xref: /llvm-project/llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp (revision 0ee250eee84da1fbcf94eaf5595420abdb8228e5)
1 //===-- SILowerI1Copies.cpp - Lower I1 Copies -----------------------------===//
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 /// i1 values are usually inserted by the CFG Structurize pass and they are
9 /// unique in that they can be copied from VALU to SALU registers.
10 /// This is not possible for any other value type.  Since there are no
11 /// MOV instructions for i1, we to use V_CMP_* and V_CNDMASK to move the i1.
12 ///
13 //===----------------------------------------------------------------------===//
14 //
15 
16 #define DEBUG_TYPE "si-i1-copies"
17 #include "AMDGPU.h"
18 #include "AMDGPUSubtarget.h"
19 #include "SIInstrInfo.h"
20 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
21 #include "llvm/CodeGen/MachineFunctionPass.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/IR/LLVMContext.h"
25 #include "llvm/IR/Function.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Target/TargetMachine.h"
28 
29 using namespace llvm;
30 
31 namespace {
32 
33 class SILowerI1Copies : public MachineFunctionPass {
34 public:
35   static char ID;
36 
37 public:
38   SILowerI1Copies() : MachineFunctionPass(ID) {
39     initializeSILowerI1CopiesPass(*PassRegistry::getPassRegistry());
40   }
41 
42   bool runOnMachineFunction(MachineFunction &MF) override;
43 
44   StringRef getPassName() const override { return "SI Lower i1 Copies"; }
45 
46   void getAnalysisUsage(AnalysisUsage &AU) const override {
47     AU.setPreservesCFG();
48     MachineFunctionPass::getAnalysisUsage(AU);
49   }
50 };
51 
52 } // End anonymous namespace.
53 
54 INITIALIZE_PASS(SILowerI1Copies, DEBUG_TYPE,
55                 "SI Lower i1 Copies", false, false)
56 
57 char SILowerI1Copies::ID = 0;
58 
59 char &llvm::SILowerI1CopiesID = SILowerI1Copies::ID;
60 
61 FunctionPass *llvm::createSILowerI1CopiesPass() {
62   return new SILowerI1Copies();
63 }
64 
65 bool SILowerI1Copies::runOnMachineFunction(MachineFunction &MF) {
66   MachineRegisterInfo &MRI = MF.getRegInfo();
67   const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
68   const SIInstrInfo *TII = ST.getInstrInfo();
69   const TargetRegisterInfo *TRI = &TII->getRegisterInfo();
70 
71   std::vector<unsigned> I1Defs;
72 
73   for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
74                                                   BI != BE; ++BI) {
75 
76     MachineBasicBlock &MBB = *BI;
77     MachineBasicBlock::iterator I, Next;
78     for (I = MBB.begin(); I != MBB.end(); I = Next) {
79       Next = std::next(I);
80       MachineInstr &MI = *I;
81 
82       if (MI.getOpcode() == AMDGPU::IMPLICIT_DEF) {
83         unsigned Reg = MI.getOperand(0).getReg();
84         const TargetRegisterClass *RC = MRI.getRegClass(Reg);
85         if (RC == &AMDGPU::VReg_1RegClass)
86           MRI.setRegClass(Reg, &AMDGPU::SReg_64RegClass);
87         continue;
88       }
89 
90       if (MI.getOpcode() != AMDGPU::COPY)
91         continue;
92 
93       const MachineOperand &Dst = MI.getOperand(0);
94       const MachineOperand &Src = MI.getOperand(1);
95 
96       if (!TargetRegisterInfo::isVirtualRegister(Src.getReg()) ||
97           !TargetRegisterInfo::isVirtualRegister(Dst.getReg()))
98         continue;
99 
100       const TargetRegisterClass *DstRC = MRI.getRegClass(Dst.getReg());
101       const TargetRegisterClass *SrcRC = MRI.getRegClass(Src.getReg());
102 
103       DebugLoc DL = MI.getDebugLoc();
104       MachineInstr *DefInst = MRI.getUniqueVRegDef(Src.getReg());
105       if (DstRC == &AMDGPU::VReg_1RegClass &&
106           TRI->getCommonSubClass(SrcRC, &AMDGPU::SGPR_64RegClass)) {
107         I1Defs.push_back(Dst.getReg());
108 
109         if (DefInst->getOpcode() == AMDGPU::S_MOV_B64) {
110           if (DefInst->getOperand(1).isImm()) {
111             I1Defs.push_back(Dst.getReg());
112 
113             int64_t Val = DefInst->getOperand(1).getImm();
114             assert(Val == 0 || Val == -1);
115 
116             BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_MOV_B32_e32))
117               .addOperand(Dst)
118               .addImm(Val);
119             MI.eraseFromParent();
120             continue;
121           }
122         }
123 
124         BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_CNDMASK_B32_e64))
125           .addOperand(Dst)
126           .addImm(0)
127           .addImm(-1)
128           .addOperand(Src);
129         MI.eraseFromParent();
130       } else if (TRI->getCommonSubClass(DstRC, &AMDGPU::SGPR_64RegClass) &&
131                  SrcRC == &AMDGPU::VReg_1RegClass) {
132         if (DefInst->getOpcode() == AMDGPU::V_CNDMASK_B32_e64 &&
133             DefInst->getOperand(1).isImm() && DefInst->getOperand(2).isImm() &&
134             DefInst->getOperand(1).getImm() == 0 &&
135             DefInst->getOperand(2).getImm() != 0 &&
136             DefInst->getOperand(3).isReg() &&
137             TargetRegisterInfo::isVirtualRegister(
138               DefInst->getOperand(3).getReg()) &&
139             TRI->getCommonSubClass(
140               MRI.getRegClass(DefInst->getOperand(3).getReg()),
141               &AMDGPU::SGPR_64RegClass)) {
142           BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_AND_B64))
143             .addOperand(Dst)
144             .addReg(AMDGPU::EXEC)
145             .addOperand(DefInst->getOperand(3));
146         } else {
147           BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_CMP_NE_U32_e64))
148             .addOperand(Dst)
149             .addOperand(Src)
150             .addImm(0);
151         }
152         MI.eraseFromParent();
153       }
154     }
155   }
156 
157   for (unsigned Reg : I1Defs)
158     MRI.setRegClass(Reg, &AMDGPU::VGPR_32RegClass);
159 
160   return false;
161 }
162