summaryrefslogtreecommitdiffstats
path: root/lib/Target/NVPTX/NVPTXLowerAggrCopies.cpp
blob: 578401a34ea20f7b11a34ffb4b487fe083cf6247 (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
//===- NVPTXLowerAggrCopies.cpp - ------------------------------*- C++ -*--===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
// Lower aggregate copies, memset, memcpy, memmov intrinsics into loops when
// the size is large or is not a compile-time constant.
//
//===----------------------------------------------------------------------===//

#include "NVPTXLowerAggrCopies.h"
#include "llvm/CodeGen/MachineFunctionAnalysis.h"
#include "llvm/CodeGen/StackProtector.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/InstIterator.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/Module.h"
#include "llvm/Support/Debug.h"

#define DEBUG_TYPE "nvptx"

using namespace llvm;

namespace {
// actual analysis class, which is a functionpass
struct NVPTXLowerAggrCopies : public FunctionPass {
  static char ID;

  NVPTXLowerAggrCopies() : FunctionPass(ID) {}

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.addPreserved<MachineFunctionAnalysis>();
    AU.addPreserved<StackProtector>();
  }

  bool runOnFunction(Function &F) override;

  static const unsigned MaxAggrCopySize = 128;

  const char *getPassName() const override {
    return "Lower aggregate copies/intrinsics into loops";
  }
};
} // namespace

char NVPTXLowerAggrCopies::ID = 0;

// Lower MemTransferInst or load-store pair to loop
static void convertTransferToLoop(
    Instruction *splitAt, Value *srcAddr, Value *dstAddr, Value *len,
    //unsigned numLoads,
    bool srcVolatile, bool dstVolatile, LLVMContext &Context, Function &F) {
  Type *indType = len->getType();

  BasicBlock *origBB = splitAt->getParent();
  BasicBlock *newBB = splitAt->getParent()->splitBasicBlock(splitAt, "split");
  BasicBlock *loopBB = BasicBlock::Create(Context, "loadstoreloop", &F, newBB);

  origBB->getTerminator()->setSuccessor(0, loopBB);
  IRBuilder<> builder(origBB, origBB->getTerminator());

  // srcAddr and dstAddr are expected to be pointer types,
  // so no check is made here.
  unsigned srcAS = dyn_cast<PointerType>(srcAddr->getType())->getAddressSpace();
  unsigned dstAS = dyn_cast<PointerType>(dstAddr->getType())->getAddressSpace();

  // Cast pointers to (char *)
  srcAddr = builder.CreateBitCast(srcAddr, Type::getInt8PtrTy(Context, srcAS));
  dstAddr = builder.CreateBitCast(dstAddr, Type::getInt8PtrTy(Context, dstAS));

  IRBuilder<> loop(loopBB);
  // The loop index (ind) is a phi node.
  PHINode *ind = loop.CreatePHI(indType, 0);
  // Incoming value for ind is 0
  ind->addIncoming(ConstantInt::get(indType, 0), origBB);

  // load from srcAddr+ind
  Value *val = loop.CreateLoad(loop.CreateGEP(srcAddr, ind), srcVolatile);
  // store at dstAddr+ind
  loop.CreateStore(val, loop.CreateGEP(dstAddr, ind), dstVolatile);

  // The value for ind coming from backedge is (ind + 1)
  Value *newind = loop.CreateAdd(ind, ConstantInt::get(indType, 1));
  ind->addIncoming(newind, loopBB);

  loop.CreateCondBr(loop.CreateICmpULT(newind, len), loopBB, newBB);
}

// Lower MemSetInst to loop
static void convertMemSetToLoop(Instruction *splitAt, Value *dstAddr,
                                Value *len, Value *val, LLVMContext &Context,
                                Function &F) {
  BasicBlock *origBB = splitAt->getParent();
  BasicBlock *newBB = splitAt->getParent()->splitBasicBlock(splitAt, "split");
  BasicBlock *loopBB = BasicBlock::Create(Context, "loadstoreloop", &F, newBB);

  origBB->getTerminator()->setSuccessor(0, loopBB);
  IRBuilder<> builder(origBB, origBB->getTerminator());

  unsigned dstAS = dyn_cast<PointerType>(dstAddr->getType())->getAddressSpace();

  // Cast pointer to the type of value getting stored
  dstAddr =
      builder.CreateBitCast(dstAddr, PointerType::get(val->getType(), dstAS));

  IRBuilder<> loop(loopBB);
  PHINode *ind = loop.CreatePHI(len->getType(), 0);
  ind->addIncoming(ConstantInt::get(len->getType(), 0), origBB);

  loop.CreateStore(val, loop.CreateGEP(dstAddr, ind), false);

  Value *newind = loop.CreateAdd(ind, ConstantInt::get(len->getType(), 1));
  ind->addIncoming(newind, loopBB);

  loop.CreateCondBr(loop.CreateICmpULT(newind, len), loopBB, newBB);
}

bool NVPTXLowerAggrCopies::runOnFunction(Function &F) {
  SmallVector<LoadInst *, 4> aggrLoads;
  SmallVector<MemTransferInst *, 4> aggrMemcpys;
  SmallVector<MemSetInst *, 4> aggrMemsets;

  const DataLayout &DL = F.getParent()->getDataLayout();
  LLVMContext &Context = F.getParent()->getContext();

  //
  // Collect all the aggrLoads, aggrMemcpys and addrMemsets.
  //
  //const BasicBlock *firstBB = &F.front();  // first BB in F
  for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
    //BasicBlock *bb = BI;
    for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;
         ++II) {
      if (LoadInst *load = dyn_cast<LoadInst>(II)) {

        if (!load->hasOneUse())
          continue;

        if (DL.getTypeStoreSize(load->getType()) < MaxAggrCopySize)
          continue;

        User *use = load->user_back();
        if (StoreInst *store = dyn_cast<StoreInst>(use)) {
          if (store->getOperand(0) != load) //getValueOperand
            continue;
          aggrLoads.push_back(load);
        }
      } else if (MemTransferInst *intr = dyn_cast<MemTransferInst>(II)) {
        Value *len = intr->getLength();
        // If the number of elements being copied is greater
        // than MaxAggrCopySize, lower it to a loop
        if (ConstantInt *len_int = dyn_cast<ConstantInt>(len)) {
          if (len_int->getZExtValue() >= MaxAggrCopySize) {
            aggrMemcpys.push_back(intr);
          }
        } else {
          // turn variable length memcpy/memmov into loop
          aggrMemcpys.push_back(intr);
        }
      } else if (MemSetInst *memsetintr = dyn_cast<MemSetInst>(II)) {
        Value *len = memsetintr->getLength();
        if (ConstantInt *len_int = dyn_cast<ConstantInt>(len)) {
          if (len_int->getZExtValue() >= MaxAggrCopySize) {
            aggrMemsets.push_back(memsetintr);
          }
        } else {
          // turn variable length memset into loop
          aggrMemsets.push_back(memsetintr);
        }
      }
    }
  }
  if ((aggrLoads.size() == 0) && (aggrMemcpys.size() == 0) &&
      (aggrMemsets.size() == 0))
    return false;

  //
  // Do the transformation of an aggr load/copy/set to a loop
  //
  for (unsigned i = 0, e = aggrLoads.size(); i != e; ++i) {
    LoadInst *load = aggrLoads[i];
    StoreInst *store = dyn_cast<StoreInst>(*load->user_begin());
    Value *srcAddr = load->getOperand(0);
    Value *dstAddr = store->getOperand(1);
    unsigned numLoads = DL.getTypeStoreSize(load->getType());
    Value *len = ConstantInt::get(Type::getInt32Ty(Context), numLoads);

    convertTransferToLoop(store, srcAddr, dstAddr, len, load->isVolatile(),
                          store->isVolatile(), Context, F);

    store->eraseFromParent();
    load->eraseFromParent();
  }

  for (unsigned i = 0, e = aggrMemcpys.size(); i != e; ++i) {
    MemTransferInst *cpy = aggrMemcpys[i];
    Value *len = cpy->getLength();
    // llvm 2.7 version of memcpy does not have volatile
    // operand yet. So always making it non-volatile
    // optimistically, so that we don't see unnecessary
    // st.volatile in ptx
    convertTransferToLoop(cpy, cpy->getSource(), cpy->getDest(), len, false,
                          false, Context, F);
    cpy->eraseFromParent();
  }

  for (unsigned i = 0, e = aggrMemsets.size(); i != e; ++i) {
    MemSetInst *memsetinst = aggrMemsets[i];
    Value *len = memsetinst->getLength();
    Value *val = memsetinst->getValue();
    convertMemSetToLoop(memsetinst, memsetinst->getDest(), len, val, Context,
                        F);
    memsetinst->eraseFromParent();
  }

  return true;
}

FunctionPass *llvm::createLowerAggrCopies() {
  return new NVPTXLowerAggrCopies();
}