parallel execution.
64   struct MulCandidate {
65     Instruction   *Root;
66     Value*        LHS;
67     Value*        RHS;
68     bool          Exchange = false;
69     bool          Paired = false;
70     SmallVector<LoadInst*, 2> VecLd;    // Container for loads to widen.
71 
72     MulCandidate(Instruction *I, Value *lhs, Value *rhs) :
73       Root(I), LHS(lhs), RHS(rhs) { }
74 
75     bool HasTwoLoadInputs() const {
76       return isa<LoadInst>(LHS) && isa<LoadInst>(RHS);
77     }
78 
79     LoadInst *getBaseLoad() const {
80       return VecLd.front();
81     }
82   };
83 
84   /// Represent a sequence of multiply-accumulate operations with the aim to
85   /// perform the multiplications in parallel.
86   class Reduction {
87     Instruction     *Root = nullptr;
88     Value           *Acc = nullptr;
89     MulCandList     Muls;
90     MulPairList        MulPairs;
91     SetVector<Instruction*> Adds;
92 
93   public:
94     Reduction() = delete;
95 
96     Reduction (Instruction *Add) : Root(Add) { }
97 
98     /// Record an Add instruction that is a part of the this reduction.
99     void InsertAdd(Instruction *I) { Adds.insert(I); }
100 
101     /// Create MulCandidates, each rooted at a Mul instruction, that is a part
102     /// of this reduction.
103     void InsertMuls() {
104       auto GetMulOperand = [](Value *V) -> Instruction* {
105         if (auto *SExt = dyn_cast<SExtInst>(V)) {
106           if (auto *I = dyn_cast<Instruction>(SExt->getOperand(0)))
107             if (I->getOpcode() == Instruction::Mul)
108               return I;
109         } else if (auto *I = dyn_cast<Instruction>(V)) {
110           if (I->getOpcode() == Instruction::Mul)
111             return I;
112         }
113         return nullptr;
114       };
115 
116       auto InsertMul = [this](Instruction *I) {
117         Value *LHS = cast<Instruction>(I->getOperand(0))->getOperand(0);
118         Value *RHS = cast<Instruction>(I->getOperand(1))->getOperand(0);
119         Muls.push_back(std::make_unique<MulCandidate>(I, LHS, RHS));
120       };
121 
122       for (auto *Add : Adds) {
123         if (Add == Acc)
124           continue;
125         if (auto *Mul = GetMulOperand(Add->getOperand(0)))
126           InsertMul(Mul);
127         if (auto *Mul = GetMulOperand(Add->getOperand(1)))
128           InsertMul(Mul);
129       }
130     }
131 
132     /// Add the incoming accumulator value, returns true if a value had not
133     /// already been added. Returning false signals to the user that this
134     /// reduction already has a value to initialise the accumulator.
135     bool InsertAcc(Value *V) {
136       if (Acc)
137         return false;
138       Acc = V;
139       return true;
140     }
141 
142     /// Set two MulCandidates, rooted at muls, that can be executed as a single
143     /// parallel operation.
144     void AddMulPair(MulCandidate *Mul0, MulCandidate *Mul1,
145                     bool Exchange = false) {
146       LLVM_DEBUG(dbgs() << "Pairing:\n"
147                  << *Mul0->Root << "\n"
148                  << *Mul1->Root << "\n");
149       Mul0->Paired = true;
150       Mul1->Paired = true;
151       if (Exchange)
152         Mul1->Exchange = true;
153       MulPairs.push_back(std::make_pair(Mul0, Mul1));
154     }
155 
156     /// Return the add instruction which is the root of the reduction.
157     Instruction *getRoot() { return Root; }
158 
159     bool is64Bit() const { return Root->getType()->isIntegerTy(64); }
160 
161     Type *getType() const { return Root->getType(); }
162 
163     /// Return the incoming value to be accumulated. This maybe null.
164     Value *getAccumulator() { return Acc; }
165 
166     /// Return the set of adds that comprise the reduction.
167     SetVector<Instruction*> &getAdds() { return Adds; }
168 
169     /// Return the MulCandidate, rooted at mul instruction, that comprise the
170     /// the reduction.
171     MulCandList &getMuls() { return Muls; }
172 
173     /// Return the MulCandidate, rooted at mul instructions, that have been
174     /// paired for parallel execution.
175     MulPairList &getMulPairs() { return MulPairs; }
176 
177     /// To finalise, replace the uses of the root with the intrinsic call.
178     void UpdateRoot(Instruction *SMLAD) {
179       Root->replaceAllUsesWith(SMLAD);
180     }
181 
182     void dump() {
183       LLVM_DEBUG(dbgs() << "Reduction:\n";
184         for (auto *Add : Adds)
185           LLVM_DEBUG(dbgs() << *Add << "\n");
186         for (auto &Mul : Muls)
187           LLVM_DEBUG(dbgs() << *Mul->Root << "\n"
188                      << "  " << *Mul->LHS << "\n"
189                      << "  " << *Mul->RHS << "\n");
190         LLVM_DEBUG(if (Acc) dbgs() << "Acc in: " << *Acc << "\n")
191       );
192     }
193   };
194 
195   class WidenedLoad {
196     LoadInst *NewLd = nullptr;
197     SmallVector<LoadInst*, 4> Loads;
198 
199   public:
200     WidenedLoad(SmallVectorImpl<LoadInst*> &Lds, LoadInst *Wide)
201       : NewLd(Wide) {
202       append_range(Loads, Lds);
203     }
204     LoadInst *getLoad() {
205       return NewLd;
206     }
207   };
208 
209   class ARMParallelDSP : public FunctionPass {
210     ScalarEvolution   *SE;
211     AliasAnalysis     *AA;
212     TargetLibraryInfo *TLI;
213     DominatorTree     *DT;
214     const DataLayout  *DL;
215     Module            *M;
216     std::map<LoadInst*, LoadInst*> LoadPairs;
217     SmallPtrSet<LoadInst*, 4> OffsetLoads;
218     std::map<LoadInst*, std::unique_ptr<WidenedLoad>> WideLoads;
219 
220     template<unsigned>
221     bool IsNarrowSequence(Value *V);
222     bool Search(Value *V, BasicBlock *BB, Reduction &R);
223     bool RecordMemoryOps(BasicBlock *BB);
224     void InsertParallelMACs(Reduction &Reduction);
225     bool AreSequentialLoads(LoadInst *Ld0, LoadInst *Ld1, MemInstList &VecMem);
226     LoadInst* CreateWideLoad(MemInstList &Loads, IntegerType *LoadTy);
227     bool CreateParallelPairs(Reduction &R);
228 
229     /// Try to match and generate: SMLAD, SMLADX - Signed Multiply Accumulate
230     /// Dual performs two signed 16x16-bit multiplications. It adds the
231     /// products to a 32-bit accumulate operand. Optionally, the instruction can
232     /// exchange the halfwords of the second operand before performing the
233     /// arithmetic.
234     bool MatchSMLAD(Function &F);
235 
236   public:
237     static char ID;
238 
239     ARMParallelDSP() : FunctionPass(ID) { }
240 
241     void getAnalysisUsage(AnalysisUsage &AU) const override {
242       FunctionPass::getAnalysisUsage(AU);
243       AU.addRequired<AssumptionCacheTracker>();
244       AU.addRequired<ScalarEvolutionWrapperPass>();
245       AU.addRequired<AAResultsWrapperPass>();
246       AU.addRequired<TargetLibraryInfoWrapperPass>();
247       AU.addRequired<DominatorTreeWrapperPass>();
248       AU.addRequired<TargetPassConfig>();
249       AU.addPreserved<ScalarEvolutionWrapperPass>();
250       AU.addPreserved<GlobalsAAWrapperPass>();
251       AU.setPreservesCFG();
252     }
253 
254     bool runOnFunction(Function &F) override {
255       if (DisableParallelDSP)
256         return false;
257       if (skipFunction(F))
258         return false;
259 
260       SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
261       AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
262       TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
263       DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
264       auto &TPC = getAnalysis<TargetPassConfig>();
265 
266       M = F.getParent();
267       DL = &M->getDataLayout();
268 
269       auto &TM = TPC.getTM<TargetMachine>();
270       auto *ST = &TM.getSubtarget<ARMSubtarget>(F);
271 
272       if (!ST->allowsUnalignedMem()) {
273         LLVM_DEBUG(dbgs() << "Unaligned memory access not supported: not "
274                              "running pass ARMParallelDSP\n");
275         return false;
276       }
277 
278       if (!ST->hasDSP()) {
279         LLVM_DEBUG(dbgs() << "DSP extension not enabled: not running pass "
280                              "ARMParallelDSP\n");
281         return false;
282       }
283 
284       if (!ST->isLittle()) {
285         LLVM_DEBUG(dbgs() << "Only supporting little endian: not running pass "
286                           << "ARMParallelDSP\n");
287         return false;
288       }
289 
290       LLVM_DEBUG(dbgs() << "\n== Parallel DSP pass ==\n");
291       LLVM_DEBUG(dbgs() << " - " << F.getName() << "\n\n");
292 
293       bool Changes = MatchSMLAD(F);
294       return Changes;
295     }
296   };
297 }
298 
299 bool ARMParallelDSP::AreSequentialLoads(LoadInst *Ld0, LoadInst *Ld1,
300                                         MemInstList &VecMem) {
301   if (!Ld0 || !Ld1)
302     return false;
303 
304   if (!LoadPairs.count(Ld0) || LoadPairs[Ld0] != Ld1)
305     return false;
306 
307   LLVM_DEBUG(dbgs() << "Loads are sequential and valid:\n";
308     dbgs() << "Ld0:"; Ld0->dump();
309     dbgs() << "Ld1:"; Ld1->dump();
310   );
311 
312   VecMem.clear();
313   VecMem.push_back(Ld0);
314   VecMem.push_back(Ld1);
315   return true;
316 }
317 
318 // MaxBitwidth: the maximum supported bitwidth of the elements in the DSP
319 // instructions, which is set to 16. So here we should collect all i8 and i16
320 // narrow operations.
321 // TODO: we currently only collect i16, and will support i8 later, so that's
322 // why we check that types are equal to MaxBitWidth, and not <= MaxBitWidth.
323 template<unsigned MaxBitWidth>
324 bool ARMParallelDSP::IsNarrowSequence(Value *V) {
325   if (auto *SExt = dyn_cast<SExtInst>(V)) {
326     if (SExt->getSrcTy()->getIntegerBitWidth() != MaxBitWidth)
327       return false;
328 
329     if (auto *Ld = dyn_cast<LoadInst>(SExt->getOperand(0))) {
330       // Check that this load could be paired.
331       return LoadPairs.count(Ld) || OffsetLoads.count(Ld);
332     }
333   }
334   return false;
335 }
336 
337 /// Iterate through the block and record base, offset pairs of loads which can
338 /// be widened into a single load.
339 bool ARMParallelDSP::RecordMemoryOps(BasicBlock *BB) {
340   SmallVector<LoadInst*, 8> Loads;
341   SmallVector<Instruction*, 8> Writes;
342   LoadPairs.clear();
343   WideLoads.clear();
344 
345   // Collect loads and instruction that may write to memory. For now we only
346   // record loads which are simple, sign-extended and have a single user.
347   // TODO: Allow zero-extended loads.
348   for (auto &I : *BB) {
349     if (I.mayWriteToMemory())
350       Writes.push_back(&I);
351     auto *Ld = dyn_cast<LoadInst>(&I);
352     if (!Ld || !Ld->isSimple() ||
353         !Ld->hasOneUse() || !isa<SExtInst>(Ld->user_back()))
354       continue;
355     Loads.push_back(Ld);
356   }
357 
358   if (Loads.empty() || Loads.size() > NumLoadLimit)
359     return false;
360 
361   using InstSet = std::set<Instruction*>;
362   using DepMap = std::map<Instruction*, InstSet>;
363   DepMap RAWDeps;
364 
365   // Record any writes that may alias a load.
366   const auto Size = LocationSize::beforeOrAfterPointer();
367   for (auto *Write : Writes) {
368     for (auto *Read : Loads) {
369       MemoryLocation ReadLoc =
370         MemoryLocation(Read->getPointerOperand(), Size);
371 
372       if (!isModOrRefSet(AA->getModRefInfo(Write, ReadLoc)))
373         continue;
374       if (Write->comesBefore(Read))
375         RAWDeps[Read].insert(Write);
376     }
377   }
378 
379   // Check whether there's not a write between the two loads which would
380   // prevent them from being safely merged.
381   auto SafeToPair = [&](LoadInst *Base, LoadInst *Offset) {
382     bool BaseFirst = Base->comesBefore(Offset);
383     LoadInst *Dominator = BaseFirst ? Base : Offset;
384     LoadInst *Dominated = BaseFirst ? Offset : Base;
385 
386     if (RAWDeps.count(Dominated)) {
387       InstSet &WritesBefore = RAWDeps[Dominated];
388 
389       for (auto *Before : WritesBefore) {
390         // We can't move the second load backward, past a write, to merge
391         // with the first load.
392         if (Dominator->comesBefore(Before))
393           return false;
394       }
395     }
396     return true;
397   };
398 
399   // Record base, offset load pairs.
400   for (auto *Base : Loads) {
401     for (auto *Offset : Loads) {
402       if (Base == Offset || OffsetLoads.count(Offset))
403         continue;
404 
405       if (isConsecutiveAccess(Base, Offset, *DL, *SE) &&
406           SafeToPair(Base, Offset)) {
407         LoadPairs[Base] = Offset;
408         OffsetLoads.insert(Offset);
409         break;
410       }
411     }
412   }
413 
414   LLVM_DEBUG(if (!LoadPairs.empty()) {
415                dbgs() << "Consecutive load pairs:\n";
416                for (auto &MapIt : LoadPairs) {
417                  LLVM_DEBUG(dbgs() << *MapIt.first << ", "
418                             << *MapIt.second << "\n");
419                }
420              });
421   return LoadPairs.size() > 1;
422 }
423 
424 // Search recursively back through the operands to find a tree of values that
425 // form a multiply-accumulate chain. The search records the Add and Mul
426 // instructions that form the reduction and allows us to find a single value
427 // to be used as the initial input to the accumlator.
428 bool ARMParallelDSP::Search(Value *V, BasicBlock *BB, Reduction &R) {
429   // If we find a non-instruction, try to use it as the initial accumulator
430   // value. This may have already been found during the search in which case
431   // this function will return false, signaling a search fail.
432   auto *I = dyn_cast<Instruction>(V);
433   if (!I)
434     return R.InsertAcc(V);
435 
436   if (I->getParent() != BB)
437     return false;
438 
439   switch (I->getOpcode()) {
440   default:
441     break;
442   case Instruction::PHI:
443     // Could be the accumulator value.
444     return R.InsertAcc(V);
445   case Instruction::Add: {
446     // Adds should be adding together two muls, or another add and a mul to
447     // be within the mac chain. One of the operands may also be the
448     // accumulator value at which point we should stop searching.
449     R.InsertAdd(I);
450     Value *LHS = I->getOperand(0);
451     Value *RHS = I->getOperand(1);
452     bool ValidLHS = Search(LHS, BB, R);
453     bool ValidRHS = Search(RHS, BB, R);
454 
455     if (ValidLHS && ValidRHS)
456       return true;
457 
458     // Ensure we don't add the root as the incoming accumulator.
459     if (R.getRoot() == I)
460       return false;
461 
462     return R.InsertAcc(I);
463   }
464   case Instruction::Mul: {
465     Value *MulOp0 = I->getOperand(0);
466     Value *MulOp1 = I->getOperand(1);
467     return IsNarrowSequence<16>(MulOp0) && IsNarrowSequence<16>(MulOp1);
468   }
469   case Instruction::SExt:
470     return Search(I->getOperand(0), BB, R);
471   }
472   return false;
473 }
474 
475 // The pass needs to identify integer add/sub reductions of 16-bit vector
476 // multiplications.
477 // To use SMLAD:
478 // 1) we first need to find integer add then look for this pattern:
479 //
480 // acc0 = ...
481 // ld0 = load i16
482 // sext0 = sext i16 %ld0 to i32
483 // ld1 = load i16
484 // sext1 = sext i16 %ld1 to i32
485 // mul0 = mul %sext0, %sext1
486 // ld2 = load i16
487 // sext2 = sext i16 %ld2 to i32
488 // ld3 = load i16
489 // sext3 = sext i16 %ld3 to i32
490 // mul1 = mul i32 %sext2, %sext3
491 // add0 = add i32 %mul0, %acc0
492 // acc1 = add i32 %add0, %mul1
493 //
494 // Which can be selected to:
495 //
496 // ldr r0
497 // ldr r1
498 // smlad r2, r0, r1, r2
499 //
500 // If constants are used instead of loads, these will need to be hoisted
501 // out and into a register.
502 //
503 // If loop invariants are used instead of loads, these need to be packed
504 // before the loop begins.
505 //
506 bool ARMParallelDSP::MatchSMLAD(Function &F) {
507   bool Changed = false;
508 
509   for (auto &BB : F) {
510     SmallPtrSet<Instruction*, 4> AllAdds;
511     if (!RecordMemoryOps(&BB))
512       continue;
513 
514     for (Instruction &I : reverse(BB)) {
515       if (I.getOpcode() != Instruction::Add)
516         continue;
517 
518       if (AllAdds.count(&I))
519         continue;
520 
521       const auto *Ty = I.getType();
522       if (!Ty->isIntegerTy(32) && !Ty->isIntegerTy(64))
523         continue;
524 
525       Reduction R(&I);
526       if (!Search(&I, &BB, R))
527         continue;
528 
529       R.InsertMuls();
530       LLVM_DEBUG(dbgs() << "After search, Reduction:\n"; R.dump());
531 
532       if (!CreateParallelPairs(R))
533         continue;
534 
535       InsertParallelMACs(R);
536       Changed = true;
537       AllAdds.insert(R.getAdds().begin(), R.getAdds().end());
538       LLVM_DEBUG(dbgs() << "BB after inserting parallel MACs:\n" << BB);
539     }
540   }
541 
542   return Changed;
543 }
544 
545 bool ARMParallelDSP::CreateParallelPairs(Reduction &R) {
546 
547   // Not enough mul operations to make a pair.
548   if (R.getMuls().size() < 2)
549     return false;
550 
551   // Check that the muls operate directly upon sign extended loads.
552   for (auto &MulCand : R.getMuls()) {
553     if (!MulCand->HasTwoLoadInputs())
554       return false;
555   }
556 
557   auto CanPair = [&](Reduction &R, MulCandidate *PMul0, MulCandidate *PMul1) {
558     // The first elements of each vector should be loads with sexts. If we
559     // find that its two pairs of consecutive loads, then these can be
560     // transformed into two wider loads and the users can be replaced with
561     // DSP intrinsics.
562     auto Ld0 = static_cast<LoadInst*>(PMul0->LHS);
563     auto Ld1 = static_cast<LoadInst*>(PMul1->LHS);
564     auto Ld2 = static_cast<LoadInst*>(PMul0->RHS);
565     auto Ld3 = static_cast<LoadInst*>(PMul1->RHS);
566 
567     // Check that each mul is operating on two different loads.
568     if (Ld0 == Ld2 || Ld1 == Ld3)
569       return false;
570 
571     if (AreSequentialLoads(Ld0, Ld1, PMul0->VecLd)) {
572       if (AreSequentialLoads(Ld2, Ld3, PMul1->VecLd)) {
573         LLVM_DEBUG(dbgs() << "OK: found two pairs of parallel loads!\n");
574         R.AddMulPair(PMul0, PMul1);
575         return true;
576       } else if (AreSequentialLoads(Ld3, Ld2, PMul1->VecLd)) {
577         LLVM_DEBUG(dbgs() << "OK: found two pairs of parallel loads!\n");
578         LLVM_DEBUG(dbgs() << "    exchanging Ld2 and Ld3\n");
579         R.AddMulPair(PMul0, PMul1, true);
580         return true;
581       }
582     } else if (AreSequentialLoads(Ld1, Ld0, PMul0->VecLd) &&
583                AreSequentialLoads(Ld2, Ld3, PMul1->VecLd)) {
584       LLVM_DEBUG(dbgs() << "OK: found two pairs of parallel loads!\n");
585       LLVM_DEBUG(dbgs() << "    exchanging Ld0 and Ld1\n");
586       LLVM_DEBUG(dbgs() << "    and swapping muls\n");
587       // Only the second operand can be exchanged, so swap the muls.
588       R.AddMulPair(PMul1, PMul0, true);
589       return true;
590     }
591     return false;
592   };
593 
594   MulCandList &Muls = R.getMuls();
595   const unsigned Elems = Muls.size();
596   for (unsigned i = 0; i < Elems; ++i) {
597     MulCandidate *PMul0 = static_cast<MulCandidate*>(Muls[i].get());
598     if (PMul0->Paired)
599       continue;
600 
601     for (unsigned j = 0; j < Elems; ++j) {
602       if (i == j)
603         continue;
604 
605       MulCandidate *PMul1 = static_cast<MulCandidate*>(Muls[j].get());
606       if (PMul1->Paired)
607         continue;
608 
609       const Instruction *Mul0 = PMul0->Root;
610       const Instruction *Mul1 = PMul1->Root;
611       if (Mul0 == Mul1)
612         continue;
613 
614       assert(PMul0 != PMul1 && "expected different chains");
615 
616       if (CanPair(R, PMul0, PMul1))
617         break;
618     }
619   }
620   return !R.getMulPairs().empty();
621 }
622 
623 void ARMParallelDSP::InsertParallelMACs(Reduction &R) {
624 
625   auto CreateSMLAD = [&](LoadInst* WideLd0, LoadInst *WideLd1,
626                          Value *Acc, bool Exchange,
627                          Instruction *InsertAfter) {
628     // Replace the reduction chain with an intrinsic call
629 
630     Value* Args[] = { WideLd0, WideLd1, Acc };
631     Function *SMLAD = nullptr;
632     if (Exchange)
633       SMLAD = Acc->getType()->isIntegerTy(32) ?
634         Intrinsic::getDeclaration(M, Intrinsic::arm_smladx) :
635         Intrinsic::getDeclaration(M, Intrinsic::arm_smlaldx);
636     else
637       SMLAD = Acc->getType()->isIntegerTy(32) ?
638         Intrinsic::getDeclaration(M, Intrinsic::arm_smlad) :
639         Intrinsic::getDeclaration(M, Intrinsic::arm_smlald);
640 
641     IRBuilder<NoFolder> Builder(InsertAfter->getParent(),
642                                 BasicBlock::iterator(InsertAfter));
643     Instruction *Call = Builder.CreateCall(SMLAD, Args);
644     NumSMLAD++;
645     return Call;
646   };
647 
648   // Return the instruction after the dominated instruction.
649   auto GetInsertPoint = [this](Value *A, Value *B) {
650     assert((isa<Instruction>(A) || isa<Instruction>(B)) &&
651            "expected at least one instruction");
652 
653     Value *V = nullptr;
654     if (!isa<Instruction>(A))
655       V = B;
656     else if (!isa<Instruction>(B))
657       V = A;
658     else
659       V = DT->dominates(cast<Instruction>(A), cast<Instruction>(B)) ? B : A;
660 
661     return &*++BasicBlock::iterator(cast<Instruction>(V));
662   };
663 
664   Value *Acc = R.getAccumulator();
665 
666   // For any muls that were discovered but not paired, accumulate their values
667   // as before.
668   IRBuilder<NoFolder> Builder(R.getRoot()->getParent());
669   MulCandList &MulCands = R.getMuls();
670   for (auto &MulCand : MulCands) {
671     if (MulCand->Paired)
672       continue;
673 
674     Instruction *Mul = cast<Instruction>(MulCand->Root);
675     LLVM_DEBUG(dbgs() << "Accumulating unpaired mul: " << *Mul << "\n");
676 
677     if (R.getType() != Mul->getType()) {
678       assert(R.is64Bit() && "expected 64-bit result");
679       Builder.SetInsertPoint(&*++BasicBlock::iterator(Mul));
680       Mul = cast<Instruction>(Builder.CreateSExt(Mul, R.getRoot()->getType()));
681     }
682 
683     if (!Acc) {
684       Acc = Mul;
685       continue;
686     }
687 
688     // If Acc is the original incoming value to the reduction, it could be a
689     // phi. But the phi will dominate Mul, meaning that Mul will be the
690     // insertion point.
691     Builder.SetInsertPoint(GetInsertPoint(Mul, Acc));
692     Acc = Builder.CreateAdd(Mul, Acc);
693   }
694 
695   if (!Acc) {
696     Acc = R.is64Bit() ?
697       ConstantInt::get(IntegerType::get(M->getContext(), 64), 0) :
698       ConstantInt::get(IntegerType::get(M->getContext(), 32), 0);
699   } else if (Acc->getType() != R.getType()) {
700     Builder.SetInsertPoint(R.getRoot());
701     Acc = Builder.CreateSExt(Acc, R.getType());
702   }
703 
704   // Roughly sort the mul pairs in their program order.
705   llvm::sort(R.getMulPairs(), [](auto &PairA, auto &PairB) {
706     const Instruction *A = PairA.first->Root;
707     const Instruction *B = PairB.first->Root;
708     return A->comesBefore(B);
709   });
710 
711   IntegerType *Ty = IntegerType::get(M->getContext(), 32);
712   for (auto &Pair : R.getMulPairs()) {
713     MulCandidate *LHSMul = Pair.first;
714     MulCandidate *RHSMul = Pair.second;
715     LoadInst *BaseLHS = LHSMul->getBaseLoad();
716     LoadInst *BaseRHS = RHSMul->getBaseLoad();
717     LoadInst *WideLHS = WideLoads.count(BaseLHS) ?
718       WideLoads[BaseLHS]->getLoad() : CreateWideLoad(LHSMul->VecLd, Ty);
719     LoadInst *WideRHS = WideLoads.count(BaseRHS) ?
720       WideLoads[BaseRHS]->getLoad() : CreateWideLoad(RHSMul->VecLd, Ty);
721 
722     Instruction *InsertAfter = GetInsertPoint(WideLHS, WideRHS);
723     InsertAfter = GetInsertPoint(InsertAfter, Acc);
724     Acc = CreateSMLAD(WideLHS, WideRHS, Acc, RHSMul->Exchange, InsertAfter);
725   }
726   R.UpdateRoot(cast<Instruction>(Acc));
727 }
728 
729 LoadInst* ARMParallelDSP::CreateWideLoad(MemInstList &Loads,
730                                          IntegerType *LoadTy) {
731   assert(Loads.size() == 2 && "currently only support widening two loads");
732 
733   LoadInst *Base = Loads[0];
734   LoadInst *Offset = Loads[1];
735 
736   Instruction *BaseSExt = dyn_cast<SExtInst>(Base->user_back());
737   Instruction *OffsetSExt = dyn_cast<SExtInst>(Offset->user_back());
738 
739   assert((BaseSExt && OffsetSExt)
740          && "Loads should have a single, extending, user");
741 
742   std::function<void(Value*, Value*)> MoveBefore =
743     [&](Value *A, Value *B) -> void {
744       if (!isa<Instruction>(A) || !isa<Instruction>(B))
745         return;
746 
747       auto *Source = cast<Instruction>(A);
748       auto *Sink = cast<Instruction>(B);
749 
750       if (DT->dominates(Source, Sink) ||
751           Source->getParent() != Sink->getParent() ||
752           isa<PHINode>(Source) || isa<PHINode>(Sink))
753         return;
754 
755       Source->moveBefore(Sink);
756       for (auto &Op : Source->operands())
757         MoveBefore(Op, Source);
758     };
759 
760   // Insert the load at the point of the original dominating load.
761   LoadInst *DomLoad = DT->dominates(Base, Offset) ? Base : Offset;
762   IRBuilder<NoFolder> IRB(DomLoad->getParent(),
763                           ++BasicBlock::iterator(DomLoad));
764 
765   // Create the wide load, while making sure to maintain the original alignment
766   // as this prevents ldrd from being generated when it could be illegal due to
767   // memory alignment.
768   Value *VecPtr = Base->getPointerOperand();
769   LoadInst *WideLoad = IRB.CreateAlignedLoad(LoadTy, VecPtr, Base->getAlign());
770 
771   // Make sure everything is in the correct order in the basic block.
772   MoveBefore(Base->getPointerOperand(), VecPtr);
773   MoveBefore(VecPtr, WideLoad);
774 
775   // From the wide load, create two values that equal the original two loads.
776   // Loads[0] needs trunc while Loads[1] needs a lshr and trunc.
777   // TODO: Support big-endian as well.
778   Value *Bottom = IRB.CreateTrunc(WideLoad, Base->getType());
779   Value *NewBaseSExt = IRB.CreateSExt(Bottom, BaseSExt->getType());
780   BaseSExt->replaceAllUsesWith(NewBaseSExt);
781 
782   IntegerType *OffsetTy = cast<IntegerType>(Offset->getType());
783   Value *ShiftVal = ConstantInt::get(LoadTy, OffsetTy->getBitWidth());
784   Value *Top = IRB.CreateLShr(WideLoad, ShiftVal);
785   Value *Trunc = IRB.CreateTrunc(Top, OffsetTy);
786   Value *NewOffsetSExt = IRB.CreateSExt(Trunc, OffsetSExt->getType());
787   OffsetSExt->replaceAllUsesWith(NewOffsetSExt);
788 
789   LLVM_DEBUG(dbgs() << "From Base and Offset:\n"
790              << *Base << "\n" << *Offset << "\n"
791              << "Created Wide Load:\n"
792              << *WideLoad << "\n"
793              << *Bottom << "\n"
794              << *NewBaseSExt << "\n"
795              << *Top << "\n"
796              << *Trunc << "\n"
797              << *NewOffsetSExt << "\n");
798   WideLoads.emplace(std::make_pair(Base,
799                                    std::make_unique<WidenedLoad>(Loads, WideLoad)));
800   return WideLoad;
801 }
802 
803 Pass *llvm::createARMParallelDSPPass() {
804   return new ARMParallelDSP();
805 }
806 
807 char ARMParallelDSP::ID = 0;
808 
809 INITIALIZE_PASS_BEGIN(ARMParallelDSP, "arm-parallel-dsp",
810                 "Transform functions to use DSP intrinsics", false, false)
811 INITIALIZE_PASS_END(ARMParallelDSP, "arm-parallel-dsp",
812                 "Transform functions to use DSP intrinsics", false, false)
813