aboutsummaryrefslogtreecommitdiffstats
path: root/lib/Target/R600/SIShrinkInstructions.cpp
blob: 51e72cdb5f9ed289b5ed585f5c93e5cbeebe2d81 (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
//===-- SIShrinkInstructions.cpp - Shrink Instructions --------------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
/// The pass tries to use the 32-bit encoding for instructions when possible.
//===----------------------------------------------------------------------===//
//

#include "AMDGPU.h"
#include "AMDGPUMCInstLower.h"
#include "AMDGPUSubtarget.h"
#include "SIInstrInfo.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetMachine.h"

#define DEBUG_TYPE "si-shrink-instructions"

STATISTIC(NumInstructionsShrunk,
          "Number of 64-bit instruction reduced to 32-bit.");
STATISTIC(NumLiteralConstantsFolded,
          "Number of literal constants folded into 32-bit instructions.");

namespace llvm {
  void initializeSIShrinkInstructionsPass(PassRegistry&);
}

using namespace llvm;

namespace {

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

public:
  SIShrinkInstructions() : MachineFunctionPass(ID) {
  }

  bool runOnMachineFunction(MachineFunction &MF) override;

  const char *getPassName() const override {
    return "SI Shrink Instructions";
  }

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

} // End anonymous namespace.

INITIALIZE_PASS_BEGIN(SIShrinkInstructions, DEBUG_TYPE,
                      "SI Lower il Copies", false, false)
INITIALIZE_PASS_END(SIShrinkInstructions, DEBUG_TYPE,
                    "SI Lower il Copies", false, false)

char SIShrinkInstructions::ID = 0;

FunctionPass *llvm::createSIShrinkInstructionsPass() {
  return new SIShrinkInstructions();
}

static bool isVGPR(const MachineOperand *MO, const SIRegisterInfo &TRI,
                   const MachineRegisterInfo &MRI) {
  if (!MO->isReg())
    return false;

  if (TargetRegisterInfo::isVirtualRegister(MO->getReg()))
    return TRI.hasVGPRs(MRI.getRegClass(MO->getReg()));

  return TRI.hasVGPRs(TRI.getPhysRegClass(MO->getReg()));
}

static bool canShrink(MachineInstr &MI, const SIInstrInfo *TII,
                      const SIRegisterInfo &TRI,
                      const MachineRegisterInfo &MRI) {

  const MachineOperand *Src2 = TII->getNamedOperand(MI, AMDGPU::OpName::src2);
  // Can't shrink instruction with three operands.
  // FIXME: v_cndmask_b32 has 3 operands and is shrinkable, but we need to add
  // a special case for it.  It can only be shrunk if the third operand
  // is vcc.  We should handle this the same way we handle vopc, by addding
  // a register allocation hint pre-regalloc and then do the shrining
  // post-regalloc.
  if (Src2)
    return false;

  const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
  const MachineOperand *Src1Mod =
      TII->getNamedOperand(MI, AMDGPU::OpName::src1_modifiers);

  if (Src1 && (!isVGPR(Src1, TRI, MRI) || (Src1Mod && Src1Mod->getImm() != 0)))
    return false;

  // We don't need to check src0, all input types are legal, so just make sure
  // src0 isn't using any modifiers.
  if (TII->hasModifiersSet(MI, AMDGPU::OpName::src0_modifiers))
    return false;

  // Check output modifiers
  if (TII->hasModifiersSet(MI, AMDGPU::OpName::omod))
    return false;

  if (TII->hasModifiersSet(MI, AMDGPU::OpName::clamp))
    return false;

  return true;
}

/// \brief This function checks \p MI for operands defined by a move immediate
/// instruction and then folds the literal constant into the instruction if it
/// can.  This function assumes that \p MI is a VOP1, VOP2, or VOPC instruction
/// and will only fold literal constants if we are still in SSA.
static void foldImmediates(MachineInstr &MI, const SIInstrInfo *TII,
                           MachineRegisterInfo &MRI, bool TryToCommute = true) {

  if (!MRI.isSSA())
    return;

  assert(TII->isVOP1(MI.getOpcode()) || TII->isVOP2(MI.getOpcode()) ||
         TII->isVOPC(MI.getOpcode()));

  const SIRegisterInfo &TRI = TII->getRegisterInfo();
  int Src0Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::src0);
  MachineOperand &Src0 = MI.getOperand(Src0Idx);

  // Only one literal constant is allowed per instruction, so if src0 is a
  // literal constant then we can't do any folding.
  if (Src0.isImm() &&
      TII->isLiteralConstant(Src0, TII->getOpSize(MI, Src0Idx)))
    return;

  // Literal constants and SGPRs can only be used in Src0, so if Src0 is an
  // SGPR, we cannot commute the instruction, so we can't fold any literal
  // constants.
  if (Src0.isReg() && !isVGPR(&Src0, TRI, MRI))
    return;

  // Try to fold Src0
  if (Src0.isReg()) {
    unsigned Reg = Src0.getReg();
    MachineInstr *Def = MRI.getUniqueVRegDef(Reg);
    if (Def && Def->isMoveImmediate()) {
      MachineOperand &MovSrc = Def->getOperand(1);
      bool ConstantFolded = false;

      if (MovSrc.isImm() && isUInt<32>(MovSrc.getImm())) {
        Src0.ChangeToImmediate(MovSrc.getImm());
        ConstantFolded = true;
      }
      if (ConstantFolded) {
        if (MRI.use_empty(Reg))
          Def->eraseFromParent();
        ++NumLiteralConstantsFolded;
        return;
      }
    }
  }

  // We have failed to fold src0, so commute the instruction and try again.
  if (TryToCommute && MI.isCommutable() && TII->commuteInstruction(&MI))
    foldImmediates(MI, TII, MRI, false);

}

bool SIShrinkInstructions::runOnMachineFunction(MachineFunction &MF) {
  MachineRegisterInfo &MRI = MF.getRegInfo();
  const SIInstrInfo *TII =
      static_cast<const SIInstrInfo *>(MF.getSubtarget().getInstrInfo());
  const SIRegisterInfo &TRI = TII->getRegisterInfo();
  std::vector<unsigned> I1Defs;

  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;

      // Try to use S_MOVK_I32, which will save 4 bytes for small immediates.
      if (MI.getOpcode() == AMDGPU::S_MOV_B32) {
        const MachineOperand &Src = MI.getOperand(1);

        if (Src.isImm()) {
          if (isInt<16>(Src.getImm()) && !TII->isInlineConstant(Src, 4))
            MI.setDesc(TII->get(AMDGPU::S_MOVK_I32));
        }

        continue;
      }

      if (!TII->hasVALU32BitEncoding(MI.getOpcode()))
        continue;

      if (!canShrink(MI, TII, TRI, MRI)) {
        // Try commuting the instruction and see if that enables us to shrink
        // it.
        if (!MI.isCommutable() || !TII->commuteInstruction(&MI) ||
            !canShrink(MI, TII, TRI, MRI))
          continue;
      }

      // getVOPe32 could be -1 here if we started with an instruction that had
      // a 32-bit encoding and then commuted it to an instruction that did not.
      if (!TII->hasVALU32BitEncoding(MI.getOpcode()))
        continue;

      int Op32 = AMDGPU::getVOPe32(MI.getOpcode());

      if (TII->isVOPC(Op32)) {
        unsigned DstReg = MI.getOperand(0).getReg();
        if (TargetRegisterInfo::isVirtualRegister(DstReg)) {
          // VOPC instructions can only write to the VCC register.  We can't
          // force them to use VCC here, because the register allocator has
          // trouble with sequences like this, which cause the allocator to run
          // out of registers if vreg0 and vreg1 belong to the VCCReg register
          // class:
          // vreg0 = VOPC;
          // vreg1 = VOPC;
          // S_AND_B64 vreg0, vreg1
          //
          // So, instead of forcing the instruction to write to VCC, we provide
          // a hint to the register allocator to use VCC and then we we will run
          // this pass again after RA and shrink it if it outputs to VCC.
          MRI.setRegAllocationHint(MI.getOperand(0).getReg(), 0, AMDGPU::VCC);
          continue;
        }
        if (DstReg != AMDGPU::VCC)
          continue;
      }

      // We can shrink this instruction
      DEBUG(dbgs() << "Shrinking "; MI.dump(); dbgs() << '\n';);

      MachineInstrBuilder Inst32 =
          BuildMI(MBB, I, MI.getDebugLoc(), TII->get(Op32));

      // dst
      Inst32.addOperand(MI.getOperand(0));

      Inst32.addOperand(*TII->getNamedOperand(MI, AMDGPU::OpName::src0));

      const MachineOperand *Src1 =
          TII->getNamedOperand(MI, AMDGPU::OpName::src1);
      if (Src1)
        Inst32.addOperand(*Src1);

      ++NumInstructionsShrunk;
      MI.eraseFromParent();

      foldImmediates(*Inst32, TII, MRI);
      DEBUG(dbgs() << "e32 MI = " << *Inst32 << '\n');


    }
  }
  return false;
}