File: GenXGEPLowering.cpp

package info (click to toggle)
intel-graphics-compiler 1.0.12504.6-1%2Bdeb12u1
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 83,912 kB
  • sloc: cpp: 910,147; lisp: 202,655; ansic: 15,197; python: 4,025; yacc: 2,241; lex: 1,570; pascal: 244; sh: 104; makefile: 25
file content (354 lines) | stat: -rw-r--r-- 12,325 bytes parent folder | download
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
/*========================== begin_copyright_notice ============================

Copyright (C) 2018-2021 Intel Corporation

SPDX-License-Identifier: MIT

============================= end_copyright_notice ===========================*/

/// GenXGEPLowering
/// ---------------
///
/// GenXGEPLowering is a function pass that lowers GEP instructions into
/// primitive ones that the rest of the GenX backend can deal with.
///
//===----------------------------------------------------------------------===//

#include "GenX.h"
#include "GenXModule.h"
#include "Probe/Assertion.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GetElementPtrTypeIterator.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Module.h"
#include "llvm/InitializePasses.h"
#include "llvm/Pass.h"

#include "llvmWrapper/IR/DerivedTypes.h"
#include "llvmWrapper/IR/Instructions.h"
#include "llvmWrapper/Support/TypeSize.h"
#include "llvm/Support/Debug.h"

#define DEBUG_TYPE "GENX_GEPLowering"

using namespace llvm;
using namespace genx;

namespace {
class GenXGEPLowering : public FunctionPass {
  const DataLayout *DL = nullptr;
  LoopInfo *LI = nullptr;
  IRBuilder<> *Builder = nullptr;
  Module *M = nullptr;

public:
  static char ID;

  GenXGEPLowering() : FunctionPass(ID) {}

  StringRef getPassName() const override { return "GenX GEP Lowering"; }

  bool runOnFunction(Function &F) override;

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.addRequired<LoopInfoWrapperPass>();
    AU.addRequired<TargetTransformInfoWrapperPass>();
    AU.setPreservesCFG();
    AU.addPreserved<GenXModule>();
    AU.addPreserved<LoopInfoWrapperPass>();
  }

private:
  bool lowerGetElementPtrInst(GetElementPtrInst *GEP,
                              BasicBlock::iterator &BBI) const;
  Value *truncExpr(Value *Val, Type *NewTy) const;
  Value *getSExtOrTrunc(Value *, Type *) const;
};

} // namespace

char GenXGEPLowering::ID = 0;
namespace llvm {
void initializeGenXGEPLoweringPass(PassRegistry &);
}
INITIALIZE_PASS_BEGIN(GenXGEPLowering, "GenXGEPLowering", "GenXGEPLowering",
                      false, false)
INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
INITIALIZE_PASS_END(GenXGEPLowering, "GenXGEPLowering", "GenXGEPLowering",
                    false, false)

FunctionPass *llvm::createGenXGEPLoweringPass() {
  initializeGenXGEPLoweringPass(*PassRegistry::getPassRegistry());
  return new GenXGEPLowering;
}

bool GenXGEPLowering::runOnFunction(Function &F) {
  LLVM_DEBUG(dbgs() << "Before GEPLowering\n");
  LLVM_DEBUG(F.dump());
  LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
  M = F.getParent();
  DL = &M->getDataLayout();

  const TargetTransformInfo &TTI =
      getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
  auto FlatAddrSpace = TTI.getFlatAddressSpace();

  IGC_ASSERT_MESSAGE(DL, "null datalayout");
#if 0
  // a good place to fix block layout
  if (LI->empty())
    LayoutBlocks(F);
  else
    LayoutBlocks(F, *LI);
#endif
  IRBuilder<> TheBuilder(F.getContext());
  Builder = &TheBuilder;

  bool Changed = false;
  for (auto &BB : F) {
    for (auto BI = BB.begin(), BE = BB.end(); BI != BE;) {
      Instruction *Inst = &(*BI++);
      Builder->SetInsertPoint(Inst);

      switch (Inst->getOpcode()) {
      default: // By default, DO NOTHING
        break;
      case Instruction::GetElementPtr:
        Changed |= lowerGetElementPtrInst(cast<GetElementPtrInst>(Inst), BI);
        break;
      case Instruction::PtrToInt:
        auto PtrV = cast<PtrToIntInst>(Inst)->getPointerOperand();
        auto AddrSpace = cast<PtrToIntInst>(Inst)->getPointerAddressSpace();
        if (AddrSpace == FlatAddrSpace) {
          if (auto PtrCast = dyn_cast<AddrSpaceCastInst>(PtrV)) {
            // this is no-op AddrSpaceCast, should be removed
            // create a new PtrToInt from the original pointer
            // bypass the AddrSpaceCast and PtrToInt
            auto P2I = Builder->CreatePtrToInt(PtrCast->getOperand(0),
                                               Inst->getType());
            Inst->replaceAllUsesWith(P2I);
            Inst->eraseFromParent();
            if (PtrCast->use_empty()) {
              PtrCast->eraseFromParent();
            }
          }
        }
        break;
      }
    }
  }
  LLVM_DEBUG(dbgs() << "After GEPLowering\n");
  LLVM_DEBUG(F.dump());
  Builder = nullptr;

  return Changed;
}

bool GenXGEPLowering::lowerGetElementPtrInst(GetElementPtrInst *GEP,
                                             BasicBlock::iterator &BBI) const {
  IGC_ASSERT(Builder);
  Value *PtrOp = GEP->getPointerOperand();

  Value *PointerValue;
  PointerType *PtrTy = dyn_cast<PointerType>(PtrOp->getType());
  Type *IntPtrTy;
  if (PtrTy) {
    IntPtrTy =
        DL->getIntPtrType(Builder->getContext(), PtrTy->getAddressSpace());
  } else {
    IGC_ASSERT(PtrOp->getType()->isVectorTy());
    auto PtrOpVTy = cast<IGCLLVM::FixedVectorType>(PtrOp->getType());
    PtrTy = cast<PointerType>(PtrOpVTy->getElementType());
    IntPtrTy =
        DL->getIntPtrType(Builder->getContext(), PtrTy->getAddressSpace());
    IntPtrTy =
        IGCLLVM::FixedVectorType::get(IntPtrTy, PtrOpVTy->getNumElements());
  }
  // Check if the pointer itself is created from IntToPtr. If it is, and if
  // the int is the same size, we can use the int directly. Otherwise, we
  // need to add PtrToInt.
  if (auto *I2PI = dyn_cast<IntToPtrInst>(PtrOp)) {
    Value *IntOp = I2PI->getOperand(0);
    if (IntOp->getType() == IntPtrTy)
      PointerValue = IntOp;
  }
  PointerValue = Builder->CreatePtrToInt(PtrOp, IntPtrTy);

  unsigned PtrMathSizeInBits =
      DL->getPointerSizeInBits(PtrTy->getAddressSpace());
  auto *PtrMathTy = IntegerType::get(Builder->getContext(), PtrMathSizeInBits);

  auto *GEPVecTy = dyn_cast<IGCLLVM::FixedVectorType>(GEP->getType());
  if (GEPVecTy && isa<PointerType>(PtrOp->getType())) {
    PointerValue = Builder->CreateVectorSplat(
        GEPVecTy->getNumElements(), PointerValue, PtrOp->getName() + ".splat");
  }

  Type *Ty = PtrOp->getType();
  gep_type_iterator GTI = gep_type_begin(GEP);
  for (auto OI = GEP->op_begin() + 1, E = GEP->op_end(); OI != E; ++OI, ++GTI) {
    Value *Idx = *OI;
    if (StructType *StTy = GTI.getStructTypeOrNull()) {
      int64_t Field = 0;
      Constant *CE = cast<Constant>(Idx);
      if (!CE->isNullValue()) {
        Field = CE->getUniqueInteger().getSExtValue();
        uint64_t Offset = DL->getStructLayout(StTy)->getElementOffset(Field);
        Value *OffsetVal = Constant::getIntegerValue(
          PointerValue->getType(), APInt(PtrMathSizeInBits, Offset));
        PointerValue = Builder->CreateAdd(PointerValue, OffsetVal);
      }

      Ty = StTy->getElementType(Field);
    } else {
      Ty = GTI.getIndexedType();
      if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
        if (!CI->isZero()) {
          uint64_t Offset = DL->getTypeAllocSize(Ty) * CI->getSExtValue();
          Value *OffsetVal = Builder->getInt(APInt(PtrMathSizeInBits, Offset));
          PointerValue = Builder->CreateAdd(PointerValue, OffsetVal);
        }
      } else if (!isa<ConstantAggregateZero>(Idx)) {
        Value *NewIdx = getSExtOrTrunc(Idx, PtrMathTy);
        APInt ElementSize = APInt(PtrMathSizeInBits, DL->getTypeAllocSize(Ty));

        if (BinaryOperator *BO = dyn_cast<BinaryOperator>(NewIdx)) {
          // Detect the pattern GEP base, a + b where base and a are both loop
          // invariant (but not b), so we could rearrange the lowered code into
          // (base + (a << shftAmt)) + (b << shftAmt).
          Loop *L = LI ? LI->getLoopFor(BO->getParent()) : nullptr;
          if (L && L->isLoopInvariant(GEP->getPointerOperand()) &&
              BO->getOpcode() == Instruction::Add) {

            auto reassociate = [&](Value *A, Value *B) {
              Value *InvVal = nullptr;
              if (ElementSize == 1)
                InvVal = A;
              else if (ElementSize.isPowerOf2())
                InvVal = Builder->CreateShl(
                    A, Constant::getIntegerValue(
                           A->getType(),
                           APInt(PtrMathSizeInBits, ElementSize.logBase2())));
              else
                InvVal = Builder->CreateMul(
                    A, Constant::getIntegerValue(A->getType(), ElementSize));
              PointerValue = Builder->CreateAdd(PointerValue, InvVal);
              NewIdx = B;
            };

            Value *LHS = BO->getOperand(0);
            Value *RHS = BO->getOperand(1);
            bool isLHSLI = L->isLoopInvariant(LHS);
            bool isRHSLI = L->isLoopInvariant(RHS);
            if (isLHSLI && !isRHSLI)
              reassociate(LHS, RHS);
            else if (!isLHSLI && isRHSLI)
              reassociate(RHS, LHS);
          }
        }
        if (ElementSize == 1) {
          // DO NOTHING.
        } else if (ElementSize.isPowerOf2()) {
          APInt ShiftAmount = APInt(PtrMathSizeInBits, ElementSize.logBase2());
          NewIdx = Builder->CreateShl(
              NewIdx,
              Constant::getIntegerValue(NewIdx->getType(), ShiftAmount));
        } else
          NewIdx = Builder->CreateMul(
              NewIdx,
              Constant::getIntegerValue(NewIdx->getType(), ElementSize));

        PointerValue = Builder->CreateAdd(PointerValue, NewIdx);
      }
    }
  }

  PointerValue = Builder->CreateIntToPtr(PointerValue, GEP->getType());
  GEP->replaceAllUsesWith(PointerValue);
  GEP->eraseFromParent();
  if (Instruction *I = dyn_cast<Instruction>(PointerValue)) {
    BBI = BasicBlock::iterator(I);
    ++BBI;
  }

  return true;
}

Value *GenXGEPLowering::getSExtOrTrunc(Value *Val, Type *NewTy) const {
  IGC_ASSERT(Builder);
  unsigned NewWidth = NewTy->getIntegerBitWidth();
  Type *OldTy = Val->getType();
  if (auto *OldVecTy = dyn_cast<IGCLLVM::FixedVectorType>(OldTy)) {
    NewTy = IGCLLVM::FixedVectorType::get(NewTy, OldVecTy->getNumElements());
    OldTy = OldVecTy->getElementType();
  }
  unsigned OldWidth = OldTy->getIntegerBitWidth();

  if (OldWidth < NewWidth) // SExt
    return Builder->CreateSExt(Val, NewTy);
  if (OldWidth > NewWidth) // Trunc
    return truncExpr(Val, NewTy);
  return Val;
}

Value *GenXGEPLowering::truncExpr(Value *Val, Type *NewTy) const {
  IGC_ASSERT(Builder);
  // Truncation on Gen could be as cheap as NOP by creating proper regions.
  // Instead of truncating the value itself, truncate how it's calculated.
  if (Constant *C = dyn_cast<Constant>(Val))
    return Builder->CreateIntCast(C, NewTy, false);

  if (!isa<Instruction>(Val))
    return Builder->CreateTrunc(Val, NewTy);

  Instruction *I = cast<Instruction>(Val);
  unsigned Opc = I->getOpcode();
  switch (Opc) {
  case Instruction::Add:
  case Instruction::Sub:
  case Instruction::Mul:
  case Instruction::And:
  case Instruction::Or:
  case Instruction::Xor: {
    BinaryOperator *BO = cast<BinaryOperator>(I);
    Value *LHS = truncExpr(BO->getOperand(0), NewTy);
    Value *RHS = truncExpr(BO->getOperand(1), NewTy);
    return Builder->CreateBinOp(BO->getOpcode(), LHS, RHS);
  }
  case Instruction::Trunc:
  case Instruction::ZExt:
  case Instruction::SExt: {
    Value *Opnd = I->getOperand(0);
    if (Opnd->getType() == NewTy)
      return Opnd;
    return Builder->CreateIntCast(Opnd, NewTy, Opc == Instruction::SExt);
  }
  case Instruction::Select: {
    Value *TVal = truncExpr(I->getOperand(1), NewTy);
    Value *FVal = truncExpr(I->getOperand(2), NewTy);
    return Builder->CreateSelect(I->getOperand(0), TVal, FVal);
  }
#if 0
  // TODO: Rewrite truncExpr into iterative one instead of recursive one to
  // easily found the loop due to phi-node.
  case Instruction::PHI: {
    PHINode *PN = cast<PHINode>(I);
    PHINode *Res = PHINode::Create(NewTy, PN->getNumIncomingValues());
    for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
      Value *V = truncExpr(PN->getIncomingValue(i), NewTy);
      Res->addIncoming(V, PN->getIncomingBlock(i));
    }
    return Res;
  }
#endif
  default:
    // Don't know truncate its calculation safely, fall back to the regular way.
    break;
  }

  return Builder->CreateTrunc(Val, NewTy);
}