1 //===-- LoongArchFrameLowering.cpp - LoongArch Frame Information -*- C++ -*-==//
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 contains the LoongArch implementation of TargetFrameLowering class.
11 //===----------------------------------------------------------------------===//
13 #include "LoongArchFrameLowering.h"
14 #include "LoongArchMachineFunctionInfo.h"
15 #include "LoongArchSubtarget.h"
16 #include "MCTargetDesc/LoongArchBaseInfo.h"
17 #include "MCTargetDesc/LoongArchMCTargetDesc.h"
18 #include "llvm/CodeGen/MachineFrameInfo.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/MachineInstrBuilder.h"
21 #include "llvm/CodeGen/MachineRegisterInfo.h"
22 #include "llvm/CodeGen/RegisterScavenging.h"
23 #include "llvm/IR/DiagnosticInfo.h"
24 #include "llvm/MC/MCDwarf.h"
28 #define DEBUG_TYPE "loongarch-frame-lowering"
30 // Return true if the specified function should have a dedicated frame
31 // pointer register. This is true if frame pointer elimination is
32 // disabled, if it needs dynamic stack realignment, if the function has
33 // variable sized allocas, or if the frame address is taken.
34 bool LoongArchFrameLowering::hasFP(const MachineFunction
&MF
) const {
35 const TargetRegisterInfo
*RegInfo
= MF
.getSubtarget().getRegisterInfo();
37 const MachineFrameInfo
&MFI
= MF
.getFrameInfo();
38 return MF
.getTarget().Options
.DisableFramePointerElim(MF
) ||
39 RegInfo
->hasStackRealignment(MF
) || MFI
.hasVarSizedObjects() ||
40 MFI
.isFrameAddressTaken();
43 bool LoongArchFrameLowering::hasBP(const MachineFunction
&MF
) const {
44 const MachineFrameInfo
&MFI
= MF
.getFrameInfo();
45 const TargetRegisterInfo
*TRI
= STI
.getRegisterInfo();
47 return MFI
.hasVarSizedObjects() && TRI
->hasStackRealignment(MF
);
50 void LoongArchFrameLowering::adjustReg(MachineBasicBlock
&MBB
,
51 MachineBasicBlock::iterator MBBI
,
52 const DebugLoc
&DL
, Register DestReg
,
53 Register SrcReg
, int64_t Val
,
54 MachineInstr::MIFlag Flag
) const {
55 const LoongArchInstrInfo
*TII
= STI
.getInstrInfo();
56 bool IsLA64
= STI
.is64Bit();
57 unsigned Addi
= IsLA64
? LoongArch::ADDI_D
: LoongArch::ADDI_W
;
59 if (DestReg
== SrcReg
&& Val
== 0)
63 // addi.w/d $DstReg, $SrcReg, Val
64 BuildMI(MBB
, MBBI
, DL
, TII
->get(Addi
), DestReg
)
71 // Try to split the offset across two ADDIs. We need to keep the stack pointer
72 // aligned after each ADDI. We need to determine the maximum value we can put
73 // in each ADDI. In the negative direction, we can use -2048 which is always
74 // sufficiently aligned. In the positive direction, we need to find the
75 // largest 12-bit immediate that is aligned. Exclude -4096 since it can be
76 // created with LU12I.W.
77 assert(getStackAlign().value() < 2048 && "Stack alignment too large");
78 int64_t MaxPosAdjStep
= 2048 - getStackAlign().value();
79 if (Val
> -4096 && Val
<= (2 * MaxPosAdjStep
)) {
80 int64_t FirstAdj
= Val
< 0 ? -2048 : MaxPosAdjStep
;
82 BuildMI(MBB
, MBBI
, DL
, TII
->get(Addi
), DestReg
)
86 BuildMI(MBB
, MBBI
, DL
, TII
->get(Addi
), DestReg
)
87 .addReg(DestReg
, RegState::Kill
)
93 unsigned Opc
= IsLA64
? LoongArch::ADD_D
: LoongArch::ADD_W
;
96 Opc
= IsLA64
? LoongArch::SUB_D
: LoongArch::SUB_W
;
99 MachineRegisterInfo
&MRI
= MBB
.getParent()->getRegInfo();
100 Register ScratchReg
= MRI
.createVirtualRegister(&LoongArch::GPRRegClass
);
101 TII
->movImm(MBB
, MBBI
, DL
, ScratchReg
, Val
, Flag
);
102 BuildMI(MBB
, MBBI
, DL
, TII
->get(Opc
), DestReg
)
104 .addReg(ScratchReg
, RegState::Kill
)
108 // Determine the size of the frame and maximum call frame size.
109 void LoongArchFrameLowering::determineFrameLayout(MachineFunction
&MF
) const {
110 MachineFrameInfo
&MFI
= MF
.getFrameInfo();
112 // Get the number of bytes to allocate from the FrameInfo.
113 uint64_t FrameSize
= MFI
.getStackSize();
115 // Make sure the frame is aligned.
116 FrameSize
= alignTo(FrameSize
, getStackAlign());
118 // Update frame info.
119 MFI
.setStackSize(FrameSize
);
122 static uint64_t estimateFunctionSizeInBytes(const LoongArchInstrInfo
*TII
,
123 const MachineFunction
&MF
) {
124 uint64_t FuncSize
= 0;
127 FuncSize
+= TII
->getInstSizeInBytes(MI
);
131 static bool needScavSlotForCFR(MachineFunction
&MF
) {
132 if (!MF
.getSubtarget
<LoongArchSubtarget
>().hasBasicF())
136 if (MI
.getOpcode() == LoongArch::PseudoST_CFR
)
141 void LoongArchFrameLowering::processFunctionBeforeFrameFinalized(
142 MachineFunction
&MF
, RegScavenger
*RS
) const {
143 const LoongArchRegisterInfo
*RI
= STI
.getRegisterInfo();
144 const TargetRegisterClass
&RC
= LoongArch::GPRRegClass
;
145 const LoongArchInstrInfo
*TII
= STI
.getInstrInfo();
146 LoongArchMachineFunctionInfo
*LAFI
=
147 MF
.getInfo
<LoongArchMachineFunctionInfo
>();
148 MachineFrameInfo
&MFI
= MF
.getFrameInfo();
150 unsigned ScavSlotsNum
= 0;
152 // Far branches beyond 27-bit offset require a spill slot for scratch register.
153 bool IsLargeFunction
= !isInt
<27>(estimateFunctionSizeInBytes(TII
, MF
));
157 // estimateStackSize has been observed to under-estimate the final stack
158 // size, so give ourselves wiggle-room by checking for stack size
159 // representable an 11-bit signed field rather than 12-bits.
160 if (!isInt
<11>(MFI
.estimateStackSize(MF
)))
161 ScavSlotsNum
= std::max(ScavSlotsNum
, 1u);
164 if (needScavSlotForCFR(MF
))
167 // Create emergency spill slots.
168 for (unsigned i
= 0; i
< ScavSlotsNum
; ++i
) {
169 int FI
= MFI
.CreateStackObject(RI
->getSpillSize(RC
), RI
->getSpillAlign(RC
),
171 RS
->addScavengingFrameIndex(FI
);
172 if (IsLargeFunction
&& LAFI
->getBranchRelaxationSpillFrameIndex() == -1)
173 LAFI
->setBranchRelaxationSpillFrameIndex(FI
);
174 LLVM_DEBUG(dbgs() << "Allocated FI(" << FI
175 << ") as the emergency spill slot.\n");
179 void LoongArchFrameLowering::emitPrologue(MachineFunction
&MF
,
180 MachineBasicBlock
&MBB
) const {
181 MachineFrameInfo
&MFI
= MF
.getFrameInfo();
182 auto *LoongArchFI
= MF
.getInfo
<LoongArchMachineFunctionInfo
>();
183 const LoongArchRegisterInfo
*RI
= STI
.getRegisterInfo();
184 const LoongArchInstrInfo
*TII
= STI
.getInstrInfo();
185 MachineBasicBlock::iterator MBBI
= MBB
.begin();
186 bool IsLA64
= STI
.is64Bit();
188 Register SPReg
= LoongArch::R3
;
189 Register FPReg
= LoongArch::R22
;
191 // Debug location must be unknown since the first debug location is used
192 // to determine the end of the prologue.
194 // All calls are tail calls in GHC calling conv, and functions have no
195 // prologue/epilogue.
196 if (MF
.getFunction().getCallingConv() == CallingConv::GHC
)
198 // Determine the correct frame layout
199 determineFrameLayout(MF
);
201 // First, compute final stack size.
202 uint64_t StackSize
= MFI
.getStackSize();
203 uint64_t RealStackSize
= StackSize
;
205 // Early exit if there is no need to allocate space in the stack.
206 if (StackSize
== 0 && !MFI
.adjustsStack())
209 uint64_t FirstSPAdjustAmount
= getFirstSPAdjustAmount(MF
, true);
210 uint64_t SecondSPAdjustAmount
= RealStackSize
- FirstSPAdjustAmount
;
211 // Split the SP adjustment to reduce the offsets of callee saved spill.
212 if (FirstSPAdjustAmount
)
213 StackSize
= FirstSPAdjustAmount
;
216 adjustReg(MBB
, MBBI
, DL
, SPReg
, SPReg
, -StackSize
, MachineInstr::FrameSetup
);
217 if (FirstSPAdjustAmount
!= 2048 || SecondSPAdjustAmount
== 0) {
218 // Emit ".cfi_def_cfa_offset StackSize".
220 MF
.addFrameInst(MCCFIInstruction::cfiDefCfaOffset(nullptr, StackSize
));
221 BuildMI(MBB
, MBBI
, DL
, TII
->get(TargetOpcode::CFI_INSTRUCTION
))
222 .addCFIIndex(CFIIndex
)
223 .setMIFlag(MachineInstr::FrameSetup
);
226 const auto &CSI
= MFI
.getCalleeSavedInfo();
228 // The frame pointer is callee-saved, and code has been generated for us to
229 // save it to the stack. We need to skip over the storing of callee-saved
230 // registers as the frame pointer must be modified after it has been saved
231 // to the stack, not before.
232 std::advance(MBBI
, CSI
.size());
234 // Iterate over list of callee-saved registers and emit .cfi_offset
236 for (const auto &Entry
: CSI
) {
237 int64_t Offset
= MFI
.getObjectOffset(Entry
.getFrameIdx());
238 unsigned CFIIndex
= MF
.addFrameInst(MCCFIInstruction::createOffset(
239 nullptr, RI
->getDwarfRegNum(Entry
.getReg(), true), Offset
));
240 BuildMI(MBB
, MBBI
, DL
, TII
->get(TargetOpcode::CFI_INSTRUCTION
))
241 .addCFIIndex(CFIIndex
)
242 .setMIFlag(MachineInstr::FrameSetup
);
247 adjustReg(MBB
, MBBI
, DL
, FPReg
, SPReg
,
248 StackSize
- LoongArchFI
->getVarArgsSaveSize(),
249 MachineInstr::FrameSetup
);
251 // Emit ".cfi_def_cfa $fp, LoongArchFI->getVarArgsSaveSize()"
252 unsigned CFIIndex
= MF
.addFrameInst(
253 MCCFIInstruction::cfiDefCfa(nullptr, RI
->getDwarfRegNum(FPReg
, true),
254 LoongArchFI
->getVarArgsSaveSize()));
255 BuildMI(MBB
, MBBI
, DL
, TII
->get(TargetOpcode::CFI_INSTRUCTION
))
256 .addCFIIndex(CFIIndex
)
257 .setMIFlag(MachineInstr::FrameSetup
);
260 // Emit the second SP adjustment after saving callee saved registers.
261 if (FirstSPAdjustAmount
&& SecondSPAdjustAmount
) {
263 assert(SecondSPAdjustAmount
> 0 &&
264 "SecondSPAdjustAmount should be greater than zero");
265 adjustReg(MBB
, MBBI
, DL
, SPReg
, SPReg
, -SecondSPAdjustAmount
,
266 MachineInstr::FrameSetup
);
268 // FIXME: RegScavenger will place the spill instruction before the
269 // prologue if a VReg is created in the prologue. This will pollute the
270 // caller's stack data. Therefore, until there is better way, we just use
271 // the `addi.w/d` instruction for stack adjustment to ensure that VReg
272 // will not be created.
273 for (int Val
= SecondSPAdjustAmount
; Val
> 0; Val
-= 2048)
274 BuildMI(MBB
, MBBI
, DL
,
275 TII
->get(IsLA64
? LoongArch::ADDI_D
: LoongArch::ADDI_W
), SPReg
)
277 .addImm(Val
< 2048 ? -Val
: -2048)
278 .setMIFlag(MachineInstr::FrameSetup
);
280 // If we are using a frame-pointer, and thus emitted ".cfi_def_cfa fp, 0",
281 // don't emit an sp-based .cfi_def_cfa_offset
282 // Emit ".cfi_def_cfa_offset RealStackSize"
283 unsigned CFIIndex
= MF
.addFrameInst(
284 MCCFIInstruction::cfiDefCfaOffset(nullptr, RealStackSize
));
285 BuildMI(MBB
, MBBI
, DL
, TII
->get(TargetOpcode::CFI_INSTRUCTION
))
286 .addCFIIndex(CFIIndex
)
287 .setMIFlag(MachineInstr::FrameSetup
);
293 if (RI
->hasStackRealignment(MF
)) {
294 unsigned Align
= Log2(MFI
.getMaxAlign());
295 assert(Align
> 0 && "The stack realignment size is invalid!");
296 BuildMI(MBB
, MBBI
, DL
,
297 TII
->get(IsLA64
? LoongArch::BSTRINS_D
: LoongArch::BSTRINS_W
),
300 .addReg(LoongArch::R0
)
303 .setMIFlag(MachineInstr::FrameSetup
);
304 // FP will be used to restore the frame in the epilogue, so we need
305 // another base register BP to record SP after re-alignment. SP will
306 // track the current stack after allocating variable sized objects.
309 BuildMI(MBB
, MBBI
, DL
, TII
->get(LoongArch::OR
),
310 LoongArchABI::getBPReg())
312 .addReg(LoongArch::R0
)
313 .setMIFlag(MachineInstr::FrameSetup
);
319 void LoongArchFrameLowering::emitEpilogue(MachineFunction
&MF
,
320 MachineBasicBlock
&MBB
) const {
321 const LoongArchRegisterInfo
*RI
= STI
.getRegisterInfo();
322 MachineFrameInfo
&MFI
= MF
.getFrameInfo();
323 auto *LoongArchFI
= MF
.getInfo
<LoongArchMachineFunctionInfo
>();
324 Register SPReg
= LoongArch::R3
;
325 // All calls are tail calls in GHC calling conv, and functions have no
326 // prologue/epilogue.
327 if (MF
.getFunction().getCallingConv() == CallingConv::GHC
)
329 MachineBasicBlock::iterator MBBI
= MBB
.getFirstTerminator();
330 DebugLoc DL
= MBBI
!= MBB
.end() ? MBBI
->getDebugLoc() : DebugLoc();
332 const auto &CSI
= MFI
.getCalleeSavedInfo();
333 // Skip to before the restores of callee-saved registers.
334 auto LastFrameDestroy
= MBBI
;
336 LastFrameDestroy
= std::prev(MBBI
, CSI
.size());
338 // Get the number of bytes from FrameInfo.
339 uint64_t StackSize
= MFI
.getStackSize();
341 // Restore the stack pointer.
342 if (RI
->hasStackRealignment(MF
) || MFI
.hasVarSizedObjects()) {
343 assert(hasFP(MF
) && "frame pointer should not have been eliminated");
344 adjustReg(MBB
, LastFrameDestroy
, DL
, SPReg
, LoongArch::R22
,
345 -StackSize
+ LoongArchFI
->getVarArgsSaveSize(),
346 MachineInstr::FrameDestroy
);
349 uint64_t FirstSPAdjustAmount
= getFirstSPAdjustAmount(MF
);
350 if (FirstSPAdjustAmount
) {
351 uint64_t SecondSPAdjustAmount
= StackSize
- FirstSPAdjustAmount
;
352 assert(SecondSPAdjustAmount
> 0 &&
353 "SecondSPAdjustAmount should be greater than zero");
355 adjustReg(MBB
, LastFrameDestroy
, DL
, SPReg
, SPReg
, SecondSPAdjustAmount
,
356 MachineInstr::FrameDestroy
);
357 StackSize
= FirstSPAdjustAmount
;
361 adjustReg(MBB
, MBBI
, DL
, SPReg
, SPReg
, StackSize
, MachineInstr::FrameDestroy
);
364 // We would like to split the SP adjustment to reduce prologue/epilogue
365 // as following instructions. In this way, the offset of the callee saved
366 // register could fit in a single store.
368 // addi.d $sp, $sp, -2032
369 // st.d $ra, $sp, 2024
370 // st.d $fp, $sp, 2016
371 // addi.d $sp, $sp, -16
373 LoongArchFrameLowering::getFirstSPAdjustAmount(const MachineFunction
&MF
,
374 bool IsPrologue
) const {
375 const MachineFrameInfo
&MFI
= MF
.getFrameInfo();
376 const std::vector
<CalleeSavedInfo
> &CSI
= MFI
.getCalleeSavedInfo();
378 // Return the FirstSPAdjustAmount if the StackSize can not fit in a signed
379 // 12-bit and there exists a callee-saved register needing to be pushed.
380 if (!isInt
<12>(MFI
.getStackSize())) {
381 // FirstSPAdjustAmount is chosen as (2048 - StackAlign) because 2048 will
382 // cause sp = sp + 2048 in the epilogue to be split into multiple
383 // instructions. Offsets smaller than 2048 can fit in a single load/store
384 // instruction, and we have to stick with the stack alignment.
385 // So (2048 - StackAlign) will satisfy the stack alignment.
387 // FIXME: This place may seem odd. When using multiple ADDI instructions to
388 // adjust the stack in Prologue, and there are no callee-saved registers, we
389 // can take advantage of the logic of split sp ajustment to reduce code
391 return CSI
.size() > 0 ? 2048 - getStackAlign().value()
392 : (IsPrologue
? 2048 : 0);
397 void LoongArchFrameLowering::determineCalleeSaves(MachineFunction
&MF
,
398 BitVector
&SavedRegs
,
399 RegScavenger
*RS
) const {
400 TargetFrameLowering::determineCalleeSaves(MF
, SavedRegs
, RS
);
401 // Unconditionally spill RA and FP only if the function uses a frame
404 SavedRegs
.set(LoongArch::R1
);
405 SavedRegs
.set(LoongArch::R22
);
407 // Mark BP as used if function has dedicated base pointer.
409 SavedRegs
.set(LoongArchABI::getBPReg());
412 // Do not preserve stack space within prologue for outgoing variables if the
413 // function contains variable size objects.
414 // Let eliminateCallFramePseudoInstr preserve stack space for it.
415 bool LoongArchFrameLowering::hasReservedCallFrame(
416 const MachineFunction
&MF
) const {
417 return !MF
.getFrameInfo().hasVarSizedObjects();
420 // Eliminate ADJCALLSTACKDOWN, ADJCALLSTACKUP pseudo instructions.
421 MachineBasicBlock::iterator
422 LoongArchFrameLowering::eliminateCallFramePseudoInstr(
423 MachineFunction
&MF
, MachineBasicBlock
&MBB
,
424 MachineBasicBlock::iterator MI
) const {
425 Register SPReg
= LoongArch::R3
;
426 DebugLoc DL
= MI
->getDebugLoc();
428 if (!hasReservedCallFrame(MF
)) {
429 // If space has not been reserved for a call frame, ADJCALLSTACKDOWN and
430 // ADJCALLSTACKUP must be converted to instructions manipulating the stack
431 // pointer. This is necessary when there is a variable length stack
432 // allocation (e.g. alloca), which means it's not possible to allocate
433 // space for outgoing arguments from within the function prologue.
434 int64_t Amount
= MI
->getOperand(0).getImm();
437 // Ensure the stack remains aligned after adjustment.
438 Amount
= alignSPAdjust(Amount
);
440 if (MI
->getOpcode() == LoongArch::ADJCALLSTACKDOWN
)
443 adjustReg(MBB
, MI
, DL
, SPReg
, SPReg
, Amount
, MachineInstr::NoFlags
);
447 return MBB
.erase(MI
);
450 bool LoongArchFrameLowering::spillCalleeSavedRegisters(
451 MachineBasicBlock
&MBB
, MachineBasicBlock::iterator MI
,
452 ArrayRef
<CalleeSavedInfo
> CSI
, const TargetRegisterInfo
*TRI
) const {
456 MachineFunction
*MF
= MBB
.getParent();
457 const TargetInstrInfo
&TII
= *MF
->getSubtarget().getInstrInfo();
459 // Insert the spill to the stack frame.
460 for (auto &CS
: CSI
) {
461 Register Reg
= CS
.getReg();
462 // If the register is RA and the return address is taken by method
463 // LoongArchTargetLowering::lowerRETURNADDR, don't set kill flag.
465 !(Reg
== LoongArch::R1
&& MF
->getFrameInfo().isReturnAddressTaken());
466 const TargetRegisterClass
*RC
= TRI
->getMinimalPhysRegClass(Reg
);
467 TII
.storeRegToStackSlot(MBB
, MI
, Reg
, IsKill
, CS
.getFrameIdx(), RC
, TRI
,
474 StackOffset
LoongArchFrameLowering::getFrameIndexReference(
475 const MachineFunction
&MF
, int FI
, Register
&FrameReg
) const {
476 const MachineFrameInfo
&MFI
= MF
.getFrameInfo();
477 const TargetRegisterInfo
*RI
= MF
.getSubtarget().getRegisterInfo();
478 auto *LoongArchFI
= MF
.getInfo
<LoongArchMachineFunctionInfo
>();
479 uint64_t StackSize
= MFI
.getStackSize();
480 uint64_t FirstSPAdjustAmount
= getFirstSPAdjustAmount(MF
);
482 // Callee-saved registers should be referenced relative to the stack
483 // pointer (positive offset), otherwise use the frame pointer (negative
485 const auto &CSI
= MFI
.getCalleeSavedInfo();
489 StackOffset::getFixed(MFI
.getObjectOffset(FI
) - getOffsetOfLocalArea() +
490 MFI
.getOffsetAdjustment());
493 MinCSFI
= CSI
[0].getFrameIdx();
494 MaxCSFI
= CSI
[CSI
.size() - 1].getFrameIdx();
497 if (FI
>= MinCSFI
&& FI
<= MaxCSFI
) {
498 FrameReg
= LoongArch::R3
;
499 if (FirstSPAdjustAmount
)
500 Offset
+= StackOffset::getFixed(FirstSPAdjustAmount
);
502 Offset
+= StackOffset::getFixed(StackSize
);
503 } else if (RI
->hasStackRealignment(MF
) && !MFI
.isFixedObjectIndex(FI
)) {
504 // If the stack was realigned, the frame pointer is set in order to allow
505 // SP to be restored, so we need another base register to record the stack
506 // after realignment.
507 FrameReg
= hasBP(MF
) ? LoongArchABI::getBPReg() : LoongArch::R3
;
508 Offset
+= StackOffset::getFixed(StackSize
);
510 FrameReg
= RI
->getFrameRegister(MF
);
512 Offset
+= StackOffset::getFixed(LoongArchFI
->getVarArgsSaveSize());
514 Offset
+= StackOffset::getFixed(StackSize
);
520 bool LoongArchFrameLowering::enableShrinkWrapping(
521 const MachineFunction
&MF
) const {
522 // Keep the conventional code flow when not optimizing.
523 if (MF
.getFunction().hasOptNone())