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
|
//===----- HexagonLoopAlign.cpp - Generate loop alignment directives -----===//
//
// 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
//
//===----------------------------------------------------------------------===//
// Inspect a basic block and if its single basic block loop with a small
// number of instructions, set the prefLoopAlignment to 32 bytes (5).
//===----------------------------------------------------------------------===//
#include "Hexagon.h"
#include "HexagonTargetMachine.h"
#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
#include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
#include "llvm/Support/Debug.h"
#define DEBUG_TYPE "hexagon-loop-align"
using namespace llvm;
static cl::opt<bool>
DisableLoopAlign("disable-hexagon-loop-align", cl::Hidden,
cl::desc("Disable Hexagon loop alignment pass"));
static cl::opt<uint32_t> HVXLoopAlignLimitUB(
"hexagon-hvx-loop-align-limit-ub", cl::Hidden, cl::init(16),
cl::desc("Set hexagon hvx loop upper bound align limit"));
static cl::opt<uint32_t> TinyLoopAlignLimitUB(
"hexagon-tiny-loop-align-limit-ub", cl::Hidden, cl::init(16),
cl::desc("Set hexagon tiny-core loop upper bound align limit"));
static cl::opt<uint32_t>
LoopAlignLimitUB("hexagon-loop-align-limit-ub", cl::Hidden, cl::init(8),
cl::desc("Set hexagon loop upper bound align limit"));
static cl::opt<uint32_t>
LoopAlignLimitLB("hexagon-loop-align-limit-lb", cl::Hidden, cl::init(4),
cl::desc("Set hexagon loop lower bound align limit"));
static cl::opt<uint32_t>
LoopBndlAlignLimit("hexagon-loop-bundle-align-limit", cl::Hidden,
cl::init(4),
cl::desc("Set hexagon loop align bundle limit"));
static cl::opt<uint32_t> TinyLoopBndlAlignLimit(
"hexagon-tiny-loop-bundle-align-limit", cl::Hidden, cl::init(8),
cl::desc("Set hexagon tiny-core loop align bundle limit"));
static cl::opt<uint32_t>
LoopEdgeThreshold("hexagon-loop-edge-threshold", cl::Hidden, cl::init(7500),
cl::desc("Set hexagon loop align edge threshold"));
namespace {
class HexagonLoopAlign : public MachineFunctionPass {
const HexagonSubtarget *HST = nullptr;
const TargetMachine *HTM = nullptr;
const HexagonInstrInfo *HII = nullptr;
public:
static char ID;
HexagonLoopAlign() : MachineFunctionPass(ID) {}
bool shouldBalignLoop(MachineBasicBlock &BB, bool AboveThres);
bool isSingleLoop(MachineBasicBlock &MBB);
bool attemptToBalignSmallLoop(MachineFunction &MF, MachineBasicBlock &MBB);
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.addRequired<MachineBranchProbabilityInfoWrapperPass>();
AU.addRequired<MachineBlockFrequencyInfoWrapperPass>();
MachineFunctionPass::getAnalysisUsage(AU);
}
StringRef getPassName() const override { return "Hexagon LoopAlign pass"; }
bool runOnMachineFunction(MachineFunction &MF) override;
};
char HexagonLoopAlign::ID = 0;
bool HexagonLoopAlign::shouldBalignLoop(MachineBasicBlock &BB,
bool AboveThres) {
bool isVec = false;
unsigned InstCnt = 0;
unsigned BndlCnt = 0;
for (MachineBasicBlock::instr_iterator II = BB.instr_begin(),
IE = BB.instr_end();
II != IE; ++II) {
// End if the instruction is endloop.
if (HII->isEndLoopN(II->getOpcode()))
break;
// Count the number of bundles.
if (II->isBundle()) {
BndlCnt++;
continue;
}
// Skip over debug instructions.
if (II->isDebugInstr())
continue;
// Check if there are any HVX instructions in loop.
isVec |= HII->isHVXVec(*II);
// Count the number of instructions.
InstCnt++;
}
LLVM_DEBUG({
dbgs() << "Bundle Count : " << BndlCnt << "\n";
dbgs() << "Instruction Count : " << InstCnt << "\n";
});
unsigned LimitUB = 0;
unsigned LimitBndl = LoopBndlAlignLimit;
// The conditions in the order of priority.
if (HST->isTinyCore()) {
LimitUB = TinyLoopAlignLimitUB;
LimitBndl = TinyLoopBndlAlignLimit;
} else if (isVec)
LimitUB = HVXLoopAlignLimitUB;
else if (AboveThres)
LimitUB = LoopAlignLimitUB;
// if the upper bound is not set to a value, implies we didn't meet
// the criteria.
if (LimitUB == 0)
return false;
return InstCnt >= LoopAlignLimitLB && InstCnt <= LimitUB &&
BndlCnt <= LimitBndl;
}
bool HexagonLoopAlign::isSingleLoop(MachineBasicBlock &MBB) {
int Succs = MBB.succ_size();
return (MBB.isSuccessor(&MBB) && (Succs == 2));
}
bool HexagonLoopAlign::attemptToBalignSmallLoop(MachineFunction &MF,
MachineBasicBlock &MBB) {
if (!isSingleLoop(MBB))
return false;
const MachineBranchProbabilityInfo *MBPI =
&getAnalysis<MachineBranchProbabilityInfoWrapperPass>().getMBPI();
const MachineBlockFrequencyInfo *MBFI =
&getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI();
// Compute frequency of back edge,
BlockFrequency BlockFreq = MBFI->getBlockFreq(&MBB);
BranchProbability BrProb = MBPI->getEdgeProbability(&MBB, &MBB);
BlockFrequency EdgeFreq = BlockFreq * BrProb;
LLVM_DEBUG({
dbgs() << "Loop Align Pass:\n";
dbgs() << "\tedge with freq(" << EdgeFreq.getFrequency() << ")\n";
});
bool AboveThres = EdgeFreq.getFrequency() > LoopEdgeThreshold;
if (shouldBalignLoop(MBB, AboveThres)) {
// We found a loop, change its alignment to be 32 (5).
MBB.setAlignment(llvm::Align(1 << 5));
return true;
}
return false;
}
// Inspect each basic block, and if its a single BB loop, see if it
// meets the criteria for increasing alignment to 32.
bool HexagonLoopAlign::runOnMachineFunction(MachineFunction &MF) {
HST = &MF.getSubtarget<HexagonSubtarget>();
HII = HST->getInstrInfo();
HTM = &MF.getTarget();
if (skipFunction(MF.getFunction()))
return false;
if (DisableLoopAlign)
return false;
// This optimization is performed at
// i) -O2 and above, and when the loop has a HVX instruction.
// ii) -O3
if (HST->useHVXOps()) {
if (HTM->getOptLevel() < CodeGenOptLevel::Default)
return false;
} else {
if (HTM->getOptLevel() < CodeGenOptLevel::Aggressive)
return false;
}
bool Changed = false;
for (MachineFunction::iterator MBBi = MF.begin(), MBBe = MF.end();
MBBi != MBBe; ++MBBi) {
MachineBasicBlock &MBB = *MBBi;
Changed |= attemptToBalignSmallLoop(MF, MBB);
}
return Changed;
}
} // namespace
INITIALIZE_PASS(HexagonLoopAlign, "hexagon-loop-align",
"Hexagon LoopAlign pass", false, false)
//===----------------------------------------------------------------------===//
// Public Constructor Functions
//===----------------------------------------------------------------------===//
FunctionPass *llvm::createHexagonLoopAlign() { return new HexagonLoopAlign(); }
|
