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
|
//===- GCNCreateVOPD.cpp - Create VOPD Instructions ----------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
/// \file
/// Combine VALU pairs into VOPD instructions
/// Only works on wave32
/// Has register requirements, we reject creating VOPD if the requirements are
/// not met.
/// shouldCombineVOPD mutator in postRA machine scheduler puts candidate
/// instructions for VOPD back-to-back
///
//
//===----------------------------------------------------------------------===//
#include "AMDGPU.h"
#include "GCNSubtarget.h"
#include "GCNVOPDUtils.h"
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
#include "SIInstrInfo.h"
#include "Utils/AMDGPUBaseInfo.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/Debug.h"
#include <utility>
#define DEBUG_TYPE "gcn-create-vopd"
STATISTIC(NumVOPDCreated, "Number of VOPD Insts Created.");
using namespace llvm;
namespace {
class GCNCreateVOPD : public MachineFunctionPass {
private:
class VOPDCombineInfo {
public:
VOPDCombineInfo() {}
VOPDCombineInfo(MachineInstr *First, MachineInstr *Second)
: FirstMI(First), SecondMI(Second) {}
MachineInstr *FirstMI;
MachineInstr *SecondMI;
};
public:
static char ID;
const GCNSubtarget *ST = nullptr;
GCNCreateVOPD() : MachineFunctionPass(ID) {}
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.setPreservesCFG();
MachineFunctionPass::getAnalysisUsage(AU);
}
StringRef getPassName() const override {
return "GCN Create VOPD Instructions";
}
bool doReplace(const SIInstrInfo *SII, VOPDCombineInfo &CI) {
auto *FirstMI = CI.FirstMI;
auto *SecondMI = CI.SecondMI;
unsigned Opc1 = FirstMI->getOpcode();
unsigned Opc2 = SecondMI->getOpcode();
unsigned EncodingFamily =
AMDGPU::getVOPDEncodingFamily(SII->getSubtarget());
int NewOpcode =
AMDGPU::getVOPDFull(AMDGPU::getVOPDOpcode(Opc1),
AMDGPU::getVOPDOpcode(Opc2), EncodingFamily);
assert(NewOpcode != -1 &&
"Should have previously determined this as a possible VOPD\n");
auto VOPDInst = BuildMI(*FirstMI->getParent(), FirstMI,
FirstMI->getDebugLoc(), SII->get(NewOpcode))
.setMIFlags(FirstMI->getFlags() | SecondMI->getFlags());
namespace VOPD = AMDGPU::VOPD;
MachineInstr *MI[] = {FirstMI, SecondMI};
auto InstInfo =
AMDGPU::getVOPDInstInfo(FirstMI->getDesc(), SecondMI->getDesc());
for (auto CompIdx : VOPD::COMPONENTS) {
auto MCOprIdx = InstInfo[CompIdx].getIndexOfDstInMCOperands();
VOPDInst.add(MI[CompIdx]->getOperand(MCOprIdx));
}
for (auto CompIdx : VOPD::COMPONENTS) {
auto CompSrcOprNum = InstInfo[CompIdx].getCompSrcOperandsNum();
for (unsigned CompSrcIdx = 0; CompSrcIdx < CompSrcOprNum; ++CompSrcIdx) {
auto MCOprIdx = InstInfo[CompIdx].getIndexOfSrcInMCOperands(CompSrcIdx);
VOPDInst.add(MI[CompIdx]->getOperand(MCOprIdx));
}
}
SII->fixImplicitOperands(*VOPDInst);
for (auto CompIdx : VOPD::COMPONENTS)
VOPDInst.copyImplicitOps(*MI[CompIdx]);
LLVM_DEBUG(dbgs() << "VOPD Fused: " << *VOPDInst << " from\tX: "
<< *CI.FirstMI << "\tY: " << *CI.SecondMI << "\n");
for (auto CompIdx : VOPD::COMPONENTS)
MI[CompIdx]->eraseFromParent();
++NumVOPDCreated;
return true;
}
bool runOnMachineFunction(MachineFunction &MF) override {
if (skipFunction(MF.getFunction()))
return false;
ST = &MF.getSubtarget<GCNSubtarget>();
if (!AMDGPU::hasVOPD(*ST) || !ST->isWave32())
return false;
LLVM_DEBUG(dbgs() << "CreateVOPD Pass:\n");
const SIInstrInfo *SII = ST->getInstrInfo();
bool Changed = false;
SmallVector<VOPDCombineInfo> ReplaceCandidates;
for (auto &MBB : MF) {
auto MII = MBB.begin(), E = MBB.end();
while (MII != E) {
auto *FirstMI = &*MII;
MII = next_nodbg(MII, MBB.end());
if (MII == MBB.end())
break;
if (FirstMI->isDebugInstr())
continue;
auto *SecondMI = &*MII;
unsigned Opc = FirstMI->getOpcode();
unsigned Opc2 = SecondMI->getOpcode();
llvm::AMDGPU::CanBeVOPD FirstCanBeVOPD = AMDGPU::getCanBeVOPD(Opc);
llvm::AMDGPU::CanBeVOPD SecondCanBeVOPD = AMDGPU::getCanBeVOPD(Opc2);
VOPDCombineInfo CI;
if (FirstCanBeVOPD.X && SecondCanBeVOPD.Y)
CI = VOPDCombineInfo(FirstMI, SecondMI);
else if (FirstCanBeVOPD.Y && SecondCanBeVOPD.X)
CI = VOPDCombineInfo(SecondMI, FirstMI);
else
continue;
// checkVOPDRegConstraints cares about program order, but doReplace
// cares about X-Y order in the constituted VOPD
if (llvm::checkVOPDRegConstraints(*SII, *FirstMI, *SecondMI)) {
ReplaceCandidates.push_back(CI);
++MII;
}
}
}
for (auto &CI : ReplaceCandidates) {
Changed |= doReplace(SII, CI);
}
return Changed;
}
};
} // namespace
char GCNCreateVOPD::ID = 0;
char &llvm::GCNCreateVOPDID = GCNCreateVOPD::ID;
INITIALIZE_PASS(GCNCreateVOPD, DEBUG_TYPE, "GCN Create VOPD Instructions",
false, false)
|