1//===-- NVPTXPeephole.cpp - NVPTX Peephole Optimiztions -------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// In NVPTX, NVPTXFrameLowering will emit following instruction at the beginning
10// of a MachineFunction.
11//
12// mov %SPL, %depot
13// cvta.local %SP, %SPL
14//
15// Because Frame Index is a generic address and alloca can only return generic
16// pointer, without this pass the instructions producing alloca'ed address will
17// be based on %SP. NVPTXLowerAlloca tends to help replace store and load on
18// this address with their .local versions, but this may introduce a lot of
19// cvta.to.local instructions. Performance can be improved if we avoid casting
20// address back and forth and directly calculate local address based on %SPL.
21// This peephole pass optimizes these cases, for example
22//
23// It will transform the following pattern
24// %0 = LEA_ADDRi64 %VRFrame64, 4
25// %1 = cvta_to_local_64 %0
26//
27// into
28// %1 = LEA_ADDRi64 %VRFrameLocal64, 4
29//
30// %VRFrameLocal64 is the virtual register name of %SPL
31//
32//===----------------------------------------------------------------------===//
33
34#include "NVPTX.h"
35#include "NVPTXRegisterInfo.h"
36#include "NVPTXSubtarget.h"
37#include "llvm/CodeGen/MachineFunctionPass.h"
38#include "llvm/CodeGen/MachineInstrBuilder.h"
39#include "llvm/CodeGen/MachineRegisterInfo.h"
40#include "llvm/CodeGen/TargetInstrInfo.h"
41#include "llvm/CodeGen/TargetRegisterInfo.h"
42
43using namespace llvm;
44
45#define DEBUG_TYPE "nvptx-peephole"
46
47namespace llvm {
48void initializeNVPTXPeepholePass(PassRegistry &);
49}
50
51namespace {
52struct NVPTXPeephole : public MachineFunctionPass {
53 public:
54 static char ID;
55 NVPTXPeephole() : MachineFunctionPass(ID) {
56 initializeNVPTXPeepholePass(*PassRegistry::getPassRegistry());
57 }
58
59 bool runOnMachineFunction(MachineFunction &MF) override;
60
61 StringRef getPassName() const override {
62 return "NVPTX optimize redundant cvta.to.local instruction";
63 }
64
65 void getAnalysisUsage(AnalysisUsage &AU) const override {
66 MachineFunctionPass::getAnalysisUsage(AU);
67 }
68};
69}
70
71char NVPTXPeephole::ID = 0;
72
73INITIALIZE_PASS(NVPTXPeephole, "nvptx-peephole", "NVPTX Peephole", false, false)
74
75static bool isCVTAToLocalCombinationCandidate(MachineInstr &Root) {
76 auto &MBB = *Root.getParent();
77 auto &MF = *MBB.getParent();
78 // Check current instruction is cvta.to.local
79 if (Root.getOpcode() != NVPTX::cvta_to_local_64 &&
80 Root.getOpcode() != NVPTX::cvta_to_local)
81 return false;
82
83 auto &Op = Root.getOperand(i: 1);
84 const auto &MRI = MF.getRegInfo();
85 MachineInstr *GenericAddrDef = nullptr;
86 if (Op.isReg() && Op.getReg().isVirtual()) {
87 GenericAddrDef = MRI.getUniqueVRegDef(Reg: Op.getReg());
88 }
89
90 // Check the register operand is uniquely defined by LEA_ADDRi instruction
91 if (!GenericAddrDef || GenericAddrDef->getParent() != &MBB ||
92 (GenericAddrDef->getOpcode() != NVPTX::LEA_ADDRi64 &&
93 GenericAddrDef->getOpcode() != NVPTX::LEA_ADDRi)) {
94 return false;
95 }
96
97 const NVPTXRegisterInfo *NRI =
98 MF.getSubtarget<NVPTXSubtarget>().getRegisterInfo();
99
100 // Check the LEA_ADDRi operand is Frame index
101 auto &BaseAddrOp = GenericAddrDef->getOperand(i: 1);
102 if (BaseAddrOp.isReg() && BaseAddrOp.getReg() == NRI->getFrameRegister(MF)) {
103 return true;
104 }
105
106 return false;
107}
108
109static void CombineCVTAToLocal(MachineInstr &Root) {
110 auto &MBB = *Root.getParent();
111 auto &MF = *MBB.getParent();
112 const auto &MRI = MF.getRegInfo();
113 const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
114 auto &Prev = *MRI.getUniqueVRegDef(Reg: Root.getOperand(i: 1).getReg());
115
116 const NVPTXRegisterInfo *NRI =
117 MF.getSubtarget<NVPTXSubtarget>().getRegisterInfo();
118
119 MachineInstrBuilder MIB =
120 BuildMI(MF, MIMD: Root.getDebugLoc(), MCID: TII->get(Opcode: Prev.getOpcode()),
121 DestReg: Root.getOperand(i: 0).getReg())
122 .addReg(RegNo: NRI->getFrameLocalRegister(MF))
123 .add(MO: Prev.getOperand(i: 2));
124
125 MBB.insert(I: (MachineBasicBlock::iterator)&Root, MI: MIB);
126
127 // Check if MRI has only one non dbg use, which is Root
128 if (MRI.hasOneNonDBGUse(RegNo: Prev.getOperand(i: 0).getReg())) {
129 Prev.eraseFromParent();
130 }
131 Root.eraseFromParent();
132}
133
134bool NVPTXPeephole::runOnMachineFunction(MachineFunction &MF) {
135 if (skipFunction(F: MF.getFunction()))
136 return false;
137
138 bool Changed = false;
139 // Loop over all of the basic blocks.
140 for (auto &MBB : MF) {
141 // Traverse the basic block.
142 auto BlockIter = MBB.begin();
143
144 while (BlockIter != MBB.end()) {
145 auto &MI = *BlockIter++;
146 if (isCVTAToLocalCombinationCandidate(Root&: MI)) {
147 CombineCVTAToLocal(Root&: MI);
148 Changed = true;
149 }
150 } // Instruction
151 } // Basic Block
152
153 const NVPTXRegisterInfo *NRI =
154 MF.getSubtarget<NVPTXSubtarget>().getRegisterInfo();
155
156 // Remove unnecessary %VRFrame = cvta.local %VRFrameLocal
157 const auto &MRI = MF.getRegInfo();
158 if (MRI.use_empty(RegNo: NRI->getFrameRegister(MF))) {
159 if (auto MI = MRI.getUniqueVRegDef(Reg: NRI->getFrameRegister(MF))) {
160 MI->eraseFromParent();
161 }
162 }
163
164 return Changed;
165}
166
167MachineFunctionPass *llvm::createNVPTXPeephole() { return new NVPTXPeephole(); }
168