forked from JuliaLang/julia
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathllvm-ptls.cpp
378 lines (346 loc) · 16.1 KB
/
llvm-ptls.cpp
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
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
// This file is a part of Julia. License is MIT: https://julialang.org/license
// LLVM pass to lower TLS access and remove references to julia intrinsics
#include "llvm-version.h"
#include "support/dtypes.h"
#include "passes.h"
#include <llvm-c/Core.h>
#include <llvm-c/Types.h>
#include <llvm/Pass.h>
#include <llvm/IR/Module.h>
#include <llvm/IR/LegacyPassManager.h>
#include <llvm/IR/Function.h>
#include <llvm/IR/Instructions.h>
#include <llvm/IR/Constants.h>
#include <llvm/IR/LLVMContext.h>
#include <llvm/IR/MDBuilder.h>
#include <llvm/IR/Verifier.h>
#include <llvm/IR/InlineAsm.h>
#include <llvm/Transforms/Utils/BasicBlockUtils.h>
#include "julia.h"
#include "julia_internal.h"
#include "llvm-codegen-shared.h"
#include "julia_assert.h"
#define DEBUG_TYPE "lower_ptls"
#undef DEBUG
using namespace llvm;
typedef Instruction TerminatorInst;
namespace {
struct LowerPTLS {
LowerPTLS(Module &M, bool imaging_mode=false)
: imaging_mode(imaging_mode), M(&M), TargetTriple(M.getTargetTriple())
{}
bool run(bool *CFGModified);
private:
const bool imaging_mode;
Module *M;
Triple TargetTriple;
MDNode *tbaa_const{nullptr};
MDNode *tbaa_gcframe{nullptr};
FunctionType *FT_pgcstack_getter{nullptr};
PointerType *T_pgcstack_getter{nullptr};
PointerType *T_pppjlvalue{nullptr};
Type *T_size{nullptr};
GlobalVariable *pgcstack_func_slot{nullptr};
GlobalVariable *pgcstack_key_slot{nullptr};
GlobalVariable *pgcstack_offset{nullptr};
void set_pgcstack_attrs(CallInst *pgcstack) const;
Instruction *emit_pgcstack_tp(Value *offset, Instruction *insertBefore) const;
template<typename T> T *add_comdat(T *G) const;
GlobalVariable *create_aliased_global(Type *T, StringRef name) const;
void fix_pgcstack_use(CallInst *pgcstack, Function *pgcstack_getter, bool or_new, bool *CFGModified);
};
void LowerPTLS::set_pgcstack_attrs(CallInst *pgcstack) const
{
addFnAttr(pgcstack, Attribute::ReadNone);
addFnAttr(pgcstack, Attribute::NoUnwind);
}
Instruction *LowerPTLS::emit_pgcstack_tp(Value *offset, Instruction *insertBefore) const
{
Value *tls;
if (TargetTriple.isX86() && insertBefore->getFunction()->callsFunctionThatReturnsTwice()) {
// Workaround LLVM bug by hiding the offset computation
// (and therefore the optimization opportunity) from LLVM.
// Ref https://github.com/JuliaLang/julia/issues/17288
std::string const_asm_str;
raw_string_ostream(const_asm_str) << (TargetTriple.getArch() == Triple::x86_64 ?
"movq %fs:0, $0;\naddq $$" : "movl %gs:0, $0;\naddl $$")
<< jl_tls_offset << ", $0";
const char *dyn_asm_str = TargetTriple.getArch() == Triple::x86_64 ?
"movq %fs:0, $0;\naddq $1, $0" :
"movl %gs:0, $0;\naddl $1, $0";
// The add instruction clobbers flags
if (offset) {
std::vector<Type*> args(0);
args.push_back(offset->getType());
auto tp = InlineAsm::get(FunctionType::get(Type::getInt8PtrTy(insertBefore->getContext()), args, false),
dyn_asm_str, "=&r,r,~{dirflag},~{fpsr},~{flags}", false);
tls = CallInst::Create(tp, offset, "pgcstack_i8", insertBefore);
}
else {
auto tp = InlineAsm::get(FunctionType::get(Type::getInt8PtrTy(insertBefore->getContext()), false),
const_asm_str.c_str(), "=r,~{dirflag},~{fpsr},~{flags}",
false);
tls = CallInst::Create(tp, "pgcstack_i8", insertBefore);
}
} else {
// AArch64/ARM doesn't seem to have this issue.
// (Possibly because there are many more registers and the offset is
// positive and small)
// It's also harder to emit the offset in a generic way on ARM/AArch64
// (need to generate one or two `add` with shift) so let llvm emit
// the add for now.
const char *asm_str;
if (TargetTriple.isAArch64()) {
asm_str = "mrs $0, tpidr_el0";
} else if (TargetTriple.isARM()) {
asm_str = "mrc p15, 0, $0, c13, c0, 3";
} else if (TargetTriple.getArch() == Triple::x86_64) {
asm_str = "movq %fs:0, $0";
} else if (TargetTriple.getArch() == Triple::x86) {
asm_str = "movl %gs:0, $0";
} else {
llvm_unreachable("Cannot emit thread pointer for this architecture.");
}
if (!offset)
offset = ConstantInt::getSigned(T_size, jl_tls_offset);
auto tp = InlineAsm::get(FunctionType::get(Type::getInt8PtrTy(insertBefore->getContext()), false), asm_str, "=r", false);
tls = CallInst::Create(tp, "thread_ptr", insertBefore);
tls = GetElementPtrInst::Create(Type::getInt8Ty(insertBefore->getContext()), tls, {offset}, "ppgcstack_i8", insertBefore);
}
tls = new BitCastInst(tls, T_pppjlvalue->getPointerTo(), "ppgcstack", insertBefore);
return new LoadInst(T_pppjlvalue, tls, "pgcstack", false, insertBefore);
}
GlobalVariable *LowerPTLS::create_aliased_global(Type *T, StringRef name) const
{
// Create a static global variable and points a global alias to it so that
// the address is visible externally but LLVM can still assume that the
// address of this variable doesn't need dynamic relocation
// (can be accessed with a single PC-rel load).
auto GV = new GlobalVariable(*M, T, false, GlobalVariable::ExternalLinkage,
nullptr, name);
return GV;
}
void LowerPTLS::fix_pgcstack_use(CallInst *pgcstack, Function *pgcstack_getter, bool or_new, bool *CFGModified)
{
if (pgcstack->use_empty()) {
pgcstack->eraseFromParent();
return;
}
if (or_new) {
// pgcstack();
// if (pgcstack != nullptr)
// last_gc_state = emit_gc_unsafe_enter(ctx);
// phi = pgcstack; // fast
// else
// last_gc_state = gc_safe;
// phi = adopt(); // slow
// use phi;
// if (!retboxed)
// foreach(retinst)
// emit_gc_unsafe_leave(ctx, last_gc_state);
auto phi = PHINode::Create(pgcstack->getType(), 2, "");
phi->insertAfter(pgcstack);
pgcstack->replaceAllUsesWith(phi);
MDBuilder MDB(pgcstack->getContext());
SmallVector<uint32_t, 2> Weights{9, 1};
TerminatorInst *fastTerm;
TerminatorInst *slowTerm;
auto cmp = new ICmpInst(phi, CmpInst::ICMP_NE, pgcstack, Constant::getNullValue(pgcstack->getType()));
SplitBlockAndInsertIfThenElse(cmp, phi, &fastTerm, &slowTerm,
MDB.createBranchWeights(Weights));
if (CFGModified)
*CFGModified = true;
// emit slow branch code
CallInst *adopt = cast<CallInst>(pgcstack->clone());
Function *adoptFunc = M->getFunction(XSTR(jl_adopt_thread));
if (adoptFunc == NULL) {
adoptFunc = Function::Create(pgcstack_getter->getFunctionType(),
pgcstack_getter->getLinkage(), pgcstack_getter->getAddressSpace(),
XSTR(jl_adopt_thread), M);
adoptFunc->copyAttributesFrom(pgcstack_getter);
adoptFunc->copyMetadata(pgcstack_getter, 0);
}
adopt->setCalledFunction(adoptFunc);
adopt->insertBefore(slowTerm);
phi->addIncoming(adopt, slowTerm->getParent());
// emit fast branch code
IRBuilder<> builder(fastTerm->getParent());
fastTerm->removeFromParent();
MDNode *tbaa = tbaa_gcframe;
Value *prior = emit_gc_unsafe_enter(builder, T_size, get_current_ptls_from_task(builder, T_size, get_current_task_from_pgcstack(builder, T_size, pgcstack), tbaa), true);
builder.Insert(fastTerm);
phi->addIncoming(pgcstack, fastTerm->getParent());
// emit pre-return cleanup
if (CountTrackedPointers(pgcstack->getParent()->getParent()->getReturnType()).count == 0) {
auto last_gc_state = PHINode::Create(Type::getInt8Ty(pgcstack->getContext()), 2, "", phi);
// if we called jl_adopt_thread, we must end this cfunction back in the safe-state
last_gc_state->addIncoming(ConstantInt::get(Type::getInt8Ty(M->getContext()), JL_GC_STATE_SAFE), slowTerm->getParent());
last_gc_state->addIncoming(prior, fastTerm->getParent());
for (auto &BB : *pgcstack->getParent()->getParent()) {
if (isa<ReturnInst>(BB.getTerminator())) {
IRBuilder<> builder(BB.getTerminator());
emit_gc_unsafe_leave(builder, T_size, get_current_ptls_from_task(builder, T_size, get_current_task_from_pgcstack(builder, T_size, phi), tbaa), last_gc_state, true);
}
}
}
}
if (imaging_mode) {
if (jl_tls_elf_support) {
// if (offset != 0)
// pgcstack = tp + offset; // fast
// else
// pgcstack = getter(); // slow
auto offset = new LoadInst(T_size, pgcstack_offset, "", false, pgcstack);
offset->setMetadata(llvm::LLVMContext::MD_tbaa, tbaa_const);
offset->setMetadata(llvm::LLVMContext::MD_invariant_load, MDNode::get(pgcstack->getContext(), None));
auto cmp = new ICmpInst(pgcstack, CmpInst::ICMP_NE, offset,
Constant::getNullValue(offset->getType()));
MDBuilder MDB(pgcstack->getContext());
SmallVector<uint32_t, 2> Weights{9, 1};
TerminatorInst *fastTerm;
TerminatorInst *slowTerm;
SplitBlockAndInsertIfThenElse(cmp, pgcstack, &fastTerm, &slowTerm,
MDB.createBranchWeights(Weights));
if (CFGModified)
*CFGModified = true;
auto fastTLS = emit_pgcstack_tp(offset, fastTerm);
auto phi = PHINode::Create(T_pppjlvalue, 2, "", pgcstack);
pgcstack->replaceAllUsesWith(phi);
pgcstack->moveBefore(slowTerm);
auto getter = new LoadInst(T_pgcstack_getter, pgcstack_func_slot, "", false, pgcstack);
getter->setMetadata(llvm::LLVMContext::MD_tbaa, tbaa_const);
getter->setMetadata(llvm::LLVMContext::MD_invariant_load, MDNode::get(pgcstack->getContext(), None));
pgcstack->setCalledFunction(pgcstack->getFunctionType(), getter);
set_pgcstack_attrs(pgcstack);
phi->addIncoming(fastTLS, fastTLS->getParent());
phi->addIncoming(pgcstack, pgcstack->getParent());
return;
}
// In imaging mode, we emit the function address as a load of a static
// variable to be filled (in `staticdata.c`) at initialization time of the sysimg.
// This way we can bypass the extra indirection in `jl_get_pgcstack`
// since we may not know which getter function to use ahead of time.
auto getter = new LoadInst(T_pgcstack_getter, pgcstack_func_slot, "", false, pgcstack);
getter->setMetadata(llvm::LLVMContext::MD_tbaa, tbaa_const);
getter->setMetadata(llvm::LLVMContext::MD_invariant_load, MDNode::get(pgcstack->getContext(), None));
if (TargetTriple.isOSDarwin()) {
auto key = new LoadInst(T_size, pgcstack_key_slot, "", false, pgcstack);
key->setMetadata(llvm::LLVMContext::MD_tbaa, tbaa_const);
key->setMetadata(llvm::LLVMContext::MD_invariant_load, MDNode::get(pgcstack->getContext(), None));
auto new_pgcstack = CallInst::Create(FT_pgcstack_getter, getter, {key}, "", pgcstack);
new_pgcstack->takeName(pgcstack);
pgcstack->replaceAllUsesWith(new_pgcstack);
pgcstack->eraseFromParent();
pgcstack = new_pgcstack;
} else {
pgcstack->setCalledFunction(pgcstack->getFunctionType(), getter);
}
set_pgcstack_attrs(pgcstack);
}
else if (jl_tls_offset != -1) {
pgcstack->replaceAllUsesWith(emit_pgcstack_tp(nullptr, pgcstack));
pgcstack->eraseFromParent();
}
else {
// use the address of the actual getter function directly
jl_get_pgcstack_func *f;
jl_pgcstack_key_t k;
jl_pgcstack_getkey(&f, &k);
Constant *val = ConstantInt::get(T_size, (uintptr_t)f);
val = ConstantExpr::getIntToPtr(val, T_pgcstack_getter);
if (TargetTriple.isOSDarwin()) {
assert(sizeof(k) == sizeof(uintptr_t));
Constant *key = ConstantInt::get(T_size, (uintptr_t)k);
auto new_pgcstack = CallInst::Create(FT_pgcstack_getter, val, {key}, "", pgcstack);
new_pgcstack->takeName(pgcstack);
pgcstack->replaceAllUsesWith(new_pgcstack);
pgcstack->eraseFromParent();
pgcstack = new_pgcstack;
} else {
pgcstack->setCalledFunction(pgcstack->getFunctionType(), val);
}
set_pgcstack_attrs(pgcstack);
}
}
bool LowerPTLS::run(bool *CFGModified)
{
bool need_init = true;
auto runOnGetter = [&](bool or_new) {
Function *pgcstack_getter = M->getFunction(or_new ? "julia.get_pgcstack_or_new" : "julia.get_pgcstack");
if (!pgcstack_getter)
return false;
if (need_init) {
tbaa_const = tbaa_make_child_with_context(M->getContext(), "jtbaa_const", nullptr, true).first;
tbaa_gcframe = tbaa_make_child_with_context(M->getContext(), "jtbaa_gcframe").first;
T_size = M->getDataLayout().getIntPtrType(M->getContext());
FT_pgcstack_getter = pgcstack_getter->getFunctionType();
if (TargetTriple.isOSDarwin()) {
assert(sizeof(jl_pgcstack_key_t) == sizeof(uintptr_t));
FT_pgcstack_getter = FunctionType::get(FT_pgcstack_getter->getReturnType(), {T_size}, false);
}
T_pgcstack_getter = FT_pgcstack_getter->getPointerTo();
T_pppjlvalue = cast<PointerType>(FT_pgcstack_getter->getReturnType());
if (imaging_mode) {
pgcstack_func_slot = create_aliased_global(T_pgcstack_getter, "jl_pgcstack_func_slot");
pgcstack_key_slot = create_aliased_global(T_size, "jl_pgcstack_key_slot"); // >= sizeof(jl_pgcstack_key_t)
pgcstack_offset = create_aliased_global(T_size, "jl_tls_offset");
}
need_init = false;
}
for (auto it = pgcstack_getter->user_begin(); it != pgcstack_getter->user_end();) {
auto call = cast<CallInst>(*it);
++it;
assert(call->getCalledOperand() == pgcstack_getter);
fix_pgcstack_use(call, pgcstack_getter, or_new, CFGModified);
}
assert(pgcstack_getter->use_empty());
pgcstack_getter->eraseFromParent();
return true;
};
return runOnGetter(false) + runOnGetter(true);
}
struct LowerPTLSLegacy: public ModulePass {
static char ID;
LowerPTLSLegacy(bool imaging_mode=false)
: ModulePass(ID),
imaging_mode(imaging_mode)
{}
bool imaging_mode;
bool runOnModule(Module &M) override {
LowerPTLS lower(M, imaging_mode);
bool modified = lower.run(nullptr);
#ifdef JL_VERIFY_PASSES
assert(!verifyModule(M, &errs()));
#endif
return modified;
}
};
char LowerPTLSLegacy::ID = 0;
static RegisterPass<LowerPTLSLegacy> X("LowerPTLS", "LowerPTLS Pass",
false /* Only looks at CFG */,
false /* Analysis Pass */);
} // anonymous namespace
PreservedAnalyses LowerPTLSPass::run(Module &M, ModuleAnalysisManager &AM) {
LowerPTLS lower(M, imaging_mode);
bool CFGModified = false;
bool modified = lower.run(&CFGModified);
#ifdef JL_VERIFY_PASSES
assert(!verifyModule(M, &errs()));
#endif
if (modified) {
if (CFGModified) {
return PreservedAnalyses::none();
} else {
return PreservedAnalyses::allInSet<CFGAnalyses>();
}
}
return PreservedAnalyses::all();
}
Pass *createLowerPTLSPass(bool imaging_mode)
{
return new LowerPTLSLegacy(imaging_mode);
}
extern "C" JL_DLLEXPORT void LLVMExtraAddLowerPTLSPass_impl(LLVMPassManagerRef PM, LLVMBool imaging_mode)
{
unwrap(PM)->add(createLowerPTLSPass(imaging_mode));
}