1 //===- LocalStackSlotAllocation.cpp - Pre-allocate locals to stack slots --===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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.
14 //
15 //===----------------------------------------------------------------------===//
16 
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"
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   class FrameRef {
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
54   public:
FrameRef(MachineBasicBlock::iterator I,int64_t Offset,int Idx)55     FrameRef(MachineBasicBlock::iterator I, int64_t Offset, int Idx) :
56       MI(I), LocalOffset(Offset), FrameIdx(Idx) {}
operator <(const FrameRef & RHS) const57     bool operator<(const FrameRef &RHS) const {
58       return LocalOffset < RHS.LocalOffset;
59     }
getMachineInstr() const60     MachineBasicBlock::iterator getMachineInstr() const { return MI; }
getLocalOffset() const61     int64_t getLocalOffset() const { return LocalOffset; }
getFrameIndex() const62     int getFrameIndex() const { return FrameIdx; }
63   };
64 
65   class LocalStackSlotPass: public MachineFunctionPass {
66     SmallVector<int64_t,16> LocalOffsets;
67     /// StackObjSet - A set of stack object indexes
68     typedef SmallSetVector<int, 8> StackObjSet;
69 
70     void AdjustStackOffset(MachineFrameInfo *MFI, int FrameIdx, int64_t &Offset,
71                            bool StackGrowsDown, unsigned &MaxAlign);
72     void AssignProtectedObjSet(const StackObjSet &UnassignedObjs,
73                                SmallSet<int, 16> &ProtectedObjs,
74                                MachineFrameInfo *MFI, bool StackGrowsDown,
75                                int64_t &Offset, unsigned &MaxAlign);
76     void calculateFrameObjectOffsets(MachineFunction &Fn);
77     bool insertFrameReferenceRegisters(MachineFunction &Fn);
78   public:
79     static char ID; // Pass identification, replacement for typeid
LocalStackSlotPass()80     explicit LocalStackSlotPass() : MachineFunctionPass(ID) {
81       initializeLocalStackSlotPassPass(*PassRegistry::getPassRegistry());
82     }
83     bool runOnMachineFunction(MachineFunction &MF) override;
84 
getAnalysisUsage(AnalysisUsage & AU) const85     void getAnalysisUsage(AnalysisUsage &AU) const override {
86       AU.setPreservesCFG();
87       AU.addRequired<StackProtector>();
88       MachineFunctionPass::getAnalysisUsage(AU);
89     }
90 
91   private:
92   };
93 } // end anonymous namespace
94 
95 char LocalStackSlotPass::ID = 0;
96 char &llvm::LocalStackSlotAllocationID = LocalStackSlotPass::ID;
97 INITIALIZE_PASS_BEGIN(LocalStackSlotPass, "localstackalloc",
98                       "Local Stack Slot Allocation", false, false)
INITIALIZE_PASS_DEPENDENCY(StackProtector)99 INITIALIZE_PASS_DEPENDENCY(StackProtector)
100 INITIALIZE_PASS_END(LocalStackSlotPass, "localstackalloc",
101                     "Local Stack Slot Allocation", false, false)
102 
103 
104 bool LocalStackSlotPass::runOnMachineFunction(MachineFunction &MF) {
105   MachineFrameInfo *MFI = MF.getFrameInfo();
106   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
107   unsigned LocalObjectCount = MFI->getObjectIndexEnd();
108 
109   // If the target doesn't want/need this pass, or if there are no locals
110   // to consider, early exit.
111   if (!TRI->requiresVirtualBaseRegisters(MF) || LocalObjectCount == 0)
112     return true;
113 
114   // Make sure we have enough space to store the local offsets.
115   LocalOffsets.resize(MFI->getObjectIndexEnd());
116 
117   // Lay out the local blob.
118   calculateFrameObjectOffsets(MF);
119 
120   // Insert virtual base registers to resolve frame index references.
121   bool UsedBaseRegs = insertFrameReferenceRegisters(MF);
122 
123   // Tell MFI whether any base registers were allocated. PEI will only
124   // want to use the local block allocations from this pass if there were any.
125   // Otherwise, PEI can do a bit better job of getting the alignment right
126   // without a hole at the start since it knows the alignment of the stack
127   // at the start of local allocation, and this pass doesn't.
128   MFI->setUseLocalStackAllocationBlock(UsedBaseRegs);
129 
130   return true;
131 }
132 
133 /// AdjustStackOffset - Helper function used to adjust the stack frame offset.
AdjustStackOffset(MachineFrameInfo * MFI,int FrameIdx,int64_t & Offset,bool StackGrowsDown,unsigned & MaxAlign)134 void LocalStackSlotPass::AdjustStackOffset(MachineFrameInfo *MFI,
135                                            int FrameIdx, int64_t &Offset,
136                                            bool StackGrowsDown,
137                                            unsigned &MaxAlign) {
138   // If the stack grows down, add the object size to find the lowest address.
139   if (StackGrowsDown)
140     Offset += MFI->getObjectSize(FrameIdx);
141 
142   unsigned Align = MFI->getObjectAlignment(FrameIdx);
143 
144   // If the alignment of this object is greater than that of the stack, then
145   // increase the stack alignment to match.
146   MaxAlign = std::max(MaxAlign, Align);
147 
148   // Adjust to alignment boundary.
149   Offset = (Offset + Align - 1) / Align * Align;
150 
151   int64_t LocalOffset = StackGrowsDown ? -Offset : Offset;
152   DEBUG(dbgs() << "Allocate FI(" << FrameIdx << ") to local offset "
153         << LocalOffset << "\n");
154   // Keep the offset available for base register allocation
155   LocalOffsets[FrameIdx] = LocalOffset;
156   // And tell MFI about it for PEI to use later
157   MFI->mapLocalFrameObject(FrameIdx, LocalOffset);
158 
159   if (!StackGrowsDown)
160     Offset += MFI->getObjectSize(FrameIdx);
161 
162   ++NumAllocations;
163 }
164 
165 /// AssignProtectedObjSet - Helper function to assign large stack objects (i.e.,
166 /// those required to be close to the Stack Protector) to stack offsets.
AssignProtectedObjSet(const StackObjSet & UnassignedObjs,SmallSet<int,16> & ProtectedObjs,MachineFrameInfo * MFI,bool StackGrowsDown,int64_t & Offset,unsigned & MaxAlign)167 void LocalStackSlotPass::AssignProtectedObjSet(const StackObjSet &UnassignedObjs,
168                                            SmallSet<int, 16> &ProtectedObjs,
169                                            MachineFrameInfo *MFI,
170                                            bool StackGrowsDown, int64_t &Offset,
171                                            unsigned &MaxAlign) {
172 
173   for (StackObjSet::const_iterator I = UnassignedObjs.begin(),
174         E = UnassignedObjs.end(); I != E; ++I) {
175     int i = *I;
176     AdjustStackOffset(MFI, i, Offset, StackGrowsDown, MaxAlign);
177     ProtectedObjs.insert(i);
178   }
179 }
180 
181 /// calculateFrameObjectOffsets - Calculate actual frame offsets for all of the
182 /// abstract stack objects.
183 ///
calculateFrameObjectOffsets(MachineFunction & Fn)184 void LocalStackSlotPass::calculateFrameObjectOffsets(MachineFunction &Fn) {
185   // Loop over all of the stack objects, assigning sequential addresses...
186   MachineFrameInfo *MFI = Fn.getFrameInfo();
187   const TargetFrameLowering &TFI = *Fn.getSubtarget().getFrameLowering();
188   bool StackGrowsDown =
189     TFI.getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
190   int64_t Offset = 0;
191   unsigned MaxAlign = 0;
192   StackProtector *SP = &getAnalysis<StackProtector>();
193 
194   // Make sure that the stack protector comes before the local variables on the
195   // stack.
196   SmallSet<int, 16> ProtectedObjs;
197   if (MFI->getStackProtectorIndex() >= 0) {
198     StackObjSet LargeArrayObjs;
199     StackObjSet SmallArrayObjs;
200     StackObjSet AddrOfObjs;
201 
202     AdjustStackOffset(MFI, MFI->getStackProtectorIndex(), Offset,
203                       StackGrowsDown, MaxAlign);
204 
205     // Assign large stack objects first.
206     for (unsigned i = 0, e = MFI->getObjectIndexEnd(); i != e; ++i) {
207       if (MFI->isDeadObjectIndex(i))
208         continue;
209       if (MFI->getStackProtectorIndex() == (int)i)
210         continue;
211 
212       switch (SP->getSSPLayout(MFI->getObjectAllocation(i))) {
213       case StackProtector::SSPLK_None:
214         continue;
215       case StackProtector::SSPLK_SmallArray:
216         SmallArrayObjs.insert(i);
217         continue;
218       case StackProtector::SSPLK_AddrOf:
219         AddrOfObjs.insert(i);
220         continue;
221       case StackProtector::SSPLK_LargeArray:
222         LargeArrayObjs.insert(i);
223         continue;
224       }
225       llvm_unreachable("Unexpected SSPLayoutKind.");
226     }
227 
228     AssignProtectedObjSet(LargeArrayObjs, ProtectedObjs, MFI, StackGrowsDown,
229                           Offset, MaxAlign);
230     AssignProtectedObjSet(SmallArrayObjs, ProtectedObjs, MFI, StackGrowsDown,
231                           Offset, MaxAlign);
232     AssignProtectedObjSet(AddrOfObjs, ProtectedObjs, MFI, StackGrowsDown,
233                           Offset, MaxAlign);
234   }
235 
236   // Then assign frame offsets to stack objects that are not used to spill
237   // callee saved registers.
238   for (unsigned i = 0, e = MFI->getObjectIndexEnd(); i != e; ++i) {
239     if (MFI->isDeadObjectIndex(i))
240       continue;
241     if (MFI->getStackProtectorIndex() == (int)i)
242       continue;
243     if (ProtectedObjs.count(i))
244       continue;
245 
246     AdjustStackOffset(MFI, i, Offset, StackGrowsDown, MaxAlign);
247   }
248 
249   // Remember how big this blob of stack space is
250   MFI->setLocalFrameSize(Offset);
251   MFI->setLocalFrameMaxAlign(MaxAlign);
252 }
253 
254 static inline bool
lookupCandidateBaseReg(unsigned BaseReg,int64_t BaseOffset,int64_t FrameSizeAdjust,int64_t LocalFrameOffset,const MachineInstr * MI,const TargetRegisterInfo * TRI)255 lookupCandidateBaseReg(unsigned BaseReg,
256                        int64_t BaseOffset,
257                        int64_t FrameSizeAdjust,
258                        int64_t LocalFrameOffset,
259                        const MachineInstr *MI,
260                        const TargetRegisterInfo *TRI) {
261   // Check if the relative offset from the where the base register references
262   // to the target address is in range for the instruction.
263   int64_t Offset = FrameSizeAdjust + LocalFrameOffset - BaseOffset;
264   return TRI->isFrameOffsetLegal(MI, BaseReg, Offset);
265 }
266 
insertFrameReferenceRegisters(MachineFunction & Fn)267 bool LocalStackSlotPass::insertFrameReferenceRegisters(MachineFunction &Fn) {
268   // Scan the function's instructions looking for frame index references.
269   // For each, ask the target if it wants a virtual base register for it
270   // based on what we can tell it about where the local will end up in the
271   // stack frame. If it wants one, re-use a suitable one we've previously
272   // allocated, or if there isn't one that fits the bill, allocate a new one
273   // and ask the target to create a defining instruction for it.
274   bool UsedBaseReg = false;
275 
276   MachineFrameInfo *MFI = Fn.getFrameInfo();
277   const TargetRegisterInfo *TRI = Fn.getSubtarget().getRegisterInfo();
278   const TargetFrameLowering &TFI = *Fn.getSubtarget().getFrameLowering();
279   bool StackGrowsDown =
280     TFI.getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
281 
282   // Collect all of the instructions in the block that reference
283   // a frame index. Also store the frame index referenced to ease later
284   // lookup. (For any insn that has more than one FI reference, we arbitrarily
285   // choose the first one).
286   SmallVector<FrameRef, 64> FrameReferenceInsns;
287 
288   for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
289     for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
290       MachineInstr *MI = I;
291 
292       // Debug value, stackmap and patchpoint instructions can't be out of
293       // range, so they don't need any updates.
294       if (MI->isDebugValue() ||
295           MI->getOpcode() == TargetOpcode::STATEPOINT ||
296           MI->getOpcode() == TargetOpcode::STACKMAP ||
297           MI->getOpcode() == TargetOpcode::PATCHPOINT)
298         continue;
299 
300       // For now, allocate the base register(s) within the basic block
301       // where they're used, and don't try to keep them around outside
302       // of that. It may be beneficial to try sharing them more broadly
303       // than that, but the increased register pressure makes that a
304       // tricky thing to balance. Investigate if re-materializing these
305       // becomes an issue.
306       for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
307         // Consider replacing all frame index operands that reference
308         // an object allocated in the local block.
309         if (MI->getOperand(i).isFI()) {
310           // Don't try this with values not in the local block.
311           if (!MFI->isObjectPreAllocated(MI->getOperand(i).getIndex()))
312             break;
313           int Idx = MI->getOperand(i).getIndex();
314           int64_t LocalOffset = LocalOffsets[Idx];
315           if (!TRI->needsFrameBaseReg(MI, LocalOffset))
316             break;
317           FrameReferenceInsns.
318             push_back(FrameRef(MI, LocalOffset, Idx));
319           break;
320         }
321       }
322     }
323   }
324 
325   // Sort the frame references by local offset
326   array_pod_sort(FrameReferenceInsns.begin(), FrameReferenceInsns.end());
327 
328   MachineBasicBlock *Entry = &Fn.front();
329 
330   unsigned BaseReg = 0;
331   int64_t BaseOffset = 0;
332 
333   // Loop through the frame references and allocate for them as necessary.
334   for (int ref = 0, e = FrameReferenceInsns.size(); ref < e ; ++ref) {
335     FrameRef &FR = FrameReferenceInsns[ref];
336     MachineBasicBlock::iterator I = FR.getMachineInstr();
337     MachineInstr *MI = I;
338     int64_t LocalOffset = FR.getLocalOffset();
339     int FrameIdx = FR.getFrameIndex();
340     assert(MFI->isObjectPreAllocated(FrameIdx) &&
341            "Only pre-allocated locals expected!");
342 
343     DEBUG(dbgs() << "Considering: " << *MI);
344 
345     unsigned idx = 0;
346     for (unsigned f = MI->getNumOperands(); idx != f; ++idx) {
347       if (!MI->getOperand(idx).isFI())
348         continue;
349 
350       if (FrameIdx == I->getOperand(idx).getIndex())
351         break;
352     }
353 
354     assert(idx < MI->getNumOperands() && "Cannot find FI operand");
355 
356     int64_t Offset = 0;
357     int64_t FrameSizeAdjust = StackGrowsDown ? MFI->getLocalFrameSize() : 0;
358 
359     DEBUG(dbgs() << "  Replacing FI in: " << *MI);
360 
361     // If we have a suitable base register available, use it; otherwise
362     // create a new one. Note that any offset encoded in the
363     // instruction itself will be taken into account by the target,
364     // so we don't have to adjust for it here when reusing a base
365     // register.
366     if (UsedBaseReg && lookupCandidateBaseReg(BaseReg, BaseOffset,
367                                               FrameSizeAdjust, LocalOffset, MI,
368                                               TRI)) {
369       DEBUG(dbgs() << "  Reusing base register " << BaseReg << "\n");
370       // We found a register to reuse.
371       Offset = FrameSizeAdjust + LocalOffset - BaseOffset;
372     } else {
373       // No previously defined register was in range, so create a // new one.
374 
375       int64_t InstrOffset = TRI->getFrameIndexInstrOffset(MI, idx);
376 
377       int64_t PrevBaseOffset = BaseOffset;
378       BaseOffset = FrameSizeAdjust + LocalOffset + InstrOffset;
379 
380       // We'd like to avoid creating single-use virtual base registers.
381       // Because the FrameRefs are in sorted order, and we've already
382       // processed all FrameRefs before this one, just check whether or not
383       // the next FrameRef will be able to reuse this new register. If not,
384       // then don't bother creating it.
385       if (ref + 1 >= e ||
386           !lookupCandidateBaseReg(
387               BaseReg, BaseOffset, FrameSizeAdjust,
388               FrameReferenceInsns[ref + 1].getLocalOffset(),
389               FrameReferenceInsns[ref + 1].getMachineInstr(), TRI)) {
390         BaseOffset = PrevBaseOffset;
391         continue;
392       }
393 
394       const MachineFunction *MF = MI->getParent()->getParent();
395       const TargetRegisterClass *RC = TRI->getPointerRegClass(*MF);
396       BaseReg = Fn.getRegInfo().createVirtualRegister(RC);
397 
398       DEBUG(dbgs() << "  Materializing base register " << BaseReg <<
399             " at frame local offset " << LocalOffset + InstrOffset << "\n");
400 
401       // Tell the target to insert the instruction to initialize
402       // the base register.
403       //            MachineBasicBlock::iterator InsertionPt = Entry->begin();
404       TRI->materializeFrameBaseRegister(Entry, BaseReg, FrameIdx,
405                                         InstrOffset);
406 
407       // The base register already includes any offset specified
408       // by the instruction, so account for that so it doesn't get
409       // applied twice.
410       Offset = -InstrOffset;
411 
412       ++NumBaseRegisters;
413       UsedBaseReg = true;
414     }
415     assert(BaseReg != 0 && "Unable to allocate virtual base register!");
416 
417     // Modify the instruction to use the new base register rather
418     // than the frame index operand.
419     TRI->resolveFrameIndex(*I, BaseReg, Offset);
420     DEBUG(dbgs() << "Resolved: " << *MI);
421 
422     ++NumReplacements;
423   }
424 
425   return UsedBaseReg;
426 }
427