| 1 | //===- RDFLiveness.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 | // Computation of the liveness information from the data-flow graph. |
| 10 | // |
| 11 | // The main functionality of this code is to compute block live-in |
| 12 | // information. With the live-in information in place, the placement |
| 13 | // of kill flags can also be recalculated. |
| 14 | // |
| 15 | // The block live-in calculation is based on the ideas from the following |
| 16 | // publication: |
| 17 | // |
| 18 | // Dibyendu Das, Ramakrishna Upadrasta, Benoit Dupont de Dinechin. |
| 19 | // "Efficient Liveness Computation Using Merge Sets and DJ-Graphs." |
| 20 | // ACM Transactions on Architecture and Code Optimization, Association for |
| 21 | // Computing Machinery, 2012, ACM TACO Special Issue on "High-Performance |
| 22 | // and Embedded Architectures and Compilers", 8 (4), |
| 23 | // <10.1145/2086696.2086706>. <hal-00647369> |
| 24 | // |
| 25 | #include "llvm/ADT/BitVector.h" |
| 26 | #include "llvm/ADT/DenseMap.h" |
| 27 | #include "llvm/ADT/STLExtras.h" |
| 28 | #include "llvm/ADT/SetVector.h" |
| 29 | #include "llvm/ADT/SmallSet.h" |
| 30 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 31 | #include "llvm/CodeGen/MachineDominanceFrontier.h" |
| 32 | #include "llvm/CodeGen/MachineDominators.h" |
| 33 | #include "llvm/CodeGen/MachineFunction.h" |
| 34 | #include "llvm/CodeGen/MachineInstr.h" |
| 35 | #include "llvm/CodeGen/RDFGraph.h" |
| 36 | #include "llvm/CodeGen/RDFLiveness.h" |
| 37 | #include "llvm/CodeGen/RDFRegisters.h" |
| 38 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 39 | #include "llvm/MC/LaneBitmask.h" |
| 40 | #include "llvm/MC/MCRegisterInfo.h" |
| 41 | #include "llvm/Support/CommandLine.h" |
| 42 | #include "llvm/Support/ErrorHandling.h" |
| 43 | #include "llvm/Support/raw_ostream.h" |
| 44 | #include <algorithm> |
| 45 | #include <cassert> |
| 46 | #include <cstdint> |
| 47 | #include <iterator> |
| 48 | #include <map> |
| 49 | #include <unordered_map> |
| 50 | #include <utility> |
| 51 | #include <vector> |
| 52 | |
| 53 | using namespace llvm; |
| 54 | |
| 55 | static cl::opt<unsigned> MaxRecNest("rdf-liveness-max-rec" , cl::init(Val: 25), |
| 56 | cl::Hidden, |
| 57 | cl::desc("Maximum recursion level" )); |
| 58 | |
| 59 | namespace llvm::rdf { |
| 60 | |
| 61 | raw_ostream &operator<<(raw_ostream &OS, const Print<Liveness::RefMap> &P) { |
| 62 | OS << '{'; |
| 63 | for (const auto &I : P.Obj) { |
| 64 | OS << ' ' << printReg(Reg: I.first, TRI: &P.G.getTRI()) << '{'; |
| 65 | for (auto J = I.second.begin(), E = I.second.end(); J != E;) { |
| 66 | OS << Print(J->first, P.G) << PrintLaneMaskShort(J->second); |
| 67 | if (++J != E) |
| 68 | OS << ','; |
| 69 | } |
| 70 | OS << '}'; |
| 71 | } |
| 72 | OS << " }" ; |
| 73 | return OS; |
| 74 | } |
| 75 | |
| 76 | // The order in the returned sequence is the order of reaching defs in the |
| 77 | // upward traversal: the first def is the closest to the given reference RefA, |
| 78 | // the next one is further up, and so on. |
| 79 | // The list ends at a reaching phi def, or when the reference from RefA is |
| 80 | // covered by the defs in the list (see FullChain). |
| 81 | // This function provides two modes of operation: |
| 82 | // (1) Returning the sequence of reaching defs for a particular reference |
| 83 | // node. This sequence will terminate at the first phi node [1]. |
| 84 | // (2) Returning a partial sequence of reaching defs, where the final goal |
| 85 | // is to traverse past phi nodes to the actual defs arising from the code |
| 86 | // itself. |
| 87 | // In mode (2), the register reference for which the search was started |
| 88 | // may be different from the reference node RefA, for which this call was |
| 89 | // made, hence the argument RefRR, which holds the original register. |
| 90 | // Also, some definitions may have already been encountered in a previous |
| 91 | // call that will influence register covering. The register references |
| 92 | // already defined are passed in through DefRRs. |
| 93 | // In mode (1), the "continuation" considerations do not apply, and the |
| 94 | // RefRR is the same as the register in RefA, and the set DefRRs is empty. |
| 95 | // |
| 96 | // [1] It is possible for multiple phi nodes to be included in the returned |
| 97 | // sequence: |
| 98 | // SubA = phi ... |
| 99 | // SubB = phi ... |
| 100 | // ... = SuperAB(rdef:SubA), SuperAB"(rdef:SubB) |
| 101 | // However, these phi nodes are independent from one another in terms of |
| 102 | // the data-flow. |
| 103 | |
| 104 | NodeList Liveness::getAllReachingDefs(RegisterRef RefRR, |
| 105 | NodeAddr<RefNode *> RefA, bool TopShadows, |
| 106 | bool FullChain, |
| 107 | const RegisterAggr &DefRRs) { |
| 108 | NodeList RDefs; // Return value. |
| 109 | SetVector<NodeId> DefQ; |
| 110 | DenseMap<MachineInstr *, uint32_t> OrdMap; |
| 111 | |
| 112 | // Dead defs will be treated as if they were live, since they are actually |
| 113 | // on the data-flow path. They cannot be ignored because even though they |
| 114 | // do not generate meaningful values, they still modify registers. |
| 115 | |
| 116 | // If the reference is undefined, there is nothing to do. |
| 117 | if (RefA.Addr->getFlags() & NodeAttrs::Undef) |
| 118 | return RDefs; |
| 119 | |
| 120 | // The initial queue should not have reaching defs for shadows. The |
| 121 | // whole point of a shadow is that it will have a reaching def that |
| 122 | // is not aliased to the reaching defs of the related shadows. |
| 123 | NodeId Start = RefA.Id; |
| 124 | auto SNA = DFG.addr<RefNode *>(N: Start); |
| 125 | if (NodeId RD = SNA.Addr->getReachingDef()) |
| 126 | DefQ.insert(X: RD); |
| 127 | if (TopShadows) { |
| 128 | for (auto S : DFG.getRelatedRefs(IA: RefA.Addr->getOwner(G: DFG), RA: RefA)) |
| 129 | if (NodeId RD = NodeAddr<RefNode *>(S).Addr->getReachingDef()) |
| 130 | DefQ.insert(X: RD); |
| 131 | } |
| 132 | |
| 133 | // Collect all the reaching defs, going up until a phi node is encountered, |
| 134 | // or there are no more reaching defs. From this set, the actual set of |
| 135 | // reaching defs will be selected. |
| 136 | // The traversal upwards must go on until a covering def is encountered. |
| 137 | // It is possible that a collection of non-covering (individually) defs |
| 138 | // will be sufficient, but keep going until a covering one is found. |
| 139 | for (unsigned i = 0; i < DefQ.size(); ++i) { |
| 140 | auto TA = DFG.addr<DefNode *>(N: DefQ[i]); |
| 141 | if (TA.Addr->getFlags() & NodeAttrs::PhiRef) |
| 142 | continue; |
| 143 | // Stop at the covering/overwriting def of the initial register reference. |
| 144 | RegisterRef RR = TA.Addr->getRegRef(G: DFG); |
| 145 | if (!DFG.IsPreservingDef(DA: TA)) |
| 146 | if (RegisterAggr::isCoverOf(RA: RR, RB: RefRR, PRI)) |
| 147 | continue; |
| 148 | // Get the next level of reaching defs. This will include multiple |
| 149 | // reaching defs for shadows. |
| 150 | for (auto S : DFG.getRelatedRefs(IA: TA.Addr->getOwner(G: DFG), RA: TA)) |
| 151 | if (NodeId RD = NodeAddr<RefNode *>(S).Addr->getReachingDef()) |
| 152 | DefQ.insert(X: RD); |
| 153 | // Don't visit sibling defs. They share the same reaching def (which |
| 154 | // will be visited anyway), but they define something not aliased to |
| 155 | // this ref. |
| 156 | } |
| 157 | |
| 158 | // Return the MachineBasicBlock containing a given instruction. |
| 159 | auto Block = [this](NodeAddr<InstrNode *> IA) -> MachineBasicBlock * { |
| 160 | if (IA.Addr->getKind() == NodeAttrs::Stmt) |
| 161 | return NodeAddr<StmtNode *>(IA).Addr->getCode()->getParent(); |
| 162 | assert(IA.Addr->getKind() == NodeAttrs::Phi); |
| 163 | NodeAddr<PhiNode *> PA = IA; |
| 164 | NodeAddr<BlockNode *> BA = PA.Addr->getOwner(G: DFG); |
| 165 | return BA.Addr->getCode(); |
| 166 | }; |
| 167 | |
| 168 | SmallSet<NodeId, 32> Defs; |
| 169 | |
| 170 | // Remove all non-phi defs that are not aliased to RefRR, and separate |
| 171 | // the the remaining defs into buckets for containing blocks. |
| 172 | std::map<NodeId, NodeAddr<InstrNode *>> Owners; |
| 173 | std::map<MachineBasicBlock *, SmallVector<NodeId, 32>> Blocks; |
| 174 | for (NodeId N : DefQ) { |
| 175 | auto TA = DFG.addr<DefNode *>(N); |
| 176 | bool IsPhi = TA.Addr->getFlags() & NodeAttrs::PhiRef; |
| 177 | if (!IsPhi && !PRI.alias(RA: RefRR, RB: TA.Addr->getRegRef(G: DFG))) |
| 178 | continue; |
| 179 | Defs.insert(V: TA.Id); |
| 180 | NodeAddr<InstrNode *> IA = TA.Addr->getOwner(G: DFG); |
| 181 | Owners[TA.Id] = IA; |
| 182 | Blocks[Block(IA)].push_back(Elt: IA.Id); |
| 183 | } |
| 184 | |
| 185 | auto Precedes = [this, &OrdMap](NodeId A, NodeId B) { |
| 186 | if (A == B) |
| 187 | return false; |
| 188 | NodeAddr<InstrNode *> OA = DFG.addr<InstrNode *>(N: A); |
| 189 | NodeAddr<InstrNode *> OB = DFG.addr<InstrNode *>(N: B); |
| 190 | bool StmtA = OA.Addr->getKind() == NodeAttrs::Stmt; |
| 191 | bool StmtB = OB.Addr->getKind() == NodeAttrs::Stmt; |
| 192 | if (StmtA && StmtB) { |
| 193 | const MachineInstr *InA = NodeAddr<StmtNode *>(OA).Addr->getCode(); |
| 194 | const MachineInstr *InB = NodeAddr<StmtNode *>(OB).Addr->getCode(); |
| 195 | assert(InA->getParent() == InB->getParent()); |
| 196 | auto FA = OrdMap.find(Val: InA); |
| 197 | if (FA != OrdMap.end()) |
| 198 | return FA->second < OrdMap.find(Val: InB)->second; |
| 199 | const MachineBasicBlock *BB = InA->getParent(); |
| 200 | for (auto It = BB->begin(), E = BB->end(); It != E; ++It) { |
| 201 | if (It == InA->getIterator()) |
| 202 | return true; |
| 203 | if (It == InB->getIterator()) |
| 204 | return false; |
| 205 | } |
| 206 | llvm_unreachable("InA and InB should be in the same block" ); |
| 207 | } |
| 208 | // One of them is a phi node. |
| 209 | if (!StmtA && !StmtB) { |
| 210 | // Both are phis, which are unordered. Break the tie by id numbers. |
| 211 | return A < B; |
| 212 | } |
| 213 | // Only one of them is a phi. Phis always precede statements. |
| 214 | return !StmtA; |
| 215 | }; |
| 216 | |
| 217 | auto GetOrder = [&OrdMap](MachineBasicBlock &B) { |
| 218 | uint32_t Pos = 0; |
| 219 | for (MachineInstr &In : B) |
| 220 | OrdMap.insert(KV: {&In, ++Pos}); |
| 221 | }; |
| 222 | |
| 223 | // For each block, sort the nodes in it. |
| 224 | std::vector<MachineBasicBlock *> TmpBB; |
| 225 | for (auto &Bucket : Blocks) { |
| 226 | TmpBB.push_back(x: Bucket.first); |
| 227 | if (Bucket.second.size() > 2) |
| 228 | GetOrder(*Bucket.first); |
| 229 | llvm::sort(C&: Bucket.second, Comp: Precedes); |
| 230 | } |
| 231 | |
| 232 | // Sort the blocks with respect to dominance. |
| 233 | llvm::sort(C&: TmpBB, |
| 234 | Comp: [this](auto A, auto B) { return MDT.properlyDominates(A, B); }); |
| 235 | |
| 236 | std::vector<NodeId> TmpInst; |
| 237 | for (MachineBasicBlock *MBB : llvm::reverse(C&: TmpBB)) { |
| 238 | auto &Bucket = Blocks[MBB]; |
| 239 | TmpInst.insert(position: TmpInst.end(), first: Bucket.rbegin(), last: Bucket.rend()); |
| 240 | } |
| 241 | |
| 242 | // The vector is a list of instructions, so that defs coming from |
| 243 | // the same instruction don't need to be artificially ordered. |
| 244 | // Then, when computing the initial segment, and iterating over an |
| 245 | // instruction, pick the defs that contribute to the covering (i.e. is |
| 246 | // not covered by previously added defs). Check the defs individually, |
| 247 | // i.e. first check each def if is covered or not (without adding them |
| 248 | // to the tracking set), and then add all the selected ones. |
| 249 | |
| 250 | // The reason for this is this example: |
| 251 | // *d1<A>, *d2<B>, ... Assume A and B are aliased (can happen in phi nodes). |
| 252 | // *d3<C> If A \incl BuC, and B \incl AuC, then *d2 would be |
| 253 | // covered if we added A first, and A would be covered |
| 254 | // if we added B first. |
| 255 | // In this example we want both A and B, because we don't want to give |
| 256 | // either one priority over the other, since they belong to the same |
| 257 | // statement. |
| 258 | |
| 259 | RegisterAggr RRs(DefRRs); |
| 260 | |
| 261 | auto DefInSet = [&Defs](NodeAddr<RefNode *> TA) -> bool { |
| 262 | return TA.Addr->getKind() == NodeAttrs::Def && Defs.count(V: TA.Id); |
| 263 | }; |
| 264 | |
| 265 | for (NodeId T : TmpInst) { |
| 266 | if (!FullChain && RRs.hasCoverOf(RR: RefRR)) |
| 267 | break; |
| 268 | auto TA = DFG.addr<InstrNode *>(N: T); |
| 269 | bool IsPhi = DFG.IsCode<NodeAttrs::Phi>(BA: TA); |
| 270 | NodeList Ds; |
| 271 | for (NodeAddr<DefNode *> DA : TA.Addr->members_if(P: DefInSet, G: DFG)) { |
| 272 | RegisterRef QR = DA.Addr->getRegRef(G: DFG); |
| 273 | // Add phi defs even if they are covered by subsequent defs. This is |
| 274 | // for cases where the reached use is not covered by any of the defs |
| 275 | // encountered so far: the phi def is needed to expose the liveness |
| 276 | // of that use to the entry of the block. |
| 277 | // Example: |
| 278 | // phi d1<R3>(,d2,), ... Phi def d1 is covered by d2. |
| 279 | // d2<R3>(d1,,u3), ... |
| 280 | // ..., u3<D1>(d2) This use needs to be live on entry. |
| 281 | if (FullChain || IsPhi || !RRs.hasCoverOf(RR: QR)) |
| 282 | Ds.push_back(Elt: DA); |
| 283 | } |
| 284 | llvm::append_range(C&: RDefs, R&: Ds); |
| 285 | for (NodeAddr<DefNode *> DA : Ds) { |
| 286 | // When collecting a full chain of definitions, do not consider phi |
| 287 | // defs to actually define a register. |
| 288 | uint16_t Flags = DA.Addr->getFlags(); |
| 289 | if (!FullChain || !(Flags & NodeAttrs::PhiRef)) |
| 290 | if (!(Flags & NodeAttrs::Preserving)) // Don't care about Undef here. |
| 291 | RRs.insert(RR: DA.Addr->getRegRef(G: DFG)); |
| 292 | } |
| 293 | } |
| 294 | |
| 295 | auto DeadP = [](const NodeAddr<DefNode *> DA) -> bool { |
| 296 | return DA.Addr->getFlags() & NodeAttrs::Dead; |
| 297 | }; |
| 298 | llvm::erase_if(C&: RDefs, P: DeadP); |
| 299 | |
| 300 | return RDefs; |
| 301 | } |
| 302 | |
| 303 | std::pair<NodeSet, bool> |
| 304 | Liveness::getAllReachingDefsRec(RegisterRef RefRR, NodeAddr<RefNode *> RefA, |
| 305 | NodeSet &Visited, const NodeSet &Defs) { |
| 306 | return getAllReachingDefsRecImpl(RefRR, RefA, Visited, Defs, Nest: 0, MaxNest: MaxRecNest); |
| 307 | } |
| 308 | |
| 309 | std::pair<NodeSet, bool> |
| 310 | Liveness::getAllReachingDefsRecImpl(RegisterRef RefRR, NodeAddr<RefNode *> RefA, |
| 311 | NodeSet &Visited, const NodeSet &Defs, |
| 312 | unsigned Nest, unsigned MaxNest) { |
| 313 | if (Nest > MaxNest) |
| 314 | return {NodeSet(), false}; |
| 315 | // Collect all defined registers. Do not consider phis to be defining |
| 316 | // anything, only collect "real" definitions. |
| 317 | RegisterAggr DefRRs(PRI); |
| 318 | for (NodeId D : Defs) { |
| 319 | const auto DA = DFG.addr<const DefNode *>(N: D); |
| 320 | if (!(DA.Addr->getFlags() & NodeAttrs::PhiRef)) |
| 321 | DefRRs.insert(RR: DA.Addr->getRegRef(G: DFG)); |
| 322 | } |
| 323 | |
| 324 | NodeList RDs = getAllReachingDefs(RefRR, RefA, TopShadows: false, FullChain: true, DefRRs); |
| 325 | if (RDs.empty()) |
| 326 | return {Defs, true}; |
| 327 | |
| 328 | // Make a copy of the preexisting definitions and add the newly found ones. |
| 329 | NodeSet TmpDefs = Defs; |
| 330 | for (NodeAddr<NodeBase *> R : RDs) |
| 331 | TmpDefs.insert(x: R.Id); |
| 332 | |
| 333 | NodeSet Result = Defs; |
| 334 | |
| 335 | for (NodeAddr<DefNode *> DA : RDs) { |
| 336 | Result.insert(x: DA.Id); |
| 337 | if (!(DA.Addr->getFlags() & NodeAttrs::PhiRef)) |
| 338 | continue; |
| 339 | NodeAddr<PhiNode *> PA = DA.Addr->getOwner(G: DFG); |
| 340 | if (!Visited.insert(x: PA.Id).second) |
| 341 | continue; |
| 342 | // Go over all phi uses and get the reaching defs for each use. |
| 343 | for (auto U : PA.Addr->members_if(P: DFG.IsRef<NodeAttrs::Use>, G: DFG)) { |
| 344 | const auto &T = getAllReachingDefsRecImpl(RefRR, RefA: U, Visited, Defs: TmpDefs, |
| 345 | Nest: Nest + 1, MaxNest); |
| 346 | if (!T.second) |
| 347 | return {T.first, false}; |
| 348 | Result.insert(first: T.first.begin(), last: T.first.end()); |
| 349 | } |
| 350 | } |
| 351 | |
| 352 | return {Result, true}; |
| 353 | } |
| 354 | |
| 355 | /// Find the nearest ref node aliased to RefRR, going upwards in the data |
| 356 | /// flow, starting from the instruction immediately preceding Inst. |
| 357 | NodeAddr<RefNode *> Liveness::getNearestAliasedRef(RegisterRef RefRR, |
| 358 | NodeAddr<InstrNode *> IA) { |
| 359 | NodeAddr<BlockNode *> BA = IA.Addr->getOwner(G: DFG); |
| 360 | NodeList Ins = BA.Addr->members(G: DFG); |
| 361 | NodeId FindId = IA.Id; |
| 362 | auto E = Ins.rend(); |
| 363 | auto B = |
| 364 | std::find_if(first: Ins.rbegin(), last: E, pred: [FindId](const NodeAddr<InstrNode *> T) { |
| 365 | return T.Id == FindId; |
| 366 | }); |
| 367 | // Do not scan IA (which is what B would point to). |
| 368 | if (B != E) |
| 369 | ++B; |
| 370 | |
| 371 | do { |
| 372 | // Process the range of instructions from B to E. |
| 373 | for (NodeAddr<InstrNode *> I : make_range(x: B, y: E)) { |
| 374 | NodeList Refs = I.Addr->members(G: DFG); |
| 375 | NodeAddr<RefNode *> Clob, Use; |
| 376 | // Scan all the refs in I aliased to RefRR, and return the one that |
| 377 | // is the closest to the output of I, i.e. def > clobber > use. |
| 378 | for (NodeAddr<RefNode *> R : Refs) { |
| 379 | if (!PRI.alias(RA: R.Addr->getRegRef(G: DFG), RB: RefRR)) |
| 380 | continue; |
| 381 | if (DFG.IsDef(BA: R)) { |
| 382 | // If it's a non-clobbering def, just return it. |
| 383 | if (!(R.Addr->getFlags() & NodeAttrs::Clobbering)) |
| 384 | return R; |
| 385 | Clob = R; |
| 386 | } else { |
| 387 | Use = R; |
| 388 | } |
| 389 | } |
| 390 | if (Clob.Id != 0) |
| 391 | return Clob; |
| 392 | if (Use.Id != 0) |
| 393 | return Use; |
| 394 | } |
| 395 | |
| 396 | // Go up to the immediate dominator, if any. |
| 397 | MachineBasicBlock *BB = BA.Addr->getCode(); |
| 398 | BA = NodeAddr<BlockNode *>(); |
| 399 | if (MachineDomTreeNode *N = MDT.getNode(BB)) { |
| 400 | if ((N = N->getIDom())) |
| 401 | BA = DFG.findBlock(BB: N->getBlock()); |
| 402 | } |
| 403 | if (!BA.Id) |
| 404 | break; |
| 405 | |
| 406 | Ins = BA.Addr->members(G: DFG); |
| 407 | B = Ins.rbegin(); |
| 408 | E = Ins.rend(); |
| 409 | } while (true); |
| 410 | |
| 411 | return NodeAddr<RefNode *>(); |
| 412 | } |
| 413 | |
| 414 | NodeSet Liveness::getAllReachedUses(RegisterRef RefRR, NodeAddr<DefNode *> DefA, |
| 415 | const RegisterAggr &DefRRs) { |
| 416 | NodeSet Uses; |
| 417 | |
| 418 | // If the original register is already covered by all the intervening |
| 419 | // defs, no more uses can be reached. |
| 420 | if (DefRRs.hasCoverOf(RR: RefRR)) |
| 421 | return Uses; |
| 422 | |
| 423 | // Add all directly reached uses. |
| 424 | // If the def is dead, it does not provide a value for any use. |
| 425 | bool IsDead = DefA.Addr->getFlags() & NodeAttrs::Dead; |
| 426 | NodeId U = !IsDead ? DefA.Addr->getReachedUse() : 0; |
| 427 | while (U != 0) { |
| 428 | auto UA = DFG.addr<UseNode *>(N: U); |
| 429 | if (!(UA.Addr->getFlags() & NodeAttrs::Undef)) { |
| 430 | RegisterRef UR = UA.Addr->getRegRef(G: DFG); |
| 431 | if (PRI.alias(RA: RefRR, RB: UR) && !DefRRs.hasCoverOf(RR: UR)) |
| 432 | Uses.insert(x: U); |
| 433 | } |
| 434 | U = UA.Addr->getSibling(); |
| 435 | } |
| 436 | |
| 437 | // Traverse all reached defs. This time dead defs cannot be ignored. |
| 438 | for (NodeId D = DefA.Addr->getReachedDef(), NextD; D != 0; D = NextD) { |
| 439 | auto DA = DFG.addr<DefNode *>(N: D); |
| 440 | NextD = DA.Addr->getSibling(); |
| 441 | RegisterRef DR = DA.Addr->getRegRef(G: DFG); |
| 442 | // If this def is already covered, it cannot reach anything new. |
| 443 | // Similarly, skip it if it is not aliased to the interesting register. |
| 444 | if (DefRRs.hasCoverOf(RR: DR) || !PRI.alias(RA: RefRR, RB: DR)) |
| 445 | continue; |
| 446 | NodeSet T; |
| 447 | if (DFG.IsPreservingDef(DA)) { |
| 448 | // If it is a preserving def, do not update the set of intervening defs. |
| 449 | T = getAllReachedUses(RefRR, DefA: DA, DefRRs); |
| 450 | } else { |
| 451 | RegisterAggr NewDefRRs = DefRRs; |
| 452 | NewDefRRs.insert(RR: DR); |
| 453 | T = getAllReachedUses(RefRR, DefA: DA, DefRRs: NewDefRRs); |
| 454 | } |
| 455 | Uses.insert(first: T.begin(), last: T.end()); |
| 456 | } |
| 457 | return Uses; |
| 458 | } |
| 459 | |
| 460 | void Liveness::computePhiInfo() { |
| 461 | RealUseMap.clear(); |
| 462 | |
| 463 | NodeList Phis; |
| 464 | NodeAddr<FuncNode *> FA = DFG.getFunc(); |
| 465 | NodeList Blocks = FA.Addr->members(G: DFG); |
| 466 | for (NodeAddr<BlockNode *> BA : Blocks) { |
| 467 | auto Ps = BA.Addr->members_if(P: DFG.IsCode<NodeAttrs::Phi>, G: DFG); |
| 468 | llvm::append_range(C&: Phis, R&: Ps); |
| 469 | } |
| 470 | |
| 471 | // phi use -> (map: reaching phi -> set of registers defined in between) |
| 472 | std::map<NodeId, std::map<NodeId, RegisterAggr>> PhiUp; |
| 473 | std::vector<NodeId> PhiUQ; // Work list of phis for upward propagation. |
| 474 | std::unordered_map<NodeId, RegisterAggr> |
| 475 | PhiDRs; // Phi -> registers defined by it. |
| 476 | |
| 477 | // Go over all phis. |
| 478 | for (NodeAddr<PhiNode *> PhiA : Phis) { |
| 479 | // Go over all defs and collect the reached uses that are non-phi uses |
| 480 | // (i.e. the "real uses"). |
| 481 | RefMap &RealUses = RealUseMap[PhiA.Id]; |
| 482 | NodeList PhiRefs = PhiA.Addr->members(G: DFG); |
| 483 | |
| 484 | // Have a work queue of defs whose reached uses need to be found. |
| 485 | // For each def, add to the queue all reached (non-phi) defs. |
| 486 | SetVector<NodeId> DefQ; |
| 487 | NodeSet PhiDefs; |
| 488 | RegisterAggr DRs(PRI); |
| 489 | for (NodeAddr<RefNode *> R : PhiRefs) { |
| 490 | if (!DFG.IsRef<NodeAttrs::Def>(BA: R)) |
| 491 | continue; |
| 492 | DRs.insert(RR: R.Addr->getRegRef(G: DFG)); |
| 493 | DefQ.insert(X: R.Id); |
| 494 | PhiDefs.insert(x: R.Id); |
| 495 | } |
| 496 | PhiDRs.insert(x: std::make_pair(x&: PhiA.Id, y&: DRs)); |
| 497 | |
| 498 | // Collect the super-set of all possible reached uses. This set will |
| 499 | // contain all uses reached from this phi, either directly from the |
| 500 | // phi defs, or (recursively) via non-phi defs reached by the phi defs. |
| 501 | // This set of uses will later be trimmed to only contain these uses that |
| 502 | // are actually reached by the phi defs. |
| 503 | for (unsigned i = 0; i < DefQ.size(); ++i) { |
| 504 | NodeAddr<DefNode *> DA = DFG.addr<DefNode *>(N: DefQ[i]); |
| 505 | // Visit all reached uses. Phi defs should not really have the "dead" |
| 506 | // flag set, but check it anyway for consistency. |
| 507 | bool IsDead = DA.Addr->getFlags() & NodeAttrs::Dead; |
| 508 | NodeId UN = !IsDead ? DA.Addr->getReachedUse() : 0; |
| 509 | while (UN != 0) { |
| 510 | NodeAddr<UseNode *> A = DFG.addr<UseNode *>(N: UN); |
| 511 | uint16_t F = A.Addr->getFlags(); |
| 512 | if ((F & (NodeAttrs::Undef | NodeAttrs::PhiRef)) == 0) { |
| 513 | RegisterRef R = A.Addr->getRegRef(G: DFG); |
| 514 | RealUses[R.Reg].insert(x: {A.Id, R.Mask}); |
| 515 | } |
| 516 | UN = A.Addr->getSibling(); |
| 517 | } |
| 518 | // Visit all reached defs, and add them to the queue. These defs may |
| 519 | // override some of the uses collected here, but that will be handled |
| 520 | // later. |
| 521 | NodeId DN = DA.Addr->getReachedDef(); |
| 522 | while (DN != 0) { |
| 523 | NodeAddr<DefNode *> A = DFG.addr<DefNode *>(N: DN); |
| 524 | for (auto T : DFG.getRelatedRefs(IA: A.Addr->getOwner(G: DFG), RA: A)) { |
| 525 | uint16_t Flags = NodeAddr<DefNode *>(T).Addr->getFlags(); |
| 526 | // Must traverse the reached-def chain. Consider: |
| 527 | // def(D0) -> def(R0) -> def(R0) -> use(D0) |
| 528 | // The reachable use of D0 passes through a def of R0. |
| 529 | if (!(Flags & NodeAttrs::PhiRef)) |
| 530 | DefQ.insert(X: T.Id); |
| 531 | } |
| 532 | DN = A.Addr->getSibling(); |
| 533 | } |
| 534 | } |
| 535 | // Filter out these uses that appear to be reachable, but really |
| 536 | // are not. For example: |
| 537 | // |
| 538 | // R1:0 = d1 |
| 539 | // = R1:0 u2 Reached by d1. |
| 540 | // R0 = d3 |
| 541 | // = R1:0 u4 Still reached by d1: indirectly through |
| 542 | // the def d3. |
| 543 | // R1 = d5 |
| 544 | // = R1:0 u6 Not reached by d1 (covered collectively |
| 545 | // by d3 and d5), but following reached |
| 546 | // defs and uses from d1 will lead here. |
| 547 | for (auto UI = RealUses.begin(), UE = RealUses.end(); UI != UE;) { |
| 548 | // For each reached register UI->first, there is a set UI->second, of |
| 549 | // uses of it. For each such use, check if it is reached by this phi, |
| 550 | // i.e. check if the set of its reaching uses intersects the set of |
| 551 | // this phi's defs. |
| 552 | NodeRefSet Uses = UI->second; |
| 553 | UI->second.clear(); |
| 554 | for (std::pair<NodeId, LaneBitmask> I : Uses) { |
| 555 | auto UA = DFG.addr<UseNode *>(N: I.first); |
| 556 | // Undef flag is checked above. |
| 557 | assert((UA.Addr->getFlags() & NodeAttrs::Undef) == 0); |
| 558 | RegisterRef UseR(UI->first, I.second); // Ref from Uses |
| 559 | // R = intersection of the ref from the phi and the ref from Uses |
| 560 | RegisterRef R = PhiDRs.at(k: PhiA.Id).intersectWith(RR: UseR); |
| 561 | if (!R) |
| 562 | continue; |
| 563 | // Calculate the exposed part of the reached use. |
| 564 | RegisterAggr Covered(PRI); |
| 565 | for (NodeAddr<DefNode *> DA : getAllReachingDefs(RefRR: R, RefA: UA)) { |
| 566 | if (PhiDefs.count(x: DA.Id)) |
| 567 | break; |
| 568 | Covered.insert(RR: DA.Addr->getRegRef(G: DFG)); |
| 569 | } |
| 570 | if (RegisterRef RC = Covered.clearIn(RR: R)) { |
| 571 | // We are updating the map for register UI->first, so we need |
| 572 | // to map RC to be expressed in terms of that register. |
| 573 | RegisterRef S = PRI.mapTo(RR: RC, R: UI->first); |
| 574 | UI->second.insert(x: {I.first, S.Mask}); |
| 575 | } |
| 576 | } |
| 577 | UI = UI->second.empty() ? RealUses.erase(position: UI) : std::next(x: UI); |
| 578 | } |
| 579 | |
| 580 | // If this phi reaches some "real" uses, add it to the queue for upward |
| 581 | // propagation. |
| 582 | if (!RealUses.empty()) |
| 583 | PhiUQ.push_back(x: PhiA.Id); |
| 584 | |
| 585 | // Go over all phi uses and check if the reaching def is another phi. |
| 586 | // Collect the phis that are among the reaching defs of these uses. |
| 587 | // While traversing the list of reaching defs for each phi use, accumulate |
| 588 | // the set of registers defined between this phi (PhiA) and the owner phi |
| 589 | // of the reaching def. |
| 590 | NodeSet SeenUses; |
| 591 | |
| 592 | for (auto I : PhiRefs) { |
| 593 | if (!DFG.IsRef<NodeAttrs::Use>(BA: I) || SeenUses.count(x: I.Id)) |
| 594 | continue; |
| 595 | NodeAddr<PhiUseNode *> PUA = I; |
| 596 | if (PUA.Addr->getReachingDef() == 0) |
| 597 | continue; |
| 598 | |
| 599 | RegisterRef UR = PUA.Addr->getRegRef(G: DFG); |
| 600 | NodeList Ds = getAllReachingDefs(RefRR: UR, RefA: PUA, TopShadows: true, FullChain: false, DefRRs: NoRegs); |
| 601 | RegisterAggr DefRRs(PRI); |
| 602 | |
| 603 | for (NodeAddr<DefNode *> D : Ds) { |
| 604 | if (D.Addr->getFlags() & NodeAttrs::PhiRef) { |
| 605 | NodeId RP = D.Addr->getOwner(G: DFG).Id; |
| 606 | auto [F, Inserted] = PhiUp[PUA.Id].try_emplace(k: RP, args&: DefRRs); |
| 607 | if (!Inserted) |
| 608 | F->second.insert(RG: DefRRs); |
| 609 | } |
| 610 | DefRRs.insert(RR: D.Addr->getRegRef(G: DFG)); |
| 611 | } |
| 612 | |
| 613 | for (NodeAddr<PhiUseNode *> T : DFG.getRelatedRefs(IA: PhiA, RA: PUA)) |
| 614 | SeenUses.insert(x: T.Id); |
| 615 | } |
| 616 | } |
| 617 | |
| 618 | if (Trace) { |
| 619 | dbgs() << "Phi-up-to-phi map with intervening defs:\n" ; |
| 620 | for (auto I : PhiUp) { |
| 621 | dbgs() << "phi " << Print(I.first, DFG) << " -> {" ; |
| 622 | for (auto R : I.second) |
| 623 | dbgs() << ' ' << Print(R.first, DFG) << Print(R.second, DFG); |
| 624 | dbgs() << " }\n" ; |
| 625 | } |
| 626 | } |
| 627 | |
| 628 | // Propagate the reached registers up in the phi chain. |
| 629 | // |
| 630 | // The following type of situation needs careful handling: |
| 631 | // |
| 632 | // phi d1<R1:0> (1) |
| 633 | // | |
| 634 | // ... d2<R1> |
| 635 | // | |
| 636 | // phi u3<R1:0> (2) |
| 637 | // | |
| 638 | // ... u4<R1> |
| 639 | // |
| 640 | // The phi node (2) defines a register pair R1:0, and reaches a "real" |
| 641 | // use u4 of just R1. The same phi node is also known to reach (upwards) |
| 642 | // the phi node (1). However, the use u4 is not reached by phi (1), |
| 643 | // because of the intervening definition d2 of R1. The data flow between |
| 644 | // phis (1) and (2) is restricted to R1:0 minus R1, i.e. R0. |
| 645 | // |
| 646 | // When propagating uses up the phi chains, get the all reaching defs |
| 647 | // for a given phi use, and traverse the list until the propagated ref |
| 648 | // is covered, or until reaching the final phi. Only assume that the |
| 649 | // reference reaches the phi in the latter case. |
| 650 | |
| 651 | // The operation "clearIn" can be expensive. For a given set of intervening |
| 652 | // defs, cache the result of subtracting these defs from a given register |
| 653 | // ref. |
| 654 | using RefHash = std::hash<RegisterRef>; |
| 655 | using RefEqual = std::equal_to<RegisterRef>; |
| 656 | using SubMap = std::unordered_map<RegisterRef, RegisterRef>; |
| 657 | std::unordered_map<RegisterAggr, SubMap> Subs; |
| 658 | auto ClearIn = [](RegisterRef RR, const RegisterAggr &Mid, SubMap &SM) { |
| 659 | if (Mid.empty()) |
| 660 | return RR; |
| 661 | auto F = SM.find(x: RR); |
| 662 | if (F != SM.end()) |
| 663 | return F->second; |
| 664 | RegisterRef S = Mid.clearIn(RR); |
| 665 | SM.insert(x: {RR, S}); |
| 666 | return S; |
| 667 | }; |
| 668 | |
| 669 | // Go over all phis. |
| 670 | for (unsigned i = 0; i < PhiUQ.size(); ++i) { |
| 671 | auto PA = DFG.addr<PhiNode *>(N: PhiUQ[i]); |
| 672 | NodeList PUs = PA.Addr->members_if(P: DFG.IsRef<NodeAttrs::Use>, G: DFG); |
| 673 | RefMap &RUM = RealUseMap[PA.Id]; |
| 674 | |
| 675 | for (NodeAddr<UseNode *> UA : PUs) { |
| 676 | std::map<NodeId, RegisterAggr> &PUM = PhiUp[UA.Id]; |
| 677 | RegisterRef UR = UA.Addr->getRegRef(G: DFG); |
| 678 | for (const std::pair<const NodeId, RegisterAggr> &P : PUM) { |
| 679 | bool Changed = false; |
| 680 | const RegisterAggr &MidDefs = P.second; |
| 681 | // Collect the set PropUp of uses that are reached by the current |
| 682 | // phi PA, and are not covered by any intervening def between the |
| 683 | // currently visited use UA and the upward phi P. |
| 684 | |
| 685 | if (MidDefs.hasCoverOf(RR: UR)) |
| 686 | continue; |
| 687 | SubMap &SM = Subs.try_emplace(k: MidDefs, args: 1, args: RefHash(), args: RefEqual(PRI)) |
| 688 | .first->second; |
| 689 | |
| 690 | // General algorithm: |
| 691 | // for each (R,U) : U is use node of R, U is reached by PA |
| 692 | // if MidDefs does not cover (R,U) |
| 693 | // then add (R-MidDefs,U) to RealUseMap[P] |
| 694 | // |
| 695 | for (const std::pair<const RegisterId, NodeRefSet> &T : RUM) { |
| 696 | RegisterRef R(T.first); |
| 697 | // The current phi (PA) could be a phi for a regmask. It could |
| 698 | // reach a whole variety of uses that are not related to the |
| 699 | // specific upward phi (P.first). |
| 700 | const RegisterAggr &DRs = PhiDRs.at(k: P.first); |
| 701 | if (!DRs.hasAliasOf(RR: R)) |
| 702 | continue; |
| 703 | R = PRI.mapTo(RR: DRs.intersectWith(RR: R), R: T.first); |
| 704 | for (std::pair<NodeId, LaneBitmask> V : T.second) { |
| 705 | LaneBitmask M = R.Mask & V.second; |
| 706 | if (M.none()) |
| 707 | continue; |
| 708 | if (RegisterRef SS = ClearIn(RegisterRef(R.Reg, M), MidDefs, SM)) { |
| 709 | NodeRefSet &RS = RealUseMap[P.first][SS.Reg]; |
| 710 | Changed |= RS.insert(x: {V.first, SS.Mask}).second; |
| 711 | } |
| 712 | } |
| 713 | } |
| 714 | |
| 715 | if (Changed) |
| 716 | PhiUQ.push_back(x: P.first); |
| 717 | } |
| 718 | } |
| 719 | } |
| 720 | |
| 721 | if (Trace) { |
| 722 | dbgs() << "Real use map:\n" ; |
| 723 | for (auto I : RealUseMap) { |
| 724 | dbgs() << "phi " << Print(I.first, DFG); |
| 725 | NodeAddr<PhiNode *> PA = DFG.addr<PhiNode *>(N: I.first); |
| 726 | NodeList Ds = PA.Addr->members_if(P: DFG.IsRef<NodeAttrs::Def>, G: DFG); |
| 727 | if (!Ds.empty()) { |
| 728 | RegisterRef RR = NodeAddr<DefNode *>(Ds[0]).Addr->getRegRef(G: DFG); |
| 729 | dbgs() << '<' << Print(RR, DFG) << '>'; |
| 730 | } else { |
| 731 | dbgs() << "<noreg>" ; |
| 732 | } |
| 733 | dbgs() << " -> " << Print(I.second, DFG) << '\n'; |
| 734 | } |
| 735 | } |
| 736 | } |
| 737 | |
| 738 | void Liveness::computeLiveIns() { |
| 739 | // Populate the node-to-block map. This speeds up the calculations |
| 740 | // significantly. |
| 741 | NBMap.clear(); |
| 742 | for (NodeAddr<BlockNode *> BA : DFG.getFunc().Addr->members(G: DFG)) { |
| 743 | MachineBasicBlock *BB = BA.Addr->getCode(); |
| 744 | for (NodeAddr<InstrNode *> IA : BA.Addr->members(G: DFG)) { |
| 745 | for (NodeAddr<RefNode *> RA : IA.Addr->members(G: DFG)) |
| 746 | NBMap.insert(KV: std::make_pair(x&: RA.Id, y&: BB)); |
| 747 | NBMap.insert(KV: std::make_pair(x&: IA.Id, y&: BB)); |
| 748 | } |
| 749 | } |
| 750 | |
| 751 | MachineFunction &MF = DFG.getMF(); |
| 752 | |
| 753 | // Compute IDF first, then the inverse. |
| 754 | decltype(IIDF) IDF; |
| 755 | for (MachineBasicBlock &B : MF) { |
| 756 | auto F1 = MDF.find(B: &B); |
| 757 | if (F1 == MDF.end()) |
| 758 | continue; |
| 759 | SetVector<MachineBasicBlock *> IDFB(llvm::from_range, F1->second); |
| 760 | for (unsigned i = 0; i < IDFB.size(); ++i) { |
| 761 | auto F2 = MDF.find(B: IDFB[i]); |
| 762 | if (F2 != MDF.end()) |
| 763 | IDFB.insert_range(R: F2->second); |
| 764 | } |
| 765 | // Add B to the IDF(B). This will put B in the IIDF(B). |
| 766 | IDFB.insert(X: &B); |
| 767 | IDF[&B].insert(first: IDFB.begin(), last: IDFB.end()); |
| 768 | } |
| 769 | |
| 770 | for (auto I : IDF) |
| 771 | for (auto *S : I.second) |
| 772 | IIDF[S].insert(x: I.first); |
| 773 | |
| 774 | computePhiInfo(); |
| 775 | |
| 776 | NodeAddr<FuncNode *> FA = DFG.getFunc(); |
| 777 | NodeList Blocks = FA.Addr->members(G: DFG); |
| 778 | |
| 779 | // Build the phi live-on-entry map. |
| 780 | for (NodeAddr<BlockNode *> BA : Blocks) { |
| 781 | MachineBasicBlock *MB = BA.Addr->getCode(); |
| 782 | RefMap &LON = PhiLON[MB]; |
| 783 | for (auto P : BA.Addr->members_if(P: DFG.IsCode<NodeAttrs::Phi>, G: DFG)) { |
| 784 | for (const RefMap::value_type &S : RealUseMap[P.Id]) |
| 785 | LON[S.first].insert(first: S.second.begin(), last: S.second.end()); |
| 786 | } |
| 787 | } |
| 788 | |
| 789 | if (Trace) { |
| 790 | dbgs() << "Phi live-on-entry map:\n" ; |
| 791 | for (auto &I : PhiLON) |
| 792 | dbgs() << "block #" << I.first->getNumber() << " -> " |
| 793 | << Print(I.second, DFG) << '\n'; |
| 794 | } |
| 795 | |
| 796 | // Build the phi live-on-exit map. Each phi node has some set of reached |
| 797 | // "real" uses. Propagate this set backwards into the block predecessors |
| 798 | // through the reaching defs of the corresponding phi uses. |
| 799 | for (NodeAddr<BlockNode *> BA : Blocks) { |
| 800 | NodeList Phis = BA.Addr->members_if(P: DFG.IsCode<NodeAttrs::Phi>, G: DFG); |
| 801 | for (NodeAddr<PhiNode *> PA : Phis) { |
| 802 | RefMap &RUs = RealUseMap[PA.Id]; |
| 803 | if (RUs.empty()) |
| 804 | continue; |
| 805 | |
| 806 | NodeSet SeenUses; |
| 807 | for (auto U : PA.Addr->members_if(P: DFG.IsRef<NodeAttrs::Use>, G: DFG)) { |
| 808 | if (!SeenUses.insert(x: U.Id).second) |
| 809 | continue; |
| 810 | NodeAddr<PhiUseNode *> PUA = U; |
| 811 | if (PUA.Addr->getReachingDef() == 0) |
| 812 | continue; |
| 813 | |
| 814 | // Each phi has some set (possibly empty) of reached "real" uses, |
| 815 | // that is, uses that are part of the compiled program. Such a use |
| 816 | // may be located in some farther block, but following a chain of |
| 817 | // reaching defs will eventually lead to this phi. |
| 818 | // Any chain of reaching defs may fork at a phi node, but there |
| 819 | // will be a path upwards that will lead to this phi. Now, this |
| 820 | // chain will need to fork at this phi, since some of the reached |
| 821 | // uses may have definitions joining in from multiple predecessors. |
| 822 | // For each reached "real" use, identify the set of reaching defs |
| 823 | // coming from each predecessor P, and add them to PhiLOX[P]. |
| 824 | // |
| 825 | auto PrA = DFG.addr<BlockNode *>(N: PUA.Addr->getPredecessor()); |
| 826 | RefMap &LOX = PhiLOX[PrA.Addr->getCode()]; |
| 827 | |
| 828 | for (const std::pair<const RegisterId, NodeRefSet> &RS : RUs) { |
| 829 | // We need to visit each individual use. |
| 830 | for (std::pair<NodeId, LaneBitmask> P : RS.second) { |
| 831 | // Create a register ref corresponding to the use, and find |
| 832 | // all reaching defs starting from the phi use, and treating |
| 833 | // all related shadows as a single use cluster. |
| 834 | RegisterRef S(RS.first, P.second); |
| 835 | NodeList Ds = getAllReachingDefs(RefRR: S, RefA: PUA, TopShadows: true, FullChain: false, DefRRs: NoRegs); |
| 836 | for (NodeAddr<DefNode *> D : Ds) { |
| 837 | // Calculate the mask corresponding to the visited def. |
| 838 | RegisterAggr TA(PRI); |
| 839 | TA.insert(RR: D.Addr->getRegRef(G: DFG)).intersect(RR: S); |
| 840 | LaneBitmask TM = TA.makeRegRef().Mask; |
| 841 | LOX[S.Reg].insert(x: {D.Id, TM}); |
| 842 | } |
| 843 | } |
| 844 | } |
| 845 | |
| 846 | for (NodeAddr<PhiUseNode *> T : DFG.getRelatedRefs(IA: PA, RA: PUA)) |
| 847 | SeenUses.insert(x: T.Id); |
| 848 | } // for U : phi uses |
| 849 | } // for P : Phis |
| 850 | } // for B : Blocks |
| 851 | |
| 852 | if (Trace) { |
| 853 | dbgs() << "Phi live-on-exit map:\n" ; |
| 854 | for (auto &I : PhiLOX) |
| 855 | dbgs() << "block #" << I.first->getNumber() << " -> " |
| 856 | << Print(I.second, DFG) << '\n'; |
| 857 | } |
| 858 | |
| 859 | RefMap LiveIn; |
| 860 | traverse(B: &MF.front(), LiveIn); |
| 861 | |
| 862 | // Add function live-ins to the live-in set of the function entry block. |
| 863 | LiveMap[&MF.front()].insert(RG: DFG.getLiveIns()); |
| 864 | |
| 865 | if (Trace) { |
| 866 | // Dump the liveness map |
| 867 | for (MachineBasicBlock &B : MF) { |
| 868 | std::vector<RegisterRef> LV; |
| 869 | for (const MachineBasicBlock::RegisterMaskPair &LI : B.liveins()) |
| 870 | LV.push_back(x: RegisterRef(LI.PhysReg, LI.LaneMask)); |
| 871 | llvm::sort(C&: LV, Comp: std::less<RegisterRef>(PRI)); |
| 872 | dbgs() << printMBBReference(MBB: B) << "\t rec = {" ; |
| 873 | for (auto I : LV) |
| 874 | dbgs() << ' ' << Print(I, DFG); |
| 875 | dbgs() << " }\n" ; |
| 876 | // dbgs() << "\tcomp = " << Print(LiveMap[&B], DFG) << '\n'; |
| 877 | |
| 878 | LV.clear(); |
| 879 | for (RegisterRef RR : LiveMap[&B].refs()) |
| 880 | LV.push_back(x: RR); |
| 881 | llvm::sort(C&: LV, Comp: std::less<RegisterRef>(PRI)); |
| 882 | dbgs() << "\tcomp = {" ; |
| 883 | for (auto I : LV) |
| 884 | dbgs() << ' ' << Print(I, DFG); |
| 885 | dbgs() << " }\n" ; |
| 886 | } |
| 887 | } |
| 888 | } |
| 889 | |
| 890 | void Liveness::resetLiveIns() { |
| 891 | for (auto &B : DFG.getMF()) { |
| 892 | // Remove all live-ins. |
| 893 | std::vector<MCRegister> T; |
| 894 | for (const MachineBasicBlock::RegisterMaskPair &LI : B.liveins()) |
| 895 | T.push_back(x: LI.PhysReg); |
| 896 | for (auto I : T) |
| 897 | B.removeLiveIn(Reg: I); |
| 898 | // Add the newly computed live-ins. |
| 899 | const RegisterAggr &LiveIns = LiveMap[&B]; |
| 900 | for (RegisterRef R : LiveIns.refs()) |
| 901 | B.addLiveIn(RegMaskPair: {MCPhysReg(R.Reg), R.Mask}); |
| 902 | } |
| 903 | } |
| 904 | |
| 905 | void Liveness::resetKills() { |
| 906 | for (auto &B : DFG.getMF()) |
| 907 | resetKills(B: &B); |
| 908 | } |
| 909 | |
| 910 | void Liveness::resetKills(MachineBasicBlock *B) { |
| 911 | auto CopyLiveIns = [this](MachineBasicBlock *B, BitVector &LV) -> void { |
| 912 | for (auto I : B->liveins()) { |
| 913 | MCSubRegIndexIterator S(I.PhysReg, &TRI); |
| 914 | if (!S.isValid()) { |
| 915 | LV.set(I.PhysReg.id()); |
| 916 | continue; |
| 917 | } |
| 918 | do { |
| 919 | LaneBitmask M = TRI.getSubRegIndexLaneMask(SubIdx: S.getSubRegIndex()); |
| 920 | if ((M & I.LaneMask).any()) |
| 921 | LV.set(S.getSubReg()); |
| 922 | ++S; |
| 923 | } while (S.isValid()); |
| 924 | } |
| 925 | }; |
| 926 | |
| 927 | BitVector LiveIn(TRI.getNumRegs()), Live(TRI.getNumRegs()); |
| 928 | CopyLiveIns(B, LiveIn); |
| 929 | for (auto *SI : B->successors()) |
| 930 | CopyLiveIns(SI, Live); |
| 931 | |
| 932 | for (MachineInstr &MI : llvm::reverse(C&: *B)) { |
| 933 | if (MI.isDebugInstr()) |
| 934 | continue; |
| 935 | |
| 936 | MI.clearKillInfo(); |
| 937 | for (auto &Op : MI.all_defs()) { |
| 938 | // An implicit def of a super-register may not necessarily start a |
| 939 | // live range of it, since an implicit use could be used to keep parts |
| 940 | // of it live. Instead of analyzing the implicit operands, ignore |
| 941 | // implicit defs. |
| 942 | if (Op.isImplicit()) |
| 943 | continue; |
| 944 | Register R = Op.getReg(); |
| 945 | if (!R.isPhysical()) |
| 946 | continue; |
| 947 | for (MCPhysReg SR : TRI.subregs_inclusive(Reg: R)) |
| 948 | Live.reset(Idx: SR); |
| 949 | } |
| 950 | for (auto &Op : MI.all_uses()) { |
| 951 | if (Op.isUndef()) |
| 952 | continue; |
| 953 | Register R = Op.getReg(); |
| 954 | if (!R.isPhysical()) |
| 955 | continue; |
| 956 | bool IsLive = false; |
| 957 | for (MCRegAliasIterator AR(R, &TRI, true); AR.isValid(); ++AR) { |
| 958 | if (!Live[(*AR).id()]) |
| 959 | continue; |
| 960 | IsLive = true; |
| 961 | break; |
| 962 | } |
| 963 | if (!IsLive) |
| 964 | Op.setIsKill(true); |
| 965 | for (MCPhysReg SR : TRI.subregs_inclusive(Reg: R)) |
| 966 | Live.set(SR); |
| 967 | } |
| 968 | } |
| 969 | } |
| 970 | |
| 971 | // Helper function to obtain the basic block containing the reaching def |
| 972 | // of the given use. |
| 973 | MachineBasicBlock *Liveness::getBlockWithRef(NodeId RN) const { |
| 974 | auto F = NBMap.find(Val: RN); |
| 975 | if (F != NBMap.end()) |
| 976 | return F->second; |
| 977 | llvm_unreachable("Node id not in map" ); |
| 978 | } |
| 979 | |
| 980 | void Liveness::traverse(MachineBasicBlock *B, RefMap &LiveIn) { |
| 981 | // The LiveIn map, for each (physical) register, contains the set of live |
| 982 | // reaching defs of that register that are live on entry to the associated |
| 983 | // block. |
| 984 | |
| 985 | // The summary of the traversal algorithm: |
| 986 | // |
| 987 | // R is live-in in B, if there exists a U(R), such that rdef(R) dom B |
| 988 | // and (U \in IDF(B) or B dom U). |
| 989 | // |
| 990 | // for (C : children) { |
| 991 | // LU = {} |
| 992 | // traverse(C, LU) |
| 993 | // LiveUses += LU |
| 994 | // } |
| 995 | // |
| 996 | // LiveUses -= Defs(B); |
| 997 | // LiveUses += UpwardExposedUses(B); |
| 998 | // for (C : IIDF[B]) |
| 999 | // for (U : LiveUses) |
| 1000 | // if (Rdef(U) dom C) |
| 1001 | // C.addLiveIn(U) |
| 1002 | // |
| 1003 | |
| 1004 | // Go up the dominator tree (depth-first). |
| 1005 | MachineDomTreeNode *N = MDT.getNode(BB: B); |
| 1006 | for (auto *I : *N) { |
| 1007 | RefMap L; |
| 1008 | MachineBasicBlock *SB = I->getBlock(); |
| 1009 | traverse(B: SB, LiveIn&: L); |
| 1010 | |
| 1011 | for (auto S : L) |
| 1012 | LiveIn[S.first].insert(first: S.second.begin(), last: S.second.end()); |
| 1013 | } |
| 1014 | |
| 1015 | if (Trace) { |
| 1016 | dbgs() << "\n-- " << printMBBReference(MBB: *B) << ": " << __func__ |
| 1017 | << " after recursion into: {" ; |
| 1018 | for (auto *I : *N) |
| 1019 | dbgs() << ' ' << I->getBlock()->getNumber(); |
| 1020 | dbgs() << " }\n" ; |
| 1021 | dbgs() << " LiveIn: " << Print(LiveIn, DFG) << '\n'; |
| 1022 | dbgs() << " Local: " << Print(LiveMap[B], DFG) << '\n'; |
| 1023 | } |
| 1024 | |
| 1025 | // Add reaching defs of phi uses that are live on exit from this block. |
| 1026 | RefMap &PUs = PhiLOX[B]; |
| 1027 | for (auto &S : PUs) |
| 1028 | LiveIn[S.first].insert(first: S.second.begin(), last: S.second.end()); |
| 1029 | |
| 1030 | if (Trace) { |
| 1031 | dbgs() << "after LOX\n" ; |
| 1032 | dbgs() << " LiveIn: " << Print(LiveIn, DFG) << '\n'; |
| 1033 | dbgs() << " Local: " << Print(LiveMap[B], DFG) << '\n'; |
| 1034 | } |
| 1035 | |
| 1036 | // The LiveIn map at this point has all defs that are live-on-exit from B, |
| 1037 | // as if they were live-on-entry to B. First, we need to filter out all |
| 1038 | // defs that are present in this block. Then we will add reaching defs of |
| 1039 | // all upward-exposed uses. |
| 1040 | |
| 1041 | // To filter out the defs, first make a copy of LiveIn, and then re-populate |
| 1042 | // LiveIn with the defs that should remain. |
| 1043 | RefMap LiveInCopy = LiveIn; |
| 1044 | LiveIn.clear(); |
| 1045 | |
| 1046 | for (const std::pair<const RegisterId, NodeRefSet> &LE : LiveInCopy) { |
| 1047 | RegisterRef LRef(LE.first); |
| 1048 | NodeRefSet &NewDefs = LiveIn[LRef.Reg]; // To be filled. |
| 1049 | const NodeRefSet &OldDefs = LE.second; |
| 1050 | for (NodeRef OR : OldDefs) { |
| 1051 | // R is a def node that was live-on-exit |
| 1052 | auto DA = DFG.addr<DefNode *>(N: OR.first); |
| 1053 | NodeAddr<InstrNode *> IA = DA.Addr->getOwner(G: DFG); |
| 1054 | NodeAddr<BlockNode *> BA = IA.Addr->getOwner(G: DFG); |
| 1055 | if (B != BA.Addr->getCode()) { |
| 1056 | // Defs from a different block need to be preserved. Defs from this |
| 1057 | // block will need to be processed further, except for phi defs, the |
| 1058 | // liveness of which is handled through the PhiLON/PhiLOX maps. |
| 1059 | NewDefs.insert(x: OR); |
| 1060 | continue; |
| 1061 | } |
| 1062 | |
| 1063 | // Defs from this block need to stop the liveness from being |
| 1064 | // propagated upwards. This only applies to non-preserving defs, |
| 1065 | // and to the parts of the register actually covered by those defs. |
| 1066 | // (Note that phi defs should always be preserving.) |
| 1067 | RegisterAggr RRs(PRI); |
| 1068 | LRef.Mask = OR.second; |
| 1069 | |
| 1070 | if (!DFG.IsPreservingDef(DA)) { |
| 1071 | assert(!(IA.Addr->getFlags() & NodeAttrs::Phi)); |
| 1072 | // DA is a non-phi def that is live-on-exit from this block, and |
| 1073 | // that is also located in this block. LRef is a register ref |
| 1074 | // whose use this def reaches. If DA covers LRef, then no part |
| 1075 | // of LRef is exposed upwards.A |
| 1076 | if (RRs.insert(RR: DA.Addr->getRegRef(G: DFG)).hasCoverOf(RR: LRef)) |
| 1077 | continue; |
| 1078 | } |
| 1079 | |
| 1080 | // DA itself was not sufficient to cover LRef. In general, it is |
| 1081 | // the last in a chain of aliased defs before the exit from this block. |
| 1082 | // There could be other defs in this block that are a part of that |
| 1083 | // chain. Check that now: accumulate the registers from these defs, |
| 1084 | // and if they all together cover LRef, it is not live-on-entry. |
| 1085 | for (NodeAddr<DefNode *> TA : getAllReachingDefs(RefA: DA)) { |
| 1086 | // DefNode -> InstrNode -> BlockNode. |
| 1087 | NodeAddr<InstrNode *> ITA = TA.Addr->getOwner(G: DFG); |
| 1088 | NodeAddr<BlockNode *> BTA = ITA.Addr->getOwner(G: DFG); |
| 1089 | // Reaching defs are ordered in the upward direction. |
| 1090 | if (BTA.Addr->getCode() != B) { |
| 1091 | // We have reached past the beginning of B, and the accumulated |
| 1092 | // registers are not covering LRef. The first def from the |
| 1093 | // upward chain will be live. |
| 1094 | // Subtract all accumulated defs (RRs) from LRef. |
| 1095 | RegisterRef T = RRs.clearIn(RR: LRef); |
| 1096 | assert(T); |
| 1097 | NewDefs.insert(x: {TA.Id, T.Mask}); |
| 1098 | break; |
| 1099 | } |
| 1100 | |
| 1101 | // TA is in B. Only add this def to the accumulated cover if it is |
| 1102 | // not preserving. |
| 1103 | if (!(TA.Addr->getFlags() & NodeAttrs::Preserving)) |
| 1104 | RRs.insert(RR: TA.Addr->getRegRef(G: DFG)); |
| 1105 | // If this is enough to cover LRef, then stop. |
| 1106 | if (RRs.hasCoverOf(RR: LRef)) |
| 1107 | break; |
| 1108 | } |
| 1109 | } |
| 1110 | } |
| 1111 | |
| 1112 | emptify(M&: LiveIn); |
| 1113 | |
| 1114 | if (Trace) { |
| 1115 | dbgs() << "after defs in block\n" ; |
| 1116 | dbgs() << " LiveIn: " << Print(LiveIn, DFG) << '\n'; |
| 1117 | dbgs() << " Local: " << Print(LiveMap[B], DFG) << '\n'; |
| 1118 | } |
| 1119 | |
| 1120 | // Scan the block for upward-exposed uses and add them to the tracking set. |
| 1121 | for (auto I : DFG.getFunc().Addr->findBlock(BB: B, G: DFG).Addr->members(G: DFG)) { |
| 1122 | NodeAddr<InstrNode *> IA = I; |
| 1123 | if (IA.Addr->getKind() != NodeAttrs::Stmt) |
| 1124 | continue; |
| 1125 | for (NodeAddr<UseNode *> UA : IA.Addr->members_if(P: DFG.IsUse, G: DFG)) { |
| 1126 | if (UA.Addr->getFlags() & NodeAttrs::Undef) |
| 1127 | continue; |
| 1128 | RegisterRef RR = UA.Addr->getRegRef(G: DFG); |
| 1129 | for (NodeAddr<DefNode *> D : getAllReachingDefs(RefA: UA)) |
| 1130 | if (getBlockWithRef(RN: D.Id) != B) |
| 1131 | LiveIn[RR.Reg].insert(x: {D.Id, RR.Mask}); |
| 1132 | } |
| 1133 | } |
| 1134 | |
| 1135 | if (Trace) { |
| 1136 | dbgs() << "after uses in block\n" ; |
| 1137 | dbgs() << " LiveIn: " << Print(LiveIn, DFG) << '\n'; |
| 1138 | dbgs() << " Local: " << Print(LiveMap[B], DFG) << '\n'; |
| 1139 | } |
| 1140 | |
| 1141 | // Phi uses should not be propagated up the dominator tree, since they |
| 1142 | // are not dominated by their corresponding reaching defs. |
| 1143 | RegisterAggr &Local = LiveMap[B]; |
| 1144 | RefMap &LON = PhiLON[B]; |
| 1145 | for (auto &R : LON) { |
| 1146 | LaneBitmask M; |
| 1147 | for (auto P : R.second) |
| 1148 | M |= P.second; |
| 1149 | Local.insert(RR: RegisterRef(R.first, M)); |
| 1150 | } |
| 1151 | |
| 1152 | if (Trace) { |
| 1153 | dbgs() << "after phi uses in block\n" ; |
| 1154 | dbgs() << " LiveIn: " << Print(LiveIn, DFG) << '\n'; |
| 1155 | dbgs() << " Local: " << Print(Local, DFG) << '\n'; |
| 1156 | } |
| 1157 | |
| 1158 | for (auto *C : IIDF[B]) { |
| 1159 | RegisterAggr &LiveC = LiveMap[C]; |
| 1160 | for (const std::pair<const RegisterId, NodeRefSet> &S : LiveIn) |
| 1161 | for (auto R : S.second) |
| 1162 | if (MDT.properlyDominates(A: getBlockWithRef(RN: R.first), B: C)) |
| 1163 | LiveC.insert(RR: RegisterRef(S.first, R.second)); |
| 1164 | } |
| 1165 | } |
| 1166 | |
| 1167 | void Liveness::emptify(RefMap &M) { |
| 1168 | for (auto I = M.begin(), E = M.end(); I != E;) |
| 1169 | I = I->second.empty() ? M.erase(position: I) : std::next(x: I); |
| 1170 | } |
| 1171 | |
| 1172 | } // namespace llvm::rdf |
| 1173 | |