| 1 | //===- AArch64LoadStoreOptimizer.cpp - AArch64 load/store opt. pass -------===// |
| 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 a pass that performs load / store related peephole |
| 10 | // optimizations. This pass should be run after register allocation. |
| 11 | // |
| 12 | // The pass runs after the PrologEpilogInserter where we emit the CFI |
| 13 | // instructions. In order to preserve the correctness of the unwind information, |
| 14 | // the pass should not change the order of any two instructions, one of which |
| 15 | // has the FrameSetup/FrameDestroy flag or, alternatively, apply an add-hoc fix |
| 16 | // to unwind information. |
| 17 | // |
| 18 | //===----------------------------------------------------------------------===// |
| 19 | |
| 20 | #include "AArch64InstrInfo.h" |
| 21 | #include "AArch64MachineFunctionInfo.h" |
| 22 | #include "AArch64Subtarget.h" |
| 23 | #include "MCTargetDesc/AArch64AddressingModes.h" |
| 24 | #include "llvm/ADT/SmallVector.h" |
| 25 | #include "llvm/ADT/Statistic.h" |
| 26 | #include "llvm/ADT/StringRef.h" |
| 27 | #include "llvm/ADT/iterator_range.h" |
| 28 | #include "llvm/Analysis/AliasAnalysis.h" |
| 29 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 30 | #include "llvm/CodeGen/MachineFunction.h" |
| 31 | #include "llvm/CodeGen/MachineFunctionPass.h" |
| 32 | #include "llvm/CodeGen/MachineInstr.h" |
| 33 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 34 | #include "llvm/CodeGen/MachineOperand.h" |
| 35 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 36 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 37 | #include "llvm/IR/DebugLoc.h" |
| 38 | #include "llvm/MC/MCAsmInfo.h" |
| 39 | #include "llvm/MC/MCDwarf.h" |
| 40 | #include "llvm/Pass.h" |
| 41 | #include "llvm/Support/CommandLine.h" |
| 42 | #include "llvm/Support/Debug.h" |
| 43 | #include "llvm/Support/DebugCounter.h" |
| 44 | #include "llvm/Support/ErrorHandling.h" |
| 45 | #include <cassert> |
| 46 | #include <cstdint> |
| 47 | #include <functional> |
| 48 | #include <iterator> |
| 49 | #include <limits> |
| 50 | #include <optional> |
| 51 | |
| 52 | using namespace llvm; |
| 53 | |
| 54 | #define DEBUG_TYPE "aarch64-ldst-opt" |
| 55 | |
| 56 | STATISTIC(NumPairCreated, "Number of load/store pair instructions generated" ); |
| 57 | STATISTIC(NumPostFolded, "Number of post-index updates folded" ); |
| 58 | STATISTIC(NumPreFolded, "Number of pre-index updates folded" ); |
| 59 | STATISTIC(NumUnscaledPairCreated, |
| 60 | "Number of load/store from unscaled generated" ); |
| 61 | STATISTIC(NumZeroStoresPromoted, "Number of narrow zero stores promoted" ); |
| 62 | STATISTIC(NumLoadsFromStoresPromoted, "Number of loads from stores promoted" ); |
| 63 | STATISTIC(NumFailedAlignmentCheck, "Number of load/store pair transformation " |
| 64 | "not passed the alignment check" ); |
| 65 | STATISTIC(NumConstOffsetFolded, |
| 66 | "Number of const offset of index address folded" ); |
| 67 | |
| 68 | DEBUG_COUNTER(RegRenamingCounter, DEBUG_TYPE "-reg-renaming" , |
| 69 | "Controls which pairs are considered for renaming" ); |
| 70 | |
| 71 | // The LdStLimit limits how far we search for load/store pairs. |
| 72 | static cl::opt<unsigned> LdStLimit("aarch64-load-store-scan-limit" , |
| 73 | cl::init(Val: 20), cl::Hidden); |
| 74 | |
| 75 | // The UpdateLimit limits how far we search for update instructions when we form |
| 76 | // pre-/post-index instructions. |
| 77 | static cl::opt<unsigned> UpdateLimit("aarch64-update-scan-limit" , cl::init(Val: 100), |
| 78 | cl::Hidden); |
| 79 | |
| 80 | // The LdStConstLimit limits how far we search for const offset instructions |
| 81 | // when we form index address load/store instructions. |
| 82 | static cl::opt<unsigned> LdStConstLimit("aarch64-load-store-const-scan-limit" , |
| 83 | cl::init(Val: 10), cl::Hidden); |
| 84 | |
| 85 | // Enable register renaming to find additional store pairing opportunities. |
| 86 | static cl::opt<bool> EnableRenaming("aarch64-load-store-renaming" , |
| 87 | cl::init(Val: true), cl::Hidden); |
| 88 | |
| 89 | #define AARCH64_LOAD_STORE_OPT_NAME "AArch64 load / store optimization pass" |
| 90 | |
| 91 | namespace { |
| 92 | |
| 93 | using LdStPairFlags = struct LdStPairFlags { |
| 94 | // If a matching instruction is found, MergeForward is set to true if the |
| 95 | // merge is to remove the first instruction and replace the second with |
| 96 | // a pair-wise insn, and false if the reverse is true. |
| 97 | bool MergeForward = false; |
| 98 | |
| 99 | // SExtIdx gives the index of the result of the load pair that must be |
| 100 | // extended. The value of SExtIdx assumes that the paired load produces the |
| 101 | // value in this order: (I, returned iterator), i.e., -1 means no value has |
| 102 | // to be extended, 0 means I, and 1 means the returned iterator. |
| 103 | int SExtIdx = -1; |
| 104 | |
| 105 | // If not none, RenameReg can be used to rename the result register of the |
| 106 | // first store in a pair. Currently this only works when merging stores |
| 107 | // forward. |
| 108 | std::optional<MCPhysReg> RenameReg; |
| 109 | |
| 110 | LdStPairFlags() = default; |
| 111 | |
| 112 | void setMergeForward(bool V = true) { MergeForward = V; } |
| 113 | bool getMergeForward() const { return MergeForward; } |
| 114 | |
| 115 | void setSExtIdx(int V) { SExtIdx = V; } |
| 116 | int getSExtIdx() const { return SExtIdx; } |
| 117 | |
| 118 | void setRenameReg(MCPhysReg R) { RenameReg = R; } |
| 119 | void clearRenameReg() { RenameReg = std::nullopt; } |
| 120 | std::optional<MCPhysReg> getRenameReg() const { return RenameReg; } |
| 121 | }; |
| 122 | |
| 123 | struct AArch64LoadStoreOpt { |
| 124 | AliasAnalysis *AA; |
| 125 | const AArch64InstrInfo *TII; |
| 126 | const TargetRegisterInfo *TRI; |
| 127 | const AArch64Subtarget *Subtarget; |
| 128 | |
| 129 | // Track which register units have been modified and used. |
| 130 | LiveRegUnits ModifiedRegUnits, UsedRegUnits; |
| 131 | LiveRegUnits DefinedInBB; |
| 132 | |
| 133 | // Scan the instructions looking for a load/store that can be combined |
| 134 | // with the current instruction into a load/store pair. |
| 135 | // Return the matching instruction if one is found, else MBB->end(). |
| 136 | MachineBasicBlock::iterator findMatchingInsn(MachineBasicBlock::iterator I, |
| 137 | LdStPairFlags &Flags, |
| 138 | unsigned Limit, |
| 139 | bool FindNarrowMerge); |
| 140 | |
| 141 | // Scan the instructions looking for a store that writes to the address from |
| 142 | // which the current load instruction reads. Return true if one is found. |
| 143 | bool findMatchingStore(MachineBasicBlock::iterator I, unsigned Limit, |
| 144 | MachineBasicBlock::iterator &StoreI); |
| 145 | |
| 146 | // Merge the two instructions indicated into a wider narrow store instruction. |
| 147 | MachineBasicBlock::iterator |
| 148 | mergeNarrowZeroStores(MachineBasicBlock::iterator I, |
| 149 | MachineBasicBlock::iterator MergeMI, |
| 150 | const LdStPairFlags &Flags); |
| 151 | |
| 152 | // Merge the two instructions indicated into a single pair-wise instruction. |
| 153 | MachineBasicBlock::iterator |
| 154 | mergePairedInsns(MachineBasicBlock::iterator I, |
| 155 | MachineBasicBlock::iterator Paired, |
| 156 | const LdStPairFlags &Flags); |
| 157 | |
| 158 | // Promote the load that reads directly from the address stored to. |
| 159 | MachineBasicBlock::iterator |
| 160 | promoteLoadFromStore(MachineBasicBlock::iterator LoadI, |
| 161 | MachineBasicBlock::iterator StoreI); |
| 162 | |
| 163 | // Scan the instruction list to find a base register update that can |
| 164 | // be combined with the current instruction (a load or store) using |
| 165 | // pre or post indexed addressing with writeback. Scan forwards. |
| 166 | MachineBasicBlock::iterator |
| 167 | findMatchingUpdateInsnForward(MachineBasicBlock::iterator I, |
| 168 | int UnscaledOffset, unsigned Limit); |
| 169 | |
| 170 | // Scan the instruction list to find a register assigned with a const |
| 171 | // value that can be combined with the current instruction (a load or store) |
| 172 | // using base addressing with writeback. Scan backwards. |
| 173 | MachineBasicBlock::iterator |
| 174 | findMatchingConstOffsetBackward(MachineBasicBlock::iterator I, unsigned Limit, |
| 175 | unsigned &Offset); |
| 176 | |
| 177 | // Scan the instruction list to find a base register update that can |
| 178 | // be combined with the current instruction (a load or store) using |
| 179 | // pre or post indexed addressing with writeback. Scan backwards. |
| 180 | // `MergeEither` is set to true if the combined instruction may be placed |
| 181 | // either at the location of the load/store instruction or at the location of |
| 182 | // the update instruction. |
| 183 | MachineBasicBlock::iterator |
| 184 | findMatchingUpdateInsnBackward(MachineBasicBlock::iterator I, unsigned Limit, |
| 185 | bool &MergeEither); |
| 186 | |
| 187 | // Find an instruction that updates the base register of the ld/st |
| 188 | // instruction. |
| 189 | bool isMatchingUpdateInsn(MachineInstr &MemMI, MachineInstr &MI, |
| 190 | unsigned BaseReg, int Offset); |
| 191 | |
| 192 | bool isMatchingMovConstInsn(MachineInstr &MemMI, MachineInstr &MI, |
| 193 | unsigned IndexReg, unsigned &Offset); |
| 194 | |
| 195 | // Merge a pre- or post-index base register update into a ld/st instruction. |
| 196 | std::optional<MachineBasicBlock::iterator> |
| 197 | mergeUpdateInsn(MachineBasicBlock::iterator I, |
| 198 | MachineBasicBlock::iterator Update, bool IsForward, |
| 199 | bool IsPreIdx, bool MergeEither); |
| 200 | |
| 201 | MachineBasicBlock::iterator |
| 202 | mergeConstOffsetInsn(MachineBasicBlock::iterator I, |
| 203 | MachineBasicBlock::iterator Update, unsigned Offset, |
| 204 | int Scale); |
| 205 | |
| 206 | // Find and merge zero store instructions. |
| 207 | bool tryToMergeZeroStInst(MachineBasicBlock::iterator &MBBI); |
| 208 | |
| 209 | // Find and pair ldr/str instructions. |
| 210 | bool tryToPairLdStInst(MachineBasicBlock::iterator &MBBI); |
| 211 | |
| 212 | // Find and promote load instructions which read directly from store. |
| 213 | bool tryToPromoteLoadFromStore(MachineBasicBlock::iterator &MBBI); |
| 214 | |
| 215 | // Find and merge a base register updates before or after a ld/st instruction. |
| 216 | bool tryToMergeLdStUpdate(MachineBasicBlock::iterator &MBBI); |
| 217 | |
| 218 | // Find and merge an index ldr/st instruction into a base ld/st instruction. |
| 219 | bool tryToMergeIndexLdSt(MachineBasicBlock::iterator &MBBI, int Scale); |
| 220 | |
| 221 | bool optimizeBlock(MachineBasicBlock &MBB, bool EnableNarrowZeroStOpt); |
| 222 | |
| 223 | bool runOnMachineFunction(MachineFunction &MF); |
| 224 | }; |
| 225 | |
| 226 | struct AArch64LoadStoreOptLegacy : public MachineFunctionPass { |
| 227 | static char ID; |
| 228 | |
| 229 | AArch64LoadStoreOptLegacy() : MachineFunctionPass(ID) {} |
| 230 | |
| 231 | bool runOnMachineFunction(MachineFunction &Fn) override; |
| 232 | |
| 233 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 234 | AU.addRequired<AAResultsWrapperPass>(); |
| 235 | MachineFunctionPass::getAnalysisUsage(AU); |
| 236 | } |
| 237 | |
| 238 | MachineFunctionProperties getRequiredProperties() const override { |
| 239 | return MachineFunctionProperties().setNoVRegs(); |
| 240 | } |
| 241 | |
| 242 | StringRef getPassName() const override { return AARCH64_LOAD_STORE_OPT_NAME; } |
| 243 | }; |
| 244 | |
| 245 | char AArch64LoadStoreOptLegacy::ID = 0; |
| 246 | |
| 247 | } // end anonymous namespace |
| 248 | |
| 249 | INITIALIZE_PASS(AArch64LoadStoreOptLegacy, "aarch64-ldst-opt" , |
| 250 | AARCH64_LOAD_STORE_OPT_NAME, false, false) |
| 251 | |
| 252 | static bool isNarrowStore(unsigned Opc) { |
| 253 | switch (Opc) { |
| 254 | default: |
| 255 | return false; |
| 256 | case AArch64::STRBBui: |
| 257 | case AArch64::STURBBi: |
| 258 | case AArch64::STRHHui: |
| 259 | case AArch64::STURHHi: |
| 260 | return true; |
| 261 | } |
| 262 | } |
| 263 | |
| 264 | // These instruction set memory tag and either keep memory contents unchanged or |
| 265 | // set it to zero, ignoring the address part of the source register. |
| 266 | static bool isTagStore(const MachineInstr &MI) { |
| 267 | switch (MI.getOpcode()) { |
| 268 | default: |
| 269 | return false; |
| 270 | case AArch64::STGi: |
| 271 | case AArch64::STZGi: |
| 272 | case AArch64::ST2Gi: |
| 273 | case AArch64::STZ2Gi: |
| 274 | return true; |
| 275 | } |
| 276 | } |
| 277 | |
| 278 | static unsigned getMatchingNonSExtOpcode(unsigned Opc, |
| 279 | bool *IsValidLdStrOpc = nullptr) { |
| 280 | if (IsValidLdStrOpc) |
| 281 | *IsValidLdStrOpc = true; |
| 282 | switch (Opc) { |
| 283 | default: |
| 284 | if (IsValidLdStrOpc) |
| 285 | *IsValidLdStrOpc = false; |
| 286 | return std::numeric_limits<unsigned>::max(); |
| 287 | case AArch64::STRDui: |
| 288 | case AArch64::STURDi: |
| 289 | case AArch64::STRDpre: |
| 290 | case AArch64::STRQui: |
| 291 | case AArch64::STURQi: |
| 292 | case AArch64::STRQpre: |
| 293 | case AArch64::STRBBui: |
| 294 | case AArch64::STURBBi: |
| 295 | case AArch64::STRHHui: |
| 296 | case AArch64::STURHHi: |
| 297 | case AArch64::STRWui: |
| 298 | case AArch64::STRWpre: |
| 299 | case AArch64::STURWi: |
| 300 | case AArch64::STRXui: |
| 301 | case AArch64::STRXpre: |
| 302 | case AArch64::STURXi: |
| 303 | case AArch64::STR_ZXI: |
| 304 | case AArch64::LDRDui: |
| 305 | case AArch64::LDURDi: |
| 306 | case AArch64::LDRDpre: |
| 307 | case AArch64::LDRQui: |
| 308 | case AArch64::LDURQi: |
| 309 | case AArch64::LDRQpre: |
| 310 | case AArch64::LDRWui: |
| 311 | case AArch64::LDURWi: |
| 312 | case AArch64::LDRWpre: |
| 313 | case AArch64::LDRXui: |
| 314 | case AArch64::LDURXi: |
| 315 | case AArch64::LDRXpre: |
| 316 | case AArch64::STRSui: |
| 317 | case AArch64::STURSi: |
| 318 | case AArch64::STRSpre: |
| 319 | case AArch64::LDRSui: |
| 320 | case AArch64::LDURSi: |
| 321 | case AArch64::LDRSpre: |
| 322 | case AArch64::LDR_ZXI: |
| 323 | return Opc; |
| 324 | case AArch64::LDRSWui: |
| 325 | return AArch64::LDRWui; |
| 326 | case AArch64::LDURSWi: |
| 327 | return AArch64::LDURWi; |
| 328 | case AArch64::LDRSWpre: |
| 329 | return AArch64::LDRWpre; |
| 330 | } |
| 331 | } |
| 332 | |
| 333 | static unsigned getMatchingWideOpcode(unsigned Opc) { |
| 334 | switch (Opc) { |
| 335 | default: |
| 336 | llvm_unreachable("Opcode has no wide equivalent!" ); |
| 337 | case AArch64::STRBBui: |
| 338 | return AArch64::STRHHui; |
| 339 | case AArch64::STRHHui: |
| 340 | return AArch64::STRWui; |
| 341 | case AArch64::STURBBi: |
| 342 | return AArch64::STURHHi; |
| 343 | case AArch64::STURHHi: |
| 344 | return AArch64::STURWi; |
| 345 | case AArch64::STURWi: |
| 346 | return AArch64::STURXi; |
| 347 | case AArch64::STRWui: |
| 348 | return AArch64::STRXui; |
| 349 | } |
| 350 | } |
| 351 | |
| 352 | static unsigned getMatchingPairOpcode(unsigned Opc) { |
| 353 | switch (Opc) { |
| 354 | default: |
| 355 | llvm_unreachable("Opcode has no pairwise equivalent!" ); |
| 356 | case AArch64::STRSui: |
| 357 | case AArch64::STURSi: |
| 358 | return AArch64::STPSi; |
| 359 | case AArch64::STRSpre: |
| 360 | return AArch64::STPSpre; |
| 361 | case AArch64::STRDui: |
| 362 | case AArch64::STURDi: |
| 363 | return AArch64::STPDi; |
| 364 | case AArch64::STRDpre: |
| 365 | return AArch64::STPDpre; |
| 366 | case AArch64::STRQui: |
| 367 | case AArch64::STURQi: |
| 368 | case AArch64::STR_ZXI: |
| 369 | return AArch64::STPQi; |
| 370 | case AArch64::STRQpre: |
| 371 | return AArch64::STPQpre; |
| 372 | case AArch64::STRWui: |
| 373 | case AArch64::STURWi: |
| 374 | return AArch64::STPWi; |
| 375 | case AArch64::STRWpre: |
| 376 | return AArch64::STPWpre; |
| 377 | case AArch64::STRXui: |
| 378 | case AArch64::STURXi: |
| 379 | return AArch64::STPXi; |
| 380 | case AArch64::STRXpre: |
| 381 | return AArch64::STPXpre; |
| 382 | case AArch64::LDRSui: |
| 383 | case AArch64::LDURSi: |
| 384 | return AArch64::LDPSi; |
| 385 | case AArch64::LDRSpre: |
| 386 | return AArch64::LDPSpre; |
| 387 | case AArch64::LDRDui: |
| 388 | case AArch64::LDURDi: |
| 389 | return AArch64::LDPDi; |
| 390 | case AArch64::LDRDpre: |
| 391 | return AArch64::LDPDpre; |
| 392 | case AArch64::LDRQui: |
| 393 | case AArch64::LDURQi: |
| 394 | case AArch64::LDR_ZXI: |
| 395 | return AArch64::LDPQi; |
| 396 | case AArch64::LDRQpre: |
| 397 | return AArch64::LDPQpre; |
| 398 | case AArch64::LDRWui: |
| 399 | case AArch64::LDURWi: |
| 400 | return AArch64::LDPWi; |
| 401 | case AArch64::LDRWpre: |
| 402 | return AArch64::LDPWpre; |
| 403 | case AArch64::LDRXui: |
| 404 | case AArch64::LDURXi: |
| 405 | return AArch64::LDPXi; |
| 406 | case AArch64::LDRXpre: |
| 407 | return AArch64::LDPXpre; |
| 408 | case AArch64::LDRSWui: |
| 409 | case AArch64::LDURSWi: |
| 410 | return AArch64::LDPSWi; |
| 411 | case AArch64::LDRSWpre: |
| 412 | return AArch64::LDPSWpre; |
| 413 | } |
| 414 | } |
| 415 | |
| 416 | static unsigned isMatchingStore(MachineInstr &LoadInst, |
| 417 | MachineInstr &StoreInst) { |
| 418 | unsigned LdOpc = LoadInst.getOpcode(); |
| 419 | unsigned StOpc = StoreInst.getOpcode(); |
| 420 | switch (LdOpc) { |
| 421 | default: |
| 422 | llvm_unreachable("Unsupported load instruction!" ); |
| 423 | case AArch64::LDRBBui: |
| 424 | return StOpc == AArch64::STRBBui || StOpc == AArch64::STRHHui || |
| 425 | StOpc == AArch64::STRWui || StOpc == AArch64::STRXui; |
| 426 | case AArch64::LDURBBi: |
| 427 | return StOpc == AArch64::STURBBi || StOpc == AArch64::STURHHi || |
| 428 | StOpc == AArch64::STURWi || StOpc == AArch64::STURXi; |
| 429 | case AArch64::LDRHHui: |
| 430 | return StOpc == AArch64::STRHHui || StOpc == AArch64::STRWui || |
| 431 | StOpc == AArch64::STRXui; |
| 432 | case AArch64::LDURHHi: |
| 433 | return StOpc == AArch64::STURHHi || StOpc == AArch64::STURWi || |
| 434 | StOpc == AArch64::STURXi; |
| 435 | case AArch64::LDRWui: |
| 436 | return StOpc == AArch64::STRWui || StOpc == AArch64::STRXui; |
| 437 | case AArch64::LDURWi: |
| 438 | return StOpc == AArch64::STURWi || StOpc == AArch64::STURXi; |
| 439 | case AArch64::LDRXui: |
| 440 | return StOpc == AArch64::STRXui; |
| 441 | case AArch64::LDURXi: |
| 442 | return StOpc == AArch64::STURXi; |
| 443 | } |
| 444 | } |
| 445 | |
| 446 | static unsigned getPreIndexedOpcode(unsigned Opc) { |
| 447 | // FIXME: We don't currently support creating pre-indexed loads/stores when |
| 448 | // the load or store is the unscaled version. If we decide to perform such an |
| 449 | // optimization in the future the cases for the unscaled loads/stores will |
| 450 | // need to be added here. |
| 451 | switch (Opc) { |
| 452 | default: |
| 453 | llvm_unreachable("Opcode has no pre-indexed equivalent!" ); |
| 454 | case AArch64::STRBui: |
| 455 | return AArch64::STRBpre; |
| 456 | case AArch64::STRHui: |
| 457 | return AArch64::STRHpre; |
| 458 | case AArch64::STRSui: |
| 459 | return AArch64::STRSpre; |
| 460 | case AArch64::STRDui: |
| 461 | return AArch64::STRDpre; |
| 462 | case AArch64::STRQui: |
| 463 | return AArch64::STRQpre; |
| 464 | case AArch64::STRBBui: |
| 465 | return AArch64::STRBBpre; |
| 466 | case AArch64::STRHHui: |
| 467 | return AArch64::STRHHpre; |
| 468 | case AArch64::STRWui: |
| 469 | return AArch64::STRWpre; |
| 470 | case AArch64::STRXui: |
| 471 | return AArch64::STRXpre; |
| 472 | case AArch64::LDRBui: |
| 473 | return AArch64::LDRBpre; |
| 474 | case AArch64::LDRHui: |
| 475 | return AArch64::LDRHpre; |
| 476 | case AArch64::LDRSui: |
| 477 | return AArch64::LDRSpre; |
| 478 | case AArch64::LDRDui: |
| 479 | return AArch64::LDRDpre; |
| 480 | case AArch64::LDRQui: |
| 481 | return AArch64::LDRQpre; |
| 482 | case AArch64::LDRBBui: |
| 483 | return AArch64::LDRBBpre; |
| 484 | case AArch64::LDRHHui: |
| 485 | return AArch64::LDRHHpre; |
| 486 | case AArch64::LDRWui: |
| 487 | return AArch64::LDRWpre; |
| 488 | case AArch64::LDRXui: |
| 489 | return AArch64::LDRXpre; |
| 490 | case AArch64::LDRSWui: |
| 491 | return AArch64::LDRSWpre; |
| 492 | case AArch64::LDPSi: |
| 493 | return AArch64::LDPSpre; |
| 494 | case AArch64::LDPSWi: |
| 495 | return AArch64::LDPSWpre; |
| 496 | case AArch64::LDPDi: |
| 497 | return AArch64::LDPDpre; |
| 498 | case AArch64::LDPQi: |
| 499 | return AArch64::LDPQpre; |
| 500 | case AArch64::LDPWi: |
| 501 | return AArch64::LDPWpre; |
| 502 | case AArch64::LDPXi: |
| 503 | return AArch64::LDPXpre; |
| 504 | case AArch64::STPSi: |
| 505 | return AArch64::STPSpre; |
| 506 | case AArch64::STPDi: |
| 507 | return AArch64::STPDpre; |
| 508 | case AArch64::STPQi: |
| 509 | return AArch64::STPQpre; |
| 510 | case AArch64::STPWi: |
| 511 | return AArch64::STPWpre; |
| 512 | case AArch64::STPXi: |
| 513 | return AArch64::STPXpre; |
| 514 | case AArch64::STGi: |
| 515 | return AArch64::STGPreIndex; |
| 516 | case AArch64::STZGi: |
| 517 | return AArch64::STZGPreIndex; |
| 518 | case AArch64::ST2Gi: |
| 519 | return AArch64::ST2GPreIndex; |
| 520 | case AArch64::STZ2Gi: |
| 521 | return AArch64::STZ2GPreIndex; |
| 522 | case AArch64::STGPi: |
| 523 | return AArch64::STGPpre; |
| 524 | } |
| 525 | } |
| 526 | |
| 527 | static unsigned getBaseAddressOpcode(unsigned Opc) { |
| 528 | // TODO: Add more index address stores. |
| 529 | switch (Opc) { |
| 530 | default: |
| 531 | llvm_unreachable("Opcode has no base address equivalent!" ); |
| 532 | case AArch64::LDRBroX: |
| 533 | return AArch64::LDRBui; |
| 534 | case AArch64::LDRBBroX: |
| 535 | return AArch64::LDRBBui; |
| 536 | case AArch64::LDRSBXroX: |
| 537 | return AArch64::LDRSBXui; |
| 538 | case AArch64::LDRSBWroX: |
| 539 | return AArch64::LDRSBWui; |
| 540 | case AArch64::LDRHroX: |
| 541 | return AArch64::LDRHui; |
| 542 | case AArch64::LDRHHroX: |
| 543 | return AArch64::LDRHHui; |
| 544 | case AArch64::LDRSHXroX: |
| 545 | return AArch64::LDRSHXui; |
| 546 | case AArch64::LDRSHWroX: |
| 547 | return AArch64::LDRSHWui; |
| 548 | case AArch64::LDRWroX: |
| 549 | return AArch64::LDRWui; |
| 550 | case AArch64::LDRSroX: |
| 551 | return AArch64::LDRSui; |
| 552 | case AArch64::LDRSWroX: |
| 553 | return AArch64::LDRSWui; |
| 554 | case AArch64::LDRDroX: |
| 555 | return AArch64::LDRDui; |
| 556 | case AArch64::LDRXroX: |
| 557 | return AArch64::LDRXui; |
| 558 | case AArch64::LDRQroX: |
| 559 | return AArch64::LDRQui; |
| 560 | } |
| 561 | } |
| 562 | |
| 563 | static unsigned getPostIndexedOpcode(unsigned Opc) { |
| 564 | switch (Opc) { |
| 565 | default: |
| 566 | llvm_unreachable("Opcode has no post-indexed wise equivalent!" ); |
| 567 | case AArch64::STRBui: |
| 568 | return AArch64::STRBpost; |
| 569 | case AArch64::STRHui: |
| 570 | return AArch64::STRHpost; |
| 571 | case AArch64::STRSui: |
| 572 | case AArch64::STURSi: |
| 573 | return AArch64::STRSpost; |
| 574 | case AArch64::STRDui: |
| 575 | case AArch64::STURDi: |
| 576 | return AArch64::STRDpost; |
| 577 | case AArch64::STRQui: |
| 578 | case AArch64::STURQi: |
| 579 | return AArch64::STRQpost; |
| 580 | case AArch64::STRBBui: |
| 581 | return AArch64::STRBBpost; |
| 582 | case AArch64::STRHHui: |
| 583 | return AArch64::STRHHpost; |
| 584 | case AArch64::STRWui: |
| 585 | case AArch64::STURWi: |
| 586 | return AArch64::STRWpost; |
| 587 | case AArch64::STRXui: |
| 588 | case AArch64::STURXi: |
| 589 | return AArch64::STRXpost; |
| 590 | case AArch64::LDRBui: |
| 591 | return AArch64::LDRBpost; |
| 592 | case AArch64::LDRHui: |
| 593 | return AArch64::LDRHpost; |
| 594 | case AArch64::LDRSui: |
| 595 | case AArch64::LDURSi: |
| 596 | return AArch64::LDRSpost; |
| 597 | case AArch64::LDRDui: |
| 598 | case AArch64::LDURDi: |
| 599 | return AArch64::LDRDpost; |
| 600 | case AArch64::LDRQui: |
| 601 | case AArch64::LDURQi: |
| 602 | return AArch64::LDRQpost; |
| 603 | case AArch64::LDRBBui: |
| 604 | return AArch64::LDRBBpost; |
| 605 | case AArch64::LDRHHui: |
| 606 | return AArch64::LDRHHpost; |
| 607 | case AArch64::LDRWui: |
| 608 | case AArch64::LDURWi: |
| 609 | return AArch64::LDRWpost; |
| 610 | case AArch64::LDRXui: |
| 611 | case AArch64::LDURXi: |
| 612 | return AArch64::LDRXpost; |
| 613 | case AArch64::LDRSWui: |
| 614 | return AArch64::LDRSWpost; |
| 615 | case AArch64::LDPSi: |
| 616 | return AArch64::LDPSpost; |
| 617 | case AArch64::LDPSWi: |
| 618 | return AArch64::LDPSWpost; |
| 619 | case AArch64::LDPDi: |
| 620 | return AArch64::LDPDpost; |
| 621 | case AArch64::LDPQi: |
| 622 | return AArch64::LDPQpost; |
| 623 | case AArch64::LDPWi: |
| 624 | return AArch64::LDPWpost; |
| 625 | case AArch64::LDPXi: |
| 626 | return AArch64::LDPXpost; |
| 627 | case AArch64::STPSi: |
| 628 | return AArch64::STPSpost; |
| 629 | case AArch64::STPDi: |
| 630 | return AArch64::STPDpost; |
| 631 | case AArch64::STPQi: |
| 632 | return AArch64::STPQpost; |
| 633 | case AArch64::STPWi: |
| 634 | return AArch64::STPWpost; |
| 635 | case AArch64::STPXi: |
| 636 | return AArch64::STPXpost; |
| 637 | case AArch64::STGi: |
| 638 | return AArch64::STGPostIndex; |
| 639 | case AArch64::STZGi: |
| 640 | return AArch64::STZGPostIndex; |
| 641 | case AArch64::ST2Gi: |
| 642 | return AArch64::ST2GPostIndex; |
| 643 | case AArch64::STZ2Gi: |
| 644 | return AArch64::STZ2GPostIndex; |
| 645 | case AArch64::STGPi: |
| 646 | return AArch64::STGPpost; |
| 647 | } |
| 648 | } |
| 649 | |
| 650 | static bool isPreLdStPairCandidate(MachineInstr &FirstMI, MachineInstr &MI) { |
| 651 | |
| 652 | unsigned OpcA = FirstMI.getOpcode(); |
| 653 | unsigned OpcB = MI.getOpcode(); |
| 654 | |
| 655 | switch (OpcA) { |
| 656 | default: |
| 657 | return false; |
| 658 | case AArch64::STRSpre: |
| 659 | return (OpcB == AArch64::STRSui) || (OpcB == AArch64::STURSi); |
| 660 | case AArch64::STRDpre: |
| 661 | return (OpcB == AArch64::STRDui) || (OpcB == AArch64::STURDi); |
| 662 | case AArch64::STRQpre: |
| 663 | return (OpcB == AArch64::STRQui) || (OpcB == AArch64::STURQi); |
| 664 | case AArch64::STRWpre: |
| 665 | return (OpcB == AArch64::STRWui) || (OpcB == AArch64::STURWi); |
| 666 | case AArch64::STRXpre: |
| 667 | return (OpcB == AArch64::STRXui) || (OpcB == AArch64::STURXi); |
| 668 | case AArch64::LDRSpre: |
| 669 | return (OpcB == AArch64::LDRSui) || (OpcB == AArch64::LDURSi); |
| 670 | case AArch64::LDRDpre: |
| 671 | return (OpcB == AArch64::LDRDui) || (OpcB == AArch64::LDURDi); |
| 672 | case AArch64::LDRQpre: |
| 673 | return (OpcB == AArch64::LDRQui) || (OpcB == AArch64::LDURQi); |
| 674 | case AArch64::LDRWpre: |
| 675 | return (OpcB == AArch64::LDRWui) || (OpcB == AArch64::LDURWi); |
| 676 | case AArch64::LDRXpre: |
| 677 | return (OpcB == AArch64::LDRXui) || (OpcB == AArch64::LDURXi); |
| 678 | case AArch64::LDRSWpre: |
| 679 | return (OpcB == AArch64::LDRSWui) || (OpcB == AArch64::LDURSWi); |
| 680 | } |
| 681 | } |
| 682 | |
| 683 | // Returns the scale and offset range of pre/post indexed variants of MI. |
| 684 | static void getPrePostIndexedMemOpInfo(const MachineInstr &MI, int &Scale, |
| 685 | int &MinOffset, int &MaxOffset) { |
| 686 | bool IsPaired = AArch64InstrInfo::isPairedLdSt(MI); |
| 687 | bool IsTagStore = isTagStore(MI); |
| 688 | // ST*G and all paired ldst have the same scale in pre/post-indexed variants |
| 689 | // as in the "unsigned offset" variant. |
| 690 | // All other pre/post indexed ldst instructions are unscaled. |
| 691 | Scale = (IsTagStore || IsPaired) ? AArch64InstrInfo::getMemScale(MI) : 1; |
| 692 | |
| 693 | if (IsPaired) { |
| 694 | MinOffset = -64; |
| 695 | MaxOffset = 63; |
| 696 | } else { |
| 697 | MinOffset = -256; |
| 698 | MaxOffset = 255; |
| 699 | } |
| 700 | } |
| 701 | |
| 702 | static MachineOperand &getLdStRegOp(MachineInstr &MI, |
| 703 | unsigned PairedRegOp = 0) { |
| 704 | assert(PairedRegOp < 2 && "Unexpected register operand idx." ); |
| 705 | bool IsPreLdSt = AArch64InstrInfo::isPreLdSt(MI); |
| 706 | if (IsPreLdSt) |
| 707 | PairedRegOp += 1; |
| 708 | unsigned Idx = |
| 709 | AArch64InstrInfo::isPairedLdSt(MI) || IsPreLdSt ? PairedRegOp : 0; |
| 710 | return MI.getOperand(i: Idx); |
| 711 | } |
| 712 | |
| 713 | static bool isLdOffsetInRangeOfSt(MachineInstr &LoadInst, |
| 714 | MachineInstr &StoreInst, |
| 715 | const AArch64InstrInfo *TII) { |
| 716 | assert(isMatchingStore(LoadInst, StoreInst) && "Expect only matched ld/st." ); |
| 717 | int LoadSize = TII->getMemScale(MI: LoadInst); |
| 718 | int StoreSize = TII->getMemScale(MI: StoreInst); |
| 719 | int UnscaledStOffset = |
| 720 | TII->hasUnscaledLdStOffset(MI&: StoreInst) |
| 721 | ? AArch64InstrInfo::getLdStOffsetOp(MI: StoreInst).getImm() |
| 722 | : AArch64InstrInfo::getLdStOffsetOp(MI: StoreInst).getImm() * StoreSize; |
| 723 | int UnscaledLdOffset = |
| 724 | TII->hasUnscaledLdStOffset(MI&: LoadInst) |
| 725 | ? AArch64InstrInfo::getLdStOffsetOp(MI: LoadInst).getImm() |
| 726 | : AArch64InstrInfo::getLdStOffsetOp(MI: LoadInst).getImm() * LoadSize; |
| 727 | return (UnscaledStOffset <= UnscaledLdOffset) && |
| 728 | (UnscaledLdOffset + LoadSize <= (UnscaledStOffset + StoreSize)); |
| 729 | } |
| 730 | |
| 731 | static bool isPromotableZeroStoreInst(MachineInstr &MI) { |
| 732 | unsigned Opc = MI.getOpcode(); |
| 733 | return (Opc == AArch64::STRWui || Opc == AArch64::STURWi || |
| 734 | isNarrowStore(Opc)) && |
| 735 | getLdStRegOp(MI).getReg() == AArch64::WZR; |
| 736 | } |
| 737 | |
| 738 | static bool isPromotableLoadFromStore(MachineInstr &MI) { |
| 739 | switch (MI.getOpcode()) { |
| 740 | default: |
| 741 | return false; |
| 742 | // Scaled instructions. |
| 743 | case AArch64::LDRBBui: |
| 744 | case AArch64::LDRHHui: |
| 745 | case AArch64::LDRWui: |
| 746 | case AArch64::LDRXui: |
| 747 | // Unscaled instructions. |
| 748 | case AArch64::LDURBBi: |
| 749 | case AArch64::LDURHHi: |
| 750 | case AArch64::LDURWi: |
| 751 | case AArch64::LDURXi: |
| 752 | return true; |
| 753 | } |
| 754 | } |
| 755 | |
| 756 | static bool isMergeableLdStUpdate(MachineInstr &MI, AArch64FunctionInfo &AFI) { |
| 757 | unsigned Opc = MI.getOpcode(); |
| 758 | switch (Opc) { |
| 759 | default: |
| 760 | return false; |
| 761 | // Scaled instructions. |
| 762 | case AArch64::STRBui: |
| 763 | case AArch64::STRHui: |
| 764 | case AArch64::STRSui: |
| 765 | case AArch64::STRDui: |
| 766 | case AArch64::STRQui: |
| 767 | case AArch64::STRXui: |
| 768 | case AArch64::STRWui: |
| 769 | case AArch64::STRHHui: |
| 770 | case AArch64::STRBBui: |
| 771 | case AArch64::LDRBui: |
| 772 | case AArch64::LDRHui: |
| 773 | case AArch64::LDRSui: |
| 774 | case AArch64::LDRDui: |
| 775 | case AArch64::LDRQui: |
| 776 | case AArch64::LDRXui: |
| 777 | case AArch64::LDRWui: |
| 778 | case AArch64::LDRHHui: |
| 779 | case AArch64::LDRBBui: |
| 780 | case AArch64::STGi: |
| 781 | case AArch64::STZGi: |
| 782 | case AArch64::ST2Gi: |
| 783 | case AArch64::STZ2Gi: |
| 784 | case AArch64::STGPi: |
| 785 | // Unscaled instructions. |
| 786 | case AArch64::STURSi: |
| 787 | case AArch64::STURDi: |
| 788 | case AArch64::STURQi: |
| 789 | case AArch64::STURWi: |
| 790 | case AArch64::STURXi: |
| 791 | case AArch64::LDURSi: |
| 792 | case AArch64::LDURDi: |
| 793 | case AArch64::LDURQi: |
| 794 | case AArch64::LDURWi: |
| 795 | case AArch64::LDURXi: |
| 796 | // Paired instructions. |
| 797 | case AArch64::LDPSi: |
| 798 | case AArch64::LDPSWi: |
| 799 | case AArch64::LDPDi: |
| 800 | case AArch64::LDPQi: |
| 801 | case AArch64::LDPWi: |
| 802 | case AArch64::LDPXi: |
| 803 | case AArch64::STPSi: |
| 804 | case AArch64::STPDi: |
| 805 | case AArch64::STPQi: |
| 806 | case AArch64::STPWi: |
| 807 | case AArch64::STPXi: |
| 808 | // Make sure this is a reg+imm (as opposed to an address reloc). |
| 809 | if (!AArch64InstrInfo::getLdStOffsetOp(MI).isImm()) |
| 810 | return false; |
| 811 | |
| 812 | // When using stack tagging, simple sp+imm loads and stores are not |
| 813 | // tag-checked, but pre- and post-indexed versions of them are, so we can't |
| 814 | // replace the former with the latter. This transformation would be valid |
| 815 | // if the load/store accesses an untagged stack slot, but we don't have |
| 816 | // that information available after frame indices have been eliminated. |
| 817 | if (AFI.isMTETagged() && |
| 818 | AArch64InstrInfo::getLdStBaseOp(MI).getReg() == AArch64::SP) |
| 819 | return false; |
| 820 | |
| 821 | return true; |
| 822 | } |
| 823 | } |
| 824 | |
| 825 | // Make sure this is a reg+reg Ld/St |
| 826 | static bool isMergeableIndexLdSt(MachineInstr &MI, int &Scale) { |
| 827 | unsigned Opc = MI.getOpcode(); |
| 828 | switch (Opc) { |
| 829 | default: |
| 830 | return false; |
| 831 | // Scaled instructions. |
| 832 | // TODO: Add more index address stores. |
| 833 | case AArch64::LDRBroX: |
| 834 | case AArch64::LDRBBroX: |
| 835 | case AArch64::LDRSBXroX: |
| 836 | case AArch64::LDRSBWroX: |
| 837 | Scale = 1; |
| 838 | return true; |
| 839 | case AArch64::LDRHroX: |
| 840 | case AArch64::LDRHHroX: |
| 841 | case AArch64::LDRSHXroX: |
| 842 | case AArch64::LDRSHWroX: |
| 843 | Scale = 2; |
| 844 | return true; |
| 845 | case AArch64::LDRWroX: |
| 846 | case AArch64::LDRSroX: |
| 847 | case AArch64::LDRSWroX: |
| 848 | Scale = 4; |
| 849 | return true; |
| 850 | case AArch64::LDRDroX: |
| 851 | case AArch64::LDRXroX: |
| 852 | Scale = 8; |
| 853 | return true; |
| 854 | case AArch64::LDRQroX: |
| 855 | Scale = 16; |
| 856 | return true; |
| 857 | } |
| 858 | } |
| 859 | |
| 860 | static bool isRewritableImplicitDef(const MachineOperand &MO) { |
| 861 | switch (MO.getParent()->getOpcode()) { |
| 862 | default: |
| 863 | return MO.isRenamable(); |
| 864 | case AArch64::ORRWrs: |
| 865 | case AArch64::ADDWri: |
| 866 | return true; |
| 867 | } |
| 868 | } |
| 869 | |
| 870 | MachineBasicBlock::iterator |
| 871 | AArch64LoadStoreOpt::mergeNarrowZeroStores(MachineBasicBlock::iterator I, |
| 872 | MachineBasicBlock::iterator MergeMI, |
| 873 | const LdStPairFlags &Flags) { |
| 874 | assert(isPromotableZeroStoreInst(*I) && isPromotableZeroStoreInst(*MergeMI) && |
| 875 | "Expected promotable zero stores." ); |
| 876 | |
| 877 | MachineBasicBlock::iterator E = I->getParent()->end(); |
| 878 | MachineBasicBlock::iterator NextI = next_nodbg(It: I, End: E); |
| 879 | // If NextI is the second of the two instructions to be merged, we need |
| 880 | // to skip one further. Either way we merge will invalidate the iterator, |
| 881 | // and we don't need to scan the new instruction, as it's a pairwise |
| 882 | // instruction, which we're not considering for further action anyway. |
| 883 | if (NextI == MergeMI) |
| 884 | NextI = next_nodbg(It: NextI, End: E); |
| 885 | |
| 886 | unsigned Opc = I->getOpcode(); |
| 887 | unsigned MergeMIOpc = MergeMI->getOpcode(); |
| 888 | bool IsScaled = !TII->hasUnscaledLdStOffset(Opc); |
| 889 | bool IsMergedMIScaled = !TII->hasUnscaledLdStOffset(Opc: MergeMIOpc); |
| 890 | int OffsetStride = IsScaled ? TII->getMemScale(MI: *I) : 1; |
| 891 | int MergeMIOffsetStride = IsMergedMIScaled ? TII->getMemScale(MI: *MergeMI) : 1; |
| 892 | |
| 893 | bool MergeForward = Flags.getMergeForward(); |
| 894 | // Insert our new paired instruction after whichever of the paired |
| 895 | // instructions MergeForward indicates. |
| 896 | MachineBasicBlock::iterator InsertionPoint = MergeForward ? MergeMI : I; |
| 897 | // Also based on MergeForward is from where we copy the base register operand |
| 898 | // so we get the flags compatible with the input code. |
| 899 | const MachineOperand &BaseRegOp = |
| 900 | MergeForward ? AArch64InstrInfo::getLdStBaseOp(MI: *MergeMI) |
| 901 | : AArch64InstrInfo::getLdStBaseOp(MI: *I); |
| 902 | |
| 903 | // Which register is Rt and which is Rt2 depends on the offset order. |
| 904 | int64_t IOffsetInBytes = |
| 905 | AArch64InstrInfo::getLdStOffsetOp(MI: *I).getImm() * OffsetStride; |
| 906 | int64_t MIOffsetInBytes = |
| 907 | AArch64InstrInfo::getLdStOffsetOp(MI: *MergeMI).getImm() * |
| 908 | MergeMIOffsetStride; |
| 909 | // Select final offset based on the offset order. |
| 910 | int64_t OffsetImm; |
| 911 | if (IOffsetInBytes > MIOffsetInBytes) |
| 912 | OffsetImm = MIOffsetInBytes; |
| 913 | else |
| 914 | OffsetImm = IOffsetInBytes; |
| 915 | |
| 916 | int NewOpcode = getMatchingWideOpcode(Opc); |
| 917 | // Adjust final offset on scaled stores because the new instruction |
| 918 | // has a different scale. |
| 919 | if (!TII->hasUnscaledLdStOffset(Opc: NewOpcode)) { |
| 920 | int NewOffsetStride = TII->getMemScale(Opc: NewOpcode); |
| 921 | assert(((OffsetImm % NewOffsetStride) == 0) && |
| 922 | "Offset should be a multiple of the store memory scale" ); |
| 923 | OffsetImm = OffsetImm / NewOffsetStride; |
| 924 | } |
| 925 | |
| 926 | // Construct the new instruction. |
| 927 | DebugLoc DL = I->getDebugLoc(); |
| 928 | MachineBasicBlock *MBB = I->getParent(); |
| 929 | MachineInstrBuilder MIB; |
| 930 | MIB = BuildMI(BB&: *MBB, I: InsertionPoint, MIMD: DL, MCID: TII->get(Opcode: NewOpcode)) |
| 931 | .addReg(RegNo: isNarrowStore(Opc) ? AArch64::WZR : AArch64::XZR) |
| 932 | .add(MO: BaseRegOp) |
| 933 | .addImm(Val: OffsetImm) |
| 934 | .cloneMergedMemRefs(OtherMIs: {&*I, &*MergeMI}) |
| 935 | .setMIFlags(I->mergeFlagsWith(Other: *MergeMI)); |
| 936 | (void)MIB; |
| 937 | |
| 938 | LLVM_DEBUG(dbgs() << "Creating wider store. Replacing instructions:\n " ); |
| 939 | LLVM_DEBUG(I->print(dbgs())); |
| 940 | LLVM_DEBUG(dbgs() << " " ); |
| 941 | LLVM_DEBUG(MergeMI->print(dbgs())); |
| 942 | LLVM_DEBUG(dbgs() << " with instruction:\n " ); |
| 943 | LLVM_DEBUG(((MachineInstr *)MIB)->print(dbgs())); |
| 944 | LLVM_DEBUG(dbgs() << "\n" ); |
| 945 | |
| 946 | // Erase the old instructions. |
| 947 | I->eraseFromParent(); |
| 948 | MergeMI->eraseFromParent(); |
| 949 | return NextI; |
| 950 | } |
| 951 | |
| 952 | // Apply Fn to all instructions between MI and the beginning of the block, until |
| 953 | // a def for DefReg is reached. Returns true, iff Fn returns true for all |
| 954 | // visited instructions. Stop after visiting Limit iterations. |
| 955 | static bool forAllMIsUntilDef(MachineInstr &MI, MCPhysReg DefReg, |
| 956 | const TargetRegisterInfo *TRI, unsigned Limit, |
| 957 | std::function<bool(MachineInstr &, bool)> &Fn) { |
| 958 | auto MBB = MI.getParent(); |
| 959 | for (MachineInstr &I : |
| 960 | instructionsWithoutDebug(It: MI.getReverseIterator(), End: MBB->instr_rend())) { |
| 961 | if (!Limit) |
| 962 | return false; |
| 963 | --Limit; |
| 964 | |
| 965 | bool isDef = any_of(Range: I.operands(), P: [DefReg, TRI](MachineOperand &MOP) { |
| 966 | return MOP.isReg() && MOP.isDef() && !MOP.isDebug() && MOP.getReg() && |
| 967 | TRI->regsOverlap(RegA: MOP.getReg(), RegB: DefReg); |
| 968 | }); |
| 969 | if (!Fn(I, isDef)) |
| 970 | return false; |
| 971 | if (isDef) |
| 972 | break; |
| 973 | } |
| 974 | return true; |
| 975 | } |
| 976 | |
| 977 | static void updateDefinedRegisters(MachineInstr &MI, LiveRegUnits &Units, |
| 978 | const TargetRegisterInfo *TRI) { |
| 979 | |
| 980 | for (const MachineOperand &MOP : phys_regs_and_masks(MI)) |
| 981 | if (MOP.isReg() && MOP.isKill()) |
| 982 | Units.removeReg(Reg: MOP.getReg()); |
| 983 | |
| 984 | for (const MachineOperand &MOP : phys_regs_and_masks(MI)) |
| 985 | if (MOP.isReg() && !MOP.isKill()) |
| 986 | Units.addReg(Reg: MOP.getReg()); |
| 987 | } |
| 988 | |
| 989 | /// This function will add a new entry into the debugValueSubstitutions table |
| 990 | /// when two instruction have been merged into a new one represented by \p |
| 991 | /// MergedInstr. |
| 992 | static void addDebugSubstitutionsToTable(MachineFunction *MF, |
| 993 | unsigned InstrNumToSet, |
| 994 | MachineInstr &OriginalInstr, |
| 995 | MachineInstr &MergedInstr) { |
| 996 | |
| 997 | // Figure out the Operand Index of the destination register of the |
| 998 | // OriginalInstr in the new MergedInstr. |
| 999 | auto Reg = OriginalInstr.getOperand(i: 0).getReg(); |
| 1000 | unsigned OperandNo = 0; |
| 1001 | bool RegFound = false; |
| 1002 | for (const auto Op : MergedInstr.operands()) { |
| 1003 | if (Op.getReg() == Reg) { |
| 1004 | RegFound = true; |
| 1005 | break; |
| 1006 | } |
| 1007 | OperandNo++; |
| 1008 | } |
| 1009 | |
| 1010 | if (RegFound) |
| 1011 | MF->makeDebugValueSubstitution({OriginalInstr.peekDebugInstrNum(), 0}, |
| 1012 | {InstrNumToSet, OperandNo}); |
| 1013 | } |
| 1014 | |
| 1015 | MachineBasicBlock::iterator |
| 1016 | AArch64LoadStoreOpt::mergePairedInsns(MachineBasicBlock::iterator I, |
| 1017 | MachineBasicBlock::iterator Paired, |
| 1018 | const LdStPairFlags &Flags) { |
| 1019 | MachineBasicBlock::iterator E = I->getParent()->end(); |
| 1020 | MachineBasicBlock::iterator NextI = next_nodbg(It: I, End: E); |
| 1021 | // If NextI is the second of the two instructions to be merged, we need |
| 1022 | // to skip one further. Either way we merge will invalidate the iterator, |
| 1023 | // and we don't need to scan the new instruction, as it's a pairwise |
| 1024 | // instruction, which we're not considering for further action anyway. |
| 1025 | if (NextI == Paired) |
| 1026 | NextI = next_nodbg(It: NextI, End: E); |
| 1027 | |
| 1028 | int SExtIdx = Flags.getSExtIdx(); |
| 1029 | unsigned Opc = |
| 1030 | SExtIdx == -1 ? I->getOpcode() : getMatchingNonSExtOpcode(Opc: I->getOpcode()); |
| 1031 | bool IsUnscaled = TII->hasUnscaledLdStOffset(Opc); |
| 1032 | int OffsetStride = IsUnscaled ? TII->getMemScale(MI: *I) : 1; |
| 1033 | |
| 1034 | bool MergeForward = Flags.getMergeForward(); |
| 1035 | |
| 1036 | std::optional<MCPhysReg> RenameReg = Flags.getRenameReg(); |
| 1037 | if (RenameReg) { |
| 1038 | MCRegister RegToRename = getLdStRegOp(MI&: *I).getReg(); |
| 1039 | DefinedInBB.addReg(Reg: *RenameReg); |
| 1040 | |
| 1041 | // Return the sub/super register for RenameReg, matching the size of |
| 1042 | // OriginalReg. |
| 1043 | auto GetMatchingSubReg = |
| 1044 | [this, RenameReg](const TargetRegisterClass *C) -> MCPhysReg { |
| 1045 | for (MCPhysReg SubOrSuper : |
| 1046 | TRI->sub_and_superregs_inclusive(Reg: *RenameReg)) { |
| 1047 | if (C->contains(Reg: SubOrSuper)) |
| 1048 | return SubOrSuper; |
| 1049 | } |
| 1050 | llvm_unreachable("Should have found matching sub or super register!" ); |
| 1051 | }; |
| 1052 | |
| 1053 | std::function<bool(MachineInstr &, bool)> UpdateMIs = |
| 1054 | [this, RegToRename, GetMatchingSubReg, MergeForward](MachineInstr &MI, |
| 1055 | bool IsDef) { |
| 1056 | if (IsDef) { |
| 1057 | bool SeenDef = false; |
| 1058 | for (unsigned OpIdx = 0; OpIdx < MI.getNumOperands(); ++OpIdx) { |
| 1059 | MachineOperand &MOP = MI.getOperand(i: OpIdx); |
| 1060 | // Rename the first explicit definition and all implicit |
| 1061 | // definitions matching RegToRename. |
| 1062 | if (MOP.isReg() && !MOP.isDebug() && MOP.getReg() && |
| 1063 | (!MergeForward || !SeenDef || |
| 1064 | (MOP.isDef() && MOP.isImplicit())) && |
| 1065 | TRI->regsOverlap(RegA: MOP.getReg(), RegB: RegToRename)) { |
| 1066 | assert((MOP.isImplicit() || |
| 1067 | (MOP.isRenamable() && !MOP.isEarlyClobber())) && |
| 1068 | "Need renamable operands" ); |
| 1069 | Register MatchingReg; |
| 1070 | if (const TargetRegisterClass *RC = |
| 1071 | MI.getRegClassConstraint(OpIdx, TII, TRI)) |
| 1072 | MatchingReg = GetMatchingSubReg(RC); |
| 1073 | else { |
| 1074 | if (!isRewritableImplicitDef(MO: MOP)) |
| 1075 | continue; |
| 1076 | MatchingReg = GetMatchingSubReg( |
| 1077 | TRI->getMinimalPhysRegClass(Reg: MOP.getReg())); |
| 1078 | } |
| 1079 | MOP.setReg(MatchingReg); |
| 1080 | SeenDef = true; |
| 1081 | } |
| 1082 | } |
| 1083 | } else { |
| 1084 | for (unsigned OpIdx = 0; OpIdx < MI.getNumOperands(); ++OpIdx) { |
| 1085 | MachineOperand &MOP = MI.getOperand(i: OpIdx); |
| 1086 | if (MOP.isReg() && !MOP.isDebug() && MOP.getReg() && |
| 1087 | TRI->regsOverlap(RegA: MOP.getReg(), RegB: RegToRename)) { |
| 1088 | assert((MOP.isImplicit() || |
| 1089 | (MOP.isRenamable() && !MOP.isEarlyClobber())) && |
| 1090 | "Need renamable operands" ); |
| 1091 | Register MatchingReg; |
| 1092 | if (const TargetRegisterClass *RC = |
| 1093 | MI.getRegClassConstraint(OpIdx, TII, TRI)) |
| 1094 | MatchingReg = GetMatchingSubReg(RC); |
| 1095 | else |
| 1096 | MatchingReg = GetMatchingSubReg( |
| 1097 | TRI->getMinimalPhysRegClass(Reg: MOP.getReg())); |
| 1098 | assert(MatchingReg != AArch64::NoRegister && |
| 1099 | "Cannot find matching regs for renaming" ); |
| 1100 | MOP.setReg(MatchingReg); |
| 1101 | } |
| 1102 | } |
| 1103 | } |
| 1104 | LLVM_DEBUG(dbgs() << "Renamed " << MI); |
| 1105 | return true; |
| 1106 | }; |
| 1107 | forAllMIsUntilDef(MI&: MergeForward ? *I : *Paired->getPrevNode(), DefReg: RegToRename, |
| 1108 | TRI, UINT32_MAX, Fn&: UpdateMIs); |
| 1109 | |
| 1110 | #if !defined(NDEBUG) |
| 1111 | // For forward merging store: |
| 1112 | // Make sure the register used for renaming is not used between the |
| 1113 | // paired instructions. That would trash the content before the new |
| 1114 | // paired instruction. |
| 1115 | MCPhysReg RegToCheck = *RenameReg; |
| 1116 | // For backward merging load: |
| 1117 | // Make sure the register being renamed is not used between the |
| 1118 | // paired instructions. That would trash the content after the new |
| 1119 | // paired instruction. |
| 1120 | if (!MergeForward) |
| 1121 | RegToCheck = RegToRename; |
| 1122 | for (auto &MI : |
| 1123 | iterator_range<MachineInstrBundleIterator<llvm::MachineInstr>>( |
| 1124 | MergeForward ? std::next(I) : I, |
| 1125 | MergeForward ? std::next(Paired) : Paired)) |
| 1126 | assert(all_of(MI.operands(), |
| 1127 | [this, RegToCheck](const MachineOperand &MOP) { |
| 1128 | return !MOP.isReg() || MOP.isDebug() || !MOP.getReg() || |
| 1129 | MOP.isUndef() || |
| 1130 | !TRI->regsOverlap(MOP.getReg(), RegToCheck); |
| 1131 | }) && |
| 1132 | "Rename register used between paired instruction, trashing the " |
| 1133 | "content" ); |
| 1134 | #endif |
| 1135 | } |
| 1136 | |
| 1137 | // Insert our new paired instruction after whichever of the paired |
| 1138 | // instructions MergeForward indicates. |
| 1139 | MachineBasicBlock::iterator InsertionPoint = MergeForward ? Paired : I; |
| 1140 | // Also based on MergeForward is from where we copy the base register operand |
| 1141 | // so we get the flags compatible with the input code. |
| 1142 | const MachineOperand &BaseRegOp = |
| 1143 | MergeForward ? AArch64InstrInfo::getLdStBaseOp(MI: *Paired) |
| 1144 | : AArch64InstrInfo::getLdStBaseOp(MI: *I); |
| 1145 | |
| 1146 | int Offset = AArch64InstrInfo::getLdStOffsetOp(MI: *I).getImm(); |
| 1147 | int PairedOffset = AArch64InstrInfo::getLdStOffsetOp(MI: *Paired).getImm(); |
| 1148 | bool PairedIsUnscaled = TII->hasUnscaledLdStOffset(Opc: Paired->getOpcode()); |
| 1149 | if (IsUnscaled != PairedIsUnscaled) { |
| 1150 | // We're trying to pair instructions that differ in how they are scaled. If |
| 1151 | // I is scaled then scale the offset of Paired accordingly. Otherwise, do |
| 1152 | // the opposite (i.e., make Paired's offset unscaled). |
| 1153 | int MemSize = TII->getMemScale(MI: *Paired); |
| 1154 | if (PairedIsUnscaled) { |
| 1155 | // If the unscaled offset isn't a multiple of the MemSize, we can't |
| 1156 | // pair the operations together. |
| 1157 | assert(!(PairedOffset % TII->getMemScale(*Paired)) && |
| 1158 | "Offset should be a multiple of the stride!" ); |
| 1159 | PairedOffset /= MemSize; |
| 1160 | } else { |
| 1161 | PairedOffset *= MemSize; |
| 1162 | } |
| 1163 | } |
| 1164 | |
| 1165 | // Which register is Rt and which is Rt2 depends on the offset order. |
| 1166 | // However, for pre load/stores the Rt should be the one of the pre |
| 1167 | // load/store. |
| 1168 | MachineInstr *RtMI, *Rt2MI; |
| 1169 | if (Offset == PairedOffset + OffsetStride && |
| 1170 | !AArch64InstrInfo::isPreLdSt(MI: *I)) { |
| 1171 | RtMI = &*Paired; |
| 1172 | Rt2MI = &*I; |
| 1173 | // Here we swapped the assumption made for SExtIdx. |
| 1174 | // I.e., we turn ldp I, Paired into ldp Paired, I. |
| 1175 | // Update the index accordingly. |
| 1176 | if (SExtIdx != -1) |
| 1177 | SExtIdx = (SExtIdx + 1) % 2; |
| 1178 | } else { |
| 1179 | RtMI = &*I; |
| 1180 | Rt2MI = &*Paired; |
| 1181 | } |
| 1182 | int OffsetImm = AArch64InstrInfo::getLdStOffsetOp(MI: *RtMI).getImm(); |
| 1183 | // Scale the immediate offset, if necessary. |
| 1184 | if (TII->hasUnscaledLdStOffset(Opc: RtMI->getOpcode())) { |
| 1185 | assert(!(OffsetImm % TII->getMemScale(*RtMI)) && |
| 1186 | "Unscaled offset cannot be scaled." ); |
| 1187 | OffsetImm /= TII->getMemScale(MI: *RtMI); |
| 1188 | } |
| 1189 | |
| 1190 | // Construct the new instruction. |
| 1191 | MachineInstrBuilder MIB; |
| 1192 | DebugLoc DL = I->getDebugLoc(); |
| 1193 | MachineBasicBlock *MBB = I->getParent(); |
| 1194 | MachineOperand RegOp0 = getLdStRegOp(MI&: *RtMI); |
| 1195 | MachineOperand RegOp1 = getLdStRegOp(MI&: *Rt2MI); |
| 1196 | MachineOperand &PairedRegOp = RtMI == &*Paired ? RegOp0 : RegOp1; |
| 1197 | // Kill flags may become invalid when moving stores for pairing. |
| 1198 | if (RegOp0.isUse()) { |
| 1199 | if (!MergeForward) { |
| 1200 | // Clear kill flags on store if moving upwards. Example: |
| 1201 | // STRWui kill %w0, ... |
| 1202 | // USE %w1 |
| 1203 | // STRWui kill %w1 ; need to clear kill flag when moving STRWui upwards |
| 1204 | // We are about to move the store of w1, so its kill flag may become |
| 1205 | // invalid; not the case for w0. |
| 1206 | // Since w1 is used between the stores, the kill flag on w1 is cleared |
| 1207 | // after merging. |
| 1208 | // STPWi kill %w0, %w1, ... |
| 1209 | // USE %w1 |
| 1210 | for (auto It = std::next(x: I); It != Paired && PairedRegOp.isKill(); ++It) |
| 1211 | if (It->readsRegister(Reg: PairedRegOp.getReg(), TRI)) |
| 1212 | PairedRegOp.setIsKill(false); |
| 1213 | } else { |
| 1214 | // Clear kill flags of the first stores register. Example: |
| 1215 | // STRWui %w1, ... |
| 1216 | // USE kill %w1 ; need to clear kill flag when moving STRWui downwards |
| 1217 | // STRW %w0 |
| 1218 | Register Reg = getLdStRegOp(MI&: *I).getReg(); |
| 1219 | for (MachineInstr &MI : |
| 1220 | make_range(x: std::next(x: I->getIterator()), y: Paired->getIterator())) |
| 1221 | MI.clearRegisterKills(Reg, RegInfo: TRI); |
| 1222 | } |
| 1223 | } |
| 1224 | |
| 1225 | unsigned int MatchPairOpcode = getMatchingPairOpcode(Opc); |
| 1226 | MIB = BuildMI(BB&: *MBB, I: InsertionPoint, MIMD: DL, MCID: TII->get(Opcode: MatchPairOpcode)); |
| 1227 | |
| 1228 | // Adds the pre-index operand for pre-indexed ld/st pairs. |
| 1229 | if (AArch64InstrInfo::isPreLdSt(MI: *RtMI)) |
| 1230 | MIB.addReg(RegNo: BaseRegOp.getReg(), Flags: RegState::Define); |
| 1231 | |
| 1232 | MIB.add(MO: RegOp0) |
| 1233 | .add(MO: RegOp1) |
| 1234 | .add(MO: BaseRegOp) |
| 1235 | .addImm(Val: OffsetImm) |
| 1236 | .cloneMergedMemRefs(OtherMIs: {&*I, &*Paired}) |
| 1237 | .setMIFlags(I->mergeFlagsWith(Other: *Paired)); |
| 1238 | |
| 1239 | (void)MIB; |
| 1240 | |
| 1241 | LLVM_DEBUG( |
| 1242 | dbgs() << "Creating pair load/store. Replacing instructions:\n " ); |
| 1243 | LLVM_DEBUG(I->print(dbgs())); |
| 1244 | LLVM_DEBUG(dbgs() << " " ); |
| 1245 | LLVM_DEBUG(Paired->print(dbgs())); |
| 1246 | LLVM_DEBUG(dbgs() << " with instruction:\n " ); |
| 1247 | if (SExtIdx != -1) { |
| 1248 | // Generate the sign extension for the proper result of the ldp. |
| 1249 | // I.e., with X1, that would be: |
| 1250 | // %w1 = KILL %w1, implicit-def %x1 |
| 1251 | // %x1 = SBFMXri killed %x1, 0, 31 |
| 1252 | MachineOperand &DstMO = MIB->getOperand(i: SExtIdx); |
| 1253 | // Right now, DstMO has the extended register, since it comes from an |
| 1254 | // extended opcode. |
| 1255 | Register DstRegX = DstMO.getReg(); |
| 1256 | // Get the W variant of that register. |
| 1257 | Register DstRegW = TRI->getSubReg(Reg: DstRegX, Idx: AArch64::sub_32); |
| 1258 | // Update the result of LDP to use the W instead of the X variant. |
| 1259 | DstMO.setReg(DstRegW); |
| 1260 | LLVM_DEBUG(((MachineInstr *)MIB)->print(dbgs())); |
| 1261 | LLVM_DEBUG(dbgs() << "\n" ); |
| 1262 | // Make the machine verifier happy by providing a definition for |
| 1263 | // the X register. |
| 1264 | // Insert this definition right after the generated LDP, i.e., before |
| 1265 | // InsertionPoint. |
| 1266 | MachineInstrBuilder MIBKill = |
| 1267 | BuildMI(BB&: *MBB, I: InsertionPoint, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::KILL), DestReg: DstRegW) |
| 1268 | .addReg(RegNo: DstRegW) |
| 1269 | .addReg(RegNo: DstRegX, Flags: RegState::Define); |
| 1270 | MIBKill->getOperand(i: 2).setImplicit(); |
| 1271 | // Create the sign extension. |
| 1272 | MachineInstrBuilder MIBSXTW = |
| 1273 | BuildMI(BB&: *MBB, I: InsertionPoint, MIMD: DL, MCID: TII->get(Opcode: AArch64::SBFMXri), DestReg: DstRegX) |
| 1274 | .addReg(RegNo: DstRegX) |
| 1275 | .addImm(Val: 0) |
| 1276 | .addImm(Val: 31); |
| 1277 | (void)MIBSXTW; |
| 1278 | |
| 1279 | // In the case of a sign-extend, where we have something like: |
| 1280 | // debugValueSubstitutions:[] |
| 1281 | // $w1 = LDRWui $x0, 1, debug-instr-number 1 |
| 1282 | // DBG_INSTR_REF !7, dbg-instr-ref(1, 0), debug-location !9 |
| 1283 | // $x0 = LDRSWui $x0, 0, debug-instr-number 2 |
| 1284 | // DBG_INSTR_REF !8, dbg-instr-ref(2, 0), debug-location !9 |
| 1285 | |
| 1286 | // It will be converted to: |
| 1287 | // debugValueSubstitutions:[] |
| 1288 | // $w0, $w1 = LDPWi $x0, 0 |
| 1289 | // $w0 = KILL $w0, implicit-def $x0 |
| 1290 | // $x0 = SBFMXri $x0, 0, 31 |
| 1291 | // DBG_INSTR_REF !7, dbg-instr-ref(1, 0), debug-location !9 |
| 1292 | // DBG_INSTR_REF !8, dbg-instr-ref(2, 0), debug-location !9 |
| 1293 | |
| 1294 | // We want the final result to look like: |
| 1295 | // debugValueSubstitutions: |
| 1296 | // - { srcinst: 1, srcop: 0, dstinst: 4, dstop: 1, subreg: 0 } |
| 1297 | // - { srcinst: 2, srcop: 0, dstinst: 3, dstop: 0, subreg: 0 } |
| 1298 | // $w0, $w1 = LDPWi $x0, 0, debug-instr-number 4 |
| 1299 | // $w0 = KILL $w0, implicit-def $x0 |
| 1300 | // $x0 = SBFMXri $x0, 0, 31, debug-instr-number 3 |
| 1301 | // DBG_INSTR_REF !7, dbg-instr-ref(1, 0), debug-location !9 |
| 1302 | // DBG_INSTR_REF !8, dbg-instr-ref(2, 0), debug-location !9 |
| 1303 | |
| 1304 | // $x0 is where the final value is stored, so the sign extend (SBFMXri) |
| 1305 | // instruction contains the final value we care about we give it a new |
| 1306 | // debug-instr-number 3. Whereas, $w1 contains the final value that we care |
| 1307 | // about, therefore the LDP instruction is also given a new |
| 1308 | // debug-instr-number 4. We have to add these substitutions to the |
| 1309 | // debugValueSubstitutions table. However, we also have to ensure that the |
| 1310 | // OpIndex that pointed to debug-instr-number 1 gets updated to 1, because |
| 1311 | // $w1 is the second operand of the LDP instruction. |
| 1312 | |
| 1313 | if (I->peekDebugInstrNum()) { |
| 1314 | // If I is the instruction which got sign extended and has a |
| 1315 | // debug-instr-number, give the SBFMXri instruction a new |
| 1316 | // debug-instr-number, and update the debugValueSubstitutions table with |
| 1317 | // the new debug-instr-number and OpIndex pair. Otherwise, give the Merged |
| 1318 | // instruction a new debug-instr-number, and update the |
| 1319 | // debugValueSubstitutions table with the new debug-instr-number and |
| 1320 | // OpIndex pair. |
| 1321 | unsigned NewInstrNum; |
| 1322 | if (DstRegX == I->getOperand(i: 0).getReg()) { |
| 1323 | NewInstrNum = MIBSXTW->getDebugInstrNum(); |
| 1324 | addDebugSubstitutionsToTable(MF: MBB->getParent(), InstrNumToSet: NewInstrNum, OriginalInstr&: *I, |
| 1325 | MergedInstr&: *MIBSXTW); |
| 1326 | } else { |
| 1327 | NewInstrNum = MIB->getDebugInstrNum(); |
| 1328 | addDebugSubstitutionsToTable(MF: MBB->getParent(), InstrNumToSet: NewInstrNum, OriginalInstr&: *I, MergedInstr&: *MIB); |
| 1329 | } |
| 1330 | } |
| 1331 | if (Paired->peekDebugInstrNum()) { |
| 1332 | // If Paired is the instruction which got sign extended and has a |
| 1333 | // debug-instr-number, give the SBFMXri instruction a new |
| 1334 | // debug-instr-number, and update the debugValueSubstitutions table with |
| 1335 | // the new debug-instr-number and OpIndex pair. Otherwise, give the Merged |
| 1336 | // instruction a new debug-instr-number, and update the |
| 1337 | // debugValueSubstitutions table with the new debug-instr-number and |
| 1338 | // OpIndex pair. |
| 1339 | unsigned NewInstrNum; |
| 1340 | if (DstRegX == Paired->getOperand(i: 0).getReg()) { |
| 1341 | NewInstrNum = MIBSXTW->getDebugInstrNum(); |
| 1342 | addDebugSubstitutionsToTable(MF: MBB->getParent(), InstrNumToSet: NewInstrNum, OriginalInstr&: *Paired, |
| 1343 | MergedInstr&: *MIBSXTW); |
| 1344 | } else { |
| 1345 | NewInstrNum = MIB->getDebugInstrNum(); |
| 1346 | addDebugSubstitutionsToTable(MF: MBB->getParent(), InstrNumToSet: NewInstrNum, OriginalInstr&: *Paired, |
| 1347 | MergedInstr&: *MIB); |
| 1348 | } |
| 1349 | } |
| 1350 | |
| 1351 | LLVM_DEBUG(dbgs() << " Extend operand:\n " ); |
| 1352 | LLVM_DEBUG(((MachineInstr *)MIBSXTW)->print(dbgs())); |
| 1353 | } else if (Opc == AArch64::LDR_ZXI || Opc == AArch64::STR_ZXI) { |
| 1354 | // We are combining SVE fill/spill to LDP/STP, so we need to use the Q |
| 1355 | // variant of the registers. |
| 1356 | MachineOperand &MOp0 = MIB->getOperand(i: 0); |
| 1357 | MachineOperand &MOp1 = MIB->getOperand(i: 1); |
| 1358 | assert(AArch64::ZPRRegClass.contains(MOp0.getReg()) && |
| 1359 | AArch64::ZPRRegClass.contains(MOp1.getReg()) && "Invalid register." ); |
| 1360 | MOp0.setReg(AArch64::Q0 + (MOp0.getReg() - AArch64::Z0)); |
| 1361 | MOp1.setReg(AArch64::Q0 + (MOp1.getReg() - AArch64::Z0)); |
| 1362 | LLVM_DEBUG(((MachineInstr *)MIB)->print(dbgs())); |
| 1363 | } else { |
| 1364 | |
| 1365 | // In the case that the merge doesn't result in a sign-extend, if we have |
| 1366 | // something like: |
| 1367 | // debugValueSubstitutions:[] |
| 1368 | // $x1 = LDRXui $x0, 1, debug-instr-number 1 |
| 1369 | // DBG_INSTR_REF !13, dbg-instr-ref(1, 0), debug-location !11 |
| 1370 | // $x0 = LDRXui killed $x0, 0, debug-instr-number 2 |
| 1371 | // DBG_INSTR_REF !14, dbg-instr-ref(2, 0), debug-location !11 |
| 1372 | |
| 1373 | // It will be converted to: |
| 1374 | // debugValueSubstitutions: [] |
| 1375 | // $x0, $x1 = LDPXi $x0, 0 |
| 1376 | // DBG_INSTR_REF !12, dbg-instr-ref(1, 0), debug-location !14 |
| 1377 | // DBG_INSTR_REF !13, dbg-instr-ref(2, 0), debug-location !14 |
| 1378 | |
| 1379 | // We want the final result to look like: |
| 1380 | // debugValueSubstitutions: |
| 1381 | // - { srcinst: 1, srcop: 0, dstinst: 3, dstop: 1, subreg: 0 } |
| 1382 | // - { srcinst: 2, srcop: 0, dstinst: 3, dstop: 0, subreg: 0 } |
| 1383 | // $x0, $x1 = LDPXi $x0, 0, debug-instr-number 3 |
| 1384 | // DBG_INSTR_REF !12, dbg-instr-ref(1, 0), debug-location !14 |
| 1385 | // DBG_INSTR_REF !12, dbg-instr-ref(2, 0), debug-location !14 |
| 1386 | |
| 1387 | // Here all that needs to be done is, that the LDP instruction needs to be |
| 1388 | // updated with a new debug-instr-number, we then need to add entries into |
| 1389 | // the debugSubstitutions table to map the old instr-refs to the new ones. |
| 1390 | |
| 1391 | // Assign new DebugInstrNum to the Paired instruction. |
| 1392 | if (I->peekDebugInstrNum()) { |
| 1393 | unsigned NewDebugInstrNum = MIB->getDebugInstrNum(); |
| 1394 | addDebugSubstitutionsToTable(MF: MBB->getParent(), InstrNumToSet: NewDebugInstrNum, OriginalInstr&: *I, |
| 1395 | MergedInstr&: *MIB); |
| 1396 | } |
| 1397 | if (Paired->peekDebugInstrNum()) { |
| 1398 | unsigned NewDebugInstrNum = MIB->getDebugInstrNum(); |
| 1399 | addDebugSubstitutionsToTable(MF: MBB->getParent(), InstrNumToSet: NewDebugInstrNum, OriginalInstr&: *Paired, |
| 1400 | MergedInstr&: *MIB); |
| 1401 | } |
| 1402 | |
| 1403 | LLVM_DEBUG(((MachineInstr *)MIB)->print(dbgs())); |
| 1404 | } |
| 1405 | LLVM_DEBUG(dbgs() << "\n" ); |
| 1406 | |
| 1407 | if (MergeForward) |
| 1408 | for (const MachineOperand &MOP : phys_regs_and_masks(MI: *I)) |
| 1409 | if (MOP.isReg() && MOP.isKill()) |
| 1410 | DefinedInBB.addReg(Reg: MOP.getReg()); |
| 1411 | |
| 1412 | // Copy over any implicit-def operands. This is like MI.copyImplicitOps, but |
| 1413 | // only copies implicit defs and makes sure that each operand is only added |
| 1414 | // once in case of duplicates. |
| 1415 | auto CopyImplicitOps = [&](MachineBasicBlock::iterator MI1, |
| 1416 | MachineBasicBlock::iterator MI2) { |
| 1417 | SmallSetVector<Register, 4> Ops; |
| 1418 | for (const MachineOperand &MO : |
| 1419 | llvm::drop_begin(RangeOrContainer: MI1->operands(), N: MI1->getDesc().getNumOperands())) |
| 1420 | if (MO.isReg() && MO.isImplicit() && MO.isDef()) |
| 1421 | Ops.insert(X: MO.getReg()); |
| 1422 | for (const MachineOperand &MO : |
| 1423 | llvm::drop_begin(RangeOrContainer: MI2->operands(), N: MI2->getDesc().getNumOperands())) |
| 1424 | if (MO.isReg() && MO.isImplicit() && MO.isDef()) |
| 1425 | Ops.insert(X: MO.getReg()); |
| 1426 | for (auto Op : Ops) |
| 1427 | MIB.addDef(RegNo: Op, Flags: RegState::Implicit); |
| 1428 | }; |
| 1429 | CopyImplicitOps(I, Paired); |
| 1430 | |
| 1431 | // Erase the old instructions. |
| 1432 | I->eraseFromParent(); |
| 1433 | Paired->eraseFromParent(); |
| 1434 | |
| 1435 | return NextI; |
| 1436 | } |
| 1437 | |
| 1438 | MachineBasicBlock::iterator |
| 1439 | AArch64LoadStoreOpt::promoteLoadFromStore(MachineBasicBlock::iterator LoadI, |
| 1440 | MachineBasicBlock::iterator StoreI) { |
| 1441 | MachineBasicBlock::iterator NextI = |
| 1442 | next_nodbg(It: LoadI, End: LoadI->getParent()->end()); |
| 1443 | |
| 1444 | int LoadSize = TII->getMemScale(MI: *LoadI); |
| 1445 | int StoreSize = TII->getMemScale(MI: *StoreI); |
| 1446 | Register LdRt = getLdStRegOp(MI&: *LoadI).getReg(); |
| 1447 | const MachineOperand &StMO = getLdStRegOp(MI&: *StoreI); |
| 1448 | Register StRt = getLdStRegOp(MI&: *StoreI).getReg(); |
| 1449 | bool IsStoreXReg = TRI->getRegClass(i: AArch64::GPR64RegClassID)->contains(Reg: StRt); |
| 1450 | |
| 1451 | assert((IsStoreXReg || |
| 1452 | TRI->getRegClass(AArch64::GPR32RegClassID)->contains(StRt)) && |
| 1453 | "Unexpected RegClass" ); |
| 1454 | |
| 1455 | MachineInstr *BitExtMI; |
| 1456 | if (LoadSize == StoreSize && (LoadSize == 4 || LoadSize == 8)) { |
| 1457 | // Remove the load, if the destination register of the loads is the same |
| 1458 | // register for stored value. |
| 1459 | if (StRt == LdRt && LoadSize == 8) { |
| 1460 | for (MachineInstr &MI : make_range(x: StoreI->getIterator(), |
| 1461 | y: LoadI->getIterator())) { |
| 1462 | if (MI.killsRegister(Reg: StRt, TRI)) { |
| 1463 | MI.clearRegisterKills(Reg: StRt, RegInfo: TRI); |
| 1464 | break; |
| 1465 | } |
| 1466 | } |
| 1467 | LLVM_DEBUG(dbgs() << "Remove load instruction:\n " ); |
| 1468 | LLVM_DEBUG(LoadI->print(dbgs())); |
| 1469 | LLVM_DEBUG(dbgs() << "\n" ); |
| 1470 | LoadI->eraseFromParent(); |
| 1471 | return NextI; |
| 1472 | } |
| 1473 | // Replace the load with a mov if the load and store are in the same size. |
| 1474 | BitExtMI = |
| 1475 | BuildMI(BB&: *LoadI->getParent(), I: LoadI, MIMD: LoadI->getDebugLoc(), |
| 1476 | MCID: TII->get(Opcode: IsStoreXReg ? AArch64::ORRXrs : AArch64::ORRWrs), DestReg: LdRt) |
| 1477 | .addReg(RegNo: IsStoreXReg ? AArch64::XZR : AArch64::WZR) |
| 1478 | .add(MO: StMO) |
| 1479 | .addImm(Val: AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: 0)) |
| 1480 | .setMIFlags(LoadI->getFlags()); |
| 1481 | } else { |
| 1482 | // FIXME: Currently we disable this transformation in big-endian targets as |
| 1483 | // performance and correctness are verified only in little-endian. |
| 1484 | if (!Subtarget->isLittleEndian()) |
| 1485 | return NextI; |
| 1486 | bool IsUnscaled = TII->hasUnscaledLdStOffset(MI&: *LoadI); |
| 1487 | assert(IsUnscaled == TII->hasUnscaledLdStOffset(*StoreI) && |
| 1488 | "Unsupported ld/st match" ); |
| 1489 | assert(LoadSize <= StoreSize && "Invalid load size" ); |
| 1490 | int UnscaledLdOffset = |
| 1491 | IsUnscaled |
| 1492 | ? AArch64InstrInfo::getLdStOffsetOp(MI: *LoadI).getImm() |
| 1493 | : AArch64InstrInfo::getLdStOffsetOp(MI: *LoadI).getImm() * LoadSize; |
| 1494 | int UnscaledStOffset = |
| 1495 | IsUnscaled |
| 1496 | ? AArch64InstrInfo::getLdStOffsetOp(MI: *StoreI).getImm() |
| 1497 | : AArch64InstrInfo::getLdStOffsetOp(MI: *StoreI).getImm() * StoreSize; |
| 1498 | int Width = LoadSize * 8; |
| 1499 | Register DestReg = |
| 1500 | IsStoreXReg ? Register(TRI->getMatchingSuperReg( |
| 1501 | Reg: LdRt, SubIdx: AArch64::sub_32, RC: &AArch64::GPR64RegClass)) |
| 1502 | : LdRt; |
| 1503 | |
| 1504 | assert((UnscaledLdOffset >= UnscaledStOffset && |
| 1505 | (UnscaledLdOffset + LoadSize) <= UnscaledStOffset + StoreSize) && |
| 1506 | "Invalid offset" ); |
| 1507 | |
| 1508 | int Immr = 8 * (UnscaledLdOffset - UnscaledStOffset); |
| 1509 | int Imms = Immr + Width - 1; |
| 1510 | if (UnscaledLdOffset == UnscaledStOffset) { |
| 1511 | uint32_t AndMaskEncoded = ((IsStoreXReg ? 1 : 0) << 12) // N |
| 1512 | | ((Immr) << 6) // immr |
| 1513 | | ((Imms) << 0) // imms |
| 1514 | ; |
| 1515 | |
| 1516 | BitExtMI = |
| 1517 | BuildMI(BB&: *LoadI->getParent(), I: LoadI, MIMD: LoadI->getDebugLoc(), |
| 1518 | MCID: TII->get(Opcode: IsStoreXReg ? AArch64::ANDXri : AArch64::ANDWri), |
| 1519 | DestReg) |
| 1520 | .add(MO: StMO) |
| 1521 | .addImm(Val: AndMaskEncoded) |
| 1522 | .setMIFlags(LoadI->getFlags()); |
| 1523 | } else if (IsStoreXReg && Imms == 31) { |
| 1524 | // Use the 32 bit variant of UBFM if it's the LSR alias of the |
| 1525 | // instruction. |
| 1526 | assert(Immr <= Imms && "Expected LSR alias of UBFM" ); |
| 1527 | BitExtMI = BuildMI(BB&: *LoadI->getParent(), I: LoadI, MIMD: LoadI->getDebugLoc(), |
| 1528 | MCID: TII->get(Opcode: AArch64::UBFMWri), |
| 1529 | DestReg: TRI->getSubReg(Reg: DestReg, Idx: AArch64::sub_32)) |
| 1530 | .addReg(RegNo: TRI->getSubReg(Reg: StRt, Idx: AArch64::sub_32)) |
| 1531 | .addImm(Val: Immr) |
| 1532 | .addImm(Val: Imms) |
| 1533 | .setMIFlags(LoadI->getFlags()); |
| 1534 | } else { |
| 1535 | BitExtMI = |
| 1536 | BuildMI(BB&: *LoadI->getParent(), I: LoadI, MIMD: LoadI->getDebugLoc(), |
| 1537 | MCID: TII->get(Opcode: IsStoreXReg ? AArch64::UBFMXri : AArch64::UBFMWri), |
| 1538 | DestReg) |
| 1539 | .add(MO: StMO) |
| 1540 | .addImm(Val: Immr) |
| 1541 | .addImm(Val: Imms) |
| 1542 | .setMIFlags(LoadI->getFlags()); |
| 1543 | } |
| 1544 | } |
| 1545 | |
| 1546 | // Clear kill flags between store and load. |
| 1547 | for (MachineInstr &MI : make_range(x: StoreI->getIterator(), |
| 1548 | y: BitExtMI->getIterator())) |
| 1549 | if (MI.killsRegister(Reg: StRt, TRI)) { |
| 1550 | MI.clearRegisterKills(Reg: StRt, RegInfo: TRI); |
| 1551 | break; |
| 1552 | } |
| 1553 | |
| 1554 | LLVM_DEBUG(dbgs() << "Promoting load by replacing :\n " ); |
| 1555 | LLVM_DEBUG(StoreI->print(dbgs())); |
| 1556 | LLVM_DEBUG(dbgs() << " " ); |
| 1557 | LLVM_DEBUG(LoadI->print(dbgs())); |
| 1558 | LLVM_DEBUG(dbgs() << " with instructions:\n " ); |
| 1559 | LLVM_DEBUG(StoreI->print(dbgs())); |
| 1560 | LLVM_DEBUG(dbgs() << " " ); |
| 1561 | LLVM_DEBUG((BitExtMI)->print(dbgs())); |
| 1562 | LLVM_DEBUG(dbgs() << "\n" ); |
| 1563 | |
| 1564 | // Erase the old instructions. |
| 1565 | LoadI->eraseFromParent(); |
| 1566 | return NextI; |
| 1567 | } |
| 1568 | |
| 1569 | static bool inBoundsForPair(bool IsUnscaled, int Offset, int OffsetStride) { |
| 1570 | // Convert the byte-offset used by unscaled into an "element" offset used |
| 1571 | // by the scaled pair load/store instructions. |
| 1572 | if (IsUnscaled) { |
| 1573 | // If the byte-offset isn't a multiple of the stride, there's no point |
| 1574 | // trying to match it. |
| 1575 | if (Offset % OffsetStride) |
| 1576 | return false; |
| 1577 | Offset /= OffsetStride; |
| 1578 | } |
| 1579 | return Offset <= 63 && Offset >= -64; |
| 1580 | } |
| 1581 | |
| 1582 | // Do alignment, specialized to power of 2 and for signed ints, |
| 1583 | // avoiding having to do a C-style cast from uint_64t to int when |
| 1584 | // using alignTo from include/llvm/Support/MathExtras.h. |
| 1585 | // FIXME: Move this function to include/MathExtras.h? |
| 1586 | static int alignTo(int Num, int PowOf2) { |
| 1587 | return (Num + PowOf2 - 1) & ~(PowOf2 - 1); |
| 1588 | } |
| 1589 | |
| 1590 | static bool mayAlias(MachineInstr &MIa, |
| 1591 | SmallVectorImpl<MachineInstr *> &MemInsns, |
| 1592 | AliasAnalysis *AA) { |
| 1593 | for (MachineInstr *MIb : MemInsns) { |
| 1594 | if (MIa.mayAlias(AA, Other: *MIb, /*UseTBAA*/ false)) { |
| 1595 | LLVM_DEBUG(dbgs() << "Aliasing with: " ; MIb->dump()); |
| 1596 | return true; |
| 1597 | } |
| 1598 | } |
| 1599 | |
| 1600 | LLVM_DEBUG(dbgs() << "No aliases found\n" ); |
| 1601 | return false; |
| 1602 | } |
| 1603 | |
| 1604 | bool AArch64LoadStoreOpt::findMatchingStore( |
| 1605 | MachineBasicBlock::iterator I, unsigned Limit, |
| 1606 | MachineBasicBlock::iterator &StoreI) { |
| 1607 | MachineBasicBlock::iterator B = I->getParent()->begin(); |
| 1608 | MachineBasicBlock::iterator MBBI = I; |
| 1609 | MachineInstr &LoadMI = *I; |
| 1610 | Register BaseReg = AArch64InstrInfo::getLdStBaseOp(MI: LoadMI).getReg(); |
| 1611 | |
| 1612 | // If the load is the first instruction in the block, there's obviously |
| 1613 | // not any matching store. |
| 1614 | if (MBBI == B) |
| 1615 | return false; |
| 1616 | |
| 1617 | // Track which register units have been modified and used between the first |
| 1618 | // insn and the second insn. |
| 1619 | ModifiedRegUnits.clear(); |
| 1620 | UsedRegUnits.clear(); |
| 1621 | |
| 1622 | unsigned Count = 0; |
| 1623 | do { |
| 1624 | MBBI = prev_nodbg(It: MBBI, Begin: B); |
| 1625 | MachineInstr &MI = *MBBI; |
| 1626 | |
| 1627 | // Don't count transient instructions towards the search limit since there |
| 1628 | // may be different numbers of them if e.g. debug information is present. |
| 1629 | if (!MI.isTransient()) |
| 1630 | ++Count; |
| 1631 | |
| 1632 | // If the load instruction reads directly from the address to which the |
| 1633 | // store instruction writes and the stored value is not modified, we can |
| 1634 | // promote the load. Since we do not handle stores with pre-/post-index, |
| 1635 | // it's unnecessary to check if BaseReg is modified by the store itself. |
| 1636 | // Also we can't handle stores without an immediate offset operand, |
| 1637 | // while the operand might be the address for a global variable. |
| 1638 | if (MI.mayStore() && isMatchingStore(LoadInst&: LoadMI, StoreInst&: MI) && |
| 1639 | BaseReg == AArch64InstrInfo::getLdStBaseOp(MI).getReg() && |
| 1640 | AArch64InstrInfo::getLdStOffsetOp(MI).isImm() && |
| 1641 | isLdOffsetInRangeOfSt(LoadInst&: LoadMI, StoreInst&: MI, TII) && |
| 1642 | ModifiedRegUnits.available(Reg: getLdStRegOp(MI).getReg())) { |
| 1643 | StoreI = MBBI; |
| 1644 | return true; |
| 1645 | } |
| 1646 | |
| 1647 | if (MI.isCall()) |
| 1648 | return false; |
| 1649 | |
| 1650 | // Update modified / uses register units. |
| 1651 | LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, TRI); |
| 1652 | |
| 1653 | // Otherwise, if the base register is modified, we have no match, so |
| 1654 | // return early. |
| 1655 | if (!ModifiedRegUnits.available(Reg: BaseReg)) |
| 1656 | return false; |
| 1657 | |
| 1658 | // If we encounter a store aliased with the load, return early. |
| 1659 | if (MI.mayStore() && LoadMI.mayAlias(AA, Other: MI, /*UseTBAA*/ false)) |
| 1660 | return false; |
| 1661 | } while (MBBI != B && Count < Limit); |
| 1662 | return false; |
| 1663 | } |
| 1664 | |
| 1665 | static bool needsWinCFI(const MachineFunction *MF) { |
| 1666 | return MF->getTarget().getMCAsmInfo()->usesWindowsCFI() && |
| 1667 | MF->getFunction().needsUnwindTableEntry(); |
| 1668 | } |
| 1669 | |
| 1670 | // Returns true if FirstMI and MI are candidates for merging or pairing. |
| 1671 | // Otherwise, returns false. |
| 1672 | static bool areCandidatesToMergeOrPair(MachineInstr &FirstMI, MachineInstr &MI, |
| 1673 | LdStPairFlags &Flags, |
| 1674 | const AArch64InstrInfo *TII) { |
| 1675 | // If this is volatile or if pairing is suppressed, not a candidate. |
| 1676 | if (MI.hasOrderedMemoryRef() || TII->isLdStPairSuppressed(MI)) |
| 1677 | return false; |
| 1678 | |
| 1679 | // We should have already checked FirstMI for pair suppression and volatility. |
| 1680 | assert(!FirstMI.hasOrderedMemoryRef() && |
| 1681 | !TII->isLdStPairSuppressed(FirstMI) && |
| 1682 | "FirstMI shouldn't get here if either of these checks are true." ); |
| 1683 | |
| 1684 | if (needsWinCFI(MF: MI.getMF()) && (MI.getFlag(Flag: MachineInstr::FrameSetup) || |
| 1685 | MI.getFlag(Flag: MachineInstr::FrameDestroy))) |
| 1686 | return false; |
| 1687 | |
| 1688 | unsigned OpcA = FirstMI.getOpcode(); |
| 1689 | unsigned OpcB = MI.getOpcode(); |
| 1690 | |
| 1691 | // Opcodes match: If the opcodes are pre ld/st there is nothing more to check. |
| 1692 | if (OpcA == OpcB) |
| 1693 | return !AArch64InstrInfo::isPreLdSt(MI: FirstMI); |
| 1694 | |
| 1695 | // Bail out if one of the opcodes is SVE fill/spill, as we currently don't |
| 1696 | // allow pairing them with other instructions. |
| 1697 | if (OpcA == AArch64::LDR_ZXI || OpcA == AArch64::STR_ZXI || |
| 1698 | OpcB == AArch64::LDR_ZXI || OpcB == AArch64::STR_ZXI) |
| 1699 | return false; |
| 1700 | |
| 1701 | // Two pre ld/st of different opcodes cannot be merged either |
| 1702 | if (AArch64InstrInfo::isPreLdSt(MI: FirstMI) && AArch64InstrInfo::isPreLdSt(MI)) |
| 1703 | return false; |
| 1704 | |
| 1705 | // Try to match a sign-extended load/store with a zero-extended load/store. |
| 1706 | bool IsValidLdStrOpc, PairIsValidLdStrOpc; |
| 1707 | unsigned NonSExtOpc = getMatchingNonSExtOpcode(Opc: OpcA, IsValidLdStrOpc: &IsValidLdStrOpc); |
| 1708 | assert(IsValidLdStrOpc && |
| 1709 | "Given Opc should be a Load or Store with an immediate" ); |
| 1710 | // OpcA will be the first instruction in the pair. |
| 1711 | if (NonSExtOpc == getMatchingNonSExtOpcode(Opc: OpcB, IsValidLdStrOpc: &PairIsValidLdStrOpc)) { |
| 1712 | Flags.setSExtIdx(NonSExtOpc == OpcA ? 1 : 0); |
| 1713 | return true; |
| 1714 | } |
| 1715 | |
| 1716 | // If the second instruction isn't even a mergable/pairable load/store, bail |
| 1717 | // out. |
| 1718 | if (!PairIsValidLdStrOpc) |
| 1719 | return false; |
| 1720 | |
| 1721 | // Narrow stores do not have a matching pair opcodes, so constrain their |
| 1722 | // merging to zero stores. |
| 1723 | if (isNarrowStore(Opc: OpcA) || isNarrowStore(Opc: OpcB)) |
| 1724 | return getLdStRegOp(MI&: FirstMI).getReg() == AArch64::WZR && |
| 1725 | getLdStRegOp(MI).getReg() == AArch64::WZR && |
| 1726 | TII->getMemScale(MI: FirstMI) == TII->getMemScale(MI); |
| 1727 | |
| 1728 | // The STR<S,D,Q,W,X>pre - STR<S,D,Q,W,X>ui and |
| 1729 | // LDR<S,D,Q,W,X,SW>pre-LDR<S,D,Q,W,X,SW>ui |
| 1730 | // are candidate pairs that can be merged. |
| 1731 | if (isPreLdStPairCandidate(FirstMI, MI)) |
| 1732 | return true; |
| 1733 | |
| 1734 | // Try to match an unscaled load/store with a scaled load/store. |
| 1735 | return TII->hasUnscaledLdStOffset(Opc: OpcA) != TII->hasUnscaledLdStOffset(Opc: OpcB) && |
| 1736 | getMatchingPairOpcode(Opc: OpcA) == getMatchingPairOpcode(Opc: OpcB); |
| 1737 | |
| 1738 | // FIXME: Can we also match a mixed sext/zext unscaled/scaled pair? |
| 1739 | } |
| 1740 | |
| 1741 | static bool canRenameMOP(const MachineOperand &MOP, |
| 1742 | const TargetRegisterInfo *TRI) { |
| 1743 | if (MOP.isReg()) { |
| 1744 | auto *RegClass = TRI->getMinimalPhysRegClass(Reg: MOP.getReg()); |
| 1745 | // Renaming registers with multiple disjunct sub-registers (e.g. the |
| 1746 | // result of a LD3) means that all sub-registers are renamed, potentially |
| 1747 | // impacting other instructions we did not check. Bail out. |
| 1748 | // Note that this relies on the structure of the AArch64 register file. In |
| 1749 | // particular, a subregister cannot be written without overwriting the |
| 1750 | // whole register. |
| 1751 | if (RegClass->HasDisjunctSubRegs && RegClass->CoveredBySubRegs && |
| 1752 | (TRI->getSubRegisterClass(SuperRC: RegClass, SubRegIdx: AArch64::dsub0) || |
| 1753 | TRI->getSubRegisterClass(SuperRC: RegClass, SubRegIdx: AArch64::qsub0) || |
| 1754 | TRI->getSubRegisterClass(SuperRC: RegClass, SubRegIdx: AArch64::zsub0))) { |
| 1755 | LLVM_DEBUG( |
| 1756 | dbgs() |
| 1757 | << " Cannot rename operands with multiple disjunct subregisters (" |
| 1758 | << MOP << ")\n" ); |
| 1759 | return false; |
| 1760 | } |
| 1761 | |
| 1762 | // We cannot rename arbitrary implicit-defs, the specific rule to rewrite |
| 1763 | // them must be known. For example, in ORRWrs the implicit-def |
| 1764 | // corresponds to the result register. |
| 1765 | if (MOP.isImplicit() && MOP.isDef()) { |
| 1766 | if (!isRewritableImplicitDef(MO: MOP)) |
| 1767 | return false; |
| 1768 | return TRI->isSuperOrSubRegisterEq( |
| 1769 | RegA: MOP.getParent()->getOperand(i: 0).getReg(), RegB: MOP.getReg()); |
| 1770 | } |
| 1771 | } |
| 1772 | return MOP.isImplicit() || |
| 1773 | (MOP.isRenamable() && !MOP.isEarlyClobber() && !MOP.isTied()); |
| 1774 | } |
| 1775 | |
| 1776 | static bool |
| 1777 | canRenameUpToDef(MachineInstr &FirstMI, LiveRegUnits &UsedInBetween, |
| 1778 | SmallPtrSetImpl<const TargetRegisterClass *> &RequiredClasses, |
| 1779 | const TargetRegisterInfo *TRI) { |
| 1780 | if (!FirstMI.mayStore()) |
| 1781 | return false; |
| 1782 | |
| 1783 | // Check if we can find an unused register which we can use to rename |
| 1784 | // the register used by the first load/store. |
| 1785 | |
| 1786 | auto RegToRename = getLdStRegOp(MI&: FirstMI).getReg(); |
| 1787 | // For now, we only rename if the store operand gets killed at the store. |
| 1788 | if (!getLdStRegOp(MI&: FirstMI).isKill() && |
| 1789 | !any_of(Range: FirstMI.operands(), |
| 1790 | P: [TRI, RegToRename](const MachineOperand &MOP) { |
| 1791 | return MOP.isReg() && !MOP.isDebug() && MOP.getReg() && |
| 1792 | MOP.isImplicit() && MOP.isKill() && |
| 1793 | TRI->regsOverlap(RegA: RegToRename, RegB: MOP.getReg()); |
| 1794 | })) { |
| 1795 | LLVM_DEBUG(dbgs() << " Operand not killed at " << FirstMI); |
| 1796 | return false; |
| 1797 | } |
| 1798 | |
| 1799 | bool FoundDef = false; |
| 1800 | |
| 1801 | // For each instruction between FirstMI and the previous def for RegToRename, |
| 1802 | // we |
| 1803 | // * check if we can rename RegToRename in this instruction |
| 1804 | // * collect the registers used and required register classes for RegToRename. |
| 1805 | std::function<bool(MachineInstr &, bool)> CheckMIs = [&](MachineInstr &MI, |
| 1806 | bool IsDef) { |
| 1807 | LLVM_DEBUG(dbgs() << "Checking " << MI); |
| 1808 | // Currently we do not try to rename across frame-setup instructions. |
| 1809 | if (MI.getFlag(Flag: MachineInstr::FrameSetup)) { |
| 1810 | LLVM_DEBUG(dbgs() << " Cannot rename framesetup instructions " |
| 1811 | << "currently\n" ); |
| 1812 | return false; |
| 1813 | } |
| 1814 | |
| 1815 | UsedInBetween.accumulate(MI); |
| 1816 | |
| 1817 | // For a definition, check that we can rename the definition and exit the |
| 1818 | // loop. |
| 1819 | FoundDef = IsDef; |
| 1820 | |
| 1821 | // For defs, check if we can rename the first def of RegToRename. |
| 1822 | if (FoundDef) { |
| 1823 | // For some pseudo instructions, we might not generate code in the end |
| 1824 | // (e.g. KILL) and we would end up without a correct def for the rename |
| 1825 | // register. |
| 1826 | // TODO: This might be overly conservative and we could handle those cases |
| 1827 | // in multiple ways: |
| 1828 | // 1. Insert an extra copy, to materialize the def. |
| 1829 | // 2. Skip pseudo-defs until we find an non-pseudo def. |
| 1830 | if (MI.isPseudo()) { |
| 1831 | LLVM_DEBUG(dbgs() << " Cannot rename pseudo/bundle instruction\n" ); |
| 1832 | return false; |
| 1833 | } |
| 1834 | |
| 1835 | for (auto &MOP : MI.operands()) { |
| 1836 | if (!MOP.isReg() || !MOP.isDef() || MOP.isDebug() || !MOP.getReg() || |
| 1837 | !TRI->regsOverlap(RegA: MOP.getReg(), RegB: RegToRename)) |
| 1838 | continue; |
| 1839 | if (!canRenameMOP(MOP, TRI)) { |
| 1840 | LLVM_DEBUG(dbgs() << " Cannot rename " << MOP << " in " << MI); |
| 1841 | return false; |
| 1842 | } |
| 1843 | RequiredClasses.insert(Ptr: TRI->getMinimalPhysRegClass(Reg: MOP.getReg())); |
| 1844 | } |
| 1845 | return true; |
| 1846 | } else { |
| 1847 | for (auto &MOP : MI.operands()) { |
| 1848 | if (!MOP.isReg() || MOP.isDebug() || !MOP.getReg() || |
| 1849 | !TRI->regsOverlap(RegA: MOP.getReg(), RegB: RegToRename)) |
| 1850 | continue; |
| 1851 | |
| 1852 | if (!canRenameMOP(MOP, TRI)) { |
| 1853 | LLVM_DEBUG(dbgs() << " Cannot rename " << MOP << " in " << MI); |
| 1854 | return false; |
| 1855 | } |
| 1856 | RequiredClasses.insert(Ptr: TRI->getMinimalPhysRegClass(Reg: MOP.getReg())); |
| 1857 | } |
| 1858 | } |
| 1859 | return true; |
| 1860 | }; |
| 1861 | |
| 1862 | if (!forAllMIsUntilDef(MI&: FirstMI, DefReg: RegToRename, TRI, Limit: LdStLimit, Fn&: CheckMIs)) |
| 1863 | return false; |
| 1864 | |
| 1865 | if (!FoundDef) { |
| 1866 | LLVM_DEBUG(dbgs() << " Did not find definition for register in BB\n" ); |
| 1867 | return false; |
| 1868 | } |
| 1869 | return true; |
| 1870 | } |
| 1871 | |
| 1872 | // We want to merge the second load into the first by rewriting the usages of |
| 1873 | // the same reg between first (incl.) and second (excl.). We don't need to care |
| 1874 | // about any insns before FirstLoad or after SecondLoad. |
| 1875 | // 1. The second load writes new value into the same reg. |
| 1876 | // - The renaming is impossible to impact later use of the reg. |
| 1877 | // - The second load always trash the value written by the first load which |
| 1878 | // means the reg must be killed before the second load. |
| 1879 | // 2. The first load must be a def for the same reg so we don't need to look |
| 1880 | // into anything before it. |
| 1881 | static bool canRenameUntilSecondLoad( |
| 1882 | MachineInstr &FirstLoad, MachineInstr &SecondLoad, |
| 1883 | LiveRegUnits &UsedInBetween, |
| 1884 | SmallPtrSetImpl<const TargetRegisterClass *> &RequiredClasses, |
| 1885 | const TargetRegisterInfo *TRI) { |
| 1886 | if (FirstLoad.isPseudo()) |
| 1887 | return false; |
| 1888 | |
| 1889 | UsedInBetween.accumulate(MI: FirstLoad); |
| 1890 | auto RegToRename = getLdStRegOp(MI&: FirstLoad).getReg(); |
| 1891 | bool Success = std::all_of( |
| 1892 | first: FirstLoad.getIterator(), last: SecondLoad.getIterator(), |
| 1893 | pred: [&](MachineInstr &MI) { |
| 1894 | LLVM_DEBUG(dbgs() << "Checking " << MI); |
| 1895 | // Currently we do not try to rename across frame-setup instructions. |
| 1896 | if (MI.getFlag(Flag: MachineInstr::FrameSetup)) { |
| 1897 | LLVM_DEBUG(dbgs() << " Cannot rename framesetup instructions " |
| 1898 | << "currently\n" ); |
| 1899 | return false; |
| 1900 | } |
| 1901 | |
| 1902 | for (auto &MOP : MI.operands()) { |
| 1903 | if (!MOP.isReg() || MOP.isDebug() || !MOP.getReg() || |
| 1904 | !TRI->regsOverlap(RegA: MOP.getReg(), RegB: RegToRename)) |
| 1905 | continue; |
| 1906 | if (!canRenameMOP(MOP, TRI)) { |
| 1907 | LLVM_DEBUG(dbgs() << " Cannot rename " << MOP << " in " << MI); |
| 1908 | return false; |
| 1909 | } |
| 1910 | RequiredClasses.insert(Ptr: TRI->getMinimalPhysRegClass(Reg: MOP.getReg())); |
| 1911 | } |
| 1912 | |
| 1913 | return true; |
| 1914 | }); |
| 1915 | return Success; |
| 1916 | } |
| 1917 | |
| 1918 | // Check if we can find a physical register for renaming \p Reg. This register |
| 1919 | // must: |
| 1920 | // * not be defined already in \p DefinedInBB; DefinedInBB must contain all |
| 1921 | // defined registers up to the point where the renamed register will be used, |
| 1922 | // * not used in \p UsedInBetween; UsedInBetween must contain all accessed |
| 1923 | // registers in the range the rename register will be used, |
| 1924 | // * is available in all used register classes (checked using RequiredClasses). |
| 1925 | static std::optional<MCPhysReg> tryToFindRegisterToRename( |
| 1926 | const MachineFunction &MF, Register Reg, LiveRegUnits &DefinedInBB, |
| 1927 | LiveRegUnits &UsedInBetween, |
| 1928 | SmallPtrSetImpl<const TargetRegisterClass *> &RequiredClasses, |
| 1929 | const TargetRegisterInfo *TRI) { |
| 1930 | const MachineRegisterInfo &RegInfo = MF.getRegInfo(); |
| 1931 | |
| 1932 | // Checks if any sub- or super-register of PR is callee saved. |
| 1933 | auto AnySubOrSuperRegCalleePreserved = [&MF, TRI](MCPhysReg PR) { |
| 1934 | return any_of(Range: TRI->sub_and_superregs_inclusive(Reg: PR), |
| 1935 | P: [&MF, TRI](MCPhysReg SubOrSuper) { |
| 1936 | return TRI->isCalleeSavedPhysReg(PhysReg: SubOrSuper, MF); |
| 1937 | }); |
| 1938 | }; |
| 1939 | |
| 1940 | // Check if PR or one of its sub- or super-registers can be used for all |
| 1941 | // required register classes. |
| 1942 | auto CanBeUsedForAllClasses = [&RequiredClasses, TRI](MCPhysReg PR) { |
| 1943 | return all_of(Range&: RequiredClasses, P: [PR, TRI](const TargetRegisterClass *C) { |
| 1944 | return any_of( |
| 1945 | Range: TRI->sub_and_superregs_inclusive(Reg: PR), |
| 1946 | P: [C](MCPhysReg SubOrSuper) { return C->contains(Reg: SubOrSuper); }); |
| 1947 | }); |
| 1948 | }; |
| 1949 | |
| 1950 | auto *RegClass = TRI->getMinimalPhysRegClass(Reg); |
| 1951 | for (const MCPhysReg &PR : *RegClass) { |
| 1952 | if (DefinedInBB.available(Reg: PR) && UsedInBetween.available(Reg: PR) && |
| 1953 | !RegInfo.isReserved(PhysReg: PR) && !AnySubOrSuperRegCalleePreserved(PR) && |
| 1954 | CanBeUsedForAllClasses(PR)) { |
| 1955 | DefinedInBB.addReg(Reg: PR); |
| 1956 | LLVM_DEBUG(dbgs() << "Found rename register " << printReg(PR, TRI) |
| 1957 | << "\n" ); |
| 1958 | return {PR}; |
| 1959 | } |
| 1960 | } |
| 1961 | LLVM_DEBUG(dbgs() << "No rename register found from " |
| 1962 | << TRI->getRegClassName(RegClass) << "\n" ); |
| 1963 | return std::nullopt; |
| 1964 | } |
| 1965 | |
| 1966 | // For store pairs: returns a register from FirstMI to the beginning of the |
| 1967 | // block that can be renamed. |
| 1968 | // For load pairs: returns a register from FirstMI to MI that can be renamed. |
| 1969 | static std::optional<MCPhysReg> findRenameRegForSameLdStRegPair( |
| 1970 | std::optional<bool> MaybeCanRename, MachineInstr &FirstMI, MachineInstr &MI, |
| 1971 | Register Reg, LiveRegUnits &DefinedInBB, LiveRegUnits &UsedInBetween, |
| 1972 | SmallPtrSetImpl<const TargetRegisterClass *> &RequiredClasses, |
| 1973 | const TargetRegisterInfo *TRI) { |
| 1974 | std::optional<MCPhysReg> RenameReg; |
| 1975 | if (!DebugCounter::shouldExecute(Counter&: RegRenamingCounter)) |
| 1976 | return RenameReg; |
| 1977 | |
| 1978 | auto *RegClass = TRI->getMinimalPhysRegClass(Reg: getLdStRegOp(MI&: FirstMI).getReg()); |
| 1979 | MachineFunction &MF = *FirstMI.getParent()->getParent(); |
| 1980 | if (!RegClass || !MF.getRegInfo().tracksLiveness()) |
| 1981 | return RenameReg; |
| 1982 | |
| 1983 | const bool IsLoad = FirstMI.mayLoad(); |
| 1984 | |
| 1985 | if (!MaybeCanRename) { |
| 1986 | if (IsLoad) |
| 1987 | MaybeCanRename = {canRenameUntilSecondLoad(FirstLoad&: FirstMI, SecondLoad&: MI, UsedInBetween, |
| 1988 | RequiredClasses, TRI)}; |
| 1989 | else |
| 1990 | MaybeCanRename = { |
| 1991 | canRenameUpToDef(FirstMI, UsedInBetween, RequiredClasses, TRI)}; |
| 1992 | } |
| 1993 | |
| 1994 | if (*MaybeCanRename) { |
| 1995 | RenameReg = tryToFindRegisterToRename(MF, Reg, DefinedInBB, UsedInBetween, |
| 1996 | RequiredClasses, TRI); |
| 1997 | } |
| 1998 | return RenameReg; |
| 1999 | } |
| 2000 | |
| 2001 | /// Scan the instructions looking for a load/store that can be combined with the |
| 2002 | /// current instruction into a wider equivalent or a load/store pair. |
| 2003 | MachineBasicBlock::iterator |
| 2004 | AArch64LoadStoreOpt::findMatchingInsn(MachineBasicBlock::iterator I, |
| 2005 | LdStPairFlags &Flags, unsigned Limit, |
| 2006 | bool FindNarrowMerge) { |
| 2007 | MachineBasicBlock::iterator E = I->getParent()->end(); |
| 2008 | MachineBasicBlock::iterator MBBI = I; |
| 2009 | MachineBasicBlock::iterator MBBIWithRenameReg; |
| 2010 | MachineInstr &FirstMI = *I; |
| 2011 | MBBI = next_nodbg(It: MBBI, End: E); |
| 2012 | |
| 2013 | bool MayLoad = FirstMI.mayLoad(); |
| 2014 | bool IsUnscaled = TII->hasUnscaledLdStOffset(MI&: FirstMI); |
| 2015 | Register Reg = getLdStRegOp(MI&: FirstMI).getReg(); |
| 2016 | Register BaseReg = AArch64InstrInfo::getLdStBaseOp(MI: FirstMI).getReg(); |
| 2017 | int Offset = AArch64InstrInfo::getLdStOffsetOp(MI: FirstMI).getImm(); |
| 2018 | int OffsetStride = IsUnscaled ? TII->getMemScale(MI: FirstMI) : 1; |
| 2019 | bool IsPromotableZeroStore = isPromotableZeroStoreInst(MI&: FirstMI); |
| 2020 | |
| 2021 | std::optional<bool> MaybeCanRename; |
| 2022 | if (!EnableRenaming) |
| 2023 | MaybeCanRename = {false}; |
| 2024 | |
| 2025 | SmallPtrSet<const TargetRegisterClass *, 5> RequiredClasses; |
| 2026 | LiveRegUnits UsedInBetween; |
| 2027 | UsedInBetween.init(TRI: *TRI); |
| 2028 | |
| 2029 | Flags.clearRenameReg(); |
| 2030 | |
| 2031 | // Track which register units have been modified and used between the first |
| 2032 | // insn (inclusive) and the second insn. |
| 2033 | ModifiedRegUnits.clear(); |
| 2034 | UsedRegUnits.clear(); |
| 2035 | |
| 2036 | // Remember any instructions that read/write memory between FirstMI and MI. |
| 2037 | SmallVector<MachineInstr *, 4> MemInsns; |
| 2038 | |
| 2039 | LLVM_DEBUG(dbgs() << "Find match for: " ; FirstMI.dump()); |
| 2040 | for (unsigned Count = 0; MBBI != E && Count < Limit; |
| 2041 | MBBI = next_nodbg(It: MBBI, End: E)) { |
| 2042 | MachineInstr &MI = *MBBI; |
| 2043 | LLVM_DEBUG(dbgs() << "Analysing 2nd insn: " ; MI.dump()); |
| 2044 | |
| 2045 | UsedInBetween.accumulate(MI); |
| 2046 | |
| 2047 | // Don't count transient instructions towards the search limit since there |
| 2048 | // may be different numbers of them if e.g. debug information is present. |
| 2049 | if (!MI.isTransient()) |
| 2050 | ++Count; |
| 2051 | |
| 2052 | Flags.setSExtIdx(-1); |
| 2053 | if (areCandidatesToMergeOrPair(FirstMI, MI, Flags, TII) && |
| 2054 | AArch64InstrInfo::getLdStOffsetOp(MI).isImm()) { |
| 2055 | assert(MI.mayLoadOrStore() && "Expected memory operation." ); |
| 2056 | // If we've found another instruction with the same opcode, check to see |
| 2057 | // if the base and offset are compatible with our starting instruction. |
| 2058 | // These instructions all have scaled immediate operands, so we just |
| 2059 | // check for +1/-1. Make sure to check the new instruction offset is |
| 2060 | // actually an immediate and not a symbolic reference destined for |
| 2061 | // a relocation. |
| 2062 | Register MIBaseReg = AArch64InstrInfo::getLdStBaseOp(MI).getReg(); |
| 2063 | int MIOffset = AArch64InstrInfo::getLdStOffsetOp(MI).getImm(); |
| 2064 | bool MIIsUnscaled = TII->hasUnscaledLdStOffset(MI); |
| 2065 | if (IsUnscaled != MIIsUnscaled) { |
| 2066 | // We're trying to pair instructions that differ in how they are scaled. |
| 2067 | // If FirstMI is scaled then scale the offset of MI accordingly. |
| 2068 | // Otherwise, do the opposite (i.e., make MI's offset unscaled). |
| 2069 | int MemSize = TII->getMemScale(MI); |
| 2070 | if (MIIsUnscaled) { |
| 2071 | // If the unscaled offset isn't a multiple of the MemSize, we can't |
| 2072 | // pair the operations together: bail and keep looking. |
| 2073 | if (MIOffset % MemSize) { |
| 2074 | LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, |
| 2075 | UsedRegUnits, TRI); |
| 2076 | MemInsns.push_back(Elt: &MI); |
| 2077 | continue; |
| 2078 | } |
| 2079 | MIOffset /= MemSize; |
| 2080 | } else { |
| 2081 | MIOffset *= MemSize; |
| 2082 | } |
| 2083 | } |
| 2084 | |
| 2085 | bool IsPreLdSt = isPreLdStPairCandidate(FirstMI, MI); |
| 2086 | |
| 2087 | if (BaseReg == MIBaseReg) { |
| 2088 | // If the offset of the second ld/st is not equal to the size of the |
| 2089 | // destination register it can’t be paired with a pre-index ld/st |
| 2090 | // pair. Additionally if the base reg is used or modified the operations |
| 2091 | // can't be paired: bail and keep looking. |
| 2092 | if (IsPreLdSt) { |
| 2093 | bool IsOutOfBounds = MIOffset != TII->getMemScale(MI); |
| 2094 | bool IsBaseRegUsed = !UsedRegUnits.available( |
| 2095 | Reg: AArch64InstrInfo::getLdStBaseOp(MI).getReg()); |
| 2096 | bool IsBaseRegModified = !ModifiedRegUnits.available( |
| 2097 | Reg: AArch64InstrInfo::getLdStBaseOp(MI).getReg()); |
| 2098 | // If the stored value and the address of the second instruction is |
| 2099 | // the same, it needs to be using the updated register and therefore |
| 2100 | // it must not be folded. |
| 2101 | bool IsMIRegTheSame = |
| 2102 | TRI->regsOverlap(RegA: getLdStRegOp(MI).getReg(), |
| 2103 | RegB: AArch64InstrInfo::getLdStBaseOp(MI).getReg()); |
| 2104 | if (IsOutOfBounds || IsBaseRegUsed || IsBaseRegModified || |
| 2105 | IsMIRegTheSame) { |
| 2106 | LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, |
| 2107 | UsedRegUnits, TRI); |
| 2108 | MemInsns.push_back(Elt: &MI); |
| 2109 | continue; |
| 2110 | } |
| 2111 | } else { |
| 2112 | if ((Offset != MIOffset + OffsetStride) && |
| 2113 | (Offset + OffsetStride != MIOffset)) { |
| 2114 | LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, |
| 2115 | UsedRegUnits, TRI); |
| 2116 | MemInsns.push_back(Elt: &MI); |
| 2117 | continue; |
| 2118 | } |
| 2119 | } |
| 2120 | |
| 2121 | int MinOffset = Offset < MIOffset ? Offset : MIOffset; |
| 2122 | if (FindNarrowMerge) { |
| 2123 | // If the alignment requirements of the scaled wide load/store |
| 2124 | // instruction can't express the offset of the scaled narrow input, |
| 2125 | // bail and keep looking. For promotable zero stores, allow only when |
| 2126 | // the stored value is the same (i.e., WZR). |
| 2127 | if ((!IsUnscaled && alignTo(Num: MinOffset, PowOf2: 2) != MinOffset) || |
| 2128 | (IsPromotableZeroStore && Reg != getLdStRegOp(MI).getReg())) { |
| 2129 | LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, |
| 2130 | UsedRegUnits, TRI); |
| 2131 | MemInsns.push_back(Elt: &MI); |
| 2132 | continue; |
| 2133 | } |
| 2134 | } else { |
| 2135 | // Pairwise instructions have a 7-bit signed offset field. Single |
| 2136 | // insns have a 12-bit unsigned offset field. If the resultant |
| 2137 | // immediate offset of merging these instructions is out of range for |
| 2138 | // a pairwise instruction, bail and keep looking. |
| 2139 | if (!inBoundsForPair(IsUnscaled, Offset: MinOffset, OffsetStride)) { |
| 2140 | LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, |
| 2141 | UsedRegUnits, TRI); |
| 2142 | MemInsns.push_back(Elt: &MI); |
| 2143 | LLVM_DEBUG(dbgs() << "Offset doesn't fit in immediate, " |
| 2144 | << "keep looking.\n" ); |
| 2145 | continue; |
| 2146 | } |
| 2147 | // If the alignment requirements of the paired (scaled) instruction |
| 2148 | // can't express the offset of the unscaled input, bail and keep |
| 2149 | // looking. |
| 2150 | if (IsUnscaled && (alignTo(Num: MinOffset, PowOf2: OffsetStride) != MinOffset)) { |
| 2151 | LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, |
| 2152 | UsedRegUnits, TRI); |
| 2153 | MemInsns.push_back(Elt: &MI); |
| 2154 | LLVM_DEBUG(dbgs() |
| 2155 | << "Offset doesn't fit due to alignment requirements, " |
| 2156 | << "keep looking.\n" ); |
| 2157 | continue; |
| 2158 | } |
| 2159 | } |
| 2160 | |
| 2161 | // If the BaseReg has been modified, then we cannot do the optimization. |
| 2162 | // For example, in the following pattern |
| 2163 | // ldr x1 [x2] |
| 2164 | // ldr x2 [x3] |
| 2165 | // ldr x4 [x2, #8], |
| 2166 | // the first and third ldr cannot be converted to ldp x1, x4, [x2] |
| 2167 | if (!ModifiedRegUnits.available(Reg: BaseReg)) |
| 2168 | return E; |
| 2169 | |
| 2170 | const bool SameLoadReg = MayLoad && TRI->isSuperOrSubRegisterEq( |
| 2171 | RegA: Reg, RegB: getLdStRegOp(MI).getReg()); |
| 2172 | |
| 2173 | // If the Rt of the second instruction (destination register of the |
| 2174 | // load) was not modified or used between the two instructions and none |
| 2175 | // of the instructions between the second and first alias with the |
| 2176 | // second, we can combine the second into the first. |
| 2177 | bool RtNotModified = |
| 2178 | ModifiedRegUnits.available(Reg: getLdStRegOp(MI).getReg()); |
| 2179 | bool RtNotUsed = !(MI.mayLoad() && !SameLoadReg && |
| 2180 | !UsedRegUnits.available(Reg: getLdStRegOp(MI).getReg())); |
| 2181 | |
| 2182 | LLVM_DEBUG(dbgs() << "Checking, can combine 2nd into 1st insn:\n" |
| 2183 | << "Reg '" << getLdStRegOp(MI) << "' not modified: " |
| 2184 | << (RtNotModified ? "true" : "false" ) << "\n" |
| 2185 | << "Reg '" << getLdStRegOp(MI) << "' not used: " |
| 2186 | << (RtNotUsed ? "true" : "false" ) << "\n" ); |
| 2187 | |
| 2188 | if (RtNotModified && RtNotUsed && !mayAlias(MIa&: MI, MemInsns, AA)) { |
| 2189 | // For pairs loading into the same reg, try to find a renaming |
| 2190 | // opportunity to allow the renaming of Reg between FirstMI and MI |
| 2191 | // and combine MI into FirstMI; otherwise bail and keep looking. |
| 2192 | if (SameLoadReg) { |
| 2193 | std::optional<MCPhysReg> RenameReg = |
| 2194 | findRenameRegForSameLdStRegPair(MaybeCanRename, FirstMI, MI, |
| 2195 | Reg, DefinedInBB, UsedInBetween, |
| 2196 | RequiredClasses, TRI); |
| 2197 | if (!RenameReg) { |
| 2198 | LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, |
| 2199 | UsedRegUnits, TRI); |
| 2200 | MemInsns.push_back(Elt: &MI); |
| 2201 | LLVM_DEBUG(dbgs() << "Can't find reg for renaming, " |
| 2202 | << "keep looking.\n" ); |
| 2203 | continue; |
| 2204 | } |
| 2205 | Flags.setRenameReg(*RenameReg); |
| 2206 | } |
| 2207 | |
| 2208 | Flags.setMergeForward(false); |
| 2209 | if (!SameLoadReg) |
| 2210 | Flags.clearRenameReg(); |
| 2211 | return MBBI; |
| 2212 | } |
| 2213 | |
| 2214 | // Likewise, if the Rt of the first instruction is not modified or used |
| 2215 | // between the two instructions and none of the instructions between the |
| 2216 | // first and the second alias with the first, we can combine the first |
| 2217 | // into the second. |
| 2218 | RtNotModified = !( |
| 2219 | MayLoad && !UsedRegUnits.available(Reg: getLdStRegOp(MI&: FirstMI).getReg())); |
| 2220 | |
| 2221 | LLVM_DEBUG(dbgs() << "Checking, can combine 1st into 2nd insn:\n" |
| 2222 | << "Reg '" << getLdStRegOp(FirstMI) |
| 2223 | << "' not modified: " |
| 2224 | << (RtNotModified ? "true" : "false" ) << "\n" ); |
| 2225 | |
| 2226 | if (RtNotModified && !mayAlias(MIa&: FirstMI, MemInsns, AA)) { |
| 2227 | if (ModifiedRegUnits.available(Reg: getLdStRegOp(MI&: FirstMI).getReg())) { |
| 2228 | Flags.setMergeForward(true); |
| 2229 | Flags.clearRenameReg(); |
| 2230 | return MBBI; |
| 2231 | } |
| 2232 | |
| 2233 | std::optional<MCPhysReg> RenameReg = findRenameRegForSameLdStRegPair( |
| 2234 | MaybeCanRename, FirstMI, MI, Reg, DefinedInBB, UsedInBetween, |
| 2235 | RequiredClasses, TRI); |
| 2236 | if (RenameReg) { |
| 2237 | Flags.setMergeForward(true); |
| 2238 | Flags.setRenameReg(*RenameReg); |
| 2239 | MBBIWithRenameReg = MBBI; |
| 2240 | } |
| 2241 | } |
| 2242 | LLVM_DEBUG(dbgs() << "Unable to combine these instructions due to " |
| 2243 | << "interference in between, keep looking.\n" ); |
| 2244 | } |
| 2245 | } |
| 2246 | |
| 2247 | if (Flags.getRenameReg()) |
| 2248 | return MBBIWithRenameReg; |
| 2249 | |
| 2250 | // If the instruction wasn't a matching load or store. Stop searching if we |
| 2251 | // encounter a call instruction that might modify memory. |
| 2252 | if (MI.isCall()) { |
| 2253 | LLVM_DEBUG(dbgs() << "Found a call, stop looking.\n" ); |
| 2254 | return E; |
| 2255 | } |
| 2256 | |
| 2257 | // Update modified / uses register units. |
| 2258 | LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, TRI); |
| 2259 | |
| 2260 | // Otherwise, if the base register is modified, we have no match, so |
| 2261 | // return early. |
| 2262 | if (!ModifiedRegUnits.available(Reg: BaseReg)) { |
| 2263 | LLVM_DEBUG(dbgs() << "Base reg is modified, stop looking.\n" ); |
| 2264 | return E; |
| 2265 | } |
| 2266 | |
| 2267 | // Update list of instructions that read/write memory. |
| 2268 | if (MI.mayLoadOrStore()) |
| 2269 | MemInsns.push_back(Elt: &MI); |
| 2270 | } |
| 2271 | return E; |
| 2272 | } |
| 2273 | |
| 2274 | static MachineBasicBlock::iterator |
| 2275 | maybeMoveCFI(MachineInstr &MI, MachineBasicBlock::iterator MaybeCFI) { |
| 2276 | assert((MI.getOpcode() == AArch64::SUBXri || |
| 2277 | MI.getOpcode() == AArch64::ADDXri) && |
| 2278 | "Expected a register update instruction" ); |
| 2279 | auto End = MI.getParent()->end(); |
| 2280 | if (MaybeCFI == End || |
| 2281 | MaybeCFI->getOpcode() != TargetOpcode::CFI_INSTRUCTION || |
| 2282 | !(MI.getFlag(Flag: MachineInstr::FrameSetup) || |
| 2283 | MI.getFlag(Flag: MachineInstr::FrameDestroy)) || |
| 2284 | MI.getOperand(i: 0).getReg() != AArch64::SP) |
| 2285 | return End; |
| 2286 | |
| 2287 | const MachineFunction &MF = *MI.getParent()->getParent(); |
| 2288 | unsigned CFIIndex = MaybeCFI->getOperand(i: 0).getCFIIndex(); |
| 2289 | const MCCFIInstruction &CFI = MF.getFrameInstructions()[CFIIndex]; |
| 2290 | switch (CFI.getOperation()) { |
| 2291 | case MCCFIInstruction::OpDefCfa: |
| 2292 | case MCCFIInstruction::OpDefCfaOffset: |
| 2293 | return MaybeCFI; |
| 2294 | default: |
| 2295 | return End; |
| 2296 | } |
| 2297 | } |
| 2298 | |
| 2299 | std::optional<MachineBasicBlock::iterator> AArch64LoadStoreOpt::mergeUpdateInsn( |
| 2300 | MachineBasicBlock::iterator I, MachineBasicBlock::iterator Update, |
| 2301 | bool IsForward, bool IsPreIdx, bool MergeEither) { |
| 2302 | assert((Update->getOpcode() == AArch64::ADDXri || |
| 2303 | Update->getOpcode() == AArch64::SUBXri) && |
| 2304 | "Unexpected base register update instruction to merge!" ); |
| 2305 | MachineBasicBlock::iterator E = I->getParent()->end(); |
| 2306 | MachineBasicBlock::iterator NextI = next_nodbg(It: I, End: E); |
| 2307 | |
| 2308 | // If updating the SP and the following instruction is CFA offset related CFI, |
| 2309 | // make sure the CFI follows the SP update either by merging at the location |
| 2310 | // of the update or by moving the CFI after the merged instruction. If unable |
| 2311 | // to do so, bail. |
| 2312 | MachineBasicBlock::iterator InsertPt = I; |
| 2313 | if (IsForward) { |
| 2314 | assert(IsPreIdx); |
| 2315 | if (auto CFI = maybeMoveCFI(MI&: *Update, MaybeCFI: next_nodbg(It: Update, End: E)); CFI != E) { |
| 2316 | if (MergeEither) { |
| 2317 | InsertPt = Update; |
| 2318 | } else { |
| 2319 | // Take care not to reorder CFIs. |
| 2320 | if (std::any_of(first: std::next(x: CFI), last: I, pred: [](const auto &Insn) { |
| 2321 | return Insn.getOpcode() == TargetOpcode::CFI_INSTRUCTION; |
| 2322 | })) |
| 2323 | return std::nullopt; |
| 2324 | |
| 2325 | MachineBasicBlock *MBB = InsertPt->getParent(); |
| 2326 | MBB->splice(Where: std::next(x: InsertPt), Other: MBB, From: CFI); |
| 2327 | } |
| 2328 | } |
| 2329 | } |
| 2330 | |
| 2331 | // Return the instruction following the merged instruction, which is |
| 2332 | // the instruction following our unmerged load. Unless that's the add/sub |
| 2333 | // instruction we're merging, in which case it's the one after that. |
| 2334 | if (NextI == Update) |
| 2335 | NextI = next_nodbg(It: NextI, End: E); |
| 2336 | |
| 2337 | int Value = Update->getOperand(i: 2).getImm(); |
| 2338 | assert(AArch64_AM::getShiftValue(Update->getOperand(3).getImm()) == 0 && |
| 2339 | "Can't merge 1 << 12 offset into pre-/post-indexed load / store" ); |
| 2340 | if (Update->getOpcode() == AArch64::SUBXri) |
| 2341 | Value = -Value; |
| 2342 | |
| 2343 | unsigned NewOpc = IsPreIdx ? getPreIndexedOpcode(Opc: I->getOpcode()) |
| 2344 | : getPostIndexedOpcode(Opc: I->getOpcode()); |
| 2345 | MachineInstrBuilder MIB; |
| 2346 | int Scale, MinOffset, MaxOffset; |
| 2347 | getPrePostIndexedMemOpInfo(MI: *I, Scale, MinOffset, MaxOffset); |
| 2348 | if (!AArch64InstrInfo::isPairedLdSt(MI: *I)) { |
| 2349 | // Non-paired instruction. |
| 2350 | MIB = BuildMI(BB&: *InsertPt->getParent(), I: InsertPt, MIMD: InsertPt->getDebugLoc(), |
| 2351 | MCID: TII->get(Opcode: NewOpc)) |
| 2352 | .add(MO: Update->getOperand(i: 0)) |
| 2353 | .add(MO: getLdStRegOp(MI&: *I)) |
| 2354 | .add(MO: AArch64InstrInfo::getLdStBaseOp(MI: *I)) |
| 2355 | .addImm(Val: Value / Scale) |
| 2356 | .setMemRefs(I->memoperands()) |
| 2357 | .setMIFlags(I->mergeFlagsWith(Other: *Update)); |
| 2358 | } else { |
| 2359 | // Paired instruction. |
| 2360 | MIB = BuildMI(BB&: *InsertPt->getParent(), I: InsertPt, MIMD: InsertPt->getDebugLoc(), |
| 2361 | MCID: TII->get(Opcode: NewOpc)) |
| 2362 | .add(MO: Update->getOperand(i: 0)) |
| 2363 | .add(MO: getLdStRegOp(MI&: *I, PairedRegOp: 0)) |
| 2364 | .add(MO: getLdStRegOp(MI&: *I, PairedRegOp: 1)) |
| 2365 | .add(MO: AArch64InstrInfo::getLdStBaseOp(MI: *I)) |
| 2366 | .addImm(Val: Value / Scale) |
| 2367 | .setMemRefs(I->memoperands()) |
| 2368 | .setMIFlags(I->mergeFlagsWith(Other: *Update)); |
| 2369 | } |
| 2370 | |
| 2371 | if (IsPreIdx) { |
| 2372 | ++NumPreFolded; |
| 2373 | LLVM_DEBUG(dbgs() << "Creating pre-indexed load/store." ); |
| 2374 | } else { |
| 2375 | ++NumPostFolded; |
| 2376 | LLVM_DEBUG(dbgs() << "Creating post-indexed load/store." ); |
| 2377 | } |
| 2378 | LLVM_DEBUG(dbgs() << " Replacing instructions:\n " ); |
| 2379 | LLVM_DEBUG(I->print(dbgs())); |
| 2380 | LLVM_DEBUG(dbgs() << " " ); |
| 2381 | LLVM_DEBUG(Update->print(dbgs())); |
| 2382 | LLVM_DEBUG(dbgs() << " with instruction:\n " ); |
| 2383 | LLVM_DEBUG(((MachineInstr *)MIB)->print(dbgs())); |
| 2384 | LLVM_DEBUG(dbgs() << "\n" ); |
| 2385 | |
| 2386 | // Erase the old instructions for the block. |
| 2387 | I->eraseFromParent(); |
| 2388 | Update->eraseFromParent(); |
| 2389 | |
| 2390 | return NextI; |
| 2391 | } |
| 2392 | |
| 2393 | MachineBasicBlock::iterator |
| 2394 | AArch64LoadStoreOpt::mergeConstOffsetInsn(MachineBasicBlock::iterator I, |
| 2395 | MachineBasicBlock::iterator Update, |
| 2396 | unsigned Offset, int Scale) { |
| 2397 | assert((Update->getOpcode() == AArch64::MOVKWi) && |
| 2398 | "Unexpected const mov instruction to merge!" ); |
| 2399 | MachineBasicBlock::iterator E = I->getParent()->end(); |
| 2400 | MachineBasicBlock::iterator NextI = next_nodbg(It: I, End: E); |
| 2401 | MachineBasicBlock::iterator PrevI = prev_nodbg(It: Update, Begin: E); |
| 2402 | MachineInstr &MemMI = *I; |
| 2403 | unsigned Mask = (1 << 12) * Scale - 1; |
| 2404 | unsigned Low = Offset & Mask; |
| 2405 | unsigned High = Offset - Low; |
| 2406 | Register BaseReg = AArch64InstrInfo::getLdStBaseOp(MI: MemMI).getReg(); |
| 2407 | Register IndexReg = AArch64InstrInfo::getLdStOffsetOp(MI: MemMI).getReg(); |
| 2408 | MachineInstrBuilder AddMIB, MemMIB; |
| 2409 | |
| 2410 | // Add IndexReg, BaseReg, High (the BaseReg may be SP) |
| 2411 | AddMIB = |
| 2412 | BuildMI(BB&: *I->getParent(), I, MIMD: I->getDebugLoc(), MCID: TII->get(Opcode: AArch64::ADDXri)) |
| 2413 | .addDef(RegNo: IndexReg) |
| 2414 | .addUse(RegNo: BaseReg) |
| 2415 | .addImm(Val: High >> 12) // shifted value |
| 2416 | .addImm(Val: 12); // shift 12 |
| 2417 | (void)AddMIB; |
| 2418 | // Ld/St DestReg, IndexReg, Imm12 |
| 2419 | unsigned NewOpc = getBaseAddressOpcode(Opc: I->getOpcode()); |
| 2420 | MemMIB = BuildMI(BB&: *I->getParent(), I, MIMD: I->getDebugLoc(), MCID: TII->get(Opcode: NewOpc)) |
| 2421 | .add(MO: getLdStRegOp(MI&: MemMI)) |
| 2422 | .add(MO: AArch64InstrInfo::getLdStOffsetOp(MI: MemMI)) |
| 2423 | .addImm(Val: Low / Scale) |
| 2424 | .setMemRefs(I->memoperands()) |
| 2425 | .setMIFlags(I->mergeFlagsWith(Other: *Update)); |
| 2426 | (void)MemMIB; |
| 2427 | |
| 2428 | ++NumConstOffsetFolded; |
| 2429 | LLVM_DEBUG(dbgs() << "Creating base address load/store.\n" ); |
| 2430 | LLVM_DEBUG(dbgs() << " Replacing instructions:\n " ); |
| 2431 | LLVM_DEBUG(PrevI->print(dbgs())); |
| 2432 | LLVM_DEBUG(dbgs() << " " ); |
| 2433 | LLVM_DEBUG(Update->print(dbgs())); |
| 2434 | LLVM_DEBUG(dbgs() << " " ); |
| 2435 | LLVM_DEBUG(I->print(dbgs())); |
| 2436 | LLVM_DEBUG(dbgs() << " with instruction:\n " ); |
| 2437 | LLVM_DEBUG(((MachineInstr *)AddMIB)->print(dbgs())); |
| 2438 | LLVM_DEBUG(dbgs() << " " ); |
| 2439 | LLVM_DEBUG(((MachineInstr *)MemMIB)->print(dbgs())); |
| 2440 | LLVM_DEBUG(dbgs() << "\n" ); |
| 2441 | |
| 2442 | // Erase the old instructions for the block. |
| 2443 | I->eraseFromParent(); |
| 2444 | PrevI->eraseFromParent(); |
| 2445 | Update->eraseFromParent(); |
| 2446 | |
| 2447 | return NextI; |
| 2448 | } |
| 2449 | |
| 2450 | bool AArch64LoadStoreOpt::isMatchingUpdateInsn(MachineInstr &MemMI, |
| 2451 | MachineInstr &MI, |
| 2452 | unsigned BaseReg, int Offset) { |
| 2453 | switch (MI.getOpcode()) { |
| 2454 | default: |
| 2455 | break; |
| 2456 | case AArch64::SUBXri: |
| 2457 | case AArch64::ADDXri: |
| 2458 | // Make sure it's a vanilla immediate operand, not a relocation or |
| 2459 | // anything else we can't handle. |
| 2460 | if (!MI.getOperand(i: 2).isImm()) |
| 2461 | break; |
| 2462 | // Watch out for 1 << 12 shifted value. |
| 2463 | if (AArch64_AM::getShiftValue(Imm: MI.getOperand(i: 3).getImm())) |
| 2464 | break; |
| 2465 | |
| 2466 | // The update instruction source and destination register must be the |
| 2467 | // same as the load/store base register. |
| 2468 | if (MI.getOperand(i: 0).getReg() != BaseReg || |
| 2469 | MI.getOperand(i: 1).getReg() != BaseReg) |
| 2470 | break; |
| 2471 | |
| 2472 | int UpdateOffset = MI.getOperand(i: 2).getImm(); |
| 2473 | if (MI.getOpcode() == AArch64::SUBXri) |
| 2474 | UpdateOffset = -UpdateOffset; |
| 2475 | |
| 2476 | // The immediate must be a multiple of the scaling factor of the pre/post |
| 2477 | // indexed instruction. |
| 2478 | int Scale, MinOffset, MaxOffset; |
| 2479 | getPrePostIndexedMemOpInfo(MI: MemMI, Scale, MinOffset, MaxOffset); |
| 2480 | if (UpdateOffset % Scale != 0) |
| 2481 | break; |
| 2482 | |
| 2483 | // Scaled offset must fit in the instruction immediate. |
| 2484 | int ScaledOffset = UpdateOffset / Scale; |
| 2485 | if (ScaledOffset > MaxOffset || ScaledOffset < MinOffset) |
| 2486 | break; |
| 2487 | |
| 2488 | // If we have a non-zero Offset, we check that it matches the amount |
| 2489 | // we're adding to the register. |
| 2490 | if (!Offset || Offset == UpdateOffset) |
| 2491 | return true; |
| 2492 | break; |
| 2493 | } |
| 2494 | return false; |
| 2495 | } |
| 2496 | |
| 2497 | bool AArch64LoadStoreOpt::isMatchingMovConstInsn(MachineInstr &MemMI, |
| 2498 | MachineInstr &MI, |
| 2499 | unsigned IndexReg, |
| 2500 | unsigned &Offset) { |
| 2501 | // The update instruction source and destination register must be the |
| 2502 | // same as the load/store index register. |
| 2503 | if (MI.getOpcode() == AArch64::MOVKWi && |
| 2504 | TRI->isSuperOrSubRegisterEq(RegA: IndexReg, RegB: MI.getOperand(i: 1).getReg())) { |
| 2505 | |
| 2506 | // movz + movk hold a large offset of a Ld/St instruction. |
| 2507 | MachineBasicBlock::iterator B = MI.getParent()->begin(); |
| 2508 | MachineBasicBlock::iterator MBBI = &MI; |
| 2509 | // Skip the scene when the MI is the first instruction of a block. |
| 2510 | if (MBBI == B) |
| 2511 | return false; |
| 2512 | MBBI = prev_nodbg(It: MBBI, Begin: B); |
| 2513 | MachineInstr &MovzMI = *MBBI; |
| 2514 | // Make sure the MOVKWi and MOVZWi set the same register. |
| 2515 | if (MovzMI.getOpcode() == AArch64::MOVZWi && |
| 2516 | MovzMI.getOperand(i: 0).getReg() == MI.getOperand(i: 0).getReg()) { |
| 2517 | unsigned Low = MovzMI.getOperand(i: 1).getImm(); |
| 2518 | unsigned High = MI.getOperand(i: 2).getImm() << MI.getOperand(i: 3).getImm(); |
| 2519 | Offset = High + Low; |
| 2520 | // 12-bit optionally shifted immediates are legal for adds. |
| 2521 | return Offset >> 24 == 0; |
| 2522 | } |
| 2523 | } |
| 2524 | return false; |
| 2525 | } |
| 2526 | |
| 2527 | MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnForward( |
| 2528 | MachineBasicBlock::iterator I, int UnscaledOffset, unsigned Limit) { |
| 2529 | MachineBasicBlock::iterator E = I->getParent()->end(); |
| 2530 | MachineInstr &MemMI = *I; |
| 2531 | MachineBasicBlock::iterator MBBI = I; |
| 2532 | |
| 2533 | Register BaseReg = AArch64InstrInfo::getLdStBaseOp(MI: MemMI).getReg(); |
| 2534 | int MIUnscaledOffset = AArch64InstrInfo::getLdStOffsetOp(MI: MemMI).getImm() * |
| 2535 | TII->getMemScale(MI: MemMI); |
| 2536 | |
| 2537 | // Scan forward looking for post-index opportunities. Updating instructions |
| 2538 | // can't be formed if the memory instruction doesn't have the offset we're |
| 2539 | // looking for. |
| 2540 | if (MIUnscaledOffset != UnscaledOffset) |
| 2541 | return E; |
| 2542 | |
| 2543 | // If the base register overlaps a source/destination register, we can't |
| 2544 | // merge the update. This does not apply to tag store instructions which |
| 2545 | // ignore the address part of the source register. |
| 2546 | // This does not apply to STGPi as well, which does not have unpredictable |
| 2547 | // behavior in this case unlike normal stores, and always performs writeback |
| 2548 | // after reading the source register value. |
| 2549 | if (!isTagStore(MI: MemMI) && MemMI.getOpcode() != AArch64::STGPi) { |
| 2550 | bool IsPairedInsn = AArch64InstrInfo::isPairedLdSt(MI: MemMI); |
| 2551 | for (unsigned i = 0, e = IsPairedInsn ? 2 : 1; i != e; ++i) { |
| 2552 | Register DestReg = getLdStRegOp(MI&: MemMI, PairedRegOp: i).getReg(); |
| 2553 | if (DestReg == BaseReg || TRI->isSubRegister(RegA: BaseReg, RegB: DestReg)) |
| 2554 | return E; |
| 2555 | } |
| 2556 | } |
| 2557 | |
| 2558 | // Track which register units have been modified and used between the first |
| 2559 | // insn (inclusive) and the second insn. |
| 2560 | ModifiedRegUnits.clear(); |
| 2561 | UsedRegUnits.clear(); |
| 2562 | MBBI = next_nodbg(It: MBBI, End: E); |
| 2563 | |
| 2564 | // We can't post-increment the stack pointer if any instruction between |
| 2565 | // the memory access (I) and the increment (MBBI) can access the memory |
| 2566 | // region defined by [SP, MBBI]. |
| 2567 | const bool BaseRegSP = BaseReg == AArch64::SP; |
| 2568 | if (BaseRegSP && needsWinCFI(MF: I->getMF())) { |
| 2569 | // FIXME: For now, we always block the optimization over SP in windows |
| 2570 | // targets as it requires to adjust the unwind/debug info, messing up |
| 2571 | // the unwind info can actually cause a miscompile. |
| 2572 | return E; |
| 2573 | } |
| 2574 | |
| 2575 | unsigned Count = 0; |
| 2576 | MachineBasicBlock *CurMBB = I->getParent(); |
| 2577 | // choice of next block to visit is liveins-based |
| 2578 | bool VisitSucc = CurMBB->getParent()->getRegInfo().tracksLiveness(); |
| 2579 | |
| 2580 | while (true) { |
| 2581 | for (MachineBasicBlock::iterator CurEnd = CurMBB->end(); |
| 2582 | MBBI != CurEnd && Count < Limit; MBBI = next_nodbg(It: MBBI, End: CurEnd)) { |
| 2583 | MachineInstr &MI = *MBBI; |
| 2584 | |
| 2585 | // Don't count transient instructions towards the search limit since there |
| 2586 | // may be different numbers of them if e.g. debug information is present. |
| 2587 | if (!MI.isTransient()) |
| 2588 | ++Count; |
| 2589 | |
| 2590 | // If we found a match, return it. |
| 2591 | if (isMatchingUpdateInsn(MemMI&: *I, MI, BaseReg, Offset: UnscaledOffset)) |
| 2592 | return MBBI; |
| 2593 | |
| 2594 | // Update the status of what the instruction clobbered and used. |
| 2595 | LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, |
| 2596 | TRI); |
| 2597 | |
| 2598 | // Otherwise, if the base register is used or modified, we have no match, |
| 2599 | // so return early. If we are optimizing SP, do not allow instructions |
| 2600 | // that may load or store in between the load and the optimized value |
| 2601 | // update. |
| 2602 | if (!ModifiedRegUnits.available(Reg: BaseReg) || |
| 2603 | !UsedRegUnits.available(Reg: BaseReg) || |
| 2604 | (BaseRegSP && MBBI->mayLoadOrStore())) |
| 2605 | return E; |
| 2606 | } |
| 2607 | |
| 2608 | if (!VisitSucc || Limit <= Count) |
| 2609 | break; |
| 2610 | |
| 2611 | // Try to go downward to successors along a CF path w/o side enters |
| 2612 | // such that BaseReg is alive along it but not at its exits |
| 2613 | MachineBasicBlock *SuccToVisit = nullptr; |
| 2614 | unsigned LiveSuccCount = 0; |
| 2615 | for (MachineBasicBlock *Succ : CurMBB->successors()) { |
| 2616 | for (MCRegAliasIterator AI(BaseReg, TRI, true); AI.isValid(); ++AI) { |
| 2617 | if (Succ->isLiveIn(Reg: *AI)) { |
| 2618 | if (LiveSuccCount++) |
| 2619 | return E; |
| 2620 | if (Succ->pred_size() == 1) |
| 2621 | SuccToVisit = Succ; |
| 2622 | break; |
| 2623 | } |
| 2624 | } |
| 2625 | } |
| 2626 | if (!SuccToVisit) |
| 2627 | break; |
| 2628 | CurMBB = SuccToVisit; |
| 2629 | MBBI = CurMBB->begin(); |
| 2630 | } |
| 2631 | |
| 2632 | return E; |
| 2633 | } |
| 2634 | |
| 2635 | MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnBackward( |
| 2636 | MachineBasicBlock::iterator I, unsigned Limit, bool &MergeEither) { |
| 2637 | MachineBasicBlock::iterator B = I->getParent()->begin(); |
| 2638 | MachineBasicBlock::iterator E = I->getParent()->end(); |
| 2639 | MachineInstr &MemMI = *I; |
| 2640 | MachineBasicBlock::iterator MBBI = I; |
| 2641 | MachineFunction &MF = *MemMI.getMF(); |
| 2642 | |
| 2643 | Register BaseReg = AArch64InstrInfo::getLdStBaseOp(MI: MemMI).getReg(); |
| 2644 | int Offset = AArch64InstrInfo::getLdStOffsetOp(MI: MemMI).getImm(); |
| 2645 | |
| 2646 | bool IsPairedInsn = AArch64InstrInfo::isPairedLdSt(MI: MemMI); |
| 2647 | Register DestReg[] = {getLdStRegOp(MI&: MemMI, PairedRegOp: 0).getReg(), |
| 2648 | IsPairedInsn ? getLdStRegOp(MI&: MemMI, PairedRegOp: 1).getReg() |
| 2649 | : AArch64::NoRegister}; |
| 2650 | |
| 2651 | // If the load/store is the first instruction in the block, there's obviously |
| 2652 | // not any matching update. Ditto if the memory offset isn't zero. |
| 2653 | if (MBBI == B || Offset != 0) |
| 2654 | return E; |
| 2655 | // If the base register overlaps a destination register, we can't |
| 2656 | // merge the update. |
| 2657 | if (!isTagStore(MI: MemMI)) { |
| 2658 | for (unsigned i = 0, e = IsPairedInsn ? 2 : 1; i != e; ++i) |
| 2659 | if (DestReg[i] == BaseReg || TRI->isSubRegister(RegA: BaseReg, RegB: DestReg[i])) |
| 2660 | return E; |
| 2661 | } |
| 2662 | |
| 2663 | const bool BaseRegSP = BaseReg == AArch64::SP; |
| 2664 | if (BaseRegSP && needsWinCFI(MF: I->getMF())) { |
| 2665 | // FIXME: For now, we always block the optimization over SP in windows |
| 2666 | // targets as it requires to adjust the unwind/debug info, messing up |
| 2667 | // the unwind info can actually cause a miscompile. |
| 2668 | return E; |
| 2669 | } |
| 2670 | |
| 2671 | const AArch64Subtarget &Subtarget = MF.getSubtarget<AArch64Subtarget>(); |
| 2672 | unsigned RedZoneSize = |
| 2673 | Subtarget.getTargetLowering()->getRedZoneSize(F: MF.getFunction()); |
| 2674 | |
| 2675 | // Track which register units have been modified and used between the first |
| 2676 | // insn (inclusive) and the second insn. |
| 2677 | ModifiedRegUnits.clear(); |
| 2678 | UsedRegUnits.clear(); |
| 2679 | unsigned Count = 0; |
| 2680 | bool MemAccessBeforeSPPreInc = false; |
| 2681 | MergeEither = true; |
| 2682 | do { |
| 2683 | MBBI = prev_nodbg(It: MBBI, Begin: B); |
| 2684 | MachineInstr &MI = *MBBI; |
| 2685 | |
| 2686 | // Don't count transient instructions towards the search limit since there |
| 2687 | // may be different numbers of them if e.g. debug information is present. |
| 2688 | if (!MI.isTransient()) |
| 2689 | ++Count; |
| 2690 | |
| 2691 | // If we found a match, return it. |
| 2692 | if (isMatchingUpdateInsn(MemMI&: *I, MI, BaseReg, Offset)) { |
| 2693 | // Check that the update value is within our red zone limit (which may be |
| 2694 | // zero). |
| 2695 | if (MemAccessBeforeSPPreInc && MBBI->getOperand(i: 2).getImm() > RedZoneSize) |
| 2696 | return E; |
| 2697 | return MBBI; |
| 2698 | } |
| 2699 | |
| 2700 | // Update the status of what the instruction clobbered and used. |
| 2701 | LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, TRI); |
| 2702 | |
| 2703 | // Otherwise, if the base register is used or modified, we have no match, so |
| 2704 | // return early. |
| 2705 | if (!ModifiedRegUnits.available(Reg: BaseReg) || |
| 2706 | !UsedRegUnits.available(Reg: BaseReg)) |
| 2707 | return E; |
| 2708 | |
| 2709 | // If we have a destination register (i.e. a load instruction) and a |
| 2710 | // destination register is used or modified, then we can only merge forward, |
| 2711 | // i.e. the combined instruction is put in the place of the memory |
| 2712 | // instruction. Same applies if we see a memory access or side effects. |
| 2713 | if (MI.mayLoadOrStore() || MI.hasUnmodeledSideEffects() || |
| 2714 | (DestReg[0] != AArch64::NoRegister && |
| 2715 | !(ModifiedRegUnits.available(Reg: DestReg[0]) && |
| 2716 | UsedRegUnits.available(Reg: DestReg[0]))) || |
| 2717 | (DestReg[1] != AArch64::NoRegister && |
| 2718 | !(ModifiedRegUnits.available(Reg: DestReg[1]) && |
| 2719 | UsedRegUnits.available(Reg: DestReg[1])))) |
| 2720 | MergeEither = false; |
| 2721 | |
| 2722 | // Keep track if we have a memory access before an SP pre-increment, in this |
| 2723 | // case we need to validate later that the update amount respects the red |
| 2724 | // zone. |
| 2725 | if (BaseRegSP && MBBI->mayLoadOrStore()) |
| 2726 | MemAccessBeforeSPPreInc = true; |
| 2727 | } while (MBBI != B && Count < Limit); |
| 2728 | return E; |
| 2729 | } |
| 2730 | |
| 2731 | MachineBasicBlock::iterator |
| 2732 | AArch64LoadStoreOpt::findMatchingConstOffsetBackward( |
| 2733 | MachineBasicBlock::iterator I, unsigned Limit, unsigned &Offset) { |
| 2734 | MachineBasicBlock::iterator B = I->getParent()->begin(); |
| 2735 | MachineBasicBlock::iterator E = I->getParent()->end(); |
| 2736 | MachineInstr &MemMI = *I; |
| 2737 | MachineBasicBlock::iterator MBBI = I; |
| 2738 | |
| 2739 | // If the load is the first instruction in the block, there's obviously |
| 2740 | // not any matching load or store. |
| 2741 | if (MBBI == B) |
| 2742 | return E; |
| 2743 | |
| 2744 | // Make sure the IndexReg is killed and the shift amount is zero. |
| 2745 | // TODO: Relex this restriction to extend, simplify processing now. |
| 2746 | if (!AArch64InstrInfo::getLdStOffsetOp(MI: MemMI).isKill() || |
| 2747 | !AArch64InstrInfo::getLdStAmountOp(MI: MemMI).isImm() || |
| 2748 | (AArch64InstrInfo::getLdStAmountOp(MI: MemMI).getImm() != 0)) |
| 2749 | return E; |
| 2750 | |
| 2751 | Register IndexReg = AArch64InstrInfo::getLdStOffsetOp(MI: MemMI).getReg(); |
| 2752 | |
| 2753 | // Track which register units have been modified and used between the first |
| 2754 | // insn (inclusive) and the second insn. |
| 2755 | ModifiedRegUnits.clear(); |
| 2756 | UsedRegUnits.clear(); |
| 2757 | unsigned Count = 0; |
| 2758 | do { |
| 2759 | MBBI = prev_nodbg(It: MBBI, Begin: B); |
| 2760 | MachineInstr &MI = *MBBI; |
| 2761 | |
| 2762 | // Don't count transient instructions towards the search limit since there |
| 2763 | // may be different numbers of them if e.g. debug information is present. |
| 2764 | if (!MI.isTransient()) |
| 2765 | ++Count; |
| 2766 | |
| 2767 | // If we found a match, return it. |
| 2768 | if (isMatchingMovConstInsn(MemMI&: *I, MI, IndexReg, Offset)) { |
| 2769 | return MBBI; |
| 2770 | } |
| 2771 | |
| 2772 | // Update the status of what the instruction clobbered and used. |
| 2773 | LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, TRI); |
| 2774 | |
| 2775 | // Otherwise, if the index register is used or modified, we have no match, |
| 2776 | // so return early. |
| 2777 | if (!ModifiedRegUnits.available(Reg: IndexReg) || |
| 2778 | !UsedRegUnits.available(Reg: IndexReg)) |
| 2779 | return E; |
| 2780 | |
| 2781 | } while (MBBI != B && Count < Limit); |
| 2782 | return E; |
| 2783 | } |
| 2784 | |
| 2785 | bool AArch64LoadStoreOpt::tryToPromoteLoadFromStore( |
| 2786 | MachineBasicBlock::iterator &MBBI) { |
| 2787 | MachineInstr &MI = *MBBI; |
| 2788 | // If this is a volatile load, don't mess with it. |
| 2789 | if (MI.hasOrderedMemoryRef()) |
| 2790 | return false; |
| 2791 | |
| 2792 | if (needsWinCFI(MF: MI.getMF()) && MI.getFlag(Flag: MachineInstr::FrameDestroy)) |
| 2793 | return false; |
| 2794 | |
| 2795 | // Make sure this is a reg+imm. |
| 2796 | // FIXME: It is possible to extend it to handle reg+reg cases. |
| 2797 | if (!AArch64InstrInfo::getLdStOffsetOp(MI).isImm()) |
| 2798 | return false; |
| 2799 | |
| 2800 | // Look backward up to LdStLimit instructions. |
| 2801 | MachineBasicBlock::iterator StoreI; |
| 2802 | if (findMatchingStore(I: MBBI, Limit: LdStLimit, StoreI)) { |
| 2803 | ++NumLoadsFromStoresPromoted; |
| 2804 | // Promote the load. Keeping the iterator straight is a |
| 2805 | // pain, so we let the merge routine tell us what the next instruction |
| 2806 | // is after it's done mucking about. |
| 2807 | MBBI = promoteLoadFromStore(LoadI: MBBI, StoreI); |
| 2808 | return true; |
| 2809 | } |
| 2810 | return false; |
| 2811 | } |
| 2812 | |
| 2813 | // Merge adjacent zero stores into a wider store. |
| 2814 | bool AArch64LoadStoreOpt::tryToMergeZeroStInst( |
| 2815 | MachineBasicBlock::iterator &MBBI) { |
| 2816 | assert(isPromotableZeroStoreInst(*MBBI) && "Expected narrow store." ); |
| 2817 | MachineInstr &MI = *MBBI; |
| 2818 | MachineBasicBlock::iterator E = MI.getParent()->end(); |
| 2819 | |
| 2820 | if (!TII->isCandidateToMergeOrPair(MI)) |
| 2821 | return false; |
| 2822 | |
| 2823 | // Look ahead up to LdStLimit instructions for a mergeable instruction. |
| 2824 | LdStPairFlags Flags; |
| 2825 | MachineBasicBlock::iterator MergeMI = |
| 2826 | findMatchingInsn(I: MBBI, Flags, Limit: LdStLimit, /* FindNarrowMerge = */ true); |
| 2827 | if (MergeMI != E) { |
| 2828 | ++NumZeroStoresPromoted; |
| 2829 | |
| 2830 | // Keeping the iterator straight is a pain, so we let the merge routine tell |
| 2831 | // us what the next instruction is after it's done mucking about. |
| 2832 | MBBI = mergeNarrowZeroStores(I: MBBI, MergeMI, Flags); |
| 2833 | return true; |
| 2834 | } |
| 2835 | return false; |
| 2836 | } |
| 2837 | |
| 2838 | // Find loads and stores that can be merged into a single load or store pair |
| 2839 | // instruction. |
| 2840 | bool AArch64LoadStoreOpt::tryToPairLdStInst(MachineBasicBlock::iterator &MBBI) { |
| 2841 | MachineInstr &MI = *MBBI; |
| 2842 | MachineBasicBlock::iterator E = MI.getParent()->end(); |
| 2843 | |
| 2844 | if (!TII->isCandidateToMergeOrPair(MI)) |
| 2845 | return false; |
| 2846 | |
| 2847 | // If disable-ldp feature is opted, do not emit ldp. |
| 2848 | if (MI.mayLoad() && Subtarget->hasDisableLdp()) |
| 2849 | return false; |
| 2850 | |
| 2851 | // If disable-stp feature is opted, do not emit stp. |
| 2852 | if (MI.mayStore() && Subtarget->hasDisableStp()) |
| 2853 | return false; |
| 2854 | |
| 2855 | // Early exit if the offset is not possible to match. (6 bits of positive |
| 2856 | // range, plus allow an extra one in case we find a later insn that matches |
| 2857 | // with Offset-1) |
| 2858 | bool IsUnscaled = TII->hasUnscaledLdStOffset(MI); |
| 2859 | int Offset = AArch64InstrInfo::getLdStOffsetOp(MI).getImm(); |
| 2860 | int OffsetStride = IsUnscaled ? TII->getMemScale(MI) : 1; |
| 2861 | // Allow one more for offset. |
| 2862 | if (Offset > 0) |
| 2863 | Offset -= OffsetStride; |
| 2864 | if (!inBoundsForPair(IsUnscaled, Offset, OffsetStride)) |
| 2865 | return false; |
| 2866 | |
| 2867 | // Look ahead up to LdStLimit instructions for a pairable instruction. |
| 2868 | LdStPairFlags Flags; |
| 2869 | MachineBasicBlock::iterator Paired = |
| 2870 | findMatchingInsn(I: MBBI, Flags, Limit: LdStLimit, /* FindNarrowMerge = */ false); |
| 2871 | if (Paired != E) { |
| 2872 | // Keeping the iterator straight is a pain, so we let the merge routine tell |
| 2873 | // us what the next instruction is after it's done mucking about. |
| 2874 | auto Prev = std::prev(x: MBBI); |
| 2875 | |
| 2876 | // Fetch the memoperand of the load/store that is a candidate for |
| 2877 | // combination. |
| 2878 | MachineMemOperand *MemOp = |
| 2879 | MI.memoperands_empty() ? nullptr : MI.memoperands().front(); |
| 2880 | |
| 2881 | // If a load/store arrives and ldp/stp-aligned-only feature is opted, check |
| 2882 | // that the alignment of the source pointer is at least double the alignment |
| 2883 | // of the type. |
| 2884 | if ((MI.mayLoad() && Subtarget->hasLdpAlignedOnly()) || |
| 2885 | (MI.mayStore() && Subtarget->hasStpAlignedOnly())) { |
| 2886 | // If there is no size/align information, cancel the transformation. |
| 2887 | if (!MemOp || !MemOp->getMemoryType().isValid()) { |
| 2888 | NumFailedAlignmentCheck++; |
| 2889 | return false; |
| 2890 | } |
| 2891 | |
| 2892 | // Get the needed alignments to check them if |
| 2893 | // ldp-aligned-only/stp-aligned-only features are opted. |
| 2894 | uint64_t MemAlignment = MemOp->getAlign().value(); |
| 2895 | uint64_t TypeAlignment = |
| 2896 | Align(MemOp->getSize().getValue().getKnownMinValue()).value(); |
| 2897 | |
| 2898 | if (MemAlignment < 2 * TypeAlignment) { |
| 2899 | NumFailedAlignmentCheck++; |
| 2900 | return false; |
| 2901 | } |
| 2902 | } |
| 2903 | |
| 2904 | ++NumPairCreated; |
| 2905 | if (TII->hasUnscaledLdStOffset(MI)) |
| 2906 | ++NumUnscaledPairCreated; |
| 2907 | |
| 2908 | MBBI = mergePairedInsns(I: MBBI, Paired, Flags); |
| 2909 | // Collect liveness info for instructions between Prev and the new position |
| 2910 | // MBBI. |
| 2911 | for (auto I = std::next(x: Prev); I != MBBI; I++) |
| 2912 | updateDefinedRegisters(MI&: *I, Units&: DefinedInBB, TRI); |
| 2913 | |
| 2914 | return true; |
| 2915 | } |
| 2916 | return false; |
| 2917 | } |
| 2918 | |
| 2919 | bool AArch64LoadStoreOpt::tryToMergeLdStUpdate |
| 2920 | (MachineBasicBlock::iterator &MBBI) { |
| 2921 | MachineInstr &MI = *MBBI; |
| 2922 | MachineBasicBlock::iterator E = MI.getParent()->end(); |
| 2923 | MachineBasicBlock::iterator Update; |
| 2924 | |
| 2925 | // Look forward to try to form a post-index instruction. For example, |
| 2926 | // ldr x0, [x20] |
| 2927 | // add x20, x20, #32 |
| 2928 | // merged into: |
| 2929 | // ldr x0, [x20], #32 |
| 2930 | Update = findMatchingUpdateInsnForward(I: MBBI, UnscaledOffset: 0, Limit: UpdateLimit); |
| 2931 | if (Update != E) { |
| 2932 | // Merge the update into the ld/st. |
| 2933 | if (auto NextI = mergeUpdateInsn(I: MBBI, Update, /*IsForward=*/false, |
| 2934 | /*IsPreIdx=*/false, |
| 2935 | /*MergeEither=*/false)) { |
| 2936 | MBBI = *NextI; |
| 2937 | return true; |
| 2938 | } |
| 2939 | } |
| 2940 | |
| 2941 | // Don't know how to handle unscaled pre/post-index versions below, so bail. |
| 2942 | if (TII->hasUnscaledLdStOffset(Opc: MI.getOpcode())) |
| 2943 | return false; |
| 2944 | |
| 2945 | // Look back to try to find a pre-index instruction. For example, |
| 2946 | // add x0, x0, #8 |
| 2947 | // ldr x1, [x0] |
| 2948 | // merged into: |
| 2949 | // ldr x1, [x0, #8]! |
| 2950 | bool MergeEither; |
| 2951 | Update = findMatchingUpdateInsnBackward(I: MBBI, Limit: UpdateLimit, MergeEither); |
| 2952 | if (Update != E) { |
| 2953 | // Merge the update into the ld/st. |
| 2954 | if (auto NextI = mergeUpdateInsn(I: MBBI, Update, /*IsForward=*/true, |
| 2955 | /*IsPreIdx=*/true, MergeEither)) { |
| 2956 | MBBI = *NextI; |
| 2957 | return true; |
| 2958 | } |
| 2959 | } |
| 2960 | |
| 2961 | // The immediate in the load/store is scaled by the size of the memory |
| 2962 | // operation. The immediate in the add we're looking for, |
| 2963 | // however, is not, so adjust here. |
| 2964 | int UnscaledOffset = |
| 2965 | AArch64InstrInfo::getLdStOffsetOp(MI).getImm() * TII->getMemScale(MI); |
| 2966 | |
| 2967 | // Look forward to try to find a pre-index instruction. For example, |
| 2968 | // ldr x1, [x0, #64] |
| 2969 | // add x0, x0, #64 |
| 2970 | // merged into: |
| 2971 | // ldr x1, [x0, #64]! |
| 2972 | Update = findMatchingUpdateInsnForward(I: MBBI, UnscaledOffset, Limit: UpdateLimit); |
| 2973 | if (Update != E) { |
| 2974 | // Merge the update into the ld/st. |
| 2975 | if (auto NextI = mergeUpdateInsn(I: MBBI, Update, /*IsForward=*/false, |
| 2976 | /*IsPreIdx=*/true, |
| 2977 | /*MergeEither=*/false)) { |
| 2978 | MBBI = *NextI; |
| 2979 | return true; |
| 2980 | } |
| 2981 | } |
| 2982 | |
| 2983 | return false; |
| 2984 | } |
| 2985 | |
| 2986 | bool AArch64LoadStoreOpt::tryToMergeIndexLdSt(MachineBasicBlock::iterator &MBBI, |
| 2987 | int Scale) { |
| 2988 | MachineInstr &MI = *MBBI; |
| 2989 | MachineBasicBlock::iterator E = MI.getParent()->end(); |
| 2990 | MachineBasicBlock::iterator Update; |
| 2991 | |
| 2992 | // Don't know how to handle unscaled pre/post-index versions below, so bail. |
| 2993 | if (TII->hasUnscaledLdStOffset(Opc: MI.getOpcode())) |
| 2994 | return false; |
| 2995 | |
| 2996 | // Look back to try to find a const offset for index LdSt instruction. For |
| 2997 | // example, |
| 2998 | // mov x8, #LargeImm ; = a * (1<<12) + imm12 |
| 2999 | // ldr x1, [x0, x8] |
| 3000 | // merged into: |
| 3001 | // add x8, x0, a * (1<<12) |
| 3002 | // ldr x1, [x8, imm12] |
| 3003 | unsigned Offset; |
| 3004 | Update = findMatchingConstOffsetBackward(I: MBBI, Limit: LdStConstLimit, Offset); |
| 3005 | if (Update != E && (Offset & (Scale - 1)) == 0) { |
| 3006 | // Merge the imm12 into the ld/st. |
| 3007 | MBBI = mergeConstOffsetInsn(I: MBBI, Update, Offset, Scale); |
| 3008 | return true; |
| 3009 | } |
| 3010 | |
| 3011 | return false; |
| 3012 | } |
| 3013 | |
| 3014 | bool AArch64LoadStoreOpt::optimizeBlock(MachineBasicBlock &MBB, |
| 3015 | bool EnableNarrowZeroStOpt) { |
| 3016 | AArch64FunctionInfo &AFI = *MBB.getParent()->getInfo<AArch64FunctionInfo>(); |
| 3017 | |
| 3018 | bool Modified = false; |
| 3019 | // Four transformations to do here: |
| 3020 | // 1) Find loads that directly read from stores and promote them by |
| 3021 | // replacing with mov instructions. If the store is wider than the load, |
| 3022 | // the load will be replaced with a bitfield extract. |
| 3023 | // e.g., |
| 3024 | // str w1, [x0, #4] |
| 3025 | // ldrh w2, [x0, #6] |
| 3026 | // ; becomes |
| 3027 | // str w1, [x0, #4] |
| 3028 | // lsr w2, w1, #16 |
| 3029 | for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end(); |
| 3030 | MBBI != E;) { |
| 3031 | if (isPromotableLoadFromStore(MI&: *MBBI) && tryToPromoteLoadFromStore(MBBI)) |
| 3032 | Modified = true; |
| 3033 | else |
| 3034 | ++MBBI; |
| 3035 | } |
| 3036 | // 2) Merge adjacent zero stores into a wider store. |
| 3037 | // e.g., |
| 3038 | // strh wzr, [x0] |
| 3039 | // strh wzr, [x0, #2] |
| 3040 | // ; becomes |
| 3041 | // str wzr, [x0] |
| 3042 | // e.g., |
| 3043 | // str wzr, [x0] |
| 3044 | // str wzr, [x0, #4] |
| 3045 | // ; becomes |
| 3046 | // str xzr, [x0] |
| 3047 | if (EnableNarrowZeroStOpt) |
| 3048 | for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end(); |
| 3049 | MBBI != E;) { |
| 3050 | if (isPromotableZeroStoreInst(MI&: *MBBI) && tryToMergeZeroStInst(MBBI)) |
| 3051 | Modified = true; |
| 3052 | else |
| 3053 | ++MBBI; |
| 3054 | } |
| 3055 | // 3) Find loads and stores that can be merged into a single load or store |
| 3056 | // pair instruction. |
| 3057 | // When compiling for SVE 128, also try to combine SVE fill/spill |
| 3058 | // instructions into LDP/STP. |
| 3059 | // e.g., |
| 3060 | // ldr x0, [x2] |
| 3061 | // ldr x1, [x2, #8] |
| 3062 | // ; becomes |
| 3063 | // ldp x0, x1, [x2] |
| 3064 | // e.g., |
| 3065 | // ldr z0, [x2] |
| 3066 | // ldr z1, [x2, #1, mul vl] |
| 3067 | // ; becomes |
| 3068 | // ldp q0, q1, [x2] |
| 3069 | |
| 3070 | if (MBB.getParent()->getRegInfo().tracksLiveness()) { |
| 3071 | DefinedInBB.clear(); |
| 3072 | DefinedInBB.addLiveIns(MBB); |
| 3073 | } |
| 3074 | |
| 3075 | for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end(); |
| 3076 | MBBI != E;) { |
| 3077 | // Track currently live registers up to this point, to help with |
| 3078 | // searching for a rename register on demand. |
| 3079 | updateDefinedRegisters(MI&: *MBBI, Units&: DefinedInBB, TRI); |
| 3080 | if (TII->isPairableLdStInst(MI: *MBBI) && tryToPairLdStInst(MBBI)) |
| 3081 | Modified = true; |
| 3082 | else |
| 3083 | ++MBBI; |
| 3084 | } |
| 3085 | // 4) Find base register updates that can be merged into the load or store |
| 3086 | // as a base-reg writeback. |
| 3087 | // e.g., |
| 3088 | // ldr x0, [x2] |
| 3089 | // add x2, x2, #4 |
| 3090 | // ; becomes |
| 3091 | // ldr x0, [x2], #4 |
| 3092 | for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end(); |
| 3093 | MBBI != E;) { |
| 3094 | if (isMergeableLdStUpdate(MI&: *MBBI, AFI) && tryToMergeLdStUpdate(MBBI)) |
| 3095 | Modified = true; |
| 3096 | else |
| 3097 | ++MBBI; |
| 3098 | } |
| 3099 | |
| 3100 | // 5) Find a register assigned with a const value that can be combined with |
| 3101 | // into the load or store. e.g., |
| 3102 | // mov x8, #LargeImm ; = a * (1<<12) + imm12 |
| 3103 | // ldr x1, [x0, x8] |
| 3104 | // ; becomes |
| 3105 | // add x8, x0, a * (1<<12) |
| 3106 | // ldr x1, [x8, imm12] |
| 3107 | for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end(); |
| 3108 | MBBI != E;) { |
| 3109 | int Scale; |
| 3110 | if (isMergeableIndexLdSt(MI&: *MBBI, Scale) && tryToMergeIndexLdSt(MBBI, Scale)) |
| 3111 | Modified = true; |
| 3112 | else |
| 3113 | ++MBBI; |
| 3114 | } |
| 3115 | |
| 3116 | return Modified; |
| 3117 | } |
| 3118 | |
| 3119 | bool AArch64LoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) { |
| 3120 | Subtarget = &Fn.getSubtarget<AArch64Subtarget>(); |
| 3121 | TII = Subtarget->getInstrInfo(); |
| 3122 | TRI = Subtarget->getRegisterInfo(); |
| 3123 | |
| 3124 | // Resize the modified and used register unit trackers. We do this once |
| 3125 | // per function and then clear the register units each time we optimize a load |
| 3126 | // or store. |
| 3127 | ModifiedRegUnits.init(TRI: *TRI); |
| 3128 | UsedRegUnits.init(TRI: *TRI); |
| 3129 | DefinedInBB.init(TRI: *TRI); |
| 3130 | |
| 3131 | bool Modified = false; |
| 3132 | bool enableNarrowZeroStOpt = !Subtarget->requiresStrictAlign(); |
| 3133 | for (auto &MBB : Fn) { |
| 3134 | auto M = optimizeBlock(MBB, EnableNarrowZeroStOpt: enableNarrowZeroStOpt); |
| 3135 | Modified |= M; |
| 3136 | } |
| 3137 | |
| 3138 | return Modified; |
| 3139 | } |
| 3140 | |
| 3141 | // FIXME: Do we need/want a pre-alloc pass like ARM has to try to keep loads and |
| 3142 | // stores near one another? Note: The pre-RA instruction scheduler already has |
| 3143 | // hooks to try and schedule pairable loads/stores together to improve pairing |
| 3144 | // opportunities. Thus, pre-RA pairing pass may not be worth the effort. |
| 3145 | |
| 3146 | // FIXME: When pairing store instructions it's very possible for this pass to |
| 3147 | // hoist a store with a KILL marker above another use (without a KILL marker). |
| 3148 | // The resulting IR is invalid, but nothing uses the KILL markers after this |
| 3149 | // pass, so it's never caused a problem in practice. |
| 3150 | |
| 3151 | bool AArch64LoadStoreOptLegacy::runOnMachineFunction(MachineFunction &MF) { |
| 3152 | if (skipFunction(F: MF.getFunction())) |
| 3153 | return false; |
| 3154 | AArch64LoadStoreOpt Impl; |
| 3155 | Impl.AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); |
| 3156 | return Impl.runOnMachineFunction(Fn&: MF); |
| 3157 | } |
| 3158 | |
| 3159 | /// createAArch64LoadStoreOptimizationPass - returns an instance of the |
| 3160 | /// load / store optimization pass. |
| 3161 | FunctionPass *llvm::createAArch64LoadStoreOptLegacyPass() { |
| 3162 | return new AArch64LoadStoreOptLegacy(); |
| 3163 | } |
| 3164 | |
| 3165 | PreservedAnalyses |
| 3166 | AArch64LoadStoreOptPass::run(MachineFunction &MF, |
| 3167 | MachineFunctionAnalysisManager &MFAM) { |
| 3168 | AArch64LoadStoreOpt Impl; |
| 3169 | Impl.AA = &MFAM.getResult<FunctionAnalysisManagerMachineFunctionProxy>(IR&: MF) |
| 3170 | .getManager() |
| 3171 | .getResult<AAManager>(IR&: MF.getFunction()); |
| 3172 | bool Changed = Impl.runOnMachineFunction(Fn&: MF); |
| 3173 | if (!Changed) |
| 3174 | return PreservedAnalyses::all(); |
| 3175 | PreservedAnalyses PA = getMachineFunctionPassPreservedAnalyses(); |
| 3176 | PA.preserveSet<CFGAnalyses>(); |
| 3177 | return PA; |
| 3178 | } |
| 3179 | |