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
|
//===-- ARMGlobalMerge.cpp - Internal globals merging --------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
// This pass merges globals with internal linkage into one. This way all the
// globals which were merged into a biggest one can be addressed using offsets
// from the same base pointer (no need for separate base pointer for each of the
// global). Such a transformation can significantly reduce the register pressure
// when many globals are involved.
//
// For example, consider the code which touches several global variables at
// once:
//
// static int foo[N], bar[N], baz[N];
//
// for (i = 0; i < N; ++i) {
// foo[i] = bar[i] * baz[i];
// }
//
// On ARM the addresses of 3 arrays should be kept in the registers, thus
// this code has quite large register pressure (loop body):
//
// ldr r1, [r5], #4
// ldr r2, [r6], #4
// mul r1, r2, r1
// str r1, [r0], #4
//
// Pass converts the code to something like:
//
// static struct {
// int foo[N];
// int bar[N];
// int baz[N];
// } merged;
//
// for (i = 0; i < N; ++i) {
// merged.foo[i] = merged.bar[i] * merged.baz[i];
// }
//
// and in ARM code this becomes:
//
// ldr r0, [r5, #40]
// ldr r1, [r5, #80]
// mul r0, r1, r0
// str r0, [r5], #4
//
// note that we saved 2 registers here almostly "for free".
// ===---------------------------------------------------------------------===//
#define DEBUG_TYPE "arm-global-merge"
#include "ARM.h"
#include "ARMTargetMachine.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/Attributes.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
#include "llvm/GlobalVariable.h"
#include "llvm/Instructions.h"
#include "llvm/Intrinsics.h"
#include "llvm/Module.h"
#include "llvm/Pass.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetLoweringObjectFile.h"
using namespace llvm;
namespace {
class ARMGlobalMerge : public FunctionPass {
/// TLI - Keep a pointer of a TargetLowering to consult for determining
/// target type sizes.
const TargetLowering *TLI;
bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
Module &M, bool isConst) const;
public:
static char ID; // Pass identification, replacement for typeid.
explicit ARMGlobalMerge(const TargetLowering *tli)
: FunctionPass(ID), TLI(tli) {}
virtual bool doInitialization(Module &M);
virtual bool runOnFunction(Function &F);
const char *getPassName() const {
return "Merge internal globals";
}
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesCFG();
FunctionPass::getAnalysisUsage(AU);
}
struct GlobalCmp {
const TargetData *TD;
GlobalCmp(const TargetData *td) : TD(td) { }
bool operator()(const GlobalVariable *GV1, const GlobalVariable *GV2) {
const Type *Ty1 = cast<PointerType>(GV1->getType())->getElementType();
const Type *Ty2 = cast<PointerType>(GV2->getType())->getElementType();
return (TD->getTypeAllocSize(Ty1) < TD->getTypeAllocSize(Ty2));
}
};
};
} // end anonymous namespace
char ARMGlobalMerge::ID = 0;
bool ARMGlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
Module &M, bool isConst) const {
const TargetData *TD = TLI->getTargetData();
// FIXME: Infer the maximum possible offset depending on the actual users
// (these max offsets are different for the users inside Thumb or ARM
// functions)
unsigned MaxOffset = TLI->getMaximalGlobalOffset();
// FIXME: Find better heuristics
std::stable_sort(Globals.begin(), Globals.end(), GlobalCmp(TD));
const Type *Int32Ty = Type::getInt32Ty(M.getContext());
for (size_t i = 0, e = Globals.size(); i != e; ) {
size_t j = 0;
uint64_t MergedSize = 0;
std::vector<const Type*> Tys;
std::vector<Constant*> Inits;
for (j = i; j != e; ++j) {
const Type *Ty = Globals[j]->getType()->getElementType();
MergedSize += TD->getTypeAllocSize(Ty);
if (MergedSize > MaxOffset) {
break;
}
Tys.push_back(Ty);
Inits.push_back(Globals[j]->getInitializer());
}
StructType *MergedTy = StructType::get(M.getContext(), Tys);
Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
GlobalVariable *MergedGV = new GlobalVariable(M, MergedTy, isConst,
GlobalValue::InternalLinkage,
MergedInit, "_MergedGlobals");
for (size_t k = i; k < j; ++k) {
Constant *Idx[2] = {
ConstantInt::get(Int32Ty, 0),
ConstantInt::get(Int32Ty, k-i)
};
Constant *GEP = ConstantExpr::getInBoundsGetElementPtr(MergedGV, Idx, 2);
Globals[k]->replaceAllUsesWith(GEP);
Globals[k]->eraseFromParent();
}
i = j;
}
return true;
}
bool ARMGlobalMerge::doInitialization(Module &M) {
SmallVector<GlobalVariable*, 16> Globals, ConstGlobals, BSSGlobals;
const TargetData *TD = TLI->getTargetData();
unsigned MaxOffset = TLI->getMaximalGlobalOffset();
bool Changed = false;
// Disable this pass on darwin. The debugger is not yet ready to extract
// variable's info from a merged global.
if (TLI->getTargetMachine().getSubtarget<ARMSubtarget>().isTargetDarwin())
return false;
// Grab all non-const globals.
for (Module::global_iterator I = M.global_begin(),
E = M.global_end(); I != E; ++I) {
// Merge is safe for "normal" internal globals only
if (!I->hasLocalLinkage() || I->isThreadLocal() || I->hasSection())
continue;
// Ignore fancy-aligned globals for now.
if (I->getAlignment() != 0)
continue;
// Ignore all 'special' globals.
if (I->getName().startswith("llvm.") ||
I->getName().startswith(".llvm."))
continue;
if (TD->getTypeAllocSize(I->getType()->getElementType()) < MaxOffset) {
const TargetLoweringObjectFile &TLOF = TLI->getObjFileLowering();
if (TLOF.getKindForGlobal(I, TLI->getTargetMachine()).isBSSLocal())
BSSGlobals.push_back(I);
else if (I->isConstant())
ConstGlobals.push_back(I);
else
Globals.push_back(I);
}
}
if (Globals.size() > 1)
Changed |= doMerge(Globals, M, false);
if (BSSGlobals.size() > 1)
Changed |= doMerge(BSSGlobals, M, false);
// FIXME: This currently breaks the EH processing due to way how the
// typeinfo detection works. We might want to detect the TIs and ignore
// them in the future.
// if (ConstGlobals.size() > 1)
// Changed |= doMerge(ConstGlobals, M, true);
return Changed;
}
bool ARMGlobalMerge::runOnFunction(Function &F) {
return false;
}
FunctionPass *llvm::createARMGlobalMergePass(const TargetLowering *tli) {
return new ARMGlobalMerge(tli);
}
|
