aboutsummaryrefslogtreecommitdiffstats
path: root/lib/Target/R600/SIFoldOperands.cpp
blob: ae4b05d66956539887165912c8324e3ffc465373 (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
//===-- SIFoldOperands.cpp - Fold operands --- ----------------------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
/// \file
//===----------------------------------------------------------------------===//
//

#include "AMDGPU.h"
#include "AMDGPUSubtarget.h"
#include "SIInstrInfo.h"
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/Function.h"
#include "llvm/Support/Debug.h"
#include "llvm/Target/TargetMachine.h"

#define DEBUG_TYPE "si-fold-operands"
using namespace llvm;

namespace {

class SIFoldOperands : public MachineFunctionPass {
public:
  static char ID;

public:
  SIFoldOperands() : MachineFunctionPass(ID) {
    initializeSIFoldOperandsPass(*PassRegistry::getPassRegistry());
  }

  bool runOnMachineFunction(MachineFunction &MF) override;

  const char *getPassName() const override {
    return "SI Fold Operands";
  }

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.addRequired<MachineDominatorTree>();
    AU.setPreservesCFG();
    MachineFunctionPass::getAnalysisUsage(AU);
  }
};

struct FoldCandidate {
  MachineInstr *UseMI;
  unsigned UseOpNo;
  MachineOperand *OpToFold;
  uint64_t ImmToFold;

  FoldCandidate(MachineInstr *MI, unsigned OpNo, MachineOperand *FoldOp) :
                UseMI(MI), UseOpNo(OpNo) {

    if (FoldOp->isImm()) {
      OpToFold = nullptr;
      ImmToFold = FoldOp->getImm();
    } else {
      assert(FoldOp->isReg());
      OpToFold = FoldOp;
    }
  }

  bool isImm() const {
    return !OpToFold;
  }
};

} // End anonymous namespace.

INITIALIZE_PASS_BEGIN(SIFoldOperands, DEBUG_TYPE,
                      "SI Fold Operands", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
INITIALIZE_PASS_END(SIFoldOperands, DEBUG_TYPE,
                    "SI Fold Operands", false, false)

char SIFoldOperands::ID = 0;

char &llvm::SIFoldOperandsID = SIFoldOperands::ID;

FunctionPass *llvm::createSIFoldOperandsPass() {
  return new SIFoldOperands();
}

static bool isSafeToFold(unsigned Opcode) {
  switch(Opcode) {
  case AMDGPU::V_MOV_B32_e32:
  case AMDGPU::V_MOV_B32_e64:
  case AMDGPU::V_MOV_B64_PSEUDO:
  case AMDGPU::S_MOV_B32:
  case AMDGPU::S_MOV_B64:
  case AMDGPU::COPY:
    return true;
  default:
    return false;
  }
}

static bool updateOperand(FoldCandidate &Fold,
                          const TargetRegisterInfo &TRI) {
  MachineInstr *MI = Fold.UseMI;
  MachineOperand &Old = MI->getOperand(Fold.UseOpNo);
  assert(Old.isReg());

  if (Fold.isImm()) {
    Old.ChangeToImmediate(Fold.ImmToFold);
    return true;
  }

  MachineOperand *New = Fold.OpToFold;
  if (TargetRegisterInfo::isVirtualRegister(Old.getReg()) &&
      TargetRegisterInfo::isVirtualRegister(New->getReg())) {
    Old.substVirtReg(New->getReg(), New->getSubReg(), TRI);
    return true;
  }

  // FIXME: Handle physical registers.

  return false;
}

static bool tryAddToFoldList(std::vector<FoldCandidate> &FoldList,
                             MachineInstr *MI, unsigned OpNo,
                             MachineOperand *OpToFold,
                             const SIInstrInfo *TII) {
  if (!TII->isOperandLegal(MI, OpNo, OpToFold)) {
    // Operand is not legal, so try to commute the instruction to
    // see if this makes it possible to fold.
    unsigned CommuteIdx0;
    unsigned CommuteIdx1;
    bool CanCommute = TII->findCommutedOpIndices(MI, CommuteIdx0, CommuteIdx1);

    if (CanCommute) {
      if (CommuteIdx0 == OpNo)
        OpNo = CommuteIdx1;
      else if (CommuteIdx1 == OpNo)
        OpNo = CommuteIdx0;
    }

    if (!CanCommute || !TII->commuteInstruction(MI))
      return false;

    if (!TII->isOperandLegal(MI, OpNo, OpToFold))
      return false;
  }

  FoldList.push_back(FoldCandidate(MI, OpNo, OpToFold));
  return true;
}

bool SIFoldOperands::runOnMachineFunction(MachineFunction &MF) {
  MachineRegisterInfo &MRI = MF.getRegInfo();
  const SIInstrInfo *TII =
      static_cast<const SIInstrInfo *>(MF.getSubtarget().getInstrInfo());
  const SIRegisterInfo &TRI = TII->getRegisterInfo();

  for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
                                                  BI != BE; ++BI) {

    MachineBasicBlock &MBB = *BI;
    MachineBasicBlock::iterator I, Next;
    for (I = MBB.begin(); I != MBB.end(); I = Next) {
      Next = std::next(I);
      MachineInstr &MI = *I;

      if (!isSafeToFold(MI.getOpcode()))
        continue;

      unsigned OpSize = TII->getOpSize(MI, 1);
      MachineOperand &OpToFold = MI.getOperand(1);
      bool FoldingImm = OpToFold.isImm();

      // FIXME: We could also be folding things like FrameIndexes and
      // TargetIndexes.
      if (!FoldingImm && !OpToFold.isReg())
        continue;

      // Folding immediates with more than one use will increase program size.
      // FIXME: This will also reduce register usage, which may be better
      // in some cases.  A better heuristic is needed.
      if (FoldingImm && !TII->isInlineConstant(OpToFold, OpSize) &&
          !MRI.hasOneUse(MI.getOperand(0).getReg()))
        continue;

      // FIXME: Fold operands with subregs.
      if (OpToFold.isReg() &&
          (!TargetRegisterInfo::isVirtualRegister(OpToFold.getReg()) ||
           OpToFold.getSubReg()))
        continue;

      std::vector<FoldCandidate> FoldList;
      for (MachineRegisterInfo::use_iterator
           Use = MRI.use_begin(MI.getOperand(0).getReg()), E = MRI.use_end();
           Use != E; ++Use) {

        MachineInstr *UseMI = Use->getParent();
        const MachineOperand &UseOp = UseMI->getOperand(Use.getOperandNo());

        // FIXME: Fold operands with subregs.
        if (UseOp.isReg() && ((UseOp.getSubReg() && OpToFold.isReg()) ||
            UseOp.isImplicit())) {
          continue;
        }

        APInt Imm;

        if (FoldingImm) {
          unsigned UseReg = UseOp.getReg();
          const TargetRegisterClass *UseRC
            = TargetRegisterInfo::isVirtualRegister(UseReg) ?
            MRI.getRegClass(UseReg) :
            TRI.getRegClass(UseReg);

          Imm = APInt(64, OpToFold.getImm());

          // Split 64-bit constants into 32-bits for folding.
          if (UseOp.getSubReg()) {
            if (UseRC->getSize() != 8)
              continue;

            if (UseOp.getSubReg() == AMDGPU::sub0) {
              Imm = Imm.getLoBits(32);
            } else {
              assert(UseOp.getSubReg() == AMDGPU::sub1);
              Imm = Imm.getHiBits(32);
            }
          }

          // In order to fold immediates into copies, we need to change the
          // copy to a MOV.
          if (UseMI->getOpcode() == AMDGPU::COPY) {
            unsigned DestReg = UseMI->getOperand(0).getReg();
            const TargetRegisterClass *DestRC
              = TargetRegisterInfo::isVirtualRegister(DestReg) ?
              MRI.getRegClass(DestReg) :
              TRI.getRegClass(DestReg);

            unsigned MovOp = TII->getMovOpcode(DestRC);
            if (MovOp == AMDGPU::COPY)
              continue;

            UseMI->setDesc(TII->get(MovOp));
          }
        }

        const MCInstrDesc &UseDesc = UseMI->getDesc();

        // Don't fold into target independent nodes.  Target independent opcodes
        // don't have defined register classes.
        if (UseDesc.isVariadic() ||
            UseDesc.OpInfo[Use.getOperandNo()].RegClass == -1)
          continue;

        if (FoldingImm) {
          MachineOperand ImmOp = MachineOperand::CreateImm(Imm.getSExtValue());
          tryAddToFoldList(FoldList, UseMI, Use.getOperandNo(), &ImmOp, TII);
          continue;
        }

        tryAddToFoldList(FoldList, UseMI, Use.getOperandNo(), &OpToFold, TII);

        // FIXME: We could try to change the instruction from 64-bit to 32-bit
        // to enable more folding opportunites.  The shrink operands pass
        // already does this.
      }

      for (FoldCandidate &Fold : FoldList) {
        if (updateOperand(Fold, TRI)) {
          // Clear kill flags.
          if (!Fold.isImm()) {
            assert(Fold.OpToFold && Fold.OpToFold->isReg());
            Fold.OpToFold->setIsKill(false);
          }
          DEBUG(dbgs() << "Folded source from " << MI << " into OpNo " <<
                Fold.UseOpNo << " of " << *Fold.UseMI << '\n');
        }
      }
    }
  }
  return false;
}