1 //===----- X86WinAllocaExpander.cpp - Expand WinAlloca pseudo instruction -===//
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 defines a pass that expands WinAlloca pseudo-instructions.
10 //
11 // It performs a conservative analysis to determine whether each allocation
12 // falls within a region of the stack that is safe to use, or whether stack
13 // probes must be emitted.
14 //
15 //===----------------------------------------------------------------------===//
16
17 #include "X86.h"
18 #include "X86InstrBuilder.h"
19 #include "X86InstrInfo.h"
20 #include "X86MachineFunctionInfo.h"
21 #include "X86Subtarget.h"
22 #include "llvm/ADT/MapVector.h"
23 #include "llvm/ADT/PostOrderIterator.h"
24 #include "llvm/CodeGen/MachineFunctionPass.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/Passes.h"
28 #include "llvm/CodeGen/TargetInstrInfo.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/Support/raw_ostream.h"
31
32 using namespace llvm;
33
34 namespace {
35
36 class X86WinAllocaExpander : public MachineFunctionPass {
37 public:
X86WinAllocaExpander()38 X86WinAllocaExpander() : MachineFunctionPass(ID) {}
39
40 bool runOnMachineFunction(MachineFunction &MF) override;
41
42 private:
43 /// Strategies for lowering a WinAlloca.
44 enum Lowering { TouchAndSub, Sub, Probe };
45
46 /// Deterministic-order map from WinAlloca instruction to desired lowering.
47 typedef MapVector<MachineInstr*, Lowering> LoweringMap;
48
49 /// Compute which lowering to use for each WinAlloca instruction.
50 void computeLowerings(MachineFunction &MF, LoweringMap& Lowerings);
51
52 /// Get the appropriate lowering based on current offset and amount.
53 Lowering getLowering(int64_t CurrentOffset, int64_t AllocaAmount);
54
55 /// Lower a WinAlloca instruction.
56 void lower(MachineInstr* MI, Lowering L);
57
58 MachineRegisterInfo *MRI = nullptr;
59 const X86Subtarget *STI = nullptr;
60 const TargetInstrInfo *TII = nullptr;
61 const X86RegisterInfo *TRI = nullptr;
62 unsigned StackPtr = 0;
63 unsigned SlotSize = 0;
64 int64_t StackProbeSize = 0;
65 bool NoStackArgProbe = false;
66
getPassName() const67 StringRef getPassName() const override { return "X86 WinAlloca Expander"; }
68 static char ID;
69 };
70
71 char X86WinAllocaExpander::ID = 0;
72
73 } // end anonymous namespace
74
createX86WinAllocaExpander()75 FunctionPass *llvm::createX86WinAllocaExpander() {
76 return new X86WinAllocaExpander();
77 }
78
79 /// Return the allocation amount for a WinAlloca instruction, or -1 if unknown.
getWinAllocaAmount(MachineInstr * MI,MachineRegisterInfo * MRI)80 static int64_t getWinAllocaAmount(MachineInstr *MI, MachineRegisterInfo *MRI) {
81 assert(MI->getOpcode() == X86::WIN_ALLOCA_32 ||
82 MI->getOpcode() == X86::WIN_ALLOCA_64);
83 assert(MI->getOperand(0).isReg());
84
85 Register AmountReg = MI->getOperand(0).getReg();
86 MachineInstr *Def = MRI->getUniqueVRegDef(AmountReg);
87
88 if (!Def ||
89 (Def->getOpcode() != X86::MOV32ri && Def->getOpcode() != X86::MOV64ri) ||
90 !Def->getOperand(1).isImm())
91 return -1;
92
93 return Def->getOperand(1).getImm();
94 }
95
96 X86WinAllocaExpander::Lowering
getLowering(int64_t CurrentOffset,int64_t AllocaAmount)97 X86WinAllocaExpander::getLowering(int64_t CurrentOffset,
98 int64_t AllocaAmount) {
99 // For a non-constant amount or a large amount, we have to probe.
100 if (AllocaAmount < 0 || AllocaAmount > StackProbeSize)
101 return Probe;
102
103 // If it fits within the safe region of the stack, just subtract.
104 if (CurrentOffset + AllocaAmount <= StackProbeSize)
105 return Sub;
106
107 // Otherwise, touch the current tip of the stack, then subtract.
108 return TouchAndSub;
109 }
110
isPushPop(const MachineInstr & MI)111 static bool isPushPop(const MachineInstr &MI) {
112 switch (MI.getOpcode()) {
113 case X86::PUSH32i8:
114 case X86::PUSH32r:
115 case X86::PUSH32rmm:
116 case X86::PUSH32rmr:
117 case X86::PUSHi32:
118 case X86::PUSH64i8:
119 case X86::PUSH64r:
120 case X86::PUSH64rmm:
121 case X86::PUSH64rmr:
122 case X86::PUSH64i32:
123 case X86::POP32r:
124 case X86::POP64r:
125 return true;
126 default:
127 return false;
128 }
129 }
130
computeLowerings(MachineFunction & MF,LoweringMap & Lowerings)131 void X86WinAllocaExpander::computeLowerings(MachineFunction &MF,
132 LoweringMap &Lowerings) {
133 // Do a one-pass reverse post-order walk of the CFG to conservatively estimate
134 // the offset between the stack pointer and the lowest touched part of the
135 // stack, and use that to decide how to lower each WinAlloca instruction.
136
137 // Initialize OutOffset[B], the stack offset at exit from B, to something big.
138 DenseMap<MachineBasicBlock *, int64_t> OutOffset;
139 for (MachineBasicBlock &MBB : MF)
140 OutOffset[&MBB] = INT32_MAX;
141
142 // Note: we don't know the offset at the start of the entry block since the
143 // prologue hasn't been inserted yet, and how much that will adjust the stack
144 // pointer depends on register spills, which have not been computed yet.
145
146 // Compute the reverse post-order.
147 ReversePostOrderTraversal<MachineFunction*> RPO(&MF);
148
149 for (MachineBasicBlock *MBB : RPO) {
150 int64_t Offset = -1;
151 for (MachineBasicBlock *Pred : MBB->predecessors())
152 Offset = std::max(Offset, OutOffset[Pred]);
153 if (Offset == -1) Offset = INT32_MAX;
154
155 for (MachineInstr &MI : *MBB) {
156 if (MI.getOpcode() == X86::WIN_ALLOCA_32 ||
157 MI.getOpcode() == X86::WIN_ALLOCA_64) {
158 // A WinAlloca moves StackPtr, and potentially touches it.
159 int64_t Amount = getWinAllocaAmount(&MI, MRI);
160 Lowering L = getLowering(Offset, Amount);
161 Lowerings[&MI] = L;
162 switch (L) {
163 case Sub:
164 Offset += Amount;
165 break;
166 case TouchAndSub:
167 Offset = Amount;
168 break;
169 case Probe:
170 Offset = 0;
171 break;
172 }
173 } else if (MI.isCall() || isPushPop(MI)) {
174 // Calls, pushes and pops touch the tip of the stack.
175 Offset = 0;
176 } else if (MI.getOpcode() == X86::ADJCALLSTACKUP32 ||
177 MI.getOpcode() == X86::ADJCALLSTACKUP64) {
178 Offset -= MI.getOperand(0).getImm();
179 } else if (MI.getOpcode() == X86::ADJCALLSTACKDOWN32 ||
180 MI.getOpcode() == X86::ADJCALLSTACKDOWN64) {
181 Offset += MI.getOperand(0).getImm();
182 } else if (MI.modifiesRegister(StackPtr, TRI)) {
183 // Any other modification of SP means we've lost track of it.
184 Offset = INT32_MAX;
185 }
186 }
187
188 OutOffset[MBB] = Offset;
189 }
190 }
191
getSubOpcode(bool Is64Bit,int64_t Amount)192 static unsigned getSubOpcode(bool Is64Bit, int64_t Amount) {
193 if (Is64Bit)
194 return isInt<8>(Amount) ? X86::SUB64ri8 : X86::SUB64ri32;
195 return isInt<8>(Amount) ? X86::SUB32ri8 : X86::SUB32ri;
196 }
197
lower(MachineInstr * MI,Lowering L)198 void X86WinAllocaExpander::lower(MachineInstr* MI, Lowering L) {
199 DebugLoc DL = MI->getDebugLoc();
200 MachineBasicBlock *MBB = MI->getParent();
201 MachineBasicBlock::iterator I = *MI;
202
203 int64_t Amount = getWinAllocaAmount(MI, MRI);
204 if (Amount == 0) {
205 MI->eraseFromParent();
206 return;
207 }
208
209 // These two variables differ on x32, which is a 64-bit target with a
210 // 32-bit alloca.
211 bool Is64Bit = STI->is64Bit();
212 bool Is64BitAlloca = MI->getOpcode() == X86::WIN_ALLOCA_64;
213 assert(SlotSize == 4 || SlotSize == 8);
214
215 switch (L) {
216 case TouchAndSub: {
217 assert(Amount >= SlotSize);
218
219 // Use a push to touch the top of the stack.
220 unsigned RegA = Is64Bit ? X86::RAX : X86::EAX;
221 BuildMI(*MBB, I, DL, TII->get(Is64Bit ? X86::PUSH64r : X86::PUSH32r))
222 .addReg(RegA, RegState::Undef);
223 Amount -= SlotSize;
224 if (!Amount)
225 break;
226
227 // Fall through to make any remaining adjustment.
228 LLVM_FALLTHROUGH;
229 }
230 case Sub:
231 assert(Amount > 0);
232 if (Amount == SlotSize) {
233 // Use push to save size.
234 unsigned RegA = Is64Bit ? X86::RAX : X86::EAX;
235 BuildMI(*MBB, I, DL, TII->get(Is64Bit ? X86::PUSH64r : X86::PUSH32r))
236 .addReg(RegA, RegState::Undef);
237 } else {
238 // Sub.
239 BuildMI(*MBB, I, DL,
240 TII->get(getSubOpcode(Is64BitAlloca, Amount)), StackPtr)
241 .addReg(StackPtr)
242 .addImm(Amount);
243 }
244 break;
245 case Probe:
246 if (!NoStackArgProbe) {
247 // The probe lowering expects the amount in RAX/EAX.
248 unsigned RegA = Is64BitAlloca ? X86::RAX : X86::EAX;
249 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::COPY), RegA)
250 .addReg(MI->getOperand(0).getReg());
251
252 // Do the probe.
253 STI->getFrameLowering()->emitStackProbe(*MBB->getParent(), *MBB, MI, DL,
254 /*InProlog=*/false);
255 } else {
256 // Sub
257 BuildMI(*MBB, I, DL,
258 TII->get(Is64BitAlloca ? X86::SUB64rr : X86::SUB32rr), StackPtr)
259 .addReg(StackPtr)
260 .addReg(MI->getOperand(0).getReg());
261 }
262 break;
263 }
264
265 Register AmountReg = MI->getOperand(0).getReg();
266 MI->eraseFromParent();
267
268 // Delete the definition of AmountReg.
269 if (MRI->use_empty(AmountReg))
270 if (MachineInstr *AmountDef = MRI->getUniqueVRegDef(AmountReg))
271 AmountDef->eraseFromParent();
272 }
273
runOnMachineFunction(MachineFunction & MF)274 bool X86WinAllocaExpander::runOnMachineFunction(MachineFunction &MF) {
275 if (!MF.getInfo<X86MachineFunctionInfo>()->hasWinAlloca())
276 return false;
277
278 MRI = &MF.getRegInfo();
279 STI = &MF.getSubtarget<X86Subtarget>();
280 TII = STI->getInstrInfo();
281 TRI = STI->getRegisterInfo();
282 StackPtr = TRI->getStackRegister();
283 SlotSize = TRI->getSlotSize();
284
285 StackProbeSize = 4096;
286 if (MF.getFunction().hasFnAttribute("stack-probe-size")) {
287 MF.getFunction()
288 .getFnAttribute("stack-probe-size")
289 .getValueAsString()
290 .getAsInteger(0, StackProbeSize);
291 }
292 NoStackArgProbe = MF.getFunction().hasFnAttribute("no-stack-arg-probe");
293 if (NoStackArgProbe)
294 StackProbeSize = INT64_MAX;
295
296 LoweringMap Lowerings;
297 computeLowerings(MF, Lowerings);
298 for (auto &P : Lowerings)
299 lower(P.first, P.second);
300
301 return true;
302 }
303