1 //===-- X86PartialReduction.cpp -------------------------------------------===//
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 looks for add instructions used by a horizontal reduction to see
10 // if we might be able to use pmaddwd or psadbw. Some cases of this require
11 // cross basic block knowledge and can't be done in SelectionDAG.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "X86.h"
16 #include "X86TargetMachine.h"
17 #include "llvm/Analysis/ValueTracking.h"
18 #include "llvm/CodeGen/TargetPassConfig.h"
19 #include "llvm/IR/Constants.h"
20 #include "llvm/IR/IRBuilder.h"
21 #include "llvm/IR/Instructions.h"
22 #include "llvm/IR/IntrinsicInst.h"
23 #include "llvm/IR/IntrinsicsX86.h"
24 #include "llvm/IR/Operator.h"
25 #include "llvm/IR/PatternMatch.h"
26 #include "llvm/Pass.h"
27 #include "llvm/Support/KnownBits.h"
28 
29 using namespace llvm;
30 
31 #define DEBUG_TYPE "x86-partial-reduction"
32 
33 namespace {
34 
35 class X86PartialReduction : public FunctionPass {
36   const DataLayout *DL = nullptr;
37   const X86Subtarget *ST = nullptr;
38 
39 public:
40   static char ID; // Pass identification, replacement for typeid.
41 
42   X86PartialReduction() : FunctionPass(ID) { }
43 
44   bool runOnFunction(Function &Fn) override;
45 
46   void getAnalysisUsage(AnalysisUsage &AU) const override {
47     AU.setPreservesCFG();
48   }
49 
50   StringRef getPassName() const override {
51     return "X86 Partial Reduction";
52   }
53 
54 private:
55   bool tryMAddReplacement(Instruction *Op, bool ReduceInOneBB);
56   bool trySADReplacement(Instruction *Op);
57 };
58 }
59 
60 FunctionPass *llvm::createX86PartialReductionPass() {
61   return new X86PartialReduction();
62 }
63 
64 char X86PartialReduction::ID = 0;
65 
66 INITIALIZE_PASS(X86PartialReduction, DEBUG_TYPE,
67                 "X86 Partial Reduction", false, false)
68 
69 // This function should be aligned with detectExtMul() in X86ISelLowering.cpp.
70 static bool matchVPDPBUSDPattern(const X86Subtarget *ST, BinaryOperator *Mul,
71                                  const DataLayout *DL) {
72   if (!ST->hasVNNI() && !ST->hasAVXVNNI())
73     return false;
74 
75   Value *LHS = Mul->getOperand(0);
76   Value *RHS = Mul->getOperand(1);
77 
78   if (isa<SExtInst>(LHS))
79     std::swap(LHS, RHS);
80 
81   auto IsFreeTruncation = [&](Value *Op) {
82     if (auto *Cast = dyn_cast<CastInst>(Op)) {
83       if (Cast->getParent() == Mul->getParent() &&
84           (Cast->getOpcode() == Instruction::SExt ||
85            Cast->getOpcode() == Instruction::ZExt) &&
86           Cast->getOperand(0)->getType()->getScalarSizeInBits() <= 8)
87         return true;
88     }
89 
90     return isa<Constant>(Op);
91   };
92 
93   // (dpbusd (zext a), (sext, b)). Since the first operand should be unsigned
94   // value, we need to check LHS is zero extended value. RHS should be signed
95   // value, so we just check the signed bits.
96   if ((IsFreeTruncation(LHS) &&
97        computeKnownBits(LHS, *DL).countMaxActiveBits() <= 8) &&
98       (IsFreeTruncation(RHS) && ComputeMaxSignificantBits(RHS, *DL) <= 8))
99     return true;
100 
101   return false;
102 }
103 
104 bool X86PartialReduction::tryMAddReplacement(Instruction *Op,
105                                              bool ReduceInOneBB) {
106   if (!ST->hasSSE2())
107     return false;
108 
109   // Need at least 8 elements.
110   if (cast<FixedVectorType>(Op->getType())->getNumElements() < 8)
111     return false;
112 
113   // Element type should be i32.
114   if (!cast<VectorType>(Op->getType())->getElementType()->isIntegerTy(32))
115     return false;
116 
117   auto *Mul = dyn_cast<BinaryOperator>(Op);
118   if (!Mul || Mul->getOpcode() != Instruction::Mul)
119     return false;
120 
121   Value *LHS = Mul->getOperand(0);
122   Value *RHS = Mul->getOperand(1);
123 
124   // If the target support VNNI, leave it to ISel to combine reduce operation
125   // to VNNI instruction.
126   // TODO: we can support transforming reduce to VNNI intrinsic for across block
127   // in this pass.
128   if (ReduceInOneBB && matchVPDPBUSDPattern(ST, Mul, DL))
129     return false;
130 
131   // LHS and RHS should be only used once or if they are the same then only
132   // used twice. Only check this when SSE4.1 is enabled and we have zext/sext
133   // instructions, otherwise we use punpck to emulate zero extend in stages. The
134   // trunc/ we need to do likely won't introduce new instructions in that case.
135   if (ST->hasSSE41()) {
136     if (LHS == RHS) {
137       if (!isa<Constant>(LHS) && !LHS->hasNUses(2))
138         return false;
139     } else {
140       if (!isa<Constant>(LHS) && !LHS->hasOneUse())
141         return false;
142       if (!isa<Constant>(RHS) && !RHS->hasOneUse())
143         return false;
144     }
145   }
146 
147   auto CanShrinkOp = [&](Value *Op) {
148     auto IsFreeTruncation = [&](Value *Op) {
149       if (auto *Cast = dyn_cast<CastInst>(Op)) {
150         if (Cast->getParent() == Mul->getParent() &&
151             (Cast->getOpcode() == Instruction::SExt ||
152              Cast->getOpcode() == Instruction::ZExt) &&
153             Cast->getOperand(0)->getType()->getScalarSizeInBits() <= 16)
154           return true;
155       }
156 
157       return isa<Constant>(Op);
158     };
159 
160     // If the operation can be freely truncated and has enough sign bits we
161     // can shrink.
162     if (IsFreeTruncation(Op) &&
163         ComputeNumSignBits(Op, *DL, 0, nullptr, Mul) > 16)
164       return true;
165 
166     // SelectionDAG has limited support for truncating through an add or sub if
167     // the inputs are freely truncatable.
168     if (auto *BO = dyn_cast<BinaryOperator>(Op)) {
169       if (BO->getParent() == Mul->getParent() &&
170           IsFreeTruncation(BO->getOperand(0)) &&
171           IsFreeTruncation(BO->getOperand(1)) &&
172           ComputeNumSignBits(Op, *DL, 0, nullptr, Mul) > 16)
173         return true;
174     }
175 
176     return false;
177   };
178 
179   // Both Ops need to be shrinkable.
180   if (!CanShrinkOp(LHS) && !CanShrinkOp(RHS))
181     return false;
182 
183   IRBuilder<> Builder(Mul);
184 
185   auto *MulTy = cast<FixedVectorType>(Op->getType());
186   unsigned NumElts = MulTy->getNumElements();
187 
188   // Extract even elements and odd elements and add them together. This will
189   // be pattern matched by SelectionDAG to pmaddwd. This instruction will be
190   // half the original width.
191   SmallVector<int, 16> EvenMask(NumElts / 2);
192   SmallVector<int, 16> OddMask(NumElts / 2);
193   for (int i = 0, e = NumElts / 2; i != e; ++i) {
194     EvenMask[i] = i * 2;
195     OddMask[i] = i * 2 + 1;
196   }
197   // Creating a new mul so the replaceAllUsesWith below doesn't replace the
198   // uses in the shuffles we're creating.
199   Value *NewMul = Builder.CreateMul(Mul->getOperand(0), Mul->getOperand(1));
200   Value *EvenElts = Builder.CreateShuffleVector(NewMul, NewMul, EvenMask);
201   Value *OddElts = Builder.CreateShuffleVector(NewMul, NewMul, OddMask);
202   Value *MAdd = Builder.CreateAdd(EvenElts, OddElts);
203 
204   // Concatenate zeroes to extend back to the original type.
205   SmallVector<int, 32> ConcatMask(NumElts);
206   std::iota(ConcatMask.begin(), ConcatMask.end(), 0);
207   Value *Zero = Constant::getNullValue(MAdd->getType());
208   Value *Concat = Builder.CreateShuffleVector(MAdd, Zero, ConcatMask);
209 
210   Mul->replaceAllUsesWith(Concat);
211   Mul->eraseFromParent();
212 
213   return true;
214 }
215 
216 bool X86PartialReduction::trySADReplacement(Instruction *Op) {
217   if (!ST->hasSSE2())
218     return false;
219 
220   // TODO: There's nothing special about i32, any integer type above i16 should
221   // work just as well.
222   if (!cast<VectorType>(Op->getType())->getElementType()->isIntegerTy(32))
223     return false;
224 
225   Value *LHS;
226   if (match(Op, PatternMatch::m_Intrinsic<Intrinsic::abs>())) {
227     LHS = Op->getOperand(0);
228   } else {
229     // Operand should be a select.
230     auto *SI = dyn_cast<SelectInst>(Op);
231     if (!SI)
232       return false;
233 
234     Value *RHS;
235     // Select needs to implement absolute value.
236     auto SPR = matchSelectPattern(SI, LHS, RHS);
237     if (SPR.Flavor != SPF_ABS)
238       return false;
239   }
240 
241   // Need a subtract of two values.
242   auto *Sub = dyn_cast<BinaryOperator>(LHS);
243   if (!Sub || Sub->getOpcode() != Instruction::Sub)
244     return false;
245 
246   // Look for zero extend from i8.
247   auto getZeroExtendedVal = [](Value *Op) -> Value * {
248     if (auto *ZExt = dyn_cast<ZExtInst>(Op))
249       if (cast<VectorType>(ZExt->getOperand(0)->getType())
250               ->getElementType()
251               ->isIntegerTy(8))
252         return ZExt->getOperand(0);
253 
254     return nullptr;
255   };
256 
257   // Both operands of the subtract should be extends from vXi8.
258   Value *Op0 = getZeroExtendedVal(Sub->getOperand(0));
259   Value *Op1 = getZeroExtendedVal(Sub->getOperand(1));
260   if (!Op0 || !Op1)
261     return false;
262 
263   IRBuilder<> Builder(Op);
264 
265   auto *OpTy = cast<FixedVectorType>(Op->getType());
266   unsigned NumElts = OpTy->getNumElements();
267 
268   unsigned IntrinsicNumElts;
269   Intrinsic::ID IID;
270   if (ST->hasBWI() && NumElts >= 64) {
271     IID = Intrinsic::x86_avx512_psad_bw_512;
272     IntrinsicNumElts = 64;
273   } else if (ST->hasAVX2() && NumElts >= 32) {
274     IID = Intrinsic::x86_avx2_psad_bw;
275     IntrinsicNumElts = 32;
276   } else {
277     IID = Intrinsic::x86_sse2_psad_bw;
278     IntrinsicNumElts = 16;
279   }
280 
281   Function *PSADBWFn = Intrinsic::getDeclaration(Op->getModule(), IID);
282 
283   if (NumElts < 16) {
284     // Pad input with zeroes.
285     SmallVector<int, 32> ConcatMask(16);
286     for (unsigned i = 0; i != NumElts; ++i)
287       ConcatMask[i] = i;
288     for (unsigned i = NumElts; i != 16; ++i)
289       ConcatMask[i] = (i % NumElts) + NumElts;
290 
291     Value *Zero = Constant::getNullValue(Op0->getType());
292     Op0 = Builder.CreateShuffleVector(Op0, Zero, ConcatMask);
293     Op1 = Builder.CreateShuffleVector(Op1, Zero, ConcatMask);
294     NumElts = 16;
295   }
296 
297   // Intrinsics produce vXi64 and need to be casted to vXi32.
298   auto *I32Ty =
299       FixedVectorType::get(Builder.getInt32Ty(), IntrinsicNumElts / 4);
300 
301   assert(NumElts % IntrinsicNumElts == 0 && "Unexpected number of elements!");
302   unsigned NumSplits = NumElts / IntrinsicNumElts;
303 
304   // First collect the pieces we need.
305   SmallVector<Value *, 4> Ops(NumSplits);
306   for (unsigned i = 0; i != NumSplits; ++i) {
307     SmallVector<int, 64> ExtractMask(IntrinsicNumElts);
308     std::iota(ExtractMask.begin(), ExtractMask.end(), i * IntrinsicNumElts);
309     Value *ExtractOp0 = Builder.CreateShuffleVector(Op0, Op0, ExtractMask);
310     Value *ExtractOp1 = Builder.CreateShuffleVector(Op1, Op0, ExtractMask);
311     Ops[i] = Builder.CreateCall(PSADBWFn, {ExtractOp0, ExtractOp1});
312     Ops[i] = Builder.CreateBitCast(Ops[i], I32Ty);
313   }
314 
315   assert(isPowerOf2_32(NumSplits) && "Expected power of 2 splits");
316   unsigned Stages = Log2_32(NumSplits);
317   for (unsigned s = Stages; s > 0; --s) {
318     unsigned NumConcatElts =
319         cast<FixedVectorType>(Ops[0]->getType())->getNumElements() * 2;
320     for (unsigned i = 0; i != 1U << (s - 1); ++i) {
321       SmallVector<int, 64> ConcatMask(NumConcatElts);
322       std::iota(ConcatMask.begin(), ConcatMask.end(), 0);
323       Ops[i] = Builder.CreateShuffleVector(Ops[i*2], Ops[i*2+1], ConcatMask);
324     }
325   }
326 
327   // At this point the final value should be in Ops[0]. Now we need to adjust
328   // it to the final original type.
329   NumElts = cast<FixedVectorType>(OpTy)->getNumElements();
330   if (NumElts == 2) {
331     // Extract down to 2 elements.
332     Ops[0] = Builder.CreateShuffleVector(Ops[0], Ops[0], ArrayRef<int>{0, 1});
333   } else if (NumElts >= 8) {
334     SmallVector<int, 32> ConcatMask(NumElts);
335     unsigned SubElts =
336         cast<FixedVectorType>(Ops[0]->getType())->getNumElements();
337     for (unsigned i = 0; i != SubElts; ++i)
338       ConcatMask[i] = i;
339     for (unsigned i = SubElts; i != NumElts; ++i)
340       ConcatMask[i] = (i % SubElts) + SubElts;
341 
342     Value *Zero = Constant::getNullValue(Ops[0]->getType());
343     Ops[0] = Builder.CreateShuffleVector(Ops[0], Zero, ConcatMask);
344   }
345 
346   Op->replaceAllUsesWith(Ops[0]);
347   Op->eraseFromParent();
348 
349   return true;
350 }
351 
352 // Walk backwards from the ExtractElementInst and determine if it is the end of
353 // a horizontal reduction. Return the input to the reduction if we find one.
354 static Value *matchAddReduction(const ExtractElementInst &EE,
355                                 bool &ReduceInOneBB) {
356   ReduceInOneBB = true;
357   // Make sure we're extracting index 0.
358   auto *Index = dyn_cast<ConstantInt>(EE.getIndexOperand());
359   if (!Index || !Index->isNullValue())
360     return nullptr;
361 
362   const auto *BO = dyn_cast<BinaryOperator>(EE.getVectorOperand());
363   if (!BO || BO->getOpcode() != Instruction::Add || !BO->hasOneUse())
364     return nullptr;
365   if (EE.getParent() != BO->getParent())
366     ReduceInOneBB = false;
367 
368   unsigned NumElems = cast<FixedVectorType>(BO->getType())->getNumElements();
369   // Ensure the reduction size is a power of 2.
370   if (!isPowerOf2_32(NumElems))
371     return nullptr;
372 
373   const Value *Op = BO;
374   unsigned Stages = Log2_32(NumElems);
375   for (unsigned i = 0; i != Stages; ++i) {
376     const auto *BO = dyn_cast<BinaryOperator>(Op);
377     if (!BO || BO->getOpcode() != Instruction::Add)
378       return nullptr;
379     if (EE.getParent() != BO->getParent())
380       ReduceInOneBB = false;
381 
382     // If this isn't the first add, then it should only have 2 users, the
383     // shuffle and another add which we checked in the previous iteration.
384     if (i != 0 && !BO->hasNUses(2))
385       return nullptr;
386 
387     Value *LHS = BO->getOperand(0);
388     Value *RHS = BO->getOperand(1);
389 
390     auto *Shuffle = dyn_cast<ShuffleVectorInst>(LHS);
391     if (Shuffle) {
392       Op = RHS;
393     } else {
394       Shuffle = dyn_cast<ShuffleVectorInst>(RHS);
395       Op = LHS;
396     }
397 
398     // The first operand of the shuffle should be the same as the other operand
399     // of the bin op.
400     if (!Shuffle || Shuffle->getOperand(0) != Op)
401       return nullptr;
402 
403     // Verify the shuffle has the expected (at this stage of the pyramid) mask.
404     unsigned MaskEnd = 1 << i;
405     for (unsigned Index = 0; Index < MaskEnd; ++Index)
406       if (Shuffle->getMaskValue(Index) != (int)(MaskEnd + Index))
407         return nullptr;
408   }
409 
410   return const_cast<Value *>(Op);
411 }
412 
413 // See if this BO is reachable from this Phi by walking forward through single
414 // use BinaryOperators with the same opcode. If we get back then we know we've
415 // found a loop and it is safe to step through this Add to find more leaves.
416 static bool isReachableFromPHI(PHINode *Phi, BinaryOperator *BO) {
417   // The PHI itself should only have one use.
418   if (!Phi->hasOneUse())
419     return false;
420 
421   Instruction *U = cast<Instruction>(*Phi->user_begin());
422   if (U == BO)
423     return true;
424 
425   while (U->hasOneUse() && U->getOpcode() == BO->getOpcode())
426     U = cast<Instruction>(*U->user_begin());
427 
428   return U == BO;
429 }
430 
431 // Collect all the leaves of the tree of adds that feeds into the horizontal
432 // reduction. Root is the Value that is used by the horizontal reduction.
433 // We look through single use phis, single use adds, or adds that are used by
434 // a phi that forms a loop with the add.
435 static void collectLeaves(Value *Root, SmallVectorImpl<Instruction *> &Leaves) {
436   SmallPtrSet<Value *, 8> Visited;
437   SmallVector<Value *, 8> Worklist;
438   Worklist.push_back(Root);
439 
440   while (!Worklist.empty()) {
441     Value *V = Worklist.pop_back_val();
442     if (!Visited.insert(V).second)
443       continue;
444 
445     if (auto *PN = dyn_cast<PHINode>(V)) {
446       // PHI node should have single use unless it is the root node, then it
447       // has 2 uses.
448       if (!PN->hasNUses(PN == Root ? 2 : 1))
449         break;
450 
451       // Push incoming values to the worklist.
452       append_range(Worklist, PN->incoming_values());
453 
454       continue;
455     }
456 
457     if (auto *BO = dyn_cast<BinaryOperator>(V)) {
458       if (BO->getOpcode() == Instruction::Add) {
459         // Simple case. Single use, just push its operands to the worklist.
460         if (BO->hasNUses(BO == Root ? 2 : 1)) {
461           append_range(Worklist, BO->operands());
462           continue;
463         }
464 
465         // If there is additional use, make sure it is an unvisited phi that
466         // gets us back to this node.
467         if (BO->hasNUses(BO == Root ? 3 : 2)) {
468           PHINode *PN = nullptr;
469           for (auto *U : BO->users())
470             if (auto *P = dyn_cast<PHINode>(U))
471               if (!Visited.count(P))
472                 PN = P;
473 
474           // If we didn't find a 2-input PHI then this isn't a case we can
475           // handle.
476           if (!PN || PN->getNumIncomingValues() != 2)
477             continue;
478 
479           // Walk forward from this phi to see if it reaches back to this add.
480           if (!isReachableFromPHI(PN, BO))
481             continue;
482 
483           // The phi forms a loop with this Add, push its operands.
484           append_range(Worklist, BO->operands());
485         }
486       }
487     }
488 
489     // Not an add or phi, make it a leaf.
490     if (auto *I = dyn_cast<Instruction>(V)) {
491       if (!V->hasNUses(I == Root ? 2 : 1))
492         continue;
493 
494       // Add this as a leaf.
495       Leaves.push_back(I);
496     }
497   }
498 }
499 
500 bool X86PartialReduction::runOnFunction(Function &F) {
501   if (skipFunction(F))
502     return false;
503 
504   auto *TPC = getAnalysisIfAvailable<TargetPassConfig>();
505   if (!TPC)
506     return false;
507 
508   auto &TM = TPC->getTM<X86TargetMachine>();
509   ST = TM.getSubtargetImpl(F);
510 
511   DL = &F.getParent()->getDataLayout();
512 
513   bool MadeChange = false;
514   for (auto &BB : F) {
515     for (auto &I : BB) {
516       auto *EE = dyn_cast<ExtractElementInst>(&I);
517       if (!EE)
518         continue;
519 
520       bool ReduceInOneBB;
521       // First find a reduction tree.
522       // FIXME: Do we need to handle other opcodes than Add?
523       Value *Root = matchAddReduction(*EE, ReduceInOneBB);
524       if (!Root)
525         continue;
526 
527       SmallVector<Instruction *, 8> Leaves;
528       collectLeaves(Root, Leaves);
529 
530       for (Instruction *I : Leaves) {
531         if (tryMAddReplacement(I, ReduceInOneBB)) {
532           MadeChange = true;
533           continue;
534         }
535 
536         // Don't do SAD matching on the root node. SelectionDAG already
537         // has support for that and currently generates better code.
538         if (I != Root && trySADReplacement(I))
539           MadeChange = true;
540       }
541     }
542   }
543 
544   return MadeChange;
545 }
546