1 | //===-- RISCVFrameLowering.cpp - RISC-V Frame Information -----------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file contains the RISC-V implementation of TargetFrameLowering class. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "RISCVFrameLowering.h" |
14 | #include "MCTargetDesc/RISCVBaseInfo.h" |
15 | #include "RISCVMachineFunctionInfo.h" |
16 | #include "RISCVSubtarget.h" |
17 | #include "llvm/BinaryFormat/Dwarf.h" |
18 | #include "llvm/CodeGen/CFIInstBuilder.h" |
19 | #include "llvm/CodeGen/LivePhysRegs.h" |
20 | #include "llvm/CodeGen/MachineFrameInfo.h" |
21 | #include "llvm/CodeGen/MachineFunction.h" |
22 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
23 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
24 | #include "llvm/CodeGen/RegisterScavenging.h" |
25 | #include "llvm/IR/DiagnosticInfo.h" |
26 | #include "llvm/MC/MCDwarf.h" |
27 | #include "llvm/Support/LEB128.h" |
28 | |
29 | #include <algorithm> |
30 | |
31 | #define DEBUG_TYPE "riscv-frame" |
32 | |
33 | using namespace llvm; |
34 | |
35 | static Align getABIStackAlignment(RISCVABI::ABI ABI) { |
36 | if (ABI == RISCVABI::ABI_ILP32E) |
37 | return Align(4); |
38 | if (ABI == RISCVABI::ABI_LP64E) |
39 | return Align(8); |
40 | return Align(16); |
41 | } |
42 | |
43 | RISCVFrameLowering::RISCVFrameLowering(const RISCVSubtarget &STI) |
44 | : TargetFrameLowering( |
45 | StackGrowsDown, getABIStackAlignment(ABI: STI.getTargetABI()), |
46 | /*LocalAreaOffset=*/0, |
47 | /*TransientStackAlignment=*/getABIStackAlignment(ABI: STI.getTargetABI())), |
48 | STI(STI) {} |
49 | |
50 | // The register used to hold the frame pointer. |
51 | static constexpr MCPhysReg FPReg = RISCV::X8; |
52 | |
53 | // The register used to hold the stack pointer. |
54 | static constexpr MCPhysReg SPReg = RISCV::X2; |
55 | |
56 | // The register used to hold the return address. |
57 | static constexpr MCPhysReg RAReg = RISCV::X1; |
58 | |
59 | // LIst of CSRs that are given a fixed location by save/restore libcalls or |
60 | // Zcmp/Xqccmp Push/Pop. The order in this table indicates the order the |
61 | // registers are saved on the stack. Zcmp uses the reverse order of save/restore |
62 | // and Xqccmp on the stack, but this is handled when offsets are calculated. |
63 | static const MCPhysReg FixedCSRFIMap[] = { |
64 | /*ra*/ RAReg, /*s0*/ FPReg, /*s1*/ RISCV::X9, |
65 | /*s2*/ RISCV::X18, /*s3*/ RISCV::X19, /*s4*/ RISCV::X20, |
66 | /*s5*/ RISCV::X21, /*s6*/ RISCV::X22, /*s7*/ RISCV::X23, |
67 | /*s8*/ RISCV::X24, /*s9*/ RISCV::X25, /*s10*/ RISCV::X26, |
68 | /*s11*/ RISCV::X27}; |
69 | |
70 | // The number of stack bytes allocated by `QC.C.MIENTER(.NEST)` and popped by |
71 | // `QC.C.MILEAVERET`. |
72 | static constexpr uint64_t QCIInterruptPushAmount = 96; |
73 | |
74 | static const std::pair<MCPhysReg, int8_t> FixedCSRFIQCIInterruptMap[] = { |
75 | /* -1 is a gap for mepc/mnepc */ |
76 | {/*fp*/ FPReg, -2}, |
77 | /* -3 is a gap for qc.mcause */ |
78 | {/*ra*/ RAReg, -4}, |
79 | /* -5 is reserved */ |
80 | {/*t0*/ RISCV::X5, -6}, |
81 | {/*t1*/ RISCV::X6, -7}, |
82 | {/*t2*/ RISCV::X7, -8}, |
83 | {/*a0*/ RISCV::X10, -9}, |
84 | {/*a1*/ RISCV::X11, -10}, |
85 | {/*a2*/ RISCV::X12, -11}, |
86 | {/*a3*/ RISCV::X13, -12}, |
87 | {/*a4*/ RISCV::X14, -13}, |
88 | {/*a5*/ RISCV::X15, -14}, |
89 | {/*a6*/ RISCV::X16, -15}, |
90 | {/*a7*/ RISCV::X17, -16}, |
91 | {/*t3*/ RISCV::X28, -17}, |
92 | {/*t4*/ RISCV::X29, -18}, |
93 | {/*t5*/ RISCV::X30, -19}, |
94 | {/*t6*/ RISCV::X31, -20}, |
95 | /* -21, -22, -23, -24 are reserved */ |
96 | }; |
97 | |
98 | // For now we use x3, a.k.a gp, as pointer to shadow call stack. |
99 | // User should not use x3 in their asm. |
100 | static void emitSCSPrologue(MachineFunction &MF, MachineBasicBlock &MBB, |
101 | MachineBasicBlock::iterator MI, |
102 | const DebugLoc &DL) { |
103 | const auto &STI = MF.getSubtarget<RISCVSubtarget>(); |
104 | bool HasHWShadowStack = MF.getFunction().hasFnAttribute(Kind: "hw-shadow-stack" ) && |
105 | STI.hasStdExtZicfiss(); |
106 | bool HasSWShadowStack = |
107 | MF.getFunction().hasFnAttribute(Kind: Attribute::ShadowCallStack); |
108 | if (!HasHWShadowStack && !HasSWShadowStack) |
109 | return; |
110 | |
111 | const llvm::RISCVRegisterInfo *TRI = STI.getRegisterInfo(); |
112 | |
113 | // Do not save RA to the SCS if it's not saved to the regular stack, |
114 | // i.e. RA is not at risk of being overwritten. |
115 | std::vector<CalleeSavedInfo> &CSI = MF.getFrameInfo().getCalleeSavedInfo(); |
116 | if (llvm::none_of( |
117 | Range&: CSI, P: [&](CalleeSavedInfo &CSR) { return CSR.getReg() == RAReg; })) |
118 | return; |
119 | |
120 | const RISCVInstrInfo *TII = STI.getInstrInfo(); |
121 | if (HasHWShadowStack) { |
122 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: RISCV::SSPUSH)).addReg(RegNo: RAReg); |
123 | return; |
124 | } |
125 | |
126 | Register SCSPReg = RISCVABI::getSCSPReg(); |
127 | |
128 | bool IsRV64 = STI.is64Bit(); |
129 | int64_t SlotSize = STI.getXLen() / 8; |
130 | // Store return address to shadow call stack |
131 | // addi gp, gp, [4|8] |
132 | // s[w|d] ra, -[4|8](gp) |
133 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: RISCV::ADDI)) |
134 | .addReg(RegNo: SCSPReg, flags: RegState::Define) |
135 | .addReg(RegNo: SCSPReg) |
136 | .addImm(Val: SlotSize) |
137 | .setMIFlag(MachineInstr::FrameSetup); |
138 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: IsRV64 ? RISCV::SD : RISCV::SW)) |
139 | .addReg(RegNo: RAReg) |
140 | .addReg(RegNo: SCSPReg) |
141 | .addImm(Val: -SlotSize) |
142 | .setMIFlag(MachineInstr::FrameSetup); |
143 | |
144 | // Emit a CFI instruction that causes SlotSize to be subtracted from the value |
145 | // of the shadow stack pointer when unwinding past this frame. |
146 | char DwarfSCSReg = TRI->getDwarfRegNum(RegNum: SCSPReg, /*IsEH*/ isEH: true); |
147 | assert(DwarfSCSReg < 32 && "SCS Register should be < 32 (X3)." ); |
148 | |
149 | char Offset = static_cast<char>(-SlotSize) & 0x7f; |
150 | const char CFIInst[] = { |
151 | dwarf::DW_CFA_val_expression, |
152 | DwarfSCSReg, // register |
153 | 2, // length |
154 | static_cast<char>(unsigned(dwarf::DW_OP_breg0 + DwarfSCSReg)), |
155 | Offset, // addend (sleb128) |
156 | }; |
157 | |
158 | CFIInstBuilder(MBB, MI, MachineInstr::FrameSetup) |
159 | .buildEscape(Bytes: StringRef(CFIInst, sizeof(CFIInst))); |
160 | } |
161 | |
162 | static void emitSCSEpilogue(MachineFunction &MF, MachineBasicBlock &MBB, |
163 | MachineBasicBlock::iterator MI, |
164 | const DebugLoc &DL) { |
165 | const auto &STI = MF.getSubtarget<RISCVSubtarget>(); |
166 | bool HasHWShadowStack = MF.getFunction().hasFnAttribute(Kind: "hw-shadow-stack" ) && |
167 | STI.hasStdExtZicfiss(); |
168 | bool HasSWShadowStack = |
169 | MF.getFunction().hasFnAttribute(Kind: Attribute::ShadowCallStack); |
170 | if (!HasHWShadowStack && !HasSWShadowStack) |
171 | return; |
172 | |
173 | // See emitSCSPrologue() above. |
174 | std::vector<CalleeSavedInfo> &CSI = MF.getFrameInfo().getCalleeSavedInfo(); |
175 | if (llvm::none_of( |
176 | Range&: CSI, P: [&](CalleeSavedInfo &CSR) { return CSR.getReg() == RAReg; })) |
177 | return; |
178 | |
179 | const RISCVInstrInfo *TII = STI.getInstrInfo(); |
180 | if (HasHWShadowStack) { |
181 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: RISCV::SSPOPCHK)).addReg(RegNo: RAReg); |
182 | return; |
183 | } |
184 | |
185 | Register SCSPReg = RISCVABI::getSCSPReg(); |
186 | |
187 | bool IsRV64 = STI.is64Bit(); |
188 | int64_t SlotSize = STI.getXLen() / 8; |
189 | // Load return address from shadow call stack |
190 | // l[w|d] ra, -[4|8](gp) |
191 | // addi gp, gp, -[4|8] |
192 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: IsRV64 ? RISCV::LD : RISCV::LW)) |
193 | .addReg(RegNo: RAReg, flags: RegState::Define) |
194 | .addReg(RegNo: SCSPReg) |
195 | .addImm(Val: -SlotSize) |
196 | .setMIFlag(MachineInstr::FrameDestroy); |
197 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: RISCV::ADDI)) |
198 | .addReg(RegNo: SCSPReg, flags: RegState::Define) |
199 | .addReg(RegNo: SCSPReg) |
200 | .addImm(Val: -SlotSize) |
201 | .setMIFlag(MachineInstr::FrameDestroy); |
202 | // Restore the SCS pointer |
203 | CFIInstBuilder(MBB, MI, MachineInstr::FrameDestroy).buildRestore(Reg: SCSPReg); |
204 | } |
205 | |
206 | // Insert instruction to swap mscratchsw with sp |
207 | static void emitSiFiveCLICStackSwap(MachineFunction &MF, MachineBasicBlock &MBB, |
208 | MachineBasicBlock::iterator MBBI, |
209 | const DebugLoc &DL) { |
210 | auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
211 | |
212 | if (!RVFI->isSiFiveStackSwapInterrupt(MF)) |
213 | return; |
214 | |
215 | const auto &STI = MF.getSubtarget<RISCVSubtarget>(); |
216 | const RISCVInstrInfo *TII = STI.getInstrInfo(); |
217 | |
218 | assert(STI.hasVendorXSfmclic() && "Stack Swapping Requires XSfmclic" ); |
219 | |
220 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: RISCV::CSRRW)) |
221 | .addReg(RegNo: SPReg, flags: RegState::Define) |
222 | .addImm(Val: RISCVSysReg::sf_mscratchcsw) |
223 | .addReg(RegNo: SPReg, flags: RegState::Kill) |
224 | .setMIFlag(MachineInstr::FrameSetup); |
225 | |
226 | // FIXME: CFI Information for this swap. |
227 | } |
228 | |
229 | static void |
230 | createSiFivePreemptibleInterruptFrameEntries(MachineFunction &MF, |
231 | RISCVMachineFunctionInfo &RVFI) { |
232 | if (!RVFI.isSiFivePreemptibleInterrupt(MF)) |
233 | return; |
234 | |
235 | const TargetRegisterClass &RC = RISCV::GPRRegClass; |
236 | const TargetRegisterInfo &TRI = |
237 | *MF.getSubtarget<RISCVSubtarget>().getRegisterInfo(); |
238 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
239 | |
240 | // Create two frame objects for spilling X8 and X9, which will be done in |
241 | // `emitSiFiveCLICPreemptibleSaves`. This is in addition to any other stack |
242 | // objects we might have for X8 and X9, as they might be saved twice. |
243 | for (int I = 0; I < 2; ++I) { |
244 | int FI = MFI.CreateStackObject(Size: TRI.getSpillSize(RC), Alignment: TRI.getSpillAlign(RC), |
245 | isSpillSlot: true); |
246 | RVFI.pushInterruptCSRFrameIndex(FI); |
247 | } |
248 | } |
249 | |
250 | static void emitSiFiveCLICPreemptibleSaves(MachineFunction &MF, |
251 | MachineBasicBlock &MBB, |
252 | MachineBasicBlock::iterator MBBI, |
253 | const DebugLoc &DL) { |
254 | auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
255 | |
256 | if (!RVFI->isSiFivePreemptibleInterrupt(MF)) |
257 | return; |
258 | |
259 | const auto &STI = MF.getSubtarget<RISCVSubtarget>(); |
260 | const RISCVInstrInfo *TII = STI.getInstrInfo(); |
261 | |
262 | // FIXME: CFI Information here is nonexistent/wrong. |
263 | |
264 | // X8 and X9 might be stored into the stack twice, initially into the |
265 | // `interruptCSRFrameIndex` here, and then maybe again into their CSI frame |
266 | // index. |
267 | // |
268 | // This is done instead of telling the register allocator that we need two |
269 | // VRegs to store the value of `mcause` and `mepc` through the instruction, |
270 | // which affects other passes. |
271 | TII->storeRegToStackSlot(MBB, MBBI, SrcReg: RISCV::X8, /* IsKill=*/true, |
272 | FrameIndex: RVFI->getInterruptCSRFrameIndex(Idx: 0), |
273 | RC: &RISCV::GPRRegClass, TRI: STI.getRegisterInfo(), |
274 | VReg: Register(), Flags: MachineInstr::FrameSetup); |
275 | TII->storeRegToStackSlot(MBB, MBBI, SrcReg: RISCV::X9, /* IsKill=*/true, |
276 | FrameIndex: RVFI->getInterruptCSRFrameIndex(Idx: 1), |
277 | RC: &RISCV::GPRRegClass, TRI: STI.getRegisterInfo(), |
278 | VReg: Register(), Flags: MachineInstr::FrameSetup); |
279 | |
280 | // Put `mcause` into X8 (s0), and `mepc` into X9 (s1). If either of these are |
281 | // used in the function, then they will appear in `getUnmanagedCSI` and will |
282 | // be saved again. |
283 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: RISCV::CSRRS)) |
284 | .addReg(RegNo: RISCV::X8, flags: RegState::Define) |
285 | .addImm(Val: RISCVSysReg::mcause) |
286 | .addReg(RegNo: RISCV::X0) |
287 | .setMIFlag(MachineInstr::FrameSetup); |
288 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: RISCV::CSRRS)) |
289 | .addReg(RegNo: RISCV::X9, flags: RegState::Define) |
290 | .addImm(Val: RISCVSysReg::mepc) |
291 | .addReg(RegNo: RISCV::X0) |
292 | .setMIFlag(MachineInstr::FrameSetup); |
293 | |
294 | // Enable interrupts. |
295 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: RISCV::CSRRSI)) |
296 | .addReg(RegNo: RISCV::X0, flags: RegState::Define) |
297 | .addImm(Val: RISCVSysReg::mstatus) |
298 | .addImm(Val: 8) |
299 | .setMIFlag(MachineInstr::FrameSetup); |
300 | } |
301 | |
302 | static void emitSiFiveCLICPreemptibleRestores(MachineFunction &MF, |
303 | MachineBasicBlock &MBB, |
304 | MachineBasicBlock::iterator MBBI, |
305 | const DebugLoc &DL) { |
306 | auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
307 | |
308 | if (!RVFI->isSiFivePreemptibleInterrupt(MF)) |
309 | return; |
310 | |
311 | const auto &STI = MF.getSubtarget<RISCVSubtarget>(); |
312 | const RISCVInstrInfo *TII = STI.getInstrInfo(); |
313 | |
314 | // FIXME: CFI Information here is nonexistent/wrong. |
315 | |
316 | // Disable interrupts. |
317 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: RISCV::CSRRCI)) |
318 | .addReg(RegNo: RISCV::X0, flags: RegState::Define) |
319 | .addImm(Val: RISCVSysReg::mstatus) |
320 | .addImm(Val: 8) |
321 | .setMIFlag(MachineInstr::FrameSetup); |
322 | |
323 | // Restore `mepc` from x9 (s1), and `mcause` from x8 (s0). If either were used |
324 | // in the function, they have already been restored once, so now have the |
325 | // value stored in `emitSiFiveCLICPreemptibleSaves`. |
326 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: RISCV::CSRRW)) |
327 | .addReg(RegNo: RISCV::X0, flags: RegState::Define) |
328 | .addImm(Val: RISCVSysReg::mepc) |
329 | .addReg(RegNo: RISCV::X9, flags: RegState::Kill) |
330 | .setMIFlag(MachineInstr::FrameSetup); |
331 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: RISCV::CSRRW)) |
332 | .addReg(RegNo: RISCV::X0, flags: RegState::Define) |
333 | .addImm(Val: RISCVSysReg::mcause) |
334 | .addReg(RegNo: RISCV::X8, flags: RegState::Kill) |
335 | .setMIFlag(MachineInstr::FrameSetup); |
336 | |
337 | // X8 and X9 need to be restored to their values on function entry, which we |
338 | // saved onto the stack in `emitSiFiveCLICPreemptibleSaves`. |
339 | TII->loadRegFromStackSlot(MBB, MBBI, DstReg: RISCV::X9, |
340 | FrameIndex: RVFI->getInterruptCSRFrameIndex(Idx: 1), |
341 | RC: &RISCV::GPRRegClass, TRI: STI.getRegisterInfo(), |
342 | VReg: Register(), Flags: MachineInstr::FrameSetup); |
343 | TII->loadRegFromStackSlot(MBB, MBBI, DstReg: RISCV::X8, |
344 | FrameIndex: RVFI->getInterruptCSRFrameIndex(Idx: 0), |
345 | RC: &RISCV::GPRRegClass, TRI: STI.getRegisterInfo(), |
346 | VReg: Register(), Flags: MachineInstr::FrameSetup); |
347 | } |
348 | |
349 | // Get the ID of the libcall used for spilling and restoring callee saved |
350 | // registers. The ID is representative of the number of registers saved or |
351 | // restored by the libcall, except it is zero-indexed - ID 0 corresponds to a |
352 | // single register. |
353 | static int getLibCallID(const MachineFunction &MF, |
354 | const std::vector<CalleeSavedInfo> &CSI) { |
355 | const auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
356 | |
357 | if (CSI.empty() || !RVFI->useSaveRestoreLibCalls(MF)) |
358 | return -1; |
359 | |
360 | MCRegister MaxReg; |
361 | for (auto &CS : CSI) |
362 | // assignCalleeSavedSpillSlots assigns negative frame indexes to |
363 | // registers which can be saved by libcall. |
364 | if (CS.getFrameIdx() < 0) |
365 | MaxReg = std::max(a: MaxReg.id(), b: CS.getReg().id()); |
366 | |
367 | if (!MaxReg) |
368 | return -1; |
369 | |
370 | switch (MaxReg.id()) { |
371 | default: |
372 | llvm_unreachable("Something has gone wrong!" ); |
373 | // clang-format off |
374 | case /*s11*/ RISCV::X27: return 12; |
375 | case /*s10*/ RISCV::X26: return 11; |
376 | case /*s9*/ RISCV::X25: return 10; |
377 | case /*s8*/ RISCV::X24: return 9; |
378 | case /*s7*/ RISCV::X23: return 8; |
379 | case /*s6*/ RISCV::X22: return 7; |
380 | case /*s5*/ RISCV::X21: return 6; |
381 | case /*s4*/ RISCV::X20: return 5; |
382 | case /*s3*/ RISCV::X19: return 4; |
383 | case /*s2*/ RISCV::X18: return 3; |
384 | case /*s1*/ RISCV::X9: return 2; |
385 | case /*s0*/ FPReg: return 1; |
386 | case /*ra*/ RAReg: return 0; |
387 | // clang-format on |
388 | } |
389 | } |
390 | |
391 | // Get the name of the libcall used for spilling callee saved registers. |
392 | // If this function will not use save/restore libcalls, then return a nullptr. |
393 | static const char * |
394 | getSpillLibCallName(const MachineFunction &MF, |
395 | const std::vector<CalleeSavedInfo> &CSI) { |
396 | static const char *const SpillLibCalls[] = { |
397 | "__riscv_save_0" , |
398 | "__riscv_save_1" , |
399 | "__riscv_save_2" , |
400 | "__riscv_save_3" , |
401 | "__riscv_save_4" , |
402 | "__riscv_save_5" , |
403 | "__riscv_save_6" , |
404 | "__riscv_save_7" , |
405 | "__riscv_save_8" , |
406 | "__riscv_save_9" , |
407 | "__riscv_save_10" , |
408 | "__riscv_save_11" , |
409 | "__riscv_save_12" |
410 | }; |
411 | |
412 | int LibCallID = getLibCallID(MF, CSI); |
413 | if (LibCallID == -1) |
414 | return nullptr; |
415 | return SpillLibCalls[LibCallID]; |
416 | } |
417 | |
418 | // Get the name of the libcall used for restoring callee saved registers. |
419 | // If this function will not use save/restore libcalls, then return a nullptr. |
420 | static const char * |
421 | getRestoreLibCallName(const MachineFunction &MF, |
422 | const std::vector<CalleeSavedInfo> &CSI) { |
423 | static const char *const RestoreLibCalls[] = { |
424 | "__riscv_restore_0" , |
425 | "__riscv_restore_1" , |
426 | "__riscv_restore_2" , |
427 | "__riscv_restore_3" , |
428 | "__riscv_restore_4" , |
429 | "__riscv_restore_5" , |
430 | "__riscv_restore_6" , |
431 | "__riscv_restore_7" , |
432 | "__riscv_restore_8" , |
433 | "__riscv_restore_9" , |
434 | "__riscv_restore_10" , |
435 | "__riscv_restore_11" , |
436 | "__riscv_restore_12" |
437 | }; |
438 | |
439 | int LibCallID = getLibCallID(MF, CSI); |
440 | if (LibCallID == -1) |
441 | return nullptr; |
442 | return RestoreLibCalls[LibCallID]; |
443 | } |
444 | |
445 | // Get the max reg of Push/Pop for restoring callee saved registers. |
446 | static unsigned getNumPushPopRegs(const std::vector<CalleeSavedInfo> &CSI) { |
447 | unsigned NumPushPopRegs = 0; |
448 | for (auto &CS : CSI) { |
449 | auto *FII = llvm::find_if(Range: FixedCSRFIMap, |
450 | P: [&](MCPhysReg P) { return P == CS.getReg(); }); |
451 | if (FII != std::end(arr: FixedCSRFIMap)) { |
452 | unsigned RegNum = std::distance(first: std::begin(arr: FixedCSRFIMap), last: FII); |
453 | NumPushPopRegs = std::max(a: NumPushPopRegs, b: RegNum + 1); |
454 | } |
455 | } |
456 | assert(NumPushPopRegs != 12 && "x26 requires x27 to also be pushed" ); |
457 | return NumPushPopRegs; |
458 | } |
459 | |
460 | // Return true if the specified function should have a dedicated frame |
461 | // pointer register. This is true if frame pointer elimination is |
462 | // disabled, if it needs dynamic stack realignment, if the function has |
463 | // variable sized allocas, or if the frame address is taken. |
464 | bool RISCVFrameLowering::hasFPImpl(const MachineFunction &MF) const { |
465 | const TargetRegisterInfo *RegInfo = MF.getSubtarget().getRegisterInfo(); |
466 | |
467 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
468 | return MF.getTarget().Options.DisableFramePointerElim(MF) || |
469 | RegInfo->hasStackRealignment(MF) || MFI.hasVarSizedObjects() || |
470 | MFI.isFrameAddressTaken(); |
471 | } |
472 | |
473 | bool RISCVFrameLowering::hasBP(const MachineFunction &MF) const { |
474 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
475 | const TargetRegisterInfo *TRI = STI.getRegisterInfo(); |
476 | |
477 | // If we do not reserve stack space for outgoing arguments in prologue, |
478 | // we will adjust the stack pointer before call instruction. After the |
479 | // adjustment, we can not use SP to access the stack objects for the |
480 | // arguments. Instead, use BP to access these stack objects. |
481 | return (MFI.hasVarSizedObjects() || |
482 | (!hasReservedCallFrame(MF) && (!MFI.isMaxCallFrameSizeComputed() || |
483 | MFI.getMaxCallFrameSize() != 0))) && |
484 | TRI->hasStackRealignment(MF); |
485 | } |
486 | |
487 | // Determines the size of the frame and maximum call frame size. |
488 | void RISCVFrameLowering::determineFrameLayout(MachineFunction &MF) const { |
489 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
490 | auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
491 | |
492 | // Get the number of bytes to allocate from the FrameInfo. |
493 | uint64_t FrameSize = MFI.getStackSize(); |
494 | |
495 | // QCI Interrupts use at least 96 bytes of stack space |
496 | if (RVFI->useQCIInterrupt(MF)) |
497 | FrameSize = std::max(a: FrameSize, b: QCIInterruptPushAmount); |
498 | |
499 | // Get the alignment. |
500 | Align StackAlign = getStackAlign(); |
501 | |
502 | // Make sure the frame is aligned. |
503 | FrameSize = alignTo(Size: FrameSize, A: StackAlign); |
504 | |
505 | // Update frame info. |
506 | MFI.setStackSize(FrameSize); |
507 | |
508 | // When using SP or BP to access stack objects, we may require extra padding |
509 | // to ensure the bottom of the RVV stack is correctly aligned within the main |
510 | // stack. We calculate this as the amount required to align the scalar local |
511 | // variable section up to the RVV alignment. |
512 | const TargetRegisterInfo *TRI = STI.getRegisterInfo(); |
513 | if (RVFI->getRVVStackSize() && (!hasFP(MF) || TRI->hasStackRealignment(MF))) { |
514 | int ScalarLocalVarSize = FrameSize - RVFI->getCalleeSavedStackSize() - |
515 | RVFI->getVarArgsSaveSize(); |
516 | if (auto RVVPadding = |
517 | offsetToAlignment(Value: ScalarLocalVarSize, Alignment: RVFI->getRVVStackAlign())) |
518 | RVFI->setRVVPadding(RVVPadding); |
519 | } |
520 | } |
521 | |
522 | // Returns the stack size including RVV padding (when required), rounded back |
523 | // up to the required stack alignment. |
524 | uint64_t RISCVFrameLowering::getStackSizeWithRVVPadding( |
525 | const MachineFunction &MF) const { |
526 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
527 | auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
528 | return alignTo(Size: MFI.getStackSize() + RVFI->getRVVPadding(), A: getStackAlign()); |
529 | } |
530 | |
531 | static SmallVector<CalleeSavedInfo, 8> |
532 | getUnmanagedCSI(const MachineFunction &MF, |
533 | const std::vector<CalleeSavedInfo> &CSI) { |
534 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
535 | SmallVector<CalleeSavedInfo, 8> NonLibcallCSI; |
536 | |
537 | for (auto &CS : CSI) { |
538 | int FI = CS.getFrameIdx(); |
539 | if (FI >= 0 && MFI.getStackID(ObjectIdx: FI) == TargetStackID::Default) |
540 | NonLibcallCSI.push_back(Elt: CS); |
541 | } |
542 | |
543 | return NonLibcallCSI; |
544 | } |
545 | |
546 | static SmallVector<CalleeSavedInfo, 8> |
547 | getRVVCalleeSavedInfo(const MachineFunction &MF, |
548 | const std::vector<CalleeSavedInfo> &CSI) { |
549 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
550 | SmallVector<CalleeSavedInfo, 8> RVVCSI; |
551 | |
552 | for (auto &CS : CSI) { |
553 | int FI = CS.getFrameIdx(); |
554 | if (FI >= 0 && MFI.getStackID(ObjectIdx: FI) == TargetStackID::ScalableVector) |
555 | RVVCSI.push_back(Elt: CS); |
556 | } |
557 | |
558 | return RVVCSI; |
559 | } |
560 | |
561 | static SmallVector<CalleeSavedInfo, 8> |
562 | getPushOrLibCallsSavedInfo(const MachineFunction &MF, |
563 | const std::vector<CalleeSavedInfo> &CSI) { |
564 | auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
565 | |
566 | SmallVector<CalleeSavedInfo, 8> PushOrLibCallsCSI; |
567 | if (!RVFI->useSaveRestoreLibCalls(MF) && !RVFI->isPushable(MF)) |
568 | return PushOrLibCallsCSI; |
569 | |
570 | for (const auto &CS : CSI) { |
571 | if (RVFI->useQCIInterrupt(MF)) { |
572 | // Some registers are saved by both `QC.C.MIENTER(.NEST)` and |
573 | // `QC.CM.PUSH(FP)`. In these cases, prioritise the CFI info that points |
574 | // to the versions saved by `QC.C.MIENTER(.NEST)` which is what FP |
575 | // unwinding would use. |
576 | if (llvm::is_contained(Range: llvm::make_first_range(c: FixedCSRFIQCIInterruptMap), |
577 | Element: CS.getReg())) |
578 | continue; |
579 | } |
580 | |
581 | if (llvm::is_contained(Range: FixedCSRFIMap, Element: CS.getReg())) |
582 | PushOrLibCallsCSI.push_back(Elt: CS); |
583 | } |
584 | |
585 | return PushOrLibCallsCSI; |
586 | } |
587 | |
588 | static SmallVector<CalleeSavedInfo, 8> |
589 | getQCISavedInfo(const MachineFunction &MF, |
590 | const std::vector<CalleeSavedInfo> &CSI) { |
591 | auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
592 | |
593 | SmallVector<CalleeSavedInfo, 8> QCIInterruptCSI; |
594 | if (!RVFI->useQCIInterrupt(MF)) |
595 | return QCIInterruptCSI; |
596 | |
597 | for (const auto &CS : CSI) { |
598 | if (llvm::is_contained(Range: llvm::make_first_range(c: FixedCSRFIQCIInterruptMap), |
599 | Element: CS.getReg())) |
600 | QCIInterruptCSI.push_back(Elt: CS); |
601 | } |
602 | |
603 | return QCIInterruptCSI; |
604 | } |
605 | |
606 | void RISCVFrameLowering::allocateAndProbeStackForRVV( |
607 | MachineFunction &MF, MachineBasicBlock &MBB, |
608 | MachineBasicBlock::iterator MBBI, const DebugLoc &DL, int64_t Amount, |
609 | MachineInstr::MIFlag Flag, bool EmitCFI, bool DynAllocation) const { |
610 | assert(Amount != 0 && "Did not need to adjust stack pointer for RVV." ); |
611 | |
612 | // Emit a variable-length allocation probing loop. |
613 | |
614 | // Get VLEN in TargetReg |
615 | const RISCVInstrInfo *TII = STI.getInstrInfo(); |
616 | Register TargetReg = RISCV::X6; |
617 | uint32_t NumOfVReg = Amount / RISCV::RVVBytesPerBlock; |
618 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: RISCV::PseudoReadVLENB), DestReg: TargetReg) |
619 | .setMIFlag(Flag); |
620 | TII->mulImm(MF, MBB, II: MBBI, DL, DestReg: TargetReg, Amt: NumOfVReg, Flag); |
621 | |
622 | CFIInstBuilder CFIBuilder(MBB, MBBI, MachineInstr::FrameSetup); |
623 | if (EmitCFI) { |
624 | // Set the CFA register to TargetReg. |
625 | CFIBuilder.buildDefCFA(Reg: TargetReg, Offset: -Amount); |
626 | } |
627 | |
628 | // It will be expanded to a probe loop in `inlineStackProbe`. |
629 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: RISCV::PROBED_STACKALLOC_RVV)) |
630 | .addReg(RegNo: TargetReg); |
631 | |
632 | if (EmitCFI) { |
633 | // Set the CFA register back to SP. |
634 | CFIBuilder.buildDefCFARegister(Reg: SPReg); |
635 | } |
636 | |
637 | // SUB SP, SP, T1 |
638 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: RISCV::SUB), DestReg: SPReg) |
639 | .addReg(RegNo: SPReg) |
640 | .addReg(RegNo: TargetReg) |
641 | .setMIFlag(Flag); |
642 | |
643 | // If we have a dynamic allocation later we need to probe any residuals. |
644 | if (DynAllocation) { |
645 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: STI.is64Bit() ? RISCV::SD : RISCV::SW)) |
646 | .addReg(RegNo: RISCV::X0) |
647 | .addReg(RegNo: SPReg) |
648 | .addImm(Val: 0) |
649 | .setMIFlags(MachineInstr::FrameSetup); |
650 | } |
651 | } |
652 | |
653 | static void appendScalableVectorExpression(const TargetRegisterInfo &TRI, |
654 | SmallVectorImpl<char> &Expr, |
655 | int FixedOffset, int ScalableOffset, |
656 | llvm::raw_string_ostream &) { |
657 | unsigned DwarfVLenB = TRI.getDwarfRegNum(RegNum: RISCV::VLENB, isEH: true); |
658 | uint8_t Buffer[16]; |
659 | if (FixedOffset) { |
660 | Expr.push_back(Elt: dwarf::DW_OP_consts); |
661 | Expr.append(in_start: Buffer, in_end: Buffer + encodeSLEB128(Value: FixedOffset, p: Buffer)); |
662 | Expr.push_back(Elt: (uint8_t)dwarf::DW_OP_plus); |
663 | Comment << (FixedOffset < 0 ? " - " : " + " ) << std::abs(x: FixedOffset); |
664 | } |
665 | |
666 | Expr.push_back(Elt: (uint8_t)dwarf::DW_OP_consts); |
667 | Expr.append(in_start: Buffer, in_end: Buffer + encodeSLEB128(Value: ScalableOffset, p: Buffer)); |
668 | |
669 | Expr.push_back(Elt: (uint8_t)dwarf::DW_OP_bregx); |
670 | Expr.append(in_start: Buffer, in_end: Buffer + encodeULEB128(Value: DwarfVLenB, p: Buffer)); |
671 | Expr.push_back(Elt: 0); |
672 | |
673 | Expr.push_back(Elt: (uint8_t)dwarf::DW_OP_mul); |
674 | Expr.push_back(Elt: (uint8_t)dwarf::DW_OP_plus); |
675 | |
676 | Comment << (ScalableOffset < 0 ? " - " : " + " ) << std::abs(x: ScalableOffset) |
677 | << " * vlenb" ; |
678 | } |
679 | |
680 | static MCCFIInstruction createDefCFAExpression(const TargetRegisterInfo &TRI, |
681 | Register Reg, |
682 | uint64_t FixedOffset, |
683 | uint64_t ScalableOffset) { |
684 | assert(ScalableOffset != 0 && "Did not need to adjust CFA for RVV" ); |
685 | SmallString<64> Expr; |
686 | std::string ; |
687 | llvm::raw_string_ostream (CommentBuffer); |
688 | // Build up the expression (Reg + FixedOffset + ScalableOffset * VLENB). |
689 | unsigned DwarfReg = TRI.getDwarfRegNum(RegNum: Reg, isEH: true); |
690 | Expr.push_back(Elt: (uint8_t)(dwarf::DW_OP_breg0 + DwarfReg)); |
691 | Expr.push_back(Elt: 0); |
692 | if (Reg == SPReg) |
693 | Comment << "sp" ; |
694 | else |
695 | Comment << printReg(Reg, TRI: &TRI); |
696 | |
697 | appendScalableVectorExpression(TRI, Expr, FixedOffset, ScalableOffset, |
698 | Comment); |
699 | |
700 | SmallString<64> DefCfaExpr; |
701 | uint8_t Buffer[16]; |
702 | DefCfaExpr.push_back(Elt: dwarf::DW_CFA_def_cfa_expression); |
703 | DefCfaExpr.append(in_start: Buffer, in_end: Buffer + encodeULEB128(Value: Expr.size(), p: Buffer)); |
704 | DefCfaExpr.append(RHS: Expr.str()); |
705 | |
706 | return MCCFIInstruction::createEscape(L: nullptr, Vals: DefCfaExpr.str(), Loc: SMLoc(), |
707 | Comment: Comment.str()); |
708 | } |
709 | |
710 | static MCCFIInstruction createDefCFAOffset(const TargetRegisterInfo &TRI, |
711 | Register Reg, uint64_t FixedOffset, |
712 | uint64_t ScalableOffset) { |
713 | assert(ScalableOffset != 0 && "Did not need to adjust CFA for RVV" ); |
714 | SmallString<64> Expr; |
715 | std::string ; |
716 | llvm::raw_string_ostream (CommentBuffer); |
717 | Comment << printReg(Reg, TRI: &TRI) << " @ cfa" ; |
718 | |
719 | // Build up the expression (FixedOffset + ScalableOffset * VLENB). |
720 | appendScalableVectorExpression(TRI, Expr, FixedOffset, ScalableOffset, |
721 | Comment); |
722 | |
723 | SmallString<64> DefCfaExpr; |
724 | uint8_t Buffer[16]; |
725 | unsigned DwarfReg = TRI.getDwarfRegNum(RegNum: Reg, isEH: true); |
726 | DefCfaExpr.push_back(Elt: dwarf::DW_CFA_expression); |
727 | DefCfaExpr.append(in_start: Buffer, in_end: Buffer + encodeULEB128(Value: DwarfReg, p: Buffer)); |
728 | DefCfaExpr.append(in_start: Buffer, in_end: Buffer + encodeULEB128(Value: Expr.size(), p: Buffer)); |
729 | DefCfaExpr.append(RHS: Expr.str()); |
730 | |
731 | return MCCFIInstruction::createEscape(L: nullptr, Vals: DefCfaExpr.str(), Loc: SMLoc(), |
732 | Comment: Comment.str()); |
733 | } |
734 | |
735 | // Allocate stack space and probe it if necessary. |
736 | void RISCVFrameLowering::allocateStack(MachineBasicBlock &MBB, |
737 | MachineBasicBlock::iterator MBBI, |
738 | MachineFunction &MF, uint64_t Offset, |
739 | uint64_t RealStackSize, bool EmitCFI, |
740 | bool NeedProbe, uint64_t ProbeSize, |
741 | bool DynAllocation) const { |
742 | DebugLoc DL; |
743 | const RISCVRegisterInfo *RI = STI.getRegisterInfo(); |
744 | const RISCVInstrInfo *TII = STI.getInstrInfo(); |
745 | bool IsRV64 = STI.is64Bit(); |
746 | CFIInstBuilder CFIBuilder(MBB, MBBI, MachineInstr::FrameSetup); |
747 | |
748 | // Simply allocate the stack if it's not big enough to require a probe. |
749 | if (!NeedProbe || Offset <= ProbeSize) { |
750 | RI->adjustReg(MBB, II: MBBI, DL, DestReg: SPReg, SrcReg: SPReg, Offset: StackOffset::getFixed(Fixed: -Offset), |
751 | Flag: MachineInstr::FrameSetup, RequiredAlign: getStackAlign()); |
752 | |
753 | if (EmitCFI) |
754 | CFIBuilder.buildDefCFAOffset(Offset: RealStackSize); |
755 | |
756 | if (NeedProbe && DynAllocation) { |
757 | // s[d|w] zero, 0(sp) |
758 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: IsRV64 ? RISCV::SD : RISCV::SW)) |
759 | .addReg(RegNo: RISCV::X0) |
760 | .addReg(RegNo: SPReg) |
761 | .addImm(Val: 0) |
762 | .setMIFlags(MachineInstr::FrameSetup); |
763 | } |
764 | |
765 | return; |
766 | } |
767 | |
768 | // Unroll the probe loop depending on the number of iterations. |
769 | if (Offset < ProbeSize * 5) { |
770 | uint64_t CurrentOffset = 0; |
771 | while (CurrentOffset + ProbeSize <= Offset) { |
772 | RI->adjustReg(MBB, II: MBBI, DL, DestReg: SPReg, SrcReg: SPReg, |
773 | Offset: StackOffset::getFixed(Fixed: -ProbeSize), Flag: MachineInstr::FrameSetup, |
774 | RequiredAlign: getStackAlign()); |
775 | // s[d|w] zero, 0(sp) |
776 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: IsRV64 ? RISCV::SD : RISCV::SW)) |
777 | .addReg(RegNo: RISCV::X0) |
778 | .addReg(RegNo: SPReg) |
779 | .addImm(Val: 0) |
780 | .setMIFlags(MachineInstr::FrameSetup); |
781 | |
782 | CurrentOffset += ProbeSize; |
783 | if (EmitCFI) |
784 | CFIBuilder.buildDefCFAOffset(Offset: CurrentOffset); |
785 | } |
786 | |
787 | uint64_t Residual = Offset - CurrentOffset; |
788 | if (Residual) { |
789 | RI->adjustReg(MBB, II: MBBI, DL, DestReg: SPReg, SrcReg: SPReg, |
790 | Offset: StackOffset::getFixed(Fixed: -Residual), Flag: MachineInstr::FrameSetup, |
791 | RequiredAlign: getStackAlign()); |
792 | if (EmitCFI) |
793 | CFIBuilder.buildDefCFAOffset(Offset); |
794 | |
795 | if (DynAllocation) { |
796 | // s[d|w] zero, 0(sp) |
797 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: IsRV64 ? RISCV::SD : RISCV::SW)) |
798 | .addReg(RegNo: RISCV::X0) |
799 | .addReg(RegNo: SPReg) |
800 | .addImm(Val: 0) |
801 | .setMIFlags(MachineInstr::FrameSetup); |
802 | } |
803 | } |
804 | |
805 | return; |
806 | } |
807 | |
808 | // Emit a variable-length allocation probing loop. |
809 | uint64_t RoundedSize = alignDown(Value: Offset, Align: ProbeSize); |
810 | uint64_t Residual = Offset - RoundedSize; |
811 | |
812 | Register TargetReg = RISCV::X6; |
813 | // SUB TargetReg, SP, RoundedSize |
814 | RI->adjustReg(MBB, II: MBBI, DL, DestReg: TargetReg, SrcReg: SPReg, |
815 | Offset: StackOffset::getFixed(Fixed: -RoundedSize), Flag: MachineInstr::FrameSetup, |
816 | RequiredAlign: getStackAlign()); |
817 | |
818 | if (EmitCFI) { |
819 | // Set the CFA register to TargetReg. |
820 | CFIBuilder.buildDefCFA(Reg: TargetReg, Offset: RoundedSize); |
821 | } |
822 | |
823 | // It will be expanded to a probe loop in `inlineStackProbe`. |
824 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: RISCV::PROBED_STACKALLOC)).addReg(RegNo: TargetReg); |
825 | |
826 | if (EmitCFI) { |
827 | // Set the CFA register back to SP. |
828 | CFIBuilder.buildDefCFARegister(Reg: SPReg); |
829 | } |
830 | |
831 | if (Residual) { |
832 | RI->adjustReg(MBB, II: MBBI, DL, DestReg: SPReg, SrcReg: SPReg, Offset: StackOffset::getFixed(Fixed: -Residual), |
833 | Flag: MachineInstr::FrameSetup, RequiredAlign: getStackAlign()); |
834 | if (DynAllocation) { |
835 | // s[d|w] zero, 0(sp) |
836 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: IsRV64 ? RISCV::SD : RISCV::SW)) |
837 | .addReg(RegNo: RISCV::X0) |
838 | .addReg(RegNo: SPReg) |
839 | .addImm(Val: 0) |
840 | .setMIFlags(MachineInstr::FrameSetup); |
841 | } |
842 | } |
843 | |
844 | if (EmitCFI) |
845 | CFIBuilder.buildDefCFAOffset(Offset); |
846 | } |
847 | |
848 | static bool isPush(unsigned Opcode) { |
849 | switch (Opcode) { |
850 | case RISCV::CM_PUSH: |
851 | case RISCV::QC_CM_PUSH: |
852 | case RISCV::QC_CM_PUSHFP: |
853 | return true; |
854 | default: |
855 | return false; |
856 | } |
857 | } |
858 | |
859 | static bool isPop(unsigned Opcode) { |
860 | // There are other pops but these are the only ones introduced during this |
861 | // pass. |
862 | switch (Opcode) { |
863 | case RISCV::CM_POP: |
864 | case RISCV::QC_CM_POP: |
865 | return true; |
866 | default: |
867 | return false; |
868 | } |
869 | } |
870 | |
871 | static unsigned getPushOpcode(RISCVMachineFunctionInfo::PushPopKind Kind, |
872 | bool UpdateFP) { |
873 | switch (Kind) { |
874 | case RISCVMachineFunctionInfo::PushPopKind::StdExtZcmp: |
875 | return RISCV::CM_PUSH; |
876 | case RISCVMachineFunctionInfo::PushPopKind::VendorXqccmp: |
877 | return UpdateFP ? RISCV::QC_CM_PUSHFP : RISCV::QC_CM_PUSH; |
878 | default: |
879 | llvm_unreachable("Unhandled PushPopKind" ); |
880 | } |
881 | } |
882 | |
883 | static unsigned getPopOpcode(RISCVMachineFunctionInfo::PushPopKind Kind) { |
884 | // There are other pops but they are introduced later by the Push/Pop |
885 | // Optimizer. |
886 | switch (Kind) { |
887 | case RISCVMachineFunctionInfo::PushPopKind::StdExtZcmp: |
888 | return RISCV::CM_POP; |
889 | case RISCVMachineFunctionInfo::PushPopKind::VendorXqccmp: |
890 | return RISCV::QC_CM_POP; |
891 | default: |
892 | llvm_unreachable("Unhandled PushPopKind" ); |
893 | } |
894 | } |
895 | |
896 | void RISCVFrameLowering::emitPrologue(MachineFunction &MF, |
897 | MachineBasicBlock &MBB) const { |
898 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
899 | auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
900 | const RISCVRegisterInfo *RI = STI.getRegisterInfo(); |
901 | MachineBasicBlock::iterator MBBI = MBB.begin(); |
902 | |
903 | Register BPReg = RISCVABI::getBPReg(); |
904 | |
905 | // Debug location must be unknown since the first debug location is used |
906 | // to determine the end of the prologue. |
907 | DebugLoc DL; |
908 | |
909 | // All calls are tail calls in GHC calling conv, and functions have no |
910 | // prologue/epilogue. |
911 | if (MF.getFunction().getCallingConv() == CallingConv::GHC) |
912 | return; |
913 | |
914 | // SiFive CLIC needs to swap `sp` into `sf.mscratchcsw` |
915 | emitSiFiveCLICStackSwap(MF, MBB, MBBI, DL); |
916 | |
917 | // Emit prologue for shadow call stack. |
918 | emitSCSPrologue(MF, MBB, MI: MBBI, DL); |
919 | |
920 | // We keep track of the first instruction because it might be a |
921 | // `(QC.)CM.PUSH(FP)`, and we may need to adjust the immediate rather than |
922 | // inserting an `addi sp, sp, -N*16` |
923 | auto PossiblePush = MBBI; |
924 | |
925 | // Skip past all callee-saved register spill instructions. |
926 | while (MBBI != MBB.end() && MBBI->getFlag(Flag: MachineInstr::FrameSetup)) |
927 | ++MBBI; |
928 | |
929 | // Determine the correct frame layout |
930 | determineFrameLayout(MF); |
931 | |
932 | const auto &CSI = MFI.getCalleeSavedInfo(); |
933 | |
934 | // Skip to before the spills of scalar callee-saved registers |
935 | // FIXME: assumes exactly one instruction is used to restore each |
936 | // callee-saved register. |
937 | MBBI = std::prev(x: MBBI, n: getRVVCalleeSavedInfo(MF, CSI).size() + |
938 | getUnmanagedCSI(MF, CSI).size()); |
939 | CFIInstBuilder CFIBuilder(MBB, MBBI, MachineInstr::FrameSetup); |
940 | |
941 | // If libcalls are used to spill and restore callee-saved registers, the frame |
942 | // has two sections; the opaque section managed by the libcalls, and the |
943 | // section managed by MachineFrameInfo which can also hold callee saved |
944 | // registers in fixed stack slots, both of which have negative frame indices. |
945 | // This gets even more complicated when incoming arguments are passed via the |
946 | // stack, as these too have negative frame indices. An example is detailed |
947 | // below: |
948 | // |
949 | // | incoming arg | <- FI[-3] |
950 | // | libcallspill | |
951 | // | calleespill | <- FI[-2] |
952 | // | calleespill | <- FI[-1] |
953 | // | this_frame | <- FI[0] |
954 | // |
955 | // For negative frame indices, the offset from the frame pointer will differ |
956 | // depending on which of these groups the frame index applies to. |
957 | // The following calculates the correct offset knowing the number of callee |
958 | // saved registers spilt by the two methods. |
959 | if (int LibCallRegs = getLibCallID(MF, CSI: MFI.getCalleeSavedInfo()) + 1) { |
960 | // Calculate the size of the frame managed by the libcall. The stack |
961 | // alignment of these libcalls should be the same as how we set it in |
962 | // getABIStackAlignment. |
963 | unsigned LibCallFrameSize = |
964 | alignTo(Size: (STI.getXLen() / 8) * LibCallRegs, A: getStackAlign()); |
965 | RVFI->setLibCallStackSize(LibCallFrameSize); |
966 | |
967 | CFIBuilder.buildDefCFAOffset(Offset: LibCallFrameSize); |
968 | for (const CalleeSavedInfo &CS : getPushOrLibCallsSavedInfo(MF, CSI)) |
969 | CFIBuilder.buildOffset(Reg: CS.getReg(), |
970 | Offset: MFI.getObjectOffset(ObjectIdx: CS.getFrameIdx())); |
971 | } |
972 | |
973 | // FIXME (note copied from Lanai): This appears to be overallocating. Needs |
974 | // investigation. Get the number of bytes to allocate from the FrameInfo. |
975 | uint64_t RealStackSize = getStackSizeWithRVVPadding(MF); |
976 | uint64_t StackSize = RealStackSize - RVFI->getReservedSpillsSize(); |
977 | uint64_t RVVStackSize = RVFI->getRVVStackSize(); |
978 | |
979 | // Early exit if there is no need to allocate on the stack |
980 | if (RealStackSize == 0 && !MFI.adjustsStack() && RVVStackSize == 0) |
981 | return; |
982 | |
983 | // If the stack pointer has been marked as reserved, then produce an error if |
984 | // the frame requires stack allocation |
985 | if (STI.isRegisterReservedByUser(i: SPReg)) |
986 | MF.getFunction().getContext().diagnose(DI: DiagnosticInfoUnsupported{ |
987 | MF.getFunction(), "Stack pointer required, but has been reserved." }); |
988 | |
989 | uint64_t FirstSPAdjustAmount = getFirstSPAdjustAmount(MF); |
990 | // Split the SP adjustment to reduce the offsets of callee saved spill. |
991 | if (FirstSPAdjustAmount) { |
992 | StackSize = FirstSPAdjustAmount; |
993 | RealStackSize = FirstSPAdjustAmount; |
994 | } |
995 | |
996 | if (RVFI->useQCIInterrupt(MF)) { |
997 | // The function starts with `QC.C.MIENTER(.NEST)`, so the `(QC.)CM.PUSH(FP)` |
998 | // could only be the next instruction. |
999 | ++PossiblePush; |
1000 | |
1001 | // Insert the CFI metadata before where we think the `(QC.)CM.PUSH(FP)` |
1002 | // could be. The PUSH will also get its own CFI metadata for its own |
1003 | // modifications, which should come after the PUSH. |
1004 | CFIInstBuilder PushCFIBuilder(MBB, PossiblePush, MachineInstr::FrameSetup); |
1005 | PushCFIBuilder.buildDefCFAOffset(Offset: QCIInterruptPushAmount); |
1006 | for (const CalleeSavedInfo &CS : getQCISavedInfo(MF, CSI)) |
1007 | PushCFIBuilder.buildOffset(Reg: CS.getReg(), |
1008 | Offset: MFI.getObjectOffset(ObjectIdx: CS.getFrameIdx())); |
1009 | } |
1010 | |
1011 | if (RVFI->isPushable(MF) && PossiblePush != MBB.end() && |
1012 | isPush(Opcode: PossiblePush->getOpcode())) { |
1013 | // Use available stack adjustment in push instruction to allocate additional |
1014 | // stack space. Align the stack size down to a multiple of 16. This is |
1015 | // needed for RVE. |
1016 | // FIXME: Can we increase the stack size to a multiple of 16 instead? |
1017 | uint64_t StackAdj = |
1018 | std::min(a: alignDown(Value: StackSize, Align: 16), b: static_cast<uint64_t>(48)); |
1019 | PossiblePush->getOperand(i: 1).setImm(StackAdj); |
1020 | StackSize -= StackAdj; |
1021 | |
1022 | CFIBuilder.buildDefCFAOffset(Offset: RealStackSize - StackSize); |
1023 | for (const CalleeSavedInfo &CS : getPushOrLibCallsSavedInfo(MF, CSI)) |
1024 | CFIBuilder.buildOffset(Reg: CS.getReg(), |
1025 | Offset: MFI.getObjectOffset(ObjectIdx: CS.getFrameIdx())); |
1026 | } |
1027 | |
1028 | // Allocate space on the stack if necessary. |
1029 | auto &Subtarget = MF.getSubtarget<RISCVSubtarget>(); |
1030 | const RISCVTargetLowering *TLI = Subtarget.getTargetLowering(); |
1031 | bool NeedProbe = TLI->hasInlineStackProbe(MF); |
1032 | uint64_t ProbeSize = TLI->getStackProbeSize(MF, StackAlign: getStackAlign()); |
1033 | bool DynAllocation = |
1034 | MF.getInfo<RISCVMachineFunctionInfo>()->hasDynamicAllocation(); |
1035 | if (StackSize != 0) |
1036 | allocateStack(MBB, MBBI, MF, Offset: StackSize, RealStackSize, /*EmitCFI=*/true, |
1037 | NeedProbe, ProbeSize, DynAllocation); |
1038 | |
1039 | // Save SiFive CLIC CSRs into Stack |
1040 | emitSiFiveCLICPreemptibleSaves(MF, MBB, MBBI, DL); |
1041 | |
1042 | // The frame pointer is callee-saved, and code has been generated for us to |
1043 | // save it to the stack. We need to skip over the storing of callee-saved |
1044 | // registers as the frame pointer must be modified after it has been saved |
1045 | // to the stack, not before. |
1046 | // FIXME: assumes exactly one instruction is used to save each callee-saved |
1047 | // register. |
1048 | std::advance(i&: MBBI, n: getUnmanagedCSI(MF, CSI).size()); |
1049 | CFIBuilder.setInsertPoint(MBBI); |
1050 | |
1051 | // Iterate over list of callee-saved registers and emit .cfi_offset |
1052 | // directives. |
1053 | for (const CalleeSavedInfo &CS : getUnmanagedCSI(MF, CSI)) |
1054 | CFIBuilder.buildOffset(Reg: CS.getReg(), Offset: MFI.getObjectOffset(ObjectIdx: CS.getFrameIdx())); |
1055 | |
1056 | // Generate new FP. |
1057 | if (hasFP(MF)) { |
1058 | if (STI.isRegisterReservedByUser(i: FPReg)) |
1059 | MF.getFunction().getContext().diagnose(DI: DiagnosticInfoUnsupported{ |
1060 | MF.getFunction(), "Frame pointer required, but has been reserved." }); |
1061 | // The frame pointer does need to be reserved from register allocation. |
1062 | assert(MF.getRegInfo().isReserved(FPReg) && "FP not reserved" ); |
1063 | |
1064 | // Some stack management variants automatically keep FP updated, so we don't |
1065 | // need an instruction to do so. |
1066 | if (!RVFI->hasImplicitFPUpdates(MF)) { |
1067 | RI->adjustReg( |
1068 | MBB, II: MBBI, DL, DestReg: FPReg, SrcReg: SPReg, |
1069 | Offset: StackOffset::getFixed(Fixed: RealStackSize - RVFI->getVarArgsSaveSize()), |
1070 | Flag: MachineInstr::FrameSetup, RequiredAlign: getStackAlign()); |
1071 | } |
1072 | |
1073 | CFIBuilder.buildDefCFA(Reg: FPReg, Offset: RVFI->getVarArgsSaveSize()); |
1074 | } |
1075 | |
1076 | uint64_t SecondSPAdjustAmount = 0; |
1077 | // Emit the second SP adjustment after saving callee saved registers. |
1078 | if (FirstSPAdjustAmount) { |
1079 | SecondSPAdjustAmount = getStackSizeWithRVVPadding(MF) - FirstSPAdjustAmount; |
1080 | assert(SecondSPAdjustAmount > 0 && |
1081 | "SecondSPAdjustAmount should be greater than zero" ); |
1082 | |
1083 | allocateStack(MBB, MBBI, MF, Offset: SecondSPAdjustAmount, |
1084 | RealStackSize: getStackSizeWithRVVPadding(MF), EmitCFI: !hasFP(MF), NeedProbe, |
1085 | ProbeSize, DynAllocation); |
1086 | } |
1087 | |
1088 | if (RVVStackSize) { |
1089 | if (NeedProbe) { |
1090 | allocateAndProbeStackForRVV(MF, MBB, MBBI, DL, Amount: RVVStackSize, |
1091 | Flag: MachineInstr::FrameSetup, EmitCFI: !hasFP(MF), |
1092 | DynAllocation); |
1093 | } else { |
1094 | // We must keep the stack pointer aligned through any intermediate |
1095 | // updates. |
1096 | RI->adjustReg(MBB, II: MBBI, DL, DestReg: SPReg, SrcReg: SPReg, |
1097 | Offset: StackOffset::getScalable(Scalable: -RVVStackSize), |
1098 | Flag: MachineInstr::FrameSetup, RequiredAlign: getStackAlign()); |
1099 | } |
1100 | |
1101 | if (!hasFP(MF)) { |
1102 | // Emit .cfi_def_cfa_expression "sp + StackSize + RVVStackSize * vlenb". |
1103 | CFIBuilder.insertCFIInst(CFIInst: createDefCFAExpression( |
1104 | TRI: *RI, Reg: SPReg, FixedOffset: getStackSizeWithRVVPadding(MF), ScalableOffset: RVVStackSize / 8)); |
1105 | } |
1106 | |
1107 | std::advance(i&: MBBI, n: getRVVCalleeSavedInfo(MF, CSI).size()); |
1108 | emitCalleeSavedRVVPrologCFI(MBB, MI: MBBI, HasFP: hasFP(MF)); |
1109 | } |
1110 | |
1111 | if (hasFP(MF)) { |
1112 | // Realign Stack |
1113 | const RISCVRegisterInfo *RI = STI.getRegisterInfo(); |
1114 | if (RI->hasStackRealignment(MF)) { |
1115 | Align MaxAlignment = MFI.getMaxAlign(); |
1116 | |
1117 | const RISCVInstrInfo *TII = STI.getInstrInfo(); |
1118 | if (isInt<12>(x: -(int)MaxAlignment.value())) { |
1119 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: RISCV::ANDI), DestReg: SPReg) |
1120 | .addReg(RegNo: SPReg) |
1121 | .addImm(Val: -(int)MaxAlignment.value()) |
1122 | .setMIFlag(MachineInstr::FrameSetup); |
1123 | } else { |
1124 | unsigned ShiftAmount = Log2(A: MaxAlignment); |
1125 | Register VR = |
1126 | MF.getRegInfo().createVirtualRegister(RegClass: &RISCV::GPRRegClass); |
1127 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: RISCV::SRLI), DestReg: VR) |
1128 | .addReg(RegNo: SPReg) |
1129 | .addImm(Val: ShiftAmount) |
1130 | .setMIFlag(MachineInstr::FrameSetup); |
1131 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: RISCV::SLLI), DestReg: SPReg) |
1132 | .addReg(RegNo: VR) |
1133 | .addImm(Val: ShiftAmount) |
1134 | .setMIFlag(MachineInstr::FrameSetup); |
1135 | } |
1136 | if (NeedProbe && RVVStackSize == 0) { |
1137 | // Do a probe if the align + size allocated just passed the probe size |
1138 | // and was not yet probed. |
1139 | if (SecondSPAdjustAmount < ProbeSize && |
1140 | SecondSPAdjustAmount + MaxAlignment.value() >= ProbeSize) { |
1141 | bool IsRV64 = STI.is64Bit(); |
1142 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: IsRV64 ? RISCV::SD : RISCV::SW)) |
1143 | .addReg(RegNo: RISCV::X0) |
1144 | .addReg(RegNo: SPReg) |
1145 | .addImm(Val: 0) |
1146 | .setMIFlags(MachineInstr::FrameSetup); |
1147 | } |
1148 | } |
1149 | // FP will be used to restore the frame in the epilogue, so we need |
1150 | // another base register BP to record SP after re-alignment. SP will |
1151 | // track the current stack after allocating variable sized objects. |
1152 | if (hasBP(MF)) { |
1153 | // move BP, SP |
1154 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: RISCV::ADDI), DestReg: BPReg) |
1155 | .addReg(RegNo: SPReg) |
1156 | .addImm(Val: 0) |
1157 | .setMIFlag(MachineInstr::FrameSetup); |
1158 | } |
1159 | } |
1160 | } |
1161 | } |
1162 | |
1163 | void RISCVFrameLowering::deallocateStack(MachineFunction &MF, |
1164 | MachineBasicBlock &MBB, |
1165 | MachineBasicBlock::iterator MBBI, |
1166 | const DebugLoc &DL, |
1167 | uint64_t &StackSize, |
1168 | int64_t CFAOffset) const { |
1169 | const RISCVRegisterInfo *RI = STI.getRegisterInfo(); |
1170 | |
1171 | RI->adjustReg(MBB, II: MBBI, DL, DestReg: SPReg, SrcReg: SPReg, Offset: StackOffset::getFixed(Fixed: StackSize), |
1172 | Flag: MachineInstr::FrameDestroy, RequiredAlign: getStackAlign()); |
1173 | StackSize = 0; |
1174 | |
1175 | CFIInstBuilder(MBB, MBBI, MachineInstr::FrameDestroy) |
1176 | .buildDefCFAOffset(Offset: CFAOffset); |
1177 | } |
1178 | |
1179 | void RISCVFrameLowering::emitEpilogue(MachineFunction &MF, |
1180 | MachineBasicBlock &MBB) const { |
1181 | const RISCVRegisterInfo *RI = STI.getRegisterInfo(); |
1182 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
1183 | auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
1184 | |
1185 | // All calls are tail calls in GHC calling conv, and functions have no |
1186 | // prologue/epilogue. |
1187 | if (MF.getFunction().getCallingConv() == CallingConv::GHC) |
1188 | return; |
1189 | |
1190 | // Get the insert location for the epilogue. If there were no terminators in |
1191 | // the block, get the last instruction. |
1192 | MachineBasicBlock::iterator MBBI = MBB.end(); |
1193 | DebugLoc DL; |
1194 | if (!MBB.empty()) { |
1195 | MBBI = MBB.getLastNonDebugInstr(); |
1196 | if (MBBI != MBB.end()) |
1197 | DL = MBBI->getDebugLoc(); |
1198 | |
1199 | MBBI = MBB.getFirstTerminator(); |
1200 | |
1201 | // Skip to before the restores of all callee-saved registers. |
1202 | while (MBBI != MBB.begin() && |
1203 | std::prev(x: MBBI)->getFlag(Flag: MachineInstr::FrameDestroy)) |
1204 | --MBBI; |
1205 | } |
1206 | |
1207 | const auto &CSI = MFI.getCalleeSavedInfo(); |
1208 | |
1209 | // Skip to before the restores of scalar callee-saved registers |
1210 | // FIXME: assumes exactly one instruction is used to restore each |
1211 | // callee-saved register. |
1212 | auto FirstScalarCSRRestoreInsn = |
1213 | std::next(x: MBBI, n: getRVVCalleeSavedInfo(MF, CSI).size()); |
1214 | CFIInstBuilder CFIBuilder(MBB, FirstScalarCSRRestoreInsn, |
1215 | MachineInstr::FrameDestroy); |
1216 | |
1217 | uint64_t FirstSPAdjustAmount = getFirstSPAdjustAmount(MF); |
1218 | uint64_t RealStackSize = FirstSPAdjustAmount ? FirstSPAdjustAmount |
1219 | : getStackSizeWithRVVPadding(MF); |
1220 | uint64_t StackSize = FirstSPAdjustAmount ? FirstSPAdjustAmount |
1221 | : getStackSizeWithRVVPadding(MF) - |
1222 | RVFI->getReservedSpillsSize(); |
1223 | uint64_t FPOffset = RealStackSize - RVFI->getVarArgsSaveSize(); |
1224 | uint64_t RVVStackSize = RVFI->getRVVStackSize(); |
1225 | |
1226 | bool RestoreSPFromFP = RI->hasStackRealignment(MF) || |
1227 | MFI.hasVarSizedObjects() || !hasReservedCallFrame(MF); |
1228 | if (RVVStackSize) { |
1229 | // If RestoreSPFromFP the stack pointer will be restored using the frame |
1230 | // pointer value. |
1231 | if (!RestoreSPFromFP) |
1232 | RI->adjustReg(MBB, II: FirstScalarCSRRestoreInsn, DL, DestReg: SPReg, SrcReg: SPReg, |
1233 | Offset: StackOffset::getScalable(Scalable: RVVStackSize), |
1234 | Flag: MachineInstr::FrameDestroy, RequiredAlign: getStackAlign()); |
1235 | |
1236 | if (!hasFP(MF)) |
1237 | CFIBuilder.buildDefCFA(Reg: SPReg, Offset: RealStackSize); |
1238 | |
1239 | emitCalleeSavedRVVEpilogCFI(MBB, MI: FirstScalarCSRRestoreInsn); |
1240 | } |
1241 | |
1242 | if (FirstSPAdjustAmount) { |
1243 | uint64_t SecondSPAdjustAmount = |
1244 | getStackSizeWithRVVPadding(MF) - FirstSPAdjustAmount; |
1245 | assert(SecondSPAdjustAmount > 0 && |
1246 | "SecondSPAdjustAmount should be greater than zero" ); |
1247 | |
1248 | // If RestoreSPFromFP the stack pointer will be restored using the frame |
1249 | // pointer value. |
1250 | if (!RestoreSPFromFP) |
1251 | RI->adjustReg(MBB, II: FirstScalarCSRRestoreInsn, DL, DestReg: SPReg, SrcReg: SPReg, |
1252 | Offset: StackOffset::getFixed(Fixed: SecondSPAdjustAmount), |
1253 | Flag: MachineInstr::FrameDestroy, RequiredAlign: getStackAlign()); |
1254 | |
1255 | if (!hasFP(MF)) |
1256 | CFIBuilder.buildDefCFAOffset(Offset: FirstSPAdjustAmount); |
1257 | } |
1258 | |
1259 | // Restore the stack pointer using the value of the frame pointer. Only |
1260 | // necessary if the stack pointer was modified, meaning the stack size is |
1261 | // unknown. |
1262 | // |
1263 | // In order to make sure the stack point is right through the EH region, |
1264 | // we also need to restore stack pointer from the frame pointer if we |
1265 | // don't preserve stack space within prologue/epilogue for outgoing variables, |
1266 | // normally it's just checking the variable sized object is present or not |
1267 | // is enough, but we also don't preserve that at prologue/epilogue when |
1268 | // have vector objects in stack. |
1269 | if (RestoreSPFromFP) { |
1270 | assert(hasFP(MF) && "frame pointer should not have been eliminated" ); |
1271 | RI->adjustReg(MBB, II: FirstScalarCSRRestoreInsn, DL, DestReg: SPReg, SrcReg: FPReg, |
1272 | Offset: StackOffset::getFixed(Fixed: -FPOffset), Flag: MachineInstr::FrameDestroy, |
1273 | RequiredAlign: getStackAlign()); |
1274 | } |
1275 | |
1276 | if (hasFP(MF)) |
1277 | CFIBuilder.buildDefCFA(Reg: SPReg, Offset: RealStackSize); |
1278 | |
1279 | // Skip to after the restores of scalar callee-saved registers |
1280 | // FIXME: assumes exactly one instruction is used to restore each |
1281 | // callee-saved register. |
1282 | MBBI = std::next(x: FirstScalarCSRRestoreInsn, n: getUnmanagedCSI(MF, CSI).size()); |
1283 | CFIBuilder.setInsertPoint(MBBI); |
1284 | |
1285 | if (getLibCallID(MF, CSI) != -1) { |
1286 | // tail __riscv_restore_[0-12] instruction is considered as a terminator, |
1287 | // therefore it is unnecessary to place any CFI instructions after it. Just |
1288 | // deallocate stack if needed and return. |
1289 | if (StackSize != 0) |
1290 | deallocateStack(MF, MBB, MBBI, DL, StackSize, |
1291 | CFAOffset: RVFI->getLibCallStackSize()); |
1292 | |
1293 | // Emit epilogue for shadow call stack. |
1294 | emitSCSEpilogue(MF, MBB, MI: MBBI, DL); |
1295 | return; |
1296 | } |
1297 | |
1298 | // Recover callee-saved registers. |
1299 | for (const CalleeSavedInfo &CS : getUnmanagedCSI(MF, CSI)) |
1300 | CFIBuilder.buildRestore(Reg: CS.getReg()); |
1301 | |
1302 | if (RVFI->isPushable(MF) && MBBI != MBB.end() && isPop(Opcode: MBBI->getOpcode())) { |
1303 | // Use available stack adjustment in pop instruction to deallocate stack |
1304 | // space. Align the stack size down to a multiple of 16. This is needed for |
1305 | // RVE. |
1306 | // FIXME: Can we increase the stack size to a multiple of 16 instead? |
1307 | uint64_t StackAdj = |
1308 | std::min(a: alignDown(Value: StackSize, Align: 16), b: static_cast<uint64_t>(48)); |
1309 | MBBI->getOperand(i: 1).setImm(StackAdj); |
1310 | StackSize -= StackAdj; |
1311 | |
1312 | if (StackSize != 0) |
1313 | deallocateStack(MF, MBB, MBBI, DL, StackSize, |
1314 | /*stack_adj of cm.pop instr*/ CFAOffset: RealStackSize - StackSize); |
1315 | |
1316 | auto NextI = next_nodbg(It: MBBI, End: MBB.end()); |
1317 | if (NextI == MBB.end() || NextI->getOpcode() != RISCV::PseudoRET) { |
1318 | ++MBBI; |
1319 | CFIBuilder.setInsertPoint(MBBI); |
1320 | |
1321 | for (const CalleeSavedInfo &CS : getPushOrLibCallsSavedInfo(MF, CSI)) |
1322 | CFIBuilder.buildRestore(Reg: CS.getReg()); |
1323 | |
1324 | // Update CFA Offset. If this is a QCI interrupt function, there will be a |
1325 | // leftover offset which is deallocated by `QC.C.MILEAVERET`, otherwise |
1326 | // getQCIInterruptStackSize() will be 0. |
1327 | CFIBuilder.buildDefCFAOffset(Offset: RVFI->getQCIInterruptStackSize()); |
1328 | } |
1329 | } |
1330 | |
1331 | emitSiFiveCLICPreemptibleRestores(MF, MBB, MBBI, DL); |
1332 | |
1333 | // Deallocate stack if StackSize isn't a zero yet. If this is a QCI interrupt |
1334 | // function, there will be a leftover offset which is deallocated by |
1335 | // `QC.C.MILEAVERET`, otherwise getQCIInterruptStackSize() will be 0. |
1336 | if (StackSize != 0) |
1337 | deallocateStack(MF, MBB, MBBI, DL, StackSize, |
1338 | CFAOffset: RVFI->getQCIInterruptStackSize()); |
1339 | |
1340 | // Emit epilogue for shadow call stack. |
1341 | emitSCSEpilogue(MF, MBB, MI: MBBI, DL); |
1342 | |
1343 | // SiFive CLIC needs to swap `sf.mscratchcsw` into `sp` |
1344 | emitSiFiveCLICStackSwap(MF, MBB, MBBI, DL); |
1345 | } |
1346 | |
1347 | StackOffset |
1348 | RISCVFrameLowering::getFrameIndexReference(const MachineFunction &MF, int FI, |
1349 | Register &FrameReg) const { |
1350 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
1351 | const TargetRegisterInfo *RI = MF.getSubtarget().getRegisterInfo(); |
1352 | const auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
1353 | |
1354 | // Callee-saved registers should be referenced relative to the stack |
1355 | // pointer (positive offset), otherwise use the frame pointer (negative |
1356 | // offset). |
1357 | const auto &CSI = getUnmanagedCSI(MF, CSI: MFI.getCalleeSavedInfo()); |
1358 | int MinCSFI = 0; |
1359 | int MaxCSFI = -1; |
1360 | StackOffset Offset; |
1361 | auto StackID = MFI.getStackID(ObjectIdx: FI); |
1362 | |
1363 | assert((StackID == TargetStackID::Default || |
1364 | StackID == TargetStackID::ScalableVector) && |
1365 | "Unexpected stack ID for the frame object." ); |
1366 | if (StackID == TargetStackID::Default) { |
1367 | assert(getOffsetOfLocalArea() == 0 && "LocalAreaOffset is not 0!" ); |
1368 | Offset = StackOffset::getFixed(Fixed: MFI.getObjectOffset(ObjectIdx: FI) + |
1369 | MFI.getOffsetAdjustment()); |
1370 | } else if (StackID == TargetStackID::ScalableVector) { |
1371 | Offset = StackOffset::getScalable(Scalable: MFI.getObjectOffset(ObjectIdx: FI)); |
1372 | } |
1373 | |
1374 | uint64_t FirstSPAdjustAmount = getFirstSPAdjustAmount(MF); |
1375 | |
1376 | if (CSI.size()) { |
1377 | MinCSFI = CSI[0].getFrameIdx(); |
1378 | MaxCSFI = CSI[CSI.size() - 1].getFrameIdx(); |
1379 | } |
1380 | |
1381 | if (FI >= MinCSFI && FI <= MaxCSFI) { |
1382 | FrameReg = SPReg; |
1383 | |
1384 | if (FirstSPAdjustAmount) |
1385 | Offset += StackOffset::getFixed(Fixed: FirstSPAdjustAmount); |
1386 | else |
1387 | Offset += StackOffset::getFixed(Fixed: getStackSizeWithRVVPadding(MF)); |
1388 | return Offset; |
1389 | } |
1390 | |
1391 | if (RI->hasStackRealignment(MF) && !MFI.isFixedObjectIndex(ObjectIdx: FI)) { |
1392 | // If the stack was realigned, the frame pointer is set in order to allow |
1393 | // SP to be restored, so we need another base register to record the stack |
1394 | // after realignment. |
1395 | // |--------------------------| -- <-- FP |
1396 | // | callee-allocated save | | <----| |
1397 | // | area for register varargs| | | |
1398 | // |--------------------------| | | |
1399 | // | callee-saved registers | | | |
1400 | // |--------------------------| -- | |
1401 | // | realignment (the size of | | | |
1402 | // | this area is not counted | | | |
1403 | // | in MFI.getStackSize()) | | | |
1404 | // |--------------------------| -- |-- MFI.getStackSize() |
1405 | // | RVV alignment padding | | | |
1406 | // | (not counted in | | | |
1407 | // | MFI.getStackSize() but | | | |
1408 | // | counted in | | | |
1409 | // | RVFI.getRVVStackSize()) | | | |
1410 | // |--------------------------| -- | |
1411 | // | RVV objects | | | |
1412 | // | (not counted in | | | |
1413 | // | MFI.getStackSize()) | | | |
1414 | // |--------------------------| -- | |
1415 | // | padding before RVV | | | |
1416 | // | (not counted in | | | |
1417 | // | MFI.getStackSize() or in | | | |
1418 | // | RVFI.getRVVStackSize()) | | | |
1419 | // |--------------------------| -- | |
1420 | // | scalar local variables | | <----' |
1421 | // |--------------------------| -- <-- BP (if var sized objects present) |
1422 | // | VarSize objects | | |
1423 | // |--------------------------| -- <-- SP |
1424 | if (hasBP(MF)) { |
1425 | FrameReg = RISCVABI::getBPReg(); |
1426 | } else { |
1427 | // VarSize objects must be empty in this case! |
1428 | assert(!MFI.hasVarSizedObjects()); |
1429 | FrameReg = SPReg; |
1430 | } |
1431 | } else { |
1432 | FrameReg = RI->getFrameRegister(MF); |
1433 | } |
1434 | |
1435 | if (FrameReg == FPReg) { |
1436 | Offset += StackOffset::getFixed(Fixed: RVFI->getVarArgsSaveSize()); |
1437 | // When using FP to access scalable vector objects, we need to minus |
1438 | // the frame size. |
1439 | // |
1440 | // |--------------------------| -- <-- FP |
1441 | // | callee-allocated save | | |
1442 | // | area for register varargs| | |
1443 | // |--------------------------| | |
1444 | // | callee-saved registers | | |
1445 | // |--------------------------| | MFI.getStackSize() |
1446 | // | scalar local variables | | |
1447 | // |--------------------------| -- (Offset of RVV objects is from here.) |
1448 | // | RVV objects | |
1449 | // |--------------------------| |
1450 | // | VarSize objects | |
1451 | // |--------------------------| <-- SP |
1452 | if (StackID == TargetStackID::ScalableVector) { |
1453 | assert(!RI->hasStackRealignment(MF) && |
1454 | "Can't index across variable sized realign" ); |
1455 | // We don't expect any extra RVV alignment padding, as the stack size |
1456 | // and RVV object sections should be correct aligned in their own |
1457 | // right. |
1458 | assert(MFI.getStackSize() == getStackSizeWithRVVPadding(MF) && |
1459 | "Inconsistent stack layout" ); |
1460 | Offset -= StackOffset::getFixed(Fixed: MFI.getStackSize()); |
1461 | } |
1462 | return Offset; |
1463 | } |
1464 | |
1465 | // This case handles indexing off both SP and BP. |
1466 | // If indexing off SP, there must not be any var sized objects |
1467 | assert(FrameReg == RISCVABI::getBPReg() || !MFI.hasVarSizedObjects()); |
1468 | |
1469 | // When using SP to access frame objects, we need to add RVV stack size. |
1470 | // |
1471 | // |--------------------------| -- <-- FP |
1472 | // | callee-allocated save | | <----| |
1473 | // | area for register varargs| | | |
1474 | // |--------------------------| | | |
1475 | // | callee-saved registers | | | |
1476 | // |--------------------------| -- | |
1477 | // | RVV alignment padding | | | |
1478 | // | (not counted in | | | |
1479 | // | MFI.getStackSize() but | | | |
1480 | // | counted in | | | |
1481 | // | RVFI.getRVVStackSize()) | | | |
1482 | // |--------------------------| -- | |
1483 | // | RVV objects | | |-- MFI.getStackSize() |
1484 | // | (not counted in | | | |
1485 | // | MFI.getStackSize()) | | | |
1486 | // |--------------------------| -- | |
1487 | // | padding before RVV | | | |
1488 | // | (not counted in | | | |
1489 | // | MFI.getStackSize()) | | | |
1490 | // |--------------------------| -- | |
1491 | // | scalar local variables | | <----' |
1492 | // |--------------------------| -- <-- BP (if var sized objects present) |
1493 | // | VarSize objects | | |
1494 | // |--------------------------| -- <-- SP |
1495 | // |
1496 | // The total amount of padding surrounding RVV objects is described by |
1497 | // RVV->getRVVPadding() and it can be zero. It allows us to align the RVV |
1498 | // objects to the required alignment. |
1499 | if (MFI.getStackID(ObjectIdx: FI) == TargetStackID::Default) { |
1500 | if (MFI.isFixedObjectIndex(ObjectIdx: FI)) { |
1501 | assert(!RI->hasStackRealignment(MF) && |
1502 | "Can't index across variable sized realign" ); |
1503 | Offset += StackOffset::get(Fixed: getStackSizeWithRVVPadding(MF), |
1504 | Scalable: RVFI->getRVVStackSize()); |
1505 | } else { |
1506 | Offset += StackOffset::getFixed(Fixed: MFI.getStackSize()); |
1507 | } |
1508 | } else if (MFI.getStackID(ObjectIdx: FI) == TargetStackID::ScalableVector) { |
1509 | // Ensure the base of the RVV stack is correctly aligned: add on the |
1510 | // alignment padding. |
1511 | int ScalarLocalVarSize = MFI.getStackSize() - |
1512 | RVFI->getCalleeSavedStackSize() - |
1513 | RVFI->getVarArgsSaveSize() + RVFI->getRVVPadding(); |
1514 | Offset += StackOffset::get(Fixed: ScalarLocalVarSize, Scalable: RVFI->getRVVStackSize()); |
1515 | } |
1516 | return Offset; |
1517 | } |
1518 | |
1519 | void RISCVFrameLowering::determineCalleeSaves(MachineFunction &MF, |
1520 | BitVector &SavedRegs, |
1521 | RegScavenger *RS) const { |
1522 | TargetFrameLowering::determineCalleeSaves(MF, SavedRegs, RS); |
1523 | // Unconditionally spill RA and FP only if the function uses a frame |
1524 | // pointer. |
1525 | if (hasFP(MF)) { |
1526 | SavedRegs.set(RAReg); |
1527 | SavedRegs.set(FPReg); |
1528 | } |
1529 | // Mark BP as used if function has dedicated base pointer. |
1530 | if (hasBP(MF)) |
1531 | SavedRegs.set(RISCVABI::getBPReg()); |
1532 | |
1533 | // When using cm.push/pop we must save X27 if we save X26. |
1534 | auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
1535 | if (RVFI->isPushable(MF) && SavedRegs.test(Idx: RISCV::X26)) |
1536 | SavedRegs.set(RISCV::X27); |
1537 | |
1538 | // SiFive Preemptible Interrupt Handlers need additional frame entries |
1539 | createSiFivePreemptibleInterruptFrameEntries(MF, RVFI&: *RVFI); |
1540 | } |
1541 | |
1542 | std::pair<int64_t, Align> |
1543 | RISCVFrameLowering::assignRVVStackObjectOffsets(MachineFunction &MF) const { |
1544 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
1545 | // Create a buffer of RVV objects to allocate. |
1546 | SmallVector<int, 8> ObjectsToAllocate; |
1547 | auto pushRVVObjects = [&](int FIBegin, int FIEnd) { |
1548 | for (int I = FIBegin, E = FIEnd; I != E; ++I) { |
1549 | unsigned StackID = MFI.getStackID(ObjectIdx: I); |
1550 | if (StackID != TargetStackID::ScalableVector) |
1551 | continue; |
1552 | if (MFI.isDeadObjectIndex(ObjectIdx: I)) |
1553 | continue; |
1554 | |
1555 | ObjectsToAllocate.push_back(Elt: I); |
1556 | } |
1557 | }; |
1558 | // First push RVV Callee Saved object, then push RVV stack object |
1559 | std::vector<CalleeSavedInfo> &CSI = MF.getFrameInfo().getCalleeSavedInfo(); |
1560 | const auto &RVVCSI = getRVVCalleeSavedInfo(MF, CSI); |
1561 | if (!RVVCSI.empty()) |
1562 | pushRVVObjects(RVVCSI[0].getFrameIdx(), |
1563 | RVVCSI[RVVCSI.size() - 1].getFrameIdx() + 1); |
1564 | pushRVVObjects(0, MFI.getObjectIndexEnd() - RVVCSI.size()); |
1565 | |
1566 | // The minimum alignment is 16 bytes. |
1567 | Align RVVStackAlign(16); |
1568 | const auto &ST = MF.getSubtarget<RISCVSubtarget>(); |
1569 | |
1570 | if (!ST.hasVInstructions()) { |
1571 | assert(ObjectsToAllocate.empty() && |
1572 | "Can't allocate scalable-vector objects without V instructions" ); |
1573 | return std::make_pair(x: 0, y&: RVVStackAlign); |
1574 | } |
1575 | |
1576 | // Allocate all RVV locals and spills |
1577 | int64_t Offset = 0; |
1578 | for (int FI : ObjectsToAllocate) { |
1579 | // ObjectSize in bytes. |
1580 | int64_t ObjectSize = MFI.getObjectSize(ObjectIdx: FI); |
1581 | auto ObjectAlign = |
1582 | std::max(a: Align(RISCV::RVVBytesPerBlock), b: MFI.getObjectAlign(ObjectIdx: FI)); |
1583 | // If the data type is the fractional vector type, reserve one vector |
1584 | // register for it. |
1585 | if (ObjectSize < RISCV::RVVBytesPerBlock) |
1586 | ObjectSize = RISCV::RVVBytesPerBlock; |
1587 | Offset = alignTo(Size: Offset + ObjectSize, A: ObjectAlign); |
1588 | MFI.setObjectOffset(ObjectIdx: FI, SPOffset: -Offset); |
1589 | // Update the maximum alignment of the RVV stack section |
1590 | RVVStackAlign = std::max(a: RVVStackAlign, b: ObjectAlign); |
1591 | } |
1592 | |
1593 | uint64_t StackSize = Offset; |
1594 | |
1595 | // Ensure the alignment of the RVV stack. Since we want the most-aligned |
1596 | // object right at the bottom (i.e., any padding at the top of the frame), |
1597 | // readjust all RVV objects down by the alignment padding. |
1598 | // Stack size and offsets are multiples of vscale, stack alignment is in |
1599 | // bytes, we can divide stack alignment by minimum vscale to get a maximum |
1600 | // stack alignment multiple of vscale. |
1601 | auto VScale = |
1602 | std::max<uint64_t>(a: ST.getRealMinVLen() / RISCV::RVVBitsPerBlock, b: 1); |
1603 | if (auto RVVStackAlignVScale = RVVStackAlign.value() / VScale) { |
1604 | if (auto AlignmentPadding = |
1605 | offsetToAlignment(Value: StackSize, Alignment: Align(RVVStackAlignVScale))) { |
1606 | StackSize += AlignmentPadding; |
1607 | for (int FI : ObjectsToAllocate) |
1608 | MFI.setObjectOffset(ObjectIdx: FI, SPOffset: MFI.getObjectOffset(ObjectIdx: FI) - AlignmentPadding); |
1609 | } |
1610 | } |
1611 | |
1612 | return std::make_pair(x&: StackSize, y&: RVVStackAlign); |
1613 | } |
1614 | |
1615 | static unsigned getScavSlotsNumForRVV(MachineFunction &MF) { |
1616 | // For RVV spill, scalable stack offsets computing requires up to two scratch |
1617 | // registers |
1618 | static constexpr unsigned ScavSlotsNumRVVSpillScalableObject = 2; |
1619 | |
1620 | // For RVV spill, non-scalable stack offsets computing requires up to one |
1621 | // scratch register. |
1622 | static constexpr unsigned ScavSlotsNumRVVSpillNonScalableObject = 1; |
1623 | |
1624 | // ADDI instruction's destination register can be used for computing |
1625 | // offsets. So Scalable stack offsets require up to one scratch register. |
1626 | static constexpr unsigned ScavSlotsADDIScalableObject = 1; |
1627 | |
1628 | static constexpr unsigned MaxScavSlotsNumKnown = |
1629 | std::max(l: {ScavSlotsADDIScalableObject, ScavSlotsNumRVVSpillScalableObject, |
1630 | ScavSlotsNumRVVSpillNonScalableObject}); |
1631 | |
1632 | unsigned MaxScavSlotsNum = 0; |
1633 | if (!MF.getSubtarget<RISCVSubtarget>().hasVInstructions()) |
1634 | return false; |
1635 | for (const MachineBasicBlock &MBB : MF) |
1636 | for (const MachineInstr &MI : MBB) { |
1637 | bool IsRVVSpill = RISCV::isRVVSpill(MI); |
1638 | for (auto &MO : MI.operands()) { |
1639 | if (!MO.isFI()) |
1640 | continue; |
1641 | bool IsScalableVectorID = MF.getFrameInfo().getStackID(ObjectIdx: MO.getIndex()) == |
1642 | TargetStackID::ScalableVector; |
1643 | if (IsRVVSpill) { |
1644 | MaxScavSlotsNum = std::max( |
1645 | a: MaxScavSlotsNum, b: IsScalableVectorID |
1646 | ? ScavSlotsNumRVVSpillScalableObject |
1647 | : ScavSlotsNumRVVSpillNonScalableObject); |
1648 | } else if (MI.getOpcode() == RISCV::ADDI && IsScalableVectorID) { |
1649 | MaxScavSlotsNum = |
1650 | std::max(a: MaxScavSlotsNum, b: ScavSlotsADDIScalableObject); |
1651 | } |
1652 | } |
1653 | if (MaxScavSlotsNum == MaxScavSlotsNumKnown) |
1654 | return MaxScavSlotsNumKnown; |
1655 | } |
1656 | return MaxScavSlotsNum; |
1657 | } |
1658 | |
1659 | static bool hasRVVFrameObject(const MachineFunction &MF) { |
1660 | // Originally, the function will scan all the stack objects to check whether |
1661 | // if there is any scalable vector object on the stack or not. However, it |
1662 | // causes errors in the register allocator. In issue 53016, it returns false |
1663 | // before RA because there is no RVV stack objects. After RA, it returns true |
1664 | // because there are spilling slots for RVV values during RA. It will not |
1665 | // reserve BP during register allocation and generate BP access in the PEI |
1666 | // pass due to the inconsistent behavior of the function. |
1667 | // |
1668 | // The function is changed to use hasVInstructions() as the return value. It |
1669 | // is not precise, but it can make the register allocation correct. |
1670 | // |
1671 | // FIXME: Find a better way to make the decision or revisit the solution in |
1672 | // D103622. |
1673 | // |
1674 | // Refer to https://github.com/llvm/llvm-project/issues/53016. |
1675 | return MF.getSubtarget<RISCVSubtarget>().hasVInstructions(); |
1676 | } |
1677 | |
1678 | static unsigned estimateFunctionSizeInBytes(const MachineFunction &MF, |
1679 | const RISCVInstrInfo &TII) { |
1680 | unsigned FnSize = 0; |
1681 | for (auto &MBB : MF) { |
1682 | for (auto &MI : MBB) { |
1683 | // Far branches over 20-bit offset will be relaxed in branch relaxation |
1684 | // pass. In the worst case, conditional branches will be relaxed into |
1685 | // the following instruction sequence. Unconditional branches are |
1686 | // relaxed in the same way, with the exception that there is no first |
1687 | // branch instruction. |
1688 | // |
1689 | // foo |
1690 | // bne t5, t6, .rev_cond # `TII->getInstSizeInBytes(MI)` bytes |
1691 | // sd s11, 0(sp) # 4 bytes, or 2 bytes with Zca |
1692 | // jump .restore, s11 # 8 bytes |
1693 | // .rev_cond |
1694 | // bar |
1695 | // j .dest_bb # 4 bytes, or 2 bytes with Zca |
1696 | // .restore: |
1697 | // ld s11, 0(sp) # 4 bytes, or 2 bytes with Zca |
1698 | // .dest: |
1699 | // baz |
1700 | if (MI.isConditionalBranch()) |
1701 | FnSize += TII.getInstSizeInBytes(MI); |
1702 | if (MI.isConditionalBranch() || MI.isUnconditionalBranch()) { |
1703 | if (MF.getSubtarget<RISCVSubtarget>().hasStdExtZca()) |
1704 | FnSize += 2 + 8 + 2 + 2; |
1705 | else |
1706 | FnSize += 4 + 8 + 4 + 4; |
1707 | continue; |
1708 | } |
1709 | |
1710 | FnSize += TII.getInstSizeInBytes(MI); |
1711 | } |
1712 | } |
1713 | return FnSize; |
1714 | } |
1715 | |
1716 | void RISCVFrameLowering::processFunctionBeforeFrameFinalized( |
1717 | MachineFunction &MF, RegScavenger *RS) const { |
1718 | const RISCVRegisterInfo *RegInfo = |
1719 | MF.getSubtarget<RISCVSubtarget>().getRegisterInfo(); |
1720 | const RISCVInstrInfo *TII = MF.getSubtarget<RISCVSubtarget>().getInstrInfo(); |
1721 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
1722 | const TargetRegisterClass *RC = &RISCV::GPRRegClass; |
1723 | auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
1724 | |
1725 | int64_t RVVStackSize; |
1726 | Align RVVStackAlign; |
1727 | std::tie(args&: RVVStackSize, args&: RVVStackAlign) = assignRVVStackObjectOffsets(MF); |
1728 | |
1729 | RVFI->setRVVStackSize(RVVStackSize); |
1730 | RVFI->setRVVStackAlign(RVVStackAlign); |
1731 | |
1732 | if (hasRVVFrameObject(MF)) { |
1733 | // Ensure the entire stack is aligned to at least the RVV requirement: some |
1734 | // scalable-vector object alignments are not considered by the |
1735 | // target-independent code. |
1736 | MFI.ensureMaxAlignment(Alignment: RVVStackAlign); |
1737 | } |
1738 | |
1739 | unsigned ScavSlotsNum = 0; |
1740 | |
1741 | // estimateStackSize has been observed to under-estimate the final stack |
1742 | // size, so give ourselves wiggle-room by checking for stack size |
1743 | // representable an 11-bit signed field rather than 12-bits. |
1744 | if (!isInt<11>(x: MFI.estimateStackSize(MF))) |
1745 | ScavSlotsNum = 1; |
1746 | |
1747 | // Far branches over 20-bit offset require a spill slot for scratch register. |
1748 | bool IsLargeFunction = !isInt<20>(x: estimateFunctionSizeInBytes(MF, TII: *TII)); |
1749 | if (IsLargeFunction) |
1750 | ScavSlotsNum = std::max(a: ScavSlotsNum, b: 1u); |
1751 | |
1752 | // RVV loads & stores have no capacity to hold the immediate address offsets |
1753 | // so we must always reserve an emergency spill slot if the MachineFunction |
1754 | // contains any RVV spills. |
1755 | ScavSlotsNum = std::max(a: ScavSlotsNum, b: getScavSlotsNumForRVV(MF)); |
1756 | |
1757 | for (unsigned I = 0; I < ScavSlotsNum; I++) { |
1758 | int FI = MFI.CreateSpillStackObject(Size: RegInfo->getSpillSize(RC: *RC), |
1759 | Alignment: RegInfo->getSpillAlign(RC: *RC)); |
1760 | RS->addScavengingFrameIndex(FI); |
1761 | |
1762 | if (IsLargeFunction && RVFI->getBranchRelaxationScratchFrameIndex() == -1) |
1763 | RVFI->setBranchRelaxationScratchFrameIndex(FI); |
1764 | } |
1765 | |
1766 | unsigned Size = RVFI->getReservedSpillsSize(); |
1767 | for (const auto &Info : MFI.getCalleeSavedInfo()) { |
1768 | int FrameIdx = Info.getFrameIdx(); |
1769 | if (FrameIdx < 0 || MFI.getStackID(ObjectIdx: FrameIdx) != TargetStackID::Default) |
1770 | continue; |
1771 | |
1772 | Size += MFI.getObjectSize(ObjectIdx: FrameIdx); |
1773 | } |
1774 | RVFI->setCalleeSavedStackSize(Size); |
1775 | } |
1776 | |
1777 | // Not preserve stack space within prologue for outgoing variables when the |
1778 | // function contains variable size objects or there are vector objects accessed |
1779 | // by the frame pointer. |
1780 | // Let eliminateCallFramePseudoInstr preserve stack space for it. |
1781 | bool RISCVFrameLowering::hasReservedCallFrame(const MachineFunction &MF) const { |
1782 | return !MF.getFrameInfo().hasVarSizedObjects() && |
1783 | !(hasFP(MF) && hasRVVFrameObject(MF)); |
1784 | } |
1785 | |
1786 | // Eliminate ADJCALLSTACKDOWN, ADJCALLSTACKUP pseudo instructions. |
1787 | MachineBasicBlock::iterator RISCVFrameLowering::eliminateCallFramePseudoInstr( |
1788 | MachineFunction &MF, MachineBasicBlock &MBB, |
1789 | MachineBasicBlock::iterator MI) const { |
1790 | DebugLoc DL = MI->getDebugLoc(); |
1791 | |
1792 | if (!hasReservedCallFrame(MF)) { |
1793 | // If space has not been reserved for a call frame, ADJCALLSTACKDOWN and |
1794 | // ADJCALLSTACKUP must be converted to instructions manipulating the stack |
1795 | // pointer. This is necessary when there is a variable length stack |
1796 | // allocation (e.g. alloca), which means it's not possible to allocate |
1797 | // space for outgoing arguments from within the function prologue. |
1798 | int64_t Amount = MI->getOperand(i: 0).getImm(); |
1799 | |
1800 | if (Amount != 0) { |
1801 | // Ensure the stack remains aligned after adjustment. |
1802 | Amount = alignSPAdjust(SPAdj: Amount); |
1803 | |
1804 | if (MI->getOpcode() == RISCV::ADJCALLSTACKDOWN) |
1805 | Amount = -Amount; |
1806 | |
1807 | const RISCVTargetLowering *TLI = |
1808 | MF.getSubtarget<RISCVSubtarget>().getTargetLowering(); |
1809 | int64_t ProbeSize = TLI->getStackProbeSize(MF, StackAlign: getStackAlign()); |
1810 | if (TLI->hasInlineStackProbe(MF) && -Amount >= ProbeSize) { |
1811 | // When stack probing is enabled, the decrement of SP may need to be |
1812 | // probed. We can handle both the decrement and the probing in |
1813 | // allocateStack. |
1814 | bool DynAllocation = |
1815 | MF.getInfo<RISCVMachineFunctionInfo>()->hasDynamicAllocation(); |
1816 | allocateStack(MBB, MBBI: MI, MF, Offset: -Amount, RealStackSize: -Amount, EmitCFI: !hasFP(MF), |
1817 | /*NeedProbe=*/true, ProbeSize, DynAllocation); |
1818 | } else { |
1819 | const RISCVRegisterInfo &RI = *STI.getRegisterInfo(); |
1820 | RI.adjustReg(MBB, II: MI, DL, DestReg: SPReg, SrcReg: SPReg, Offset: StackOffset::getFixed(Fixed: Amount), |
1821 | Flag: MachineInstr::NoFlags, RequiredAlign: getStackAlign()); |
1822 | } |
1823 | } |
1824 | } |
1825 | |
1826 | return MBB.erase(I: MI); |
1827 | } |
1828 | |
1829 | // We would like to split the SP adjustment to reduce prologue/epilogue |
1830 | // as following instructions. In this way, the offset of the callee saved |
1831 | // register could fit in a single store. Supposed that the first sp adjust |
1832 | // amount is 2032. |
1833 | // add sp,sp,-2032 |
1834 | // sw ra,2028(sp) |
1835 | // sw s0,2024(sp) |
1836 | // sw s1,2020(sp) |
1837 | // sw s3,2012(sp) |
1838 | // sw s4,2008(sp) |
1839 | // add sp,sp,-64 |
1840 | uint64_t |
1841 | RISCVFrameLowering::getFirstSPAdjustAmount(const MachineFunction &MF) const { |
1842 | const auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
1843 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
1844 | const std::vector<CalleeSavedInfo> &CSI = MFI.getCalleeSavedInfo(); |
1845 | uint64_t StackSize = getStackSizeWithRVVPadding(MF); |
1846 | |
1847 | // Disable SplitSPAdjust if save-restore libcall, push/pop or QCI interrupts |
1848 | // are used. The callee-saved registers will be pushed by the save-restore |
1849 | // libcalls, so we don't have to split the SP adjustment in this case. |
1850 | if (RVFI->getReservedSpillsSize()) |
1851 | return 0; |
1852 | |
1853 | // Return the FirstSPAdjustAmount if the StackSize can not fit in a signed |
1854 | // 12-bit and there exists a callee-saved register needing to be pushed. |
1855 | if (!isInt<12>(x: StackSize) && (CSI.size() > 0)) { |
1856 | // FirstSPAdjustAmount is chosen at most as (2048 - StackAlign) because |
1857 | // 2048 will cause sp = sp + 2048 in the epilogue to be split into multiple |
1858 | // instructions. Offsets smaller than 2048 can fit in a single load/store |
1859 | // instruction, and we have to stick with the stack alignment. 2048 has |
1860 | // 16-byte alignment. The stack alignment for RV32 and RV64 is 16 and for |
1861 | // RV32E it is 4. So (2048 - StackAlign) will satisfy the stack alignment. |
1862 | const uint64_t StackAlign = getStackAlign().value(); |
1863 | |
1864 | // Amount of (2048 - StackAlign) will prevent callee saved and restored |
1865 | // instructions be compressed, so try to adjust the amount to the largest |
1866 | // offset that stack compression instructions accept when target supports |
1867 | // compression instructions. |
1868 | if (STI.hasStdExtZca()) { |
1869 | // The compression extensions may support the following instructions: |
1870 | // riscv32: c.lwsp rd, offset[7:2] => 2^(6 + 2) |
1871 | // c.swsp rs2, offset[7:2] => 2^(6 + 2) |
1872 | // c.flwsp rd, offset[7:2] => 2^(6 + 2) |
1873 | // c.fswsp rs2, offset[7:2] => 2^(6 + 2) |
1874 | // riscv64: c.ldsp rd, offset[8:3] => 2^(6 + 3) |
1875 | // c.sdsp rs2, offset[8:3] => 2^(6 + 3) |
1876 | // c.fldsp rd, offset[8:3] => 2^(6 + 3) |
1877 | // c.fsdsp rs2, offset[8:3] => 2^(6 + 3) |
1878 | const uint64_t RVCompressLen = STI.getXLen() * 8; |
1879 | // Compared with amount (2048 - StackAlign), StackSize needs to |
1880 | // satisfy the following conditions to avoid using more instructions |
1881 | // to adjust the sp after adjusting the amount, such as |
1882 | // StackSize meets the condition (StackSize <= 2048 + RVCompressLen), |
1883 | // case1: Amount is 2048 - StackAlign: use addi + addi to adjust sp. |
1884 | // case2: Amount is RVCompressLen: use addi + addi to adjust sp. |
1885 | auto CanCompress = [&](uint64_t CompressLen) -> bool { |
1886 | if (StackSize <= 2047 + CompressLen || |
1887 | (StackSize > 2048 * 2 - StackAlign && |
1888 | StackSize <= 2047 * 2 + CompressLen) || |
1889 | StackSize > 2048 * 3 - StackAlign) |
1890 | return true; |
1891 | |
1892 | return false; |
1893 | }; |
1894 | // In the epilogue, addi sp, sp, 496 is used to recover the sp and it |
1895 | // can be compressed(C.ADDI16SP, offset can be [-512, 496]), but |
1896 | // addi sp, sp, 512 can not be compressed. So try to use 496 first. |
1897 | const uint64_t ADDI16SPCompressLen = 496; |
1898 | if (STI.is64Bit() && CanCompress(ADDI16SPCompressLen)) |
1899 | return ADDI16SPCompressLen; |
1900 | if (CanCompress(RVCompressLen)) |
1901 | return RVCompressLen; |
1902 | } |
1903 | return 2048 - StackAlign; |
1904 | } |
1905 | return 0; |
1906 | } |
1907 | |
1908 | bool RISCVFrameLowering::assignCalleeSavedSpillSlots( |
1909 | MachineFunction &MF, const TargetRegisterInfo *TRI, |
1910 | std::vector<CalleeSavedInfo> &CSI, unsigned &MinCSFrameIndex, |
1911 | unsigned &MaxCSFrameIndex) const { |
1912 | auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
1913 | |
1914 | // Preemptible Interrupts have two additional Callee-save Frame Indexes, |
1915 | // not tracked by `CSI`. |
1916 | if (RVFI->isSiFivePreemptibleInterrupt(MF)) { |
1917 | for (int I = 0; I < 2; ++I) { |
1918 | int FI = RVFI->getInterruptCSRFrameIndex(Idx: I); |
1919 | MinCSFrameIndex = std::min<unsigned>(a: MinCSFrameIndex, b: FI); |
1920 | MaxCSFrameIndex = std::max<unsigned>(a: MaxCSFrameIndex, b: FI); |
1921 | } |
1922 | } |
1923 | |
1924 | // Early exit if no callee saved registers are modified! |
1925 | if (CSI.empty()) |
1926 | return true; |
1927 | |
1928 | if (RVFI->useQCIInterrupt(MF)) { |
1929 | RVFI->setQCIInterruptStackSize(QCIInterruptPushAmount); |
1930 | } |
1931 | |
1932 | if (RVFI->isPushable(MF)) { |
1933 | // Determine how many GPRs we need to push and save it to RVFI. |
1934 | unsigned PushedRegNum = getNumPushPopRegs(CSI); |
1935 | |
1936 | // `QC.C.MIENTER(.NEST)` will save `ra` and `s0`, so we should only push if |
1937 | // we want to push more than 2 registers. Otherwise, we should push if we |
1938 | // want to push more than 0 registers. |
1939 | unsigned OnlyPushIfMoreThan = RVFI->useQCIInterrupt(MF) ? 2 : 0; |
1940 | if (PushedRegNum > OnlyPushIfMoreThan) { |
1941 | RVFI->setRVPushRegs(PushedRegNum); |
1942 | RVFI->setRVPushStackSize(alignTo(Value: (STI.getXLen() / 8) * PushedRegNum, Align: 16)); |
1943 | } |
1944 | } |
1945 | |
1946 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
1947 | const TargetRegisterInfo *RegInfo = MF.getSubtarget().getRegisterInfo(); |
1948 | |
1949 | for (auto &CS : CSI) { |
1950 | MCRegister Reg = CS.getReg(); |
1951 | const TargetRegisterClass *RC = RegInfo->getMinimalPhysRegClass(Reg); |
1952 | unsigned Size = RegInfo->getSpillSize(RC: *RC); |
1953 | |
1954 | if (RVFI->useQCIInterrupt(MF)) { |
1955 | const auto *FFI = llvm::find_if(Range: FixedCSRFIQCIInterruptMap, P: [&](auto P) { |
1956 | return P.first == CS.getReg(); |
1957 | }); |
1958 | if (FFI != std::end(arr: FixedCSRFIQCIInterruptMap)) { |
1959 | int64_t Offset = FFI->second * (int64_t)Size; |
1960 | |
1961 | int FrameIdx = MFI.CreateFixedSpillStackObject(Size, SPOffset: Offset); |
1962 | assert(FrameIdx < 0); |
1963 | CS.setFrameIdx(FrameIdx); |
1964 | continue; |
1965 | } |
1966 | } |
1967 | |
1968 | if (RVFI->useSaveRestoreLibCalls(MF) || RVFI->isPushable(MF)) { |
1969 | const auto *FII = llvm::find_if( |
1970 | Range: FixedCSRFIMap, P: [&](MCPhysReg P) { return P == CS.getReg(); }); |
1971 | unsigned RegNum = std::distance(first: std::begin(arr: FixedCSRFIMap), last: FII); |
1972 | |
1973 | if (FII != std::end(arr: FixedCSRFIMap)) { |
1974 | int64_t Offset; |
1975 | if (RVFI->getPushPopKind(MF) == |
1976 | RISCVMachineFunctionInfo::PushPopKind::StdExtZcmp) |
1977 | Offset = -int64_t(RVFI->getRVPushRegs() - RegNum) * Size; |
1978 | else |
1979 | Offset = -int64_t(RegNum + 1) * Size; |
1980 | |
1981 | if (RVFI->useQCIInterrupt(MF)) |
1982 | Offset -= QCIInterruptPushAmount; |
1983 | |
1984 | int FrameIdx = MFI.CreateFixedSpillStackObject(Size, SPOffset: Offset); |
1985 | assert(FrameIdx < 0); |
1986 | CS.setFrameIdx(FrameIdx); |
1987 | continue; |
1988 | } |
1989 | } |
1990 | |
1991 | // Not a fixed slot. |
1992 | Align Alignment = RegInfo->getSpillAlign(RC: *RC); |
1993 | // We may not be able to satisfy the desired alignment specification of |
1994 | // the TargetRegisterClass if the stack alignment is smaller. Use the |
1995 | // min. |
1996 | Alignment = std::min(a: Alignment, b: getStackAlign()); |
1997 | int FrameIdx = MFI.CreateStackObject(Size, Alignment, isSpillSlot: true); |
1998 | if ((unsigned)FrameIdx < MinCSFrameIndex) |
1999 | MinCSFrameIndex = FrameIdx; |
2000 | if ((unsigned)FrameIdx > MaxCSFrameIndex) |
2001 | MaxCSFrameIndex = FrameIdx; |
2002 | CS.setFrameIdx(FrameIdx); |
2003 | if (RISCVRegisterInfo::isRVVRegClass(RC)) |
2004 | MFI.setStackID(ObjectIdx: FrameIdx, ID: TargetStackID::ScalableVector); |
2005 | } |
2006 | |
2007 | if (RVFI->useQCIInterrupt(MF)) { |
2008 | // Allocate a fixed object that covers the entire QCI stack allocation, |
2009 | // because there are gaps which are reserved for future use. |
2010 | MFI.CreateFixedSpillStackObject( |
2011 | Size: QCIInterruptPushAmount, SPOffset: -static_cast<int64_t>(QCIInterruptPushAmount)); |
2012 | } |
2013 | |
2014 | if (RVFI->isPushable(MF)) { |
2015 | int64_t QCIOffset = RVFI->useQCIInterrupt(MF) ? QCIInterruptPushAmount : 0; |
2016 | // Allocate a fixed object that covers the full push. |
2017 | if (int64_t PushSize = RVFI->getRVPushStackSize()) |
2018 | MFI.CreateFixedSpillStackObject(Size: PushSize, SPOffset: -PushSize - QCIOffset); |
2019 | } else if (int LibCallRegs = getLibCallID(MF, CSI) + 1) { |
2020 | int64_t LibCallFrameSize = |
2021 | alignTo(Size: (STI.getXLen() / 8) * LibCallRegs, A: getStackAlign()); |
2022 | MFI.CreateFixedSpillStackObject(Size: LibCallFrameSize, SPOffset: -LibCallFrameSize); |
2023 | } |
2024 | |
2025 | return true; |
2026 | } |
2027 | |
2028 | bool RISCVFrameLowering::spillCalleeSavedRegisters( |
2029 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, |
2030 | ArrayRef<CalleeSavedInfo> CSI, const TargetRegisterInfo *TRI) const { |
2031 | if (CSI.empty()) |
2032 | return true; |
2033 | |
2034 | MachineFunction *MF = MBB.getParent(); |
2035 | const TargetInstrInfo &TII = *MF->getSubtarget().getInstrInfo(); |
2036 | DebugLoc DL; |
2037 | if (MI != MBB.end() && !MI->isDebugInstr()) |
2038 | DL = MI->getDebugLoc(); |
2039 | |
2040 | RISCVMachineFunctionInfo *RVFI = MF->getInfo<RISCVMachineFunctionInfo>(); |
2041 | if (RVFI->useQCIInterrupt(MF: *MF)) { |
2042 | // Emit QC.C.MIENTER(.NEST) |
2043 | BuildMI( |
2044 | BB&: MBB, I: MI, MIMD: DL, |
2045 | MCID: TII.get(Opcode: RVFI->getInterruptStackKind(MF: *MF) == |
2046 | RISCVMachineFunctionInfo::InterruptStackKind::QCINest |
2047 | ? RISCV::QC_C_MIENTER_NEST |
2048 | : RISCV::QC_C_MIENTER)) |
2049 | .setMIFlag(MachineInstr::FrameSetup); |
2050 | |
2051 | for (auto [Reg, _Offset] : FixedCSRFIQCIInterruptMap) |
2052 | MBB.addLiveIn(PhysReg: Reg); |
2053 | } |
2054 | |
2055 | if (RVFI->isPushable(MF: *MF)) { |
2056 | // Emit CM.PUSH with base StackAdj & evaluate Push stack |
2057 | unsigned PushedRegNum = RVFI->getRVPushRegs(); |
2058 | if (PushedRegNum > 0) { |
2059 | // Use encoded number to represent registers to spill. |
2060 | unsigned Opcode = getPushOpcode( |
2061 | Kind: RVFI->getPushPopKind(MF: *MF), UpdateFP: hasFP(MF: *MF) && !RVFI->useQCIInterrupt(MF: *MF)); |
2062 | unsigned RegEnc = RISCVZC::encodeRegListNumRegs(NumRegs: PushedRegNum); |
2063 | MachineInstrBuilder PushBuilder = |
2064 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII.get(Opcode)) |
2065 | .setMIFlag(MachineInstr::FrameSetup); |
2066 | PushBuilder.addImm(Val: RegEnc); |
2067 | PushBuilder.addImm(Val: 0); |
2068 | |
2069 | for (unsigned i = 0; i < PushedRegNum; i++) |
2070 | PushBuilder.addUse(RegNo: FixedCSRFIMap[i], Flags: RegState::Implicit); |
2071 | } |
2072 | } else if (const char *SpillLibCall = getSpillLibCallName(MF: *MF, CSI)) { |
2073 | // Add spill libcall via non-callee-saved register t0. |
2074 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII.get(Opcode: RISCV::PseudoCALLReg), DestReg: RISCV::X5) |
2075 | .addExternalSymbol(FnName: SpillLibCall, TargetFlags: RISCVII::MO_CALL) |
2076 | .setMIFlag(MachineInstr::FrameSetup); |
2077 | |
2078 | // Add registers spilled in libcall as liveins. |
2079 | for (auto &CS : CSI) |
2080 | MBB.addLiveIn(PhysReg: CS.getReg()); |
2081 | } |
2082 | |
2083 | // Manually spill values not spilled by libcall & Push/Pop. |
2084 | const auto &UnmanagedCSI = getUnmanagedCSI(MF: *MF, CSI); |
2085 | const auto &RVVCSI = getRVVCalleeSavedInfo(MF: *MF, CSI); |
2086 | |
2087 | auto storeRegsToStackSlots = [&](decltype(UnmanagedCSI) CSInfo) { |
2088 | for (auto &CS : CSInfo) { |
2089 | // Insert the spill to the stack frame. |
2090 | MCRegister Reg = CS.getReg(); |
2091 | const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); |
2092 | TII.storeRegToStackSlot(MBB, MI, SrcReg: Reg, isKill: !MBB.isLiveIn(Reg), |
2093 | FrameIndex: CS.getFrameIdx(), RC, TRI, VReg: Register(), |
2094 | Flags: MachineInstr::FrameSetup); |
2095 | } |
2096 | }; |
2097 | storeRegsToStackSlots(UnmanagedCSI); |
2098 | storeRegsToStackSlots(RVVCSI); |
2099 | |
2100 | return true; |
2101 | } |
2102 | |
2103 | static unsigned getCalleeSavedRVVNumRegs(const Register &BaseReg) { |
2104 | return RISCV::VRRegClass.contains(Reg: BaseReg) ? 1 |
2105 | : RISCV::VRM2RegClass.contains(Reg: BaseReg) ? 2 |
2106 | : RISCV::VRM4RegClass.contains(Reg: BaseReg) ? 4 |
2107 | : 8; |
2108 | } |
2109 | |
2110 | static MCRegister getRVVBaseRegister(const RISCVRegisterInfo &TRI, |
2111 | const Register &Reg) { |
2112 | MCRegister BaseReg = TRI.getSubReg(Reg, Idx: RISCV::sub_vrm1_0); |
2113 | // If it's not a grouped vector register, it doesn't have subregister, so |
2114 | // the base register is just itself. |
2115 | if (BaseReg == RISCV::NoRegister) |
2116 | BaseReg = Reg; |
2117 | return BaseReg; |
2118 | } |
2119 | |
2120 | void RISCVFrameLowering::emitCalleeSavedRVVPrologCFI( |
2121 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, bool HasFP) const { |
2122 | MachineFunction *MF = MBB.getParent(); |
2123 | const MachineFrameInfo &MFI = MF->getFrameInfo(); |
2124 | RISCVMachineFunctionInfo *RVFI = MF->getInfo<RISCVMachineFunctionInfo>(); |
2125 | const RISCVRegisterInfo &TRI = *STI.getRegisterInfo(); |
2126 | |
2127 | const auto &RVVCSI = getRVVCalleeSavedInfo(MF: *MF, CSI: MFI.getCalleeSavedInfo()); |
2128 | if (RVVCSI.empty()) |
2129 | return; |
2130 | |
2131 | uint64_t FixedSize = getStackSizeWithRVVPadding(MF: *MF); |
2132 | if (!HasFP) { |
2133 | uint64_t ScalarLocalVarSize = |
2134 | MFI.getStackSize() - RVFI->getCalleeSavedStackSize() - |
2135 | RVFI->getVarArgsSaveSize() + RVFI->getRVVPadding(); |
2136 | FixedSize -= ScalarLocalVarSize; |
2137 | } |
2138 | |
2139 | CFIInstBuilder CFIBuilder(MBB, MI, MachineInstr::FrameSetup); |
2140 | for (auto &CS : RVVCSI) { |
2141 | // Insert the spill to the stack frame. |
2142 | int FI = CS.getFrameIdx(); |
2143 | MCRegister BaseReg = getRVVBaseRegister(TRI, Reg: CS.getReg()); |
2144 | unsigned NumRegs = getCalleeSavedRVVNumRegs(BaseReg: CS.getReg()); |
2145 | for (unsigned i = 0; i < NumRegs; ++i) { |
2146 | CFIBuilder.insertCFIInst(CFIInst: createDefCFAOffset( |
2147 | TRI, Reg: BaseReg + i, FixedOffset: -FixedSize, ScalableOffset: MFI.getObjectOffset(ObjectIdx: FI) / 8 + i)); |
2148 | } |
2149 | } |
2150 | } |
2151 | |
2152 | void RISCVFrameLowering::emitCalleeSavedRVVEpilogCFI( |
2153 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MI) const { |
2154 | MachineFunction *MF = MBB.getParent(); |
2155 | const MachineFrameInfo &MFI = MF->getFrameInfo(); |
2156 | const RISCVRegisterInfo &TRI = *STI.getRegisterInfo(); |
2157 | |
2158 | CFIInstBuilder CFIHelper(MBB, MI, MachineInstr::FrameDestroy); |
2159 | const auto &RVVCSI = getRVVCalleeSavedInfo(MF: *MF, CSI: MFI.getCalleeSavedInfo()); |
2160 | for (auto &CS : RVVCSI) { |
2161 | MCRegister BaseReg = getRVVBaseRegister(TRI, Reg: CS.getReg()); |
2162 | unsigned NumRegs = getCalleeSavedRVVNumRegs(BaseReg: CS.getReg()); |
2163 | for (unsigned i = 0; i < NumRegs; ++i) |
2164 | CFIHelper.buildRestore(Reg: BaseReg + i); |
2165 | } |
2166 | } |
2167 | |
2168 | bool RISCVFrameLowering::restoreCalleeSavedRegisters( |
2169 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, |
2170 | MutableArrayRef<CalleeSavedInfo> CSI, const TargetRegisterInfo *TRI) const { |
2171 | if (CSI.empty()) |
2172 | return true; |
2173 | |
2174 | MachineFunction *MF = MBB.getParent(); |
2175 | const TargetInstrInfo &TII = *MF->getSubtarget().getInstrInfo(); |
2176 | DebugLoc DL; |
2177 | if (MI != MBB.end() && !MI->isDebugInstr()) |
2178 | DL = MI->getDebugLoc(); |
2179 | |
2180 | // Manually restore values not restored by libcall & Push/Pop. |
2181 | // Reverse the restore order in epilog. In addition, the return |
2182 | // address will be restored first in the epilogue. It increases |
2183 | // the opportunity to avoid the load-to-use data hazard between |
2184 | // loading RA and return by RA. loadRegFromStackSlot can insert |
2185 | // multiple instructions. |
2186 | const auto &UnmanagedCSI = getUnmanagedCSI(MF: *MF, CSI); |
2187 | const auto &RVVCSI = getRVVCalleeSavedInfo(MF: *MF, CSI); |
2188 | |
2189 | auto loadRegFromStackSlot = [&](decltype(UnmanagedCSI) CSInfo) { |
2190 | for (auto &CS : CSInfo) { |
2191 | MCRegister Reg = CS.getReg(); |
2192 | const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); |
2193 | TII.loadRegFromStackSlot(MBB, MI, DestReg: Reg, FrameIndex: CS.getFrameIdx(), RC, TRI, |
2194 | VReg: Register(), Flags: MachineInstr::FrameDestroy); |
2195 | assert(MI != MBB.begin() && |
2196 | "loadRegFromStackSlot didn't insert any code!" ); |
2197 | } |
2198 | }; |
2199 | loadRegFromStackSlot(RVVCSI); |
2200 | loadRegFromStackSlot(UnmanagedCSI); |
2201 | |
2202 | RISCVMachineFunctionInfo *RVFI = MF->getInfo<RISCVMachineFunctionInfo>(); |
2203 | if (RVFI->useQCIInterrupt(MF: *MF)) { |
2204 | // Don't emit anything here because restoration is handled by |
2205 | // QC.C.MILEAVERET which we already inserted to return. |
2206 | assert(MI->getOpcode() == RISCV::QC_C_MILEAVERET && |
2207 | "Unexpected QCI Interrupt Return Instruction" ); |
2208 | } |
2209 | |
2210 | if (RVFI->isPushable(MF: *MF)) { |
2211 | unsigned PushedRegNum = RVFI->getRVPushRegs(); |
2212 | if (PushedRegNum > 0) { |
2213 | unsigned Opcode = getPopOpcode(Kind: RVFI->getPushPopKind(MF: *MF)); |
2214 | unsigned RegEnc = RISCVZC::encodeRegListNumRegs(NumRegs: PushedRegNum); |
2215 | MachineInstrBuilder PopBuilder = |
2216 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII.get(Opcode)) |
2217 | .setMIFlag(MachineInstr::FrameDestroy); |
2218 | // Use encoded number to represent registers to restore. |
2219 | PopBuilder.addImm(Val: RegEnc); |
2220 | PopBuilder.addImm(Val: 0); |
2221 | |
2222 | for (unsigned i = 0; i < RVFI->getRVPushRegs(); i++) |
2223 | PopBuilder.addDef(RegNo: FixedCSRFIMap[i], Flags: RegState::ImplicitDefine); |
2224 | } |
2225 | } else { |
2226 | const char *RestoreLibCall = getRestoreLibCallName(MF: *MF, CSI); |
2227 | if (RestoreLibCall) { |
2228 | // Add restore libcall via tail call. |
2229 | MachineBasicBlock::iterator NewMI = |
2230 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII.get(Opcode: RISCV::PseudoTAIL)) |
2231 | .addExternalSymbol(FnName: RestoreLibCall, TargetFlags: RISCVII::MO_CALL) |
2232 | .setMIFlag(MachineInstr::FrameDestroy); |
2233 | |
2234 | // Remove trailing returns, since the terminator is now a tail call to the |
2235 | // restore function. |
2236 | if (MI != MBB.end() && MI->getOpcode() == RISCV::PseudoRET) { |
2237 | NewMI->copyImplicitOps(MF&: *MF, MI: *MI); |
2238 | MI->eraseFromParent(); |
2239 | } |
2240 | } |
2241 | } |
2242 | return true; |
2243 | } |
2244 | |
2245 | bool RISCVFrameLowering::enableShrinkWrapping(const MachineFunction &MF) const { |
2246 | // Keep the conventional code flow when not optimizing. |
2247 | if (MF.getFunction().hasOptNone()) |
2248 | return false; |
2249 | |
2250 | return true; |
2251 | } |
2252 | |
2253 | bool RISCVFrameLowering::canUseAsPrologue(const MachineBasicBlock &MBB) const { |
2254 | MachineBasicBlock *TmpMBB = const_cast<MachineBasicBlock *>(&MBB); |
2255 | const MachineFunction *MF = MBB.getParent(); |
2256 | const auto *RVFI = MF->getInfo<RISCVMachineFunctionInfo>(); |
2257 | |
2258 | // Make sure VTYPE and VL are not live-in since we will use vsetvli in the |
2259 | // prologue to get the VLEN, and that will clobber these registers. |
2260 | // |
2261 | // We may do also check the stack contains objects with scalable vector type, |
2262 | // but this will require iterating over all the stack objects, but this may |
2263 | // not worth since the situation is rare, we could do further check in future |
2264 | // if we find it is necessary. |
2265 | if (STI.preferVsetvliOverReadVLENB() && |
2266 | (MBB.isLiveIn(Reg: RISCV::VTYPE) || MBB.isLiveIn(Reg: RISCV::VL))) |
2267 | return false; |
2268 | |
2269 | if (!RVFI->useSaveRestoreLibCalls(MF: *MF)) |
2270 | return true; |
2271 | |
2272 | // Inserting a call to a __riscv_save libcall requires the use of the register |
2273 | // t0 (X5) to hold the return address. Therefore if this register is already |
2274 | // used we can't insert the call. |
2275 | |
2276 | RegScavenger RS; |
2277 | RS.enterBasicBlock(MBB&: *TmpMBB); |
2278 | return !RS.isRegUsed(Reg: RISCV::X5); |
2279 | } |
2280 | |
2281 | bool RISCVFrameLowering::canUseAsEpilogue(const MachineBasicBlock &MBB) const { |
2282 | const MachineFunction *MF = MBB.getParent(); |
2283 | MachineBasicBlock *TmpMBB = const_cast<MachineBasicBlock *>(&MBB); |
2284 | const auto *RVFI = MF->getInfo<RISCVMachineFunctionInfo>(); |
2285 | |
2286 | // We do not want QC.C.MILEAVERET to be subject to shrink-wrapping - it must |
2287 | // come in the final block of its function as it both pops and returns. |
2288 | if (RVFI->useQCIInterrupt(MF: *MF)) |
2289 | return MBB.succ_empty(); |
2290 | |
2291 | if (!RVFI->useSaveRestoreLibCalls(MF: *MF)) |
2292 | return true; |
2293 | |
2294 | // Using the __riscv_restore libcalls to restore CSRs requires a tail call. |
2295 | // This means if we still need to continue executing code within this function |
2296 | // the restore cannot take place in this basic block. |
2297 | |
2298 | if (MBB.succ_size() > 1) |
2299 | return false; |
2300 | |
2301 | MachineBasicBlock *SuccMBB = |
2302 | MBB.succ_empty() ? TmpMBB->getFallThrough() : *MBB.succ_begin(); |
2303 | |
2304 | // Doing a tail call should be safe if there are no successors, because either |
2305 | // we have a returning block or the end of the block is unreachable, so the |
2306 | // restore will be eliminated regardless. |
2307 | if (!SuccMBB) |
2308 | return true; |
2309 | |
2310 | // The successor can only contain a return, since we would effectively be |
2311 | // replacing the successor with our own tail return at the end of our block. |
2312 | return SuccMBB->isReturnBlock() && SuccMBB->size() == 1; |
2313 | } |
2314 | |
2315 | bool RISCVFrameLowering::isSupportedStackID(TargetStackID::Value ID) const { |
2316 | switch (ID) { |
2317 | case TargetStackID::Default: |
2318 | case TargetStackID::ScalableVector: |
2319 | return true; |
2320 | case TargetStackID::NoAlloc: |
2321 | case TargetStackID::SGPRSpill: |
2322 | case TargetStackID::WasmLocal: |
2323 | return false; |
2324 | } |
2325 | llvm_unreachable("Invalid TargetStackID::Value" ); |
2326 | } |
2327 | |
2328 | TargetStackID::Value RISCVFrameLowering::getStackIDForScalableVectors() const { |
2329 | return TargetStackID::ScalableVector; |
2330 | } |
2331 | |
2332 | // Synthesize the probe loop. |
2333 | static void emitStackProbeInline(MachineBasicBlock::iterator MBBI, DebugLoc DL, |
2334 | Register TargetReg, bool IsRVV) { |
2335 | assert(TargetReg != RISCV::X2 && "New top of stack cannot already be in SP" ); |
2336 | |
2337 | MachineBasicBlock &MBB = *MBBI->getParent(); |
2338 | MachineFunction &MF = *MBB.getParent(); |
2339 | |
2340 | auto &Subtarget = MF.getSubtarget<RISCVSubtarget>(); |
2341 | const RISCVInstrInfo *TII = Subtarget.getInstrInfo(); |
2342 | bool IsRV64 = Subtarget.is64Bit(); |
2343 | Align StackAlign = Subtarget.getFrameLowering()->getStackAlign(); |
2344 | const RISCVTargetLowering *TLI = Subtarget.getTargetLowering(); |
2345 | uint64_t ProbeSize = TLI->getStackProbeSize(MF, StackAlign); |
2346 | |
2347 | MachineFunction::iterator MBBInsertPoint = std::next(x: MBB.getIterator()); |
2348 | MachineBasicBlock *LoopTestMBB = |
2349 | MF.CreateMachineBasicBlock(BB: MBB.getBasicBlock()); |
2350 | MF.insert(MBBI: MBBInsertPoint, MBB: LoopTestMBB); |
2351 | MachineBasicBlock *ExitMBB = MF.CreateMachineBasicBlock(BB: MBB.getBasicBlock()); |
2352 | MF.insert(MBBI: MBBInsertPoint, MBB: ExitMBB); |
2353 | MachineInstr::MIFlag Flags = MachineInstr::FrameSetup; |
2354 | Register ScratchReg = RISCV::X7; |
2355 | |
2356 | // ScratchReg = ProbeSize |
2357 | TII->movImm(MBB, MBBI, DL, DstReg: ScratchReg, Val: ProbeSize, Flag: Flags); |
2358 | |
2359 | // LoopTest: |
2360 | // SUB SP, SP, ProbeSize |
2361 | BuildMI(BB&: *LoopTestMBB, I: LoopTestMBB->end(), MIMD: DL, MCID: TII->get(Opcode: RISCV::SUB), DestReg: SPReg) |
2362 | .addReg(RegNo: SPReg) |
2363 | .addReg(RegNo: ScratchReg) |
2364 | .setMIFlags(Flags); |
2365 | |
2366 | // s[d|w] zero, 0(sp) |
2367 | BuildMI(BB&: *LoopTestMBB, I: LoopTestMBB->end(), MIMD: DL, |
2368 | MCID: TII->get(Opcode: IsRV64 ? RISCV::SD : RISCV::SW)) |
2369 | .addReg(RegNo: RISCV::X0) |
2370 | .addReg(RegNo: SPReg) |
2371 | .addImm(Val: 0) |
2372 | .setMIFlags(Flags); |
2373 | |
2374 | if (IsRVV) { |
2375 | // SUB TargetReg, TargetReg, ProbeSize |
2376 | BuildMI(BB&: *LoopTestMBB, I: LoopTestMBB->end(), MIMD: DL, MCID: TII->get(Opcode: RISCV::SUB), |
2377 | DestReg: TargetReg) |
2378 | .addReg(RegNo: TargetReg) |
2379 | .addReg(RegNo: ScratchReg) |
2380 | .setMIFlags(Flags); |
2381 | |
2382 | // BGE TargetReg, ProbeSize, LoopTest |
2383 | BuildMI(BB&: *LoopTestMBB, I: LoopTestMBB->end(), MIMD: DL, MCID: TII->get(Opcode: RISCV::BGE)) |
2384 | .addReg(RegNo: TargetReg) |
2385 | .addReg(RegNo: ScratchReg) |
2386 | .addMBB(MBB: LoopTestMBB) |
2387 | .setMIFlags(Flags); |
2388 | |
2389 | } else { |
2390 | // BNE SP, TargetReg, LoopTest |
2391 | BuildMI(BB&: *LoopTestMBB, I: LoopTestMBB->end(), MIMD: DL, MCID: TII->get(Opcode: RISCV::BNE)) |
2392 | .addReg(RegNo: SPReg) |
2393 | .addReg(RegNo: TargetReg) |
2394 | .addMBB(MBB: LoopTestMBB) |
2395 | .setMIFlags(Flags); |
2396 | } |
2397 | |
2398 | ExitMBB->splice(Where: ExitMBB->end(), Other: &MBB, From: std::next(x: MBBI), To: MBB.end()); |
2399 | ExitMBB->transferSuccessorsAndUpdatePHIs(FromMBB: &MBB); |
2400 | |
2401 | LoopTestMBB->addSuccessor(Succ: ExitMBB); |
2402 | LoopTestMBB->addSuccessor(Succ: LoopTestMBB); |
2403 | MBB.addSuccessor(Succ: LoopTestMBB); |
2404 | // Update liveins. |
2405 | fullyRecomputeLiveIns(MBBs: {ExitMBB, LoopTestMBB}); |
2406 | } |
2407 | |
2408 | void RISCVFrameLowering::inlineStackProbe(MachineFunction &MF, |
2409 | MachineBasicBlock &MBB) const { |
2410 | // Get the instructions that need to be replaced. We emit at most two of |
2411 | // these. Remember them in order to avoid complications coming from the need |
2412 | // to traverse the block while potentially creating more blocks. |
2413 | SmallVector<MachineInstr *, 4> ToReplace; |
2414 | for (MachineInstr &MI : MBB) { |
2415 | unsigned Opc = MI.getOpcode(); |
2416 | if (Opc == RISCV::PROBED_STACKALLOC || |
2417 | Opc == RISCV::PROBED_STACKALLOC_RVV) { |
2418 | ToReplace.push_back(Elt: &MI); |
2419 | } |
2420 | } |
2421 | |
2422 | for (MachineInstr *MI : ToReplace) { |
2423 | if (MI->getOpcode() == RISCV::PROBED_STACKALLOC || |
2424 | MI->getOpcode() == RISCV::PROBED_STACKALLOC_RVV) { |
2425 | MachineBasicBlock::iterator MBBI = MI->getIterator(); |
2426 | DebugLoc DL = MBB.findDebugLoc(MBBI); |
2427 | Register TargetReg = MI->getOperand(i: 0).getReg(); |
2428 | emitStackProbeInline(MBBI, DL, TargetReg, |
2429 | IsRVV: (MI->getOpcode() == RISCV::PROBED_STACKALLOC_RVV)); |
2430 | MBBI->eraseFromParent(); |
2431 | } |
2432 | } |
2433 | } |
2434 | |