1 //===- LocalStackSlotAllocation.cpp - Pre-allocate locals to stack slots --===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass assigns local frame indices to stack slots relative to one another
11 // and allocates additional base registers to access them when the target
12 // estimates they are likely to be out of range of stack pointer and frame
13 // pointer relative addressing.
15 //===----------------------------------------------------------------------===//
17 #include "llvm/CodeGen/Passes.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/ADT/SmallSet.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineFunctionPass.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/StackProtector.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/IR/LLVMContext.h"
32 #include "llvm/IR/Module.h"
33 #include "llvm/Pass.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/Target/TargetFrameLowering.h"
38 #include "llvm/Target/TargetRegisterInfo.h"
39 #include "llvm/Target/TargetSubtargetInfo.h"
43 #define DEBUG_TYPE "localstackalloc"
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");
51 MachineBasicBlock::iterator MI; // Instr referencing the frame
52 int64_t LocalOffset; // Local offset of the frame idx referenced
53 int FrameIdx; // The frame index
55 // Order reference instruction appears in program. Used to ensure
56 // deterministic order when multiple instructions may reference the same
61 FrameRef(MachineInstr *I, int64_t Offset, int Idx, unsigned Ord) :
62 MI(I), LocalOffset(Offset), FrameIdx(Idx), Order(Ord) {}
64 bool operator<(const FrameRef &RHS) const {
65 return std::tie(LocalOffset, FrameIdx, Order) <
66 std::tie(RHS.LocalOffset, RHS.FrameIdx, RHS.Order);
69 MachineBasicBlock::iterator getMachineInstr() const { return MI; }
70 int64_t getLocalOffset() const { return LocalOffset; }
71 int getFrameIndex() const { return FrameIdx; }
74 class LocalStackSlotPass: public MachineFunctionPass {
75 SmallVector<int64_t,16> LocalOffsets;
76 /// StackObjSet - A set of stack object indexes
77 typedef SmallSetVector<int, 8> StackObjSet;
79 void AdjustStackOffset(MachineFrameInfo &MFI, int FrameIdx, int64_t &Offset,
80 bool StackGrowsDown, unsigned &MaxAlign);
81 void AssignProtectedObjSet(const StackObjSet &UnassignedObjs,
82 SmallSet<int, 16> &ProtectedObjs,
83 MachineFrameInfo &MFI, bool StackGrowsDown,
84 int64_t &Offset, unsigned &MaxAlign);
85 void calculateFrameObjectOffsets(MachineFunction &Fn);
86 bool insertFrameReferenceRegisters(MachineFunction &Fn);
88 static char ID; // Pass identification, replacement for typeid
89 explicit LocalStackSlotPass() : MachineFunctionPass(ID) {
90 initializeLocalStackSlotPassPass(*PassRegistry::getPassRegistry());
92 bool runOnMachineFunction(MachineFunction &MF) override;
94 void getAnalysisUsage(AnalysisUsage &AU) const override {
96 AU.addRequired<StackProtector>();
97 MachineFunctionPass::getAnalysisUsage(AU);
102 } // end anonymous namespace
104 char LocalStackSlotPass::ID = 0;
105 char &llvm::LocalStackSlotAllocationID = LocalStackSlotPass::ID;
106 INITIALIZE_PASS_BEGIN(LocalStackSlotPass, "localstackalloc",
107 "Local Stack Slot Allocation", false, false)
108 INITIALIZE_PASS_DEPENDENCY(StackProtector)
109 INITIALIZE_PASS_END(LocalStackSlotPass, "localstackalloc",
110 "Local Stack Slot Allocation", false, false)
113 bool LocalStackSlotPass::runOnMachineFunction(MachineFunction &MF) {
114 MachineFrameInfo &MFI = MF.getFrameInfo();
115 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
116 unsigned LocalObjectCount = MFI.getObjectIndexEnd();
118 // If the target doesn't want/need this pass, or if there are no locals
119 // to consider, early exit.
120 if (!TRI->requiresVirtualBaseRegisters(MF) || LocalObjectCount == 0)
123 // Make sure we have enough space to store the local offsets.
124 LocalOffsets.resize(MFI.getObjectIndexEnd());
126 // Lay out the local blob.
127 calculateFrameObjectOffsets(MF);
129 // Insert virtual base registers to resolve frame index references.
130 bool UsedBaseRegs = insertFrameReferenceRegisters(MF);
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);
142 /// AdjustStackOffset - Helper function used to adjust the stack frame offset.
143 void LocalStackSlotPass::AdjustStackOffset(MachineFrameInfo &MFI,
144 int FrameIdx, int64_t &Offset,
146 unsigned &MaxAlign) {
147 // If the stack grows down, add the object size to find the lowest address.
149 Offset += MFI.getObjectSize(FrameIdx);
151 unsigned Align = MFI.getObjectAlignment(FrameIdx);
153 // If the alignment of this object is greater than that of the stack, then
154 // increase the stack alignment to match.
155 MaxAlign = std::max(MaxAlign, Align);
157 // Adjust to alignment boundary.
158 Offset = (Offset + Align - 1) / Align * Align;
160 int64_t LocalOffset = StackGrowsDown ? -Offset : Offset;
161 DEBUG(dbgs() << "Allocate FI(" << FrameIdx << ") to local offset "
162 << LocalOffset << "\n");
163 // Keep the offset available for base register allocation
164 LocalOffsets[FrameIdx] = LocalOffset;
165 // And tell MFI about it for PEI to use later
166 MFI.mapLocalFrameObject(FrameIdx, LocalOffset);
169 Offset += MFI.getObjectSize(FrameIdx);
174 /// AssignProtectedObjSet - Helper function to assign large stack objects (i.e.,
175 /// those required to be close to the Stack Protector) to stack offsets.
176 void LocalStackSlotPass::AssignProtectedObjSet(const StackObjSet &UnassignedObjs,
177 SmallSet<int, 16> &ProtectedObjs,
178 MachineFrameInfo &MFI,
179 bool StackGrowsDown, int64_t &Offset,
180 unsigned &MaxAlign) {
182 for (StackObjSet::const_iterator I = UnassignedObjs.begin(),
183 E = UnassignedObjs.end(); I != E; ++I) {
185 AdjustStackOffset(MFI, i, Offset, StackGrowsDown, MaxAlign);
186 ProtectedObjs.insert(i);
190 /// calculateFrameObjectOffsets - Calculate actual frame offsets for all of the
191 /// abstract stack objects.
193 void LocalStackSlotPass::calculateFrameObjectOffsets(MachineFunction &Fn) {
194 // Loop over all of the stack objects, assigning sequential addresses...
195 MachineFrameInfo &MFI = Fn.getFrameInfo();
196 const TargetFrameLowering &TFI = *Fn.getSubtarget().getFrameLowering();
197 bool StackGrowsDown =
198 TFI.getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
200 unsigned MaxAlign = 0;
201 StackProtector *SP = &getAnalysis<StackProtector>();
203 // Make sure that the stack protector comes before the local variables on the
205 SmallSet<int, 16> ProtectedObjs;
206 if (MFI.getStackProtectorIndex() >= 0) {
207 StackObjSet LargeArrayObjs;
208 StackObjSet SmallArrayObjs;
209 StackObjSet AddrOfObjs;
211 AdjustStackOffset(MFI, MFI.getStackProtectorIndex(), Offset,
212 StackGrowsDown, MaxAlign);
214 // Assign large stack objects first.
215 for (unsigned i = 0, e = MFI.getObjectIndexEnd(); i != e; ++i) {
216 if (MFI.isDeadObjectIndex(i))
218 if (MFI.getStackProtectorIndex() == (int)i)
221 switch (SP->getSSPLayout(MFI.getObjectAllocation(i))) {
222 case StackProtector::SSPLK_None:
224 case StackProtector::SSPLK_SmallArray:
225 SmallArrayObjs.insert(i);
227 case StackProtector::SSPLK_AddrOf:
228 AddrOfObjs.insert(i);
230 case StackProtector::SSPLK_LargeArray:
231 LargeArrayObjs.insert(i);
234 llvm_unreachable("Unexpected SSPLayoutKind.");
237 AssignProtectedObjSet(LargeArrayObjs, ProtectedObjs, MFI, StackGrowsDown,
239 AssignProtectedObjSet(SmallArrayObjs, ProtectedObjs, MFI, StackGrowsDown,
241 AssignProtectedObjSet(AddrOfObjs, ProtectedObjs, MFI, StackGrowsDown,
245 // Then assign frame offsets to stack objects that are not used to spill
246 // callee saved registers.
247 for (unsigned i = 0, e = MFI.getObjectIndexEnd(); i != e; ++i) {
248 if (MFI.isDeadObjectIndex(i))
250 if (MFI.getStackProtectorIndex() == (int)i)
252 if (ProtectedObjs.count(i))
255 AdjustStackOffset(MFI, i, Offset, StackGrowsDown, MaxAlign);
258 // Remember how big this blob of stack space is
259 MFI.setLocalFrameSize(Offset);
260 MFI.setLocalFrameMaxAlign(MaxAlign);
264 lookupCandidateBaseReg(unsigned BaseReg,
266 int64_t FrameSizeAdjust,
267 int64_t LocalFrameOffset,
268 const MachineInstr &MI,
269 const TargetRegisterInfo *TRI) {
270 // Check if the relative offset from the where the base register references
271 // to the target address is in range for the instruction.
272 int64_t Offset = FrameSizeAdjust + LocalFrameOffset - BaseOffset;
273 return TRI->isFrameOffsetLegal(&MI, BaseReg, Offset);
276 bool LocalStackSlotPass::insertFrameReferenceRegisters(MachineFunction &Fn) {
277 // Scan the function's instructions looking for frame index references.
278 // For each, ask the target if it wants a virtual base register for it
279 // based on what we can tell it about where the local will end up in the
280 // stack frame. If it wants one, re-use a suitable one we've previously
281 // allocated, or if there isn't one that fits the bill, allocate a new one
282 // and ask the target to create a defining instruction for it.
283 bool UsedBaseReg = false;
285 MachineFrameInfo &MFI = Fn.getFrameInfo();
286 const TargetRegisterInfo *TRI = Fn.getSubtarget().getRegisterInfo();
287 const TargetFrameLowering &TFI = *Fn.getSubtarget().getFrameLowering();
288 bool StackGrowsDown =
289 TFI.getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
291 // Collect all of the instructions in the block that reference
292 // a frame index. Also store the frame index referenced to ease later
293 // lookup. (For any insn that has more than one FI reference, we arbitrarily
294 // choose the first one).
295 SmallVector<FrameRef, 64> FrameReferenceInsns;
299 for (MachineBasicBlock &BB : Fn) {
300 for (MachineInstr &MI : BB) {
301 // Debug value, stackmap and patchpoint instructions can't be out of
302 // range, so they don't need any updates.
303 if (MI.isDebugValue() || MI.getOpcode() == TargetOpcode::STATEPOINT ||
304 MI.getOpcode() == TargetOpcode::STACKMAP ||
305 MI.getOpcode() == TargetOpcode::PATCHPOINT)
308 // For now, allocate the base register(s) within the basic block
309 // where they're used, and don't try to keep them around outside
310 // of that. It may be beneficial to try sharing them more broadly
311 // than that, but the increased register pressure makes that a
312 // tricky thing to balance. Investigate if re-materializing these
314 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
315 // Consider replacing all frame index operands that reference
316 // an object allocated in the local block.
317 if (MI.getOperand(i).isFI()) {
318 // Don't try this with values not in the local block.
319 if (!MFI.isObjectPreAllocated(MI.getOperand(i).getIndex()))
321 int Idx = MI.getOperand(i).getIndex();
322 int64_t LocalOffset = LocalOffsets[Idx];
323 if (!TRI->needsFrameBaseReg(&MI, LocalOffset))
325 FrameReferenceInsns.push_back(FrameRef(&MI, LocalOffset, Idx, Order++));
332 // Sort the frame references by local offset.
333 // Use frame index as a tie-breaker in case MI's have the same offset.
334 std::sort(FrameReferenceInsns.begin(), FrameReferenceInsns.end());
336 MachineBasicBlock *Entry = &Fn.front();
338 unsigned BaseReg = 0;
339 int64_t BaseOffset = 0;
341 // Loop through the frame references and allocate for them as necessary.
342 for (int ref = 0, e = FrameReferenceInsns.size(); ref < e ; ++ref) {
343 FrameRef &FR = FrameReferenceInsns[ref];
344 MachineInstr &MI = *FR.getMachineInstr();
345 int64_t LocalOffset = FR.getLocalOffset();
346 int FrameIdx = FR.getFrameIndex();
347 assert(MFI.isObjectPreAllocated(FrameIdx) &&
348 "Only pre-allocated locals expected!");
350 DEBUG(dbgs() << "Considering: " << MI);
353 for (unsigned f = MI.getNumOperands(); idx != f; ++idx) {
354 if (!MI.getOperand(idx).isFI())
357 if (FrameIdx == MI.getOperand(idx).getIndex())
361 assert(idx < MI.getNumOperands() && "Cannot find FI operand");
364 int64_t FrameSizeAdjust = StackGrowsDown ? MFI.getLocalFrameSize() : 0;
366 DEBUG(dbgs() << " Replacing FI in: " << MI);
368 // If we have a suitable base register available, use it; otherwise
369 // create a new one. Note that any offset encoded in the
370 // instruction itself will be taken into account by the target,
371 // so we don't have to adjust for it here when reusing a base
374 lookupCandidateBaseReg(BaseReg, BaseOffset, FrameSizeAdjust,
375 LocalOffset, MI, TRI)) {
376 DEBUG(dbgs() << " Reusing base register " << BaseReg << "\n");
377 // We found a register to reuse.
378 Offset = FrameSizeAdjust + LocalOffset - BaseOffset;
380 // No previously defined register was in range, so create a new one.
381 int64_t InstrOffset = TRI->getFrameIndexInstrOffset(&MI, idx);
383 int64_t PrevBaseOffset = BaseOffset;
384 BaseOffset = FrameSizeAdjust + LocalOffset + InstrOffset;
386 // We'd like to avoid creating single-use virtual base registers.
387 // Because the FrameRefs are in sorted order, and we've already
388 // processed all FrameRefs before this one, just check whether or not
389 // the next FrameRef will be able to reuse this new register. If not,
390 // then don't bother creating it.
392 !lookupCandidateBaseReg(
393 BaseReg, BaseOffset, FrameSizeAdjust,
394 FrameReferenceInsns[ref + 1].getLocalOffset(),
395 *FrameReferenceInsns[ref + 1].getMachineInstr(), TRI)) {
396 BaseOffset = PrevBaseOffset;
400 const MachineFunction *MF = MI.getParent()->getParent();
401 const TargetRegisterClass *RC = TRI->getPointerRegClass(*MF);
402 BaseReg = Fn.getRegInfo().createVirtualRegister(RC);
404 DEBUG(dbgs() << " Materializing base register " << BaseReg <<
405 " at frame local offset " << LocalOffset + InstrOffset << "\n");
407 // Tell the target to insert the instruction to initialize
408 // the base register.
409 // MachineBasicBlock::iterator InsertionPt = Entry->begin();
410 TRI->materializeFrameBaseRegister(Entry, BaseReg, FrameIdx,
413 // The base register already includes any offset specified
414 // by the instruction, so account for that so it doesn't get
416 Offset = -InstrOffset;
421 assert(BaseReg != 0 && "Unable to allocate virtual base register!");
423 // Modify the instruction to use the new base register rather
424 // than the frame index operand.
425 TRI->resolveFrameIndex(MI, BaseReg, Offset);
426 DEBUG(dbgs() << "Resolved: " << MI);