1 | //===- SIMachineFunctionInfo.cpp - SI Machine Function Info ---------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | |
9 | #include "SIMachineFunctionInfo.h" |
10 | #include "AMDGPUSubtarget.h" |
11 | #include "GCNSubtarget.h" |
12 | #include "MCTargetDesc/AMDGPUMCTargetDesc.h" |
13 | #include "SIRegisterInfo.h" |
14 | #include "Utils/AMDGPUBaseInfo.h" |
15 | #include "llvm/CodeGen/LiveIntervals.h" |
16 | #include "llvm/CodeGen/MIRParser/MIParser.h" |
17 | #include "llvm/CodeGen/MachineBasicBlock.h" |
18 | #include "llvm/CodeGen/MachineFrameInfo.h" |
19 | #include "llvm/CodeGen/MachineFunction.h" |
20 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
21 | #include "llvm/IR/CallingConv.h" |
22 | #include "llvm/IR/DiagnosticInfo.h" |
23 | #include "llvm/IR/Function.h" |
24 | #include <cassert> |
25 | #include <optional> |
26 | #include <vector> |
27 | |
28 | enum { MAX_LANES = 64 }; |
29 | |
30 | using namespace llvm; |
31 | |
32 | const GCNTargetMachine &getTM(const GCNSubtarget *STI) { |
33 | const SITargetLowering *TLI = STI->getTargetLowering(); |
34 | return static_cast<const GCNTargetMachine &>(TLI->getTargetMachine()); |
35 | } |
36 | |
37 | SIMachineFunctionInfo::SIMachineFunctionInfo(const Function &F, |
38 | const GCNSubtarget *STI) |
39 | : AMDGPUMachineFunction(F, *STI), Mode(F, *STI), GWSResourcePSV(getTM(STI)), |
40 | UserSGPRInfo(F, *STI), WorkGroupIDX(false), WorkGroupIDY(false), |
41 | WorkGroupIDZ(false), WorkGroupInfo(false), LDSKernelId(false), |
42 | PrivateSegmentWaveByteOffset(false), WorkItemIDX(false), |
43 | WorkItemIDY(false), WorkItemIDZ(false), ImplicitArgPtr(false), |
44 | GITPtrHigh(0xffffffff), HighBitsOf32BitAddress(0) { |
45 | const GCNSubtarget &ST = *static_cast<const GCNSubtarget *>(STI); |
46 | FlatWorkGroupSizes = ST.getFlatWorkGroupSizes(F); |
47 | WavesPerEU = ST.getWavesPerEU(F); |
48 | MaxNumWorkGroups = ST.getMaxNumWorkGroups(F); |
49 | assert(MaxNumWorkGroups.size() == 3); |
50 | |
51 | // Temporarily check both the attribute and the subtarget feature, until the |
52 | // latter is completely removed. |
53 | DynamicVGPRBlockSize = AMDGPU::getDynamicVGPRBlockSize(F); |
54 | if (DynamicVGPRBlockSize == 0 && ST.isDynamicVGPREnabled()) |
55 | DynamicVGPRBlockSize = ST.getDynamicVGPRBlockSize(); |
56 | |
57 | Occupancy = ST.computeOccupancy(F, LDSSize: getLDSSize()).second; |
58 | CallingConv::ID CC = F.getCallingConv(); |
59 | |
60 | VRegFlags.reserve(s: 1024); |
61 | |
62 | const bool IsKernel = CC == CallingConv::AMDGPU_KERNEL || |
63 | CC == CallingConv::SPIR_KERNEL; |
64 | |
65 | if (IsKernel) { |
66 | WorkGroupIDX = true; |
67 | WorkItemIDX = true; |
68 | } else if (CC == CallingConv::AMDGPU_PS) { |
69 | PSInputAddr = AMDGPU::getInitialPSInputAddr(F); |
70 | } |
71 | |
72 | MayNeedAGPRs = ST.hasMAIInsts(); |
73 | if (ST.hasGFX90AInsts() && |
74 | ST.getMaxNumVGPRs(F) <= AMDGPU::VGPR_32RegClass.getNumRegs() && |
75 | !mayUseAGPRs(F)) |
76 | MayNeedAGPRs = false; // We will select all MAI with VGPR operands. |
77 | |
78 | if (AMDGPU::isChainCC(CC)) { |
79 | // Chain functions don't receive an SP from their caller, but are free to |
80 | // set one up. For now, we can use s32 to match what amdgpu_gfx functions |
81 | // would use if called, but this can be revisited. |
82 | // FIXME: Only reserve this if we actually need it. |
83 | StackPtrOffsetReg = AMDGPU::SGPR32; |
84 | |
85 | ScratchRSrcReg = AMDGPU::SGPR48_SGPR49_SGPR50_SGPR51; |
86 | |
87 | ArgInfo.PrivateSegmentBuffer = |
88 | ArgDescriptor::createRegister(Reg: ScratchRSrcReg); |
89 | |
90 | ImplicitArgPtr = false; |
91 | } else if (!isEntryFunction()) { |
92 | if (CC != CallingConv::AMDGPU_Gfx) |
93 | ArgInfo = AMDGPUArgumentUsageInfo::FixedABIFunctionInfo; |
94 | |
95 | FrameOffsetReg = AMDGPU::SGPR33; |
96 | StackPtrOffsetReg = AMDGPU::SGPR32; |
97 | |
98 | if (!ST.enableFlatScratch()) { |
99 | // Non-entry functions have no special inputs for now, other registers |
100 | // required for scratch access. |
101 | ScratchRSrcReg = AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3; |
102 | |
103 | ArgInfo.PrivateSegmentBuffer = |
104 | ArgDescriptor::createRegister(Reg: ScratchRSrcReg); |
105 | } |
106 | |
107 | if (!F.hasFnAttribute(Kind: "amdgpu-no-implicitarg-ptr" )) |
108 | ImplicitArgPtr = true; |
109 | } else { |
110 | ImplicitArgPtr = false; |
111 | MaxKernArgAlign = |
112 | std::max(a: ST.getAlignmentForImplicitArgPtr(), b: MaxKernArgAlign); |
113 | } |
114 | |
115 | if (!AMDGPU::isGraphics(CC) || |
116 | ((CC == CallingConv::AMDGPU_CS || CC == CallingConv::AMDGPU_Gfx) && |
117 | ST.hasArchitectedSGPRs())) { |
118 | if (IsKernel || !F.hasFnAttribute(Kind: "amdgpu-no-workgroup-id-x" )) |
119 | WorkGroupIDX = true; |
120 | |
121 | if (!F.hasFnAttribute(Kind: "amdgpu-no-workgroup-id-y" )) |
122 | WorkGroupIDY = true; |
123 | |
124 | if (!F.hasFnAttribute(Kind: "amdgpu-no-workgroup-id-z" )) |
125 | WorkGroupIDZ = true; |
126 | } |
127 | |
128 | if (!AMDGPU::isGraphics(CC)) { |
129 | if (IsKernel || !F.hasFnAttribute(Kind: "amdgpu-no-workitem-id-x" )) |
130 | WorkItemIDX = true; |
131 | |
132 | if (!F.hasFnAttribute(Kind: "amdgpu-no-workitem-id-y" ) && |
133 | ST.getMaxWorkitemID(Kernel: F, Dimension: 1) != 0) |
134 | WorkItemIDY = true; |
135 | |
136 | if (!F.hasFnAttribute(Kind: "amdgpu-no-workitem-id-z" ) && |
137 | ST.getMaxWorkitemID(Kernel: F, Dimension: 2) != 0) |
138 | WorkItemIDZ = true; |
139 | |
140 | if (!IsKernel && !F.hasFnAttribute(Kind: "amdgpu-no-lds-kernel-id" )) |
141 | LDSKernelId = true; |
142 | } |
143 | |
144 | if (isEntryFunction()) { |
145 | // X, XY, and XYZ are the only supported combinations, so make sure Y is |
146 | // enabled if Z is. |
147 | if (WorkItemIDZ) |
148 | WorkItemIDY = true; |
149 | |
150 | if (!ST.flatScratchIsArchitected()) { |
151 | PrivateSegmentWaveByteOffset = true; |
152 | |
153 | // HS and GS always have the scratch wave offset in SGPR5 on GFX9. |
154 | if (ST.getGeneration() >= AMDGPUSubtarget::GFX9 && |
155 | (CC == CallingConv::AMDGPU_HS || CC == CallingConv::AMDGPU_GS)) |
156 | ArgInfo.PrivateSegmentWaveByteOffset = |
157 | ArgDescriptor::createRegister(Reg: AMDGPU::SGPR5); |
158 | } |
159 | } |
160 | |
161 | Attribute A = F.getFnAttribute(Kind: "amdgpu-git-ptr-high" ); |
162 | StringRef S = A.getValueAsString(); |
163 | if (!S.empty()) |
164 | S.consumeInteger(Radix: 0, Result&: GITPtrHigh); |
165 | |
166 | A = F.getFnAttribute(Kind: "amdgpu-32bit-address-high-bits" ); |
167 | S = A.getValueAsString(); |
168 | if (!S.empty()) |
169 | S.consumeInteger(Radix: 0, Result&: HighBitsOf32BitAddress); |
170 | |
171 | MaxMemoryClusterDWords = F.getFnAttributeAsParsedInteger( |
172 | Kind: "amdgpu-max-memory-cluster-dwords" , Default: DefaultMemoryClusterDWordsLimit); |
173 | |
174 | // On GFX908, in order to guarantee copying between AGPRs, we need a scratch |
175 | // VGPR available at all times. For now, reserve highest available VGPR. After |
176 | // RA, shift it to the lowest available unused VGPR if the one exist. |
177 | if (ST.hasMAIInsts() && !ST.hasGFX90AInsts()) { |
178 | VGPRForAGPRCopy = |
179 | AMDGPU::VGPR_32RegClass.getRegister(i: ST.getMaxNumVGPRs(F) - 1); |
180 | } |
181 | } |
182 | |
183 | MachineFunctionInfo *SIMachineFunctionInfo::clone( |
184 | BumpPtrAllocator &Allocator, MachineFunction &DestMF, |
185 | const DenseMap<MachineBasicBlock *, MachineBasicBlock *> &Src2DstMBB) |
186 | const { |
187 | return DestMF.cloneInfo<SIMachineFunctionInfo>(Old: *this); |
188 | } |
189 | |
190 | void SIMachineFunctionInfo::limitOccupancy(const MachineFunction &MF) { |
191 | limitOccupancy(Limit: getMaxWavesPerEU()); |
192 | const GCNSubtarget& ST = MF.getSubtarget<GCNSubtarget>(); |
193 | limitOccupancy(Limit: ST.getOccupancyWithWorkGroupSizes(MF).second); |
194 | } |
195 | |
196 | Register SIMachineFunctionInfo::addPrivateSegmentBuffer( |
197 | const SIRegisterInfo &TRI) { |
198 | ArgInfo.PrivateSegmentBuffer = |
199 | ArgDescriptor::createRegister(Reg: TRI.getMatchingSuperReg( |
200 | Reg: getNextUserSGPR(), SubIdx: AMDGPU::sub0, RC: &AMDGPU::SGPR_128RegClass)); |
201 | NumUserSGPRs += 4; |
202 | return ArgInfo.PrivateSegmentBuffer.getRegister(); |
203 | } |
204 | |
205 | Register SIMachineFunctionInfo::addDispatchPtr(const SIRegisterInfo &TRI) { |
206 | ArgInfo.DispatchPtr = ArgDescriptor::createRegister(Reg: TRI.getMatchingSuperReg( |
207 | Reg: getNextUserSGPR(), SubIdx: AMDGPU::sub0, RC: &AMDGPU::SReg_64RegClass)); |
208 | NumUserSGPRs += 2; |
209 | return ArgInfo.DispatchPtr.getRegister(); |
210 | } |
211 | |
212 | Register SIMachineFunctionInfo::addQueuePtr(const SIRegisterInfo &TRI) { |
213 | ArgInfo.QueuePtr = ArgDescriptor::createRegister(Reg: TRI.getMatchingSuperReg( |
214 | Reg: getNextUserSGPR(), SubIdx: AMDGPU::sub0, RC: &AMDGPU::SReg_64RegClass)); |
215 | NumUserSGPRs += 2; |
216 | return ArgInfo.QueuePtr.getRegister(); |
217 | } |
218 | |
219 | Register SIMachineFunctionInfo::addKernargSegmentPtr(const SIRegisterInfo &TRI) { |
220 | ArgInfo.KernargSegmentPtr |
221 | = ArgDescriptor::createRegister(Reg: TRI.getMatchingSuperReg( |
222 | Reg: getNextUserSGPR(), SubIdx: AMDGPU::sub0, RC: &AMDGPU::SReg_64RegClass)); |
223 | NumUserSGPRs += 2; |
224 | return ArgInfo.KernargSegmentPtr.getRegister(); |
225 | } |
226 | |
227 | Register SIMachineFunctionInfo::addDispatchID(const SIRegisterInfo &TRI) { |
228 | ArgInfo.DispatchID = ArgDescriptor::createRegister(Reg: TRI.getMatchingSuperReg( |
229 | Reg: getNextUserSGPR(), SubIdx: AMDGPU::sub0, RC: &AMDGPU::SReg_64RegClass)); |
230 | NumUserSGPRs += 2; |
231 | return ArgInfo.DispatchID.getRegister(); |
232 | } |
233 | |
234 | Register SIMachineFunctionInfo::addFlatScratchInit(const SIRegisterInfo &TRI) { |
235 | ArgInfo.FlatScratchInit = ArgDescriptor::createRegister(Reg: TRI.getMatchingSuperReg( |
236 | Reg: getNextUserSGPR(), SubIdx: AMDGPU::sub0, RC: &AMDGPU::SReg_64RegClass)); |
237 | NumUserSGPRs += 2; |
238 | return ArgInfo.FlatScratchInit.getRegister(); |
239 | } |
240 | |
241 | Register SIMachineFunctionInfo::addPrivateSegmentSize(const SIRegisterInfo &TRI) { |
242 | ArgInfo.PrivateSegmentSize = ArgDescriptor::createRegister(Reg: getNextUserSGPR()); |
243 | NumUserSGPRs += 1; |
244 | return ArgInfo.PrivateSegmentSize.getRegister(); |
245 | } |
246 | |
247 | Register SIMachineFunctionInfo::addImplicitBufferPtr(const SIRegisterInfo &TRI) { |
248 | ArgInfo.ImplicitBufferPtr = ArgDescriptor::createRegister(Reg: TRI.getMatchingSuperReg( |
249 | Reg: getNextUserSGPR(), SubIdx: AMDGPU::sub0, RC: &AMDGPU::SReg_64RegClass)); |
250 | NumUserSGPRs += 2; |
251 | return ArgInfo.ImplicitBufferPtr.getRegister(); |
252 | } |
253 | |
254 | Register SIMachineFunctionInfo::addLDSKernelId() { |
255 | ArgInfo.LDSKernelId = ArgDescriptor::createRegister(Reg: getNextUserSGPR()); |
256 | NumUserSGPRs += 1; |
257 | return ArgInfo.LDSKernelId.getRegister(); |
258 | } |
259 | |
260 | SmallVectorImpl<MCRegister> *SIMachineFunctionInfo::addPreloadedKernArg( |
261 | const SIRegisterInfo &TRI, const TargetRegisterClass *RC, |
262 | unsigned AllocSizeDWord, int KernArgIdx, int PaddingSGPRs) { |
263 | auto [It, Inserted] = ArgInfo.PreloadKernArgs.try_emplace(Key: KernArgIdx); |
264 | assert(Inserted && "Preload kernel argument allocated twice." ); |
265 | NumUserSGPRs += PaddingSGPRs; |
266 | // If the available register tuples are aligned with the kernarg to be |
267 | // preloaded use that register, otherwise we need to use a set of SGPRs and |
268 | // merge them. |
269 | if (!ArgInfo.FirstKernArgPreloadReg) |
270 | ArgInfo.FirstKernArgPreloadReg = getNextUserSGPR(); |
271 | Register PreloadReg = |
272 | TRI.getMatchingSuperReg(Reg: getNextUserSGPR(), SubIdx: AMDGPU::sub0, RC); |
273 | auto &Regs = It->second.Regs; |
274 | if (PreloadReg && |
275 | (RC == &AMDGPU::SReg_32RegClass || RC == &AMDGPU::SReg_64RegClass)) { |
276 | Regs.push_back(Elt: PreloadReg); |
277 | NumUserSGPRs += AllocSizeDWord; |
278 | } else { |
279 | Regs.reserve(N: AllocSizeDWord); |
280 | for (unsigned I = 0; I < AllocSizeDWord; ++I) { |
281 | Regs.push_back(Elt: getNextUserSGPR()); |
282 | NumUserSGPRs++; |
283 | } |
284 | } |
285 | |
286 | // Track the actual number of SGPRs that HW will preload to. |
287 | UserSGPRInfo.allocKernargPreloadSGPRs(NumSGPRs: AllocSizeDWord + PaddingSGPRs); |
288 | return &Regs; |
289 | } |
290 | |
291 | void SIMachineFunctionInfo::allocateWWMSpill(MachineFunction &MF, Register VGPR, |
292 | uint64_t Size, Align Alignment) { |
293 | // Skip if it is an entry function or the register is already added. |
294 | if (isEntryFunction() || WWMSpills.count(Key: VGPR)) |
295 | return; |
296 | |
297 | // Skip if this is a function with the amdgpu_cs_chain or |
298 | // amdgpu_cs_chain_preserve calling convention and this is a scratch register. |
299 | // We never need to allocate a spill for these because we don't even need to |
300 | // restore the inactive lanes for them (they're scratchier than the usual |
301 | // scratch registers). We only need to do this if we have calls to |
302 | // llvm.amdgcn.cs.chain (otherwise there's no one to save them for, since |
303 | // chain functions do not return) and the function did not contain a call to |
304 | // llvm.amdgcn.init.whole.wave (since in that case there are no inactive lanes |
305 | // when entering the function). |
306 | if (isChainFunction() && |
307 | (SIRegisterInfo::isChainScratchRegister(VGPR) || |
308 | !MF.getFrameInfo().hasTailCall() || hasInitWholeWave())) |
309 | return; |
310 | |
311 | WWMSpills.insert(KV: std::make_pair( |
312 | x&: VGPR, y: MF.getFrameInfo().CreateSpillStackObject(Size, Alignment))); |
313 | } |
314 | |
315 | // Separate out the callee-saved and scratch registers. |
316 | void SIMachineFunctionInfo::splitWWMSpillRegisters( |
317 | MachineFunction &MF, |
318 | SmallVectorImpl<std::pair<Register, int>> &CalleeSavedRegs, |
319 | SmallVectorImpl<std::pair<Register, int>> &ScratchRegs) const { |
320 | const MCPhysReg *CSRegs = MF.getRegInfo().getCalleeSavedRegs(); |
321 | for (auto &Reg : WWMSpills) { |
322 | if (isCalleeSavedReg(CSRegs, Reg: Reg.first)) |
323 | CalleeSavedRegs.push_back(Elt: Reg); |
324 | else |
325 | ScratchRegs.push_back(Elt: Reg); |
326 | } |
327 | } |
328 | |
329 | bool SIMachineFunctionInfo::isCalleeSavedReg(const MCPhysReg *CSRegs, |
330 | MCPhysReg Reg) const { |
331 | for (unsigned I = 0; CSRegs[I]; ++I) { |
332 | if (CSRegs[I] == Reg) |
333 | return true; |
334 | } |
335 | |
336 | return false; |
337 | } |
338 | |
339 | void SIMachineFunctionInfo::shiftWwmVGPRsToLowestRange( |
340 | MachineFunction &MF, SmallVectorImpl<Register> &WWMVGPRs, |
341 | BitVector &SavedVGPRs) { |
342 | const SIRegisterInfo *TRI = MF.getSubtarget<GCNSubtarget>().getRegisterInfo(); |
343 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
344 | for (unsigned I = 0, E = WWMVGPRs.size(); I < E; ++I) { |
345 | Register Reg = WWMVGPRs[I]; |
346 | Register NewReg = |
347 | TRI->findUnusedRegister(MRI, RC: &AMDGPU::VGPR_32RegClass, MF); |
348 | if (!NewReg || NewReg >= Reg) |
349 | break; |
350 | |
351 | MRI.replaceRegWith(FromReg: Reg, ToReg: NewReg); |
352 | |
353 | // Update various tables with the new VGPR. |
354 | WWMVGPRs[I] = NewReg; |
355 | WWMReservedRegs.remove(X: Reg); |
356 | WWMReservedRegs.insert(X: NewReg); |
357 | MRI.reserveReg(PhysReg: NewReg, TRI); |
358 | |
359 | // Replace the register in SpillPhysVGPRs. This is needed to look for free |
360 | // lanes while spilling special SGPRs like FP, BP, etc. during PEI. |
361 | auto *RegItr = llvm::find(Range&: SpillPhysVGPRs, Val: Reg); |
362 | if (RegItr != SpillPhysVGPRs.end()) { |
363 | unsigned Idx = std::distance(first: SpillPhysVGPRs.begin(), last: RegItr); |
364 | SpillPhysVGPRs[Idx] = NewReg; |
365 | } |
366 | |
367 | // The generic `determineCalleeSaves` might have set the old register if it |
368 | // is in the CSR range. |
369 | SavedVGPRs.reset(Idx: Reg); |
370 | |
371 | for (MachineBasicBlock &MBB : MF) { |
372 | MBB.removeLiveIn(Reg); |
373 | MBB.sortUniqueLiveIns(); |
374 | } |
375 | |
376 | Reg = NewReg; |
377 | } |
378 | } |
379 | |
380 | bool SIMachineFunctionInfo::allocateVirtualVGPRForSGPRSpills( |
381 | MachineFunction &MF, int FI, unsigned LaneIndex) { |
382 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
383 | Register LaneVGPR; |
384 | if (!LaneIndex) { |
385 | LaneVGPR = MRI.createVirtualRegister(RegClass: &AMDGPU::VGPR_32RegClass); |
386 | SpillVGPRs.push_back(Elt: LaneVGPR); |
387 | } else { |
388 | LaneVGPR = SpillVGPRs.back(); |
389 | } |
390 | |
391 | SGPRSpillsToVirtualVGPRLanes[FI].emplace_back(args&: LaneVGPR, args&: LaneIndex); |
392 | return true; |
393 | } |
394 | |
395 | bool SIMachineFunctionInfo::allocatePhysicalVGPRForSGPRSpills( |
396 | MachineFunction &MF, int FI, unsigned LaneIndex, bool IsPrologEpilog) { |
397 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
398 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
399 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
400 | Register LaneVGPR; |
401 | if (!LaneIndex) { |
402 | // Find the highest available register if called before RA to ensure the |
403 | // lowest registers are available for allocation. The LaneVGPR, in that |
404 | // case, will be shifted back to the lowest range after VGPR allocation. |
405 | LaneVGPR = TRI->findUnusedRegister(MRI, RC: &AMDGPU::VGPR_32RegClass, MF, |
406 | ReserveHighestVGPR: !IsPrologEpilog); |
407 | if (LaneVGPR == AMDGPU::NoRegister) { |
408 | // We have no VGPRs left for spilling SGPRs. Reset because we will not |
409 | // partially spill the SGPR to VGPRs. |
410 | SGPRSpillsToPhysicalVGPRLanes.erase(Val: FI); |
411 | return false; |
412 | } |
413 | |
414 | if (IsPrologEpilog) |
415 | allocateWWMSpill(MF, VGPR: LaneVGPR); |
416 | |
417 | reserveWWMRegister(Reg: LaneVGPR); |
418 | for (MachineBasicBlock &MBB : MF) { |
419 | MBB.addLiveIn(PhysReg: LaneVGPR); |
420 | MBB.sortUniqueLiveIns(); |
421 | } |
422 | SpillPhysVGPRs.push_back(Elt: LaneVGPR); |
423 | } else { |
424 | LaneVGPR = SpillPhysVGPRs.back(); |
425 | } |
426 | |
427 | SGPRSpillsToPhysicalVGPRLanes[FI].emplace_back(args&: LaneVGPR, args&: LaneIndex); |
428 | return true; |
429 | } |
430 | |
431 | bool SIMachineFunctionInfo::allocateSGPRSpillToVGPRLane( |
432 | MachineFunction &MF, int FI, bool SpillToPhysVGPRLane, |
433 | bool IsPrologEpilog) { |
434 | std::vector<SIRegisterInfo::SpilledReg> &SpillLanes = |
435 | SpillToPhysVGPRLane ? SGPRSpillsToPhysicalVGPRLanes[FI] |
436 | : SGPRSpillsToVirtualVGPRLanes[FI]; |
437 | |
438 | // This has already been allocated. |
439 | if (!SpillLanes.empty()) |
440 | return true; |
441 | |
442 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
443 | MachineFrameInfo &FrameInfo = MF.getFrameInfo(); |
444 | unsigned WaveSize = ST.getWavefrontSize(); |
445 | |
446 | unsigned Size = FrameInfo.getObjectSize(ObjectIdx: FI); |
447 | unsigned NumLanes = Size / 4; |
448 | |
449 | if (NumLanes > WaveSize) |
450 | return false; |
451 | |
452 | assert(Size >= 4 && "invalid sgpr spill size" ); |
453 | assert(ST.getRegisterInfo()->spillSGPRToVGPR() && |
454 | "not spilling SGPRs to VGPRs" ); |
455 | |
456 | unsigned &NumSpillLanes = SpillToPhysVGPRLane ? NumPhysicalVGPRSpillLanes |
457 | : NumVirtualVGPRSpillLanes; |
458 | |
459 | for (unsigned I = 0; I < NumLanes; ++I, ++NumSpillLanes) { |
460 | unsigned LaneIndex = (NumSpillLanes % WaveSize); |
461 | |
462 | bool Allocated = SpillToPhysVGPRLane |
463 | ? allocatePhysicalVGPRForSGPRSpills(MF, FI, LaneIndex, |
464 | IsPrologEpilog) |
465 | : allocateVirtualVGPRForSGPRSpills(MF, FI, LaneIndex); |
466 | if (!Allocated) { |
467 | NumSpillLanes -= I; |
468 | return false; |
469 | } |
470 | } |
471 | |
472 | return true; |
473 | } |
474 | |
475 | /// Reserve AGPRs or VGPRs to support spilling for FrameIndex \p FI. |
476 | /// Either AGPR is spilled to VGPR to vice versa. |
477 | /// Returns true if a \p FI can be eliminated completely. |
478 | bool SIMachineFunctionInfo::allocateVGPRSpillToAGPR(MachineFunction &MF, |
479 | int FI, |
480 | bool isAGPRtoVGPR) { |
481 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
482 | MachineFrameInfo &FrameInfo = MF.getFrameInfo(); |
483 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
484 | |
485 | assert(ST.hasMAIInsts() && FrameInfo.isSpillSlotObjectIndex(FI)); |
486 | |
487 | auto &Spill = VGPRToAGPRSpills[FI]; |
488 | |
489 | // This has already been allocated. |
490 | if (!Spill.Lanes.empty()) |
491 | return Spill.FullyAllocated; |
492 | |
493 | unsigned Size = FrameInfo.getObjectSize(ObjectIdx: FI); |
494 | unsigned NumLanes = Size / 4; |
495 | Spill.Lanes.resize(N: NumLanes, NV: AMDGPU::NoRegister); |
496 | |
497 | const TargetRegisterClass &RC = |
498 | isAGPRtoVGPR ? AMDGPU::VGPR_32RegClass : AMDGPU::AGPR_32RegClass; |
499 | auto Regs = RC.getRegisters(); |
500 | |
501 | auto &SpillRegs = isAGPRtoVGPR ? SpillAGPR : SpillVGPR; |
502 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
503 | Spill.FullyAllocated = true; |
504 | |
505 | // FIXME: Move allocation logic out of MachineFunctionInfo and initialize |
506 | // once. |
507 | BitVector OtherUsedRegs; |
508 | OtherUsedRegs.resize(N: TRI->getNumRegs()); |
509 | |
510 | const uint32_t *CSRMask = |
511 | TRI->getCallPreservedMask(MF, MF.getFunction().getCallingConv()); |
512 | if (CSRMask) |
513 | OtherUsedRegs.setBitsInMask(Mask: CSRMask); |
514 | |
515 | // TODO: Should include register tuples, but doesn't matter with current |
516 | // usage. |
517 | for (MCPhysReg Reg : SpillAGPR) |
518 | OtherUsedRegs.set(Reg); |
519 | for (MCPhysReg Reg : SpillVGPR) |
520 | OtherUsedRegs.set(Reg); |
521 | |
522 | SmallVectorImpl<MCPhysReg>::const_iterator NextSpillReg = Regs.begin(); |
523 | for (int I = NumLanes - 1; I >= 0; --I) { |
524 | NextSpillReg = std::find_if( |
525 | first: NextSpillReg, last: Regs.end(), pred: [&MRI, &OtherUsedRegs](MCPhysReg Reg) { |
526 | return MRI.isAllocatable(PhysReg: Reg) && !MRI.isPhysRegUsed(PhysReg: Reg) && |
527 | !OtherUsedRegs[Reg]; |
528 | }); |
529 | |
530 | if (NextSpillReg == Regs.end()) { // Registers exhausted |
531 | Spill.FullyAllocated = false; |
532 | break; |
533 | } |
534 | |
535 | OtherUsedRegs.set(*NextSpillReg); |
536 | SpillRegs.push_back(Elt: *NextSpillReg); |
537 | MRI.reserveReg(PhysReg: *NextSpillReg, TRI); |
538 | Spill.Lanes[I] = *NextSpillReg++; |
539 | } |
540 | |
541 | return Spill.FullyAllocated; |
542 | } |
543 | |
544 | bool SIMachineFunctionInfo::removeDeadFrameIndices( |
545 | MachineFrameInfo &MFI, bool ResetSGPRSpillStackIDs) { |
546 | // Remove dead frame indices from function frame, however keep FP & BP since |
547 | // spills for them haven't been inserted yet. And also make sure to remove the |
548 | // frame indices from `SGPRSpillsToVirtualVGPRLanes` data structure, |
549 | // otherwise, it could result in an unexpected side effect and bug, in case of |
550 | // any re-mapping of freed frame indices by later pass(es) like "stack slot |
551 | // coloring". |
552 | for (auto &R : make_early_inc_range(Range&: SGPRSpillsToVirtualVGPRLanes)) { |
553 | MFI.RemoveStackObject(ObjectIdx: R.first); |
554 | SGPRSpillsToVirtualVGPRLanes.erase(Val: R.first); |
555 | } |
556 | |
557 | // Remove the dead frame indices of CSR SGPRs which are spilled to physical |
558 | // VGPR lanes during SILowerSGPRSpills pass. |
559 | if (!ResetSGPRSpillStackIDs) { |
560 | for (auto &R : make_early_inc_range(Range&: SGPRSpillsToPhysicalVGPRLanes)) { |
561 | MFI.RemoveStackObject(ObjectIdx: R.first); |
562 | SGPRSpillsToPhysicalVGPRLanes.erase(Val: R.first); |
563 | } |
564 | } |
565 | bool HaveSGPRToMemory = false; |
566 | |
567 | if (ResetSGPRSpillStackIDs) { |
568 | // All other SGPRs must be allocated on the default stack, so reset the |
569 | // stack ID. |
570 | for (int I = MFI.getObjectIndexBegin(), E = MFI.getObjectIndexEnd(); I != E; |
571 | ++I) { |
572 | if (!checkIndexInPrologEpilogSGPRSpills(FI: I)) { |
573 | if (MFI.getStackID(ObjectIdx: I) == TargetStackID::SGPRSpill) { |
574 | MFI.setStackID(ObjectIdx: I, ID: TargetStackID::Default); |
575 | HaveSGPRToMemory = true; |
576 | } |
577 | } |
578 | } |
579 | } |
580 | |
581 | for (auto &R : VGPRToAGPRSpills) { |
582 | if (R.second.IsDead) |
583 | MFI.RemoveStackObject(ObjectIdx: R.first); |
584 | } |
585 | |
586 | return HaveSGPRToMemory; |
587 | } |
588 | |
589 | int SIMachineFunctionInfo::getScavengeFI(MachineFrameInfo &MFI, |
590 | const SIRegisterInfo &TRI) { |
591 | if (ScavengeFI) |
592 | return *ScavengeFI; |
593 | |
594 | ScavengeFI = |
595 | MFI.CreateStackObject(Size: TRI.getSpillSize(RC: AMDGPU::SGPR_32RegClass), |
596 | Alignment: TRI.getSpillAlign(RC: AMDGPU::SGPR_32RegClass), isSpillSlot: false); |
597 | return *ScavengeFI; |
598 | } |
599 | |
600 | MCPhysReg SIMachineFunctionInfo::getNextUserSGPR() const { |
601 | assert(NumSystemSGPRs == 0 && "System SGPRs must be added after user SGPRs" ); |
602 | return AMDGPU::SGPR0 + NumUserSGPRs; |
603 | } |
604 | |
605 | MCPhysReg SIMachineFunctionInfo::getNextSystemSGPR() const { |
606 | return AMDGPU::SGPR0 + NumUserSGPRs + NumSystemSGPRs; |
607 | } |
608 | |
609 | void SIMachineFunctionInfo::MRI_NoteNewVirtualRegister(Register Reg) { |
610 | VRegFlags.grow(n: Reg); |
611 | } |
612 | |
613 | void SIMachineFunctionInfo::MRI_NoteCloneVirtualRegister(Register NewReg, |
614 | Register SrcReg) { |
615 | VRegFlags.grow(n: NewReg); |
616 | VRegFlags[NewReg] = VRegFlags[SrcReg]; |
617 | } |
618 | |
619 | Register |
620 | SIMachineFunctionInfo::getGITPtrLoReg(const MachineFunction &MF) const { |
621 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
622 | if (!ST.isAmdPalOS()) |
623 | return Register(); |
624 | Register GitPtrLo = AMDGPU::SGPR0; // Low GIT address passed in |
625 | if (ST.hasMergedShaders()) { |
626 | switch (MF.getFunction().getCallingConv()) { |
627 | case CallingConv::AMDGPU_HS: |
628 | case CallingConv::AMDGPU_GS: |
629 | // Low GIT address is passed in s8 rather than s0 for an LS+HS or |
630 | // ES+GS merged shader on gfx9+. |
631 | GitPtrLo = AMDGPU::SGPR8; |
632 | return GitPtrLo; |
633 | default: |
634 | return GitPtrLo; |
635 | } |
636 | } |
637 | return GitPtrLo; |
638 | } |
639 | |
640 | static yaml::StringValue regToString(Register Reg, |
641 | const TargetRegisterInfo &TRI) { |
642 | yaml::StringValue Dest; |
643 | { |
644 | raw_string_ostream OS(Dest.Value); |
645 | OS << printReg(Reg, TRI: &TRI); |
646 | } |
647 | return Dest; |
648 | } |
649 | |
650 | static std::optional<yaml::SIArgumentInfo> |
651 | convertArgumentInfo(const AMDGPUFunctionArgInfo &ArgInfo, |
652 | const TargetRegisterInfo &TRI) { |
653 | yaml::SIArgumentInfo AI; |
654 | |
655 | auto convertArg = [&](std::optional<yaml::SIArgument> &A, |
656 | const ArgDescriptor &Arg) { |
657 | if (!Arg) |
658 | return false; |
659 | |
660 | // Create a register or stack argument. |
661 | yaml::SIArgument SA = yaml::SIArgument::createArgument(IsReg: Arg.isRegister()); |
662 | if (Arg.isRegister()) { |
663 | raw_string_ostream OS(SA.RegisterName.Value); |
664 | OS << printReg(Reg: Arg.getRegister(), TRI: &TRI); |
665 | } else |
666 | SA.StackOffset = Arg.getStackOffset(); |
667 | // Check and update the optional mask. |
668 | if (Arg.isMasked()) |
669 | SA.Mask = Arg.getMask(); |
670 | |
671 | A = SA; |
672 | return true; |
673 | }; |
674 | |
675 | // TODO: Need to serialize kernarg preloads. |
676 | bool Any = false; |
677 | Any |= convertArg(AI.PrivateSegmentBuffer, ArgInfo.PrivateSegmentBuffer); |
678 | Any |= convertArg(AI.DispatchPtr, ArgInfo.DispatchPtr); |
679 | Any |= convertArg(AI.QueuePtr, ArgInfo.QueuePtr); |
680 | Any |= convertArg(AI.KernargSegmentPtr, ArgInfo.KernargSegmentPtr); |
681 | Any |= convertArg(AI.DispatchID, ArgInfo.DispatchID); |
682 | Any |= convertArg(AI.FlatScratchInit, ArgInfo.FlatScratchInit); |
683 | Any |= convertArg(AI.LDSKernelId, ArgInfo.LDSKernelId); |
684 | Any |= convertArg(AI.PrivateSegmentSize, ArgInfo.PrivateSegmentSize); |
685 | Any |= convertArg(AI.WorkGroupIDX, ArgInfo.WorkGroupIDX); |
686 | Any |= convertArg(AI.WorkGroupIDY, ArgInfo.WorkGroupIDY); |
687 | Any |= convertArg(AI.WorkGroupIDZ, ArgInfo.WorkGroupIDZ); |
688 | Any |= convertArg(AI.WorkGroupInfo, ArgInfo.WorkGroupInfo); |
689 | Any |= convertArg(AI.PrivateSegmentWaveByteOffset, |
690 | ArgInfo.PrivateSegmentWaveByteOffset); |
691 | Any |= convertArg(AI.ImplicitArgPtr, ArgInfo.ImplicitArgPtr); |
692 | Any |= convertArg(AI.ImplicitBufferPtr, ArgInfo.ImplicitBufferPtr); |
693 | Any |= convertArg(AI.WorkItemIDX, ArgInfo.WorkItemIDX); |
694 | Any |= convertArg(AI.WorkItemIDY, ArgInfo.WorkItemIDY); |
695 | Any |= convertArg(AI.WorkItemIDZ, ArgInfo.WorkItemIDZ); |
696 | |
697 | if (Any) |
698 | return AI; |
699 | |
700 | return std::nullopt; |
701 | } |
702 | |
703 | yaml::SIMachineFunctionInfo::SIMachineFunctionInfo( |
704 | const llvm::SIMachineFunctionInfo &MFI, const TargetRegisterInfo &TRI, |
705 | const llvm::MachineFunction &MF) |
706 | : ExplicitKernArgSize(MFI.getExplicitKernArgSize()), |
707 | MaxKernArgAlign(MFI.getMaxKernArgAlign()), LDSSize(MFI.getLDSSize()), |
708 | GDSSize(MFI.getGDSSize()), DynLDSAlign(MFI.getDynLDSAlign()), |
709 | IsEntryFunction(MFI.isEntryFunction()), |
710 | NoSignedZerosFPMath(MFI.hasNoSignedZerosFPMath()), |
711 | MemoryBound(MFI.isMemoryBound()), WaveLimiter(MFI.needsWaveLimiter()), |
712 | HasSpilledSGPRs(MFI.hasSpilledSGPRs()), |
713 | HasSpilledVGPRs(MFI.hasSpilledVGPRs()), |
714 | HighBitsOf32BitAddress(MFI.get32BitAddressHighBits()), |
715 | Occupancy(MFI.getOccupancy()), |
716 | ScratchRSrcReg(regToString(Reg: MFI.getScratchRSrcReg(), TRI)), |
717 | FrameOffsetReg(regToString(Reg: MFI.getFrameOffsetReg(), TRI)), |
718 | StackPtrOffsetReg(regToString(Reg: MFI.getStackPtrOffsetReg(), TRI)), |
719 | BytesInStackArgArea(MFI.getBytesInStackArgArea()), |
720 | ReturnsVoid(MFI.returnsVoid()), |
721 | ArgInfo(convertArgumentInfo(ArgInfo: MFI.getArgInfo(), TRI)), |
722 | PSInputAddr(MFI.getPSInputAddr()), PSInputEnable(MFI.getPSInputEnable()), |
723 | MaxMemoryClusterDWords(MFI.getMaxMemoryClusterDWords()), |
724 | Mode(MFI.getMode()), HasInitWholeWave(MFI.hasInitWholeWave()), |
725 | DynamicVGPRBlockSize(MFI.getDynamicVGPRBlockSize()), |
726 | ScratchReservedForDynamicVGPRs(MFI.getScratchReservedForDynamicVGPRs()) { |
727 | for (Register Reg : MFI.getSGPRSpillPhysVGPRs()) |
728 | SpillPhysVGPRS.push_back(Elt: regToString(Reg, TRI)); |
729 | |
730 | for (Register Reg : MFI.getWWMReservedRegs()) |
731 | WWMReservedRegs.push_back(Elt: regToString(Reg, TRI)); |
732 | |
733 | if (MFI.getLongBranchReservedReg()) |
734 | LongBranchReservedReg = regToString(Reg: MFI.getLongBranchReservedReg(), TRI); |
735 | if (MFI.getVGPRForAGPRCopy()) |
736 | VGPRForAGPRCopy = regToString(Reg: MFI.getVGPRForAGPRCopy(), TRI); |
737 | |
738 | if (MFI.getSGPRForEXECCopy()) |
739 | SGPRForEXECCopy = regToString(Reg: MFI.getSGPRForEXECCopy(), TRI); |
740 | |
741 | auto SFI = MFI.getOptionalScavengeFI(); |
742 | if (SFI) |
743 | ScavengeFI = yaml::FrameIndex(*SFI, MF.getFrameInfo()); |
744 | } |
745 | |
746 | void yaml::SIMachineFunctionInfo::mappingImpl(yaml::IO &YamlIO) { |
747 | MappingTraits<SIMachineFunctionInfo>::mapping(YamlIO, MFI&: *this); |
748 | } |
749 | |
750 | bool SIMachineFunctionInfo::initializeBaseYamlFields( |
751 | const yaml::SIMachineFunctionInfo &YamlMFI, const MachineFunction &MF, |
752 | PerFunctionMIParsingState &PFS, SMDiagnostic &Error, SMRange &SourceRange) { |
753 | ExplicitKernArgSize = YamlMFI.ExplicitKernArgSize; |
754 | MaxKernArgAlign = YamlMFI.MaxKernArgAlign; |
755 | LDSSize = YamlMFI.LDSSize; |
756 | GDSSize = YamlMFI.GDSSize; |
757 | DynLDSAlign = YamlMFI.DynLDSAlign; |
758 | PSInputAddr = YamlMFI.PSInputAddr; |
759 | PSInputEnable = YamlMFI.PSInputEnable; |
760 | MaxMemoryClusterDWords = YamlMFI.MaxMemoryClusterDWords; |
761 | HighBitsOf32BitAddress = YamlMFI.HighBitsOf32BitAddress; |
762 | Occupancy = YamlMFI.Occupancy; |
763 | IsEntryFunction = YamlMFI.IsEntryFunction; |
764 | NoSignedZerosFPMath = YamlMFI.NoSignedZerosFPMath; |
765 | MemoryBound = YamlMFI.MemoryBound; |
766 | WaveLimiter = YamlMFI.WaveLimiter; |
767 | HasSpilledSGPRs = YamlMFI.HasSpilledSGPRs; |
768 | HasSpilledVGPRs = YamlMFI.HasSpilledVGPRs; |
769 | BytesInStackArgArea = YamlMFI.BytesInStackArgArea; |
770 | ReturnsVoid = YamlMFI.ReturnsVoid; |
771 | |
772 | if (YamlMFI.ScavengeFI) { |
773 | auto FIOrErr = YamlMFI.ScavengeFI->getFI(MFI: MF.getFrameInfo()); |
774 | if (!FIOrErr) { |
775 | // Create a diagnostic for a the frame index. |
776 | const MemoryBuffer &Buffer = |
777 | *PFS.SM->getMemoryBuffer(i: PFS.SM->getMainFileID()); |
778 | |
779 | Error = SMDiagnostic(*PFS.SM, SMLoc(), Buffer.getBufferIdentifier(), 1, 1, |
780 | SourceMgr::DK_Error, toString(E: FIOrErr.takeError()), |
781 | "" , {}, {}); |
782 | SourceRange = YamlMFI.ScavengeFI->SourceRange; |
783 | return true; |
784 | } |
785 | ScavengeFI = *FIOrErr; |
786 | } else { |
787 | ScavengeFI = std::nullopt; |
788 | } |
789 | return false; |
790 | } |
791 | |
792 | bool SIMachineFunctionInfo::mayUseAGPRs(const Function &F) const { |
793 | auto [MinNumAGPR, MaxNumAGPR] = |
794 | AMDGPU::getIntegerPairAttribute(F, Name: "amdgpu-agpr-alloc" , Default: {~0u, ~0u}, |
795 | /*OnlyFirstRequired=*/true); |
796 | return MinNumAGPR != 0u; |
797 | } |
798 | |