R600/SI: Refactor SIFoldOperands to simplify immediate folding
[oota-llvm.git] / lib / Target / R600 / SIFoldOperands.cpp
1 //===-- SIFoldOperands.cpp - Fold operands --- ----------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 /// \file
9 //===----------------------------------------------------------------------===//
10 //
11
12 #include "AMDGPU.h"
13 #include "AMDGPUSubtarget.h"
14 #include "SIInstrInfo.h"
15 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
16 #include "llvm/CodeGen/MachineDominators.h"
17 #include "llvm/CodeGen/MachineFunctionPass.h"
18 #include "llvm/CodeGen/MachineInstrBuilder.h"
19 #include "llvm/CodeGen/MachineRegisterInfo.h"
20 #include "llvm/IR/LLVMContext.h"
21 #include "llvm/IR/Function.h"
22 #include "llvm/Support/Debug.h"
23 #include "llvm/Target/TargetMachine.h"
24
25 #define DEBUG_TYPE "si-fold-operands"
26 using namespace llvm;
27
28 namespace {
29
30 class SIFoldOperands : public MachineFunctionPass {
31 public:
32   static char ID;
33
34 public:
35   SIFoldOperands() : MachineFunctionPass(ID) {
36     initializeSIFoldOperandsPass(*PassRegistry::getPassRegistry());
37   }
38
39   bool runOnMachineFunction(MachineFunction &MF) override;
40
41   const char *getPassName() const override {
42     return "SI Fold Operands";
43   }
44
45   void getAnalysisUsage(AnalysisUsage &AU) const override {
46     AU.addRequired<MachineDominatorTree>();
47     AU.setPreservesCFG();
48     MachineFunctionPass::getAnalysisUsage(AU);
49   }
50 };
51
52 struct FoldCandidate {
53   MachineInstr *UseMI;
54   unsigned UseOpNo;
55   MachineOperand *OpToFold;
56   uint64_t ImmToFold;
57
58   FoldCandidate(MachineInstr *MI, unsigned OpNo, MachineOperand *FoldOp) :
59       UseMI(MI), UseOpNo(OpNo), OpToFold(FoldOp), ImmToFold(0) { }
60
61   FoldCandidate(MachineInstr *MI, unsigned OpNo, uint64_t Imm) :
62       UseMI(MI), UseOpNo(OpNo), OpToFold(nullptr), ImmToFold(Imm) { }
63
64   bool isImm() const {
65     return !OpToFold;
66   }
67 };
68
69 } // End anonymous namespace.
70
71 INITIALIZE_PASS_BEGIN(SIFoldOperands, DEBUG_TYPE,
72                       "SI Fold Operands", false, false)
73 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
74 INITIALIZE_PASS_END(SIFoldOperands, DEBUG_TYPE,
75                     "SI Fold Operands", false, false)
76
77 char SIFoldOperands::ID = 0;
78
79 char &llvm::SIFoldOperandsID = SIFoldOperands::ID;
80
81 FunctionPass *llvm::createSIFoldOperandsPass() {
82   return new SIFoldOperands();
83 }
84
85 static bool isSafeToFold(unsigned Opcode) {
86   switch(Opcode) {
87   case AMDGPU::V_MOV_B32_e32:
88   case AMDGPU::V_MOV_B32_e64:
89   case AMDGPU::S_MOV_B32:
90   case AMDGPU::S_MOV_B64:
91   case AMDGPU::COPY:
92     return true;
93   default:
94     return false;
95   }
96 }
97
98 static bool updateOperand(FoldCandidate &Fold,
99                           const TargetRegisterInfo &TRI) {
100   MachineInstr *MI = Fold.UseMI;
101   MachineOperand &Old = MI->getOperand(Fold.UseOpNo);
102   assert(Old.isReg());
103
104   if (Fold.isImm()) {
105     Old.ChangeToImmediate(Fold.ImmToFold);
106     return true;
107   }
108
109   MachineOperand *New = Fold.OpToFold;
110   if (TargetRegisterInfo::isVirtualRegister(Old.getReg()) &&
111       TargetRegisterInfo::isVirtualRegister(New->getReg())) {
112     Old.substVirtReg(New->getReg(), New->getSubReg(), TRI);
113     return true;
114   }
115
116   // FIXME: Handle physical registers.
117
118   return false;
119 }
120
121 bool SIFoldOperands::runOnMachineFunction(MachineFunction &MF) {
122   MachineRegisterInfo &MRI = MF.getRegInfo();
123   const SIInstrInfo *TII =
124       static_cast<const SIInstrInfo *>(MF.getSubtarget().getInstrInfo());
125   const SIRegisterInfo &TRI = TII->getRegisterInfo();
126
127   for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
128                                                   BI != BE; ++BI) {
129
130     MachineBasicBlock &MBB = *BI;
131     MachineBasicBlock::iterator I, Next;
132     for (I = MBB.begin(); I != MBB.end(); I = Next) {
133       Next = std::next(I);
134       MachineInstr &MI = *I;
135
136       if (!isSafeToFold(MI.getOpcode()))
137         continue;
138
139       MachineOperand &OpToFold = MI.getOperand(1);
140
141       // FIXME: Fold operands with subregs.
142       if (OpToFold.isReg() &&
143           (!TargetRegisterInfo::isVirtualRegister(OpToFold.getReg()) ||
144            OpToFold.getSubReg()))
145         continue;
146
147       std::vector<FoldCandidate> FoldList;
148       for (MachineRegisterInfo::use_iterator
149            Use = MRI.use_begin(MI.getOperand(0).getReg()), E = MRI.use_end();
150            Use != E; ++Use) {
151
152         MachineInstr *UseMI = Use->getParent();
153         const MachineOperand &UseOp = UseMI->getOperand(Use.getOperandNo());
154
155         // FIXME: Fold operands with subregs.
156         if (UseOp.isReg() && UseOp.getSubReg()) {
157           continue;
158         }
159
160         bool FoldingImm = OpToFold.isImm() || OpToFold.isFPImm();
161
162         // In order to fold immediates into copies, we need to change the
163         // copy to a MOV.
164         if (FoldingImm && UseMI->getOpcode() == AMDGPU::COPY) {
165           const TargetRegisterClass *TRC =
166               MRI.getRegClass(UseMI->getOperand(0).getReg());
167
168           if (TRC->getSize() == 4) {
169             if (TRI.isSGPRClass(TRC))
170               UseMI->setDesc(TII->get(AMDGPU::S_MOV_B32));
171             else
172               UseMI->setDesc(TII->get(AMDGPU::V_MOV_B32_e32));
173           } else if (TRC->getSize() == 8 && TRI.isSGPRClass(TRC)) {
174             UseMI->setDesc(TII->get(AMDGPU::S_MOV_B64));
175           } else {
176             continue;
177           }
178         }
179
180         const MCInstrDesc &UseDesc = UseMI->getDesc();
181
182         // Don't fold into target independent nodes.  Target independent opcodes
183         // don't have defined register classes.
184         if (UseDesc.isVariadic() ||
185             UseDesc.OpInfo[Use.getOperandNo()].RegClass == -1)
186           continue;
187
188         if (FoldingImm) {
189           uint64_t Imm;
190           if (OpToFold.isFPImm()) {
191             Imm = OpToFold.getFPImm()->getValueAPF().bitcastToAPInt().getSExtValue();
192           } else {
193             Imm = OpToFold.getImm();
194           }
195
196           const MachineOperand ImmOp = MachineOperand::CreateImm(Imm);
197           if (TII->isOperandLegal(UseMI, Use.getOperandNo(), &ImmOp)) {
198             FoldList.push_back(FoldCandidate(UseMI, Use.getOperandNo(), Imm));
199             continue;
200           }
201         }
202
203         // Normal substitution with registers
204         if (TII->isOperandLegal(UseMI, Use.getOperandNo(), &OpToFold)) {
205           FoldList.push_back(FoldCandidate(UseMI, Use.getOperandNo(), &OpToFold));
206           continue;
207         }
208
209         // FIXME: We could commute the instruction to create more opportunites
210         // for folding.  This will only be useful if we have 32-bit instructions.
211
212         // FIXME: We could try to change the instruction from 64-bit to 32-bit
213         // to enable more folding opportunites.  The shrink operands pass
214         // already does this.
215       }
216
217       for (FoldCandidate &Fold : FoldList) {
218         if (updateOperand(Fold, TRI)) {
219           // Clear kill flags.
220           if (!Fold.isImm()) {
221             assert(Fold.OpToFold && Fold.OpToFold->isReg());
222             Fold.OpToFold->setIsKill(false);
223           }
224           DEBUG(dbgs() << "Folded source from " << MI << " into OpNo " <<
225                 Fold.UseOpNo << " of " << *Fold.UseMI << '\n');
226         }
227       }
228     }
229   }
230   return false;
231 }