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 
parseOverrideOption(const std::string & TextOpt)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.
anyExpandVPOverridesSet()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.
isAllTrueMask(Value * MaskVal)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.
getSafeDivisor(Type * DivTy)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.
transferDecorations(Value & NewVal,VPIntrinsic & VPI)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.
replaceOperation(Value & NewOp,VPIntrinsic & OldOp)115 static void replaceOperation(Value &NewOp, VPIntrinsic &OldOp) {
116   transferDecorations(NewOp, OldOp);
117   OldOp.replaceAllUsesWith(&NewOp);
118   OldOp.eraseFromParent();
119 }
120 
maySpeculateLanes(VPIntrinsic & VPI)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   if (auto IntrID = VPI.getFunctionalIntrinsicID())
127     return Intrinsic::getAttributes(VPI.getContext(), *IntrID)
128         .hasFnAttr(Attribute::AttrKind::Speculatable);
129   if (auto Opc = VPI.getFunctionalOpcode())
130     return isSafeToSpeculativelyExecuteWithOpcode(*Opc, &VPI);
131   return false;
132 }
133 
134 //// } Helpers
135 
136 namespace {
137 
138 // Expansion pass state at function scope.
139 struct CachingVPExpander {
140   Function &F;
141   const TargetTransformInfo &TTI;
142 
143   /// \returns A (fixed length) vector with ascending integer indices
144   /// (<0, 1, ..., NumElems-1>).
145   /// \p Builder
146   ///    Used for instruction creation.
147   /// \p LaneTy
148   ///    Integer element type of the result vector.
149   /// \p NumElems
150   ///    Number of vector elements.
151   Value *createStepVector(IRBuilder<> &Builder, Type *LaneTy,
152                           unsigned NumElems);
153 
154   /// \returns A bitmask that is true where the lane position is less-than \p
155   /// EVLParam
156   ///
157   /// \p Builder
158   ///    Used for instruction creation.
159   /// \p VLParam
160   ///    The explicit vector length parameter to test against the lane
161   ///    positions.
162   /// \p ElemCount
163   ///    Static (potentially scalable) number of vector elements.
164   Value *convertEVLToMask(IRBuilder<> &Builder, Value *EVLParam,
165                           ElementCount ElemCount);
166 
167   Value *foldEVLIntoMask(VPIntrinsic &VPI);
168 
169   /// "Remove" the %evl parameter of \p PI by setting it to the static vector
170   /// length of the operation.
171   void discardEVLParameter(VPIntrinsic &PI);
172 
173   /// Lower this VP binary operator to a unpredicated binary operator.
174   Value *expandPredicationInBinaryOperator(IRBuilder<> &Builder,
175                                            VPIntrinsic &PI);
176 
177   /// Lower this VP int call to a unpredicated int call.
178   Value *expandPredicationToIntCall(IRBuilder<> &Builder, VPIntrinsic &PI,
179                                     unsigned UnpredicatedIntrinsicID);
180 
181   /// Lower this VP fp call to a unpredicated fp call.
182   Value *expandPredicationToFPCall(IRBuilder<> &Builder, VPIntrinsic &PI,
183                                    unsigned UnpredicatedIntrinsicID);
184 
185   /// Lower this VP reduction to a call to an unpredicated reduction intrinsic.
186   Value *expandPredicationInReduction(IRBuilder<> &Builder,
187                                       VPReductionIntrinsic &PI);
188 
189   /// Lower this VP cast operation to a non-VP intrinsic.
190   Value *expandPredicationToCastIntrinsic(IRBuilder<> &Builder,
191                                           VPIntrinsic &VPI);
192 
193   /// Lower this VP memory operation to a non-VP intrinsic.
194   Value *expandPredicationInMemoryIntrinsic(IRBuilder<> &Builder,
195                                             VPIntrinsic &VPI);
196 
197   /// Lower this VP comparison to a call to an unpredicated comparison.
198   Value *expandPredicationInComparison(IRBuilder<> &Builder,
199                                        VPCmpIntrinsic &PI);
200 
201   /// Query TTI and expand the vector predication in \p P accordingly.
202   Value *expandPredication(VPIntrinsic &PI);
203 
204   /// Determine how and whether the VPIntrinsic \p VPI shall be expanded. This
205   /// overrides TTI with the cl::opts listed at the top of this file.
206   VPLegalization getVPLegalizationStrategy(const VPIntrinsic &VPI) const;
207   bool UsingTTIOverrides;
208 
209 public:
CachingVPExpander__anone938bf610111::CachingVPExpander210   CachingVPExpander(Function &F, const TargetTransformInfo &TTI)
211       : F(F), TTI(TTI), UsingTTIOverrides(anyExpandVPOverridesSet()) {}
212 
213   bool expandVectorPredication();
214 };
215 
216 //// CachingVPExpander {
217 
createStepVector(IRBuilder<> & Builder,Type * LaneTy,unsigned NumElems)218 Value *CachingVPExpander::createStepVector(IRBuilder<> &Builder, Type *LaneTy,
219                                            unsigned NumElems) {
220   // TODO add caching
221   SmallVector<Constant *, 16> ConstElems;
222 
223   for (unsigned Idx = 0; Idx < NumElems; ++Idx)
224     ConstElems.push_back(ConstantInt::get(LaneTy, Idx, false));
225 
226   return ConstantVector::get(ConstElems);
227 }
228 
convertEVLToMask(IRBuilder<> & Builder,Value * EVLParam,ElementCount ElemCount)229 Value *CachingVPExpander::convertEVLToMask(IRBuilder<> &Builder,
230                                            Value *EVLParam,
231                                            ElementCount ElemCount) {
232   // TODO add caching
233   // Scalable vector %evl conversion.
234   if (ElemCount.isScalable()) {
235     auto *M = Builder.GetInsertBlock()->getModule();
236     Type *BoolVecTy = VectorType::get(Builder.getInt1Ty(), ElemCount);
237     Function *ActiveMaskFunc = Intrinsic::getDeclaration(
238         M, Intrinsic::get_active_lane_mask, {BoolVecTy, EVLParam->getType()});
239     // `get_active_lane_mask` performs an implicit less-than comparison.
240     Value *ConstZero = Builder.getInt32(0);
241     return Builder.CreateCall(ActiveMaskFunc, {ConstZero, EVLParam});
242   }
243 
244   // Fixed vector %evl conversion.
245   Type *LaneTy = EVLParam->getType();
246   unsigned NumElems = ElemCount.getFixedValue();
247   Value *VLSplat = Builder.CreateVectorSplat(NumElems, EVLParam);
248   Value *IdxVec = createStepVector(Builder, LaneTy, NumElems);
249   return Builder.CreateICmp(CmpInst::ICMP_ULT, IdxVec, VLSplat);
250 }
251 
252 Value *
expandPredicationInBinaryOperator(IRBuilder<> & Builder,VPIntrinsic & VPI)253 CachingVPExpander::expandPredicationInBinaryOperator(IRBuilder<> &Builder,
254                                                      VPIntrinsic &VPI) {
255   assert((maySpeculateLanes(VPI) || VPI.canIgnoreVectorLengthParam()) &&
256          "Implicitly dropping %evl in non-speculatable operator!");
257 
258   auto OC = static_cast<Instruction::BinaryOps>(*VPI.getFunctionalOpcode());
259   assert(Instruction::isBinaryOp(OC));
260 
261   Value *Op0 = VPI.getOperand(0);
262   Value *Op1 = VPI.getOperand(1);
263   Value *Mask = VPI.getMaskParam();
264 
265   // Blend in safe operands.
266   if (Mask && !isAllTrueMask(Mask)) {
267     switch (OC) {
268     default:
269       // Can safely ignore the predicate.
270       break;
271 
272     // Division operators need a safe divisor on masked-off lanes (1).
273     case Instruction::UDiv:
274     case Instruction::SDiv:
275     case Instruction::URem:
276     case Instruction::SRem:
277       // 2nd operand must not be zero.
278       Value *SafeDivisor = getSafeDivisor(VPI.getType());
279       Op1 = Builder.CreateSelect(Mask, Op1, SafeDivisor);
280     }
281   }
282 
283   Value *NewBinOp = Builder.CreateBinOp(OC, Op0, Op1, VPI.getName());
284 
285   replaceOperation(*NewBinOp, VPI);
286   return NewBinOp;
287 }
288 
expandPredicationToIntCall(IRBuilder<> & Builder,VPIntrinsic & VPI,unsigned UnpredicatedIntrinsicID)289 Value *CachingVPExpander::expandPredicationToIntCall(
290     IRBuilder<> &Builder, VPIntrinsic &VPI, unsigned UnpredicatedIntrinsicID) {
291   switch (UnpredicatedIntrinsicID) {
292   case Intrinsic::abs:
293   case Intrinsic::smax:
294   case Intrinsic::smin:
295   case Intrinsic::umax:
296   case Intrinsic::umin: {
297     Value *Op0 = VPI.getOperand(0);
298     Value *Op1 = VPI.getOperand(1);
299     Function *Fn = Intrinsic::getDeclaration(
300         VPI.getModule(), UnpredicatedIntrinsicID, {VPI.getType()});
301     Value *NewOp = Builder.CreateCall(Fn, {Op0, Op1}, VPI.getName());
302     replaceOperation(*NewOp, VPI);
303     return NewOp;
304   }
305   case Intrinsic::bswap:
306   case Intrinsic::bitreverse: {
307     Value *Op = VPI.getOperand(0);
308     Function *Fn = Intrinsic::getDeclaration(
309         VPI.getModule(), UnpredicatedIntrinsicID, {VPI.getType()});
310     Value *NewOp = Builder.CreateCall(Fn, {Op}, VPI.getName());
311     replaceOperation(*NewOp, VPI);
312     return NewOp;
313   }
314   }
315   return nullptr;
316 }
317 
expandPredicationToFPCall(IRBuilder<> & Builder,VPIntrinsic & VPI,unsigned UnpredicatedIntrinsicID)318 Value *CachingVPExpander::expandPredicationToFPCall(
319     IRBuilder<> &Builder, VPIntrinsic &VPI, unsigned UnpredicatedIntrinsicID) {
320   assert((maySpeculateLanes(VPI) || VPI.canIgnoreVectorLengthParam()) &&
321          "Implicitly dropping %evl in non-speculatable operator!");
322 
323   switch (UnpredicatedIntrinsicID) {
324   case Intrinsic::fabs:
325   case Intrinsic::sqrt: {
326     Value *Op0 = VPI.getOperand(0);
327     Function *Fn = Intrinsic::getDeclaration(
328         VPI.getModule(), UnpredicatedIntrinsicID, {VPI.getType()});
329     Value *NewOp = Builder.CreateCall(Fn, {Op0}, VPI.getName());
330     replaceOperation(*NewOp, VPI);
331     return NewOp;
332   }
333   case Intrinsic::maxnum:
334   case Intrinsic::minnum: {
335     Value *Op0 = VPI.getOperand(0);
336     Value *Op1 = VPI.getOperand(1);
337     Function *Fn = Intrinsic::getDeclaration(
338         VPI.getModule(), UnpredicatedIntrinsicID, {VPI.getType()});
339     Value *NewOp = Builder.CreateCall(Fn, {Op0, Op1}, VPI.getName());
340     replaceOperation(*NewOp, VPI);
341     return NewOp;
342   }
343   case Intrinsic::experimental_constrained_fma:
344   case Intrinsic::experimental_constrained_fmuladd: {
345     Value *Op0 = VPI.getOperand(0);
346     Value *Op1 = VPI.getOperand(1);
347     Value *Op2 = VPI.getOperand(2);
348     Function *Fn = Intrinsic::getDeclaration(
349         VPI.getModule(), UnpredicatedIntrinsicID, {VPI.getType()});
350     Value *NewOp =
351         Builder.CreateConstrainedFPCall(Fn, {Op0, Op1, Op2}, VPI.getName());
352     replaceOperation(*NewOp, VPI);
353     return NewOp;
354   }
355   }
356 
357   return nullptr;
358 }
359 
getNeutralReductionElement(const VPReductionIntrinsic & VPI,Type * EltTy)360 static Value *getNeutralReductionElement(const VPReductionIntrinsic &VPI,
361                                          Type *EltTy) {
362   bool Negative = false;
363   unsigned EltBits = EltTy->getScalarSizeInBits();
364   switch (VPI.getIntrinsicID()) {
365   default:
366     llvm_unreachable("Expecting a VP reduction intrinsic");
367   case Intrinsic::vp_reduce_add:
368   case Intrinsic::vp_reduce_or:
369   case Intrinsic::vp_reduce_xor:
370   case Intrinsic::vp_reduce_umax:
371     return Constant::getNullValue(EltTy);
372   case Intrinsic::vp_reduce_mul:
373     return ConstantInt::get(EltTy, 1, /*IsSigned*/ false);
374   case Intrinsic::vp_reduce_and:
375   case Intrinsic::vp_reduce_umin:
376     return ConstantInt::getAllOnesValue(EltTy);
377   case Intrinsic::vp_reduce_smin:
378     return ConstantInt::get(EltTy->getContext(),
379                             APInt::getSignedMaxValue(EltBits));
380   case Intrinsic::vp_reduce_smax:
381     return ConstantInt::get(EltTy->getContext(),
382                             APInt::getSignedMinValue(EltBits));
383   case Intrinsic::vp_reduce_fmax:
384     Negative = true;
385     [[fallthrough]];
386   case Intrinsic::vp_reduce_fmin: {
387     FastMathFlags Flags = VPI.getFastMathFlags();
388     const fltSemantics &Semantics = EltTy->getFltSemantics();
389     return !Flags.noNaNs() ? ConstantFP::getQNaN(EltTy, Negative)
390            : !Flags.noInfs()
391                ? ConstantFP::getInfinity(EltTy, Negative)
392                : ConstantFP::get(EltTy,
393                                  APFloat::getLargest(Semantics, Negative));
394   }
395   case Intrinsic::vp_reduce_fadd:
396     return ConstantFP::getNegativeZero(EltTy);
397   case Intrinsic::vp_reduce_fmul:
398     return ConstantFP::get(EltTy, 1.0);
399   }
400 }
401 
402 Value *
expandPredicationInReduction(IRBuilder<> & Builder,VPReductionIntrinsic & VPI)403 CachingVPExpander::expandPredicationInReduction(IRBuilder<> &Builder,
404                                                 VPReductionIntrinsic &VPI) {
405   assert((maySpeculateLanes(VPI) || VPI.canIgnoreVectorLengthParam()) &&
406          "Implicitly dropping %evl in non-speculatable operator!");
407 
408   Value *Mask = VPI.getMaskParam();
409   Value *RedOp = VPI.getOperand(VPI.getVectorParamPos());
410 
411   // Insert neutral element in masked-out positions
412   if (Mask && !isAllTrueMask(Mask)) {
413     auto *NeutralElt = getNeutralReductionElement(VPI, VPI.getType());
414     auto *NeutralVector = Builder.CreateVectorSplat(
415         cast<VectorType>(RedOp->getType())->getElementCount(), NeutralElt);
416     RedOp = Builder.CreateSelect(Mask, RedOp, NeutralVector);
417   }
418 
419   Value *Reduction;
420   Value *Start = VPI.getOperand(VPI.getStartParamPos());
421 
422   switch (VPI.getIntrinsicID()) {
423   default:
424     llvm_unreachable("Impossible reduction kind");
425   case Intrinsic::vp_reduce_add:
426     Reduction = Builder.CreateAddReduce(RedOp);
427     Reduction = Builder.CreateAdd(Reduction, Start);
428     break;
429   case Intrinsic::vp_reduce_mul:
430     Reduction = Builder.CreateMulReduce(RedOp);
431     Reduction = Builder.CreateMul(Reduction, Start);
432     break;
433   case Intrinsic::vp_reduce_and:
434     Reduction = Builder.CreateAndReduce(RedOp);
435     Reduction = Builder.CreateAnd(Reduction, Start);
436     break;
437   case Intrinsic::vp_reduce_or:
438     Reduction = Builder.CreateOrReduce(RedOp);
439     Reduction = Builder.CreateOr(Reduction, Start);
440     break;
441   case Intrinsic::vp_reduce_xor:
442     Reduction = Builder.CreateXorReduce(RedOp);
443     Reduction = Builder.CreateXor(Reduction, Start);
444     break;
445   case Intrinsic::vp_reduce_smax:
446     Reduction = Builder.CreateIntMaxReduce(RedOp, /*IsSigned*/ true);
447     Reduction =
448         Builder.CreateBinaryIntrinsic(Intrinsic::smax, Reduction, Start);
449     break;
450   case Intrinsic::vp_reduce_smin:
451     Reduction = Builder.CreateIntMinReduce(RedOp, /*IsSigned*/ true);
452     Reduction =
453         Builder.CreateBinaryIntrinsic(Intrinsic::smin, Reduction, Start);
454     break;
455   case Intrinsic::vp_reduce_umax:
456     Reduction = Builder.CreateIntMaxReduce(RedOp, /*IsSigned*/ false);
457     Reduction =
458         Builder.CreateBinaryIntrinsic(Intrinsic::umax, Reduction, Start);
459     break;
460   case Intrinsic::vp_reduce_umin:
461     Reduction = Builder.CreateIntMinReduce(RedOp, /*IsSigned*/ false);
462     Reduction =
463         Builder.CreateBinaryIntrinsic(Intrinsic::umin, Reduction, Start);
464     break;
465   case Intrinsic::vp_reduce_fmax:
466     Reduction = Builder.CreateFPMaxReduce(RedOp);
467     transferDecorations(*Reduction, VPI);
468     Reduction =
469         Builder.CreateBinaryIntrinsic(Intrinsic::maxnum, Reduction, Start);
470     break;
471   case Intrinsic::vp_reduce_fmin:
472     Reduction = Builder.CreateFPMinReduce(RedOp);
473     transferDecorations(*Reduction, VPI);
474     Reduction =
475         Builder.CreateBinaryIntrinsic(Intrinsic::minnum, Reduction, Start);
476     break;
477   case Intrinsic::vp_reduce_fadd:
478     Reduction = Builder.CreateFAddReduce(Start, RedOp);
479     break;
480   case Intrinsic::vp_reduce_fmul:
481     Reduction = Builder.CreateFMulReduce(Start, RedOp);
482     break;
483   }
484 
485   replaceOperation(*Reduction, VPI);
486   return Reduction;
487 }
488 
expandPredicationToCastIntrinsic(IRBuilder<> & Builder,VPIntrinsic & VPI)489 Value *CachingVPExpander::expandPredicationToCastIntrinsic(IRBuilder<> &Builder,
490                                                            VPIntrinsic &VPI) {
491   Value *CastOp = nullptr;
492   switch (VPI.getIntrinsicID()) {
493   default:
494     llvm_unreachable("Not a VP cast intrinsic");
495   case Intrinsic::vp_sext:
496     CastOp =
497         Builder.CreateSExt(VPI.getOperand(0), VPI.getType(), VPI.getName());
498     break;
499   case Intrinsic::vp_zext:
500     CastOp =
501         Builder.CreateZExt(VPI.getOperand(0), VPI.getType(), VPI.getName());
502     break;
503   case Intrinsic::vp_trunc:
504     CastOp =
505         Builder.CreateTrunc(VPI.getOperand(0), VPI.getType(), VPI.getName());
506     break;
507   case Intrinsic::vp_inttoptr:
508     CastOp =
509         Builder.CreateIntToPtr(VPI.getOperand(0), VPI.getType(), VPI.getName());
510     break;
511   case Intrinsic::vp_ptrtoint:
512     CastOp =
513         Builder.CreatePtrToInt(VPI.getOperand(0), VPI.getType(), VPI.getName());
514     break;
515   case Intrinsic::vp_fptosi:
516     CastOp =
517         Builder.CreateFPToSI(VPI.getOperand(0), VPI.getType(), VPI.getName());
518     break;
519 
520   case Intrinsic::vp_fptoui:
521     CastOp =
522         Builder.CreateFPToUI(VPI.getOperand(0), VPI.getType(), VPI.getName());
523     break;
524   case Intrinsic::vp_sitofp:
525     CastOp =
526         Builder.CreateSIToFP(VPI.getOperand(0), VPI.getType(), VPI.getName());
527     break;
528   case Intrinsic::vp_uitofp:
529     CastOp =
530         Builder.CreateUIToFP(VPI.getOperand(0), VPI.getType(), VPI.getName());
531     break;
532   case Intrinsic::vp_fptrunc:
533     CastOp =
534         Builder.CreateFPTrunc(VPI.getOperand(0), VPI.getType(), VPI.getName());
535     break;
536   case Intrinsic::vp_fpext:
537     CastOp =
538         Builder.CreateFPExt(VPI.getOperand(0), VPI.getType(), VPI.getName());
539     break;
540   }
541   replaceOperation(*CastOp, VPI);
542   return CastOp;
543 }
544 
545 Value *
expandPredicationInMemoryIntrinsic(IRBuilder<> & Builder,VPIntrinsic & VPI)546 CachingVPExpander::expandPredicationInMemoryIntrinsic(IRBuilder<> &Builder,
547                                                       VPIntrinsic &VPI) {
548   assert(VPI.canIgnoreVectorLengthParam());
549 
550   const auto &DL = F.getParent()->getDataLayout();
551 
552   Value *MaskParam = VPI.getMaskParam();
553   Value *PtrParam = VPI.getMemoryPointerParam();
554   Value *DataParam = VPI.getMemoryDataParam();
555   bool IsUnmasked = isAllTrueMask(MaskParam);
556 
557   MaybeAlign AlignOpt = VPI.getPointerAlignment();
558 
559   Value *NewMemoryInst = nullptr;
560   switch (VPI.getIntrinsicID()) {
561   default:
562     llvm_unreachable("Not a VP memory intrinsic");
563   case Intrinsic::vp_store:
564     if (IsUnmasked) {
565       StoreInst *NewStore =
566           Builder.CreateStore(DataParam, PtrParam, /*IsVolatile*/ false);
567       if (AlignOpt.has_value())
568         NewStore->setAlignment(*AlignOpt);
569       NewMemoryInst = NewStore;
570     } else
571       NewMemoryInst = Builder.CreateMaskedStore(
572           DataParam, PtrParam, AlignOpt.valueOrOne(), MaskParam);
573 
574     break;
575   case Intrinsic::vp_load:
576     if (IsUnmasked) {
577       LoadInst *NewLoad =
578           Builder.CreateLoad(VPI.getType(), PtrParam, /*IsVolatile*/ false);
579       if (AlignOpt.has_value())
580         NewLoad->setAlignment(*AlignOpt);
581       NewMemoryInst = NewLoad;
582     } else
583       NewMemoryInst = Builder.CreateMaskedLoad(
584           VPI.getType(), PtrParam, AlignOpt.valueOrOne(), MaskParam);
585 
586     break;
587   case Intrinsic::vp_scatter: {
588     auto *ElementType =
589         cast<VectorType>(DataParam->getType())->getElementType();
590     NewMemoryInst = Builder.CreateMaskedScatter(
591         DataParam, PtrParam,
592         AlignOpt.value_or(DL.getPrefTypeAlign(ElementType)), MaskParam);
593     break;
594   }
595   case Intrinsic::vp_gather: {
596     auto *ElementType = cast<VectorType>(VPI.getType())->getElementType();
597     NewMemoryInst = Builder.CreateMaskedGather(
598         VPI.getType(), PtrParam,
599         AlignOpt.value_or(DL.getPrefTypeAlign(ElementType)), MaskParam, nullptr,
600         VPI.getName());
601     break;
602   }
603   }
604 
605   assert(NewMemoryInst);
606   replaceOperation(*NewMemoryInst, VPI);
607   return NewMemoryInst;
608 }
609 
expandPredicationInComparison(IRBuilder<> & Builder,VPCmpIntrinsic & VPI)610 Value *CachingVPExpander::expandPredicationInComparison(IRBuilder<> &Builder,
611                                                         VPCmpIntrinsic &VPI) {
612   assert((maySpeculateLanes(VPI) || VPI.canIgnoreVectorLengthParam()) &&
613          "Implicitly dropping %evl in non-speculatable operator!");
614 
615   assert(*VPI.getFunctionalOpcode() == Instruction::ICmp ||
616          *VPI.getFunctionalOpcode() == Instruction::FCmp);
617 
618   Value *Op0 = VPI.getOperand(0);
619   Value *Op1 = VPI.getOperand(1);
620   auto Pred = VPI.getPredicate();
621 
622   auto *NewCmp = Builder.CreateCmp(Pred, Op0, Op1);
623 
624   replaceOperation(*NewCmp, VPI);
625   return NewCmp;
626 }
627 
discardEVLParameter(VPIntrinsic & VPI)628 void CachingVPExpander::discardEVLParameter(VPIntrinsic &VPI) {
629   LLVM_DEBUG(dbgs() << "Discard EVL parameter in " << VPI << "\n");
630 
631   if (VPI.canIgnoreVectorLengthParam())
632     return;
633 
634   Value *EVLParam = VPI.getVectorLengthParam();
635   if (!EVLParam)
636     return;
637 
638   ElementCount StaticElemCount = VPI.getStaticVectorLength();
639   Value *MaxEVL = nullptr;
640   Type *Int32Ty = Type::getInt32Ty(VPI.getContext());
641   if (StaticElemCount.isScalable()) {
642     // TODO add caching
643     auto *M = VPI.getModule();
644     Function *VScaleFunc =
645         Intrinsic::getDeclaration(M, Intrinsic::vscale, Int32Ty);
646     IRBuilder<> Builder(VPI.getParent(), VPI.getIterator());
647     Value *FactorConst = Builder.getInt32(StaticElemCount.getKnownMinValue());
648     Value *VScale = Builder.CreateCall(VScaleFunc, {}, "vscale");
649     MaxEVL = Builder.CreateMul(VScale, FactorConst, "scalable_size",
650                                /*NUW*/ true, /*NSW*/ false);
651   } else {
652     MaxEVL = ConstantInt::get(Int32Ty, StaticElemCount.getFixedValue(), false);
653   }
654   VPI.setVectorLengthParam(MaxEVL);
655 }
656 
foldEVLIntoMask(VPIntrinsic & VPI)657 Value *CachingVPExpander::foldEVLIntoMask(VPIntrinsic &VPI) {
658   LLVM_DEBUG(dbgs() << "Folding vlen for " << VPI << '\n');
659 
660   IRBuilder<> Builder(&VPI);
661 
662   // Ineffective %evl parameter and so nothing to do here.
663   if (VPI.canIgnoreVectorLengthParam())
664     return &VPI;
665 
666   // Only VP intrinsics can have an %evl parameter.
667   Value *OldMaskParam = VPI.getMaskParam();
668   Value *OldEVLParam = VPI.getVectorLengthParam();
669   assert(OldMaskParam && "no mask param to fold the vl param into");
670   assert(OldEVLParam && "no EVL param to fold away");
671 
672   LLVM_DEBUG(dbgs() << "OLD evl: " << *OldEVLParam << '\n');
673   LLVM_DEBUG(dbgs() << "OLD mask: " << *OldMaskParam << '\n');
674 
675   // Convert the %evl predication into vector mask predication.
676   ElementCount ElemCount = VPI.getStaticVectorLength();
677   Value *VLMask = convertEVLToMask(Builder, OldEVLParam, ElemCount);
678   Value *NewMaskParam = Builder.CreateAnd(VLMask, OldMaskParam);
679   VPI.setMaskParam(NewMaskParam);
680 
681   // Drop the %evl parameter.
682   discardEVLParameter(VPI);
683   assert(VPI.canIgnoreVectorLengthParam() &&
684          "transformation did not render the evl param ineffective!");
685 
686   // Reassess the modified instruction.
687   return &VPI;
688 }
689 
expandPredication(VPIntrinsic & VPI)690 Value *CachingVPExpander::expandPredication(VPIntrinsic &VPI) {
691   LLVM_DEBUG(dbgs() << "Lowering to unpredicated op: " << VPI << '\n');
692 
693   IRBuilder<> Builder(&VPI);
694 
695   // Try lowering to a LLVM instruction first.
696   auto OC = VPI.getFunctionalOpcode();
697 
698   if (OC && Instruction::isBinaryOp(*OC))
699     return expandPredicationInBinaryOperator(Builder, VPI);
700 
701   if (auto *VPRI = dyn_cast<VPReductionIntrinsic>(&VPI))
702     return expandPredicationInReduction(Builder, *VPRI);
703 
704   if (auto *VPCmp = dyn_cast<VPCmpIntrinsic>(&VPI))
705     return expandPredicationInComparison(Builder, *VPCmp);
706 
707   if (VPCastIntrinsic::isVPCast(VPI.getIntrinsicID())) {
708     return expandPredicationToCastIntrinsic(Builder, VPI);
709   }
710 
711   switch (VPI.getIntrinsicID()) {
712   default:
713     break;
714   case Intrinsic::vp_fneg: {
715     Value *NewNegOp = Builder.CreateFNeg(VPI.getOperand(0), VPI.getName());
716     replaceOperation(*NewNegOp, VPI);
717     return NewNegOp;
718   }
719   case Intrinsic::vp_abs:
720   case Intrinsic::vp_smax:
721   case Intrinsic::vp_smin:
722   case Intrinsic::vp_umax:
723   case Intrinsic::vp_umin:
724   case Intrinsic::vp_bswap:
725   case Intrinsic::vp_bitreverse:
726     return expandPredicationToIntCall(Builder, VPI,
727                                       VPI.getFunctionalIntrinsicID().value());
728   case Intrinsic::vp_fabs:
729   case Intrinsic::vp_sqrt:
730   case Intrinsic::vp_maxnum:
731   case Intrinsic::vp_minnum:
732   case Intrinsic::vp_maximum:
733   case Intrinsic::vp_minimum:
734     return expandPredicationToFPCall(Builder, VPI,
735                                      VPI.getFunctionalIntrinsicID().value());
736   case Intrinsic::vp_load:
737   case Intrinsic::vp_store:
738   case Intrinsic::vp_gather:
739   case Intrinsic::vp_scatter:
740     return expandPredicationInMemoryIntrinsic(Builder, VPI);
741   }
742 
743   if (auto CID = VPI.getConstrainedIntrinsicID())
744     if (Value *Call = expandPredicationToFPCall(Builder, VPI, *CID))
745       return Call;
746 
747   return &VPI;
748 }
749 
750 //// } CachingVPExpander
751 
752 struct TransformJob {
753   VPIntrinsic *PI;
754   TargetTransformInfo::VPLegalization Strategy;
TransformJob__anone938bf610111::TransformJob755   TransformJob(VPIntrinsic *PI, TargetTransformInfo::VPLegalization InitStrat)
756       : PI(PI), Strategy(InitStrat) {}
757 
isDone__anone938bf610111::TransformJob758   bool isDone() const { return Strategy.shouldDoNothing(); }
759 };
760 
sanitizeStrategy(VPIntrinsic & VPI,VPLegalization & LegalizeStrat)761 void sanitizeStrategy(VPIntrinsic &VPI, VPLegalization &LegalizeStrat) {
762   // Operations with speculatable lanes do not strictly need predication.
763   if (maySpeculateLanes(VPI)) {
764     // Converting a speculatable VP intrinsic means dropping %mask and %evl.
765     // No need to expand %evl into the %mask only to ignore that code.
766     if (LegalizeStrat.OpStrategy == VPLegalization::Convert)
767       LegalizeStrat.EVLParamStrategy = VPLegalization::Discard;
768     return;
769   }
770 
771   // We have to preserve the predicating effect of %evl for this
772   // non-speculatable VP intrinsic.
773   // 1) Never discard %evl.
774   // 2) If this VP intrinsic will be expanded to non-VP code, make sure that
775   //    %evl gets folded into %mask.
776   if ((LegalizeStrat.EVLParamStrategy == VPLegalization::Discard) ||
777       (LegalizeStrat.OpStrategy == VPLegalization::Convert)) {
778     LegalizeStrat.EVLParamStrategy = VPLegalization::Convert;
779   }
780 }
781 
782 VPLegalization
getVPLegalizationStrategy(const VPIntrinsic & VPI) const783 CachingVPExpander::getVPLegalizationStrategy(const VPIntrinsic &VPI) const {
784   auto VPStrat = TTI.getVPLegalizationStrategy(VPI);
785   if (LLVM_LIKELY(!UsingTTIOverrides)) {
786     // No overrides - we are in production.
787     return VPStrat;
788   }
789 
790   // Overrides set - we are in testing, the following does not need to be
791   // efficient.
792   VPStrat.EVLParamStrategy = parseOverrideOption(EVLTransformOverride);
793   VPStrat.OpStrategy = parseOverrideOption(MaskTransformOverride);
794   return VPStrat;
795 }
796 
797 /// Expand llvm.vp.* intrinsics as requested by \p TTI.
expandVectorPredication()798 bool CachingVPExpander::expandVectorPredication() {
799   SmallVector<TransformJob, 16> Worklist;
800 
801   // Collect all VPIntrinsics that need expansion and determine their expansion
802   // strategy.
803   for (auto &I : instructions(F)) {
804     auto *VPI = dyn_cast<VPIntrinsic>(&I);
805     if (!VPI)
806       continue;
807     auto VPStrat = getVPLegalizationStrategy(*VPI);
808     sanitizeStrategy(*VPI, VPStrat);
809     if (!VPStrat.shouldDoNothing())
810       Worklist.emplace_back(VPI, VPStrat);
811   }
812   if (Worklist.empty())
813     return false;
814 
815   // Transform all VPIntrinsics on the worklist.
816   LLVM_DEBUG(dbgs() << "\n:::: Transforming " << Worklist.size()
817                     << " instructions ::::\n");
818   for (TransformJob Job : Worklist) {
819     // Transform the EVL parameter.
820     switch (Job.Strategy.EVLParamStrategy) {
821     case VPLegalization::Legal:
822       break;
823     case VPLegalization::Discard:
824       discardEVLParameter(*Job.PI);
825       break;
826     case VPLegalization::Convert:
827       if (foldEVLIntoMask(*Job.PI))
828         ++NumFoldedVL;
829       break;
830     }
831     Job.Strategy.EVLParamStrategy = VPLegalization::Legal;
832 
833     // Replace with a non-predicated operation.
834     switch (Job.Strategy.OpStrategy) {
835     case VPLegalization::Legal:
836       break;
837     case VPLegalization::Discard:
838       llvm_unreachable("Invalid strategy for operators.");
839     case VPLegalization::Convert:
840       expandPredication(*Job.PI);
841       ++NumLoweredVPOps;
842       break;
843     }
844     Job.Strategy.OpStrategy = VPLegalization::Legal;
845 
846     assert(Job.isDone() && "incomplete transformation");
847   }
848 
849   return true;
850 }
851 class ExpandVectorPredication : public FunctionPass {
852 public:
853   static char ID;
ExpandVectorPredication()854   ExpandVectorPredication() : FunctionPass(ID) {
855     initializeExpandVectorPredicationPass(*PassRegistry::getPassRegistry());
856   }
857 
runOnFunction(Function & F)858   bool runOnFunction(Function &F) override {
859     const auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
860     CachingVPExpander VPExpander(F, *TTI);
861     return VPExpander.expandVectorPredication();
862   }
863 
getAnalysisUsage(AnalysisUsage & AU) const864   void getAnalysisUsage(AnalysisUsage &AU) const override {
865     AU.addRequired<TargetTransformInfoWrapperPass>();
866     AU.setPreservesCFG();
867   }
868 };
869 } // namespace
870 
871 char ExpandVectorPredication::ID;
872 INITIALIZE_PASS_BEGIN(ExpandVectorPredication, "expandvp",
873                       "Expand vector predication intrinsics", false, false)
INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)874 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
875 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
876 INITIALIZE_PASS_END(ExpandVectorPredication, "expandvp",
877                     "Expand vector predication intrinsics", false, false)
878 
879 FunctionPass *llvm::createExpandVectorPredicationPass() {
880   return new ExpandVectorPredication();
881 }
882 
883 PreservedAnalyses
run(Function & F,FunctionAnalysisManager & AM)884 ExpandVectorPredicationPass::run(Function &F, FunctionAnalysisManager &AM) {
885   const auto &TTI = AM.getResult<TargetIRAnalysis>(F);
886   CachingVPExpander VPExpander(F, TTI);
887   if (!VPExpander.expandVectorPredication())
888     return PreservedAnalyses::all();
889   PreservedAnalyses PA;
890   PA.preserveSet<CFGAnalyses>();
891   return PA;
892 }
893