summaryrefslogtreecommitdiff
path: root/backend/src/llvm/llvm_scalarize.cpp
blob: f7ef37520ae928a702bf74e314690605ed717aa8 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
/*
 * Copyright © 2012 Intel Corporation
 *
 * This library is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2 of the License, or (at your option) any later version.
 *
 * This library is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with this library. If not, see <http://www.gnu.org/licenses/>.
 */

/**
 * \file llvm_scalarize.cpp
 * \author Yang Rong <rong.r.yang@intel.com>
 *
 * This file is derived from:
 *  https://code.google.com/p/lunarglass/source/browse/trunk/Core/Passes/Transforms/Scalarize.cpp?r=605
 */

//===- Scalarize.cpp - Scalarize LunarGLASS IR ----------------------------===//
//
// LunarGLASS: An Open Modular Shader Compiler Architecture
// Copyright (C) 2010-2011 LunarG, Inc.
//
// This program is free software; you can redistribute it and/or
// modify it under the terms of the GNU General Public License
// as published by the Free Software Foundation; version 2 of the
// License.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with this program; if not, write to the Free Software
// Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
// 02110-1301, USA.
//
//===----------------------------------------------------------------------===//
//
// Author: Michael Ilseman, LunarG
//
//===----------------------------------------------------------------------===//
//
// Scalarize the IR.
//   * Loads of uniforms become multiple loadComponent calls
//
//   * Reads/writes become read/writeComponent calls
//
//   * Component-wise operations become multiple ops over each component
//
//   * Texture call become recomponsed texture calls
//
//   * Vector ops disappear, with their users referring to the scalarized
//   * components
//
//===----------------------------------------------------------------------===//

#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/PostOrderIterator.h"
#include "llvm/Function.h"
#include "llvm/InstrTypes.h"
#include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/Module.h"
#include "llvm/Pass.h"
#include "llvm/IRBuilder.h"
#include "llvm/Support/CallSite.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/raw_ostream.h"

#include "llvm/llvm_gen_backend.hpp"
#include "ir/unit.hpp"
#include "sys/map.hpp"


using namespace llvm;

namespace gbe {

  struct VectorValues {
    VectorValues() : vals()
    { }

    void setComponent(int c, llvm::Value* val)
    {
      assert(c >= 0 && c < 16 && "Out of bounds component");
      vals[c] = val;
    }
    llvm::Value* getComponent(int c)
    {
      assert(c >= 0 && c < 16 && "Out of bounds component");
      assert(vals[c] && "Requesting non-existing component");
      return vals[c];
    }

    // {Value* x, Value* y, Value* z, Value* w}
    llvm::Value* vals[16];
  };

  class Scalarize : public FunctionPass {

  public:
    // Standard pass stuff
    static char ID;

    Scalarize(ir::Unit& unit) : FunctionPass(ID), unit(unit)
    {
      initializeLoopInfoPass(*PassRegistry::getPassRegistry());
      initializeDominatorTreePass(*PassRegistry::getPassRegistry());
    }

    virtual bool runOnFunction(Function&);
    void print(raw_ostream&, const Module* = 0) const;
    virtual void getAnalysisUsage(AnalysisUsage&) const;

  protected:
    // An instruction is valid post-scalarization iff it is fully scalar or it
    // is a gla_loadn
    bool isValid(const Instruction*);

    // Take an instruction that produces a vector, and scalarize it
    bool scalarize(Instruction*);
    bool scalarizePerComponent(Instruction*);
    bool scalarizeFuncCall(CallInst *);
    bool scalarizeLoad(LoadInst*);
    bool scalarizeStore(StoreInst*);
    //bool scalarizeIntrinsic(IntrinsicInst*);
    bool scalarizeExtract(ExtractElementInst*);
    bool scalarizeInsert(InsertElementInst*);
    bool scalarizeShuffleVector(ShuffleVectorInst*);
    bool scalarizePHI(PHINode*);
    void scalarizeArgs(Function& F);
    // ...

    // Helpers to make the actual multiple scalar calls, one per
    // component. Updates the given VectorValues's components with the new
    // Values.
    void makeScalarizedCalls(Function*, ArrayRef<Value*>, int numComponents, VectorValues&);

    void makePerComponentScalarizedCalls(Instruction*, ArrayRef<Value*>);

    // Makes a scalar form of the given instruction: replaces the operands
    // and chooses a correct return type
    Instruction* createScalarInstruction(Instruction* inst, ArrayRef<Value*>);

    // Gather the specified components in the given values. Returns the
    // component if the given value is a vector, or the scalar itself.
    void gatherComponents(int component, ArrayRef<Value*> args, SmallVectorImpl<Value*>& componentArgs);

    // Get the assigned component for that value. If the value is a scalar,
    // returns the scalar. If it's a constant, returns that component. If
    // it's an instruction, returns the vectorValues of that instruction for
    // that component
    Value* getComponent(int component, Value*);

    // Used for assertion purposes. Whether we can get the component out with
    // a getComponent call
    bool canGetComponent(Value*);

    // Used for assertion purposes. Whether for every operand we can get
    // components with a getComponent call
    bool canGetComponentArgs(User*);

    // Delete the instruction in the deadList
    void dce();


    int GetConstantInt(const Value* value);
    bool IsPerComponentOp(const Instruction* inst);
    bool IsPerComponentOp(const Value* value);

    //these function used to add extract and insert instructions when load/store etc.
    void extractFromeVector(Value* insn);
    Value* InsertToVector(Value* insn, Value* vecValue);

    Type* GetBasicType(Value* value) {
      return GetBasicType(value->getType());
    }

    Type* GetBasicType(Type* type) {
      switch(type->getTypeID()) {
      case Type::VectorTyID:
      case Type::ArrayTyID:
        return GetBasicType(type->getContainedType(0));
      default:
        break;
      }
      return type;
    }

    int GetComponentCount(const Type* type)  {
      if (type->getTypeID() == Type::VectorTyID)
        return llvm::dyn_cast<VectorType>(type)->getNumElements();
      else
        return 1;
    }

    int GetComponentCount(const Value* value) {
      return GetComponentCount(value->getType());
    }

    DenseMap<Value*, VectorValues> vectorVals;
    Module* module;
    IRBuilder<>* builder;

    Type* intTy;
    Type* floatTy;
    ir::Unit &unit;

    std::vector<Instruction*> deadList;

    // List of vector phis that were not completely scalarized because some
    // of their operands hadn't before been visited (i.e. loop variant
    // variables)
    SmallVector<PHINode*, 16> incompletePhis;
  };

  Value* Scalarize::getComponent(int component, Value* v)
  {
    assert(canGetComponent(v) && "getComponent called on unhandled vector");

    if (v->getType()->isVectorTy()) {
      if (ConstantDataVector* c = dyn_cast<ConstantDataVector>(v)) {
        return c->getElementAsConstant(component);
      } else if (ConstantVector* c = dyn_cast<ConstantVector>(v)) {
        return c->getOperand(component);
      } else if (isa<ConstantAggregateZero>(v)) {
        return Constant::getNullValue(GetBasicType(v));
      } else if (isa<UndefValue>(v)) {
        return UndefValue::get(GetBasicType(v));
      } else {
        return vectorVals[v].getComponent(component);
      }
    } else {
      return v;
    }
  }

  bool IsPerComponentOp(const llvm::Value* value)
  {
    const llvm::Instruction* inst = llvm::dyn_cast<const llvm::Instruction>(value);
    return inst && IsPerComponentOp(inst);
  }

  bool Scalarize::IsPerComponentOp(const Instruction* inst)
  {
    //if (const IntrinsicInst* intr = dyn_cast<const IntrinsicInst>(inst))
    //    return IsPerComponentOp(intr);

    if (inst->isTerminator())
        return false;

    switch (inst->getOpcode()) {

    // Cast ops are only per-component if they cast back to the same vector
    // width
    case Instruction::Trunc:
    case Instruction::ZExt:
    case Instruction::SExt:
    case Instruction::FPToUI:
    case Instruction::FPToSI:
    case Instruction::UIToFP:
    case Instruction::SIToFP:
    case Instruction::FPTrunc:
    case Instruction::FPExt:
    case Instruction::PtrToInt:
    case Instruction::IntToPtr:
    case Instruction::BitCast:
      return GetComponentCount(inst->getOperand(0)) == GetComponentCount(inst);

    // Vector ops
    case Instruction::InsertElement:
    case Instruction::ExtractElement:
    case Instruction::ShuffleVector:

    // Ways of accessing/loading/storing vectors
    case Instruction::ExtractValue:
    case Instruction::InsertValue:

    // Memory ops
    case Instruction::Alloca:
    case Instruction::Load:
    case Instruction::Store:
    case Instruction::GetElementPtr:
    // Phis are a little special. We consider them not to be per-component
    // because the mechanism of choice is a single value (what path we took to
    // get here), and doesn't choose per-component (as select would). The caller
    // should know to handle phis specially
    case Instruction::PHI:
    // Call insts, conservatively are no per-component
    case Instruction::Call:
    // Misc
    case Instruction::LandingPad:  //--- 3.0
    case Instruction::VAArg:
      return false;
    } // end of switch (inst->getOpcode())

    return true;
  }
  int Scalarize::GetConstantInt(const Value* value)
  {
    const ConstantInt *constantInt = dyn_cast<ConstantInt>(value);

    // this might still be a constant expression, rather than a numeric constant,
    // e.g., expression with undef's in it, so it was not folded
    if (! constantInt)
      NOT_IMPLEMENTED; //gla::UnsupportedFunctionality("non-simple constant");

    return constantInt->getValue().getSExtValue();
  }
  bool Scalarize::canGetComponent(Value* v)
  {
    if (v->getType()->isVectorTy()) {
      if (isa<ConstantDataVector>(v) || isa<ConstantVector>(v) || isa<ConstantAggregateZero>(v) || isa<UndefValue>(v)) {
        return true;
      } else {
        assert((isa<Instruction>(v) || isa<Argument>(v)) && "Non-constant non-instuction?");
        return vectorVals.count(v);
      }
    } else {
      return true;
    }
  }

  bool Scalarize::canGetComponentArgs(User* u)
  {
    if (PHINode* phi = dyn_cast<PHINode>(u)) {
      for (unsigned int i = 0; i < phi->getNumIncomingValues(); ++i)
        if (!canGetComponent(phi->getIncomingValue(i)))
          return false;
    } else {
      for (User::op_iterator i = u->op_begin(), e = u->op_end(); i != e; ++i)
        if (!canGetComponent(*i))
          return false;
    }
    return true;
  }

  void Scalarize::gatherComponents(int component, ArrayRef<Value*> args, SmallVectorImpl<Value*>& componentArgs)
  {
    componentArgs.clear();
    for (ArrayRef<Value*>::iterator i = args.begin(), e = args.end(); i != e; ++i)
      componentArgs.push_back(getComponent(component, *i));
  }

  Instruction* Scalarize::createScalarInstruction(Instruction* inst, ArrayRef<Value*> args)
  {
    // TODO: Refine the below into one large switch

    unsigned op = inst->getOpcode();
    if (inst->isCast()) {
      assert(args.size() == 1 && "incorrect number of arguments for cast op");
      return CastInst::Create((Instruction::CastOps)op, args[0], GetBasicType(inst));
    }

    if (inst->isBinaryOp()) {
      assert(args.size() == 2 && "incorrect number of arguments for binary op");
      return BinaryOperator::Create((Instruction::BinaryOps)op, args[0], args[1]);
    }

    if (PHINode* phi = dyn_cast<PHINode>(inst)) {
      PHINode* res = PHINode::Create(GetBasicType(inst), phi->getNumIncomingValues());
      assert(args.size() % 2 == 0 && "Odd number of arguments for a PHI");

      // Loop over pairs of operands: [Value*, BasicBlock*]
      for (unsigned int i = 0; i < args.size(); i++) {
        BasicBlock* bb = phi->getIncomingBlock(i); //dyn_cast<BasicBlock>(args[i+1]);
        //assert(bb && "Non-basic block incoming block?");
        res->addIncoming(args[i], bb);
      }

      return res;
    }

    if (CmpInst* cmpInst = dyn_cast<CmpInst>(inst)) {
      assert(args.size() == 2 && "incorrect number of arguments for comparison");
      return CmpInst::Create(cmpInst->getOpcode(), cmpInst->getPredicate(), args[0], args[1]);
    }

    if (isa<SelectInst>(inst)) {
      assert(args.size() == 3 && "incorrect number of arguments for select");
      return SelectInst::Create(args[0], args[1], args[2]);
    }

    if (IntrinsicInst* intr = dyn_cast<IntrinsicInst>(inst)) {
      if (! IsPerComponentOp(inst))
        NOT_IMPLEMENTED; //gla::UnsupportedFunctionality("Scalarize instruction on a non-per-component intrinsic");

      // TODO: Assumption is that all per-component intrinsics have all their
      // arguments be overloadable. Need to find some way to assert on this
      // assumption. This is due to how getDeclaration operates; it only takes
      // a list of types that fit overloadable slots.
      SmallVector<Type*, 8> tys(1, GetBasicType(inst->getType()));
      // Call instructions have the decl as a last argument, so skip it
      for (ArrayRef<Value*>::iterator i = args.begin(), e = args.end() - 1; i != e; ++i) {
        tys.push_back(GetBasicType((*i)->getType()));
      }

      Function* f = Intrinsic::getDeclaration(module, intr->getIntrinsicID(), tys);
      return CallInst::Create(f, args);
    }

    NOT_IMPLEMENTED; //gla::UnsupportedFunctionality("Currently unsupported instruction: ", inst->getOpcode(),
                     //             inst->getOpcodeName());
    return 0;

  }


  void Scalarize::makeScalarizedCalls(Function* f, ArrayRef<Value*> args, int count, VectorValues& vVals)
  {
    assert(count > 0 && count <= 16 && "invalid number of vector components");
    for (int i = 0; i < count; ++i) {
      Value* res;
      SmallVector<Value*, 8> callArgs(args.begin(), args.end());
      callArgs.push_back(ConstantInt::get(intTy, i));

      res = builder->CreateCall(f, callArgs);
      vVals.setComponent(i, res);
    }
  }

  void Scalarize::makePerComponentScalarizedCalls(Instruction* inst, ArrayRef<Value*> args)
  {
    int count = GetComponentCount(inst);
    assert(count > 0 && count <= 16 && "invalid number of vector components");
    assert((inst->getNumOperands() == args.size() || isa<PHINode>(inst))
           && "not enough arguments passed for instruction");

    VectorValues& vVals = vectorVals[inst];

    for (int i = 0; i < count; ++i) {
      // Set this component of each arg
      SmallVector<Value*, 8> callArgs(args.size(), 0);
      gatherComponents(i, args, callArgs);

      Instruction* res = createScalarInstruction(inst, callArgs);

      vVals.setComponent(i, res);
      builder->Insert(res);
    }
  }

  bool Scalarize::isValid(const Instruction* inst)
  {
    // The result
    if (inst->getType()->isVectorTy())
        return false;

    // The arguments
    for (Instruction::const_op_iterator i = inst->op_begin(), e = inst->op_end(); i != e; ++i) {
      const Value* v = (*i);
      assert(v);
      if (v->getType()->isVectorTy())
        return false;
    }

    return true;
  }

  bool Scalarize::scalarize(Instruction* inst)
  {
    if (isValid(inst))
        return false;

    assert(! vectorVals.count(inst) && "We've already scalarized this somehow?");
    assert((canGetComponentArgs(inst) || isa<PHINode>(inst)) &&
           "Scalarizing an op whose arguments haven't been scalarized ");
    builder->SetInsertPoint(inst);

    if (IsPerComponentOp(inst))
      return scalarizePerComponent(inst);

    if (LoadInst* ld = dyn_cast<LoadInst>(inst))
      return scalarizeLoad(ld);

    if (CallInst* call = dyn_cast<CallInst>(inst))
      return scalarizeFuncCall(call);

    if (ExtractElementInst* extr = dyn_cast<ExtractElementInst>(inst))
      return scalarizeExtract(extr);

    if (InsertElementInst* ins = dyn_cast<InsertElementInst>(inst))
      return scalarizeInsert(ins);

    if (ShuffleVectorInst* sv = dyn_cast<ShuffleVectorInst>(inst))
      return scalarizeShuffleVector(sv);

    if (PHINode* phi = dyn_cast<PHINode>(inst))
      return scalarizePHI(phi);

    if (isa<ExtractValueInst>(inst) || isa<InsertValueInst>(inst))
      // TODO: need to come up with a struct/array model for scalarization
      NOT_IMPLEMENTED; //gla::UnsupportedFunctionality("Scalarizing struct/array ops");

    if (StoreInst* st = dyn_cast<StoreInst>(inst))
      return scalarizeStore(st);

    NOT_IMPLEMENTED; //gla::UnsupportedFunctionality("Currently unhandled instruction ", inst->getOpcode(), inst->getOpcodeName());
    return false;
  }

  bool Scalarize::scalarizeShuffleVector(ShuffleVectorInst* sv)
  {
    //     %res = shuffleVector <n x ty> %foo, <n x ty> bar, <n x i32> <...>
    // ==> nothing (just make a new VectorValues with the new components)
    VectorValues& vVals = vectorVals[sv];

    int size = GetComponentCount(sv);
    int srcSize = GetComponentCount(sv->getOperand(0)->getType());

    for (int i = 0; i < size; ++i) {
      int select = sv->getMaskValue(i);

      if (select < 0) {
        vVals.setComponent(i, UndefValue::get(GetBasicType(sv->getOperand(0))));
        continue;
      }

      // Otherwise look up the corresponding component from the correct
      // source.
      Value* selectee;
      if (select < srcSize) {
        selectee = sv->getOperand(0);
      } else {
        // Choose from the second operand
        select -= srcSize;
        selectee = sv->getOperand(1);
      }

      vVals.setComponent(i, getComponent(select, selectee));
    }

    return true;
  }

  bool Scalarize::scalarizePerComponent(Instruction* inst)
  {
    //     dst  = op <n x ty> %foo, <n x ty> %bar
    // ==> dstx = op ty %foox, ty %barx
    //     dsty = op ty %fooy, ty %bary
    //     ...

    SmallVector<Value*, 16> args(inst->op_begin(), inst->op_end());

    makePerComponentScalarizedCalls(inst, args);

    return true;
  }

  bool Scalarize::scalarizePHI(PHINode* phi)
  {
    //     dst = phi <n x ty> [ %foo, %bb1 ], [ %bar, %bb2], ...
    // ==> dstx = phi ty [ %foox, %bb1 ], [ %barx, %bb2], ...
    //     dsty = phi ty [ %fooy, %bb1 ], [ %bary, %bb2], ...

    // If the scalar values are all known up-front, then just make the full
    // phinode now. If they are not yet known (phinode for a loop variant
    // variable), then deferr the arguments until later

    if (canGetComponentArgs(phi)) {
      SmallVector<Value*, 8> args(phi->op_begin(), phi->op_end());
      makePerComponentScalarizedCalls(phi, args);
    } else {
      makePerComponentScalarizedCalls(phi, ArrayRef<Value*>());
      incompletePhis.push_back(phi);
    }

    return true;
  }

  void Scalarize::extractFromeVector(Value* insn) {
    VectorValues& vVals = vectorVals[insn];

    for (int i = 0; i < GetComponentCount(insn); ++i) {
      Value *cv = ConstantInt::get(intTy, i);
      Value *EI = builder->CreateExtractElement(insn, cv);
      vVals.setComponent(i, EI);
      //unit.fakeInsnMap[EI] = insn;
      unit.newValueProxy(insn, EI, i, 0);
    }
  }

  Value* Scalarize::InsertToVector(Value * insn, Value* vecValue) {
    //VectorValues& vVals = vectorVals[writeValue];
    //unit.vecValuesMap[call] = vectorVals[writeValue];

    //add fake insert instructions to avoid removed
    Value *II = NULL;
    for (int i = 0; i < GetComponentCount(vecValue); ++i) {
      Value *vec = II ? II : UndefValue::get(vecValue->getType());
      Value *cv = ConstantInt::get(intTy, i);
      II = builder->CreateInsertElement(vec, getComponent(i, vecValue), cv);
      //unit.vecValuesMap[insn].setComponent(i, getComponent(i, writeValue));
      //unit.newValueProxy(getComponent(i, vecValue), vecValue, 0, i);
      //unit.fakeInsnMap[II] = insn;
    }

    for (int i = 0; i < GetComponentCount(vecValue); ++i) {
      unit.newValueProxy(getComponent(i, vecValue), II, 0, i);
    }
    return II;
  }

  bool Scalarize::scalarizeFuncCall(CallInst* call) {
    if (Function *F = call->getCalledFunction()) {
      if (F->getIntrinsicID() != 0) {   //Intrinsic functions
        NOT_IMPLEMENTED;
      } else {
        Value *Callee = call->getCalledValue();
        const std::string fnName = Callee->getName();
        auto it = instrinsicMap.map.find(fnName);
        GBE_ASSERT(it != instrinsicMap.map.end());

        // Get the function arguments
        CallSite CS(call);
        CallSite::arg_iterator CI = CS.arg_begin() + 3;

        switch (it->second) {
          default: break;
          case GEN_OCL_READ_IMAGE0:
          case GEN_OCL_READ_IMAGE1:
          case GEN_OCL_READ_IMAGE2:
          case GEN_OCL_READ_IMAGE3:
          case GEN_OCL_READ_IMAGE4:
          case GEN_OCL_READ_IMAGE5:
          case GEN_OCL_READ_IMAGE10:
          case GEN_OCL_READ_IMAGE11:
          case GEN_OCL_READ_IMAGE12:
          case GEN_OCL_READ_IMAGE13:
          case GEN_OCL_READ_IMAGE14:
          case GEN_OCL_READ_IMAGE15:
          case GEN_OCL_GET_IMAGE_WIDTH:
          case GEN_OCL_GET_IMAGE_HEIGHT:
          {
            extractFromeVector(call);
            break;
          }
          case GEN_OCL_WRITE_IMAGE10:
          case GEN_OCL_WRITE_IMAGE11:
          case GEN_OCL_WRITE_IMAGE12:
          case GEN_OCL_WRITE_IMAGE13:
          case GEN_OCL_WRITE_IMAGE14:
          case GEN_OCL_WRITE_IMAGE15:
            CI++;
          case GEN_OCL_WRITE_IMAGE0:
          case GEN_OCL_WRITE_IMAGE1:
          case GEN_OCL_WRITE_IMAGE2:
          case GEN_OCL_WRITE_IMAGE3:
          case GEN_OCL_WRITE_IMAGE4:
          case GEN_OCL_WRITE_IMAGE5:
          {
            *CI = InsertToVector(call, *CI);
            break;
          }
        }
      }
    }
    return false;
  }

  bool Scalarize::scalarizeLoad(LoadInst* ld)
  {
    extractFromeVector(ld);
    return false;
  }

  bool Scalarize::scalarizeStore(StoreInst* st) {
    st->setOperand(0, InsertToVector(st, st->getValueOperand()));
    return false;
  }

  bool Scalarize::scalarizeExtract(ExtractElementInst* extr)
  {
    //     %res = extractelement <n X ty> %foo, %i
    // ==> nothing (just use %foo's %ith component instead of %res)

    if (! isa<Constant>(extr->getOperand(1))) {
        // TODO: Variably referenced components. Probably handle/emulate through
        // a series of selects.
        NOT_IMPLEMENTED; //gla::UnsupportedFunctionality("Variably referenced vector components");
    }
    //if (isa<Argument>(extr->getOperand(0)))
    //  return false;
    int component = GetConstantInt(extr->getOperand(1));
    Value* v = getComponent(component, extr->getOperand(0));
    if(extr == v)
      return false;
    extr->replaceAllUsesWith(v);

    return true;
  }

  bool Scalarize::scalarizeInsert(InsertElementInst* ins)
  {
    //     %res = insertValue <n x ty> %foo, %i
    // ==> nothing (just make a new VectorValues with the new component)

    if (! isa<Constant>(ins->getOperand(2))) {
      // TODO: Variably referenced components. Probably handle/emulate through
      // a series of selects.
      NOT_IMPLEMENTED;   //gla::UnsupportedFunctionality("Variably referenced vector components");
    }

    int component = GetConstantInt(ins->getOperand(2));

    VectorValues& vVals = vectorVals[ins];
    for (int i = 0; i < GetComponentCount(ins); ++i) {
      vVals.setComponent(i, i == component ? ins->getOperand(1)
                                           : getComponent(i, ins->getOperand(0)));
    }

    return true;
  }

  void Scalarize::scalarizeArgs(Function& F)  {
    if (F.arg_empty())
      return;
    ReversePostOrderTraversal<Function*> rpot(&F);
    BasicBlock::iterator instI = (*rpot.begin())->begin();
    builder->SetInsertPoint(instI);

    Function::arg_iterator I = F.arg_begin(), E = F.arg_end();

#if LLVM_VERSION_MINOR <= 1
    const AttrListPtr &PAL = F.getAttributes();
    uint32_t argID = 1; // Start at one actually
    for (; I != E; ++I, ++argID) {
#else
    for (; I != E; ++I) {
#endif /* LLVM_VERSION_MINOR <= 1 */
      Type *type = I->getType();

      if(type->isVectorTy())
        extractFromeVector(I);
    }
    return;
  }

  bool Scalarize::runOnFunction(Function& F)
  {
    switch (F.getCallingConv()) {
    case CallingConv::PTX_Device:
      return false;
    case CallingConv::PTX_Kernel:
      break;
    default: GBE_ASSERTM(false, "Unsupported calling convention");
    }

    bool changed = false;
    module = F.getParent();
    intTy = IntegerType::get(module->getContext(), 32);
    floatTy = Type::getFloatTy(module->getContext());
    builder = new IRBuilder<>(module->getContext());

    scalarizeArgs(F);

    typedef ReversePostOrderTraversal<Function*> RPOTType;
    RPOTType rpot(&F);
    for (RPOTType::rpo_iterator bbI = rpot.begin(), bbE = rpot.end(); bbI != bbE; ++bbI) {
      for (BasicBlock::iterator instI = (*bbI)->begin(), instE = (*bbI)->end(); instI != instE; ++instI) {
        bool scalarized = scalarize(instI);
        if (scalarized) {
          changed = true;
          // TODO: uncomment when done
          deadList.push_back(instI);
        }
      }
    }

    // Fill in the incomplete phis
    for (SmallVectorImpl<PHINode*>::iterator phiI = incompletePhis.begin(), phiE = incompletePhis.end();
       phiI != phiE; ++phiI) {
      assert(canGetComponentArgs(*phiI) && "Phi's operands never scalarized");

      // Fill in each component of this phi
      VectorValues& vVals = vectorVals[*phiI];
      for (int c = 0; c < GetComponentCount(*phiI); ++c) {
        PHINode* compPhi = dyn_cast<PHINode>(vVals.getComponent(c));
        assert(compPhi && "Vector phi got scalarized to non-phis?");

        // Loop over pairs of operands: [Value*, BasicBlock*]
        for (unsigned int i = 0; i < (*phiI)->getNumOperands(); i++) {
          BasicBlock* bb = (*phiI)->getIncomingBlock(i);
          assert(bb && "Non-basic block incoming block?");
          compPhi->addIncoming(getComponent(c, (*phiI)->getOperand(i)), bb);
        }
      }
    }

    dce();

    incompletePhis.clear();
    vectorVals.clear();

    delete builder;
    builder = 0;

    return changed;
  }

  void Scalarize::dce()
  {
    //two passes delete for some phinode
    for (std::vector<Instruction*>::reverse_iterator i = deadList.rbegin(), e = deadList.rend(); i != e; ++i) {
      (*i)->dropAllReferences();
      if((*i)->use_empty())
        (*i)->eraseFromParent();
    }
    for (std::vector<Instruction*>::reverse_iterator i = deadList.rbegin(), e = deadList.rend(); i != e; ++i) {
      if((*i)->getParent())
        (*i)->eraseFromParent();
    }
    deadList.clear();
  }

  void Scalarize::getAnalysisUsage(AnalysisUsage& AU) const
  {
  }

  void Scalarize::print(raw_ostream&, const Module*) const
  {
      return;
  }
  FunctionPass* createScalarizePass(ir::Unit &unit)
  {
      return new Scalarize(unit);
  }
  char Scalarize::ID = 0;

} // end namespace