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
|
// LLVM function pass that adds implicit barriers to loops if it sees
// beneficial.
//
// Copyright (c) 2012-2014 Pekka Jääskeläinen / TUT
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.
#include "config.h"
#include "ImplicitLoopBarriers.h"
#include "Barrier.h"
#include "Workgroup.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#if (defined LLVM_3_1 or defined LLVM_3_2)
#include "llvm/Constants.h"
#include "llvm/Instructions.h"
#include "llvm/Module.h"
#else
#include "llvm/IR/Constants.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Module.h"
#endif
#include "VariableUniformityAnalysis.h"
#include <iostream>
//#define DEBUG_ILOOP_BARRIERS
using namespace llvm;
using namespace pocl;
namespace {
static
RegisterPass<ImplicitLoopBarriers> X("implicit-loop-barriers",
"Adds implicit barriers to loops");
}
char ImplicitLoopBarriers::ID = 0;
void
ImplicitLoopBarriers::getAnalysisUsage(AnalysisUsage &AU) const
{
AU.addRequired<DominatorTree>();
AU.addPreserved<DominatorTree>();
AU.addRequired<VariableUniformityAnalysis>();
AU.addPreserved<VariableUniformityAnalysis>();
}
bool
ImplicitLoopBarriers::runOnLoop(Loop *L, LPPassManager &LPM)
{
if (!Workgroup::isKernelToProcess(*L->getHeader()->getParent()))
return false;
return ProcessLoop(L, LPM);
}
/**
* Adds a barrier to the first BB of each loop.
*
* Note: it's not safe to do this in case the loop is not executed
* by all work items. Therefore this is not enabled by default.
*/
bool
ImplicitLoopBarriers::ProcessLoop(Loop *L, LPPassManager &LPM)
{
bool isBLoop = false;
for (Loop::block_iterator i = L->block_begin(), e = L->block_end();
i != e && !isBLoop; ++i) {
for (BasicBlock::iterator j = (*i)->begin(), e = (*i)->end();
j != e; ++j) {
if (isa<Barrier>(j)) {
isBLoop = true;
break;
}
}
}
if (isBLoop) return false;
return AddInnerLoopBarrier(L, LPM);
}
/**
* Adds a barrier to the beginning of the loop body to force its treatment
* similarly to a loop with work-group barriers.
*
* This allows parallelizing work-items across the work-group per kernel
* for-loop iteration, potentially leading to easier horizontal vectorization.
* The idea is similar to loop switching where the work-item loop is
* switched with the kernel for-loop.
*
* We need to make sure it is legal to add the barrier, though. The
* OpenCL barrier semantics require either all or none of the WIs to
* reach the barrier at each iteration. This is satisfied at least when
*
* a) loop exit condition does not depend on the WI and
* b) all or none of the WIs always enter the loop
*/
bool
ImplicitLoopBarriers::AddInnerLoopBarrier(llvm::Loop *L, llvm::LPPassManager &LPM) {
/* Only add barriers to the innermost loops. */
if (L->getSubLoops().size() > 0)
return false;
#ifdef DEBUG_ILOOP_BARRIERS
std::cerr << "### trying to add a loop barrier to force horizontal parallelization"
<< std::endl;
#endif
BasicBlock *brexit = L->getExitingBlock();
if (brexit == NULL) return false; /* Multiple exit points */
llvm::BasicBlock *loopEntry = L->getHeader();
if (loopEntry == NULL) return false; /* Multiple entries blocks? */
llvm::Function *f = brexit->getParent();
VariableUniformityAnalysis &VUA =
getAnalysis<VariableUniformityAnalysis>();
/* Check if the whole loop construct is executed by all or none of the
work-items. */
if (!VUA.isUniform(f, loopEntry)) {
#ifdef DEBUG_ILOOP_BARRIERS
std::cerr << "### the loop is not uniform because loop entry '"
<< loopEntry->getName().str() << "' is not uniform" << std::endl;
#endif
return false;
}
/* Check the branch condition predicate. If it is uniform, we know the loop
is executed the same number of times for all WIs. */
llvm::BranchInst *br = dyn_cast<llvm::BranchInst>(brexit->getTerminator());
if (br && br->isConditional() &&
VUA.isUniform(f, br->getCondition())) {
Barrier::Create(brexit->getTerminator());
#ifdef DEBUG_ILOOP_BARRIERS
std::cerr << "### added an inner-loop barrier to the loop" << std::endl << std::endl;
#endif
return true;
} else {
#ifdef DEBUG_ILOOP_BARRIERS
if (br && br->isConditional() && !VUA.isUniform(f, br->getCondition())) {
std::cerr << "### loop condition not uniform" << std::endl;
br->getCondition()->dump();
}
#endif
}
#ifdef DEBUG_ILOOP_BARRIERS
std::cerr << "### cannot add an inner-loop barrier to the loop" << std::endl << std::endl;
#endif
return false;
}
|
