[x86] fix assert with horizontal math + broadcast of vector (PR43402)
[llvm-core.git] / lib / Target / X86 / X86WinAllocaExpander.cpp
blobae72c6427588ee40e0202954844cf7e472f2b0ed
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.
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.
15 //===----------------------------------------------------------------------===//
17 #include "X86.h"
18 #include "X86InstrBuilder.h"
19 #include "X86InstrInfo.h"
20 #include "X86MachineFunctionInfo.h"
21 #include "X86Subtarget.h"
22 #include "llvm/ADT/PostOrderIterator.h"
23 #include "llvm/CodeGen/MachineFunctionPass.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/Passes.h"
27 #include "llvm/CodeGen/TargetInstrInfo.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/Support/raw_ostream.h"
31 using namespace llvm;
33 namespace {
35 class X86WinAllocaExpander : public MachineFunctionPass {
36 public:
37 X86WinAllocaExpander() : MachineFunctionPass(ID) {}
39 bool runOnMachineFunction(MachineFunction &MF) override;
41 private:
42 /// Strategies for lowering a WinAlloca.
43 enum Lowering { TouchAndSub, Sub, Probe };
45 /// Deterministic-order map from WinAlloca instruction to desired lowering.
46 typedef MapVector<MachineInstr*, Lowering> LoweringMap;
48 /// Compute which lowering to use for each WinAlloca instruction.
49 void computeLowerings(MachineFunction &MF, LoweringMap& Lowerings);
51 /// Get the appropriate lowering based on current offset and amount.
52 Lowering getLowering(int64_t CurrentOffset, int64_t AllocaAmount);
54 /// Lower a WinAlloca instruction.
55 void lower(MachineInstr* MI, Lowering L);
57 MachineRegisterInfo *MRI;
58 const X86Subtarget *STI;
59 const TargetInstrInfo *TII;
60 const X86RegisterInfo *TRI;
61 unsigned StackPtr;
62 unsigned SlotSize;
63 int64_t StackProbeSize;
64 bool NoStackArgProbe;
66 StringRef getPassName() const override { return "X86 WinAlloca Expander"; }
67 static char ID;
70 char X86WinAllocaExpander::ID = 0;
72 } // end anonymous namespace
74 FunctionPass *llvm::createX86WinAllocaExpander() {
75 return new X86WinAllocaExpander();
78 /// Return the allocation amount for a WinAlloca instruction, or -1 if unknown.
79 static int64_t getWinAllocaAmount(MachineInstr *MI, MachineRegisterInfo *MRI) {
80 assert(MI->getOpcode() == X86::WIN_ALLOCA_32 ||
81 MI->getOpcode() == X86::WIN_ALLOCA_64);
82 assert(MI->getOperand(0).isReg());
84 Register AmountReg = MI->getOperand(0).getReg();
85 MachineInstr *Def = MRI->getUniqueVRegDef(AmountReg);
87 if (!Def ||
88 (Def->getOpcode() != X86::MOV32ri && Def->getOpcode() != X86::MOV64ri) ||
89 !Def->getOperand(1).isImm())
90 return -1;
92 return Def->getOperand(1).getImm();
95 X86WinAllocaExpander::Lowering
96 X86WinAllocaExpander::getLowering(int64_t CurrentOffset,
97 int64_t AllocaAmount) {
98 // For a non-constant amount or a large amount, we have to probe.
99 if (AllocaAmount < 0 || AllocaAmount > StackProbeSize)
100 return Probe;
102 // If it fits within the safe region of the stack, just subtract.
103 if (CurrentOffset + AllocaAmount <= StackProbeSize)
104 return Sub;
106 // Otherwise, touch the current tip of the stack, then subtract.
107 return TouchAndSub;
110 static bool isPushPop(const MachineInstr &MI) {
111 switch (MI.getOpcode()) {
112 case X86::PUSH32i8:
113 case X86::PUSH32r:
114 case X86::PUSH32rmm:
115 case X86::PUSH32rmr:
116 case X86::PUSHi32:
117 case X86::PUSH64i8:
118 case X86::PUSH64r:
119 case X86::PUSH64rmm:
120 case X86::PUSH64rmr:
121 case X86::PUSH64i32:
122 case X86::POP32r:
123 case X86::POP64r:
124 return true;
125 default:
126 return false;
130 void X86WinAllocaExpander::computeLowerings(MachineFunction &MF,
131 LoweringMap &Lowerings) {
132 // Do a one-pass reverse post-order walk of the CFG to conservatively estimate
133 // the offset between the stack pointer and the lowest touched part of the
134 // stack, and use that to decide how to lower each WinAlloca instruction.
136 // Initialize OutOffset[B], the stack offset at exit from B, to something big.
137 DenseMap<MachineBasicBlock *, int64_t> OutOffset;
138 for (MachineBasicBlock &MBB : MF)
139 OutOffset[&MBB] = INT32_MAX;
141 // Note: we don't know the offset at the start of the entry block since the
142 // prologue hasn't been inserted yet, and how much that will adjust the stack
143 // pointer depends on register spills, which have not been computed yet.
145 // Compute the reverse post-order.
146 ReversePostOrderTraversal<MachineFunction*> RPO(&MF);
148 for (MachineBasicBlock *MBB : RPO) {
149 int64_t Offset = -1;
150 for (MachineBasicBlock *Pred : MBB->predecessors())
151 Offset = std::max(Offset, OutOffset[Pred]);
152 if (Offset == -1) Offset = INT32_MAX;
154 for (MachineInstr &MI : *MBB) {
155 if (MI.getOpcode() == X86::WIN_ALLOCA_32 ||
156 MI.getOpcode() == X86::WIN_ALLOCA_64) {
157 // A WinAlloca moves StackPtr, and potentially touches it.
158 int64_t Amount = getWinAllocaAmount(&MI, MRI);
159 Lowering L = getLowering(Offset, Amount);
160 Lowerings[&MI] = L;
161 switch (L) {
162 case Sub:
163 Offset += Amount;
164 break;
165 case TouchAndSub:
166 Offset = Amount;
167 break;
168 case Probe:
169 Offset = 0;
170 break;
172 } else if (MI.isCall() || isPushPop(MI)) {
173 // Calls, pushes and pops touch the tip of the stack.
174 Offset = 0;
175 } else if (MI.getOpcode() == X86::ADJCALLSTACKUP32 ||
176 MI.getOpcode() == X86::ADJCALLSTACKUP64) {
177 Offset -= MI.getOperand(0).getImm();
178 } else if (MI.getOpcode() == X86::ADJCALLSTACKDOWN32 ||
179 MI.getOpcode() == X86::ADJCALLSTACKDOWN64) {
180 Offset += MI.getOperand(0).getImm();
181 } else if (MI.modifiesRegister(StackPtr, TRI)) {
182 // Any other modification of SP means we've lost track of it.
183 Offset = INT32_MAX;
187 OutOffset[MBB] = Offset;
191 static unsigned getSubOpcode(bool Is64Bit, int64_t Amount) {
192 if (Is64Bit)
193 return isInt<8>(Amount) ? X86::SUB64ri8 : X86::SUB64ri32;
194 return isInt<8>(Amount) ? X86::SUB32ri8 : X86::SUB32ri;
197 void X86WinAllocaExpander::lower(MachineInstr* MI, Lowering L) {
198 DebugLoc DL = MI->getDebugLoc();
199 MachineBasicBlock *MBB = MI->getParent();
200 MachineBasicBlock::iterator I = *MI;
202 int64_t Amount = getWinAllocaAmount(MI, MRI);
203 if (Amount == 0) {
204 MI->eraseFromParent();
205 return;
208 // These two variables differ on x32, which is a 64-bit target with a
209 // 32-bit alloca.
210 bool Is64Bit = STI->is64Bit();
211 bool Is64BitAlloca = MI->getOpcode() == X86::WIN_ALLOCA_64;
212 assert(SlotSize == 4 || SlotSize == 8);
214 switch (L) {
215 case TouchAndSub: {
216 assert(Amount >= SlotSize);
218 // Use a push to touch the top of the stack.
219 unsigned RegA = Is64Bit ? X86::RAX : X86::EAX;
220 BuildMI(*MBB, I, DL, TII->get(Is64Bit ? X86::PUSH64r : X86::PUSH32r))
221 .addReg(RegA, RegState::Undef);
222 Amount -= SlotSize;
223 if (!Amount)
224 break;
226 // Fall through to make any remaining adjustment.
227 LLVM_FALLTHROUGH;
229 case Sub:
230 assert(Amount > 0);
231 if (Amount == SlotSize) {
232 // Use push to save size.
233 unsigned RegA = Is64Bit ? X86::RAX : X86::EAX;
234 BuildMI(*MBB, I, DL, TII->get(Is64Bit ? X86::PUSH64r : X86::PUSH32r))
235 .addReg(RegA, RegState::Undef);
236 } else {
237 // Sub.
238 BuildMI(*MBB, I, DL,
239 TII->get(getSubOpcode(Is64BitAlloca, Amount)), StackPtr)
240 .addReg(StackPtr)
241 .addImm(Amount);
243 break;
244 case Probe:
245 if (!NoStackArgProbe) {
246 // The probe lowering expects the amount in RAX/EAX.
247 unsigned RegA = Is64BitAlloca ? X86::RAX : X86::EAX;
248 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::COPY), RegA)
249 .addReg(MI->getOperand(0).getReg());
251 // Do the probe.
252 STI->getFrameLowering()->emitStackProbe(*MBB->getParent(), *MBB, MI, DL,
253 /*InProlog=*/false);
254 } else {
255 // Sub
256 BuildMI(*MBB, I, DL,
257 TII->get(Is64BitAlloca ? X86::SUB64rr : X86::SUB32rr), StackPtr)
258 .addReg(StackPtr)
259 .addReg(MI->getOperand(0).getReg());
261 break;
264 Register AmountReg = MI->getOperand(0).getReg();
265 MI->eraseFromParent();
267 // Delete the definition of AmountReg.
268 if (MRI->use_empty(AmountReg))
269 if (MachineInstr *AmountDef = MRI->getUniqueVRegDef(AmountReg))
270 AmountDef->eraseFromParent();
273 bool X86WinAllocaExpander::runOnMachineFunction(MachineFunction &MF) {
274 if (!MF.getInfo<X86MachineFunctionInfo>()->hasWinAlloca())
275 return false;
277 MRI = &MF.getRegInfo();
278 STI = &MF.getSubtarget<X86Subtarget>();
279 TII = STI->getInstrInfo();
280 TRI = STI->getRegisterInfo();
281 StackPtr = TRI->getStackRegister();
282 SlotSize = TRI->getSlotSize();
284 StackProbeSize = 4096;
285 if (MF.getFunction().hasFnAttribute("stack-probe-size")) {
286 MF.getFunction()
287 .getFnAttribute("stack-probe-size")
288 .getValueAsString()
289 .getAsInteger(0, StackProbeSize);
291 NoStackArgProbe = MF.getFunction().hasFnAttribute("no-stack-arg-probe");
292 if (NoStackArgProbe)
293 StackProbeSize = INT64_MAX;
295 LoweringMap Lowerings;
296 computeLowerings(MF, Lowerings);
297 for (auto &P : Lowerings)
298 lower(P.first, P.second);
300 return true;