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
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
|
//===----- AMDILBarrierDetect.cpp - Barrier Detect pass -*- C++ -*- ------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//==-----------------------------------------------------------------------===//
#define DEBUG_TYPE "BarrierDetect"
#ifdef DEBUG
#define DEBUGME (DebugFlag && isCurrentDebugType(DEBUG_TYPE))
#else
#define DEBUGME 0
#endif
#include "AMDILAlgorithms.tpp"
#include "AMDILCompilerWarnings.h"
#include "AMDILDevices.h"
#include "AMDILMachineFunctionInfo.h"
#include "AMDILSubtarget.h"
#include "AMDILTargetMachine.h"
#include "llvm/BasicBlock.h"
#include "llvm/CodeGen/MachineFunctionAnalysis.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/Function.h"
#include "llvm/Instructions.h"
#include "llvm/Module.h"
#include "llvm/Target/TargetMachine.h"
using namespace llvm;
// The barrier detect pass determines if a barrier has been duplicated in the
// source program which can cause undefined behaviour if more than a single
// wavefront is executed in a group. This is because LLVM does not have an
// execution barrier and if this barrier function gets duplicated, undefined
// behaviour can occur. In order to work around this, we detect the duplicated
// barrier and then make the work-group execute in a single wavefront mode,
// essentially making the barrier a no-op.
namespace
{
class LLVM_LIBRARY_VISIBILITY AMDILBarrierDetect : public FunctionPass
{
TargetMachine &TM;
static char ID;
public:
AMDILBarrierDetect(TargetMachine &TM AMDIL_OPT_LEVEL_DECL);
~AMDILBarrierDetect();
const char *getPassName() const;
bool runOnFunction(Function &F);
bool doInitialization(Module &M);
bool doFinalization(Module &M);
void getAnalysisUsage(AnalysisUsage &AU) const;
private:
bool detectBarrier(BasicBlock::iterator *BBI);
bool detectMemFence(BasicBlock::iterator *BBI);
bool mChanged;
SmallVector<int64_t, DEFAULT_VEC_SLOTS> bVecMap;
const AMDILSubtarget *mStm;
// Constants used to define memory type.
static const unsigned int LOCAL_MEM_FENCE = 1<<0;
static const unsigned int GLOBAL_MEM_FENCE = 1<<1;
static const unsigned int REGION_MEM_FENCE = 1<<2;
};
char AMDILBarrierDetect::ID = 0;
} // anonymouse namespace
namespace llvm
{
FunctionPass *
createAMDILBarrierDetect(TargetMachine &TM AMDIL_OPT_LEVEL_DECL)
{
return new AMDILBarrierDetect(TM AMDIL_OPT_LEVEL_VAR);
}
} // llvm namespace
AMDILBarrierDetect::AMDILBarrierDetect(TargetMachine &TM
AMDIL_OPT_LEVEL_DECL)
:
FunctionPass(ID),
TM(TM)
{
}
AMDILBarrierDetect::~AMDILBarrierDetect()
{
}
bool AMDILBarrierDetect::detectBarrier(BasicBlock::iterator *BBI)
{
SmallVector<int64_t, DEFAULT_VEC_SLOTS>::iterator bIter;
int64_t bID;
Instruction *inst = (*BBI);
CallInst *CI = dyn_cast<CallInst>(inst);
if (!CI || !CI->getNumOperands()) {
return false;
}
const Value *funcVal = CI->getOperand(CI->getNumOperands() - 1);
if (funcVal && strncmp(funcVal->getName().data(), "__amd_barrier", 13)) {
return false;
}
if (inst->getNumOperands() >= 3) {
const Value *V = inst->getOperand(0);
const ConstantInt *Cint = dyn_cast<ConstantInt>(V);
bID = Cint->getSExtValue();
bIter = std::find(bVecMap.begin(), bVecMap.end(), bID);
if (bIter == bVecMap.end()) {
bVecMap.push_back(bID);
} else {
if (mStm->device()->isSupported(AMDILDeviceInfo::BarrierDetect)) {
AMDILMachineFunctionInfo *MFI =
getAnalysis<MachineFunctionAnalysis>().getMF()
.getInfo<AMDILMachineFunctionInfo>();
MFI->addMetadata(";limitgroupsize");
MFI->addErrorMsg(amd::CompilerWarningMessage[BAD_BARRIER_OPT]);
}
}
}
if (mStm->device()->getGeneration() == AMDILDeviceInfo::HD4XXX) {
AMDILMachineFunctionInfo *MFI =
getAnalysis<MachineFunctionAnalysis>().getMF()
.getInfo<AMDILMachineFunctionInfo>();
MFI->addErrorMsg(amd::CompilerWarningMessage[LIMIT_BARRIER]);
MFI->addMetadata(";limitgroupsize");
MFI->setUsesLocal();
}
const Value *V = inst->getOperand(inst->getNumOperands()-2);
const ConstantInt *Cint = dyn_cast<ConstantInt>(V);
Function *iF = dyn_cast<Function>(inst->getOperand(inst->getNumOperands()-1));
Module *M = iF->getParent();
bID = Cint->getSExtValue();
if (bID > 0) {
const char *name = "barrier";
if (bID == GLOBAL_MEM_FENCE) {
name = "barrierGlobal";
} else if (bID == LOCAL_MEM_FENCE
&& mStm->device()->usesHardware(AMDILDeviceInfo::LocalMem)) {
name = "barrierLocal";
} else if (bID == REGION_MEM_FENCE
&& mStm->device()->usesHardware(AMDILDeviceInfo::RegionMem)) {
name = "barrierRegion";
}
Function *nF =
dyn_cast<Function>(M->getOrInsertFunction(name, iF->getFunctionType()));
inst->setOperand(inst->getNumOperands()-1, nF);
return false;
}
return false;
}
bool AMDILBarrierDetect::detectMemFence(BasicBlock::iterator *BBI)
{
int64_t bID;
Instruction *inst = (*BBI);
CallInst *CI = dyn_cast<CallInst>(inst);
if (!CI || CI->getNumOperands() != 2) {
return false;
}
const Value *V = inst->getOperand(inst->getNumOperands()-2);
const ConstantInt *Cint = dyn_cast<ConstantInt>(V);
Function *iF = dyn_cast<Function>(inst->getOperand(inst->getNumOperands()-1));
const char *fence_local_name;
const char *fence_global_name;
const char *fence_region_name;
const char* fence_name = "mem_fence";
if (!iF) {
return false;
}
if (strncmp(iF->getName().data(), "mem_fence", 9) == 0) {
fence_local_name = "mem_fence_local";
fence_global_name = "mem_fence_global";
fence_region_name = "mem_fence_region";
} else if (strncmp(iF->getName().data(), "read_mem_fence", 14) == 0) {
fence_local_name = "read_mem_fence_local";
fence_global_name = "read_mem_fence_global";
fence_region_name = "read_mem_fence_region";
} else if (strncmp(iF->getName().data(), "write_mem_fence", 15) == 0) {
fence_local_name = "write_mem_fence_local";
fence_global_name = "write_mem_fence_global";
fence_region_name = "write_mem_fence_region";
} else {
return false;
}
Module *M = iF->getParent();
bID = Cint->getSExtValue();
if (bID > 0) {
const char *name = fence_name;
if (bID == GLOBAL_MEM_FENCE) {
name = fence_global_name;
} else if (bID == LOCAL_MEM_FENCE
&& mStm->device()->usesHardware(AMDILDeviceInfo::LocalMem)) {
name = fence_local_name;
} else if (bID == REGION_MEM_FENCE
&& mStm->device()->usesHardware(AMDILDeviceInfo::RegionMem)) {
name = fence_region_name;
}
Function *nF =
dyn_cast<Function>(M->getOrInsertFunction(name, iF->getFunctionType()));
inst->setOperand(inst->getNumOperands()-1, nF);
return false;
}
return false;
}
bool AMDILBarrierDetect::runOnFunction(Function &MF)
{
mChanged = false;
bVecMap.clear();
mStm = &TM.getSubtarget<AMDILSubtarget>();
Function *F = &MF;
safeNestedForEach(F->begin(), F->end(), F->begin()->begin(),
std::bind1st(
std::mem_fun(
&AMDILBarrierDetect::detectBarrier), this));
safeNestedForEach(F->begin(), F->end(), F->begin()->begin(),
std::bind1st(
std::mem_fun(
&AMDILBarrierDetect::detectMemFence), this));
return mChanged;
}
const char* AMDILBarrierDetect::getPassName() const
{
return "AMDIL Barrier Detect Pass";
}
bool AMDILBarrierDetect::doInitialization(Module &M)
{
return false;
}
bool AMDILBarrierDetect::doFinalization(Module &M)
{
return false;
}
void AMDILBarrierDetect::getAnalysisUsage(AnalysisUsage &AU) const
{
AU.addRequired<MachineFunctionAnalysis>();
FunctionPass::getAnalysisUsage(AU);
AU.setPreservesAll();
}
|