File: | polly/lib/CodeGen/IslNodeBuilder.cpp |
Warning: | line 1173, column 10 Value stored to 'size' during its initialization is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- IslNodeBuilder.cpp - Translate an isl AST into a LLVM-IR AST -------===// |
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 | // This file contains the IslNodeBuilder, a class to translate an isl AST into |
10 | // a LLVM-IR AST. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "polly/CodeGen/IslNodeBuilder.h" |
15 | #include "polly/CodeGen/BlockGenerators.h" |
16 | #include "polly/CodeGen/CodeGeneration.h" |
17 | #include "polly/CodeGen/IslAst.h" |
18 | #include "polly/CodeGen/IslExprBuilder.h" |
19 | #include "polly/CodeGen/LoopGeneratorsGOMP.h" |
20 | #include "polly/CodeGen/LoopGeneratorsKMP.h" |
21 | #include "polly/CodeGen/RuntimeDebugBuilder.h" |
22 | #include "polly/Options.h" |
23 | #include "polly/ScopInfo.h" |
24 | #include "polly/Support/ISLTools.h" |
25 | #include "polly/Support/SCEVValidator.h" |
26 | #include "polly/Support/ScopHelper.h" |
27 | #include "llvm/ADT/APInt.h" |
28 | #include "llvm/ADT/PostOrderIterator.h" |
29 | #include "llvm/ADT/SetVector.h" |
30 | #include "llvm/ADT/SmallPtrSet.h" |
31 | #include "llvm/ADT/Statistic.h" |
32 | #include "llvm/Analysis/LoopInfo.h" |
33 | #include "llvm/Analysis/RegionInfo.h" |
34 | #include "llvm/Analysis/ScalarEvolution.h" |
35 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
36 | #include "llvm/IR/BasicBlock.h" |
37 | #include "llvm/IR/Constant.h" |
38 | #include "llvm/IR/Constants.h" |
39 | #include "llvm/IR/DataLayout.h" |
40 | #include "llvm/IR/DerivedTypes.h" |
41 | #include "llvm/IR/Dominators.h" |
42 | #include "llvm/IR/Function.h" |
43 | #include "llvm/IR/InstrTypes.h" |
44 | #include "llvm/IR/Instruction.h" |
45 | #include "llvm/IR/Instructions.h" |
46 | #include "llvm/IR/Type.h" |
47 | #include "llvm/IR/Value.h" |
48 | #include "llvm/Support/Casting.h" |
49 | #include "llvm/Support/CommandLine.h" |
50 | #include "llvm/Support/ErrorHandling.h" |
51 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
52 | #include "isl/aff.h" |
53 | #include "isl/aff_type.h" |
54 | #include "isl/ast.h" |
55 | #include "isl/ast_build.h" |
56 | #include "isl/isl-noexceptions.h" |
57 | #include "isl/map.h" |
58 | #include "isl/set.h" |
59 | #include "isl/union_map.h" |
60 | #include "isl/union_set.h" |
61 | #include "isl/val.h" |
62 | #include <algorithm> |
63 | #include <cassert> |
64 | #include <cstdint> |
65 | #include <cstring> |
66 | #include <string> |
67 | #include <utility> |
68 | #include <vector> |
69 | |
70 | using namespace llvm; |
71 | using namespace polly; |
72 | |
73 | #define DEBUG_TYPE"polly-codegen" "polly-codegen" |
74 | |
75 | STATISTIC(VersionedScops, "Number of SCoPs that required versioning.")static llvm::Statistic VersionedScops = {"polly-codegen", "VersionedScops" , "Number of SCoPs that required versioning."}; |
76 | |
77 | STATISTIC(SequentialLoops, "Number of generated sequential for-loops")static llvm::Statistic SequentialLoops = {"polly-codegen", "SequentialLoops" , "Number of generated sequential for-loops"}; |
78 | STATISTIC(ParallelLoops, "Number of generated parallel for-loops")static llvm::Statistic ParallelLoops = {"polly-codegen", "ParallelLoops" , "Number of generated parallel for-loops"}; |
79 | STATISTIC(VectorLoops, "Number of generated vector for-loops")static llvm::Statistic VectorLoops = {"polly-codegen", "VectorLoops" , "Number of generated vector for-loops"}; |
80 | STATISTIC(IfConditions, "Number of generated if-conditions")static llvm::Statistic IfConditions = {"polly-codegen", "IfConditions" , "Number of generated if-conditions"}; |
81 | |
82 | /// OpenMP backend options |
83 | enum class OpenMPBackend { GNU, LLVM }; |
84 | |
85 | static cl::opt<bool> PollyGenerateRTCPrint( |
86 | "polly-codegen-emit-rtc-print", |
87 | cl::desc("Emit code that prints the runtime check result dynamically."), |
88 | cl::Hidden, cl::init(false), cl::ZeroOrMore, cl::cat(PollyCategory)); |
89 | |
90 | // If this option is set we always use the isl AST generator to regenerate |
91 | // memory accesses. Without this option set we regenerate expressions using the |
92 | // original SCEV expressions and only generate new expressions in case the |
93 | // access relation has been changed and consequently must be regenerated. |
94 | static cl::opt<bool> PollyGenerateExpressions( |
95 | "polly-codegen-generate-expressions", |
96 | cl::desc("Generate AST expressions for unmodified and modified accesses"), |
97 | cl::Hidden, cl::init(false), cl::ZeroOrMore, cl::cat(PollyCategory)); |
98 | |
99 | static cl::opt<int> PollyTargetFirstLevelCacheLineSize( |
100 | "polly-target-first-level-cache-line-size", |
101 | cl::desc("The size of the first level cache line size specified in bytes."), |
102 | cl::Hidden, cl::init(64), cl::ZeroOrMore, cl::cat(PollyCategory)); |
103 | |
104 | static cl::opt<OpenMPBackend> PollyOmpBackend( |
105 | "polly-omp-backend", cl::desc("Choose the OpenMP library to use:"), |
106 | cl::values(clEnumValN(OpenMPBackend::GNU, "GNU", "GNU OpenMP")llvm::cl::OptionEnumValue { "GNU", int(OpenMPBackend::GNU), "GNU OpenMP" }, |
107 | clEnumValN(OpenMPBackend::LLVM, "LLVM", "LLVM OpenMP")llvm::cl::OptionEnumValue { "LLVM", int(OpenMPBackend::LLVM), "LLVM OpenMP" }), |
108 | cl::Hidden, cl::init(OpenMPBackend::GNU), cl::cat(PollyCategory)); |
109 | |
110 | isl::ast_expr IslNodeBuilder::getUpperBound(isl::ast_node For, |
111 | ICmpInst::Predicate &Predicate) { |
112 | isl::ast_expr Cond = For.for_get_cond(); |
113 | isl::ast_expr Iterator = For.for_get_iterator(); |
114 | assert(isl_ast_expr_get_type(Cond.get()) == isl_ast_expr_op &&((isl_ast_expr_get_type(Cond.get()) == isl_ast_expr_op && "conditional expression is not an atomic upper bound") ? static_cast <void> (0) : __assert_fail ("isl_ast_expr_get_type(Cond.get()) == isl_ast_expr_op && \"conditional expression is not an atomic upper bound\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 115, __PRETTY_FUNCTION__)) |
115 | "conditional expression is not an atomic upper bound")((isl_ast_expr_get_type(Cond.get()) == isl_ast_expr_op && "conditional expression is not an atomic upper bound") ? static_cast <void> (0) : __assert_fail ("isl_ast_expr_get_type(Cond.get()) == isl_ast_expr_op && \"conditional expression is not an atomic upper bound\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 115, __PRETTY_FUNCTION__)); |
116 | |
117 | isl_ast_op_typeisl_ast_expr_op_type OpType = isl_ast_expr_get_op_type(Cond.get()); |
118 | |
119 | switch (OpType) { |
120 | case isl_ast_op_leisl_ast_expr_op_le: |
121 | Predicate = ICmpInst::ICMP_SLE; |
122 | break; |
123 | case isl_ast_op_ltisl_ast_expr_op_lt: |
124 | Predicate = ICmpInst::ICMP_SLT; |
125 | break; |
126 | default: |
127 | llvm_unreachable("Unexpected comparison type in loop condition")::llvm::llvm_unreachable_internal("Unexpected comparison type in loop condition" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 127); |
128 | } |
129 | |
130 | isl::ast_expr Arg0 = Cond.get_op_arg(0); |
131 | |
132 | assert(isl_ast_expr_get_type(Arg0.get()) == isl_ast_expr_id &&((isl_ast_expr_get_type(Arg0.get()) == isl_ast_expr_id && "conditional expression is not an atomic upper bound") ? static_cast <void> (0) : __assert_fail ("isl_ast_expr_get_type(Arg0.get()) == isl_ast_expr_id && \"conditional expression is not an atomic upper bound\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 133, __PRETTY_FUNCTION__)) |
133 | "conditional expression is not an atomic upper bound")((isl_ast_expr_get_type(Arg0.get()) == isl_ast_expr_id && "conditional expression is not an atomic upper bound") ? static_cast <void> (0) : __assert_fail ("isl_ast_expr_get_type(Arg0.get()) == isl_ast_expr_id && \"conditional expression is not an atomic upper bound\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 133, __PRETTY_FUNCTION__)); |
134 | |
135 | isl::id UBID = Arg0.get_id(); |
136 | |
137 | assert(isl_ast_expr_get_type(Iterator.get()) == isl_ast_expr_id &&((isl_ast_expr_get_type(Iterator.get()) == isl_ast_expr_id && "Could not get the iterator") ? static_cast<void> (0) : __assert_fail ("isl_ast_expr_get_type(Iterator.get()) == isl_ast_expr_id && \"Could not get the iterator\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 138, __PRETTY_FUNCTION__)) |
138 | "Could not get the iterator")((isl_ast_expr_get_type(Iterator.get()) == isl_ast_expr_id && "Could not get the iterator") ? static_cast<void> (0) : __assert_fail ("isl_ast_expr_get_type(Iterator.get()) == isl_ast_expr_id && \"Could not get the iterator\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 138, __PRETTY_FUNCTION__)); |
139 | |
140 | isl::id IteratorID = Iterator.get_id(); |
141 | |
142 | assert(UBID.get() == IteratorID.get() &&((UBID.get() == IteratorID.get() && "conditional expression is not an atomic upper bound" ) ? static_cast<void> (0) : __assert_fail ("UBID.get() == IteratorID.get() && \"conditional expression is not an atomic upper bound\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 143, __PRETTY_FUNCTION__)) |
143 | "conditional expression is not an atomic upper bound")((UBID.get() == IteratorID.get() && "conditional expression is not an atomic upper bound" ) ? static_cast<void> (0) : __assert_fail ("UBID.get() == IteratorID.get() && \"conditional expression is not an atomic upper bound\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 143, __PRETTY_FUNCTION__)); |
144 | |
145 | return Cond.get_op_arg(1); |
146 | } |
147 | |
148 | /// Return true if a return value of Predicate is true for the value represented |
149 | /// by passed isl_ast_expr_int. |
150 | static bool checkIslAstExprInt(__isl_take isl_ast_expr *Expr, |
151 | isl_bool (*Predicate)(__isl_keep isl_val *)) { |
152 | if (isl_ast_expr_get_type(Expr) != isl_ast_expr_int) { |
153 | isl_ast_expr_free(Expr); |
154 | return false; |
155 | } |
156 | auto ExprVal = isl_ast_expr_get_val(Expr); |
157 | isl_ast_expr_free(Expr); |
158 | if (Predicate(ExprVal) != isl_bool_true) { |
159 | isl_val_free(ExprVal); |
160 | return false; |
161 | } |
162 | isl_val_free(ExprVal); |
163 | return true; |
164 | } |
165 | |
166 | int IslNodeBuilder::getNumberOfIterations(isl::ast_node For) { |
167 | assert(isl_ast_node_get_type(For.get()) == isl_ast_node_for)((isl_ast_node_get_type(For.get()) == isl_ast_node_for) ? static_cast <void> (0) : __assert_fail ("isl_ast_node_get_type(For.get()) == isl_ast_node_for" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 167, __PRETTY_FUNCTION__)); |
168 | isl::ast_node Body = For.for_get_body(); |
169 | |
170 | // First, check if we can actually handle this code. |
171 | switch (isl_ast_node_get_type(Body.get())) { |
172 | case isl_ast_node_user: |
173 | break; |
174 | case isl_ast_node_block: { |
175 | isl::ast_node_list List = Body.block_get_children(); |
176 | for (isl::ast_node Node : List) { |
177 | isl_ast_node_type NodeType = isl_ast_node_get_type(Node.get()); |
178 | if (NodeType != isl_ast_node_user) |
179 | return -1; |
180 | } |
181 | break; |
182 | } |
183 | default: |
184 | return -1; |
185 | } |
186 | |
187 | isl::ast_expr Init = For.for_get_init(); |
188 | if (!checkIslAstExprInt(Init.release(), isl_val_is_zero)) |
189 | return -1; |
190 | isl::ast_expr Inc = For.for_get_inc(); |
191 | if (!checkIslAstExprInt(Inc.release(), isl_val_is_one)) |
192 | return -1; |
193 | CmpInst::Predicate Predicate; |
194 | isl::ast_expr UB = getUpperBound(For, Predicate); |
195 | if (isl_ast_expr_get_type(UB.get()) != isl_ast_expr_int) |
196 | return -1; |
197 | isl::val UpVal = UB.get_val(); |
198 | int NumberIterations = UpVal.get_num_si(); |
199 | if (NumberIterations < 0) |
200 | return -1; |
201 | if (Predicate == CmpInst::ICMP_SLT) |
202 | return NumberIterations; |
203 | else |
204 | return NumberIterations + 1; |
205 | } |
206 | |
207 | /// Extract the values and SCEVs needed to generate code for a block. |
208 | static int findReferencesInBlock(struct SubtreeReferences &References, |
209 | const ScopStmt *Stmt, BasicBlock *BB) { |
210 | for (Instruction &Inst : *BB) { |
211 | // Include invariant loads |
212 | if (isa<LoadInst>(Inst)) |
213 | if (Value *InvariantLoad = References.GlobalMap.lookup(&Inst)) |
214 | References.Values.insert(InvariantLoad); |
215 | |
216 | for (Value *SrcVal : Inst.operands()) { |
217 | auto *Scope = References.LI.getLoopFor(BB); |
218 | if (canSynthesize(SrcVal, References.S, &References.SE, Scope)) { |
219 | References.SCEVs.insert(References.SE.getSCEVAtScope(SrcVal, Scope)); |
220 | continue; |
221 | } else if (Value *NewVal = References.GlobalMap.lookup(SrcVal)) |
222 | References.Values.insert(NewVal); |
223 | } |
224 | } |
225 | return 0; |
226 | } |
227 | |
228 | void polly::addReferencesFromStmt(const ScopStmt *Stmt, void *UserPtr, |
229 | bool CreateScalarRefs) { |
230 | auto &References = *static_cast<struct SubtreeReferences *>(UserPtr); |
231 | |
232 | if (Stmt->isBlockStmt()) |
233 | findReferencesInBlock(References, Stmt, Stmt->getBasicBlock()); |
234 | else if (Stmt->isRegionStmt()) { |
235 | for (BasicBlock *BB : Stmt->getRegion()->blocks()) |
236 | findReferencesInBlock(References, Stmt, BB); |
237 | } else { |
238 | assert(Stmt->isCopyStmt())((Stmt->isCopyStmt()) ? static_cast<void> (0) : __assert_fail ("Stmt->isCopyStmt()", "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 238, __PRETTY_FUNCTION__)); |
239 | // Copy Stmts have no instructions that we need to consider. |
240 | } |
241 | |
242 | for (auto &Access : *Stmt) { |
243 | if (References.ParamSpace) { |
244 | isl::space ParamSpace = Access->getLatestAccessRelation().get_space(); |
245 | (*References.ParamSpace) = |
246 | References.ParamSpace->align_params(ParamSpace); |
247 | } |
248 | |
249 | if (Access->isLatestArrayKind()) { |
250 | auto *BasePtr = Access->getLatestScopArrayInfo()->getBasePtr(); |
251 | if (Instruction *OpInst = dyn_cast<Instruction>(BasePtr)) |
252 | if (Stmt->getParent()->contains(OpInst)) |
253 | continue; |
254 | |
255 | References.Values.insert(BasePtr); |
256 | continue; |
257 | } |
258 | |
259 | if (CreateScalarRefs) |
260 | References.Values.insert(References.BlockGen.getOrCreateAlloca(*Access)); |
261 | } |
262 | } |
263 | |
264 | /// Extract the out-of-scop values and SCEVs referenced from a set describing |
265 | /// a ScopStmt. |
266 | /// |
267 | /// This includes the SCEVUnknowns referenced by the SCEVs used in the |
268 | /// statement and the base pointers of the memory accesses. For scalar |
269 | /// statements we force the generation of alloca memory locations and list |
270 | /// these locations in the set of out-of-scop values as well. |
271 | /// |
272 | /// @param Set A set which references the ScopStmt we are interested in. |
273 | /// @param UserPtr A void pointer that can be casted to a SubtreeReferences |
274 | /// structure. |
275 | static void addReferencesFromStmtSet(isl::set Set, |
276 | struct SubtreeReferences *UserPtr) { |
277 | isl::id Id = Set.get_tuple_id(); |
278 | auto *Stmt = static_cast<const ScopStmt *>(Id.get_user()); |
279 | return addReferencesFromStmt(Stmt, UserPtr); |
280 | } |
281 | |
282 | /// Extract the out-of-scop values and SCEVs referenced from a union set |
283 | /// referencing multiple ScopStmts. |
284 | /// |
285 | /// This includes the SCEVUnknowns referenced by the SCEVs used in the |
286 | /// statement and the base pointers of the memory accesses. For scalar |
287 | /// statements we force the generation of alloca memory locations and list |
288 | /// these locations in the set of out-of-scop values as well. |
289 | /// |
290 | /// @param USet A union set referencing the ScopStmts we are interested |
291 | /// in. |
292 | /// @param References The SubtreeReferences data structure through which |
293 | /// results are returned and further information is |
294 | /// provided. |
295 | static void |
296 | addReferencesFromStmtUnionSet(isl::union_set USet, |
297 | struct SubtreeReferences &References) { |
298 | |
299 | for (isl::set Set : USet.get_set_list()) |
300 | addReferencesFromStmtSet(Set, &References); |
301 | } |
302 | |
303 | __isl_give isl_union_map * |
304 | IslNodeBuilder::getScheduleForAstNode(__isl_keep isl_ast_node *For) { |
305 | return IslAstInfo::getSchedule(For); |
306 | } |
307 | |
308 | void IslNodeBuilder::getReferencesInSubtree(__isl_keep isl_ast_node *For, |
309 | SetVector<Value *> &Values, |
310 | SetVector<const Loop *> &Loops) { |
311 | SetVector<const SCEV *> SCEVs; |
312 | struct SubtreeReferences References = { |
313 | LI, SE, S, ValueMap, Values, SCEVs, getBlockGenerator(), nullptr}; |
314 | |
315 | for (const auto &I : IDToValue) |
316 | Values.insert(I.second); |
317 | |
318 | // NOTE: this is populated in IslNodeBuilder::addParameters |
319 | for (const auto &I : OutsideLoopIterations) |
320 | Values.insert(cast<SCEVUnknown>(I.second)->getValue()); |
321 | |
322 | isl::union_set Schedule = |
323 | isl::manage(isl_union_map_domain(getScheduleForAstNode(For))); |
324 | addReferencesFromStmtUnionSet(Schedule, References); |
325 | |
326 | for (const SCEV *Expr : SCEVs) { |
327 | findValues(Expr, SE, Values); |
328 | findLoops(Expr, Loops); |
329 | } |
330 | |
331 | Values.remove_if([](const Value *V) { return isa<GlobalValue>(V); }); |
332 | |
333 | /// Note: Code generation of induction variables of loops outside Scops |
334 | /// |
335 | /// Remove loops that contain the scop or that are part of the scop, as they |
336 | /// are considered local. This leaves only loops that are before the scop, but |
337 | /// do not contain the scop itself. |
338 | /// We ignore loops perfectly contained in the Scop because these are already |
339 | /// generated at `IslNodeBuilder::addParameters`. These `Loops` are loops |
340 | /// whose induction variables are referred to by the Scop, but the Scop is not |
341 | /// fully contained in these Loops. Since there can be many of these, |
342 | /// we choose to codegen these on-demand. |
343 | /// @see IslNodeBuilder::materializeNonScopLoopInductionVariable. |
344 | Loops.remove_if([this](const Loop *L) { |
345 | return S.contains(L) || L->contains(S.getEntry()); |
346 | }); |
347 | |
348 | // Contains Values that may need to be replaced with other values |
349 | // due to replacements from the ValueMap. We should make sure |
350 | // that we return correctly remapped values. |
351 | // NOTE: this code path is tested by: |
352 | // 1. test/Isl/CodeGen/OpenMP/single_loop_with_loop_invariant_baseptr.ll |
353 | // 2. test/Isl/CodeGen/OpenMP/loop-body-references-outer-values-3.ll |
354 | SetVector<Value *> ReplacedValues; |
355 | for (Value *V : Values) { |
356 | ReplacedValues.insert(getLatestValue(V)); |
357 | } |
358 | Values = ReplacedValues; |
359 | } |
360 | |
361 | void IslNodeBuilder::updateValues(ValueMapT &NewValues) { |
362 | SmallPtrSet<Value *, 5> Inserted; |
363 | |
364 | for (const auto &I : IDToValue) { |
365 | IDToValue[I.first] = NewValues[I.second]; |
366 | Inserted.insert(I.second); |
367 | } |
368 | |
369 | for (const auto &I : NewValues) { |
370 | if (Inserted.count(I.first)) |
371 | continue; |
372 | |
373 | ValueMap[I.first] = I.second; |
374 | } |
375 | } |
376 | |
377 | Value *IslNodeBuilder::getLatestValue(Value *Original) const { |
378 | auto It = ValueMap.find(Original); |
379 | if (It == ValueMap.end()) |
380 | return Original; |
381 | return It->second; |
382 | } |
383 | |
384 | void IslNodeBuilder::createUserVector(__isl_take isl_ast_node *User, |
385 | std::vector<Value *> &IVS, |
386 | __isl_take isl_id *IteratorID, |
387 | __isl_take isl_union_map *Schedule) { |
388 | isl_ast_expr *Expr = isl_ast_node_user_get_expr(User); |
389 | isl_ast_expr *StmtExpr = isl_ast_expr_get_op_arg(Expr, 0); |
390 | isl_id *Id = isl_ast_expr_get_id(StmtExpr); |
391 | isl_ast_expr_free(StmtExpr); |
392 | ScopStmt *Stmt = (ScopStmt *)isl_id_get_user(Id); |
393 | std::vector<LoopToScevMapT> VLTS(IVS.size()); |
394 | |
395 | isl_union_set *Domain = isl_union_set_from_set(Stmt->getDomain().release()); |
396 | Schedule = isl_union_map_intersect_domain(Schedule, Domain); |
397 | isl_map *S = isl_map_from_union_map(Schedule); |
398 | |
399 | auto *NewAccesses = createNewAccesses(Stmt, User); |
400 | createSubstitutionsVector(Expr, Stmt, VLTS, IVS, IteratorID); |
401 | VectorBlockGenerator::generate(BlockGen, *Stmt, VLTS, S, NewAccesses); |
402 | isl_id_to_ast_expr_free(NewAccesses); |
403 | isl_map_free(S); |
404 | isl_id_free(Id); |
405 | isl_ast_node_free(User); |
406 | } |
407 | |
408 | void IslNodeBuilder::createMark(__isl_take isl_ast_node *Node) { |
409 | auto *Id = isl_ast_node_mark_get_id(Node); |
410 | auto Child = isl_ast_node_mark_get_node(Node); |
411 | isl_ast_node_free(Node); |
412 | // If a child node of a 'SIMD mark' is a loop that has a single iteration, |
413 | // it will be optimized away and we should skip it. |
414 | if (strcmp(isl_id_get_name(Id), "SIMD") == 0 && |
415 | isl_ast_node_get_type(Child) == isl_ast_node_for) { |
416 | bool Vector = PollyVectorizerChoice == VECTORIZER_POLLY; |
417 | int VectorWidth = getNumberOfIterations(isl::manage_copy(Child)); |
418 | if (Vector && 1 < VectorWidth && VectorWidth <= 16) |
419 | createForVector(Child, VectorWidth); |
420 | else |
421 | createForSequential(isl::manage(Child), true); |
422 | isl_id_free(Id); |
423 | return; |
424 | } |
425 | if (strcmp(isl_id_get_name(Id), "Inter iteration alias-free") == 0) { |
426 | auto *BasePtr = static_cast<Value *>(isl_id_get_user(Id)); |
427 | Annotator.addInterIterationAliasFreeBasePtr(BasePtr); |
428 | } |
429 | |
430 | BandAttr *ChildLoopAttr = getLoopAttr(isl::manage_copy(Id)); |
431 | BandAttr *AncestorLoopAttr; |
432 | if (ChildLoopAttr) { |
433 | // Save current LoopAttr environment to restore again when leaving this |
434 | // subtree. This means there was no loop between the ancestor LoopAttr and |
435 | // this mark, i.e. the ancestor LoopAttr did not directly mark a loop. This |
436 | // can happen e.g. if the AST build peeled or unrolled the loop. |
437 | AncestorLoopAttr = Annotator.getStagingAttrEnv(); |
438 | |
439 | Annotator.getStagingAttrEnv() = ChildLoopAttr; |
440 | } |
441 | |
442 | create(Child); |
443 | |
444 | if (ChildLoopAttr) { |
445 | assert(Annotator.getStagingAttrEnv() == ChildLoopAttr &&((Annotator.getStagingAttrEnv() == ChildLoopAttr && "Nest must not overwrite loop attr environment" ) ? static_cast<void> (0) : __assert_fail ("Annotator.getStagingAttrEnv() == ChildLoopAttr && \"Nest must not overwrite loop attr environment\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 446, __PRETTY_FUNCTION__)) |
446 | "Nest must not overwrite loop attr environment")((Annotator.getStagingAttrEnv() == ChildLoopAttr && "Nest must not overwrite loop attr environment" ) ? static_cast<void> (0) : __assert_fail ("Annotator.getStagingAttrEnv() == ChildLoopAttr && \"Nest must not overwrite loop attr environment\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 446, __PRETTY_FUNCTION__)); |
447 | Annotator.getStagingAttrEnv() = AncestorLoopAttr; |
448 | } |
449 | |
450 | isl_id_free(Id); |
451 | } |
452 | |
453 | void IslNodeBuilder::createForVector(__isl_take isl_ast_node *For, |
454 | int VectorWidth) { |
455 | isl_ast_node *Body = isl_ast_node_for_get_body(For); |
456 | isl_ast_expr *Init = isl_ast_node_for_get_init(For); |
457 | isl_ast_expr *Inc = isl_ast_node_for_get_inc(For); |
458 | isl_ast_expr *Iterator = isl_ast_node_for_get_iterator(For); |
459 | isl_id *IteratorID = isl_ast_expr_get_id(Iterator); |
460 | |
461 | Value *ValueLB = ExprBuilder.create(Init); |
462 | Value *ValueInc = ExprBuilder.create(Inc); |
463 | |
464 | Type *MaxType = ExprBuilder.getType(Iterator); |
465 | MaxType = ExprBuilder.getWidestType(MaxType, ValueLB->getType()); |
466 | MaxType = ExprBuilder.getWidestType(MaxType, ValueInc->getType()); |
467 | |
468 | if (MaxType != ValueLB->getType()) |
469 | ValueLB = Builder.CreateSExt(ValueLB, MaxType); |
470 | if (MaxType != ValueInc->getType()) |
471 | ValueInc = Builder.CreateSExt(ValueInc, MaxType); |
472 | |
473 | std::vector<Value *> IVS(VectorWidth); |
474 | IVS[0] = ValueLB; |
475 | |
476 | for (int i = 1; i < VectorWidth; i++) |
477 | IVS[i] = Builder.CreateAdd(IVS[i - 1], ValueInc, "p_vector_iv"); |
478 | |
479 | isl_union_map *Schedule = getScheduleForAstNode(For); |
480 | assert(Schedule && "For statement annotation does not contain its schedule")((Schedule && "For statement annotation does not contain its schedule" ) ? static_cast<void> (0) : __assert_fail ("Schedule && \"For statement annotation does not contain its schedule\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 480, __PRETTY_FUNCTION__)); |
481 | |
482 | IDToValue[IteratorID] = ValueLB; |
483 | |
484 | switch (isl_ast_node_get_type(Body)) { |
485 | case isl_ast_node_user: |
486 | createUserVector(Body, IVS, isl_id_copy(IteratorID), |
487 | isl_union_map_copy(Schedule)); |
488 | break; |
489 | case isl_ast_node_block: { |
490 | isl_ast_node_list *List = isl_ast_node_block_get_children(Body); |
491 | |
492 | for (int i = 0; i < isl_ast_node_list_n_ast_node(List); ++i) |
493 | createUserVector(isl_ast_node_list_get_ast_node(List, i), IVS, |
494 | isl_id_copy(IteratorID), isl_union_map_copy(Schedule)); |
495 | |
496 | isl_ast_node_free(Body); |
497 | isl_ast_node_list_free(List); |
498 | break; |
499 | } |
500 | default: |
501 | isl_ast_node_dump(Body); |
502 | llvm_unreachable("Unhandled isl_ast_node in vectorizer")::llvm::llvm_unreachable_internal("Unhandled isl_ast_node in vectorizer" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 502); |
503 | } |
504 | |
505 | IDToValue.erase(IDToValue.find(IteratorID)); |
506 | isl_id_free(IteratorID); |
507 | isl_union_map_free(Schedule); |
508 | |
509 | isl_ast_node_free(For); |
510 | isl_ast_expr_free(Iterator); |
511 | |
512 | VectorLoops++; |
513 | } |
514 | |
515 | /// Restore the initial ordering of dimensions of the band node |
516 | /// |
517 | /// In case the band node represents all the dimensions of the iteration |
518 | /// domain, recreate the band node to restore the initial ordering of the |
519 | /// dimensions. |
520 | /// |
521 | /// @param Node The band node to be modified. |
522 | /// @return The modified schedule node. |
523 | static bool IsLoopVectorizerDisabled(isl::ast_node Node) { |
524 | assert(isl_ast_node_get_type(Node.get()) == isl_ast_node_for)((isl_ast_node_get_type(Node.get()) == isl_ast_node_for) ? static_cast <void> (0) : __assert_fail ("isl_ast_node_get_type(Node.get()) == isl_ast_node_for" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 524, __PRETTY_FUNCTION__)); |
525 | auto Body = Node.for_get_body(); |
526 | if (isl_ast_node_get_type(Body.get()) != isl_ast_node_mark) |
527 | return false; |
528 | auto Id = Body.mark_get_id(); |
529 | if (strcmp(Id.get_name().c_str(), "Loop Vectorizer Disabled") == 0) |
530 | return true; |
531 | return false; |
532 | } |
533 | |
534 | void IslNodeBuilder::createForSequential(isl::ast_node For, bool MarkParallel) { |
535 | Value *ValueLB, *ValueUB, *ValueInc; |
536 | Type *MaxType; |
537 | BasicBlock *ExitBlock; |
538 | Value *IV; |
539 | CmpInst::Predicate Predicate; |
540 | |
541 | bool LoopVectorizerDisabled = IsLoopVectorizerDisabled(For); |
542 | |
543 | isl::ast_node Body = For.for_get_body(); |
544 | |
545 | // isl_ast_node_for_is_degenerate(For) |
546 | // |
547 | // TODO: For degenerated loops we could generate a plain assignment. |
548 | // However, for now we just reuse the logic for normal loops, which will |
549 | // create a loop with a single iteration. |
550 | |
551 | isl::ast_expr Init = For.for_get_init(); |
552 | isl::ast_expr Inc = For.for_get_inc(); |
553 | isl::ast_expr Iterator = For.for_get_iterator(); |
554 | isl::id IteratorID = Iterator.get_id(); |
555 | isl::ast_expr UB = getUpperBound(For, Predicate); |
556 | |
557 | ValueLB = ExprBuilder.create(Init.release()); |
558 | ValueUB = ExprBuilder.create(UB.release()); |
559 | ValueInc = ExprBuilder.create(Inc.release()); |
560 | |
561 | MaxType = ExprBuilder.getType(Iterator.get()); |
562 | MaxType = ExprBuilder.getWidestType(MaxType, ValueLB->getType()); |
563 | MaxType = ExprBuilder.getWidestType(MaxType, ValueUB->getType()); |
564 | MaxType = ExprBuilder.getWidestType(MaxType, ValueInc->getType()); |
565 | |
566 | if (MaxType != ValueLB->getType()) |
567 | ValueLB = Builder.CreateSExt(ValueLB, MaxType); |
568 | if (MaxType != ValueUB->getType()) |
569 | ValueUB = Builder.CreateSExt(ValueUB, MaxType); |
570 | if (MaxType != ValueInc->getType()) |
571 | ValueInc = Builder.CreateSExt(ValueInc, MaxType); |
572 | |
573 | // If we can show that LB <Predicate> UB holds at least once, we can |
574 | // omit the GuardBB in front of the loop. |
575 | bool UseGuardBB = |
576 | !SE.isKnownPredicate(Predicate, SE.getSCEV(ValueLB), SE.getSCEV(ValueUB)); |
577 | IV = createLoop(ValueLB, ValueUB, ValueInc, Builder, LI, DT, ExitBlock, |
578 | Predicate, &Annotator, MarkParallel, UseGuardBB, |
579 | LoopVectorizerDisabled); |
580 | IDToValue[IteratorID.get()] = IV; |
581 | |
582 | create(Body.release()); |
583 | |
584 | Annotator.popLoop(MarkParallel); |
585 | |
586 | IDToValue.erase(IDToValue.find(IteratorID.get())); |
587 | |
588 | Builder.SetInsertPoint(&ExitBlock->front()); |
589 | |
590 | SequentialLoops++; |
591 | } |
592 | |
593 | /// Remove the BBs contained in a (sub)function from the dominator tree. |
594 | /// |
595 | /// This function removes the basic blocks that are part of a subfunction from |
596 | /// the dominator tree. Specifically, when generating code it may happen that at |
597 | /// some point the code generation continues in a new sub-function (e.g., when |
598 | /// generating OpenMP code). The basic blocks that are created in this |
599 | /// sub-function are then still part of the dominator tree of the original |
600 | /// function, such that the dominator tree reaches over function boundaries. |
601 | /// This is not only incorrect, but also causes crashes. This function now |
602 | /// removes from the dominator tree all basic blocks that are dominated (and |
603 | /// consequently reachable) from the entry block of this (sub)function. |
604 | /// |
605 | /// FIXME: A LLVM (function or region) pass should not touch anything outside of |
606 | /// the function/region it runs on. Hence, the pure need for this function shows |
607 | /// that we do not comply to this rule. At the moment, this does not cause any |
608 | /// issues, but we should be aware that such issues may appear. Unfortunately |
609 | /// the current LLVM pass infrastructure does not allow to make Polly a module |
610 | /// or call-graph pass to solve this issue, as such a pass would not have access |
611 | /// to the per-function analyses passes needed by Polly. A future pass manager |
612 | /// infrastructure is supposed to enable such kind of access possibly allowing |
613 | /// us to create a cleaner solution here. |
614 | /// |
615 | /// FIXME: Instead of adding the dominance information and then dropping it |
616 | /// later on, we should try to just not add it in the first place. This requires |
617 | /// some careful testing to make sure this does not break in interaction with |
618 | /// the SCEVBuilder and SplitBlock which may rely on the dominator tree or |
619 | /// which may try to update it. |
620 | /// |
621 | /// @param F The function which contains the BBs to removed. |
622 | /// @param DT The dominator tree from which to remove the BBs. |
623 | static void removeSubFuncFromDomTree(Function *F, DominatorTree &DT) { |
624 | DomTreeNode *N = DT.getNode(&F->getEntryBlock()); |
625 | std::vector<BasicBlock *> Nodes; |
626 | |
627 | // We can only remove an element from the dominator tree, if all its children |
628 | // have been removed. To ensure this we obtain the list of nodes to remove |
629 | // using a post-order tree traversal. |
630 | for (po_iterator<DomTreeNode *> I = po_begin(N), E = po_end(N); I != E; ++I) |
631 | Nodes.push_back(I->getBlock()); |
632 | |
633 | for (BasicBlock *BB : Nodes) |
634 | DT.eraseNode(BB); |
635 | } |
636 | |
637 | void IslNodeBuilder::createForParallel(__isl_take isl_ast_node *For) { |
638 | isl_ast_node *Body; |
639 | isl_ast_expr *Init, *Inc, *Iterator, *UB; |
640 | isl_id *IteratorID; |
641 | Value *ValueLB, *ValueUB, *ValueInc; |
642 | Type *MaxType; |
643 | Value *IV; |
644 | CmpInst::Predicate Predicate; |
645 | |
646 | // The preamble of parallel code interacts different than normal code with |
647 | // e.g., scalar initialization. Therefore, we ensure the parallel code is |
648 | // separated from the last basic block. |
649 | BasicBlock *ParBB = SplitBlock(Builder.GetInsertBlock(), |
650 | &*Builder.GetInsertPoint(), &DT, &LI); |
651 | ParBB->setName("polly.parallel.for"); |
652 | Builder.SetInsertPoint(&ParBB->front()); |
653 | |
654 | Body = isl_ast_node_for_get_body(For); |
655 | Init = isl_ast_node_for_get_init(For); |
656 | Inc = isl_ast_node_for_get_inc(For); |
657 | Iterator = isl_ast_node_for_get_iterator(For); |
658 | IteratorID = isl_ast_expr_get_id(Iterator); |
659 | UB = getUpperBound(isl::manage_copy(For), Predicate).release(); |
660 | |
661 | ValueLB = ExprBuilder.create(Init); |
662 | ValueUB = ExprBuilder.create(UB); |
663 | ValueInc = ExprBuilder.create(Inc); |
664 | |
665 | // OpenMP always uses SLE. In case the isl generated AST uses a SLT |
666 | // expression, we need to adjust the loop bound by one. |
667 | if (Predicate == CmpInst::ICMP_SLT) |
668 | ValueUB = Builder.CreateAdd( |
669 | ValueUB, Builder.CreateSExt(Builder.getTrue(), ValueUB->getType())); |
670 | |
671 | MaxType = ExprBuilder.getType(Iterator); |
672 | MaxType = ExprBuilder.getWidestType(MaxType, ValueLB->getType()); |
673 | MaxType = ExprBuilder.getWidestType(MaxType, ValueUB->getType()); |
674 | MaxType = ExprBuilder.getWidestType(MaxType, ValueInc->getType()); |
675 | |
676 | if (MaxType != ValueLB->getType()) |
677 | ValueLB = Builder.CreateSExt(ValueLB, MaxType); |
678 | if (MaxType != ValueUB->getType()) |
679 | ValueUB = Builder.CreateSExt(ValueUB, MaxType); |
680 | if (MaxType != ValueInc->getType()) |
681 | ValueInc = Builder.CreateSExt(ValueInc, MaxType); |
682 | |
683 | BasicBlock::iterator LoopBody; |
684 | |
685 | SetVector<Value *> SubtreeValues; |
686 | SetVector<const Loop *> Loops; |
687 | |
688 | getReferencesInSubtree(For, SubtreeValues, Loops); |
689 | |
690 | // Create for all loops we depend on values that contain the current loop |
691 | // iteration. These values are necessary to generate code for SCEVs that |
692 | // depend on such loops. As a result we need to pass them to the subfunction. |
693 | // See [Code generation of induction variables of loops outside Scops] |
694 | for (const Loop *L : Loops) { |
695 | Value *LoopInductionVar = materializeNonScopLoopInductionVariable(L); |
696 | SubtreeValues.insert(LoopInductionVar); |
697 | } |
698 | |
699 | ValueMapT NewValues; |
700 | |
701 | std::unique_ptr<ParallelLoopGenerator> ParallelLoopGenPtr; |
702 | |
703 | switch (PollyOmpBackend) { |
704 | case OpenMPBackend::GNU: |
705 | ParallelLoopGenPtr.reset( |
706 | new ParallelLoopGeneratorGOMP(Builder, LI, DT, DL)); |
707 | break; |
708 | case OpenMPBackend::LLVM: |
709 | ParallelLoopGenPtr.reset(new ParallelLoopGeneratorKMP(Builder, LI, DT, DL)); |
710 | break; |
711 | } |
712 | |
713 | IV = ParallelLoopGenPtr->createParallelLoop( |
714 | ValueLB, ValueUB, ValueInc, SubtreeValues, NewValues, &LoopBody); |
715 | BasicBlock::iterator AfterLoop = Builder.GetInsertPoint(); |
716 | Builder.SetInsertPoint(&*LoopBody); |
717 | |
718 | // Remember the parallel subfunction |
719 | ParallelSubfunctions.push_back(LoopBody->getFunction()); |
720 | |
721 | // Save the current values. |
722 | auto ValueMapCopy = ValueMap; |
723 | IslExprBuilder::IDToValueTy IDToValueCopy = IDToValue; |
724 | |
725 | updateValues(NewValues); |
726 | IDToValue[IteratorID] = IV; |
727 | |
728 | ValueMapT NewValuesReverse; |
729 | |
730 | for (auto P : NewValues) |
731 | NewValuesReverse[P.second] = P.first; |
732 | |
733 | Annotator.addAlternativeAliasBases(NewValuesReverse); |
734 | |
735 | create(Body); |
736 | |
737 | Annotator.resetAlternativeAliasBases(); |
738 | // Restore the original values. |
739 | ValueMap = ValueMapCopy; |
740 | IDToValue = IDToValueCopy; |
741 | |
742 | Builder.SetInsertPoint(&*AfterLoop); |
743 | removeSubFuncFromDomTree((*LoopBody).getParent()->getParent(), DT); |
744 | |
745 | for (const Loop *L : Loops) |
746 | OutsideLoopIterations.erase(L); |
747 | |
748 | isl_ast_node_free(For); |
749 | isl_ast_expr_free(Iterator); |
750 | isl_id_free(IteratorID); |
751 | |
752 | ParallelLoops++; |
753 | } |
754 | |
755 | /// Return whether any of @p Node's statements contain partial accesses. |
756 | /// |
757 | /// Partial accesses are not supported by Polly's vector code generator. |
758 | static bool hasPartialAccesses(__isl_take isl_ast_node *Node) { |
759 | return isl_ast_node_foreach_descendant_top_down( |
760 | Node, |
761 | [](isl_ast_node *Node, void *User) -> isl_bool { |
762 | if (isl_ast_node_get_type(Node) != isl_ast_node_user) |
763 | return isl_bool_true; |
764 | |
765 | isl::ast_expr Expr = |
766 | isl::manage(isl_ast_node_user_get_expr(Node)); |
767 | isl::ast_expr StmtExpr = Expr.get_op_arg(0); |
768 | isl::id Id = StmtExpr.get_id(); |
769 | |
770 | ScopStmt *Stmt = |
771 | static_cast<ScopStmt *>(isl_id_get_user(Id.get())); |
772 | isl::set StmtDom = Stmt->getDomain(); |
773 | for (auto *MA : *Stmt) { |
774 | if (MA->isLatestPartialAccess()) |
775 | return isl_bool_error; |
776 | } |
777 | return isl_bool_true; |
778 | }, |
779 | nullptr) == isl_stat_error; |
780 | } |
781 | |
782 | void IslNodeBuilder::createFor(__isl_take isl_ast_node *For) { |
783 | bool Vector = PollyVectorizerChoice == VECTORIZER_POLLY; |
784 | |
785 | if (Vector && IslAstInfo::isInnermostParallel(isl::manage_copy(For)) && |
786 | !IslAstInfo::isReductionParallel(For)) { |
787 | int VectorWidth = getNumberOfIterations(isl::manage_copy(For)); |
788 | if (1 < VectorWidth && VectorWidth <= 16 && !hasPartialAccesses(For)) { |
789 | createForVector(For, VectorWidth); |
790 | return; |
791 | } |
792 | } |
793 | |
794 | if (IslAstInfo::isExecutedInParallel(For)) { |
795 | createForParallel(For); |
796 | return; |
797 | } |
798 | bool Parallel = |
799 | (IslAstInfo::isParallel(For) && !IslAstInfo::isReductionParallel(For)); |
800 | createForSequential(isl::manage(For), Parallel); |
801 | } |
802 | |
803 | void IslNodeBuilder::createIf(__isl_take isl_ast_node *If) { |
804 | isl_ast_expr *Cond = isl_ast_node_if_get_cond(If); |
805 | |
806 | Function *F = Builder.GetInsertBlock()->getParent(); |
807 | LLVMContext &Context = F->getContext(); |
808 | |
809 | BasicBlock *CondBB = SplitBlock(Builder.GetInsertBlock(), |
810 | &*Builder.GetInsertPoint(), &DT, &LI); |
811 | CondBB->setName("polly.cond"); |
812 | BasicBlock *MergeBB = SplitBlock(CondBB, &CondBB->front(), &DT, &LI); |
813 | MergeBB->setName("polly.merge"); |
814 | BasicBlock *ThenBB = BasicBlock::Create(Context, "polly.then", F); |
815 | BasicBlock *ElseBB = BasicBlock::Create(Context, "polly.else", F); |
816 | |
817 | DT.addNewBlock(ThenBB, CondBB); |
818 | DT.addNewBlock(ElseBB, CondBB); |
819 | DT.changeImmediateDominator(MergeBB, CondBB); |
820 | |
821 | Loop *L = LI.getLoopFor(CondBB); |
822 | if (L) { |
823 | L->addBasicBlockToLoop(ThenBB, LI); |
824 | L->addBasicBlockToLoop(ElseBB, LI); |
825 | } |
826 | |
827 | CondBB->getTerminator()->eraseFromParent(); |
828 | |
829 | Builder.SetInsertPoint(CondBB); |
830 | Value *Predicate = ExprBuilder.create(Cond); |
831 | Builder.CreateCondBr(Predicate, ThenBB, ElseBB); |
832 | Builder.SetInsertPoint(ThenBB); |
833 | Builder.CreateBr(MergeBB); |
834 | Builder.SetInsertPoint(ElseBB); |
835 | Builder.CreateBr(MergeBB); |
836 | Builder.SetInsertPoint(&ThenBB->front()); |
837 | |
838 | create(isl_ast_node_if_get_then(If)); |
839 | |
840 | Builder.SetInsertPoint(&ElseBB->front()); |
841 | |
842 | if (isl_ast_node_if_has_else(If)) |
843 | create(isl_ast_node_if_get_else(If)); |
844 | |
845 | Builder.SetInsertPoint(&MergeBB->front()); |
846 | |
847 | isl_ast_node_free(If); |
848 | |
849 | IfConditions++; |
850 | } |
851 | |
852 | __isl_give isl_id_to_ast_expr * |
853 | IslNodeBuilder::createNewAccesses(ScopStmt *Stmt, |
854 | __isl_keep isl_ast_node *Node) { |
855 | isl_id_to_ast_expr *NewAccesses = |
856 | isl_id_to_ast_expr_alloc(Stmt->getParent()->getIslCtx().get(), 0); |
857 | |
858 | auto *Build = IslAstInfo::getBuild(Node); |
859 | assert(Build && "Could not obtain isl_ast_build from user node")((Build && "Could not obtain isl_ast_build from user node" ) ? static_cast<void> (0) : __assert_fail ("Build && \"Could not obtain isl_ast_build from user node\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 859, __PRETTY_FUNCTION__)); |
860 | Stmt->setAstBuild(isl::manage_copy(Build)); |
861 | |
862 | for (auto *MA : *Stmt) { |
863 | if (!MA->hasNewAccessRelation()) { |
864 | if (PollyGenerateExpressions) { |
865 | if (!MA->isAffine()) |
866 | continue; |
867 | if (MA->getLatestScopArrayInfo()->getBasePtrOriginSAI()) |
868 | continue; |
869 | |
870 | auto *BasePtr = |
871 | dyn_cast<Instruction>(MA->getLatestScopArrayInfo()->getBasePtr()); |
872 | if (BasePtr && Stmt->getParent()->getRegion().contains(BasePtr)) |
873 | continue; |
874 | } else { |
875 | continue; |
876 | } |
877 | } |
878 | assert(MA->isAffine() &&((MA->isAffine() && "Only affine memory accesses can be code generated" ) ? static_cast<void> (0) : __assert_fail ("MA->isAffine() && \"Only affine memory accesses can be code generated\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 879, __PRETTY_FUNCTION__)) |
879 | "Only affine memory accesses can be code generated")((MA->isAffine() && "Only affine memory accesses can be code generated" ) ? static_cast<void> (0) : __assert_fail ("MA->isAffine() && \"Only affine memory accesses can be code generated\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 879, __PRETTY_FUNCTION__)); |
880 | |
881 | auto Schedule = isl_ast_build_get_schedule(Build); |
882 | |
883 | #ifndef NDEBUG |
884 | if (MA->isRead()) { |
885 | auto Dom = Stmt->getDomain().release(); |
886 | auto SchedDom = isl_set_from_union_set( |
887 | isl_union_map_domain(isl_union_map_copy(Schedule))); |
888 | auto AccDom = isl_map_domain(MA->getAccessRelation().release()); |
889 | Dom = isl_set_intersect_params(Dom, |
890 | Stmt->getParent()->getContext().release()); |
891 | SchedDom = isl_set_intersect_params( |
892 | SchedDom, Stmt->getParent()->getContext().release()); |
893 | assert(isl_set_is_subset(SchedDom, AccDom) &&((isl_set_is_subset(SchedDom, AccDom) && "Access relation not defined on full schedule domain" ) ? static_cast<void> (0) : __assert_fail ("isl_set_is_subset(SchedDom, AccDom) && \"Access relation not defined on full schedule domain\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 894, __PRETTY_FUNCTION__)) |
894 | "Access relation not defined on full schedule domain")((isl_set_is_subset(SchedDom, AccDom) && "Access relation not defined on full schedule domain" ) ? static_cast<void> (0) : __assert_fail ("isl_set_is_subset(SchedDom, AccDom) && \"Access relation not defined on full schedule domain\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 894, __PRETTY_FUNCTION__)); |
895 | assert(isl_set_is_subset(Dom, AccDom) &&((isl_set_is_subset(Dom, AccDom) && "Access relation not defined on full domain" ) ? static_cast<void> (0) : __assert_fail ("isl_set_is_subset(Dom, AccDom) && \"Access relation not defined on full domain\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 896, __PRETTY_FUNCTION__)) |
896 | "Access relation not defined on full domain")((isl_set_is_subset(Dom, AccDom) && "Access relation not defined on full domain" ) ? static_cast<void> (0) : __assert_fail ("isl_set_is_subset(Dom, AccDom) && \"Access relation not defined on full domain\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 896, __PRETTY_FUNCTION__)); |
897 | isl_set_free(AccDom); |
898 | isl_set_free(SchedDom); |
899 | isl_set_free(Dom); |
900 | } |
901 | #endif |
902 | |
903 | auto PWAccRel = |
904 | MA->applyScheduleToAccessRelation(isl::manage(Schedule)).release(); |
905 | |
906 | // isl cannot generate an index expression for access-nothing accesses. |
907 | isl::set AccDomain = |
908 | isl::manage(isl_pw_multi_aff_domain(isl_pw_multi_aff_copy(PWAccRel))); |
909 | isl::set Context = S.getContext(); |
910 | AccDomain = AccDomain.intersect_params(Context); |
911 | if (AccDomain.is_empty()) { |
912 | isl_pw_multi_aff_free(PWAccRel); |
913 | continue; |
914 | } |
915 | |
916 | auto AccessExpr = isl_ast_build_access_from_pw_multi_aff(Build, PWAccRel); |
917 | NewAccesses = |
918 | isl_id_to_ast_expr_set(NewAccesses, MA->getId().release(), AccessExpr); |
919 | } |
920 | |
921 | return NewAccesses; |
922 | } |
923 | |
924 | void IslNodeBuilder::createSubstitutions(__isl_take isl_ast_expr *Expr, |
925 | ScopStmt *Stmt, LoopToScevMapT <S) { |
926 | assert(isl_ast_expr_get_type(Expr) == isl_ast_expr_op &&((isl_ast_expr_get_type(Expr) == isl_ast_expr_op && "Expression of type 'op' expected" ) ? static_cast<void> (0) : __assert_fail ("isl_ast_expr_get_type(Expr) == isl_ast_expr_op && \"Expression of type 'op' expected\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 927, __PRETTY_FUNCTION__)) |
927 | "Expression of type 'op' expected")((isl_ast_expr_get_type(Expr) == isl_ast_expr_op && "Expression of type 'op' expected" ) ? static_cast<void> (0) : __assert_fail ("isl_ast_expr_get_type(Expr) == isl_ast_expr_op && \"Expression of type 'op' expected\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 927, __PRETTY_FUNCTION__)); |
928 | assert(isl_ast_expr_get_op_type(Expr) == isl_ast_op_call &&((isl_ast_expr_get_op_type(Expr) == isl_ast_expr_op_call && "Operation of type 'call' expected") ? static_cast<void> (0) : __assert_fail ("isl_ast_expr_get_op_type(Expr) == isl_ast_op_call && \"Operation of type 'call' expected\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 929, __PRETTY_FUNCTION__)) |
929 | "Operation of type 'call' expected")((isl_ast_expr_get_op_type(Expr) == isl_ast_expr_op_call && "Operation of type 'call' expected") ? static_cast<void> (0) : __assert_fail ("isl_ast_expr_get_op_type(Expr) == isl_ast_op_call && \"Operation of type 'call' expected\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 929, __PRETTY_FUNCTION__)); |
930 | for (int i = 0; i < isl_ast_expr_get_op_n_arg(Expr) - 1; ++i) { |
931 | isl_ast_expr *SubExpr; |
932 | Value *V; |
933 | |
934 | SubExpr = isl_ast_expr_get_op_arg(Expr, i + 1); |
935 | V = ExprBuilder.create(SubExpr); |
936 | ScalarEvolution *SE = Stmt->getParent()->getSE(); |
937 | LTS[Stmt->getLoopForDimension(i)] = SE->getUnknown(V); |
938 | } |
939 | |
940 | isl_ast_expr_free(Expr); |
941 | } |
942 | |
943 | void IslNodeBuilder::createSubstitutionsVector( |
944 | __isl_take isl_ast_expr *Expr, ScopStmt *Stmt, |
945 | std::vector<LoopToScevMapT> &VLTS, std::vector<Value *> &IVS, |
946 | __isl_take isl_id *IteratorID) { |
947 | int i = 0; |
948 | |
949 | Value *OldValue = IDToValue[IteratorID]; |
950 | for (Value *IV : IVS) { |
951 | IDToValue[IteratorID] = IV; |
952 | createSubstitutions(isl_ast_expr_copy(Expr), Stmt, VLTS[i]); |
953 | i++; |
954 | } |
955 | |
956 | IDToValue[IteratorID] = OldValue; |
957 | isl_id_free(IteratorID); |
958 | isl_ast_expr_free(Expr); |
959 | } |
960 | |
961 | void IslNodeBuilder::generateCopyStmt( |
962 | ScopStmt *Stmt, __isl_keep isl_id_to_ast_expr *NewAccesses) { |
963 | assert(Stmt->size() == 2)((Stmt->size() == 2) ? static_cast<void> (0) : __assert_fail ("Stmt->size() == 2", "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 963, __PRETTY_FUNCTION__)); |
964 | auto ReadAccess = Stmt->begin(); |
965 | auto WriteAccess = ReadAccess++; |
966 | assert((*ReadAccess)->isRead() && (*WriteAccess)->isMustWrite())(((*ReadAccess)->isRead() && (*WriteAccess)->isMustWrite ()) ? static_cast<void> (0) : __assert_fail ("(*ReadAccess)->isRead() && (*WriteAccess)->isMustWrite()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 966, __PRETTY_FUNCTION__)); |
967 | assert((*ReadAccess)->getElementType() == (*WriteAccess)->getElementType() &&(((*ReadAccess)->getElementType() == (*WriteAccess)->getElementType () && "Accesses use the same data type") ? static_cast <void> (0) : __assert_fail ("(*ReadAccess)->getElementType() == (*WriteAccess)->getElementType() && \"Accesses use the same data type\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 968, __PRETTY_FUNCTION__)) |
968 | "Accesses use the same data type")(((*ReadAccess)->getElementType() == (*WriteAccess)->getElementType () && "Accesses use the same data type") ? static_cast <void> (0) : __assert_fail ("(*ReadAccess)->getElementType() == (*WriteAccess)->getElementType() && \"Accesses use the same data type\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 968, __PRETTY_FUNCTION__)); |
969 | assert((*ReadAccess)->isArrayKind() && (*WriteAccess)->isArrayKind())(((*ReadAccess)->isArrayKind() && (*WriteAccess)-> isArrayKind()) ? static_cast<void> (0) : __assert_fail ( "(*ReadAccess)->isArrayKind() && (*WriteAccess)->isArrayKind()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 969, __PRETTY_FUNCTION__)); |
970 | auto *AccessExpr = |
971 | isl_id_to_ast_expr_get(NewAccesses, (*ReadAccess)->getId().release()); |
972 | auto *LoadValue = ExprBuilder.create(AccessExpr); |
973 | AccessExpr = |
974 | isl_id_to_ast_expr_get(NewAccesses, (*WriteAccess)->getId().release()); |
975 | auto *StoreAddr = ExprBuilder.createAccessAddress(AccessExpr).first; |
976 | Builder.CreateStore(LoadValue, StoreAddr); |
977 | } |
978 | |
979 | Value *IslNodeBuilder::materializeNonScopLoopInductionVariable(const Loop *L) { |
980 | assert(OutsideLoopIterations.find(L) == OutsideLoopIterations.end() &&((OutsideLoopIterations.find(L) == OutsideLoopIterations.end( ) && "trying to materialize loop induction variable twice" ) ? static_cast<void> (0) : __assert_fail ("OutsideLoopIterations.find(L) == OutsideLoopIterations.end() && \"trying to materialize loop induction variable twice\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 981, __PRETTY_FUNCTION__)) |
981 | "trying to materialize loop induction variable twice")((OutsideLoopIterations.find(L) == OutsideLoopIterations.end( ) && "trying to materialize loop induction variable twice" ) ? static_cast<void> (0) : __assert_fail ("OutsideLoopIterations.find(L) == OutsideLoopIterations.end() && \"trying to materialize loop induction variable twice\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 981, __PRETTY_FUNCTION__)); |
982 | const SCEV *OuterLIV = SE.getAddRecExpr(SE.getUnknown(Builder.getInt64(0)), |
983 | SE.getUnknown(Builder.getInt64(1)), L, |
984 | SCEV::FlagAnyWrap); |
985 | Value *V = generateSCEV(OuterLIV); |
986 | OutsideLoopIterations[L] = SE.getUnknown(V); |
987 | return V; |
988 | } |
989 | |
990 | void IslNodeBuilder::createUser(__isl_take isl_ast_node *User) { |
991 | LoopToScevMapT LTS; |
992 | isl_id *Id; |
993 | ScopStmt *Stmt; |
994 | |
995 | isl_ast_expr *Expr = isl_ast_node_user_get_expr(User); |
996 | isl_ast_expr *StmtExpr = isl_ast_expr_get_op_arg(Expr, 0); |
997 | Id = isl_ast_expr_get_id(StmtExpr); |
998 | isl_ast_expr_free(StmtExpr); |
999 | |
1000 | LTS.insert(OutsideLoopIterations.begin(), OutsideLoopIterations.end()); |
1001 | |
1002 | Stmt = (ScopStmt *)isl_id_get_user(Id); |
1003 | auto *NewAccesses = createNewAccesses(Stmt, User); |
1004 | if (Stmt->isCopyStmt()) { |
1005 | generateCopyStmt(Stmt, NewAccesses); |
1006 | isl_ast_expr_free(Expr); |
1007 | } else { |
1008 | createSubstitutions(Expr, Stmt, LTS); |
1009 | |
1010 | if (Stmt->isBlockStmt()) |
1011 | BlockGen.copyStmt(*Stmt, LTS, NewAccesses); |
1012 | else |
1013 | RegionGen.copyStmt(*Stmt, LTS, NewAccesses); |
1014 | } |
1015 | |
1016 | isl_id_to_ast_expr_free(NewAccesses); |
1017 | isl_ast_node_free(User); |
1018 | isl_id_free(Id); |
1019 | } |
1020 | |
1021 | void IslNodeBuilder::createBlock(__isl_take isl_ast_node *Block) { |
1022 | isl_ast_node_list *List = isl_ast_node_block_get_children(Block); |
1023 | |
1024 | for (int i = 0; i < isl_ast_node_list_n_ast_node(List); ++i) |
1025 | create(isl_ast_node_list_get_ast_node(List, i)); |
1026 | |
1027 | isl_ast_node_free(Block); |
1028 | isl_ast_node_list_free(List); |
1029 | } |
1030 | |
1031 | void IslNodeBuilder::create(__isl_take isl_ast_node *Node) { |
1032 | switch (isl_ast_node_get_type(Node)) { |
1033 | case isl_ast_node_error: |
1034 | llvm_unreachable("code generation error")::llvm::llvm_unreachable_internal("code generation error", "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1034); |
1035 | case isl_ast_node_mark: |
1036 | createMark(Node); |
1037 | return; |
1038 | case isl_ast_node_for: |
1039 | createFor(Node); |
1040 | return; |
1041 | case isl_ast_node_if: |
1042 | createIf(Node); |
1043 | return; |
1044 | case isl_ast_node_user: |
1045 | createUser(Node); |
1046 | return; |
1047 | case isl_ast_node_block: |
1048 | createBlock(Node); |
1049 | return; |
1050 | } |
1051 | |
1052 | llvm_unreachable("Unknown isl_ast_node type")::llvm::llvm_unreachable_internal("Unknown isl_ast_node type" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1052); |
1053 | } |
1054 | |
1055 | bool IslNodeBuilder::materializeValue(isl_id *Id) { |
1056 | // If the Id is already mapped, skip it. |
1057 | if (!IDToValue.count(Id)) { |
1058 | auto *ParamSCEV = (const SCEV *)isl_id_get_user(Id); |
1059 | Value *V = nullptr; |
1060 | |
1061 | // Parameters could refer to invariant loads that need to be |
1062 | // preloaded before we can generate code for the parameter. Thus, |
1063 | // check if any value referred to in ParamSCEV is an invariant load |
1064 | // and if so make sure its equivalence class is preloaded. |
1065 | SetVector<Value *> Values; |
1066 | findValues(ParamSCEV, SE, Values); |
1067 | for (auto *Val : Values) { |
1068 | // Check if the value is an instruction in a dead block within the SCoP |
1069 | // and if so do not code generate it. |
1070 | if (auto *Inst = dyn_cast<Instruction>(Val)) { |
1071 | if (S.contains(Inst)) { |
1072 | bool IsDead = true; |
1073 | |
1074 | // Check for "undef" loads first, then if there is a statement for |
1075 | // the parent of Inst and lastly if the parent of Inst has an empty |
1076 | // domain. In the first and last case the instruction is dead but if |
1077 | // there is a statement or the domain is not empty Inst is not dead. |
1078 | auto MemInst = MemAccInst::dyn_cast(Inst); |
1079 | auto Address = MemInst ? MemInst.getPointerOperand() : nullptr; |
1080 | if (Address && SE.getUnknown(UndefValue::get(Address->getType())) == |
1081 | SE.getPointerBase(SE.getSCEV(Address))) { |
1082 | } else if (S.getStmtFor(Inst)) { |
1083 | IsDead = false; |
1084 | } else { |
1085 | auto *Domain = S.getDomainConditions(Inst->getParent()).release(); |
1086 | IsDead = isl_set_is_empty(Domain); |
1087 | isl_set_free(Domain); |
1088 | } |
1089 | |
1090 | if (IsDead) { |
1091 | V = UndefValue::get(ParamSCEV->getType()); |
1092 | break; |
1093 | } |
1094 | } |
1095 | } |
1096 | |
1097 | if (auto *IAClass = S.lookupInvariantEquivClass(Val)) { |
1098 | // Check if this invariant access class is empty, hence if we never |
1099 | // actually added a loads instruction to it. In that case it has no |
1100 | // (meaningful) users and we should not try to code generate it. |
1101 | if (IAClass->InvariantAccesses.empty()) |
1102 | V = UndefValue::get(ParamSCEV->getType()); |
1103 | |
1104 | if (!preloadInvariantEquivClass(*IAClass)) { |
1105 | isl_id_free(Id); |
1106 | return false; |
1107 | } |
1108 | } |
1109 | } |
1110 | |
1111 | V = V ? V : generateSCEV(ParamSCEV); |
1112 | IDToValue[Id] = V; |
1113 | } |
1114 | |
1115 | isl_id_free(Id); |
1116 | return true; |
1117 | } |
1118 | |
1119 | bool IslNodeBuilder::materializeParameters(isl_set *Set) { |
1120 | for (unsigned i = 0, e = isl_set_dim(Set, isl_dim_param); i < e; ++i) { |
1121 | if (!isl_set_involves_dims(Set, isl_dim_param, i, 1)) |
1122 | continue; |
1123 | isl_id *Id = isl_set_get_dim_id(Set, isl_dim_param, i); |
1124 | if (!materializeValue(Id)) |
1125 | return false; |
1126 | } |
1127 | return true; |
1128 | } |
1129 | |
1130 | bool IslNodeBuilder::materializeParameters() { |
1131 | for (const SCEV *Param : S.parameters()) { |
1132 | isl_id *Id = S.getIdForParam(Param).release(); |
1133 | if (!materializeValue(Id)) |
1134 | return false; |
1135 | } |
1136 | return true; |
1137 | } |
1138 | |
1139 | /// Generate the computation of the size of the outermost dimension from the |
1140 | /// Fortran array descriptor (in this case, `@g_arr`). The final `%size` |
1141 | /// contains the size of the array. |
1142 | /// |
1143 | /// %arrty = type { i8*, i64, i64, [3 x %desc.dimensionty] } |
1144 | /// %desc.dimensionty = type { i64, i64, i64 } |
1145 | /// @g_arr = global %arrty zeroinitializer, align 32 |
1146 | /// ... |
1147 | /// %0 = load i64, i64* getelementptr inbounds |
1148 | /// (%arrty, %arrty* @g_arr, i64 0, i32 3, i64 0, i32 2) |
1149 | /// %1 = load i64, i64* getelementptr inbounds |
1150 | /// (%arrty, %arrty* @g_arr, i64 0, i32 3, i64 0, i32 1) |
1151 | /// %2 = sub nsw i64 %0, %1 |
1152 | /// %size = add nsw i64 %2, 1 |
1153 | static Value *buildFADOutermostDimensionLoad(Value *GlobalDescriptor, |
1154 | PollyIRBuilder &Builder, |
1155 | std::string ArrayName) { |
1156 | assert(GlobalDescriptor && "invalid global descriptor given")((GlobalDescriptor && "invalid global descriptor given" ) ? static_cast<void> (0) : __assert_fail ("GlobalDescriptor && \"invalid global descriptor given\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1156, __PRETTY_FUNCTION__)); |
1157 | |
1158 | Value *endIdx[4] = {Builder.getInt64(0), Builder.getInt32(3), |
1159 | Builder.getInt64(0), Builder.getInt32(2)}; |
1160 | Value *endPtr = Builder.CreateInBoundsGEP(GlobalDescriptor, endIdx, |
1161 | ArrayName + "_end_ptr"); |
1162 | Type *type = cast<GEPOperator>(endPtr)->getResultElementType(); |
1163 | assert(isa<IntegerType>(type) && "expected type of end to be integral")((isa<IntegerType>(type) && "expected type of end to be integral" ) ? static_cast<void> (0) : __assert_fail ("isa<IntegerType>(type) && \"expected type of end to be integral\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1163, __PRETTY_FUNCTION__)); |
1164 | |
1165 | Value *end = Builder.CreateLoad(type, endPtr, ArrayName + "_end"); |
1166 | |
1167 | Value *beginIdx[4] = {Builder.getInt64(0), Builder.getInt32(3), |
1168 | Builder.getInt64(0), Builder.getInt32(1)}; |
1169 | Value *beginPtr = Builder.CreateInBoundsGEP(GlobalDescriptor, beginIdx, |
1170 | ArrayName + "_begin_ptr"); |
1171 | Value *begin = Builder.CreateLoad(type, beginPtr, ArrayName + "_begin"); |
1172 | |
1173 | Value *size = |
Value stored to 'size' during its initialization is never read | |
1174 | Builder.CreateNSWSub(end, begin, ArrayName + "_end_begin_delta"); |
1175 | |
1176 | size = Builder.CreateNSWAdd( |
1177 | end, ConstantInt::get(type, 1, /* signed = */ true), ArrayName + "_size"); |
1178 | |
1179 | return size; |
1180 | } |
1181 | |
1182 | bool IslNodeBuilder::materializeFortranArrayOutermostDimension() { |
1183 | for (ScopArrayInfo *Array : S.arrays()) { |
1184 | if (Array->getNumberOfDimensions() == 0) |
1185 | continue; |
1186 | |
1187 | Value *FAD = Array->getFortranArrayDescriptor(); |
1188 | if (!FAD) |
1189 | continue; |
1190 | |
1191 | isl_pw_aff *ParametricPwAff = Array->getDimensionSizePw(0).release(); |
1192 | assert(ParametricPwAff && "parametric pw_aff corresponding "((ParametricPwAff && "parametric pw_aff corresponding " "to outermost dimension does not " "exist") ? static_cast< void> (0) : __assert_fail ("ParametricPwAff && \"parametric pw_aff corresponding \" \"to outermost dimension does not \" \"exist\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1194, __PRETTY_FUNCTION__)) |
1193 | "to outermost dimension does not "((ParametricPwAff && "parametric pw_aff corresponding " "to outermost dimension does not " "exist") ? static_cast< void> (0) : __assert_fail ("ParametricPwAff && \"parametric pw_aff corresponding \" \"to outermost dimension does not \" \"exist\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1194, __PRETTY_FUNCTION__)) |
1194 | "exist")((ParametricPwAff && "parametric pw_aff corresponding " "to outermost dimension does not " "exist") ? static_cast< void> (0) : __assert_fail ("ParametricPwAff && \"parametric pw_aff corresponding \" \"to outermost dimension does not \" \"exist\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1194, __PRETTY_FUNCTION__)); |
1195 | |
1196 | isl_id *Id = isl_pw_aff_get_dim_id(ParametricPwAff, isl_dim_param, 0); |
1197 | isl_pw_aff_free(ParametricPwAff); |
1198 | |
1199 | assert(Id && "pw_aff is not parametric")((Id && "pw_aff is not parametric") ? static_cast< void> (0) : __assert_fail ("Id && \"pw_aff is not parametric\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1199, __PRETTY_FUNCTION__)); |
1200 | |
1201 | if (IDToValue.count(Id)) { |
1202 | isl_id_free(Id); |
1203 | continue; |
1204 | } |
1205 | |
1206 | Value *FinalValue = |
1207 | buildFADOutermostDimensionLoad(FAD, Builder, Array->getName()); |
1208 | assert(FinalValue && "unable to build Fortran array "((FinalValue && "unable to build Fortran array " "descriptor load of outermost dimension" ) ? static_cast<void> (0) : __assert_fail ("FinalValue && \"unable to build Fortran array \" \"descriptor load of outermost dimension\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1209, __PRETTY_FUNCTION__)) |
1209 | "descriptor load of outermost dimension")((FinalValue && "unable to build Fortran array " "descriptor load of outermost dimension" ) ? static_cast<void> (0) : __assert_fail ("FinalValue && \"unable to build Fortran array \" \"descriptor load of outermost dimension\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1209, __PRETTY_FUNCTION__)); |
1210 | IDToValue[Id] = FinalValue; |
1211 | isl_id_free(Id); |
1212 | } |
1213 | return true; |
1214 | } |
1215 | |
1216 | Value *IslNodeBuilder::preloadUnconditionally(isl_set *AccessRange, |
1217 | isl_ast_build *Build, |
1218 | Instruction *AccInst) { |
1219 | isl_pw_multi_aff *PWAccRel = isl_pw_multi_aff_from_set(AccessRange); |
1220 | isl_ast_expr *Access = |
1221 | isl_ast_build_access_from_pw_multi_aff(Build, PWAccRel); |
1222 | auto *Address = isl_ast_expr_address_of(Access); |
1223 | auto *AddressValue = ExprBuilder.create(Address); |
1224 | Value *PreloadVal; |
1225 | |
1226 | // Correct the type as the SAI might have a different type than the user |
1227 | // expects, especially if the base pointer is a struct. |
1228 | Type *Ty = AccInst->getType(); |
1229 | |
1230 | auto *Ptr = AddressValue; |
1231 | auto Name = Ptr->getName(); |
1232 | auto AS = Ptr->getType()->getPointerAddressSpace(); |
1233 | Ptr = Builder.CreatePointerCast(Ptr, Ty->getPointerTo(AS), Name + ".cast"); |
1234 | PreloadVal = Builder.CreateLoad(Ty, Ptr, Name + ".load"); |
1235 | if (LoadInst *PreloadInst = dyn_cast<LoadInst>(PreloadVal)) |
1236 | PreloadInst->setAlignment(cast<LoadInst>(AccInst)->getAlign()); |
1237 | |
1238 | // TODO: This is only a hot fix for SCoP sequences that use the same load |
1239 | // instruction contained and hoisted by one of the SCoPs. |
1240 | if (SE.isSCEVable(Ty)) |
1241 | SE.forgetValue(AccInst); |
1242 | |
1243 | return PreloadVal; |
1244 | } |
1245 | |
1246 | Value *IslNodeBuilder::preloadInvariantLoad(const MemoryAccess &MA, |
1247 | isl_set *Domain) { |
1248 | isl_set *AccessRange = isl_map_range(MA.getAddressFunction().release()); |
1249 | AccessRange = isl_set_gist_params(AccessRange, S.getContext().release()); |
1250 | |
1251 | if (!materializeParameters(AccessRange)) { |
1252 | isl_set_free(AccessRange); |
1253 | isl_set_free(Domain); |
1254 | return nullptr; |
1255 | } |
1256 | |
1257 | auto *Build = |
1258 | isl_ast_build_from_context(isl_set_universe(S.getParamSpace().release())); |
1259 | isl_set *Universe = isl_set_universe(isl_set_get_space(Domain)); |
1260 | bool AlwaysExecuted = isl_set_is_equal(Domain, Universe); |
1261 | isl_set_free(Universe); |
1262 | |
1263 | Instruction *AccInst = MA.getAccessInstruction(); |
1264 | Type *AccInstTy = AccInst->getType(); |
1265 | |
1266 | Value *PreloadVal = nullptr; |
1267 | if (AlwaysExecuted) { |
1268 | PreloadVal = preloadUnconditionally(AccessRange, Build, AccInst); |
1269 | isl_ast_build_free(Build); |
1270 | isl_set_free(Domain); |
1271 | return PreloadVal; |
1272 | } |
1273 | |
1274 | if (!materializeParameters(Domain)) { |
1275 | isl_ast_build_free(Build); |
1276 | isl_set_free(AccessRange); |
1277 | isl_set_free(Domain); |
1278 | return nullptr; |
1279 | } |
1280 | |
1281 | isl_ast_expr *DomainCond = isl_ast_build_expr_from_set(Build, Domain); |
1282 | Domain = nullptr; |
1283 | |
1284 | ExprBuilder.setTrackOverflow(true); |
1285 | Value *Cond = ExprBuilder.create(DomainCond); |
1286 | Value *OverflowHappened = Builder.CreateNot(ExprBuilder.getOverflowState(), |
1287 | "polly.preload.cond.overflown"); |
1288 | Cond = Builder.CreateAnd(Cond, OverflowHappened, "polly.preload.cond.result"); |
1289 | ExprBuilder.setTrackOverflow(false); |
1290 | |
1291 | if (!Cond->getType()->isIntegerTy(1)) |
1292 | Cond = Builder.CreateIsNotNull(Cond); |
1293 | |
1294 | BasicBlock *CondBB = SplitBlock(Builder.GetInsertBlock(), |
1295 | &*Builder.GetInsertPoint(), &DT, &LI); |
1296 | CondBB->setName("polly.preload.cond"); |
1297 | |
1298 | BasicBlock *MergeBB = SplitBlock(CondBB, &CondBB->front(), &DT, &LI); |
1299 | MergeBB->setName("polly.preload.merge"); |
1300 | |
1301 | Function *F = Builder.GetInsertBlock()->getParent(); |
1302 | LLVMContext &Context = F->getContext(); |
1303 | BasicBlock *ExecBB = BasicBlock::Create(Context, "polly.preload.exec", F); |
1304 | |
1305 | DT.addNewBlock(ExecBB, CondBB); |
1306 | if (Loop *L = LI.getLoopFor(CondBB)) |
1307 | L->addBasicBlockToLoop(ExecBB, LI); |
1308 | |
1309 | auto *CondBBTerminator = CondBB->getTerminator(); |
1310 | Builder.SetInsertPoint(CondBBTerminator); |
1311 | Builder.CreateCondBr(Cond, ExecBB, MergeBB); |
1312 | CondBBTerminator->eraseFromParent(); |
1313 | |
1314 | Builder.SetInsertPoint(ExecBB); |
1315 | Builder.CreateBr(MergeBB); |
1316 | |
1317 | Builder.SetInsertPoint(ExecBB->getTerminator()); |
1318 | Value *PreAccInst = preloadUnconditionally(AccessRange, Build, AccInst); |
1319 | Builder.SetInsertPoint(MergeBB->getTerminator()); |
1320 | auto *MergePHI = Builder.CreatePHI( |
1321 | AccInstTy, 2, "polly.preload." + AccInst->getName() + ".merge"); |
1322 | PreloadVal = MergePHI; |
1323 | |
1324 | if (!PreAccInst) { |
1325 | PreloadVal = nullptr; |
1326 | PreAccInst = UndefValue::get(AccInstTy); |
1327 | } |
1328 | |
1329 | MergePHI->addIncoming(PreAccInst, ExecBB); |
1330 | MergePHI->addIncoming(Constant::getNullValue(AccInstTy), CondBB); |
1331 | |
1332 | isl_ast_build_free(Build); |
1333 | return PreloadVal; |
1334 | } |
1335 | |
1336 | bool IslNodeBuilder::preloadInvariantEquivClass( |
1337 | InvariantEquivClassTy &IAClass) { |
1338 | // For an equivalence class of invariant loads we pre-load the representing |
1339 | // element with the unified execution context. However, we have to map all |
1340 | // elements of the class to the one preloaded load as they are referenced |
1341 | // during the code generation and therefor need to be mapped. |
1342 | const MemoryAccessList &MAs = IAClass.InvariantAccesses; |
1343 | if (MAs.empty()) |
1344 | return true; |
1345 | |
1346 | MemoryAccess *MA = MAs.front(); |
1347 | assert(MA->isArrayKind() && MA->isRead())((MA->isArrayKind() && MA->isRead()) ? static_cast <void> (0) : __assert_fail ("MA->isArrayKind() && MA->isRead()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1347, __PRETTY_FUNCTION__)); |
1348 | |
1349 | // If the access function was already mapped, the preload of this equivalence |
1350 | // class was triggered earlier already and doesn't need to be done again. |
1351 | if (ValueMap.count(MA->getAccessInstruction())) |
1352 | return true; |
1353 | |
1354 | // Check for recursion which can be caused by additional constraints, e.g., |
1355 | // non-finite loop constraints. In such a case we have to bail out and insert |
1356 | // a "false" runtime check that will cause the original code to be executed. |
1357 | auto PtrId = std::make_pair(IAClass.IdentifyingPointer, IAClass.AccessType); |
1358 | if (!PreloadedPtrs.insert(PtrId).second) |
1359 | return false; |
1360 | |
1361 | // The execution context of the IAClass. |
1362 | isl::set &ExecutionCtx = IAClass.ExecutionContext; |
1363 | |
1364 | // If the base pointer of this class is dependent on another one we have to |
1365 | // make sure it was preloaded already. |
1366 | auto *SAI = MA->getScopArrayInfo(); |
1367 | if (auto *BaseIAClass = S.lookupInvariantEquivClass(SAI->getBasePtr())) { |
1368 | if (!preloadInvariantEquivClass(*BaseIAClass)) |
1369 | return false; |
1370 | |
1371 | // After we preloaded the BaseIAClass we adjusted the BaseExecutionCtx and |
1372 | // we need to refine the ExecutionCtx. |
1373 | isl::set BaseExecutionCtx = BaseIAClass->ExecutionContext; |
1374 | ExecutionCtx = ExecutionCtx.intersect(BaseExecutionCtx); |
1375 | } |
1376 | |
1377 | // If the size of a dimension is dependent on another class, make sure it is |
1378 | // preloaded. |
1379 | for (unsigned i = 1, e = SAI->getNumberOfDimensions(); i < e; ++i) { |
1380 | const SCEV *Dim = SAI->getDimensionSize(i); |
1381 | SetVector<Value *> Values; |
1382 | findValues(Dim, SE, Values); |
1383 | for (auto *Val : Values) { |
1384 | if (auto *BaseIAClass = S.lookupInvariantEquivClass(Val)) { |
1385 | if (!preloadInvariantEquivClass(*BaseIAClass)) |
1386 | return false; |
1387 | |
1388 | // After we preloaded the BaseIAClass we adjusted the BaseExecutionCtx |
1389 | // and we need to refine the ExecutionCtx. |
1390 | isl::set BaseExecutionCtx = BaseIAClass->ExecutionContext; |
1391 | ExecutionCtx = ExecutionCtx.intersect(BaseExecutionCtx); |
1392 | } |
1393 | } |
1394 | } |
1395 | |
1396 | Instruction *AccInst = MA->getAccessInstruction(); |
1397 | Type *AccInstTy = AccInst->getType(); |
1398 | |
1399 | Value *PreloadVal = preloadInvariantLoad(*MA, ExecutionCtx.copy()); |
1400 | if (!PreloadVal) |
1401 | return false; |
1402 | |
1403 | for (const MemoryAccess *MA : MAs) { |
1404 | Instruction *MAAccInst = MA->getAccessInstruction(); |
1405 | assert(PreloadVal->getType() == MAAccInst->getType())((PreloadVal->getType() == MAAccInst->getType()) ? static_cast <void> (0) : __assert_fail ("PreloadVal->getType() == MAAccInst->getType()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1405, __PRETTY_FUNCTION__)); |
1406 | ValueMap[MAAccInst] = PreloadVal; |
1407 | } |
1408 | |
1409 | if (SE.isSCEVable(AccInstTy)) { |
1410 | isl_id *ParamId = S.getIdForParam(SE.getSCEV(AccInst)).release(); |
1411 | if (ParamId) |
1412 | IDToValue[ParamId] = PreloadVal; |
1413 | isl_id_free(ParamId); |
1414 | } |
1415 | |
1416 | BasicBlock *EntryBB = &Builder.GetInsertBlock()->getParent()->getEntryBlock(); |
1417 | auto *Alloca = new AllocaInst(AccInstTy, DL.getAllocaAddrSpace(), |
1418 | AccInst->getName() + ".preload.s2a", |
1419 | &*EntryBB->getFirstInsertionPt()); |
1420 | Builder.CreateStore(PreloadVal, Alloca); |
1421 | ValueMapT PreloadedPointer; |
1422 | PreloadedPointer[PreloadVal] = AccInst; |
1423 | Annotator.addAlternativeAliasBases(PreloadedPointer); |
1424 | |
1425 | for (auto *DerivedSAI : SAI->getDerivedSAIs()) { |
1426 | Value *BasePtr = DerivedSAI->getBasePtr(); |
1427 | |
1428 | for (const MemoryAccess *MA : MAs) { |
1429 | // As the derived SAI information is quite coarse, any load from the |
1430 | // current SAI could be the base pointer of the derived SAI, however we |
1431 | // should only change the base pointer of the derived SAI if we actually |
1432 | // preloaded it. |
1433 | if (BasePtr == MA->getOriginalBaseAddr()) { |
1434 | assert(BasePtr->getType() == PreloadVal->getType())((BasePtr->getType() == PreloadVal->getType()) ? static_cast <void> (0) : __assert_fail ("BasePtr->getType() == PreloadVal->getType()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1434, __PRETTY_FUNCTION__)); |
1435 | DerivedSAI->setBasePtr(PreloadVal); |
1436 | } |
1437 | |
1438 | // For scalar derived SAIs we remap the alloca used for the derived value. |
1439 | if (BasePtr == MA->getAccessInstruction()) |
1440 | ScalarMap[DerivedSAI] = Alloca; |
1441 | } |
1442 | } |
1443 | |
1444 | for (const MemoryAccess *MA : MAs) { |
1445 | Instruction *MAAccInst = MA->getAccessInstruction(); |
1446 | // Use the escape system to get the correct value to users outside the SCoP. |
1447 | BlockGenerator::EscapeUserVectorTy EscapeUsers; |
1448 | for (auto *U : MAAccInst->users()) |
1449 | if (Instruction *UI = dyn_cast<Instruction>(U)) |
1450 | if (!S.contains(UI)) |
1451 | EscapeUsers.push_back(UI); |
1452 | |
1453 | if (EscapeUsers.empty()) |
1454 | continue; |
1455 | |
1456 | EscapeMap[MA->getAccessInstruction()] = |
1457 | std::make_pair(Alloca, std::move(EscapeUsers)); |
1458 | } |
1459 | |
1460 | return true; |
1461 | } |
1462 | |
1463 | void IslNodeBuilder::allocateNewArrays(BBPair StartExitBlocks) { |
1464 | for (auto &SAI : S.arrays()) { |
1465 | if (SAI->getBasePtr()) |
1466 | continue; |
1467 | |
1468 | assert(SAI->getNumberOfDimensions() > 0 && SAI->getDimensionSize(0) &&((SAI->getNumberOfDimensions() > 0 && SAI->getDimensionSize (0) && "The size of the outermost dimension is used to declare newly " "created arrays that require memory allocation.") ? static_cast <void> (0) : __assert_fail ("SAI->getNumberOfDimensions() > 0 && SAI->getDimensionSize(0) && \"The size of the outermost dimension is used to declare newly \" \"created arrays that require memory allocation.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1470, __PRETTY_FUNCTION__)) |
1469 | "The size of the outermost dimension is used to declare newly "((SAI->getNumberOfDimensions() > 0 && SAI->getDimensionSize (0) && "The size of the outermost dimension is used to declare newly " "created arrays that require memory allocation.") ? static_cast <void> (0) : __assert_fail ("SAI->getNumberOfDimensions() > 0 && SAI->getDimensionSize(0) && \"The size of the outermost dimension is used to declare newly \" \"created arrays that require memory allocation.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1470, __PRETTY_FUNCTION__)) |
1470 | "created arrays that require memory allocation.")((SAI->getNumberOfDimensions() > 0 && SAI->getDimensionSize (0) && "The size of the outermost dimension is used to declare newly " "created arrays that require memory allocation.") ? static_cast <void> (0) : __assert_fail ("SAI->getNumberOfDimensions() > 0 && SAI->getDimensionSize(0) && \"The size of the outermost dimension is used to declare newly \" \"created arrays that require memory allocation.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1470, __PRETTY_FUNCTION__)); |
1471 | |
1472 | Type *NewArrayType = nullptr; |
1473 | |
1474 | // Get the size of the array = size(dim_1)*...*size(dim_n) |
1475 | uint64_t ArraySizeInt = 1; |
1476 | for (int i = SAI->getNumberOfDimensions() - 1; i >= 0; i--) { |
1477 | auto *DimSize = SAI->getDimensionSize(i); |
1478 | unsigned UnsignedDimSize = static_cast<const SCEVConstant *>(DimSize) |
1479 | ->getAPInt() |
1480 | .getLimitedValue(); |
1481 | |
1482 | if (!NewArrayType) |
1483 | NewArrayType = SAI->getElementType(); |
1484 | |
1485 | NewArrayType = ArrayType::get(NewArrayType, UnsignedDimSize); |
1486 | ArraySizeInt *= UnsignedDimSize; |
1487 | } |
1488 | |
1489 | if (SAI->isOnHeap()) { |
1490 | LLVMContext &Ctx = NewArrayType->getContext(); |
1491 | |
1492 | // Get the IntPtrTy from the Datalayout |
1493 | auto IntPtrTy = DL.getIntPtrType(Ctx); |
1494 | |
1495 | // Get the size of the element type in bits |
1496 | unsigned Size = SAI->getElemSizeInBytes(); |
1497 | |
1498 | // Insert the malloc call at polly.start |
1499 | auto InstIt = std::get<0>(StartExitBlocks)->getTerminator(); |
1500 | auto *CreatedArray = CallInst::CreateMalloc( |
1501 | &*InstIt, IntPtrTy, SAI->getElementType(), |
1502 | ConstantInt::get(Type::getInt64Ty(Ctx), Size), |
1503 | ConstantInt::get(Type::getInt64Ty(Ctx), ArraySizeInt), nullptr, |
1504 | SAI->getName()); |
1505 | |
1506 | SAI->setBasePtr(CreatedArray); |
1507 | |
1508 | // Insert the free call at polly.exiting |
1509 | CallInst::CreateFree(CreatedArray, |
1510 | std::get<1>(StartExitBlocks)->getTerminator()); |
1511 | } else { |
1512 | auto InstIt = Builder.GetInsertBlock() |
1513 | ->getParent() |
1514 | ->getEntryBlock() |
1515 | .getTerminator(); |
1516 | |
1517 | auto *CreatedArray = new AllocaInst(NewArrayType, DL.getAllocaAddrSpace(), |
1518 | SAI->getName(), &*InstIt); |
1519 | if (PollyTargetFirstLevelCacheLineSize) |
1520 | CreatedArray->setAlignment(Align(PollyTargetFirstLevelCacheLineSize)); |
1521 | SAI->setBasePtr(CreatedArray); |
1522 | } |
1523 | } |
1524 | } |
1525 | |
1526 | bool IslNodeBuilder::preloadInvariantLoads() { |
1527 | auto &InvariantEquivClasses = S.getInvariantAccesses(); |
1528 | if (InvariantEquivClasses.empty()) |
1529 | return true; |
1530 | |
1531 | BasicBlock *PreLoadBB = SplitBlock(Builder.GetInsertBlock(), |
1532 | &*Builder.GetInsertPoint(), &DT, &LI); |
1533 | PreLoadBB->setName("polly.preload.begin"); |
1534 | Builder.SetInsertPoint(&PreLoadBB->front()); |
1535 | |
1536 | for (auto &IAClass : InvariantEquivClasses) |
1537 | if (!preloadInvariantEquivClass(IAClass)) |
1538 | return false; |
1539 | |
1540 | return true; |
1541 | } |
1542 | |
1543 | void IslNodeBuilder::addParameters(__isl_take isl_set *Context) { |
1544 | // Materialize values for the parameters of the SCoP. |
1545 | materializeParameters(); |
1546 | |
1547 | // materialize the outermost dimension parameters for a Fortran array. |
1548 | // NOTE: materializeParameters() does not work since it looks through |
1549 | // the SCEVs. We don't have a corresponding SCEV for the array size |
1550 | // parameter |
1551 | materializeFortranArrayOutermostDimension(); |
1552 | |
1553 | // Generate values for the current loop iteration for all surrounding loops. |
1554 | // |
1555 | // We may also reference loops outside of the scop which do not contain the |
1556 | // scop itself, but as the number of such scops may be arbitrarily large we do |
1557 | // not generate code for them here, but only at the point of code generation |
1558 | // where these values are needed. |
1559 | Loop *L = LI.getLoopFor(S.getEntry()); |
1560 | |
1561 | while (L != nullptr && S.contains(L)) |
1562 | L = L->getParentLoop(); |
1563 | |
1564 | while (L != nullptr) { |
1565 | materializeNonScopLoopInductionVariable(L); |
1566 | L = L->getParentLoop(); |
1567 | } |
1568 | |
1569 | isl_set_free(Context); |
1570 | } |
1571 | |
1572 | Value *IslNodeBuilder::generateSCEV(const SCEV *Expr) { |
1573 | /// We pass the insert location of our Builder, as Polly ensures during IR |
1574 | /// generation that there is always a valid CFG into which instructions are |
1575 | /// inserted. As a result, the insertpoint is known to be always followed by a |
1576 | /// terminator instruction. This means the insert point may be specified by a |
1577 | /// terminator instruction, but it can never point to an ->end() iterator |
1578 | /// which does not have a corresponding instruction. Hence, dereferencing |
1579 | /// the insertpoint to obtain an instruction is known to be save. |
1580 | /// |
1581 | /// We also do not need to update the Builder here, as new instructions are |
1582 | /// always inserted _before_ the given InsertLocation. As a result, the |
1583 | /// insert location remains valid. |
1584 | assert(Builder.GetInsertBlock()->end() != Builder.GetInsertPoint() &&((Builder.GetInsertBlock()->end() != Builder.GetInsertPoint () && "Insert location points after last valid instruction" ) ? static_cast<void> (0) : __assert_fail ("Builder.GetInsertBlock()->end() != Builder.GetInsertPoint() && \"Insert location points after last valid instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1585, __PRETTY_FUNCTION__)) |
1585 | "Insert location points after last valid instruction")((Builder.GetInsertBlock()->end() != Builder.GetInsertPoint () && "Insert location points after last valid instruction" ) ? static_cast<void> (0) : __assert_fail ("Builder.GetInsertBlock()->end() != Builder.GetInsertPoint() && \"Insert location points after last valid instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/polly/lib/CodeGen/IslNodeBuilder.cpp" , 1585, __PRETTY_FUNCTION__)); |
1586 | Instruction *InsertLocation = &*Builder.GetInsertPoint(); |
1587 | return expandCodeFor(S, SE, DL, "polly", Expr, Expr->getType(), |
1588 | InsertLocation, &ValueMap, |
1589 | StartBlock->getSinglePredecessor()); |
1590 | } |
1591 | |
1592 | /// The AST expression we generate to perform the run-time check assumes |
1593 | /// computations on integer types of infinite size. As we only use 64-bit |
1594 | /// arithmetic we check for overflows, in case of which we set the result |
1595 | /// of this run-time check to false to be conservatively correct, |
1596 | Value *IslNodeBuilder::createRTC(isl_ast_expr *Condition) { |
1597 | auto ExprBuilder = getExprBuilder(); |
1598 | |
1599 | // In case the AST expression has integers larger than 64 bit, bail out. The |
1600 | // resulting LLVM-IR will contain operations on types that use more than 64 |
1601 | // bits. These are -- in case wrapping intrinsics are used -- translated to |
1602 | // runtime library calls that are not available on all systems (e.g., Android) |
1603 | // and consequently will result in linker errors. |
1604 | if (ExprBuilder.hasLargeInts(isl::manage_copy(Condition))) { |
1605 | isl_ast_expr_free(Condition); |
1606 | return Builder.getFalse(); |
1607 | } |
1608 | |
1609 | ExprBuilder.setTrackOverflow(true); |
1610 | Value *RTC = ExprBuilder.create(Condition); |
1611 | if (!RTC->getType()->isIntegerTy(1)) |
1612 | RTC = Builder.CreateIsNotNull(RTC); |
1613 | Value *OverflowHappened = |
1614 | Builder.CreateNot(ExprBuilder.getOverflowState(), "polly.rtc.overflown"); |
1615 | |
1616 | if (PollyGenerateRTCPrint) { |
1617 | auto *F = Builder.GetInsertBlock()->getParent(); |
1618 | RuntimeDebugBuilder::createCPUPrinter( |
1619 | Builder, |
1620 | "F: " + F->getName().str() + " R: " + S.getRegion().getNameStr() + |
1621 | "RTC: ", |
1622 | RTC, " Overflow: ", OverflowHappened, |
1623 | "\n" |
1624 | " (0 failed, -1 succeeded)\n" |
1625 | " (if one or both are 0 falling back to original code, if both are -1 " |
1626 | "executing Polly code)\n"); |
1627 | } |
1628 | |
1629 | RTC = Builder.CreateAnd(RTC, OverflowHappened, "polly.rtc.result"); |
1630 | ExprBuilder.setTrackOverflow(false); |
1631 | |
1632 | if (!isa<ConstantInt>(RTC)) |
1633 | VersionedScops++; |
1634 | |
1635 | return RTC; |
1636 | } |