1 //===- ARMParallelDSP.cpp - Parallel DSP Pass -----------------------------===//
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
7 //===----------------------------------------------------------------------===//
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.
15 //===----------------------------------------------------------------------===//
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/LoopAccessAnalysis.h"
21 #include "llvm/Analysis/OrderedBasicBlock.h"
22 #include "llvm/IR/Instructions.h"
23 #include "llvm/IR/NoFolder.h"
24 #include "llvm/Transforms/Scalar.h"
25 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
26 #include "llvm/Pass.h"
27 #include "llvm/PassRegistry.h"
28 #include "llvm/PassSupport.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/IR/PatternMatch.h"
31 #include "llvm/CodeGen/TargetPassConfig.h"
33 #include "ARMSubtarget.h"
36 using namespace PatternMatch
;
38 #define DEBUG_TYPE "arm-parallel-dsp"
40 STATISTIC(NumSMLAD
, "Number of smlad instructions generated");
43 DisableParallelDSP("disable-arm-parallel-dsp", cl::Hidden
, cl::init(false),
44 cl::desc("Disable the ARM Parallel DSP pass"));
46 static cl::opt
<unsigned>
47 NumLoadLimit("arm-parallel-dsp-load-limit", cl::Hidden
, cl::init(16),
48 cl::desc("Limit the number of loads analysed"));
54 using MulCandList
= SmallVector
<std::unique_ptr
<MulCandidate
>, 8>;
55 using MemInstList
= SmallVectorImpl
<LoadInst
*>;
56 using MulPairList
= SmallVector
<std::pair
<MulCandidate
*, MulCandidate
*>, 8>;
58 // 'MulCandidate' holds the multiplication instructions that are candidates
59 // for parallel execution.
64 bool Exchange
= false;
67 SmallVector
<LoadInst
*, 2> VecLd
; // Container for loads to widen.
69 MulCandidate(Instruction
*I
, Value
*lhs
, Value
*rhs
) :
70 Root(I
), LHS(lhs
), RHS(rhs
) { }
72 bool HasTwoLoadInputs() const {
73 return isa
<LoadInst
>(LHS
) && isa
<LoadInst
>(RHS
);
76 LoadInst
*getBaseLoad() const {
81 /// Represent a sequence of multiply-accumulate operations with the aim to
82 /// perform the multiplications in parallel.
84 Instruction
*Root
= nullptr;
88 SetVector
<Instruction
*> Adds
;
93 Reduction (Instruction
*Add
) : Root(Add
) { }
95 /// Record an Add instruction that is a part of the this reduction.
96 void InsertAdd(Instruction
*I
) { Adds
.insert(I
); }
98 /// Create MulCandidates, each rooted at a Mul instruction, that is a part
99 /// of this reduction.
101 auto GetMulOperand
= [](Value
*V
) -> Instruction
* {
102 if (auto *SExt
= dyn_cast
<SExtInst
>(V
)) {
103 if (auto *I
= dyn_cast
<Instruction
>(SExt
->getOperand(0)))
104 if (I
->getOpcode() == Instruction::Mul
)
106 } else if (auto *I
= dyn_cast
<Instruction
>(V
)) {
107 if (I
->getOpcode() == Instruction::Mul
)
113 auto InsertMul
= [this](Instruction
*I
) {
114 Value
*LHS
= cast
<Instruction
>(I
->getOperand(0))->getOperand(0);
115 Value
*RHS
= cast
<Instruction
>(I
->getOperand(1))->getOperand(0);
116 Muls
.push_back(std::make_unique
<MulCandidate
>(I
, LHS
, RHS
));
119 for (auto *Add
: Adds
) {
122 if (auto *Mul
= GetMulOperand(Add
->getOperand(0)))
124 if (auto *Mul
= GetMulOperand(Add
->getOperand(1)))
129 /// Add the incoming accumulator value, returns true if a value had not
130 /// already been added. Returning false signals to the user that this
131 /// reduction already has a value to initialise the accumulator.
132 bool InsertAcc(Value
*V
) {
139 /// Set two MulCandidates, rooted at muls, that can be executed as a single
140 /// parallel operation.
141 void AddMulPair(MulCandidate
*Mul0
, MulCandidate
*Mul1
,
142 bool Exchange
= false) {
143 LLVM_DEBUG(dbgs() << "Pairing:\n"
144 << *Mul0
->Root
<< "\n"
145 << *Mul1
->Root
<< "\n");
149 Mul1
->Exchange
= true;
150 MulPairs
.push_back(std::make_pair(Mul0
, Mul1
));
153 /// Return true if enough mul operations are found that can be executed in
155 bool CreateParallelPairs();
157 /// Return the add instruction which is the root of the reduction.
158 Instruction
*getRoot() { return Root
; }
160 bool is64Bit() const { return Root
->getType()->isIntegerTy(64); }
162 Type
*getType() const { return Root
->getType(); }
164 /// Return the incoming value to be accumulated. This maybe null.
165 Value
*getAccumulator() { return Acc
; }
167 /// Return the set of adds that comprise the reduction.
168 SetVector
<Instruction
*> &getAdds() { return Adds
; }
170 /// Return the MulCandidate, rooted at mul instruction, that comprise the
172 MulCandList
&getMuls() { return Muls
; }
174 /// Return the MulCandidate, rooted at mul instructions, that have been
175 /// paired for parallel execution.
176 MulPairList
&getMulPairs() { return MulPairs
; }
178 /// To finalise, replace the uses of the root with the intrinsic call.
179 void UpdateRoot(Instruction
*SMLAD
) {
180 Root
->replaceAllUsesWith(SMLAD
);
184 LLVM_DEBUG(dbgs() << "Reduction:\n";
185 for (auto *Add
: Adds
)
186 LLVM_DEBUG(dbgs() << *Add
<< "\n");
187 for (auto &Mul
: Muls
)
188 LLVM_DEBUG(dbgs() << *Mul
->Root
<< "\n"
189 << " " << *Mul
->LHS
<< "\n"
190 << " " << *Mul
->RHS
<< "\n");
191 LLVM_DEBUG(if (Acc
) dbgs() << "Acc in: " << *Acc
<< "\n")
197 LoadInst
*NewLd
= nullptr;
198 SmallVector
<LoadInst
*, 4> Loads
;
201 WidenedLoad(SmallVectorImpl
<LoadInst
*> &Lds
, LoadInst
*Wide
)
206 LoadInst
*getLoad() {
211 class ARMParallelDSP
: public FunctionPass
{
214 TargetLibraryInfo
*TLI
;
216 const DataLayout
*DL
;
218 std::map
<LoadInst
*, LoadInst
*> LoadPairs
;
219 SmallPtrSet
<LoadInst
*, 4> OffsetLoads
;
220 std::map
<LoadInst
*, std::unique_ptr
<WidenedLoad
>> WideLoads
;
223 bool IsNarrowSequence(Value
*V
);
224 bool Search(Value
*V
, BasicBlock
*BB
, Reduction
&R
);
225 bool RecordMemoryOps(BasicBlock
*BB
);
226 void InsertParallelMACs(Reduction
&Reduction
);
227 bool AreSequentialLoads(LoadInst
*Ld0
, LoadInst
*Ld1
, MemInstList
&VecMem
);
228 LoadInst
* CreateWideLoad(MemInstList
&Loads
, IntegerType
*LoadTy
);
229 bool CreateParallelPairs(Reduction
&R
);
231 /// Try to match and generate: SMLAD, SMLADX - Signed Multiply Accumulate
232 /// Dual performs two signed 16x16-bit multiplications. It adds the
233 /// products to a 32-bit accumulate operand. Optionally, the instruction can
234 /// exchange the halfwords of the second operand before performing the
236 bool MatchSMLAD(Function
&F
);
241 ARMParallelDSP() : FunctionPass(ID
) { }
243 void getAnalysisUsage(AnalysisUsage
&AU
) const override
{
244 FunctionPass::getAnalysisUsage(AU
);
245 AU
.addRequired
<AssumptionCacheTracker
>();
246 AU
.addRequired
<ScalarEvolutionWrapperPass
>();
247 AU
.addRequired
<AAResultsWrapperPass
>();
248 AU
.addRequired
<TargetLibraryInfoWrapperPass
>();
249 AU
.addRequired
<DominatorTreeWrapperPass
>();
250 AU
.addRequired
<TargetPassConfig
>();
251 AU
.addPreserved
<ScalarEvolutionWrapperPass
>();
252 AU
.addPreserved
<GlobalsAAWrapperPass
>();
253 AU
.setPreservesCFG();
256 bool runOnFunction(Function
&F
) override
{
257 if (DisableParallelDSP
)
262 SE
= &getAnalysis
<ScalarEvolutionWrapperPass
>().getSE();
263 AA
= &getAnalysis
<AAResultsWrapperPass
>().getAAResults();
264 TLI
= &getAnalysis
<TargetLibraryInfoWrapperPass
>().getTLI(F
);
265 DT
= &getAnalysis
<DominatorTreeWrapperPass
>().getDomTree();
266 auto &TPC
= getAnalysis
<TargetPassConfig
>();
269 DL
= &M
->getDataLayout();
271 auto &TM
= TPC
.getTM
<TargetMachine
>();
272 auto *ST
= &TM
.getSubtarget
<ARMSubtarget
>(F
);
274 if (!ST
->allowsUnalignedMem()) {
275 LLVM_DEBUG(dbgs() << "Unaligned memory access not supported: not "
276 "running pass ARMParallelDSP\n");
281 LLVM_DEBUG(dbgs() << "DSP extension not enabled: not running pass "
286 if (!ST
->isLittle()) {
287 LLVM_DEBUG(dbgs() << "Only supporting little endian: not running pass "
288 << "ARMParallelDSP\n");
292 LLVM_DEBUG(dbgs() << "\n== Parallel DSP pass ==\n");
293 LLVM_DEBUG(dbgs() << " - " << F
.getName() << "\n\n");
295 bool Changes
= MatchSMLAD(F
);
301 template<typename MemInst
>
302 static bool AreSequentialAccesses(MemInst
*MemOp0
, MemInst
*MemOp1
,
303 const DataLayout
&DL
, ScalarEvolution
&SE
) {
304 if (isConsecutiveAccess(MemOp0
, MemOp1
, DL
, SE
))
309 bool ARMParallelDSP::AreSequentialLoads(LoadInst
*Ld0
, LoadInst
*Ld1
,
310 MemInstList
&VecMem
) {
314 if (!LoadPairs
.count(Ld0
) || LoadPairs
[Ld0
] != Ld1
)
317 LLVM_DEBUG(dbgs() << "Loads are sequential and valid:\n";
318 dbgs() << "Ld0:"; Ld0
->dump();
319 dbgs() << "Ld1:"; Ld1
->dump();
323 VecMem
.push_back(Ld0
);
324 VecMem
.push_back(Ld1
);
328 // MaxBitwidth: the maximum supported bitwidth of the elements in the DSP
329 // instructions, which is set to 16. So here we should collect all i8 and i16
330 // narrow operations.
331 // TODO: we currently only collect i16, and will support i8 later, so that's
332 // why we check that types are equal to MaxBitWidth, and not <= MaxBitWidth.
333 template<unsigned MaxBitWidth
>
334 bool ARMParallelDSP::IsNarrowSequence(Value
*V
) {
335 if (auto *SExt
= dyn_cast
<SExtInst
>(V
)) {
336 if (SExt
->getSrcTy()->getIntegerBitWidth() != MaxBitWidth
)
339 if (auto *Ld
= dyn_cast
<LoadInst
>(SExt
->getOperand(0))) {
340 // Check that this load could be paired.
341 return LoadPairs
.count(Ld
) || OffsetLoads
.count(Ld
);
347 /// Iterate through the block and record base, offset pairs of loads which can
348 /// be widened into a single load.
349 bool ARMParallelDSP::RecordMemoryOps(BasicBlock
*BB
) {
350 SmallVector
<LoadInst
*, 8> Loads
;
351 SmallVector
<Instruction
*, 8> Writes
;
354 OrderedBasicBlock
OrderedBB(BB
);
356 // Collect loads and instruction that may write to memory. For now we only
357 // record loads which are simple, sign-extended and have a single user.
358 // TODO: Allow zero-extended loads.
359 for (auto &I
: *BB
) {
360 if (I
.mayWriteToMemory())
361 Writes
.push_back(&I
);
362 auto *Ld
= dyn_cast
<LoadInst
>(&I
);
363 if (!Ld
|| !Ld
->isSimple() ||
364 !Ld
->hasOneUse() || !isa
<SExtInst
>(Ld
->user_back()))
369 if (Loads
.empty() || Loads
.size() > NumLoadLimit
)
372 using InstSet
= std::set
<Instruction
*>;
373 using DepMap
= std::map
<Instruction
*, InstSet
>;
376 // Record any writes that may alias a load.
377 const auto Size
= LocationSize::unknown();
378 for (auto Write
: Writes
) {
379 for (auto Read
: Loads
) {
380 MemoryLocation ReadLoc
=
381 MemoryLocation(Read
->getPointerOperand(), Size
);
383 if (!isModOrRefSet(intersectModRef(AA
->getModRefInfo(Write
, ReadLoc
),
384 ModRefInfo::ModRef
)))
386 if (OrderedBB
.dominates(Write
, Read
))
387 RAWDeps
[Read
].insert(Write
);
391 // Check whether there's not a write between the two loads which would
392 // prevent them from being safely merged.
393 auto SafeToPair
= [&](LoadInst
*Base
, LoadInst
*Offset
) {
394 LoadInst
*Dominator
= OrderedBB
.dominates(Base
, Offset
) ? Base
: Offset
;
395 LoadInst
*Dominated
= OrderedBB
.dominates(Base
, Offset
) ? Offset
: Base
;
397 if (RAWDeps
.count(Dominated
)) {
398 InstSet
&WritesBefore
= RAWDeps
[Dominated
];
400 for (auto Before
: WritesBefore
) {
401 // We can't move the second load backward, past a write, to merge
402 // with the first load.
403 if (OrderedBB
.dominates(Dominator
, Before
))
410 // Record base, offset load pairs.
411 for (auto *Base
: Loads
) {
412 for (auto *Offset
: Loads
) {
413 if (Base
== Offset
|| OffsetLoads
.count(Offset
))
416 if (AreSequentialAccesses
<LoadInst
>(Base
, Offset
, *DL
, *SE
) &&
417 SafeToPair(Base
, Offset
)) {
418 LoadPairs
[Base
] = Offset
;
419 OffsetLoads
.insert(Offset
);
425 LLVM_DEBUG(if (!LoadPairs
.empty()) {
426 dbgs() << "Consecutive load pairs:\n";
427 for (auto &MapIt
: LoadPairs
) {
428 LLVM_DEBUG(dbgs() << *MapIt
.first
<< ", "
429 << *MapIt
.second
<< "\n");
432 return LoadPairs
.size() > 1;
435 // Search recursively back through the operands to find a tree of values that
436 // form a multiply-accumulate chain. The search records the Add and Mul
437 // instructions that form the reduction and allows us to find a single value
438 // to be used as the initial input to the accumlator.
439 bool ARMParallelDSP::Search(Value
*V
, BasicBlock
*BB
, Reduction
&R
) {
440 // If we find a non-instruction, try to use it as the initial accumulator
441 // value. This may have already been found during the search in which case
442 // this function will return false, signaling a search fail.
443 auto *I
= dyn_cast
<Instruction
>(V
);
445 return R
.InsertAcc(V
);
447 if (I
->getParent() != BB
)
450 switch (I
->getOpcode()) {
453 case Instruction::PHI
:
454 // Could be the accumulator value.
455 return R
.InsertAcc(V
);
456 case Instruction::Add
: {
457 // Adds should be adding together two muls, or another add and a mul to
458 // be within the mac chain. One of the operands may also be the
459 // accumulator value at which point we should stop searching.
461 Value
*LHS
= I
->getOperand(0);
462 Value
*RHS
= I
->getOperand(1);
463 bool ValidLHS
= Search(LHS
, BB
, R
);
464 bool ValidRHS
= Search(RHS
, BB
, R
);
466 if (ValidLHS
&& ValidRHS
)
469 return R
.InsertAcc(I
);
471 case Instruction::Mul
: {
472 Value
*MulOp0
= I
->getOperand(0);
473 Value
*MulOp1
= I
->getOperand(1);
474 return IsNarrowSequence
<16>(MulOp0
) && IsNarrowSequence
<16>(MulOp1
);
476 case Instruction::SExt
:
477 return Search(I
->getOperand(0), BB
, R
);
482 // The pass needs to identify integer add/sub reductions of 16-bit vector
485 // 1) we first need to find integer add then look for this pattern:
489 // sext0 = sext i16 %ld0 to i32
491 // sext1 = sext i16 %ld1 to i32
492 // mul0 = mul %sext0, %sext1
494 // sext2 = sext i16 %ld2 to i32
496 // sext3 = sext i16 %ld3 to i32
497 // mul1 = mul i32 %sext2, %sext3
498 // add0 = add i32 %mul0, %acc0
499 // acc1 = add i32 %add0, %mul1
501 // Which can be selected to:
505 // smlad r2, r0, r1, r2
507 // If constants are used instead of loads, these will need to be hoisted
508 // out and into a register.
510 // If loop invariants are used instead of loads, these need to be packed
511 // before the loop begins.
513 bool ARMParallelDSP::MatchSMLAD(Function
&F
) {
514 bool Changed
= false;
517 SmallPtrSet
<Instruction
*, 4> AllAdds
;
518 if (!RecordMemoryOps(&BB
))
521 for (Instruction
&I
: reverse(BB
)) {
522 if (I
.getOpcode() != Instruction::Add
)
525 if (AllAdds
.count(&I
))
528 const auto *Ty
= I
.getType();
529 if (!Ty
->isIntegerTy(32) && !Ty
->isIntegerTy(64))
533 if (!Search(&I
, &BB
, R
))
537 LLVM_DEBUG(dbgs() << "After search, Reduction:\n"; R
.dump());
539 if (!CreateParallelPairs(R
))
542 InsertParallelMACs(R
);
544 AllAdds
.insert(R
.getAdds().begin(), R
.getAdds().end());
551 bool ARMParallelDSP::CreateParallelPairs(Reduction
&R
) {
553 // Not enough mul operations to make a pair.
554 if (R
.getMuls().size() < 2)
557 // Check that the muls operate directly upon sign extended loads.
558 for (auto &MulCand
: R
.getMuls()) {
559 if (!MulCand
->HasTwoLoadInputs())
563 auto CanPair
= [&](Reduction
&R
, MulCandidate
*PMul0
, MulCandidate
*PMul1
) {
564 // The first elements of each vector should be loads with sexts. If we
565 // find that its two pairs of consecutive loads, then these can be
566 // transformed into two wider loads and the users can be replaced with
568 auto Ld0
= static_cast<LoadInst
*>(PMul0
->LHS
);
569 auto Ld1
= static_cast<LoadInst
*>(PMul1
->LHS
);
570 auto Ld2
= static_cast<LoadInst
*>(PMul0
->RHS
);
571 auto Ld3
= static_cast<LoadInst
*>(PMul1
->RHS
);
573 if (AreSequentialLoads(Ld0
, Ld1
, PMul0
->VecLd
)) {
574 if (AreSequentialLoads(Ld2
, Ld3
, PMul1
->VecLd
)) {
575 LLVM_DEBUG(dbgs() << "OK: found two pairs of parallel loads!\n");
576 R
.AddMulPair(PMul0
, PMul1
);
578 } else if (AreSequentialLoads(Ld3
, Ld2
, PMul1
->VecLd
)) {
579 LLVM_DEBUG(dbgs() << "OK: found two pairs of parallel loads!\n");
580 LLVM_DEBUG(dbgs() << " exchanging Ld2 and Ld3\n");
581 R
.AddMulPair(PMul0
, PMul1
, true);
584 } else if (AreSequentialLoads(Ld1
, Ld0
, PMul0
->VecLd
) &&
585 AreSequentialLoads(Ld2
, Ld3
, PMul1
->VecLd
)) {
586 LLVM_DEBUG(dbgs() << "OK: found two pairs of parallel loads!\n");
587 LLVM_DEBUG(dbgs() << " exchanging Ld0 and Ld1\n");
588 LLVM_DEBUG(dbgs() << " and swapping muls\n");
589 // Only the second operand can be exchanged, so swap the muls.
590 R
.AddMulPair(PMul1
, PMul0
, true);
596 MulCandList
&Muls
= R
.getMuls();
597 const unsigned Elems
= Muls
.size();
598 for (unsigned i
= 0; i
< Elems
; ++i
) {
599 MulCandidate
*PMul0
= static_cast<MulCandidate
*>(Muls
[i
].get());
603 for (unsigned j
= 0; j
< Elems
; ++j
) {
607 MulCandidate
*PMul1
= static_cast<MulCandidate
*>(Muls
[j
].get());
611 const Instruction
*Mul0
= PMul0
->Root
;
612 const Instruction
*Mul1
= PMul1
->Root
;
616 assert(PMul0
!= PMul1
&& "expected different chains");
618 if (CanPair(R
, PMul0
, PMul1
))
622 return !R
.getMulPairs().empty();
625 void ARMParallelDSP::InsertParallelMACs(Reduction
&R
) {
627 auto CreateSMLAD
= [&](LoadInst
* WideLd0
, LoadInst
*WideLd1
,
628 Value
*Acc
, bool Exchange
,
629 Instruction
*InsertAfter
) {
630 // Replace the reduction chain with an intrinsic call
632 Value
* Args
[] = { WideLd0
, WideLd1
, Acc
};
633 Function
*SMLAD
= nullptr;
635 SMLAD
= Acc
->getType()->isIntegerTy(32) ?
636 Intrinsic::getDeclaration(M
, Intrinsic::arm_smladx
) :
637 Intrinsic::getDeclaration(M
, Intrinsic::arm_smlaldx
);
639 SMLAD
= Acc
->getType()->isIntegerTy(32) ?
640 Intrinsic::getDeclaration(M
, Intrinsic::arm_smlad
) :
641 Intrinsic::getDeclaration(M
, Intrinsic::arm_smlald
);
643 IRBuilder
<NoFolder
> Builder(InsertAfter
->getParent(),
644 BasicBlock::iterator(InsertAfter
));
645 Instruction
*Call
= Builder
.CreateCall(SMLAD
, Args
);
650 // Return the instruction after the dominated instruction.
651 auto GetInsertPoint
= [this](Value
*A
, Value
*B
) {
652 assert((isa
<Instruction
>(A
) || isa
<Instruction
>(B
)) &&
653 "expected at least one instruction");
656 if (!isa
<Instruction
>(A
))
658 else if (!isa
<Instruction
>(B
))
661 V
= DT
->dominates(cast
<Instruction
>(A
), cast
<Instruction
>(B
)) ? B
: A
;
663 return &*++BasicBlock::iterator(cast
<Instruction
>(V
));
666 Value
*Acc
= R
.getAccumulator();
668 // For any muls that were discovered but not paired, accumulate their values
670 IRBuilder
<NoFolder
> Builder(R
.getRoot()->getParent());
671 MulCandList
&MulCands
= R
.getMuls();
672 for (auto &MulCand
: MulCands
) {
676 Instruction
*Mul
= cast
<Instruction
>(MulCand
->Root
);
677 LLVM_DEBUG(dbgs() << "Accumulating unpaired mul: " << *Mul
<< "\n");
679 if (R
.getType() != Mul
->getType()) {
680 assert(R
.is64Bit() && "expected 64-bit result");
681 Builder
.SetInsertPoint(&*++BasicBlock::iterator(Mul
));
682 Mul
= cast
<Instruction
>(Builder
.CreateSExt(Mul
, R
.getRoot()->getType()));
690 // If Acc is the original incoming value to the reduction, it could be a
691 // phi. But the phi will dominate Mul, meaning that Mul will be the
693 Builder
.SetInsertPoint(GetInsertPoint(Mul
, Acc
));
694 Acc
= Builder
.CreateAdd(Mul
, Acc
);
699 ConstantInt::get(IntegerType::get(M
->getContext(), 64), 0) :
700 ConstantInt::get(IntegerType::get(M
->getContext(), 32), 0);
701 } else if (Acc
->getType() != R
.getType()) {
702 Builder
.SetInsertPoint(R
.getRoot());
703 Acc
= Builder
.CreateSExt(Acc
, R
.getType());
706 // Roughly sort the mul pairs in their program order.
707 OrderedBasicBlock
OrderedBB(R
.getRoot()->getParent());
708 llvm::sort(R
.getMulPairs(), [&OrderedBB
](auto &PairA
, auto &PairB
) {
709 const Instruction
*A
= PairA
.first
->Root
;
710 const Instruction
*B
= PairB
.first
->Root
;
711 return OrderedBB
.dominates(A
, B
);
714 IntegerType
*Ty
= IntegerType::get(M
->getContext(), 32);
715 for (auto &Pair
: R
.getMulPairs()) {
716 MulCandidate
*LHSMul
= Pair
.first
;
717 MulCandidate
*RHSMul
= Pair
.second
;
718 LoadInst
*BaseLHS
= LHSMul
->getBaseLoad();
719 LoadInst
*BaseRHS
= RHSMul
->getBaseLoad();
720 LoadInst
*WideLHS
= WideLoads
.count(BaseLHS
) ?
721 WideLoads
[BaseLHS
]->getLoad() : CreateWideLoad(LHSMul
->VecLd
, Ty
);
722 LoadInst
*WideRHS
= WideLoads
.count(BaseRHS
) ?
723 WideLoads
[BaseRHS
]->getLoad() : CreateWideLoad(RHSMul
->VecLd
, Ty
);
725 Instruction
*InsertAfter
= GetInsertPoint(WideLHS
, WideRHS
);
726 InsertAfter
= GetInsertPoint(InsertAfter
, Acc
);
727 Acc
= CreateSMLAD(WideLHS
, WideRHS
, Acc
, RHSMul
->Exchange
, InsertAfter
);
729 R
.UpdateRoot(cast
<Instruction
>(Acc
));
732 LoadInst
* ARMParallelDSP::CreateWideLoad(MemInstList
&Loads
,
733 IntegerType
*LoadTy
) {
734 assert(Loads
.size() == 2 && "currently only support widening two loads");
736 LoadInst
*Base
= Loads
[0];
737 LoadInst
*Offset
= Loads
[1];
739 Instruction
*BaseSExt
= dyn_cast
<SExtInst
>(Base
->user_back());
740 Instruction
*OffsetSExt
= dyn_cast
<SExtInst
>(Offset
->user_back());
742 assert((BaseSExt
&& OffsetSExt
)
743 && "Loads should have a single, extending, user");
745 std::function
<void(Value
*, Value
*)> MoveBefore
=
746 [&](Value
*A
, Value
*B
) -> void {
747 if (!isa
<Instruction
>(A
) || !isa
<Instruction
>(B
))
750 auto *Source
= cast
<Instruction
>(A
);
751 auto *Sink
= cast
<Instruction
>(B
);
753 if (DT
->dominates(Source
, Sink
) ||
754 Source
->getParent() != Sink
->getParent() ||
755 isa
<PHINode
>(Source
) || isa
<PHINode
>(Sink
))
758 Source
->moveBefore(Sink
);
759 for (auto &Op
: Source
->operands())
760 MoveBefore(Op
, Source
);
763 // Insert the load at the point of the original dominating load.
764 LoadInst
*DomLoad
= DT
->dominates(Base
, Offset
) ? Base
: Offset
;
765 IRBuilder
<NoFolder
> IRB(DomLoad
->getParent(),
766 ++BasicBlock::iterator(DomLoad
));
768 // Bitcast the pointer to a wider type and create the wide load, while making
769 // sure to maintain the original alignment as this prevents ldrd from being
770 // generated when it could be illegal due to memory alignment.
771 const unsigned AddrSpace
= DomLoad
->getPointerAddressSpace();
772 Value
*VecPtr
= IRB
.CreateBitCast(Base
->getPointerOperand(),
773 LoadTy
->getPointerTo(AddrSpace
));
774 LoadInst
*WideLoad
= IRB
.CreateAlignedLoad(LoadTy
, VecPtr
,
775 Base
->getAlignment());
777 // Make sure everything is in the correct order in the basic block.
778 MoveBefore(Base
->getPointerOperand(), VecPtr
);
779 MoveBefore(VecPtr
, WideLoad
);
781 // From the wide load, create two values that equal the original two loads.
782 // Loads[0] needs trunc while Loads[1] needs a lshr and trunc.
783 // TODO: Support big-endian as well.
784 Value
*Bottom
= IRB
.CreateTrunc(WideLoad
, Base
->getType());
785 Value
*NewBaseSExt
= IRB
.CreateSExt(Bottom
, BaseSExt
->getType());
786 BaseSExt
->replaceAllUsesWith(NewBaseSExt
);
788 IntegerType
*OffsetTy
= cast
<IntegerType
>(Offset
->getType());
789 Value
*ShiftVal
= ConstantInt::get(LoadTy
, OffsetTy
->getBitWidth());
790 Value
*Top
= IRB
.CreateLShr(WideLoad
, ShiftVal
);
791 Value
*Trunc
= IRB
.CreateTrunc(Top
, OffsetTy
);
792 Value
*NewOffsetSExt
= IRB
.CreateSExt(Trunc
, OffsetSExt
->getType());
793 OffsetSExt
->replaceAllUsesWith(NewOffsetSExt
);
795 LLVM_DEBUG(dbgs() << "From Base and Offset:\n"
796 << *Base
<< "\n" << *Offset
<< "\n"
797 << "Created Wide Load:\n"
800 << *NewBaseSExt
<< "\n"
803 << *NewOffsetSExt
<< "\n");
804 WideLoads
.emplace(std::make_pair(Base
,
805 std::make_unique
<WidenedLoad
>(Loads
, WideLoad
)));
809 Pass
*llvm::createARMParallelDSPPass() {
810 return new ARMParallelDSP();
813 char ARMParallelDSP::ID
= 0;
815 INITIALIZE_PASS_BEGIN(ARMParallelDSP
, "arm-parallel-dsp",
816 "Transform functions to use DSP intrinsics", false, false)
817 INITIALIZE_PASS_END(ARMParallelDSP
, "arm-parallel-dsp",
818 "Transform functions to use DSP intrinsics", false, false)