1 //===- LocalStackSlotAllocation.cpp - Pre-allocate locals to stack slots --===// 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 pass assigns local frame indices to stack slots relative to one another 10 // and allocates additional base registers to access them when the target 11 // estimates they are likely to be out of range of stack pointer and frame 12 // pointer relative addressing. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #include "llvm/ADT/SetVector.h" 17 #include "llvm/ADT/SmallSet.h" 18 #include "llvm/ADT/SmallVector.h" 19 #include "llvm/ADT/Statistic.h" 20 #include "llvm/CodeGen/MachineBasicBlock.h" 21 #include "llvm/CodeGen/MachineFrameInfo.h" 22 #include "llvm/CodeGen/MachineFunction.h" 23 #include "llvm/CodeGen/MachineFunctionPass.h" 24 #include "llvm/CodeGen/MachineInstr.h" 25 #include "llvm/CodeGen/MachineOperand.h" 26 #include "llvm/CodeGen/MachineRegisterInfo.h" 27 #include "llvm/CodeGen/TargetFrameLowering.h" 28 #include "llvm/CodeGen/TargetOpcodes.h" 29 #include "llvm/CodeGen/TargetRegisterInfo.h" 30 #include "llvm/CodeGen/TargetSubtargetInfo.h" 31 #include "llvm/InitializePasses.h" 32 #include "llvm/Pass.h" 33 #include "llvm/Support/Debug.h" 34 #include "llvm/Support/ErrorHandling.h" 35 #include "llvm/Support/raw_ostream.h" 36 #include <algorithm> 37 #include <cassert> 38 #include <cstdint> 39 #include <tuple> 40 41 using namespace llvm; 42 43 #define DEBUG_TYPE "localstackalloc" 44 45 STATISTIC(NumAllocations, "Number of frame indices allocated into local block"); 46 STATISTIC(NumBaseRegisters, "Number of virtual frame base registers allocated"); 47 STATISTIC(NumReplacements, "Number of frame indices references replaced"); 48 49 namespace { 50 51 class FrameRef { 52 MachineBasicBlock::iterator MI; // Instr referencing the frame 53 int64_t LocalOffset; // Local offset of the frame idx referenced 54 int FrameIdx; // The frame index 55 56 // Order reference instruction appears in program. Used to ensure 57 // deterministic order when multiple instructions may reference the same 58 // location. 59 unsigned Order; 60 61 public: 62 FrameRef(MachineInstr *I, int64_t Offset, int Idx, unsigned Ord) : 63 MI(I), LocalOffset(Offset), FrameIdx(Idx), Order(Ord) {} 64 65 bool operator<(const FrameRef &RHS) const { 66 return std::tie(LocalOffset, FrameIdx, Order) < 67 std::tie(RHS.LocalOffset, RHS.FrameIdx, RHS.Order); 68 } 69 70 MachineBasicBlock::iterator getMachineInstr() const { return MI; } 71 int64_t getLocalOffset() const { return LocalOffset; } 72 int getFrameIndex() const { return FrameIdx; } 73 }; 74 75 class LocalStackSlotPass: public MachineFunctionPass { 76 SmallVector<int64_t, 16> LocalOffsets; 77 78 /// StackObjSet - A set of stack object indexes 79 using StackObjSet = SmallSetVector<int, 8>; 80 81 void AdjustStackOffset(MachineFrameInfo &MFI, int FrameIdx, int64_t &Offset, 82 bool StackGrowsDown, Align &MaxAlign); 83 void AssignProtectedObjSet(const StackObjSet &UnassignedObjs, 84 SmallSet<int, 16> &ProtectedObjs, 85 MachineFrameInfo &MFI, bool StackGrowsDown, 86 int64_t &Offset, Align &MaxAlign); 87 void calculateFrameObjectOffsets(MachineFunction &Fn); 88 bool insertFrameReferenceRegisters(MachineFunction &Fn); 89 90 public: 91 static char ID; // Pass identification, replacement for typeid 92 93 explicit LocalStackSlotPass() : MachineFunctionPass(ID) { 94 initializeLocalStackSlotPassPass(*PassRegistry::getPassRegistry()); 95 } 96 97 bool runOnMachineFunction(MachineFunction &MF) override; 98 99 void getAnalysisUsage(AnalysisUsage &AU) const override { 100 AU.setPreservesCFG(); 101 MachineFunctionPass::getAnalysisUsage(AU); 102 } 103 }; 104 105 } // end anonymous namespace 106 107 char LocalStackSlotPass::ID = 0; 108 109 char &llvm::LocalStackSlotAllocationID = LocalStackSlotPass::ID; 110 INITIALIZE_PASS(LocalStackSlotPass, DEBUG_TYPE, 111 "Local Stack Slot Allocation", false, false) 112 113 bool LocalStackSlotPass::runOnMachineFunction(MachineFunction &MF) { 114 MachineFrameInfo &MFI = MF.getFrameInfo(); 115 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 116 unsigned LocalObjectCount = MFI.getObjectIndexEnd(); 117 118 // If the target doesn't want/need this pass, or if there are no locals 119 // to consider, early exit. 120 if (LocalObjectCount == 0 || !TRI->requiresVirtualBaseRegisters(MF)) 121 return true; 122 123 // Make sure we have enough space to store the local offsets. 124 LocalOffsets.resize(MFI.getObjectIndexEnd()); 125 126 // Lay out the local blob. 127 calculateFrameObjectOffsets(MF); 128 129 // Insert virtual base registers to resolve frame index references. 130 bool UsedBaseRegs = insertFrameReferenceRegisters(MF); 131 132 // Tell MFI whether any base registers were allocated. PEI will only 133 // want to use the local block allocations from this pass if there were any. 134 // Otherwise, PEI can do a bit better job of getting the alignment right 135 // without a hole at the start since it knows the alignment of the stack 136 // at the start of local allocation, and this pass doesn't. 137 MFI.setUseLocalStackAllocationBlock(UsedBaseRegs); 138 139 return true; 140 } 141 142 /// AdjustStackOffset - Helper function used to adjust the stack frame offset. 143 void LocalStackSlotPass::AdjustStackOffset(MachineFrameInfo &MFI, int FrameIdx, 144 int64_t &Offset, bool StackGrowsDown, 145 Align &MaxAlign) { 146 // If the stack grows down, add the object size to find the lowest address. 147 if (StackGrowsDown) 148 Offset += MFI.getObjectSize(FrameIdx); 149 150 Align Alignment = MFI.getObjectAlign(FrameIdx); 151 152 // If the alignment of this object is greater than that of the stack, then 153 // increase the stack alignment to match. 154 MaxAlign = std::max(MaxAlign, Alignment); 155 156 // Adjust to alignment boundary. 157 Offset = alignTo(Offset, Alignment); 158 159 int64_t LocalOffset = StackGrowsDown ? -Offset : Offset; 160 LLVM_DEBUG(dbgs() << "Allocate FI(" << FrameIdx << ") to local offset " 161 << LocalOffset << "\n"); 162 // Keep the offset available for base register allocation 163 LocalOffsets[FrameIdx] = LocalOffset; 164 // And tell MFI about it for PEI to use later 165 MFI.mapLocalFrameObject(FrameIdx, LocalOffset); 166 167 if (!StackGrowsDown) 168 Offset += MFI.getObjectSize(FrameIdx); 169 170 ++NumAllocations; 171 } 172 173 /// AssignProtectedObjSet - Helper function to assign large stack objects (i.e., 174 /// those required to be close to the Stack Protector) to stack offsets. 175 void LocalStackSlotPass::AssignProtectedObjSet( 176 const StackObjSet &UnassignedObjs, SmallSet<int, 16> &ProtectedObjs, 177 MachineFrameInfo &MFI, bool StackGrowsDown, int64_t &Offset, 178 Align &MaxAlign) { 179 for (int i : UnassignedObjs) { 180 AdjustStackOffset(MFI, i, Offset, StackGrowsDown, MaxAlign); 181 ProtectedObjs.insert(i); 182 } 183 } 184 185 /// calculateFrameObjectOffsets - Calculate actual frame offsets for all of the 186 /// abstract stack objects. 187 void LocalStackSlotPass::calculateFrameObjectOffsets(MachineFunction &Fn) { 188 // Loop over all of the stack objects, assigning sequential addresses... 189 MachineFrameInfo &MFI = Fn.getFrameInfo(); 190 const TargetFrameLowering &TFI = *Fn.getSubtarget().getFrameLowering(); 191 bool StackGrowsDown = 192 TFI.getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown; 193 int64_t Offset = 0; 194 Align MaxAlign; 195 196 // Make sure that the stack protector comes before the local variables on the 197 // stack. 198 SmallSet<int, 16> ProtectedObjs; 199 if (MFI.hasStackProtectorIndex()) { 200 int StackProtectorFI = MFI.getStackProtectorIndex(); 201 202 // We need to make sure we didn't pre-allocate the stack protector when 203 // doing this. 204 // If we already have a stack protector, this will re-assign it to a slot 205 // that is **not** covering the protected objects. 206 assert(!MFI.isObjectPreAllocated(StackProtectorFI) && 207 "Stack protector pre-allocated in LocalStackSlotAllocation"); 208 209 StackObjSet LargeArrayObjs; 210 StackObjSet SmallArrayObjs; 211 StackObjSet AddrOfObjs; 212 213 // Only place the stack protector in the local stack area if the target 214 // allows it. 215 if (TFI.isStackIdSafeForLocalArea(MFI.getStackID(StackProtectorFI))) 216 AdjustStackOffset(MFI, StackProtectorFI, Offset, StackGrowsDown, 217 MaxAlign); 218 219 // Assign large stack objects first. 220 for (unsigned i = 0, e = MFI.getObjectIndexEnd(); i != e; ++i) { 221 if (MFI.isDeadObjectIndex(i)) 222 continue; 223 if (StackProtectorFI == (int)i) 224 continue; 225 if (!TFI.isStackIdSafeForLocalArea(MFI.getStackID(i))) 226 continue; 227 228 switch (MFI.getObjectSSPLayout(i)) { 229 case MachineFrameInfo::SSPLK_None: 230 continue; 231 case MachineFrameInfo::SSPLK_SmallArray: 232 SmallArrayObjs.insert(i); 233 continue; 234 case MachineFrameInfo::SSPLK_AddrOf: 235 AddrOfObjs.insert(i); 236 continue; 237 case MachineFrameInfo::SSPLK_LargeArray: 238 LargeArrayObjs.insert(i); 239 continue; 240 } 241 llvm_unreachable("Unexpected SSPLayoutKind."); 242 } 243 244 AssignProtectedObjSet(LargeArrayObjs, ProtectedObjs, MFI, StackGrowsDown, 245 Offset, MaxAlign); 246 AssignProtectedObjSet(SmallArrayObjs, ProtectedObjs, MFI, StackGrowsDown, 247 Offset, MaxAlign); 248 AssignProtectedObjSet(AddrOfObjs, ProtectedObjs, MFI, StackGrowsDown, 249 Offset, MaxAlign); 250 } 251 252 // Then assign frame offsets to stack objects that are not used to spill 253 // callee saved registers. 254 for (unsigned i = 0, e = MFI.getObjectIndexEnd(); i != e; ++i) { 255 if (MFI.isDeadObjectIndex(i)) 256 continue; 257 if (MFI.getStackProtectorIndex() == (int)i) 258 continue; 259 if (ProtectedObjs.count(i)) 260 continue; 261 if (!TFI.isStackIdSafeForLocalArea(MFI.getStackID(i))) 262 continue; 263 264 AdjustStackOffset(MFI, i, Offset, StackGrowsDown, MaxAlign); 265 } 266 267 // Remember how big this blob of stack space is 268 MFI.setLocalFrameSize(Offset); 269 MFI.setLocalFrameMaxAlign(MaxAlign); 270 } 271 272 static inline bool 273 lookupCandidateBaseReg(unsigned BaseReg, 274 int64_t BaseOffset, 275 int64_t FrameSizeAdjust, 276 int64_t LocalFrameOffset, 277 const MachineInstr &MI, 278 const TargetRegisterInfo *TRI) { 279 // Check if the relative offset from the where the base register references 280 // to the target address is in range for the instruction. 281 int64_t Offset = FrameSizeAdjust + LocalFrameOffset - BaseOffset; 282 return TRI->isFrameOffsetLegal(&MI, BaseReg, Offset); 283 } 284 285 bool LocalStackSlotPass::insertFrameReferenceRegisters(MachineFunction &Fn) { 286 // Scan the function's instructions looking for frame index references. 287 // For each, ask the target if it wants a virtual base register for it 288 // based on what we can tell it about where the local will end up in the 289 // stack frame. If it wants one, re-use a suitable one we've previously 290 // allocated, or if there isn't one that fits the bill, allocate a new one 291 // and ask the target to create a defining instruction for it. 292 bool UsedBaseReg = false; 293 294 MachineFrameInfo &MFI = Fn.getFrameInfo(); 295 const TargetRegisterInfo *TRI = Fn.getSubtarget().getRegisterInfo(); 296 const TargetFrameLowering &TFI = *Fn.getSubtarget().getFrameLowering(); 297 bool StackGrowsDown = 298 TFI.getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown; 299 300 // Collect all of the instructions in the block that reference 301 // a frame index. Also store the frame index referenced to ease later 302 // lookup. (For any insn that has more than one FI reference, we arbitrarily 303 // choose the first one). 304 SmallVector<FrameRef, 64> FrameReferenceInsns; 305 306 unsigned Order = 0; 307 308 for (MachineBasicBlock &BB : Fn) { 309 for (MachineInstr &MI : BB) { 310 // Debug value, stackmap and patchpoint instructions can't be out of 311 // range, so they don't need any updates. 312 if (MI.isDebugInstr() || MI.getOpcode() == TargetOpcode::STATEPOINT || 313 MI.getOpcode() == TargetOpcode::STACKMAP || 314 MI.getOpcode() == TargetOpcode::PATCHPOINT) 315 continue; 316 317 // For now, allocate the base register(s) within the basic block 318 // where they're used, and don't try to keep them around outside 319 // of that. It may be beneficial to try sharing them more broadly 320 // than that, but the increased register pressure makes that a 321 // tricky thing to balance. Investigate if re-materializing these 322 // becomes an issue. 323 for (const MachineOperand &MO : MI.operands()) { 324 // Consider replacing all frame index operands that reference 325 // an object allocated in the local block. 326 if (MO.isFI()) { 327 // Don't try this with values not in the local block. 328 if (!MFI.isObjectPreAllocated(MO.getIndex())) 329 break; 330 int Idx = MO.getIndex(); 331 int64_t LocalOffset = LocalOffsets[Idx]; 332 if (!TRI->needsFrameBaseReg(&MI, LocalOffset)) 333 break; 334 FrameReferenceInsns.push_back(FrameRef(&MI, LocalOffset, Idx, Order++)); 335 break; 336 } 337 } 338 } 339 } 340 341 // Sort the frame references by local offset. 342 // Use frame index as a tie-breaker in case MI's have the same offset. 343 llvm::sort(FrameReferenceInsns); 344 345 MachineBasicBlock *Entry = &Fn.front(); 346 347 unsigned BaseReg = 0; 348 int64_t BaseOffset = 0; 349 350 // Loop through the frame references and allocate for them as necessary. 351 for (int ref = 0, e = FrameReferenceInsns.size(); ref < e ; ++ref) { 352 FrameRef &FR = FrameReferenceInsns[ref]; 353 MachineInstr &MI = *FR.getMachineInstr(); 354 int64_t LocalOffset = FR.getLocalOffset(); 355 int FrameIdx = FR.getFrameIndex(); 356 assert(MFI.isObjectPreAllocated(FrameIdx) && 357 "Only pre-allocated locals expected!"); 358 359 // We need to keep the references to the stack protector slot through frame 360 // index operands so that it gets resolved by PEI rather than this pass. 361 // This avoids accesses to the stack protector though virtual base 362 // registers, and forces PEI to address it using fp/sp/bp. 363 if (MFI.hasStackProtectorIndex() && 364 FrameIdx == MFI.getStackProtectorIndex()) 365 continue; 366 367 LLVM_DEBUG(dbgs() << "Considering: " << MI); 368 369 unsigned idx = 0; 370 for (unsigned f = MI.getNumOperands(); idx != f; ++idx) { 371 if (!MI.getOperand(idx).isFI()) 372 continue; 373 374 if (FrameIdx == MI.getOperand(idx).getIndex()) 375 break; 376 } 377 378 assert(idx < MI.getNumOperands() && "Cannot find FI operand"); 379 380 int64_t Offset = 0; 381 int64_t FrameSizeAdjust = StackGrowsDown ? MFI.getLocalFrameSize() : 0; 382 383 LLVM_DEBUG(dbgs() << " Replacing FI in: " << MI); 384 385 // If we have a suitable base register available, use it; otherwise 386 // create a new one. Note that any offset encoded in the 387 // instruction itself will be taken into account by the target, 388 // so we don't have to adjust for it here when reusing a base 389 // register. 390 if (UsedBaseReg && 391 lookupCandidateBaseReg(BaseReg, BaseOffset, FrameSizeAdjust, 392 LocalOffset, MI, TRI)) { 393 LLVM_DEBUG(dbgs() << " Reusing base register " << BaseReg << "\n"); 394 // We found a register to reuse. 395 Offset = FrameSizeAdjust + LocalOffset - BaseOffset; 396 } else { 397 // No previously defined register was in range, so create a new one. 398 int64_t InstrOffset = TRI->getFrameIndexInstrOffset(&MI, idx); 399 400 int64_t PrevBaseOffset = BaseOffset; 401 BaseOffset = FrameSizeAdjust + LocalOffset + InstrOffset; 402 403 // We'd like to avoid creating single-use virtual base registers. 404 // Because the FrameRefs are in sorted order, and we've already 405 // processed all FrameRefs before this one, just check whether or not 406 // the next FrameRef will be able to reuse this new register. If not, 407 // then don't bother creating it. 408 if (ref + 1 >= e || 409 !lookupCandidateBaseReg( 410 BaseReg, BaseOffset, FrameSizeAdjust, 411 FrameReferenceInsns[ref + 1].getLocalOffset(), 412 *FrameReferenceInsns[ref + 1].getMachineInstr(), TRI)) { 413 BaseOffset = PrevBaseOffset; 414 continue; 415 } 416 417 const MachineFunction *MF = MI.getMF(); 418 const TargetRegisterClass *RC = TRI->getPointerRegClass(*MF); 419 BaseReg = Fn.getRegInfo().createVirtualRegister(RC); 420 421 LLVM_DEBUG(dbgs() << " Materializing base register" 422 << " at frame local offset " 423 << LocalOffset + InstrOffset); 424 425 // Tell the target to insert the instruction to initialize 426 // the base register. 427 // MachineBasicBlock::iterator InsertionPt = Entry->begin(); 428 BaseReg = TRI->materializeFrameBaseRegister(Entry, FrameIdx, InstrOffset); 429 430 LLVM_DEBUG(dbgs() << " into " << printReg(BaseReg, TRI) << '\n'); 431 432 // The base register already includes any offset specified 433 // by the instruction, so account for that so it doesn't get 434 // applied twice. 435 Offset = -InstrOffset; 436 437 ++NumBaseRegisters; 438 UsedBaseReg = true; 439 } 440 assert(BaseReg != 0 && "Unable to allocate virtual base register!"); 441 442 // Modify the instruction to use the new base register rather 443 // than the frame index operand. 444 TRI->resolveFrameIndex(MI, BaseReg, Offset); 445 LLVM_DEBUG(dbgs() << "Resolved: " << MI); 446 447 ++NumReplacements; 448 } 449 450 return UsedBaseReg; 451 } 452