@@ -996,12 +996,15 @@ class LoopVectorizationCostModel {
996
996
// / Holds the maximum number of concurrent live intervals in the loop.
997
997
// / The key is ClassID of target-provided register class.
998
998
SmallMapVector<unsigned , unsigned , 4 > MaxLocalUsers;
999
- };
1000
999
1001
- // / \return Returns information about the register usages of the loop for the
1002
- // / given vectorization factors.
1003
- SmallVector<RegisterUsage, 8 >
1004
- calculateRegisterUsage (ArrayRef<ElementCount> VFs);
1000
+ // / Check if any of the tracked live intervals exceeds the number of
1001
+ // / available registers for the target.
1002
+ bool exceedsMaxNumRegs (const TargetTransformInfo &TTI) const {
1003
+ return any_of (MaxLocalUsers, [&TTI](auto &LU) {
1004
+ return LU.second > TTI.getNumberOfRegisters (LU.first );
1005
+ });
1006
+ }
1007
+ };
1005
1008
1006
1009
// / Collect values we want to ignore in the cost model.
1007
1010
void collectValuesToIgnore ();
@@ -4013,29 +4016,8 @@ ElementCount LoopVectorizationCostModel::getMaximizedVFForTarget(
4013
4016
auto MaxVectorElementCountMaxBW = ElementCount::get (
4014
4017
llvm::bit_floor (WidestRegister.getKnownMinValue () / SmallestType),
4015
4018
ComputeScalableMaxVF);
4016
- MaxVectorElementCountMaxBW = MinVF (MaxVectorElementCountMaxBW, MaxSafeVF);
4017
-
4018
- // Collect all viable vectorization factors larger than the default MaxVF
4019
- // (i.e. MaxVectorElementCount).
4020
- SmallVector<ElementCount, 8 > VFs;
4021
- for (ElementCount VS = MaxVectorElementCount * 2 ;
4022
- ElementCount::isKnownLE (VS, MaxVectorElementCountMaxBW); VS *= 2 )
4023
- VFs.push_back (VS);
4024
-
4025
- // For each VF calculate its register usage.
4026
- auto RUs = calculateRegisterUsage (VFs);
4027
-
4028
- // Select the largest VF which doesn't require more registers than existing
4029
- // ones.
4030
- for (int I = RUs.size () - 1 ; I >= 0 ; --I) {
4031
- const auto &MLU = RUs[I].MaxLocalUsers ;
4032
- if (all_of (MLU, [&](decltype (MLU.front ()) &LU) {
4033
- return LU.second <= TTI.getNumberOfRegisters (LU.first );
4034
- })) {
4035
- MaxVF = VFs[I];
4036
- break ;
4037
- }
4038
- }
4019
+ MaxVF = MinVF (MaxVectorElementCountMaxBW, MaxSafeVF);
4020
+
4039
4021
if (ElementCount MinVF =
4040
4022
TTI.getMinimumVF (SmallestType, ComputeScalableMaxVF)) {
4041
4023
if (ElementCount::isKnownLT (MaxVF, MinVF)) {
@@ -4360,6 +4342,15 @@ static bool hasReplicatorRegion(VPlan &Plan) {
4360
4342
}
4361
4343
4362
4344
#ifndef NDEBUG
4345
+ // / Estimate the register usage for \p Plan and vectorization factors in \p VFs
4346
+ // / by calculating the highest number of values that are live at a single
4347
+ // / location as a rough estimate. Returns the register usage for each VF in \p
4348
+ // / VFs.
4349
+ static SmallVector<LoopVectorizationCostModel::RegisterUsage, 8 >
4350
+ calculateRegisterUsage (VPlan &Plan, ArrayRef<ElementCount> VFs,
4351
+ const TargetTransformInfo &TTI,
4352
+ const SmallPtrSetImpl<const Value *> &ValuesToIgnore);
4353
+
4363
4354
VectorizationFactor LoopVectorizationPlanner::selectVectorizationFactor () {
4364
4355
InstructionCost ExpectedCost = CM.expectedCost (ElementCount::getFixed (1 ));
4365
4356
LLVM_DEBUG (dbgs () << " LV: Scalar loop costs: " << ExpectedCost << " .\n " );
@@ -4383,11 +4374,19 @@ VectorizationFactor LoopVectorizationPlanner::selectVectorizationFactor() {
4383
4374
}
4384
4375
4385
4376
for (auto &P : VPlans) {
4386
- for (ElementCount VF : P->vectorFactors ()) {
4377
+ ArrayRef<ElementCount> VFs (P->vectorFactors ().begin (),
4378
+ P->vectorFactors ().end ());
4379
+ auto RUs = ::calculateRegisterUsage (*P, VFs, TTI, CM.ValuesToIgnore );
4380
+ for (auto [VF, RU] : zip_equal (VFs, RUs)) {
4387
4381
// The cost for scalar VF=1 is already calculated, so ignore it.
4388
4382
if (VF.isScalar ())
4389
4383
continue ;
4390
4384
4385
+ // / Don't consider the VF if it exceeds the number of registers for the
4386
+ // / target.
4387
+ if (RU.exceedsMaxNumRegs (TTI))
4388
+ continue ;
4389
+
4391
4390
InstructionCost C = CM.expectedCost (VF);
4392
4391
4393
4392
// Add on other costs that are modelled in VPlan, but not in the legacy
@@ -4859,9 +4858,13 @@ calculateRegisterUsage(VPlan &Plan, ArrayRef<ElementCount> VFs,
4859
4858
isa<VPCanonicalIVPHIRecipe, VPReplicateRecipe, VPDerivedIVRecipe,
4860
4859
VPScalarIVStepsRecipe>(R) ||
4861
4860
(isa<VPInstruction>(R) &&
4862
- all_of (cast<VPSingleDefRecipe>(R)->users (), [&](VPUser *U) {
4863
- return cast<VPRecipeBase>(U)->usesScalars (R->getVPSingleValue ());
4864
- }))) {
4861
+ all_of (cast<VPSingleDefRecipe>(R)->users (),
4862
+ [&](VPUser *U) {
4863
+ return cast<VPRecipeBase>(U)->usesScalars (
4864
+ R->getVPSingleValue ());
4865
+ })) ||
4866
+ (isa<VPReductionPHIRecipe>(R) &&
4867
+ (cast<VPReductionPHIRecipe>(R))->isInLoop ())) {
4865
4868
unsigned ClassID = TTI.getRegisterClassForType (
4866
4869
false , TypeInfo.inferScalarType (R->getVPSingleValue ()));
4867
4870
// FIXME: The target might use more than one register for the type
@@ -5234,213 +5237,6 @@ LoopVectorizationCostModel::selectInterleaveCount(VPlan &Plan, ElementCount VF,
5234
5237
return 1 ;
5235
5238
}
5236
5239
5237
- SmallVector<LoopVectorizationCostModel::RegisterUsage, 8 >
5238
- LoopVectorizationCostModel::calculateRegisterUsage (ArrayRef<ElementCount> VFs) {
5239
- // This function calculates the register usage by measuring the highest number
5240
- // of values that are alive at a single location. Obviously, this is a very
5241
- // rough estimation. We scan the loop in a topological order in order and
5242
- // assign a number to each instruction. We use RPO to ensure that defs are
5243
- // met before their users. We assume that each instruction that has in-loop
5244
- // users starts an interval. We record every time that an in-loop value is
5245
- // used, so we have a list of the first and last occurrences of each
5246
- // instruction. Next, we transpose this data structure into a multi map that
5247
- // holds the list of intervals that *end* at a specific location. This multi
5248
- // map allows us to perform a linear search. We scan the instructions linearly
5249
- // and record each time that a new interval starts, by placing it in a set.
5250
- // If we find this value in the multi-map then we remove it from the set.
5251
- // The max register usage is the maximum size of the set.
5252
- // We also search for instructions that are defined outside the loop, but are
5253
- // used inside the loop. We need this number separately from the max-interval
5254
- // usage number because when we unroll, loop-invariant values do not take
5255
- // more registers.
5256
- LoopBlocksDFS DFS (TheLoop);
5257
- DFS.perform (LI);
5258
-
5259
- RegisterUsage RU;
5260
-
5261
- // Each 'key' in the map opens a new interval. The values
5262
- // of the map are the index of the 'last seen' usage of the
5263
- // instruction that is the key.
5264
- using IntervalMap = SmallDenseMap<Instruction *, unsigned , 16 >;
5265
-
5266
- // Maps instruction to its index.
5267
- SmallVector<Instruction *, 64 > IdxToInstr;
5268
- // Marks the end of each interval.
5269
- IntervalMap EndPoint;
5270
- // Saves the list of instruction indices that are used in the loop.
5271
- SmallPtrSet<Instruction *, 8 > Ends;
5272
- // Saves the list of values that are used in the loop but are defined outside
5273
- // the loop (not including non-instruction values such as arguments and
5274
- // constants).
5275
- SmallSetVector<Instruction *, 8 > LoopInvariants;
5276
-
5277
- for (BasicBlock *BB : make_range (DFS.beginRPO (), DFS.endRPO ())) {
5278
- for (Instruction &I : BB->instructionsWithoutDebug ()) {
5279
- IdxToInstr.push_back (&I);
5280
-
5281
- // Save the end location of each USE.
5282
- for (Value *U : I.operands ()) {
5283
- auto *Instr = dyn_cast<Instruction>(U);
5284
-
5285
- // Ignore non-instruction values such as arguments, constants, etc.
5286
- // FIXME: Might need some motivation why these values are ignored. If
5287
- // for example an argument is used inside the loop it will increase the
5288
- // register pressure (so shouldn't we add it to LoopInvariants).
5289
- if (!Instr)
5290
- continue ;
5291
-
5292
- // If this instruction is outside the loop then record it and continue.
5293
- if (!TheLoop->contains (Instr)) {
5294
- LoopInvariants.insert (Instr);
5295
- continue ;
5296
- }
5297
-
5298
- // Overwrite previous end points.
5299
- EndPoint[Instr] = IdxToInstr.size ();
5300
- Ends.insert (Instr);
5301
- }
5302
- }
5303
- }
5304
-
5305
- // Saves the list of intervals that end with the index in 'key'.
5306
- using InstrList = SmallVector<Instruction *, 2 >;
5307
- SmallDenseMap<unsigned , InstrList, 16 > TransposeEnds;
5308
-
5309
- // Transpose the EndPoints to a list of values that end at each index.
5310
- for (auto &Interval : EndPoint)
5311
- TransposeEnds[Interval.second ].push_back (Interval.first );
5312
-
5313
- SmallPtrSet<Instruction *, 8 > OpenIntervals;
5314
- SmallVector<RegisterUsage, 8 > RUs (VFs.size ());
5315
- SmallVector<SmallMapVector<unsigned , unsigned , 4 >, 8 > MaxUsages (VFs.size ());
5316
-
5317
- LLVM_DEBUG (dbgs () << " LV(REG): Calculating max register usage:\n " );
5318
-
5319
- const auto &TTICapture = TTI;
5320
- auto GetRegUsage = [&TTICapture](Type *Ty, ElementCount VF) -> unsigned {
5321
- if (Ty->isTokenTy () || !VectorType::isValidElementType (Ty) ||
5322
- (VF.isScalable () &&
5323
- !TTICapture.isElementTypeLegalForScalableVector (Ty)))
5324
- return 0 ;
5325
- return TTICapture.getRegUsageForType (VectorType::get (Ty, VF));
5326
- };
5327
-
5328
- collectInLoopReductions ();
5329
-
5330
- for (unsigned int Idx = 0 , Sz = IdxToInstr.size (); Idx < Sz; ++Idx) {
5331
- Instruction *I = IdxToInstr[Idx];
5332
-
5333
- // Remove all of the instructions that end at this location.
5334
- InstrList &List = TransposeEnds[Idx];
5335
- for (Instruction *ToRemove : List)
5336
- OpenIntervals.erase (ToRemove);
5337
-
5338
- // Ignore instructions that are never used within the loop and do not have
5339
- // side-effects.
5340
- if (!Ends.count (I) && !I->mayHaveSideEffects ())
5341
- continue ;
5342
-
5343
- // Skip ignored values.
5344
- if (ValuesToIgnore.count (I))
5345
- continue ;
5346
-
5347
- // For each VF find the maximum usage of registers.
5348
- for (unsigned J = 0 , E = VFs.size (); J < E; ++J) {
5349
- // Count the number of registers used, per register class, given all open
5350
- // intervals.
5351
- // Note that elements in this SmallMapVector will be default constructed
5352
- // as 0. So we can use "RegUsage[ClassID] += n" in the code below even if
5353
- // there is no previous entry for ClassID.
5354
- SmallMapVector<unsigned , unsigned , 4 > RegUsage;
5355
-
5356
- if (VFs[J].isScalar ()) {
5357
- for (auto *Inst : OpenIntervals) {
5358
- unsigned ClassID =
5359
- TTI.getRegisterClassForType (false , Inst->getType ());
5360
- // FIXME: The target might use more than one register for the type
5361
- // even in the scalar case.
5362
- RegUsage[ClassID] += 1 ;
5363
- }
5364
- } else {
5365
- collectNonVectorizedAndSetWideningDecisions (VFs[J]);
5366
- for (auto *Inst : OpenIntervals) {
5367
- // Skip ignored values for VF > 1.
5368
- if (VecValuesToIgnore.count (Inst))
5369
- continue ;
5370
- if (isScalarAfterVectorization (Inst, VFs[J])) {
5371
- unsigned ClassID =
5372
- TTI.getRegisterClassForType (false , Inst->getType ());
5373
- // FIXME: The target might use more than one register for the type
5374
- // even in the scalar case.
5375
- RegUsage[ClassID] += 1 ;
5376
- } else {
5377
- unsigned ClassID =
5378
- TTI.getRegisterClassForType (true , Inst->getType ());
5379
- RegUsage[ClassID] += GetRegUsage (Inst->getType (), VFs[J]);
5380
- }
5381
- }
5382
- }
5383
-
5384
- for (const auto &Pair : RegUsage) {
5385
- auto &Entry = MaxUsages[J][Pair.first ];
5386
- Entry = std::max (Entry, Pair.second );
5387
- }
5388
- }
5389
-
5390
- LLVM_DEBUG (dbgs () << " LV(REG): At #" << Idx << " Interval # "
5391
- << OpenIntervals.size () << ' \n ' );
5392
-
5393
- // Add the current instruction to the list of open intervals.
5394
- OpenIntervals.insert (I);
5395
- }
5396
-
5397
- for (unsigned Idx = 0 , End = VFs.size (); Idx < End; ++Idx) {
5398
- // Note that elements in this SmallMapVector will be default constructed
5399
- // as 0. So we can use "Invariant[ClassID] += n" in the code below even if
5400
- // there is no previous entry for ClassID.
5401
- SmallMapVector<unsigned , unsigned , 4 > Invariant;
5402
-
5403
- for (auto *Inst : LoopInvariants) {
5404
- // FIXME: The target might use more than one register for the type
5405
- // even in the scalar case.
5406
- bool IsScalar = all_of (Inst->users (), [&](User *U) {
5407
- auto *I = cast<Instruction>(U);
5408
- return TheLoop != LI->getLoopFor (I->getParent ()) ||
5409
- isScalarAfterVectorization (I, VFs[Idx]);
5410
- });
5411
-
5412
- ElementCount VF = IsScalar ? ElementCount::getFixed (1 ) : VFs[Idx];
5413
- unsigned ClassID =
5414
- TTI.getRegisterClassForType (VF.isVector (), Inst->getType ());
5415
- Invariant[ClassID] += GetRegUsage (Inst->getType (), VF);
5416
- }
5417
-
5418
- LLVM_DEBUG ({
5419
- dbgs () << " LV(REG): VF = " << VFs[Idx] << ' \n ' ;
5420
- dbgs () << " LV(REG): Found max usage: " << MaxUsages[Idx].size ()
5421
- << " item\n " ;
5422
- for (const auto &pair : MaxUsages[Idx]) {
5423
- dbgs () << " LV(REG): RegisterClass: "
5424
- << TTI.getRegisterClassName (pair.first ) << " , " << pair.second
5425
- << " registers\n " ;
5426
- }
5427
- dbgs () << " LV(REG): Found invariant usage: " << Invariant.size ()
5428
- << " item\n " ;
5429
- for (const auto &pair : Invariant) {
5430
- dbgs () << " LV(REG): RegisterClass: "
5431
- << TTI.getRegisterClassName (pair.first ) << " , " << pair.second
5432
- << " registers\n " ;
5433
- }
5434
- });
5435
-
5436
- RU.LoopInvariantRegs = Invariant;
5437
- RU.MaxLocalUsers = MaxUsages[Idx];
5438
- RUs[Idx] = RU;
5439
- }
5440
-
5441
- return RUs;
5442
- }
5443
-
5444
5240
bool LoopVectorizationCostModel::useEmulatedMaskMemRefHack (Instruction *I,
5445
5241
ElementCount VF) {
5446
5242
// TODO: Cost model for emulated masked load/store is completely
@@ -7621,7 +7417,10 @@ VectorizationFactor LoopVectorizationPlanner::computeBestVF() {
7621
7417
}
7622
7418
7623
7419
for (auto &P : VPlans) {
7624
- for (ElementCount VF : P->vectorFactors ()) {
7420
+ ArrayRef<ElementCount> VFs (P->vectorFactors ().begin (),
7421
+ P->vectorFactors ().end ());
7422
+ auto RUs = ::calculateRegisterUsage (*P, VFs, TTI, CM.ValuesToIgnore );
7423
+ for (auto [VF, RU] : zip_equal (VFs, RUs)) {
7625
7424
if (VF.isScalar ())
7626
7425
continue ;
7627
7426
if (!ForceVectorization && !willGenerateVectors (*P, VF, TTI)) {
@@ -7642,6 +7441,13 @@ VectorizationFactor LoopVectorizationPlanner::computeBestVF() {
7642
7441
7643
7442
InstructionCost Cost = cost (*P, VF);
7644
7443
VectorizationFactor CurrentFactor (VF, Cost, ScalarCost);
7444
+
7445
+ if (RU.exceedsMaxNumRegs (TTI)) {
7446
+ LLVM_DEBUG (dbgs () << " LV(REG): Not considering vector loop of width "
7447
+ << VF << " because it uses too many registers\n " );
7448
+ continue ;
7449
+ }
7450
+
7645
7451
if (isMoreProfitable (CurrentFactor, BestFactor, P->hasScalarTail ()))
7646
7452
BestFactor = CurrentFactor;
7647
7453
0 commit comments