1 //===----- X86WinAllocaExpander.cpp - Expand WinAlloca pseudo instruction -===//
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
7 //===----------------------------------------------------------------------===//
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 //===----------------------------------------------------------------------===//
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"
35 class X86WinAllocaExpander
: public MachineFunctionPass
{
37 X86WinAllocaExpander() : MachineFunctionPass(ID
) {}
39 bool runOnMachineFunction(MachineFunction
&MF
) override
;
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
;
63 int64_t StackProbeSize
;
66 StringRef
getPassName() const override
{ return "X86 WinAlloca Expander"; }
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 unsigned AmountReg
= MI
->getOperand(0).getReg();
85 MachineInstr
*Def
= MRI
->getUniqueVRegDef(AmountReg
);
87 // Look through copies.
88 while (Def
&& Def
->isCopy() && Def
->getOperand(1).isReg())
89 Def
= MRI
->getUniqueVRegDef(Def
->getOperand(1).getReg());
92 (Def
->getOpcode() != X86::MOV32ri
&& Def
->getOpcode() != X86::MOV64ri
) ||
93 !Def
->getOperand(1).isImm())
96 return Def
->getOperand(1).getImm();
99 X86WinAllocaExpander::Lowering
100 X86WinAllocaExpander::getLowering(int64_t CurrentOffset
,
101 int64_t AllocaAmount
) {
102 // For a non-constant amount or a large amount, we have to probe.
103 if (AllocaAmount
< 0 || AllocaAmount
> StackProbeSize
)
106 // If it fits within the safe region of the stack, just subtract.
107 if (CurrentOffset
+ AllocaAmount
<= StackProbeSize
)
110 // Otherwise, touch the current tip of the stack, then subtract.
114 static bool isPushPop(const MachineInstr
&MI
) {
115 switch (MI
.getOpcode()) {
134 void X86WinAllocaExpander::computeLowerings(MachineFunction
&MF
,
135 LoweringMap
&Lowerings
) {
136 // Do a one-pass reverse post-order walk of the CFG to conservatively estimate
137 // the offset between the stack pointer and the lowest touched part of the
138 // stack, and use that to decide how to lower each WinAlloca instruction.
140 // Initialize OutOffset[B], the stack offset at exit from B, to something big.
141 DenseMap
<MachineBasicBlock
*, int64_t> OutOffset
;
142 for (MachineBasicBlock
&MBB
: MF
)
143 OutOffset
[&MBB
] = INT32_MAX
;
145 // Note: we don't know the offset at the start of the entry block since the
146 // prologue hasn't been inserted yet, and how much that will adjust the stack
147 // pointer depends on register spills, which have not been computed yet.
149 // Compute the reverse post-order.
150 ReversePostOrderTraversal
<MachineFunction
*> RPO(&MF
);
152 for (MachineBasicBlock
*MBB
: RPO
) {
154 for (MachineBasicBlock
*Pred
: MBB
->predecessors())
155 Offset
= std::max(Offset
, OutOffset
[Pred
]);
156 if (Offset
== -1) Offset
= INT32_MAX
;
158 for (MachineInstr
&MI
: *MBB
) {
159 if (MI
.getOpcode() == X86::WIN_ALLOCA_32
||
160 MI
.getOpcode() == X86::WIN_ALLOCA_64
) {
161 // A WinAlloca moves StackPtr, and potentially touches it.
162 int64_t Amount
= getWinAllocaAmount(&MI
, MRI
);
163 Lowering L
= getLowering(Offset
, Amount
);
176 } else if (MI
.isCall() || isPushPop(MI
)) {
177 // Calls, pushes and pops touch the tip of the stack.
179 } else if (MI
.getOpcode() == X86::ADJCALLSTACKUP32
||
180 MI
.getOpcode() == X86::ADJCALLSTACKUP64
) {
181 Offset
-= MI
.getOperand(0).getImm();
182 } else if (MI
.getOpcode() == X86::ADJCALLSTACKDOWN32
||
183 MI
.getOpcode() == X86::ADJCALLSTACKDOWN64
) {
184 Offset
+= MI
.getOperand(0).getImm();
185 } else if (MI
.modifiesRegister(StackPtr
, TRI
)) {
186 // Any other modification of SP means we've lost track of it.
191 OutOffset
[MBB
] = Offset
;
195 static unsigned getSubOpcode(bool Is64Bit
, int64_t Amount
) {
197 return isInt
<8>(Amount
) ? X86::SUB64ri8
: X86::SUB64ri32
;
198 return isInt
<8>(Amount
) ? X86::SUB32ri8
: X86::SUB32ri
;
201 void X86WinAllocaExpander::lower(MachineInstr
* MI
, Lowering L
) {
202 DebugLoc DL
= MI
->getDebugLoc();
203 MachineBasicBlock
*MBB
= MI
->getParent();
204 MachineBasicBlock::iterator I
= *MI
;
206 int64_t Amount
= getWinAllocaAmount(MI
, MRI
);
208 MI
->eraseFromParent();
212 bool Is64Bit
= STI
->is64Bit();
213 assert(SlotSize
== 4 || SlotSize
== 8);
214 unsigned RegA
= (SlotSize
== 8) ? X86::RAX
: X86::EAX
;
218 assert(Amount
>= SlotSize
);
220 // Use a push to touch the top of the stack.
221 BuildMI(*MBB
, I
, DL
, TII
->get(Is64Bit
? X86::PUSH64r
: X86::PUSH32r
))
222 .addReg(RegA
, RegState::Undef
);
227 // Fall through to make any remaining adjustment.
231 if (Amount
== SlotSize
) {
232 // Use push to save size.
233 BuildMI(*MBB
, I
, DL
, TII
->get(Is64Bit
? X86::PUSH64r
: X86::PUSH32r
))
234 .addReg(RegA
, RegState::Undef
);
237 BuildMI(*MBB
, I
, DL
, TII
->get(getSubOpcode(Is64Bit
, Amount
)), StackPtr
)
243 if (!NoStackArgProbe
) {
244 // The probe lowering expects the amount in RAX/EAX.
245 BuildMI(*MBB
, MI
, DL
, TII
->get(TargetOpcode::COPY
), RegA
)
246 .addReg(MI
->getOperand(0).getReg());
249 STI
->getFrameLowering()->emitStackProbe(*MBB
->getParent(), *MBB
, MI
, DL
,
250 /*InPrologue=*/false);
253 BuildMI(*MBB
, I
, DL
, TII
->get(Is64Bit
? X86::SUB64rr
: X86::SUB32rr
),
256 .addReg(MI
->getOperand(0).getReg());
261 unsigned AmountReg
= MI
->getOperand(0).getReg();
262 MI
->eraseFromParent();
264 // Delete the definition of AmountReg, possibly walking a chain of copies.
266 if (!MRI
->use_empty(AmountReg
))
268 MachineInstr
*AmountDef
= MRI
->getUniqueVRegDef(AmountReg
);
271 if (AmountDef
->isCopy() && AmountDef
->getOperand(1).isReg())
272 AmountReg
= AmountDef
->getOperand(1).isReg();
273 AmountDef
->eraseFromParent();
278 bool X86WinAllocaExpander::runOnMachineFunction(MachineFunction
&MF
) {
279 if (!MF
.getInfo
<X86MachineFunctionInfo
>()->hasWinAlloca())
282 MRI
= &MF
.getRegInfo();
283 STI
= &MF
.getSubtarget
<X86Subtarget
>();
284 TII
= STI
->getInstrInfo();
285 TRI
= STI
->getRegisterInfo();
286 StackPtr
= TRI
->getStackRegister();
287 SlotSize
= TRI
->getSlotSize();
289 StackProbeSize
= 4096;
290 if (MF
.getFunction().hasFnAttribute("stack-probe-size")) {
292 .getFnAttribute("stack-probe-size")
294 .getAsInteger(0, StackProbeSize
);
296 NoStackArgProbe
= MF
.getFunction().hasFnAttribute("no-stack-arg-probe");
298 StackProbeSize
= INT64_MAX
;
300 LoweringMap Lowerings
;
301 computeLowerings(MF
, Lowerings
);
302 for (auto &P
: Lowerings
)
303 lower(P
.first
, P
.second
);