1 | //===------ PollyIRBuilder.cpp --------------------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // The Polly IRBuilder file contains Polly specific extensions for the IRBuilder |
10 | // that are used e.g. to emit the llvm.loop.parallel metadata. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "polly/CodeGen/IRBuilder.h" |
15 | #include "polly/ScopInfo.h" |
16 | #include "polly/Support/ScopHelper.h" |
17 | #include "llvm/ADT/SmallVector.h" |
18 | #include "llvm/IR/Metadata.h" |
19 | |
20 | using namespace llvm; |
21 | using namespace polly; |
22 | |
23 | static const int MaxArraysInAliasScops = 10; |
24 | |
25 | /// Get a self referencing id metadata node. |
26 | /// |
27 | /// The MDNode looks like this (if arg0/arg1 are not null): |
28 | /// |
29 | /// '!n = distinct !{!n, arg0, arg1}' |
30 | /// |
31 | /// @return The self referencing id metadata node. |
32 | static MDNode *getID(LLVMContext &Ctx, Metadata *arg0 = nullptr, |
33 | Metadata *arg1 = nullptr) { |
34 | MDNode *ID; |
35 | SmallVector<Metadata *, 3> Args; |
36 | // Reserve operand 0 for loop id self reference. |
37 | Args.push_back(Elt: nullptr); |
38 | |
39 | if (arg0) |
40 | Args.push_back(Elt: arg0); |
41 | if (arg1) |
42 | Args.push_back(Elt: arg1); |
43 | |
44 | ID = MDNode::getDistinct(Context&: Ctx, MDs: Args); |
45 | ID->replaceOperandWith(I: 0, New: ID); |
46 | return ID; |
47 | } |
48 | |
49 | ScopAnnotator::ScopAnnotator() : SE(nullptr), AliasScopeDomain(nullptr) { |
50 | // Push an empty staging BandAttr. |
51 | LoopAttrEnv.emplace_back(); |
52 | } |
53 | |
54 | ScopAnnotator::~ScopAnnotator() { |
55 | assert(LoopAttrEnv.size() == 1 && "Loop stack imbalance" ); |
56 | assert(!getStagingAttrEnv() && "Forgot to clear staging attr env" ); |
57 | } |
58 | |
59 | void ScopAnnotator::buildAliasScopes(Scop &S) { |
60 | SE = S.getSE(); |
61 | |
62 | LLVMContext &Ctx = SE->getContext(); |
63 | AliasScopeDomain = getID(Ctx, arg0: MDString::get(Context&: Ctx, Str: "polly.alias.scope.domain" )); |
64 | |
65 | AliasScopeMap.clear(); |
66 | OtherAliasScopeListMap.clear(); |
67 | |
68 | // We are only interested in arrays, but no scalar references. Scalars should |
69 | // be handled easily by basicaa. |
70 | SmallVector<ScopArrayInfo *, 10> Arrays; |
71 | for (ScopArrayInfo *Array : S.arrays()) |
72 | if (Array->isArrayKind()) |
73 | Arrays.push_back(Elt: Array); |
74 | |
75 | // The construction of alias scopes is quadratic in the number of arrays |
76 | // involved. In case of too many arrays, skip the construction of alias |
77 | // information to avoid quadratic increases in compile time and code size. |
78 | if (Arrays.size() > MaxArraysInAliasScops) |
79 | return; |
80 | |
81 | std::string AliasScopeStr = "polly.alias.scope." ; |
82 | for (const ScopArrayInfo *Array : Arrays) { |
83 | assert(Array->getBasePtr() && "Base pointer must be present" ); |
84 | AliasScopeMap[Array->getBasePtr()] = |
85 | getID(Ctx, arg0: AliasScopeDomain, |
86 | arg1: MDString::get(Context&: Ctx, Str: (AliasScopeStr + Array->getName()).c_str())); |
87 | } |
88 | |
89 | for (const ScopArrayInfo *Array : Arrays) { |
90 | MDNode *AliasScopeList = MDNode::get(Context&: Ctx, MDs: {}); |
91 | for (const auto &AliasScopePair : AliasScopeMap) { |
92 | if (Array->getBasePtr() == AliasScopePair.first) |
93 | continue; |
94 | |
95 | Metadata *Args = {AliasScopePair.second}; |
96 | AliasScopeList = |
97 | MDNode::concatenate(A: AliasScopeList, B: MDNode::get(Context&: Ctx, MDs: Args)); |
98 | } |
99 | |
100 | OtherAliasScopeListMap[Array->getBasePtr()] = AliasScopeList; |
101 | } |
102 | } |
103 | |
104 | void ScopAnnotator::pushLoop(Loop *L, bool IsParallel) { |
105 | ActiveLoops.push_back(Elt: L); |
106 | |
107 | if (IsParallel) { |
108 | LLVMContext &Ctx = SE->getContext(); |
109 | MDNode *AccessGroup = MDNode::getDistinct(Context&: Ctx, MDs: {}); |
110 | ParallelLoops.push_back(Elt: AccessGroup); |
111 | } |
112 | |
113 | // Open an empty BandAttr context for loops nested in this one. |
114 | LoopAttrEnv.emplace_back(); |
115 | } |
116 | |
117 | void ScopAnnotator::popLoop(bool IsParallel) { |
118 | ActiveLoops.pop_back(); |
119 | |
120 | if (IsParallel) { |
121 | assert(!ParallelLoops.empty() && "Expected a parallel loop to pop" ); |
122 | ParallelLoops.pop_back(); |
123 | } |
124 | |
125 | // Exit the subloop context. |
126 | assert(!getStagingAttrEnv() && "Forgot to clear staging attr env" ); |
127 | assert(LoopAttrEnv.size() >= 2 && "Popped too many" ); |
128 | LoopAttrEnv.pop_back(); |
129 | } |
130 | |
131 | void ScopAnnotator::annotateLoopLatch(BranchInst *B, Loop *L, bool IsParallel, |
132 | bool IsLoopVectorizerDisabled) const { |
133 | LLVMContext &Ctx = SE->getContext(); |
134 | SmallVector<Metadata *, 3> Args; |
135 | |
136 | // For the LoopID self-reference. |
137 | Args.push_back(Elt: nullptr); |
138 | |
139 | // Add the user-defined loop properties to the annotation, if any. Any |
140 | // additional properties are appended. |
141 | // FIXME: What to do if these conflict? |
142 | MDNode *MData = nullptr; |
143 | if (BandAttr *AttrEnv = getActiveAttrEnv()) { |
144 | MData = AttrEnv->Metadata; |
145 | if (MData) |
146 | llvm::append_range(C&: Args, R: drop_begin(RangeOrContainer: MData->operands(), N: 1)); |
147 | } |
148 | |
149 | if (IsLoopVectorizerDisabled) { |
150 | MDString *PropName = MDString::get(Context&: Ctx, Str: "llvm.loop.vectorize.enable" ); |
151 | ConstantInt *FalseValue = ConstantInt::get(Ty: Type::getInt1Ty(C&: Ctx), V: 0); |
152 | ValueAsMetadata *PropValue = ValueAsMetadata::get(V: FalseValue); |
153 | Args.push_back(Elt: MDNode::get(Context&: Ctx, MDs: {PropName, PropValue})); |
154 | } |
155 | |
156 | if (IsParallel) { |
157 | MDString *PropName = MDString::get(Context&: Ctx, Str: "llvm.loop.parallel_accesses" ); |
158 | MDNode *AccGroup = ParallelLoops.back(); |
159 | Args.push_back(Elt: MDNode::get(Context&: Ctx, MDs: {PropName, AccGroup})); |
160 | } |
161 | |
162 | // No metadata to annotate. |
163 | if (!MData && Args.size() <= 1) |
164 | return; |
165 | |
166 | // Reuse the MData node if possible, this will avoid having to create another |
167 | // one that cannot be merged because LoopIDs are 'distinct'. However, we have |
168 | // to create a new one if we add properties. |
169 | if (!MData || Args.size() > MData->getNumOperands()) { |
170 | MData = MDNode::getDistinct(Context&: Ctx, MDs: Args); |
171 | MData->replaceOperandWith(I: 0, New: MData); |
172 | } |
173 | B->setMetadata(KindID: LLVMContext::MD_loop, Node: MData); |
174 | } |
175 | |
176 | /// Get the pointer operand |
177 | /// |
178 | /// @param Inst The instruction to be analyzed. |
179 | /// @return the pointer operand in case @p Inst is a memory access |
180 | /// instruction and nullptr otherwise. |
181 | static llvm::Value *getMemAccInstPointerOperand(Instruction *Inst) { |
182 | auto MemInst = MemAccInst::dyn_cast(V: Inst); |
183 | if (!MemInst) |
184 | return nullptr; |
185 | |
186 | return MemInst.getPointerOperand(); |
187 | } |
188 | |
189 | /// Find the base pointer of an array access. |
190 | /// |
191 | /// This should be equivalent to ScalarEvolution::getPointerBase, which we |
192 | /// cannot use here the IR is still under construction which ScalarEvolution |
193 | /// assumes to not be modified. |
194 | static Value *findBasePtr(Value *Val) { |
195 | while (true) { |
196 | if (auto *Gep = dyn_cast<GEPOperator>(Val)) { |
197 | Val = Gep->getPointerOperand(); |
198 | continue; |
199 | } |
200 | if (auto *Cast = dyn_cast<BitCastOperator>(Val)) { |
201 | Val = Cast->getOperand(i_nocapture: 0); |
202 | continue; |
203 | } |
204 | |
205 | break; |
206 | } |
207 | |
208 | return Val; |
209 | } |
210 | |
211 | void ScopAnnotator::annotate(Instruction *Inst) { |
212 | if (!Inst->mayReadOrWriteMemory()) |
213 | return; |
214 | |
215 | switch (ParallelLoops.size()) { |
216 | case 0: |
217 | // Not parallel to anything: no access group needed. |
218 | break; |
219 | case 1: |
220 | // Single parallel loop: use directly. |
221 | Inst->setMetadata(KindID: LLVMContext::MD_access_group, |
222 | Node: cast<MDNode>(Val: ParallelLoops.front())); |
223 | break; |
224 | default: |
225 | // Parallel to multiple loops: refer to list of access groups. |
226 | Inst->setMetadata(KindID: LLVMContext::MD_access_group, |
227 | Node: MDNode::get(Context&: SE->getContext(), |
228 | MDs: ArrayRef<Metadata *>( |
229 | (Metadata *const *)ParallelLoops.data(), |
230 | ParallelLoops.size()))); |
231 | break; |
232 | } |
233 | |
234 | // TODO: Use the ScopArrayInfo once available here. |
235 | if (!AliasScopeDomain) |
236 | return; |
237 | |
238 | // Do not apply annotations on memory operations that take more than one |
239 | // pointer. It would be ambiguous to which pointer the annotation applies. |
240 | // FIXME: How can we specify annotations for all pointer arguments? |
241 | if (isa<CallInst>(Val: Inst) && !isa<MemSetInst>(Val: Inst)) |
242 | return; |
243 | |
244 | auto *Ptr = getMemAccInstPointerOperand(Inst); |
245 | if (!Ptr) |
246 | return; |
247 | |
248 | Value *BasePtr = findBasePtr(Val: Ptr); |
249 | if (!BasePtr) |
250 | return; |
251 | |
252 | auto AliasScope = AliasScopeMap.lookup(Key: BasePtr); |
253 | |
254 | if (!AliasScope) { |
255 | BasePtr = AlternativeAliasBases.lookup(Val: BasePtr); |
256 | if (!BasePtr) |
257 | return; |
258 | |
259 | AliasScope = AliasScopeMap.lookup(Key: BasePtr); |
260 | if (!AliasScope) |
261 | return; |
262 | } |
263 | |
264 | assert(OtherAliasScopeListMap.count(BasePtr) && |
265 | "BasePtr either expected in AliasScopeMap and OtherAlias...Map" ); |
266 | auto *OtherAliasScopeList = OtherAliasScopeListMap[BasePtr]; |
267 | |
268 | Inst->setMetadata(Kind: "alias.scope" , Node: MDNode::get(Context&: SE->getContext(), MDs: AliasScope)); |
269 | Inst->setMetadata(Kind: "noalias" , Node: OtherAliasScopeList); |
270 | } |
271 | |