1 | //===-- SystemZFrameLowering.cpp - Frame lowering for SystemZ -------------===// |
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 | #include "SystemZFrameLowering.h" |
10 | #include "SystemZCallingConv.h" |
11 | #include "SystemZInstrInfo.h" |
12 | #include "SystemZMachineFunctionInfo.h" |
13 | #include "SystemZRegisterInfo.h" |
14 | #include "SystemZSubtarget.h" |
15 | #include "llvm/CodeGen/LivePhysRegs.h" |
16 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
17 | #include "llvm/CodeGen/MachineModuleInfo.h" |
18 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
19 | #include "llvm/CodeGen/RegisterScavenging.h" |
20 | #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h" |
21 | #include "llvm/IR/CallingConv.h" |
22 | #include "llvm/IR/Function.h" |
23 | #include "llvm/IR/Module.h" |
24 | #include "llvm/Target/TargetMachine.h" |
25 | |
26 | using namespace llvm; |
27 | |
28 | namespace { |
29 | // The ABI-defined register save slots, relative to the CFA (i.e. |
30 | // incoming stack pointer + SystemZMC::ELFCallFrameSize). |
31 | static const TargetFrameLowering::SpillSlot ELFSpillOffsetTable[] = { |
32 | { .Reg: SystemZ::R2D, .Offset: 0x10 }, |
33 | { .Reg: SystemZ::R3D, .Offset: 0x18 }, |
34 | { .Reg: SystemZ::R4D, .Offset: 0x20 }, |
35 | { .Reg: SystemZ::R5D, .Offset: 0x28 }, |
36 | { .Reg: SystemZ::R6D, .Offset: 0x30 }, |
37 | { .Reg: SystemZ::R7D, .Offset: 0x38 }, |
38 | { .Reg: SystemZ::R8D, .Offset: 0x40 }, |
39 | { .Reg: SystemZ::R9D, .Offset: 0x48 }, |
40 | { .Reg: SystemZ::R10D, .Offset: 0x50 }, |
41 | { .Reg: SystemZ::R11D, .Offset: 0x58 }, |
42 | { .Reg: SystemZ::R12D, .Offset: 0x60 }, |
43 | { .Reg: SystemZ::R13D, .Offset: 0x68 }, |
44 | { .Reg: SystemZ::R14D, .Offset: 0x70 }, |
45 | { .Reg: SystemZ::R15D, .Offset: 0x78 }, |
46 | { .Reg: SystemZ::F0D, .Offset: 0x80 }, |
47 | { .Reg: SystemZ::F2D, .Offset: 0x88 }, |
48 | { .Reg: SystemZ::F4D, .Offset: 0x90 }, |
49 | { .Reg: SystemZ::F6D, .Offset: 0x98 } |
50 | }; |
51 | |
52 | static const TargetFrameLowering::SpillSlot XPLINKSpillOffsetTable[] = { |
53 | {.Reg: SystemZ::R4D, .Offset: 0x00}, {.Reg: SystemZ::R5D, .Offset: 0x08}, {.Reg: SystemZ::R6D, .Offset: 0x10}, |
54 | {.Reg: SystemZ::R7D, .Offset: 0x18}, {.Reg: SystemZ::R8D, .Offset: 0x20}, {.Reg: SystemZ::R9D, .Offset: 0x28}, |
55 | {.Reg: SystemZ::R10D, .Offset: 0x30}, {.Reg: SystemZ::R11D, .Offset: 0x38}, {.Reg: SystemZ::R12D, .Offset: 0x40}, |
56 | {.Reg: SystemZ::R13D, .Offset: 0x48}, {.Reg: SystemZ::R14D, .Offset: 0x50}, {.Reg: SystemZ::R15D, .Offset: 0x58}}; |
57 | } // end anonymous namespace |
58 | |
59 | SystemZFrameLowering::SystemZFrameLowering(StackDirection D, Align StackAl, |
60 | int LAO, Align TransAl, |
61 | bool StackReal, unsigned PointerSize) |
62 | : TargetFrameLowering(D, StackAl, LAO, TransAl, StackReal), |
63 | PointerSize(PointerSize) {} |
64 | |
65 | std::unique_ptr<SystemZFrameLowering> |
66 | SystemZFrameLowering::create(const SystemZSubtarget &STI) { |
67 | unsigned PtrSz = |
68 | STI.getTargetLowering()->getTargetMachine().getPointerSize(AS: 0); |
69 | if (STI.isTargetXPLINK64()) |
70 | return std::make_unique<SystemZXPLINKFrameLowering>(args&: PtrSz); |
71 | return std::make_unique<SystemZELFFrameLowering>(args&: PtrSz); |
72 | } |
73 | |
74 | namespace { |
75 | struct SZFrameSortingObj { |
76 | bool IsValid = false; // True if we care about this Object. |
77 | uint32_t ObjectIndex = 0; // Index of Object into MFI list. |
78 | uint64_t ObjectSize = 0; // Size of Object in bytes. |
79 | uint32_t D12Count = 0; // 12-bit displacement only. |
80 | uint32_t DPairCount = 0; // 12 or 20 bit displacement. |
81 | }; |
82 | typedef std::vector<SZFrameSortingObj> SZFrameObjVec; |
83 | } // namespace |
84 | |
85 | // TODO: Move to base class. |
86 | void SystemZELFFrameLowering::orderFrameObjects( |
87 | const MachineFunction &MF, SmallVectorImpl<int> &ObjectsToAllocate) const { |
88 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
89 | auto *TII = MF.getSubtarget<SystemZSubtarget>().getInstrInfo(); |
90 | |
91 | // Make a vector of sorting objects to track all MFI objects and mark those |
92 | // to be sorted as valid. |
93 | if (ObjectsToAllocate.size() <= 1) |
94 | return; |
95 | SZFrameObjVec SortingObjects(MFI.getObjectIndexEnd()); |
96 | for (auto &Obj : ObjectsToAllocate) { |
97 | SortingObjects[Obj].IsValid = true; |
98 | SortingObjects[Obj].ObjectIndex = Obj; |
99 | SortingObjects[Obj].ObjectSize = MFI.getObjectSize(ObjectIdx: Obj); |
100 | } |
101 | |
102 | // Examine uses for each object and record short (12-bit) and "pair" |
103 | // displacement types. |
104 | for (auto &MBB : MF) |
105 | for (auto &MI : MBB) { |
106 | if (MI.isDebugInstr()) |
107 | continue; |
108 | for (unsigned I = 0, E = MI.getNumOperands(); I != E; ++I) { |
109 | const MachineOperand &MO = MI.getOperand(i: I); |
110 | if (!MO.isFI()) |
111 | continue; |
112 | int Index = MO.getIndex(); |
113 | if (Index >= 0 && Index < MFI.getObjectIndexEnd() && |
114 | SortingObjects[Index].IsValid) { |
115 | if (TII->hasDisplacementPairInsn(Opcode: MI.getOpcode())) |
116 | SortingObjects[Index].DPairCount++; |
117 | else if (!(MI.getDesc().TSFlags & SystemZII::Has20BitOffset)) |
118 | SortingObjects[Index].D12Count++; |
119 | } |
120 | } |
121 | } |
122 | |
123 | // Sort all objects for short/paired displacements, which should be |
124 | // sufficient as it seems like all frame objects typically are within the |
125 | // long displacement range. Sorting works by computing the "density" as |
126 | // Count / ObjectSize. The comparisons of two such fractions are refactored |
127 | // by multiplying both sides with A.ObjectSize * B.ObjectSize, in order to |
128 | // eliminate the (fp) divisions. A higher density object needs to go after |
129 | // in the list in order for it to end up lower on the stack. |
130 | auto CmpD12 = [](const SZFrameSortingObj &A, const SZFrameSortingObj &B) { |
131 | // Put all invalid and variable sized objects at the end. |
132 | if (!A.IsValid || !B.IsValid) |
133 | return A.IsValid; |
134 | if (!A.ObjectSize || !B.ObjectSize) |
135 | return A.ObjectSize > 0; |
136 | uint64_t ADensityCmp = A.D12Count * B.ObjectSize; |
137 | uint64_t BDensityCmp = B.D12Count * A.ObjectSize; |
138 | if (ADensityCmp != BDensityCmp) |
139 | return ADensityCmp < BDensityCmp; |
140 | return A.DPairCount * B.ObjectSize < B.DPairCount * A.ObjectSize; |
141 | }; |
142 | llvm::stable_sort(Range&: SortingObjects, C: CmpD12); |
143 | |
144 | // Now modify the original list to represent the final order that |
145 | // we want. |
146 | unsigned Idx = 0; |
147 | for (auto &Obj : SortingObjects) { |
148 | // All invalid items are sorted at the end, so it's safe to stop. |
149 | if (!Obj.IsValid) |
150 | break; |
151 | ObjectsToAllocate[Idx++] = Obj.ObjectIndex; |
152 | } |
153 | } |
154 | |
155 | bool SystemZFrameLowering::hasReservedCallFrame( |
156 | const MachineFunction &MF) const { |
157 | // The ELF ABI requires us to allocate 160 bytes of stack space for the |
158 | // callee, with any outgoing stack arguments being placed above that. It |
159 | // seems better to make that area a permanent feature of the frame even if |
160 | // we're using a frame pointer. Similarly, 64-bit XPLINK requires 96 bytes |
161 | // of stack space for the register save area. |
162 | return true; |
163 | } |
164 | |
165 | bool SystemZELFFrameLowering::assignCalleeSavedSpillSlots( |
166 | MachineFunction &MF, const TargetRegisterInfo *TRI, |
167 | std::vector<CalleeSavedInfo> &CSI) const { |
168 | SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); |
169 | MachineFrameInfo &MFFrame = MF.getFrameInfo(); |
170 | bool IsVarArg = MF.getFunction().isVarArg(); |
171 | if (CSI.empty()) |
172 | return true; // Early exit if no callee saved registers are modified! |
173 | |
174 | unsigned LowGPR = 0; |
175 | unsigned HighGPR = SystemZ::R15D; |
176 | int StartSPOffset = SystemZMC::ELFCallFrameSize; |
177 | for (auto &CS : CSI) { |
178 | MCRegister Reg = CS.getReg(); |
179 | int Offset = getRegSpillOffset(MF, Reg); |
180 | if (Offset) { |
181 | if (SystemZ::GR64BitRegClass.contains(Reg) && StartSPOffset > Offset) { |
182 | LowGPR = Reg; |
183 | StartSPOffset = Offset; |
184 | } |
185 | Offset -= SystemZMC::ELFCallFrameSize; |
186 | int FrameIdx = |
187 | MFFrame.CreateFixedSpillStackObject(Size: getPointerSize(), SPOffset: Offset); |
188 | CS.setFrameIdx(FrameIdx); |
189 | } else |
190 | CS.setFrameIdx(INT32_MAX); |
191 | } |
192 | |
193 | // Save the range of call-saved registers, for use by the |
194 | // prologue/epilogue inserters. |
195 | ZFI->setRestoreGPRRegs(Low: LowGPR, High: HighGPR, Offs: StartSPOffset); |
196 | if (IsVarArg) { |
197 | // Also save the GPR varargs, if any. R6D is call-saved, so would |
198 | // already be included, but we also need to handle the call-clobbered |
199 | // argument registers. |
200 | Register FirstGPR = ZFI->getVarArgsFirstGPR(); |
201 | if (FirstGPR < SystemZ::ELFNumArgGPRs) { |
202 | unsigned Reg = SystemZ::ELFArgGPRs[FirstGPR]; |
203 | int Offset = getRegSpillOffset(MF, Reg); |
204 | if (StartSPOffset > Offset) { |
205 | LowGPR = Reg; StartSPOffset = Offset; |
206 | } |
207 | } |
208 | } |
209 | ZFI->setSpillGPRRegs(Low: LowGPR, High: HighGPR, Offs: StartSPOffset); |
210 | |
211 | // Create fixed stack objects for the remaining registers. |
212 | int CurrOffset = -SystemZMC::ELFCallFrameSize; |
213 | if (usePackedStack(MF)) |
214 | CurrOffset += StartSPOffset; |
215 | |
216 | for (auto &CS : CSI) { |
217 | if (CS.getFrameIdx() != INT32_MAX) |
218 | continue; |
219 | MCRegister Reg = CS.getReg(); |
220 | const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); |
221 | unsigned Size = TRI->getSpillSize(RC: *RC); |
222 | CurrOffset -= Size; |
223 | assert(CurrOffset % 8 == 0 && |
224 | "8-byte alignment required for for all register save slots" ); |
225 | int FrameIdx = MFFrame.CreateFixedSpillStackObject(Size, SPOffset: CurrOffset); |
226 | CS.setFrameIdx(FrameIdx); |
227 | } |
228 | |
229 | return true; |
230 | } |
231 | |
232 | void SystemZELFFrameLowering::determineCalleeSaves(MachineFunction &MF, |
233 | BitVector &SavedRegs, |
234 | RegScavenger *RS) const { |
235 | TargetFrameLowering::determineCalleeSaves(MF, SavedRegs, RS); |
236 | |
237 | MachineFrameInfo &MFFrame = MF.getFrameInfo(); |
238 | const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); |
239 | bool HasFP = hasFP(MF); |
240 | SystemZMachineFunctionInfo *MFI = MF.getInfo<SystemZMachineFunctionInfo>(); |
241 | bool IsVarArg = MF.getFunction().isVarArg(); |
242 | |
243 | // va_start stores incoming FPR varargs in the normal way, but delegates |
244 | // the saving of incoming GPR varargs to spillCalleeSavedRegisters(). |
245 | // Record these pending uses, which typically include the call-saved |
246 | // argument register R6D. |
247 | if (IsVarArg) |
248 | for (unsigned I = MFI->getVarArgsFirstGPR(); I < SystemZ::ELFNumArgGPRs; ++I) |
249 | SavedRegs.set(SystemZ::ELFArgGPRs[I]); |
250 | |
251 | // If there are any landing pads, entering them will modify r6/r7. |
252 | if (!MF.getLandingPads().empty()) { |
253 | SavedRegs.set(SystemZ::R6D); |
254 | SavedRegs.set(SystemZ::R7D); |
255 | } |
256 | |
257 | // If the function requires a frame pointer, record that the hard |
258 | // frame pointer will be clobbered. |
259 | if (HasFP) |
260 | SavedRegs.set(SystemZ::R11D); |
261 | |
262 | // If the function calls other functions, record that the return |
263 | // address register will be clobbered. |
264 | if (MFFrame.hasCalls()) |
265 | SavedRegs.set(SystemZ::R14D); |
266 | |
267 | // If we are saving GPRs other than the stack pointer, we might as well |
268 | // save and restore the stack pointer at the same time, via STMG and LMG. |
269 | // This allows the deallocation to be done by the LMG, rather than needing |
270 | // a separate %r15 addition. |
271 | const MCPhysReg *CSRegs = TRI->getCalleeSavedRegs(MF: &MF); |
272 | for (unsigned I = 0; CSRegs[I]; ++I) { |
273 | unsigned Reg = CSRegs[I]; |
274 | if (SystemZ::GR64BitRegClass.contains(Reg) && SavedRegs.test(Idx: Reg)) { |
275 | SavedRegs.set(SystemZ::R15D); |
276 | break; |
277 | } |
278 | } |
279 | } |
280 | |
281 | SystemZELFFrameLowering::SystemZELFFrameLowering(unsigned PointerSize) |
282 | : SystemZFrameLowering(TargetFrameLowering::StackGrowsDown, Align(8), 0, |
283 | Align(8), /* StackRealignable */ false, PointerSize), |
284 | RegSpillOffsets(0) { |
285 | |
286 | // Due to the SystemZ ABI, the DWARF CFA (Canonical Frame Address) is not |
287 | // equal to the incoming stack pointer, but to incoming stack pointer plus |
288 | // 160. Instead of using a Local Area Offset, the Register save area will |
289 | // be occupied by fixed frame objects, and all offsets are actually |
290 | // relative to CFA. |
291 | |
292 | // Create a mapping from register number to save slot offset. |
293 | // These offsets are relative to the start of the register save area. |
294 | RegSpillOffsets.grow(n: SystemZ::NUM_TARGET_REGS); |
295 | for (const auto &Entry : ELFSpillOffsetTable) |
296 | RegSpillOffsets[Entry.Reg] = Entry.Offset; |
297 | } |
298 | |
299 | // Add GPR64 to the save instruction being built by MIB, which is in basic |
300 | // block MBB. IsImplicit says whether this is an explicit operand to the |
301 | // instruction, or an implicit one that comes between the explicit start |
302 | // and end registers. |
303 | static void addSavedGPR(MachineBasicBlock &MBB, MachineInstrBuilder &MIB, |
304 | unsigned GPR64, bool IsImplicit) { |
305 | const TargetRegisterInfo *RI = |
306 | MBB.getParent()->getSubtarget().getRegisterInfo(); |
307 | Register GPR32 = RI->getSubReg(Reg: GPR64, Idx: SystemZ::subreg_l32); |
308 | bool IsLive = MBB.isLiveIn(Reg: GPR64) || MBB.isLiveIn(Reg: GPR32); |
309 | if (!IsLive || !IsImplicit) { |
310 | MIB.addReg(RegNo: GPR64, flags: getImplRegState(B: IsImplicit) | getKillRegState(B: !IsLive)); |
311 | if (!IsLive) |
312 | MBB.addLiveIn(PhysReg: GPR64); |
313 | } |
314 | } |
315 | |
316 | bool SystemZELFFrameLowering::spillCalleeSavedRegisters( |
317 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, |
318 | ArrayRef<CalleeSavedInfo> CSI, const TargetRegisterInfo *TRI) const { |
319 | if (CSI.empty()) |
320 | return false; |
321 | |
322 | MachineFunction &MF = *MBB.getParent(); |
323 | const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); |
324 | SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); |
325 | bool IsVarArg = MF.getFunction().isVarArg(); |
326 | DebugLoc DL; |
327 | |
328 | // Save GPRs |
329 | SystemZ::GPRRegs SpillGPRs = ZFI->getSpillGPRRegs(); |
330 | if (SpillGPRs.LowGPR) { |
331 | assert(SpillGPRs.LowGPR != SpillGPRs.HighGPR && |
332 | "Should be saving %r15 and something else" ); |
333 | |
334 | // Build an STMG instruction. |
335 | MachineInstrBuilder MIB = BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: SystemZ::STMG)); |
336 | |
337 | // Add the explicit register operands. |
338 | addSavedGPR(MBB, MIB, GPR64: SpillGPRs.LowGPR, IsImplicit: false); |
339 | addSavedGPR(MBB, MIB, GPR64: SpillGPRs.HighGPR, IsImplicit: false); |
340 | |
341 | // Add the address. |
342 | MIB.addReg(RegNo: SystemZ::R15D).addImm(Val: SpillGPRs.GPROffset); |
343 | |
344 | // Make sure all call-saved GPRs are included as operands and are |
345 | // marked as live on entry. |
346 | for (const CalleeSavedInfo &I : CSI) { |
347 | MCRegister Reg = I.getReg(); |
348 | if (SystemZ::GR64BitRegClass.contains(Reg)) |
349 | addSavedGPR(MBB, MIB, GPR64: Reg, IsImplicit: true); |
350 | } |
351 | |
352 | // ...likewise GPR varargs. |
353 | if (IsVarArg) |
354 | for (unsigned I = ZFI->getVarArgsFirstGPR(); I < SystemZ::ELFNumArgGPRs; ++I) |
355 | addSavedGPR(MBB, MIB, GPR64: SystemZ::ELFArgGPRs[I], IsImplicit: true); |
356 | } |
357 | |
358 | // Save FPRs/VRs in the normal TargetInstrInfo way. |
359 | for (const CalleeSavedInfo &I : CSI) { |
360 | MCRegister Reg = I.getReg(); |
361 | if (SystemZ::FP64BitRegClass.contains(Reg)) { |
362 | MBB.addLiveIn(PhysReg: Reg); |
363 | TII->storeRegToStackSlot(MBB, MI: MBBI, SrcReg: Reg, isKill: true, FrameIndex: I.getFrameIdx(), |
364 | RC: &SystemZ::FP64BitRegClass, TRI, VReg: Register()); |
365 | } |
366 | if (SystemZ::VR128BitRegClass.contains(Reg)) { |
367 | MBB.addLiveIn(PhysReg: Reg); |
368 | TII->storeRegToStackSlot(MBB, MI: MBBI, SrcReg: Reg, isKill: true, FrameIndex: I.getFrameIdx(), |
369 | RC: &SystemZ::VR128BitRegClass, TRI, VReg: Register()); |
370 | } |
371 | } |
372 | |
373 | return true; |
374 | } |
375 | |
376 | bool SystemZELFFrameLowering::restoreCalleeSavedRegisters( |
377 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, |
378 | MutableArrayRef<CalleeSavedInfo> CSI, const TargetRegisterInfo *TRI) const { |
379 | if (CSI.empty()) |
380 | return false; |
381 | |
382 | MachineFunction &MF = *MBB.getParent(); |
383 | const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); |
384 | SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); |
385 | bool HasFP = hasFP(MF); |
386 | DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); |
387 | |
388 | // Restore FPRs/VRs in the normal TargetInstrInfo way. |
389 | for (const CalleeSavedInfo &I : CSI) { |
390 | MCRegister Reg = I.getReg(); |
391 | if (SystemZ::FP64BitRegClass.contains(Reg)) |
392 | TII->loadRegFromStackSlot(MBB, MI: MBBI, DestReg: Reg, FrameIndex: I.getFrameIdx(), |
393 | RC: &SystemZ::FP64BitRegClass, TRI, VReg: Register()); |
394 | if (SystemZ::VR128BitRegClass.contains(Reg)) |
395 | TII->loadRegFromStackSlot(MBB, MI: MBBI, DestReg: Reg, FrameIndex: I.getFrameIdx(), |
396 | RC: &SystemZ::VR128BitRegClass, TRI, VReg: Register()); |
397 | } |
398 | |
399 | // Restore call-saved GPRs (but not call-clobbered varargs, which at |
400 | // this point might hold return values). |
401 | SystemZ::GPRRegs RestoreGPRs = ZFI->getRestoreGPRRegs(); |
402 | if (RestoreGPRs.LowGPR) { |
403 | // If we saved any of %r2-%r5 as varargs, we should also be saving |
404 | // and restoring %r6. If we're saving %r6 or above, we should be |
405 | // restoring it too. |
406 | assert(RestoreGPRs.LowGPR != RestoreGPRs.HighGPR && |
407 | "Should be loading %r15 and something else" ); |
408 | |
409 | // Build an LMG instruction. |
410 | MachineInstrBuilder MIB = BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: SystemZ::LMG)); |
411 | |
412 | // Add the explicit register operands. |
413 | MIB.addReg(RegNo: RestoreGPRs.LowGPR, flags: RegState::Define); |
414 | MIB.addReg(RegNo: RestoreGPRs.HighGPR, flags: RegState::Define); |
415 | |
416 | // Add the address. |
417 | MIB.addReg(RegNo: HasFP ? SystemZ::R11D : SystemZ::R15D); |
418 | MIB.addImm(Val: RestoreGPRs.GPROffset); |
419 | |
420 | // Do a second scan adding regs as being defined by instruction |
421 | for (const CalleeSavedInfo &I : CSI) { |
422 | MCRegister Reg = I.getReg(); |
423 | if (Reg != RestoreGPRs.LowGPR && Reg != RestoreGPRs.HighGPR && |
424 | SystemZ::GR64BitRegClass.contains(Reg)) |
425 | MIB.addReg(RegNo: Reg, flags: RegState::ImplicitDefine); |
426 | } |
427 | } |
428 | |
429 | return true; |
430 | } |
431 | |
432 | void SystemZELFFrameLowering::processFunctionBeforeFrameFinalized( |
433 | MachineFunction &MF, RegScavenger *RS) const { |
434 | MachineFrameInfo &MFFrame = MF.getFrameInfo(); |
435 | SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); |
436 | MachineRegisterInfo *MRI = &MF.getRegInfo(); |
437 | bool BackChain = MF.getSubtarget<SystemZSubtarget>().hasBackChain(); |
438 | |
439 | if (!usePackedStack(MF) || BackChain) |
440 | // Create the incoming register save area. |
441 | getOrCreateFramePointerSaveIndex(MF); |
442 | |
443 | // Get the size of our stack frame to be allocated ... |
444 | uint64_t StackSize = (MFFrame.estimateStackSize(MF) + |
445 | SystemZMC::ELFCallFrameSize); |
446 | // ... and the maximum offset we may need to reach into the |
447 | // caller's frame to access the save area or stack arguments. |
448 | int64_t MaxArgOffset = 0; |
449 | for (int I = MFFrame.getObjectIndexBegin(); I != 0; ++I) |
450 | if (MFFrame.getObjectOffset(ObjectIdx: I) >= 0) { |
451 | int64_t ArgOffset = MFFrame.getObjectOffset(ObjectIdx: I) + |
452 | MFFrame.getObjectSize(ObjectIdx: I); |
453 | MaxArgOffset = std::max(a: MaxArgOffset, b: ArgOffset); |
454 | } |
455 | |
456 | uint64_t MaxReach = StackSize + MaxArgOffset; |
457 | if (!isUInt<12>(x: MaxReach)) { |
458 | // We may need register scavenging slots if some parts of the frame |
459 | // are outside the reach of an unsigned 12-bit displacement. |
460 | // Create 2 for the case where both addresses in an MVC are |
461 | // out of range. |
462 | RS->addScavengingFrameIndex( |
463 | FI: MFFrame.CreateSpillStackObject(Size: getPointerSize(), Alignment: Align(8))); |
464 | RS->addScavengingFrameIndex( |
465 | FI: MFFrame.CreateSpillStackObject(Size: getPointerSize(), Alignment: Align(8))); |
466 | } |
467 | |
468 | // If R6 is used as an argument register it is still callee saved. If it in |
469 | // this case is not clobbered (and restored) it should never be marked as |
470 | // killed. |
471 | if (MF.front().isLiveIn(Reg: SystemZ::R6D) && |
472 | ZFI->getRestoreGPRRegs().LowGPR != SystemZ::R6D) |
473 | for (auto &MO : MRI->use_nodbg_operands(Reg: SystemZ::R6D)) |
474 | MO.setIsKill(false); |
475 | } |
476 | |
477 | // Emit instructions before MBBI (in MBB) to add NumBytes to Reg. |
478 | static void emitIncrement(MachineBasicBlock &MBB, |
479 | MachineBasicBlock::iterator &MBBI, const DebugLoc &DL, |
480 | Register Reg, int64_t NumBytes, |
481 | const TargetInstrInfo *TII) { |
482 | while (NumBytes) { |
483 | unsigned Opcode; |
484 | int64_t ThisVal = NumBytes; |
485 | if (isInt<16>(x: NumBytes)) |
486 | Opcode = SystemZ::AGHI; |
487 | else { |
488 | Opcode = SystemZ::AGFI; |
489 | // Make sure we maintain 8-byte stack alignment. |
490 | int64_t MinVal = -uint64_t(1) << 31; |
491 | int64_t MaxVal = (int64_t(1) << 31) - 8; |
492 | if (ThisVal < MinVal) |
493 | ThisVal = MinVal; |
494 | else if (ThisVal > MaxVal) |
495 | ThisVal = MaxVal; |
496 | } |
497 | MachineInstr *MI = BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode), DestReg: Reg) |
498 | .addReg(RegNo: Reg).addImm(Val: ThisVal); |
499 | // The CC implicit def is dead. |
500 | MI->getOperand(i: 3).setIsDead(); |
501 | NumBytes -= ThisVal; |
502 | } |
503 | } |
504 | |
505 | // Add CFI for the new CFA offset. |
506 | static void buildCFAOffs(MachineBasicBlock &MBB, |
507 | MachineBasicBlock::iterator MBBI, |
508 | const DebugLoc &DL, int Offset, |
509 | const SystemZInstrInfo *ZII) { |
510 | unsigned CFIIndex = MBB.getParent()->addFrameInst( |
511 | Inst: MCCFIInstruction::cfiDefCfaOffset(L: nullptr, Offset: -Offset)); |
512 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: TargetOpcode::CFI_INSTRUCTION)) |
513 | .addCFIIndex(CFIIndex); |
514 | } |
515 | |
516 | // Add CFI for the new frame location. |
517 | static void buildDefCFAReg(MachineBasicBlock &MBB, |
518 | MachineBasicBlock::iterator MBBI, |
519 | const DebugLoc &DL, unsigned Reg, |
520 | const SystemZInstrInfo *ZII) { |
521 | MachineFunction &MF = *MBB.getParent(); |
522 | const MCRegisterInfo *MRI = MF.getContext().getRegisterInfo(); |
523 | unsigned RegNum = MRI->getDwarfRegNum(RegNum: Reg, isEH: true); |
524 | unsigned CFIIndex = MF.addFrameInst( |
525 | Inst: MCCFIInstruction::createDefCfaRegister(L: nullptr, Register: RegNum)); |
526 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: TargetOpcode::CFI_INSTRUCTION)) |
527 | .addCFIIndex(CFIIndex); |
528 | } |
529 | |
530 | void SystemZELFFrameLowering::emitPrologue(MachineFunction &MF, |
531 | MachineBasicBlock &MBB) const { |
532 | assert(&MF.front() == &MBB && "Shrink-wrapping not yet supported" ); |
533 | const SystemZSubtarget &STI = MF.getSubtarget<SystemZSubtarget>(); |
534 | const SystemZTargetLowering &TLI = *STI.getTargetLowering(); |
535 | MachineFrameInfo &MFFrame = MF.getFrameInfo(); |
536 | auto *ZII = static_cast<const SystemZInstrInfo *>(STI.getInstrInfo()); |
537 | SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); |
538 | MachineBasicBlock::iterator MBBI = MBB.begin(); |
539 | const MCRegisterInfo *MRI = MF.getContext().getRegisterInfo(); |
540 | const std::vector<CalleeSavedInfo> &CSI = MFFrame.getCalleeSavedInfo(); |
541 | bool HasFP = hasFP(MF); |
542 | |
543 | // In GHC calling convention C stack space, including the ABI-defined |
544 | // 160-byte base area, is (de)allocated by GHC itself. This stack space may |
545 | // be used by LLVM as spill slots for the tail recursive GHC functions. Thus |
546 | // do not allocate stack space here, too. |
547 | if (MF.getFunction().getCallingConv() == CallingConv::GHC) { |
548 | if (MFFrame.getStackSize() > 2048 * sizeof(long)) { |
549 | report_fatal_error( |
550 | reason: "Pre allocated stack space for GHC function is too small" ); |
551 | } |
552 | if (HasFP) { |
553 | report_fatal_error( |
554 | reason: "In GHC calling convention a frame pointer is not supported" ); |
555 | } |
556 | MFFrame.setStackSize(MFFrame.getStackSize() + SystemZMC::ELFCallFrameSize); |
557 | return; |
558 | } |
559 | |
560 | // Debug location must be unknown since the first debug location is used |
561 | // to determine the end of the prologue. |
562 | DebugLoc DL; |
563 | // Add mcount instrumentation if necessary. |
564 | if (MF.getFunction() |
565 | .getFnAttribute(Kind: "systemz-instrument-function-entry" ) |
566 | .getValueAsString() == "mcount" ) { |
567 | |
568 | // Store return address 8 bytes above stack pointer. |
569 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::STG)) |
570 | .addReg(RegNo: SystemZ::R14D) |
571 | .addReg(RegNo: SystemZ::R15D) |
572 | .addImm(Val: 8) |
573 | .addReg(RegNo: 0); |
574 | |
575 | // Call mcount (Regmask from CC AnyReg since mcount preserves all normal |
576 | // argument registers). |
577 | FunctionCallee FC = MF.getFunction().getParent()->getOrInsertFunction( |
578 | Name: "mcount" , RetTy: Type::getVoidTy(C&: MF.getFunction().getContext())); |
579 | const uint32_t *Mask = MF.getSubtarget<SystemZSubtarget>() |
580 | .getSpecialRegisters() |
581 | ->getCallPreservedMask(MF, CC: CallingConv::AnyReg); |
582 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::CallBRASL)) |
583 | .addGlobalAddress(GV: dyn_cast<Function>(Val: FC.getCallee())) |
584 | .addRegMask(Mask); |
585 | |
586 | // Reload return address from 8 bytes above stack pointer. |
587 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::LG)) |
588 | .addReg(RegNo: SystemZ::R14D, flags: RegState::Define) |
589 | .addReg(RegNo: SystemZ::R15D) |
590 | .addImm(Val: 8) |
591 | .addReg(RegNo: 0); |
592 | } |
593 | |
594 | // The current offset of the stack pointer from the CFA. |
595 | int64_t SPOffsetFromCFA = -SystemZMC::ELFCFAOffsetFromInitialSP; |
596 | |
597 | if (ZFI->getSpillGPRRegs().LowGPR) { |
598 | // Skip over the GPR saves. |
599 | if (MBBI != MBB.end() && MBBI->getOpcode() == SystemZ::STMG) |
600 | ++MBBI; |
601 | else |
602 | llvm_unreachable("Couldn't skip over GPR saves" ); |
603 | |
604 | // Add CFI for the GPR saves. |
605 | for (auto &Save : CSI) { |
606 | MCRegister Reg = Save.getReg(); |
607 | if (SystemZ::GR64BitRegClass.contains(Reg)) { |
608 | int FI = Save.getFrameIdx(); |
609 | int64_t Offset = MFFrame.getObjectOffset(ObjectIdx: FI); |
610 | unsigned CFIIndex = MF.addFrameInst(Inst: MCCFIInstruction::createOffset( |
611 | L: nullptr, Register: MRI->getDwarfRegNum(RegNum: Reg, isEH: true), Offset)); |
612 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: TargetOpcode::CFI_INSTRUCTION)) |
613 | .addCFIIndex(CFIIndex); |
614 | } |
615 | } |
616 | } |
617 | |
618 | uint64_t StackSize = MFFrame.getStackSize(); |
619 | // We need to allocate the ABI-defined 160-byte base area whenever |
620 | // we allocate stack space for our own use and whenever we call another |
621 | // function. |
622 | bool HasStackObject = false; |
623 | for (unsigned i = 0, e = MFFrame.getObjectIndexEnd(); i != e; ++i) |
624 | if (!MFFrame.isDeadObjectIndex(ObjectIdx: i)) { |
625 | HasStackObject = true; |
626 | break; |
627 | } |
628 | if (HasStackObject || MFFrame.hasCalls()) |
629 | StackSize += SystemZMC::ELFCallFrameSize; |
630 | // Don't allocate the incoming reg save area. |
631 | StackSize = StackSize > SystemZMC::ELFCallFrameSize |
632 | ? StackSize - SystemZMC::ELFCallFrameSize |
633 | : 0; |
634 | MFFrame.setStackSize(StackSize); |
635 | |
636 | if (StackSize) { |
637 | // Allocate StackSize bytes. |
638 | int64_t Delta = -int64_t(StackSize); |
639 | const unsigned ProbeSize = TLI.getStackProbeSize(MF); |
640 | bool FreeProbe = (ZFI->getSpillGPRRegs().GPROffset && |
641 | (ZFI->getSpillGPRRegs().GPROffset + StackSize) < ProbeSize); |
642 | if (!FreeProbe && |
643 | MF.getSubtarget().getTargetLowering()->hasInlineStackProbe(MF)) { |
644 | // Stack probing may involve looping, but splitting the prologue block |
645 | // is not possible at this point since it would invalidate the |
646 | // SaveBlocks / RestoreBlocks sets of PEI in the single block function |
647 | // case. Build a pseudo to be handled later by inlineStackProbe(). |
648 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::PROBED_STACKALLOC)) |
649 | .addImm(Val: StackSize); |
650 | } |
651 | else { |
652 | bool StoreBackchain = MF.getSubtarget<SystemZSubtarget>().hasBackChain(); |
653 | // If we need backchain, save current stack pointer. R1 is free at |
654 | // this point. |
655 | if (StoreBackchain) |
656 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::LGR)) |
657 | .addReg(RegNo: SystemZ::R1D, flags: RegState::Define).addReg(RegNo: SystemZ::R15D); |
658 | emitIncrement(MBB, MBBI, DL, Reg: SystemZ::R15D, NumBytes: Delta, TII: ZII); |
659 | buildCFAOffs(MBB, MBBI, DL, Offset: SPOffsetFromCFA + Delta, ZII); |
660 | if (StoreBackchain) |
661 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::STG)) |
662 | .addReg(RegNo: SystemZ::R1D, flags: RegState::Kill).addReg(RegNo: SystemZ::R15D) |
663 | .addImm(Val: getBackchainOffset(MF)).addReg(RegNo: 0); |
664 | } |
665 | SPOffsetFromCFA += Delta; |
666 | } |
667 | |
668 | if (HasFP) { |
669 | // Copy the base of the frame to R11. |
670 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::LGR), DestReg: SystemZ::R11D) |
671 | .addReg(RegNo: SystemZ::R15D); |
672 | |
673 | // Add CFI for the new frame location. |
674 | buildDefCFAReg(MBB, MBBI, DL, Reg: SystemZ::R11D, ZII); |
675 | |
676 | // Mark the FramePtr as live at the beginning of every block except |
677 | // the entry block. (We'll have marked R11 as live on entry when |
678 | // saving the GPRs.) |
679 | for (MachineBasicBlock &MBBJ : llvm::drop_begin(RangeOrContainer&: MF)) |
680 | MBBJ.addLiveIn(PhysReg: SystemZ::R11D); |
681 | } |
682 | |
683 | // Skip over the FPR/VR saves. |
684 | SmallVector<unsigned, 8> CFIIndexes; |
685 | for (auto &Save : CSI) { |
686 | MCRegister Reg = Save.getReg(); |
687 | if (SystemZ::FP64BitRegClass.contains(Reg)) { |
688 | if (MBBI != MBB.end() && |
689 | (MBBI->getOpcode() == SystemZ::STD || |
690 | MBBI->getOpcode() == SystemZ::STDY)) |
691 | ++MBBI; |
692 | else |
693 | llvm_unreachable("Couldn't skip over FPR save" ); |
694 | } else if (SystemZ::VR128BitRegClass.contains(Reg)) { |
695 | if (MBBI != MBB.end() && |
696 | MBBI->getOpcode() == SystemZ::VST) |
697 | ++MBBI; |
698 | else |
699 | llvm_unreachable("Couldn't skip over VR save" ); |
700 | } else |
701 | continue; |
702 | |
703 | // Add CFI for the this save. |
704 | unsigned DwarfReg = MRI->getDwarfRegNum(RegNum: Reg, isEH: true); |
705 | Register IgnoredFrameReg; |
706 | int64_t Offset = |
707 | getFrameIndexReference(MF, FI: Save.getFrameIdx(), FrameReg&: IgnoredFrameReg) |
708 | .getFixed(); |
709 | |
710 | unsigned CFIIndex = MF.addFrameInst(Inst: MCCFIInstruction::createOffset( |
711 | L: nullptr, Register: DwarfReg, Offset: SPOffsetFromCFA + Offset)); |
712 | CFIIndexes.push_back(Elt: CFIIndex); |
713 | } |
714 | // Complete the CFI for the FPR/VR saves, modelling them as taking effect |
715 | // after the last save. |
716 | for (auto CFIIndex : CFIIndexes) { |
717 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: TargetOpcode::CFI_INSTRUCTION)) |
718 | .addCFIIndex(CFIIndex); |
719 | } |
720 | } |
721 | |
722 | void SystemZELFFrameLowering::emitEpilogue(MachineFunction &MF, |
723 | MachineBasicBlock &MBB) const { |
724 | MachineBasicBlock::iterator MBBI = MBB.getLastNonDebugInstr(); |
725 | auto *ZII = |
726 | static_cast<const SystemZInstrInfo *>(MF.getSubtarget().getInstrInfo()); |
727 | SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); |
728 | MachineFrameInfo &MFFrame = MF.getFrameInfo(); |
729 | |
730 | // See SystemZELFFrameLowering::emitPrologue |
731 | if (MF.getFunction().getCallingConv() == CallingConv::GHC) |
732 | return; |
733 | |
734 | // Skip the return instruction. |
735 | assert(MBBI->isReturn() && "Can only insert epilogue into returning blocks" ); |
736 | |
737 | uint64_t StackSize = MFFrame.getStackSize(); |
738 | if (ZFI->getRestoreGPRRegs().LowGPR) { |
739 | --MBBI; |
740 | unsigned Opcode = MBBI->getOpcode(); |
741 | if (Opcode != SystemZ::LMG) |
742 | llvm_unreachable("Expected to see callee-save register restore code" ); |
743 | |
744 | unsigned AddrOpNo = 2; |
745 | DebugLoc DL = MBBI->getDebugLoc(); |
746 | uint64_t Offset = StackSize + MBBI->getOperand(i: AddrOpNo + 1).getImm(); |
747 | unsigned NewOpcode = ZII->getOpcodeForOffset(Opcode, Offset); |
748 | |
749 | // If the offset is too large, use the largest stack-aligned offset |
750 | // and add the rest to the base register (the stack or frame pointer). |
751 | if (!NewOpcode) { |
752 | uint64_t NumBytes = Offset - 0x7fff8; |
753 | emitIncrement(MBB, MBBI, DL, Reg: MBBI->getOperand(i: AddrOpNo).getReg(), |
754 | NumBytes, TII: ZII); |
755 | Offset -= NumBytes; |
756 | NewOpcode = ZII->getOpcodeForOffset(Opcode, Offset); |
757 | assert(NewOpcode && "No restore instruction available" ); |
758 | } |
759 | |
760 | MBBI->setDesc(ZII->get(Opcode: NewOpcode)); |
761 | MBBI->getOperand(i: AddrOpNo + 1).ChangeToImmediate(ImmVal: Offset); |
762 | } else if (StackSize) { |
763 | DebugLoc DL = MBBI->getDebugLoc(); |
764 | emitIncrement(MBB, MBBI, DL, Reg: SystemZ::R15D, NumBytes: StackSize, TII: ZII); |
765 | } |
766 | } |
767 | |
768 | void SystemZELFFrameLowering::inlineStackProbe( |
769 | MachineFunction &MF, MachineBasicBlock &PrologMBB) const { |
770 | auto *ZII = |
771 | static_cast<const SystemZInstrInfo *>(MF.getSubtarget().getInstrInfo()); |
772 | const SystemZSubtarget &STI = MF.getSubtarget<SystemZSubtarget>(); |
773 | const SystemZTargetLowering &TLI = *STI.getTargetLowering(); |
774 | |
775 | MachineInstr *StackAllocMI = nullptr; |
776 | for (MachineInstr &MI : PrologMBB) |
777 | if (MI.getOpcode() == SystemZ::PROBED_STACKALLOC) { |
778 | StackAllocMI = &MI; |
779 | break; |
780 | } |
781 | if (StackAllocMI == nullptr) |
782 | return; |
783 | uint64_t StackSize = StackAllocMI->getOperand(i: 0).getImm(); |
784 | const unsigned ProbeSize = TLI.getStackProbeSize(MF); |
785 | uint64_t NumFullBlocks = StackSize / ProbeSize; |
786 | uint64_t Residual = StackSize % ProbeSize; |
787 | int64_t SPOffsetFromCFA = -SystemZMC::ELFCFAOffsetFromInitialSP; |
788 | MachineBasicBlock *MBB = &PrologMBB; |
789 | MachineBasicBlock::iterator MBBI = StackAllocMI; |
790 | const DebugLoc DL = StackAllocMI->getDebugLoc(); |
791 | |
792 | // Allocate a block of Size bytes on the stack and probe it. |
793 | auto allocateAndProbe = [&](MachineBasicBlock &InsMBB, |
794 | MachineBasicBlock::iterator InsPt, unsigned Size, |
795 | bool EmitCFI) -> void { |
796 | emitIncrement(MBB&: InsMBB, MBBI&: InsPt, DL, Reg: SystemZ::R15D, NumBytes: -int64_t(Size), TII: ZII); |
797 | if (EmitCFI) { |
798 | SPOffsetFromCFA -= Size; |
799 | buildCFAOffs(MBB&: InsMBB, MBBI: InsPt, DL, Offset: SPOffsetFromCFA, ZII); |
800 | } |
801 | // Probe by means of a volatile compare. |
802 | MachineMemOperand *MMO = MF.getMachineMemOperand(PtrInfo: MachinePointerInfo(), |
803 | F: MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad, Size: 8, BaseAlignment: Align(1)); |
804 | BuildMI(BB&: InsMBB, I: InsPt, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::CG)) |
805 | .addReg(RegNo: SystemZ::R0D, flags: RegState::Undef) |
806 | .addReg(RegNo: SystemZ::R15D).addImm(Val: Size - 8).addReg(RegNo: 0) |
807 | .addMemOperand(MMO); |
808 | }; |
809 | |
810 | bool StoreBackchain = MF.getSubtarget<SystemZSubtarget>().hasBackChain(); |
811 | if (StoreBackchain) |
812 | BuildMI(BB&: *MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::LGR)) |
813 | .addReg(RegNo: SystemZ::R1D, flags: RegState::Define).addReg(RegNo: SystemZ::R15D); |
814 | |
815 | MachineBasicBlock *DoneMBB = nullptr; |
816 | MachineBasicBlock *LoopMBB = nullptr; |
817 | if (NumFullBlocks < 3) { |
818 | // Emit unrolled probe statements. |
819 | for (unsigned int i = 0; i < NumFullBlocks; i++) |
820 | allocateAndProbe(*MBB, MBBI, ProbeSize, true/*EmitCFI*/); |
821 | } else { |
822 | // Emit a loop probing the pages. |
823 | uint64_t LoopAlloc = ProbeSize * NumFullBlocks; |
824 | SPOffsetFromCFA -= LoopAlloc; |
825 | |
826 | // Use R0D to hold the exit value. |
827 | BuildMI(BB&: *MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::LGR), DestReg: SystemZ::R0D) |
828 | .addReg(RegNo: SystemZ::R15D); |
829 | buildDefCFAReg(MBB&: *MBB, MBBI, DL, Reg: SystemZ::R0D, ZII); |
830 | emitIncrement(MBB&: *MBB, MBBI, DL, Reg: SystemZ::R0D, NumBytes: -int64_t(LoopAlloc), TII: ZII); |
831 | buildCFAOffs(MBB&: *MBB, MBBI, DL, Offset: -int64_t(SystemZMC::ELFCallFrameSize + LoopAlloc), |
832 | ZII); |
833 | |
834 | DoneMBB = SystemZ::splitBlockBefore(MI: MBBI, MBB); |
835 | LoopMBB = SystemZ::emitBlockAfter(MBB); |
836 | MBB->addSuccessor(Succ: LoopMBB); |
837 | LoopMBB->addSuccessor(Succ: LoopMBB); |
838 | LoopMBB->addSuccessor(Succ: DoneMBB); |
839 | |
840 | MBB = LoopMBB; |
841 | allocateAndProbe(*MBB, MBB->end(), ProbeSize, false/*EmitCFI*/); |
842 | BuildMI(BB&: *MBB, I: MBB->end(), MIMD: DL, MCID: ZII->get(Opcode: SystemZ::CLGR)) |
843 | .addReg(RegNo: SystemZ::R15D).addReg(RegNo: SystemZ::R0D); |
844 | BuildMI(BB&: *MBB, I: MBB->end(), MIMD: DL, MCID: ZII->get(Opcode: SystemZ::BRC)) |
845 | .addImm(Val: SystemZ::CCMASK_ICMP).addImm(Val: SystemZ::CCMASK_CMP_GT).addMBB(MBB); |
846 | |
847 | MBB = DoneMBB; |
848 | MBBI = DoneMBB->begin(); |
849 | buildDefCFAReg(MBB&: *MBB, MBBI, DL, Reg: SystemZ::R15D, ZII); |
850 | } |
851 | |
852 | if (Residual) |
853 | allocateAndProbe(*MBB, MBBI, Residual, true/*EmitCFI*/); |
854 | |
855 | if (StoreBackchain) |
856 | BuildMI(BB&: *MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::STG)) |
857 | .addReg(RegNo: SystemZ::R1D, flags: RegState::Kill).addReg(RegNo: SystemZ::R15D) |
858 | .addImm(Val: getBackchainOffset(MF)).addReg(RegNo: 0); |
859 | |
860 | StackAllocMI->eraseFromParent(); |
861 | if (DoneMBB != nullptr) { |
862 | // Compute the live-in lists for the new blocks. |
863 | fullyRecomputeLiveIns(MBBs: {DoneMBB, LoopMBB}); |
864 | } |
865 | } |
866 | |
867 | bool SystemZELFFrameLowering::hasFPImpl(const MachineFunction &MF) const { |
868 | return (MF.getTarget().Options.DisableFramePointerElim(MF) || |
869 | MF.getFrameInfo().hasVarSizedObjects()); |
870 | } |
871 | |
872 | StackOffset SystemZELFFrameLowering::getFrameIndexReference( |
873 | const MachineFunction &MF, int FI, Register &FrameReg) const { |
874 | // Our incoming SP is actually SystemZMC::ELFCallFrameSize below the CFA, so |
875 | // add that difference here. |
876 | StackOffset Offset = |
877 | TargetFrameLowering::getFrameIndexReference(MF, FI, FrameReg); |
878 | return Offset + StackOffset::getFixed(Fixed: SystemZMC::ELFCallFrameSize); |
879 | } |
880 | |
881 | unsigned SystemZELFFrameLowering::getRegSpillOffset(MachineFunction &MF, |
882 | Register Reg) const { |
883 | bool IsVarArg = MF.getFunction().isVarArg(); |
884 | const SystemZSubtarget &Subtarget = MF.getSubtarget<SystemZSubtarget>(); |
885 | bool BackChain = Subtarget.hasBackChain(); |
886 | bool SoftFloat = Subtarget.hasSoftFloat(); |
887 | unsigned Offset = RegSpillOffsets[Reg]; |
888 | if (usePackedStack(MF) && !(IsVarArg && !SoftFloat)) { |
889 | if (SystemZ::GR64BitRegClass.contains(Reg)) |
890 | // Put all GPRs at the top of the Register save area with packed |
891 | // stack. Make room for the backchain if needed. |
892 | Offset += BackChain ? 24 : 32; |
893 | else |
894 | Offset = 0; |
895 | } |
896 | return Offset; |
897 | } |
898 | |
899 | int SystemZELFFrameLowering::getOrCreateFramePointerSaveIndex( |
900 | MachineFunction &MF) const { |
901 | SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); |
902 | int FI = ZFI->getFramePointerSaveIndex(); |
903 | if (!FI) { |
904 | MachineFrameInfo &MFFrame = MF.getFrameInfo(); |
905 | int Offset = getBackchainOffset(MF) - SystemZMC::ELFCallFrameSize; |
906 | FI = MFFrame.CreateFixedObject(Size: getPointerSize(), SPOffset: Offset, IsImmutable: false); |
907 | ZFI->setFramePointerSaveIndex(FI); |
908 | } |
909 | return FI; |
910 | } |
911 | |
912 | bool SystemZELFFrameLowering::usePackedStack(MachineFunction &MF) const { |
913 | bool HasPackedStackAttr = MF.getFunction().hasFnAttribute(Kind: "packed-stack" ); |
914 | const SystemZSubtarget &Subtarget = MF.getSubtarget<SystemZSubtarget>(); |
915 | bool BackChain = Subtarget.hasBackChain(); |
916 | bool SoftFloat = Subtarget.hasSoftFloat(); |
917 | if (HasPackedStackAttr && BackChain && !SoftFloat) |
918 | report_fatal_error(reason: "packed-stack + backchain + hard-float is unsupported." ); |
919 | bool CallConv = MF.getFunction().getCallingConv() != CallingConv::GHC; |
920 | return HasPackedStackAttr && CallConv; |
921 | } |
922 | |
923 | SystemZXPLINKFrameLowering::SystemZXPLINKFrameLowering(unsigned PointerSize) |
924 | : SystemZFrameLowering(TargetFrameLowering::StackGrowsDown, Align(32), 0, |
925 | Align(32), /* StackRealignable */ false, |
926 | PointerSize), |
927 | RegSpillOffsets(-1) { |
928 | |
929 | // Create a mapping from register number to save slot offset. |
930 | // These offsets are relative to the start of the local are area. |
931 | RegSpillOffsets.grow(n: SystemZ::NUM_TARGET_REGS); |
932 | for (const auto &Entry : XPLINKSpillOffsetTable) |
933 | RegSpillOffsets[Entry.Reg] = Entry.Offset; |
934 | } |
935 | |
936 | int SystemZXPLINKFrameLowering::getOrCreateFramePointerSaveIndex( |
937 | MachineFunction &MF) const { |
938 | SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); |
939 | int FI = ZFI->getFramePointerSaveIndex(); |
940 | if (!FI) { |
941 | MachineFrameInfo &MFFrame = MF.getFrameInfo(); |
942 | FI = MFFrame.CreateFixedObject(Size: getPointerSize(), SPOffset: 0, IsImmutable: false); |
943 | MFFrame.setStackID(ObjectIdx: FI, ID: TargetStackID::NoAlloc); |
944 | ZFI->setFramePointerSaveIndex(FI); |
945 | } |
946 | return FI; |
947 | } |
948 | |
949 | // Checks if the function is a potential candidate for being a XPLeaf routine. |
950 | static bool isXPLeafCandidate(const MachineFunction &MF) { |
951 | const MachineFrameInfo &MFFrame = MF.getFrameInfo(); |
952 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
953 | const SystemZSubtarget &Subtarget = MF.getSubtarget<SystemZSubtarget>(); |
954 | auto *Regs = |
955 | static_cast<SystemZXPLINK64Registers *>(Subtarget.getSpecialRegisters()); |
956 | |
957 | // If function calls other functions including alloca, then it is not a XPLeaf |
958 | // routine. |
959 | if (MFFrame.hasCalls()) |
960 | return false; |
961 | |
962 | // If the function has var Sized Objects, then it is not a XPLeaf routine. |
963 | if (MFFrame.hasVarSizedObjects()) |
964 | return false; |
965 | |
966 | // If the function adjusts the stack, then it is not a XPLeaf routine. |
967 | if (MFFrame.adjustsStack()) |
968 | return false; |
969 | |
970 | // If function modifies the stack pointer register, then it is not a XPLeaf |
971 | // routine. |
972 | if (MRI.isPhysRegModified(PhysReg: Regs->getStackPointerRegister())) |
973 | return false; |
974 | |
975 | // If function modifies the ADA register, then it is not a XPLeaf routine. |
976 | if (MRI.isPhysRegModified(PhysReg: Regs->getAddressOfCalleeRegister())) |
977 | return false; |
978 | |
979 | // If function modifies the return address register, then it is not a XPLeaf |
980 | // routine. |
981 | if (MRI.isPhysRegModified(PhysReg: Regs->getReturnFunctionAddressRegister())) |
982 | return false; |
983 | |
984 | // If the backchain pointer should be stored, then it is not a XPLeaf routine. |
985 | if (MF.getSubtarget<SystemZSubtarget>().hasBackChain()) |
986 | return false; |
987 | |
988 | // If function acquires its own stack frame, then it is not a XPLeaf routine. |
989 | // At the time this function is called, only slots for local variables are |
990 | // allocated, so this is a very rough estimate. |
991 | if (MFFrame.estimateStackSize(MF) > 0) |
992 | return false; |
993 | |
994 | return true; |
995 | } |
996 | |
997 | bool SystemZXPLINKFrameLowering::assignCalleeSavedSpillSlots( |
998 | MachineFunction &MF, const TargetRegisterInfo *TRI, |
999 | std::vector<CalleeSavedInfo> &CSI) const { |
1000 | MachineFrameInfo &MFFrame = MF.getFrameInfo(); |
1001 | SystemZMachineFunctionInfo *MFI = MF.getInfo<SystemZMachineFunctionInfo>(); |
1002 | const SystemZSubtarget &Subtarget = MF.getSubtarget<SystemZSubtarget>(); |
1003 | auto &Regs = Subtarget.getSpecialRegisters<SystemZXPLINK64Registers>(); |
1004 | auto &GRRegClass = SystemZ::GR64BitRegClass; |
1005 | |
1006 | // At this point, the result of isXPLeafCandidate() is not accurate because |
1007 | // the size of the save area has not yet been determined. If |
1008 | // isXPLeafCandidate() indicates a potential leaf function, and there are no |
1009 | // callee-save registers, then it is indeed a leaf function, and we can early |
1010 | // exit. |
1011 | // TODO: It is possible for leaf functions to use callee-saved registers. |
1012 | // It can use the 0-2k range between R4 and the caller's stack frame without |
1013 | // acquiring its own stack frame. |
1014 | bool IsLeaf = CSI.empty() && isXPLeafCandidate(MF); |
1015 | if (IsLeaf) |
1016 | return true; |
1017 | |
1018 | // For non-leaf functions: |
1019 | // - the address of callee (entry point) register R6 must be saved |
1020 | CSI.push_back(x: CalleeSavedInfo(Regs.getAddressOfCalleeRegister())); |
1021 | CSI.back().setRestored(false); |
1022 | |
1023 | // The return address register R7 must be saved and restored. |
1024 | CSI.push_back(x: CalleeSavedInfo(Regs.getReturnFunctionAddressRegister())); |
1025 | |
1026 | // If the function needs a frame pointer, or if the backchain pointer should |
1027 | // be stored, then save the stack pointer register R4. |
1028 | if (hasFP(MF) || Subtarget.hasBackChain()) |
1029 | CSI.push_back(x: CalleeSavedInfo(Regs.getStackPointerRegister())); |
1030 | |
1031 | // If this function has an associated personality function then the |
1032 | // environment register R5 must be saved in the DSA. |
1033 | if (!MF.getLandingPads().empty()) |
1034 | CSI.push_back(x: CalleeSavedInfo(Regs.getADARegister())); |
1035 | |
1036 | // Scan the call-saved GPRs and find the bounds of the register spill area. |
1037 | Register LowRestoreGPR = 0; |
1038 | int LowRestoreOffset = INT32_MAX; |
1039 | Register LowSpillGPR = 0; |
1040 | int LowSpillOffset = INT32_MAX; |
1041 | Register HighGPR = 0; |
1042 | int HighOffset = -1; |
1043 | |
1044 | // Query index of the saved frame pointer. |
1045 | int FPSI = MFI->getFramePointerSaveIndex(); |
1046 | |
1047 | for (auto &CS : CSI) { |
1048 | MCRegister Reg = CS.getReg(); |
1049 | int Offset = RegSpillOffsets[Reg]; |
1050 | if (Offset >= 0) { |
1051 | if (GRRegClass.contains(Reg)) { |
1052 | if (LowSpillOffset > Offset) { |
1053 | LowSpillOffset = Offset; |
1054 | LowSpillGPR = Reg; |
1055 | } |
1056 | if (CS.isRestored() && LowRestoreOffset > Offset) { |
1057 | LowRestoreOffset = Offset; |
1058 | LowRestoreGPR = Reg; |
1059 | } |
1060 | |
1061 | if (Offset > HighOffset) { |
1062 | HighOffset = Offset; |
1063 | HighGPR = Reg; |
1064 | } |
1065 | // Non-volatile GPRs are saved in the dedicated register save area at |
1066 | // the bottom of the stack and are not truly part of the "normal" stack |
1067 | // frame. Mark the frame index as NoAlloc to indicate it as such. |
1068 | unsigned RegSize = getPointerSize(); |
1069 | int FrameIdx = |
1070 | (FPSI && Offset == 0) |
1071 | ? FPSI |
1072 | : MFFrame.CreateFixedSpillStackObject(Size: RegSize, SPOffset: Offset); |
1073 | CS.setFrameIdx(FrameIdx); |
1074 | MFFrame.setStackID(ObjectIdx: FrameIdx, ID: TargetStackID::NoAlloc); |
1075 | } |
1076 | } else { |
1077 | MCRegister Reg = CS.getReg(); |
1078 | const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); |
1079 | Align Alignment = TRI->getSpillAlign(RC: *RC); |
1080 | unsigned Size = TRI->getSpillSize(RC: *RC); |
1081 | Alignment = std::min(a: Alignment, b: getStackAlign()); |
1082 | int FrameIdx = MFFrame.CreateStackObject(Size, Alignment, isSpillSlot: true); |
1083 | CS.setFrameIdx(FrameIdx); |
1084 | } |
1085 | } |
1086 | |
1087 | // Save the range of call-saved registers, for use by the |
1088 | // prologue/epilogue inserters. |
1089 | if (LowRestoreGPR) |
1090 | MFI->setRestoreGPRRegs(Low: LowRestoreGPR, High: HighGPR, Offs: LowRestoreOffset); |
1091 | |
1092 | // Save the range of call-saved registers, for use by the epilogue inserter. |
1093 | assert(LowSpillGPR && "Expected registers to spill" ); |
1094 | MFI->setSpillGPRRegs(Low: LowSpillGPR, High: HighGPR, Offs: LowSpillOffset); |
1095 | |
1096 | return true; |
1097 | } |
1098 | |
1099 | void SystemZXPLINKFrameLowering::determineCalleeSaves(MachineFunction &MF, |
1100 | BitVector &SavedRegs, |
1101 | RegScavenger *RS) const { |
1102 | TargetFrameLowering::determineCalleeSaves(MF, SavedRegs, RS); |
1103 | |
1104 | bool HasFP = hasFP(MF); |
1105 | const SystemZSubtarget &Subtarget = MF.getSubtarget<SystemZSubtarget>(); |
1106 | auto &Regs = Subtarget.getSpecialRegisters<SystemZXPLINK64Registers>(); |
1107 | |
1108 | // If the function requires a frame pointer, record that the hard |
1109 | // frame pointer will be clobbered. |
1110 | if (HasFP) |
1111 | SavedRegs.set(Regs.getFramePointerRegister()); |
1112 | } |
1113 | |
1114 | bool SystemZXPLINKFrameLowering::spillCalleeSavedRegisters( |
1115 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, |
1116 | ArrayRef<CalleeSavedInfo> CSI, const TargetRegisterInfo *TRI) const { |
1117 | if (CSI.empty()) |
1118 | return true; |
1119 | |
1120 | MachineFunction &MF = *MBB.getParent(); |
1121 | SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); |
1122 | const SystemZSubtarget &Subtarget = MF.getSubtarget<SystemZSubtarget>(); |
1123 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
1124 | auto &Regs = Subtarget.getSpecialRegisters<SystemZXPLINK64Registers>(); |
1125 | SystemZ::GPRRegs SpillGPRs = ZFI->getSpillGPRRegs(); |
1126 | DebugLoc DL; |
1127 | |
1128 | // Save GPRs |
1129 | if (SpillGPRs.LowGPR) { |
1130 | assert(SpillGPRs.LowGPR != SpillGPRs.HighGPR && |
1131 | "Should be saving multiple registers" ); |
1132 | |
1133 | // Build an STM/STMG instruction. |
1134 | MachineInstrBuilder MIB = BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: SystemZ::STMG)); |
1135 | |
1136 | // Add the explicit register operands. |
1137 | addSavedGPR(MBB, MIB, GPR64: SpillGPRs.LowGPR, IsImplicit: false); |
1138 | addSavedGPR(MBB, MIB, GPR64: SpillGPRs.HighGPR, IsImplicit: false); |
1139 | |
1140 | // Add the address r4 |
1141 | MIB.addReg(RegNo: Regs.getStackPointerRegister()); |
1142 | |
1143 | // Add the partial offset |
1144 | // We cannot add the actual offset as, at the stack is not finalized |
1145 | MIB.addImm(Val: SpillGPRs.GPROffset); |
1146 | |
1147 | // Make sure all call-saved GPRs are included as operands and are |
1148 | // marked as live on entry. |
1149 | auto &GRRegClass = SystemZ::GR64BitRegClass; |
1150 | for (const CalleeSavedInfo &I : CSI) { |
1151 | MCRegister Reg = I.getReg(); |
1152 | if (GRRegClass.contains(Reg)) |
1153 | addSavedGPR(MBB, MIB, GPR64: Reg, IsImplicit: true); |
1154 | } |
1155 | } |
1156 | |
1157 | // Spill FPRs to the stack in the normal TargetInstrInfo way |
1158 | for (const CalleeSavedInfo &I : CSI) { |
1159 | MCRegister Reg = I.getReg(); |
1160 | if (SystemZ::FP64BitRegClass.contains(Reg)) { |
1161 | MBB.addLiveIn(PhysReg: Reg); |
1162 | TII->storeRegToStackSlot(MBB, MI: MBBI, SrcReg: Reg, isKill: true, FrameIndex: I.getFrameIdx(), |
1163 | RC: &SystemZ::FP64BitRegClass, TRI, VReg: Register()); |
1164 | } |
1165 | if (SystemZ::VR128BitRegClass.contains(Reg)) { |
1166 | MBB.addLiveIn(PhysReg: Reg); |
1167 | TII->storeRegToStackSlot(MBB, MI: MBBI, SrcReg: Reg, isKill: true, FrameIndex: I.getFrameIdx(), |
1168 | RC: &SystemZ::VR128BitRegClass, TRI, VReg: Register()); |
1169 | } |
1170 | } |
1171 | |
1172 | return true; |
1173 | } |
1174 | |
1175 | bool SystemZXPLINKFrameLowering::restoreCalleeSavedRegisters( |
1176 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, |
1177 | MutableArrayRef<CalleeSavedInfo> CSI, const TargetRegisterInfo *TRI) const { |
1178 | |
1179 | if (CSI.empty()) |
1180 | return false; |
1181 | |
1182 | MachineFunction &MF = *MBB.getParent(); |
1183 | SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); |
1184 | const SystemZSubtarget &Subtarget = MF.getSubtarget<SystemZSubtarget>(); |
1185 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
1186 | auto &Regs = Subtarget.getSpecialRegisters<SystemZXPLINK64Registers>(); |
1187 | |
1188 | DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); |
1189 | |
1190 | // Restore FPRs in the normal TargetInstrInfo way. |
1191 | for (const CalleeSavedInfo &I : CSI) { |
1192 | MCRegister Reg = I.getReg(); |
1193 | if (SystemZ::FP64BitRegClass.contains(Reg)) |
1194 | TII->loadRegFromStackSlot(MBB, MI: MBBI, DestReg: Reg, FrameIndex: I.getFrameIdx(), |
1195 | RC: &SystemZ::FP64BitRegClass, TRI, VReg: Register()); |
1196 | if (SystemZ::VR128BitRegClass.contains(Reg)) |
1197 | TII->loadRegFromStackSlot(MBB, MI: MBBI, DestReg: Reg, FrameIndex: I.getFrameIdx(), |
1198 | RC: &SystemZ::VR128BitRegClass, TRI, VReg: Register()); |
1199 | } |
1200 | |
1201 | // Restore call-saved GPRs (but not call-clobbered varargs, which at |
1202 | // this point might hold return values). |
1203 | SystemZ::GPRRegs RestoreGPRs = ZFI->getRestoreGPRRegs(); |
1204 | if (RestoreGPRs.LowGPR) { |
1205 | assert(isInt<20>(Regs.getStackPointerBias() + RestoreGPRs.GPROffset)); |
1206 | if (RestoreGPRs.LowGPR == RestoreGPRs.HighGPR) |
1207 | // Build an LG/L instruction. |
1208 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: SystemZ::LG), DestReg: RestoreGPRs.LowGPR) |
1209 | .addReg(RegNo: Regs.getStackPointerRegister()) |
1210 | .addImm(Val: Regs.getStackPointerBias() + RestoreGPRs.GPROffset) |
1211 | .addReg(RegNo: 0); |
1212 | else { |
1213 | // Build an LMG/LM instruction. |
1214 | MachineInstrBuilder MIB = BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: SystemZ::LMG)); |
1215 | |
1216 | // Add the explicit register operands. |
1217 | MIB.addReg(RegNo: RestoreGPRs.LowGPR, flags: RegState::Define); |
1218 | MIB.addReg(RegNo: RestoreGPRs.HighGPR, flags: RegState::Define); |
1219 | |
1220 | // Add the address. |
1221 | MIB.addReg(RegNo: Regs.getStackPointerRegister()); |
1222 | MIB.addImm(Val: Regs.getStackPointerBias() + RestoreGPRs.GPROffset); |
1223 | |
1224 | // Do a second scan adding regs as being defined by instruction |
1225 | for (const CalleeSavedInfo &I : CSI) { |
1226 | MCRegister Reg = I.getReg(); |
1227 | if (Reg > RestoreGPRs.LowGPR && Reg < RestoreGPRs.HighGPR) |
1228 | MIB.addReg(RegNo: Reg, flags: RegState::ImplicitDefine); |
1229 | } |
1230 | } |
1231 | } |
1232 | |
1233 | return true; |
1234 | } |
1235 | |
1236 | void SystemZXPLINKFrameLowering::emitPrologue(MachineFunction &MF, |
1237 | MachineBasicBlock &MBB) const { |
1238 | assert(&MF.front() == &MBB && "Shrink-wrapping not yet supported" ); |
1239 | const SystemZSubtarget &Subtarget = MF.getSubtarget<SystemZSubtarget>(); |
1240 | SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); |
1241 | MachineBasicBlock::iterator MBBI = MBB.begin(); |
1242 | auto *ZII = static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo()); |
1243 | auto &Regs = Subtarget.getSpecialRegisters<SystemZXPLINK64Registers>(); |
1244 | MachineFrameInfo &MFFrame = MF.getFrameInfo(); |
1245 | MachineInstr *StoreInstr = nullptr; |
1246 | |
1247 | determineFrameLayout(MF); |
1248 | |
1249 | bool HasFP = hasFP(MF); |
1250 | // Debug location must be unknown since the first debug location is used |
1251 | // to determine the end of the prologue. |
1252 | DebugLoc DL; |
1253 | uint64_t Offset = 0; |
1254 | |
1255 | const uint64_t StackSize = MFFrame.getStackSize(); |
1256 | |
1257 | if (ZFI->getSpillGPRRegs().LowGPR) { |
1258 | // Skip over the GPR saves. |
1259 | if ((MBBI != MBB.end()) && ((MBBI->getOpcode() == SystemZ::STMG))) { |
1260 | const int Operand = 3; |
1261 | // Now we can set the offset for the operation, since now the Stack |
1262 | // has been finalized. |
1263 | Offset = Regs.getStackPointerBias() + MBBI->getOperand(i: Operand).getImm(); |
1264 | // Maximum displacement for STMG instruction. |
1265 | if (isInt<20>(x: Offset - StackSize)) |
1266 | Offset -= StackSize; |
1267 | else |
1268 | StoreInstr = &*MBBI; |
1269 | MBBI->getOperand(i: Operand).setImm(Offset); |
1270 | ++MBBI; |
1271 | } else |
1272 | llvm_unreachable("Couldn't skip over GPR saves" ); |
1273 | } |
1274 | |
1275 | if (StackSize) { |
1276 | MachineBasicBlock::iterator InsertPt = StoreInstr ? StoreInstr : MBBI; |
1277 | // Allocate StackSize bytes. |
1278 | int64_t Delta = -int64_t(StackSize); |
1279 | |
1280 | // In case the STM(G) instruction also stores SP (R4), but the displacement |
1281 | // is too large, the SP register is manipulated first before storing, |
1282 | // resulting in the wrong value stored and retrieved later. In this case, we |
1283 | // need to temporarily save the value of SP, and store it later to memory. |
1284 | if (StoreInstr && HasFP) { |
1285 | // Insert LR r0,r4 before STMG instruction. |
1286 | BuildMI(BB&: MBB, I: InsertPt, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::LGR)) |
1287 | .addReg(RegNo: SystemZ::R0D, flags: RegState::Define) |
1288 | .addReg(RegNo: SystemZ::R4D); |
1289 | // Insert ST r0,xxx(,r4) after STMG instruction. |
1290 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::STG)) |
1291 | .addReg(RegNo: SystemZ::R0D, flags: RegState::Kill) |
1292 | .addReg(RegNo: SystemZ::R4D) |
1293 | .addImm(Val: Offset) |
1294 | .addReg(RegNo: 0); |
1295 | } |
1296 | |
1297 | emitIncrement(MBB, MBBI&: InsertPt, DL, Reg: Regs.getStackPointerRegister(), NumBytes: Delta, |
1298 | TII: ZII); |
1299 | |
1300 | // If the requested stack size is larger than the guard page, then we need |
1301 | // to check if we need to call the stack extender. This requires adding a |
1302 | // conditional branch, but splitting the prologue block is not possible at |
1303 | // this point since it would invalidate the SaveBlocks / RestoreBlocks sets |
1304 | // of PEI in the single block function case. Build a pseudo to be handled |
1305 | // later by inlineStackProbe(). |
1306 | const uint64_t GuardPageSize = 1024 * 1024; |
1307 | if (StackSize > GuardPageSize) { |
1308 | assert(StoreInstr && "Wrong insertion point" ); |
1309 | BuildMI(BB&: MBB, I: InsertPt, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::XPLINK_STACKALLOC)); |
1310 | } |
1311 | } |
1312 | |
1313 | if (HasFP) { |
1314 | // Copy the base of the frame to Frame Pointer Register. |
1315 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::LGR), |
1316 | DestReg: Regs.getFramePointerRegister()) |
1317 | .addReg(RegNo: Regs.getStackPointerRegister()); |
1318 | |
1319 | // Mark the FramePtr as live at the beginning of every block except |
1320 | // the entry block. (We'll have marked R8 as live on entry when |
1321 | // saving the GPRs.) |
1322 | for (MachineBasicBlock &B : llvm::drop_begin(RangeOrContainer&: MF)) |
1323 | B.addLiveIn(PhysReg: Regs.getFramePointerRegister()); |
1324 | } |
1325 | |
1326 | // Save GPRs used for varargs, if any. |
1327 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
1328 | bool IsVarArg = MF.getFunction().isVarArg(); |
1329 | |
1330 | if (IsVarArg) { |
1331 | // FixedRegs is the number of used registers, accounting for shadow |
1332 | // registers. |
1333 | unsigned FixedRegs = ZFI->getVarArgsFirstGPR() + ZFI->getVarArgsFirstFPR(); |
1334 | auto &GPRs = SystemZ::XPLINK64ArgGPRs; |
1335 | for (unsigned I = FixedRegs; I < SystemZ::XPLINK64NumArgGPRs; I++) { |
1336 | uint64_t StartOffset = MFFrame.getOffsetAdjustment() + |
1337 | MFFrame.getStackSize() + Regs.getCallFrameSize() + |
1338 | getOffsetOfLocalArea() + I * getPointerSize(); |
1339 | unsigned Reg = GPRs[I]; |
1340 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: SystemZ::STG)) |
1341 | .addReg(RegNo: Reg) |
1342 | .addReg(RegNo: Regs.getStackPointerRegister()) |
1343 | .addImm(Val: StartOffset) |
1344 | .addReg(RegNo: 0); |
1345 | if (!MBB.isLiveIn(Reg)) |
1346 | MBB.addLiveIn(PhysReg: Reg); |
1347 | } |
1348 | } |
1349 | } |
1350 | |
1351 | void SystemZXPLINKFrameLowering::emitEpilogue(MachineFunction &MF, |
1352 | MachineBasicBlock &MBB) const { |
1353 | const SystemZSubtarget &Subtarget = MF.getSubtarget<SystemZSubtarget>(); |
1354 | MachineBasicBlock::iterator MBBI = MBB.getLastNonDebugInstr(); |
1355 | SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); |
1356 | MachineFrameInfo &MFFrame = MF.getFrameInfo(); |
1357 | auto *ZII = static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo()); |
1358 | auto &Regs = Subtarget.getSpecialRegisters<SystemZXPLINK64Registers>(); |
1359 | |
1360 | // Skip the return instruction. |
1361 | assert(MBBI->isReturn() && "Can only insert epilogue into returning blocks" ); |
1362 | |
1363 | uint64_t StackSize = MFFrame.getStackSize(); |
1364 | if (StackSize) { |
1365 | unsigned SPReg = Regs.getStackPointerRegister(); |
1366 | if (ZFI->getRestoreGPRRegs().LowGPR != SPReg) { |
1367 | DebugLoc DL = MBBI->getDebugLoc(); |
1368 | emitIncrement(MBB, MBBI, DL, Reg: SPReg, NumBytes: StackSize, TII: ZII); |
1369 | } |
1370 | } |
1371 | } |
1372 | |
1373 | // Emit a compare of the stack pointer against the stack floor, and a call to |
1374 | // the LE stack extender if needed. |
1375 | void SystemZXPLINKFrameLowering::inlineStackProbe( |
1376 | MachineFunction &MF, MachineBasicBlock &PrologMBB) const { |
1377 | auto *ZII = |
1378 | static_cast<const SystemZInstrInfo *>(MF.getSubtarget().getInstrInfo()); |
1379 | |
1380 | MachineInstr *StackAllocMI = nullptr; |
1381 | for (MachineInstr &MI : PrologMBB) |
1382 | if (MI.getOpcode() == SystemZ::XPLINK_STACKALLOC) { |
1383 | StackAllocMI = &MI; |
1384 | break; |
1385 | } |
1386 | if (StackAllocMI == nullptr) |
1387 | return; |
1388 | |
1389 | bool NeedSaveSP = hasFP(MF); |
1390 | bool NeedSaveArg = PrologMBB.isLiveIn(Reg: SystemZ::R3D); |
1391 | const int64_t SaveSlotR3 = 2192; |
1392 | |
1393 | MachineBasicBlock &MBB = PrologMBB; |
1394 | const DebugLoc DL = StackAllocMI->getDebugLoc(); |
1395 | |
1396 | // The 2nd half of block MBB after split. |
1397 | MachineBasicBlock *NextMBB; |
1398 | |
1399 | // Add new basic block for the call to the stack overflow function. |
1400 | MachineBasicBlock *StackExtMBB = |
1401 | MF.CreateMachineBasicBlock(BB: MBB.getBasicBlock()); |
1402 | MF.push_back(MBB: StackExtMBB); |
1403 | |
1404 | // LG r3,72(,r3) |
1405 | BuildMI(BB: StackExtMBB, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::LG), DestReg: SystemZ::R3D) |
1406 | .addReg(RegNo: SystemZ::R3D) |
1407 | .addImm(Val: 72) |
1408 | .addReg(RegNo: 0); |
1409 | // BASR r3,r3 |
1410 | BuildMI(BB: StackExtMBB, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::CallBASR_STACKEXT)) |
1411 | .addReg(RegNo: SystemZ::R3D); |
1412 | if (NeedSaveArg) { |
1413 | if (!NeedSaveSP) { |
1414 | // LGR r0,r3 |
1415 | BuildMI(BB&: MBB, I: StackAllocMI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::LGR)) |
1416 | .addReg(RegNo: SystemZ::R0D, flags: RegState::Define) |
1417 | .addReg(RegNo: SystemZ::R3D); |
1418 | } else { |
1419 | // In this case, the incoming value of r4 is saved in r0 so the |
1420 | // latter register is unavailable. Store r3 in its corresponding |
1421 | // slot in the parameter list instead. Do this at the start of |
1422 | // the prolog before r4 is manipulated by anything else. |
1423 | // STG r3, 2192(r4) |
1424 | BuildMI(BB&: MBB, I: MBB.begin(), MIMD: DL, MCID: ZII->get(Opcode: SystemZ::STG)) |
1425 | .addReg(RegNo: SystemZ::R3D) |
1426 | .addReg(RegNo: SystemZ::R4D) |
1427 | .addImm(Val: SaveSlotR3) |
1428 | .addReg(RegNo: 0); |
1429 | } |
1430 | } |
1431 | // LLGT r3,1208 |
1432 | BuildMI(BB&: MBB, I: StackAllocMI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::LLGT), DestReg: SystemZ::R3D) |
1433 | .addReg(RegNo: 0) |
1434 | .addImm(Val: 1208) |
1435 | .addReg(RegNo: 0); |
1436 | // CG r4,64(,r3) |
1437 | BuildMI(BB&: MBB, I: StackAllocMI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::CG)) |
1438 | .addReg(RegNo: SystemZ::R4D) |
1439 | .addReg(RegNo: SystemZ::R3D) |
1440 | .addImm(Val: 64) |
1441 | .addReg(RegNo: 0); |
1442 | // JLL b'0100',F'37' |
1443 | BuildMI(BB&: MBB, I: StackAllocMI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::BRC)) |
1444 | .addImm(Val: SystemZ::CCMASK_ICMP) |
1445 | .addImm(Val: SystemZ::CCMASK_CMP_LT) |
1446 | .addMBB(MBB: StackExtMBB); |
1447 | |
1448 | NextMBB = SystemZ::splitBlockBefore(MI: StackAllocMI, MBB: &MBB); |
1449 | MBB.addSuccessor(Succ: NextMBB); |
1450 | MBB.addSuccessor(Succ: StackExtMBB); |
1451 | if (NeedSaveArg) { |
1452 | if (!NeedSaveSP) { |
1453 | // LGR r3, r0 |
1454 | BuildMI(BB&: *NextMBB, I: StackAllocMI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::LGR)) |
1455 | .addReg(RegNo: SystemZ::R3D, flags: RegState::Define) |
1456 | .addReg(RegNo: SystemZ::R0D, flags: RegState::Kill); |
1457 | } else { |
1458 | // In this case, the incoming value of r4 is saved in r0 so the |
1459 | // latter register is unavailable. We stored r3 in its corresponding |
1460 | // slot in the parameter list instead and we now restore it from there. |
1461 | // LGR r3, r0 |
1462 | BuildMI(BB&: *NextMBB, I: StackAllocMI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::LGR)) |
1463 | .addReg(RegNo: SystemZ::R3D, flags: RegState::Define) |
1464 | .addReg(RegNo: SystemZ::R0D); |
1465 | // LG r3, 2192(r3) |
1466 | BuildMI(BB&: *NextMBB, I: StackAllocMI, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::LG)) |
1467 | .addReg(RegNo: SystemZ::R3D, flags: RegState::Define) |
1468 | .addReg(RegNo: SystemZ::R3D) |
1469 | .addImm(Val: SaveSlotR3) |
1470 | .addReg(RegNo: 0); |
1471 | } |
1472 | } |
1473 | |
1474 | // Add jump back from stack extension BB. |
1475 | BuildMI(BB: StackExtMBB, MIMD: DL, MCID: ZII->get(Opcode: SystemZ::J)).addMBB(MBB: NextMBB); |
1476 | StackExtMBB->addSuccessor(Succ: NextMBB); |
1477 | |
1478 | StackAllocMI->eraseFromParent(); |
1479 | |
1480 | // Compute the live-in lists for the new blocks. |
1481 | fullyRecomputeLiveIns(MBBs: {StackExtMBB, NextMBB}); |
1482 | } |
1483 | |
1484 | bool SystemZXPLINKFrameLowering::hasFPImpl(const MachineFunction &MF) const { |
1485 | return (MF.getFrameInfo().hasVarSizedObjects()); |
1486 | } |
1487 | |
1488 | void SystemZXPLINKFrameLowering::processFunctionBeforeFrameFinalized( |
1489 | MachineFunction &MF, RegScavenger *RS) const { |
1490 | MachineFrameInfo &MFFrame = MF.getFrameInfo(); |
1491 | const SystemZSubtarget &Subtarget = MF.getSubtarget<SystemZSubtarget>(); |
1492 | auto &Regs = Subtarget.getSpecialRegisters<SystemZXPLINK64Registers>(); |
1493 | |
1494 | // Setup stack frame offset |
1495 | MFFrame.setOffsetAdjustment(Regs.getStackPointerBias()); |
1496 | |
1497 | // Nothing to do for leaf functions. |
1498 | uint64_t StackSize = MFFrame.estimateStackSize(MF); |
1499 | if (StackSize == 0 && MFFrame.getCalleeSavedInfo().empty()) |
1500 | return; |
1501 | |
1502 | // Although the XPLINK specifications for AMODE64 state that minimum size |
1503 | // of the param area is minimum 32 bytes and no rounding is otherwise |
1504 | // specified, we round this area in 64 bytes increments to be compatible |
1505 | // with existing compilers. |
1506 | MFFrame.setMaxCallFrameSize( |
1507 | std::max(a: 64U, b: (unsigned)alignTo(Value: MFFrame.getMaxCallFrameSize(), Align: 64))); |
1508 | |
1509 | // Add frame values with positive object offsets. Since the displacement from |
1510 | // the SP/FP is calculated by ObjectOffset + StackSize + Bias, object offsets |
1511 | // with positive values are in the caller's stack frame. We need to include |
1512 | // that since it is accessed by displacement to SP/FP. |
1513 | int64_t LargestArgOffset = 0; |
1514 | for (int I = MFFrame.getObjectIndexBegin(); I != 0; ++I) { |
1515 | if (MFFrame.getObjectOffset(ObjectIdx: I) >= 0) { |
1516 | int64_t ObjOffset = MFFrame.getObjectOffset(ObjectIdx: I) + MFFrame.getObjectSize(ObjectIdx: I); |
1517 | LargestArgOffset = std::max(a: ObjOffset, b: LargestArgOffset); |
1518 | } |
1519 | } |
1520 | |
1521 | uint64_t MaxReach = (StackSize + Regs.getCallFrameSize() + |
1522 | Regs.getStackPointerBias() + LargestArgOffset); |
1523 | |
1524 | if (!isUInt<12>(x: MaxReach)) { |
1525 | // We may need register scavenging slots if some parts of the frame |
1526 | // are outside the reach of an unsigned 12-bit displacement. |
1527 | RS->addScavengingFrameIndex(FI: MFFrame.CreateSpillStackObject(Size: 8, Alignment: Align(8))); |
1528 | RS->addScavengingFrameIndex(FI: MFFrame.CreateSpillStackObject(Size: 8, Alignment: Align(8))); |
1529 | } |
1530 | } |
1531 | |
1532 | // Determines the size of the frame, and creates the deferred spill objects. |
1533 | void SystemZXPLINKFrameLowering::determineFrameLayout( |
1534 | MachineFunction &MF) const { |
1535 | MachineFrameInfo &MFFrame = MF.getFrameInfo(); |
1536 | const SystemZSubtarget &Subtarget = MF.getSubtarget<SystemZSubtarget>(); |
1537 | auto *Regs = |
1538 | static_cast<SystemZXPLINK64Registers *>(Subtarget.getSpecialRegisters()); |
1539 | |
1540 | uint64_t StackSize = MFFrame.getStackSize(); |
1541 | if (StackSize == 0) |
1542 | return; |
1543 | |
1544 | // Add the size of the register save area and the reserved area to the size. |
1545 | StackSize += Regs->getCallFrameSize(); |
1546 | MFFrame.setStackSize(StackSize); |
1547 | |
1548 | // We now know the stack size. Update the stack objects for the register save |
1549 | // area now. This has no impact on the stack frame layout, as this is already |
1550 | // computed. However, it makes sure that all callee saved registers have a |
1551 | // valid offset assigned. |
1552 | for (int FrameIdx = MFFrame.getObjectIndexBegin(); FrameIdx != 0; |
1553 | ++FrameIdx) { |
1554 | if (MFFrame.getStackID(ObjectIdx: FrameIdx) == TargetStackID::NoAlloc) { |
1555 | int64_t SPOffset = MFFrame.getObjectOffset(ObjectIdx: FrameIdx); |
1556 | SPOffset -= StackSize; |
1557 | MFFrame.setObjectOffset(ObjectIdx: FrameIdx, SPOffset); |
1558 | } |
1559 | } |
1560 | } |
1561 | |