xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/ExpandVectorPredication.cpp (revision 7ef62cebc2f965b0f640263e179276928885e33d) 
1 //===----- CodeGen/ExpandVectorPredication.cpp - Expand VP intrinsics -----===//
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 implements IR expansion for vector predication intrinsics, allowing
10 // targets to enable vector predication until just before codegen.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/CodeGen/ExpandVectorPredication.h"
15 #include "llvm/ADT/Statistic.h"
16 #include "llvm/Analysis/TargetTransformInfo.h"
17 #include "llvm/Analysis/ValueTracking.h"
18 #include "llvm/Analysis/VectorUtils.h"
19 #include "llvm/CodeGen/Passes.h"
20 #include "llvm/IR/Constants.h"
21 #include "llvm/IR/Function.h"
22 #include "llvm/IR/IRBuilder.h"
23 #include "llvm/IR/InstIterator.h"
24 #include "llvm/IR/Instructions.h"
25 #include "llvm/IR/IntrinsicInst.h"
26 #include "llvm/IR/Intrinsics.h"
27 #include "llvm/InitializePasses.h"
28 #include "llvm/Pass.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Compiler.h"
31 #include "llvm/Support/Debug.h"
32 #include <optional>
33 
34 using namespace llvm;
35 
36 using VPLegalization = TargetTransformInfo::VPLegalization;
37 using VPTransform = TargetTransformInfo::VPLegalization::VPTransform;
38 
39 // Keep this in sync with TargetTransformInfo::VPLegalization.
40 #define VPINTERNAL_VPLEGAL_CASES                                               \
41   VPINTERNAL_CASE(Legal)                                                       \
42   VPINTERNAL_CASE(Discard)                                                     \
43   VPINTERNAL_CASE(Convert)
44 
45 #define VPINTERNAL_CASE(X) "|" #X
46 
47 // Override options.
48 static cl::opt<std::string> EVLTransformOverride(
49     "expandvp-override-evl-transform", cl::init(""), cl::Hidden,
50     cl::desc("Options: <empty>" VPINTERNAL_VPLEGAL_CASES
51              ". If non-empty, ignore "
52              "TargetTransformInfo and "
53              "always use this transformation for the %evl parameter (Used in "
54              "testing)."));
55 
56 static cl::opt<std::string> MaskTransformOverride(
57     "expandvp-override-mask-transform", cl::init(""), cl::Hidden,
58     cl::desc("Options: <empty>" VPINTERNAL_VPLEGAL_CASES
59              ". If non-empty, Ignore "
60              "TargetTransformInfo and "
61              "always use this transformation for the %mask parameter (Used in "
62              "testing)."));
63 
64 #undef VPINTERNAL_CASE
65 #define VPINTERNAL_CASE(X) .Case(#X, VPLegalization::X)
66 
67 static VPTransform parseOverrideOption(const std::string &TextOpt) {
68   return StringSwitch<VPTransform>(TextOpt) VPINTERNAL_VPLEGAL_CASES;
69 }
70 
71 #undef VPINTERNAL_VPLEGAL_CASES
72 
73 // Whether any override options are set.
74 static bool anyExpandVPOverridesSet() {
75   return !EVLTransformOverride.empty() || !MaskTransformOverride.empty();
76 }
77 
78 #define DEBUG_TYPE "expandvp"
79 
80 STATISTIC(NumFoldedVL, "Number of folded vector length params");
81 STATISTIC(NumLoweredVPOps, "Number of folded vector predication operations");
82 
83 ///// Helpers {
84 
85 /// \returns Whether the vector mask \p MaskVal has all lane bits set.
86 static bool isAllTrueMask(Value *MaskVal) {
87   if (Value *SplattedVal = getSplatValue(MaskVal))
88     if (auto *ConstValue = dyn_cast<Constant>(SplattedVal))
89       return ConstValue->isAllOnesValue();
90 
91   return false;
92 }
93 
94 /// \returns A non-excepting divisor constant for this type.
95 static Constant *getSafeDivisor(Type *DivTy) {
96   assert(DivTy->isIntOrIntVectorTy() && "Unsupported divisor type");
97   return ConstantInt::get(DivTy, 1u, false);
98 }
99 
100 /// Transfer operation properties from \p OldVPI to \p NewVal.
101 static void transferDecorations(Value &NewVal, VPIntrinsic &VPI) {
102   auto *NewInst = dyn_cast<Instruction>(&NewVal);
103   if (!NewInst || !isa<FPMathOperator>(NewVal))
104     return;
105 
106   auto *OldFMOp = dyn_cast<FPMathOperator>(&VPI);
107   if (!OldFMOp)
108     return;
109 
110   NewInst->setFastMathFlags(OldFMOp->getFastMathFlags());
111 }
112 
113 /// Transfer all properties from \p OldOp to \p NewOp and replace all uses.
114 /// OldVP gets erased.
115 static void replaceOperation(Value &NewOp, VPIntrinsic &OldOp) {
116   transferDecorations(NewOp, OldOp);
117   OldOp.replaceAllUsesWith(&NewOp);
118   OldOp.eraseFromParent();
119 }
120 
121 static bool maySpeculateLanes(VPIntrinsic &VPI) {
122   // The result of VP reductions depends on the mask and evl.
123   if (isa<VPReductionIntrinsic>(VPI))
124     return false;
125   // Fallback to whether the intrinsic is speculatable.
126   std::optional<unsigned> OpcOpt = VPI.getFunctionalOpcode();
127   unsigned FunctionalOpc = OpcOpt.value_or((unsigned)Instruction::Call);
128   return isSafeToSpeculativelyExecuteWithOpcode(FunctionalOpc, &VPI);
129 }
130 
131 //// } Helpers
132 
133 namespace {
134 
135 // Expansion pass state at function scope.
136 struct CachingVPExpander {
137   Function &F;
138   const TargetTransformInfo &TTI;
139 
140   /// \returns A (fixed length) vector with ascending integer indices
141   /// (<0, 1, ..., NumElems-1>).
142   /// \p Builder
143   ///    Used for instruction creation.
144   /// \p LaneTy
145   ///    Integer element type of the result vector.
146   /// \p NumElems
147   ///    Number of vector elements.
148   Value *createStepVector(IRBuilder<> &Builder, Type *LaneTy,
149                           unsigned NumElems);
150 
151   /// \returns A bitmask that is true where the lane position is less-than \p
152   /// EVLParam
153   ///
154   /// \p Builder
155   ///    Used for instruction creation.
156   /// \p VLParam
157   ///    The explicit vector length parameter to test against the lane
158   ///    positions.
159   /// \p ElemCount
160   ///    Static (potentially scalable) number of vector elements.
161   Value *convertEVLToMask(IRBuilder<> &Builder, Value *EVLParam,
162                           ElementCount ElemCount);
163 
164   Value *foldEVLIntoMask(VPIntrinsic &VPI);
165 
166   /// "Remove" the %evl parameter of \p PI by setting it to the static vector
167   /// length of the operation.
168   void discardEVLParameter(VPIntrinsic &PI);
169 
170   /// Lower this VP binary operator to a unpredicated binary operator.
171   Value *expandPredicationInBinaryOperator(IRBuilder<> &Builder,
172                                            VPIntrinsic &PI);
173 
174   /// Lower this VP reduction to a call to an unpredicated reduction intrinsic.
175   Value *expandPredicationInReduction(IRBuilder<> &Builder,
176                                       VPReductionIntrinsic &PI);
177 
178   /// Lower this VP memory operation to a non-VP intrinsic.
179   Value *expandPredicationInMemoryIntrinsic(IRBuilder<> &Builder,
180                                             VPIntrinsic &VPI);
181 
182   /// Lower this VP comparison to a call to an unpredicated comparison.
183   Value *expandPredicationInComparison(IRBuilder<> &Builder,
184                                        VPCmpIntrinsic &PI);
185 
186   /// Query TTI and expand the vector predication in \p P accordingly.
187   Value *expandPredication(VPIntrinsic &PI);
188 
189   /// Determine how and whether the VPIntrinsic \p VPI shall be expanded. This
190   /// overrides TTI with the cl::opts listed at the top of this file.
191   VPLegalization getVPLegalizationStrategy(const VPIntrinsic &VPI) const;
192   bool UsingTTIOverrides;
193 
194 public:
195   CachingVPExpander(Function &F, const TargetTransformInfo &TTI)
196       : F(F), TTI(TTI), UsingTTIOverrides(anyExpandVPOverridesSet()) {}
197 
198   bool expandVectorPredication();
199 };
200 
201 //// CachingVPExpander {
202 
203 Value *CachingVPExpander::createStepVector(IRBuilder<> &Builder, Type *LaneTy,
204                                            unsigned NumElems) {
205   // TODO add caching
206   SmallVector<Constant *, 16> ConstElems;
207 
208   for (unsigned Idx = 0; Idx < NumElems; ++Idx)
209     ConstElems.push_back(ConstantInt::get(LaneTy, Idx, false));
210 
211   return ConstantVector::get(ConstElems);
212 }
213 
214 Value *CachingVPExpander::convertEVLToMask(IRBuilder<> &Builder,
215                                            Value *EVLParam,
216                                            ElementCount ElemCount) {
217   // TODO add caching
218   // Scalable vector %evl conversion.
219   if (ElemCount.isScalable()) {
220     auto *M = Builder.GetInsertBlock()->getModule();
221     Type *BoolVecTy = VectorType::get(Builder.getInt1Ty(), ElemCount);
222     Function *ActiveMaskFunc = Intrinsic::getDeclaration(
223         M, Intrinsic::get_active_lane_mask, {BoolVecTy, EVLParam->getType()});
224     // `get_active_lane_mask` performs an implicit less-than comparison.
225     Value *ConstZero = Builder.getInt32(0);
226     return Builder.CreateCall(ActiveMaskFunc, {ConstZero, EVLParam});
227   }
228 
229   // Fixed vector %evl conversion.
230   Type *LaneTy = EVLParam->getType();
231   unsigned NumElems = ElemCount.getFixedValue();
232   Value *VLSplat = Builder.CreateVectorSplat(NumElems, EVLParam);
233   Value *IdxVec = createStepVector(Builder, LaneTy, NumElems);
234   return Builder.CreateICmp(CmpInst::ICMP_ULT, IdxVec, VLSplat);
235 }
236 
237 Value *
238 CachingVPExpander::expandPredicationInBinaryOperator(IRBuilder<> &Builder,
239                                                      VPIntrinsic &VPI) {
240   assert((maySpeculateLanes(VPI) || VPI.canIgnoreVectorLengthParam()) &&
241          "Implicitly dropping %evl in non-speculatable operator!");
242 
243   auto OC = static_cast<Instruction::BinaryOps>(*VPI.getFunctionalOpcode());
244   assert(Instruction::isBinaryOp(OC));
245 
246   Value *Op0 = VPI.getOperand(0);
247   Value *Op1 = VPI.getOperand(1);
248   Value *Mask = VPI.getMaskParam();
249 
250   // Blend in safe operands.
251   if (Mask && !isAllTrueMask(Mask)) {
252     switch (OC) {
253     default:
254       // Can safely ignore the predicate.
255       break;
256 
257     // Division operators need a safe divisor on masked-off lanes (1).
258     case Instruction::UDiv:
259     case Instruction::SDiv:
260     case Instruction::URem:
261     case Instruction::SRem:
262       // 2nd operand must not be zero.
263       Value *SafeDivisor = getSafeDivisor(VPI.getType());
264       Op1 = Builder.CreateSelect(Mask, Op1, SafeDivisor);
265     }
266   }
267 
268   Value *NewBinOp = Builder.CreateBinOp(OC, Op0, Op1, VPI.getName());
269 
270   replaceOperation(*NewBinOp, VPI);
271   return NewBinOp;
272 }
273 
274 static Value *getNeutralReductionElement(const VPReductionIntrinsic &VPI,
275                                          Type *EltTy) {
276   bool Negative = false;
277   unsigned EltBits = EltTy->getScalarSizeInBits();
278   switch (VPI.getIntrinsicID()) {
279   default:
280     llvm_unreachable("Expecting a VP reduction intrinsic");
281   case Intrinsic::vp_reduce_add:
282   case Intrinsic::vp_reduce_or:
283   case Intrinsic::vp_reduce_xor:
284   case Intrinsic::vp_reduce_umax:
285     return Constant::getNullValue(EltTy);
286   case Intrinsic::vp_reduce_mul:
287     return ConstantInt::get(EltTy, 1, /*IsSigned*/ false);
288   case Intrinsic::vp_reduce_and:
289   case Intrinsic::vp_reduce_umin:
290     return ConstantInt::getAllOnesValue(EltTy);
291   case Intrinsic::vp_reduce_smin:
292     return ConstantInt::get(EltTy->getContext(),
293                             APInt::getSignedMaxValue(EltBits));
294   case Intrinsic::vp_reduce_smax:
295     return ConstantInt::get(EltTy->getContext(),
296                             APInt::getSignedMinValue(EltBits));
297   case Intrinsic::vp_reduce_fmax:
298     Negative = true;
299     [[fallthrough]];
300   case Intrinsic::vp_reduce_fmin: {
301     FastMathFlags Flags = VPI.getFastMathFlags();
302     const fltSemantics &Semantics = EltTy->getFltSemantics();
303     return !Flags.noNaNs() ? ConstantFP::getQNaN(EltTy, Negative)
304            : !Flags.noInfs()
305                ? ConstantFP::getInfinity(EltTy, Negative)
306                : ConstantFP::get(EltTy,
307                                  APFloat::getLargest(Semantics, Negative));
308   }
309   case Intrinsic::vp_reduce_fadd:
310     return ConstantFP::getNegativeZero(EltTy);
311   case Intrinsic::vp_reduce_fmul:
312     return ConstantFP::get(EltTy, 1.0);
313   }
314 }
315 
316 Value *
317 CachingVPExpander::expandPredicationInReduction(IRBuilder<> &Builder,
318                                                 VPReductionIntrinsic &VPI) {
319   assert((maySpeculateLanes(VPI) || VPI.canIgnoreVectorLengthParam()) &&
320          "Implicitly dropping %evl in non-speculatable operator!");
321 
322   Value *Mask = VPI.getMaskParam();
323   Value *RedOp = VPI.getOperand(VPI.getVectorParamPos());
324 
325   // Insert neutral element in masked-out positions
326   if (Mask && !isAllTrueMask(Mask)) {
327     auto *NeutralElt = getNeutralReductionElement(VPI, VPI.getType());
328     auto *NeutralVector = Builder.CreateVectorSplat(
329         cast<VectorType>(RedOp->getType())->getElementCount(), NeutralElt);
330     RedOp = Builder.CreateSelect(Mask, RedOp, NeutralVector);
331   }
332 
333   Value *Reduction;
334   Value *Start = VPI.getOperand(VPI.getStartParamPos());
335 
336   switch (VPI.getIntrinsicID()) {
337   default:
338     llvm_unreachable("Impossible reduction kind");
339   case Intrinsic::vp_reduce_add:
340     Reduction = Builder.CreateAddReduce(RedOp);
341     Reduction = Builder.CreateAdd(Reduction, Start);
342     break;
343   case Intrinsic::vp_reduce_mul:
344     Reduction = Builder.CreateMulReduce(RedOp);
345     Reduction = Builder.CreateMul(Reduction, Start);
346     break;
347   case Intrinsic::vp_reduce_and:
348     Reduction = Builder.CreateAndReduce(RedOp);
349     Reduction = Builder.CreateAnd(Reduction, Start);
350     break;
351   case Intrinsic::vp_reduce_or:
352     Reduction = Builder.CreateOrReduce(RedOp);
353     Reduction = Builder.CreateOr(Reduction, Start);
354     break;
355   case Intrinsic::vp_reduce_xor:
356     Reduction = Builder.CreateXorReduce(RedOp);
357     Reduction = Builder.CreateXor(Reduction, Start);
358     break;
359   case Intrinsic::vp_reduce_smax:
360     Reduction = Builder.CreateIntMaxReduce(RedOp, /*IsSigned*/ true);
361     Reduction =
362         Builder.CreateBinaryIntrinsic(Intrinsic::smax, Reduction, Start);
363     break;
364   case Intrinsic::vp_reduce_smin:
365     Reduction = Builder.CreateIntMinReduce(RedOp, /*IsSigned*/ true);
366     Reduction =
367         Builder.CreateBinaryIntrinsic(Intrinsic::smin, Reduction, Start);
368     break;
369   case Intrinsic::vp_reduce_umax:
370     Reduction = Builder.CreateIntMaxReduce(RedOp, /*IsSigned*/ false);
371     Reduction =
372         Builder.CreateBinaryIntrinsic(Intrinsic::umax, Reduction, Start);
373     break;
374   case Intrinsic::vp_reduce_umin:
375     Reduction = Builder.CreateIntMinReduce(RedOp, /*IsSigned*/ false);
376     Reduction =
377         Builder.CreateBinaryIntrinsic(Intrinsic::umin, Reduction, Start);
378     break;
379   case Intrinsic::vp_reduce_fmax:
380     Reduction = Builder.CreateFPMaxReduce(RedOp);
381     transferDecorations(*Reduction, VPI);
382     Reduction =
383         Builder.CreateBinaryIntrinsic(Intrinsic::maxnum, Reduction, Start);
384     break;
385   case Intrinsic::vp_reduce_fmin:
386     Reduction = Builder.CreateFPMinReduce(RedOp);
387     transferDecorations(*Reduction, VPI);
388     Reduction =
389         Builder.CreateBinaryIntrinsic(Intrinsic::minnum, Reduction, Start);
390     break;
391   case Intrinsic::vp_reduce_fadd:
392     Reduction = Builder.CreateFAddReduce(Start, RedOp);
393     break;
394   case Intrinsic::vp_reduce_fmul:
395     Reduction = Builder.CreateFMulReduce(Start, RedOp);
396     break;
397   }
398 
399   replaceOperation(*Reduction, VPI);
400   return Reduction;
401 }
402 
403 Value *
404 CachingVPExpander::expandPredicationInMemoryIntrinsic(IRBuilder<> &Builder,
405                                                       VPIntrinsic &VPI) {
406   assert(VPI.canIgnoreVectorLengthParam());
407 
408   const auto &DL = F.getParent()->getDataLayout();
409 
410   Value *MaskParam = VPI.getMaskParam();
411   Value *PtrParam = VPI.getMemoryPointerParam();
412   Value *DataParam = VPI.getMemoryDataParam();
413   bool IsUnmasked = isAllTrueMask(MaskParam);
414 
415   MaybeAlign AlignOpt = VPI.getPointerAlignment();
416 
417   Value *NewMemoryInst = nullptr;
418   switch (VPI.getIntrinsicID()) {
419   default:
420     llvm_unreachable("Not a VP memory intrinsic");
421   case Intrinsic::vp_store:
422     if (IsUnmasked) {
423       StoreInst *NewStore =
424           Builder.CreateStore(DataParam, PtrParam, /*IsVolatile*/ false);
425       if (AlignOpt.has_value())
426         NewStore->setAlignment(*AlignOpt);
427       NewMemoryInst = NewStore;
428     } else
429       NewMemoryInst = Builder.CreateMaskedStore(
430           DataParam, PtrParam, AlignOpt.valueOrOne(), MaskParam);
431 
432     break;
433   case Intrinsic::vp_load:
434     if (IsUnmasked) {
435       LoadInst *NewLoad =
436           Builder.CreateLoad(VPI.getType(), PtrParam, /*IsVolatile*/ false);
437       if (AlignOpt.has_value())
438         NewLoad->setAlignment(*AlignOpt);
439       NewMemoryInst = NewLoad;
440     } else
441       NewMemoryInst = Builder.CreateMaskedLoad(
442           VPI.getType(), PtrParam, AlignOpt.valueOrOne(), MaskParam);
443 
444     break;
445   case Intrinsic::vp_scatter: {
446     auto *ElementType =
447         cast<VectorType>(DataParam->getType())->getElementType();
448     NewMemoryInst = Builder.CreateMaskedScatter(
449         DataParam, PtrParam,
450         AlignOpt.value_or(DL.getPrefTypeAlign(ElementType)), MaskParam);
451     break;
452   }
453   case Intrinsic::vp_gather: {
454     auto *ElementType = cast<VectorType>(VPI.getType())->getElementType();
455     NewMemoryInst = Builder.CreateMaskedGather(
456         VPI.getType(), PtrParam,
457         AlignOpt.value_or(DL.getPrefTypeAlign(ElementType)), MaskParam, nullptr,
458         VPI.getName());
459     break;
460   }
461   }
462 
463   assert(NewMemoryInst);
464   replaceOperation(*NewMemoryInst, VPI);
465   return NewMemoryInst;
466 }
467 
468 Value *CachingVPExpander::expandPredicationInComparison(IRBuilder<> &Builder,
469                                                         VPCmpIntrinsic &VPI) {
470   assert((maySpeculateLanes(VPI) || VPI.canIgnoreVectorLengthParam()) &&
471          "Implicitly dropping %evl in non-speculatable operator!");
472 
473   assert(*VPI.getFunctionalOpcode() == Instruction::ICmp ||
474          *VPI.getFunctionalOpcode() == Instruction::FCmp);
475 
476   Value *Op0 = VPI.getOperand(0);
477   Value *Op1 = VPI.getOperand(1);
478   auto Pred = VPI.getPredicate();
479 
480   auto *NewCmp = Builder.CreateCmp(Pred, Op0, Op1);
481 
482   replaceOperation(*NewCmp, VPI);
483   return NewCmp;
484 }
485 
486 void CachingVPExpander::discardEVLParameter(VPIntrinsic &VPI) {
487   LLVM_DEBUG(dbgs() << "Discard EVL parameter in " << VPI << "\n");
488 
489   if (VPI.canIgnoreVectorLengthParam())
490     return;
491 
492   Value *EVLParam = VPI.getVectorLengthParam();
493   if (!EVLParam)
494     return;
495 
496   ElementCount StaticElemCount = VPI.getStaticVectorLength();
497   Value *MaxEVL = nullptr;
498   Type *Int32Ty = Type::getInt32Ty(VPI.getContext());
499   if (StaticElemCount.isScalable()) {
500     // TODO add caching
501     auto *M = VPI.getModule();
502     Function *VScaleFunc =
503         Intrinsic::getDeclaration(M, Intrinsic::vscale, Int32Ty);
504     IRBuilder<> Builder(VPI.getParent(), VPI.getIterator());
505     Value *FactorConst = Builder.getInt32(StaticElemCount.getKnownMinValue());
506     Value *VScale = Builder.CreateCall(VScaleFunc, {}, "vscale");
507     MaxEVL = Builder.CreateMul(VScale, FactorConst, "scalable_size",
508                                /*NUW*/ true, /*NSW*/ false);
509   } else {
510     MaxEVL = ConstantInt::get(Int32Ty, StaticElemCount.getFixedValue(), false);
511   }
512   VPI.setVectorLengthParam(MaxEVL);
513 }
514 
515 Value *CachingVPExpander::foldEVLIntoMask(VPIntrinsic &VPI) {
516   LLVM_DEBUG(dbgs() << "Folding vlen for " << VPI << '\n');
517 
518   IRBuilder<> Builder(&VPI);
519 
520   // Ineffective %evl parameter and so nothing to do here.
521   if (VPI.canIgnoreVectorLengthParam())
522     return &VPI;
523 
524   // Only VP intrinsics can have an %evl parameter.
525   Value *OldMaskParam = VPI.getMaskParam();
526   Value *OldEVLParam = VPI.getVectorLengthParam();
527   assert(OldMaskParam && "no mask param to fold the vl param into");
528   assert(OldEVLParam && "no EVL param to fold away");
529 
530   LLVM_DEBUG(dbgs() << "OLD evl: " << *OldEVLParam << '\n');
531   LLVM_DEBUG(dbgs() << "OLD mask: " << *OldMaskParam << '\n');
532 
533   // Convert the %evl predication into vector mask predication.
534   ElementCount ElemCount = VPI.getStaticVectorLength();
535   Value *VLMask = convertEVLToMask(Builder, OldEVLParam, ElemCount);
536   Value *NewMaskParam = Builder.CreateAnd(VLMask, OldMaskParam);
537   VPI.setMaskParam(NewMaskParam);
538 
539   // Drop the %evl parameter.
540   discardEVLParameter(VPI);
541   assert(VPI.canIgnoreVectorLengthParam() &&
542          "transformation did not render the evl param ineffective!");
543 
544   // Reassess the modified instruction.
545   return &VPI;
546 }
547 
548 Value *CachingVPExpander::expandPredication(VPIntrinsic &VPI) {
549   LLVM_DEBUG(dbgs() << "Lowering to unpredicated op: " << VPI << '\n');
550 
551   IRBuilder<> Builder(&VPI);
552 
553   // Try lowering to a LLVM instruction first.
554   auto OC = VPI.getFunctionalOpcode();
555 
556   if (OC && Instruction::isBinaryOp(*OC))
557     return expandPredicationInBinaryOperator(Builder, VPI);
558 
559   if (auto *VPRI = dyn_cast<VPReductionIntrinsic>(&VPI))
560     return expandPredicationInReduction(Builder, *VPRI);
561 
562   if (auto *VPCmp = dyn_cast<VPCmpIntrinsic>(&VPI))
563     return expandPredicationInComparison(Builder, *VPCmp);
564 
565   switch (VPI.getIntrinsicID()) {
566   default:
567     break;
568   case Intrinsic::vp_load:
569   case Intrinsic::vp_store:
570   case Intrinsic::vp_gather:
571   case Intrinsic::vp_scatter:
572     return expandPredicationInMemoryIntrinsic(Builder, VPI);
573   }
574 
575   return &VPI;
576 }
577 
578 //// } CachingVPExpander
579 
580 struct TransformJob {
581   VPIntrinsic *PI;
582   TargetTransformInfo::VPLegalization Strategy;
583   TransformJob(VPIntrinsic *PI, TargetTransformInfo::VPLegalization InitStrat)
584       : PI(PI), Strategy(InitStrat) {}
585 
586   bool isDone() const { return Strategy.shouldDoNothing(); }
587 };
588 
589 void sanitizeStrategy(VPIntrinsic &VPI, VPLegalization &LegalizeStrat) {
590   // Operations with speculatable lanes do not strictly need predication.
591   if (maySpeculateLanes(VPI)) {
592     // Converting a speculatable VP intrinsic means dropping %mask and %evl.
593     // No need to expand %evl into the %mask only to ignore that code.
594     if (LegalizeStrat.OpStrategy == VPLegalization::Convert)
595       LegalizeStrat.EVLParamStrategy = VPLegalization::Discard;
596     return;
597   }
598 
599   // We have to preserve the predicating effect of %evl for this
600   // non-speculatable VP intrinsic.
601   // 1) Never discard %evl.
602   // 2) If this VP intrinsic will be expanded to non-VP code, make sure that
603   //    %evl gets folded into %mask.
604   if ((LegalizeStrat.EVLParamStrategy == VPLegalization::Discard) ||
605       (LegalizeStrat.OpStrategy == VPLegalization::Convert)) {
606     LegalizeStrat.EVLParamStrategy = VPLegalization::Convert;
607   }
608 }
609 
610 VPLegalization
611 CachingVPExpander::getVPLegalizationStrategy(const VPIntrinsic &VPI) const {
612   auto VPStrat = TTI.getVPLegalizationStrategy(VPI);
613   if (LLVM_LIKELY(!UsingTTIOverrides)) {
614     // No overrides - we are in production.
615     return VPStrat;
616   }
617 
618   // Overrides set - we are in testing, the following does not need to be
619   // efficient.
620   VPStrat.EVLParamStrategy = parseOverrideOption(EVLTransformOverride);
621   VPStrat.OpStrategy = parseOverrideOption(MaskTransformOverride);
622   return VPStrat;
623 }
624 
625 /// Expand llvm.vp.* intrinsics as requested by \p TTI.
626 bool CachingVPExpander::expandVectorPredication() {
627   SmallVector<TransformJob, 16> Worklist;
628 
629   // Collect all VPIntrinsics that need expansion and determine their expansion
630   // strategy.
631   for (auto &I : instructions(F)) {
632     auto *VPI = dyn_cast<VPIntrinsic>(&I);
633     if (!VPI)
634       continue;
635     auto VPStrat = getVPLegalizationStrategy(*VPI);
636     sanitizeStrategy(*VPI, VPStrat);
637     if (!VPStrat.shouldDoNothing())
638       Worklist.emplace_back(VPI, VPStrat);
639   }
640   if (Worklist.empty())
641     return false;
642 
643   // Transform all VPIntrinsics on the worklist.
644   LLVM_DEBUG(dbgs() << "\n:::: Transforming " << Worklist.size()
645                     << " instructions ::::\n");
646   for (TransformJob Job : Worklist) {
647     // Transform the EVL parameter.
648     switch (Job.Strategy.EVLParamStrategy) {
649     case VPLegalization::Legal:
650       break;
651     case VPLegalization::Discard:
652       discardEVLParameter(*Job.PI);
653       break;
654     case VPLegalization::Convert:
655       if (foldEVLIntoMask(*Job.PI))
656         ++NumFoldedVL;
657       break;
658     }
659     Job.Strategy.EVLParamStrategy = VPLegalization::Legal;
660 
661     // Replace with a non-predicated operation.
662     switch (Job.Strategy.OpStrategy) {
663     case VPLegalization::Legal:
664       break;
665     case VPLegalization::Discard:
666       llvm_unreachable("Invalid strategy for operators.");
667     case VPLegalization::Convert:
668       expandPredication(*Job.PI);
669       ++NumLoweredVPOps;
670       break;
671     }
672     Job.Strategy.OpStrategy = VPLegalization::Legal;
673 
674     assert(Job.isDone() && "incomplete transformation");
675   }
676 
677   return true;
678 }
679 class ExpandVectorPredication : public FunctionPass {
680 public:
681   static char ID;
682   ExpandVectorPredication() : FunctionPass(ID) {
683     initializeExpandVectorPredicationPass(*PassRegistry::getPassRegistry());
684   }
685 
686   bool runOnFunction(Function &F) override {
687     const auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
688     CachingVPExpander VPExpander(F, *TTI);
689     return VPExpander.expandVectorPredication();
690   }
691 
692   void getAnalysisUsage(AnalysisUsage &AU) const override {
693     AU.addRequired<TargetTransformInfoWrapperPass>();
694     AU.setPreservesCFG();
695   }
696 };
697 } // namespace
698 
699 char ExpandVectorPredication::ID;
700 INITIALIZE_PASS_BEGIN(ExpandVectorPredication, "expandvp",
701                       "Expand vector predication intrinsics", false, false)
702 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
703 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
704 INITIALIZE_PASS_END(ExpandVectorPredication, "expandvp",
705                     "Expand vector predication intrinsics", false, false)
706 
707 FunctionPass *llvm::createExpandVectorPredicationPass() {
708   return new ExpandVectorPredication();
709 }
710 
711 PreservedAnalyses
712 ExpandVectorPredicationPass::run(Function &F, FunctionAnalysisManager &AM) {
713   const auto &TTI = AM.getResult<TargetIRAnalysis>(F);
714   CachingVPExpander VPExpander(F, TTI);
715   if (!VPExpander.expandVectorPredication())
716     return PreservedAnalyses::all();
717   PreservedAnalyses PA;
718   PA.preserveSet<CFGAnalyses>();
719   return PA;
720 }
721