• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===----------------------- SIFrameLowering.cpp --------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //==-----------------------------------------------------------------------===//
9 
10 #include "SIFrameLowering.h"
11 #include "SIInstrInfo.h"
12 #include "SIMachineFunctionInfo.h"
13 #include "SIRegisterInfo.h"
14 #include "AMDGPUSubtarget.h"
15 
16 #include "llvm/CodeGen/MachineFrameInfo.h"
17 #include "llvm/CodeGen/MachineFunction.h"
18 #include "llvm/CodeGen/MachineInstrBuilder.h"
19 #include "llvm/CodeGen/RegisterScavenging.h"
20 
21 using namespace llvm;
22 
23 
hasOnlySGPRSpills(const SIMachineFunctionInfo * FuncInfo,const MachineFrameInfo * FrameInfo)24 static bool hasOnlySGPRSpills(const SIMachineFunctionInfo *FuncInfo,
25                               const MachineFrameInfo *FrameInfo) {
26   return FuncInfo->hasSpilledSGPRs() &&
27     (!FuncInfo->hasSpilledVGPRs() && !FuncInfo->hasNonSpillStackObjects());
28 }
29 
getAllSGPR128()30 static ArrayRef<MCPhysReg> getAllSGPR128() {
31   return makeArrayRef(AMDGPU::SGPR_128RegClass.begin(),
32                       AMDGPU::SGPR_128RegClass.getNumRegs());
33 }
34 
getAllSGPRs()35 static ArrayRef<MCPhysReg> getAllSGPRs() {
36   return makeArrayRef(AMDGPU::SGPR_32RegClass.begin(),
37                       AMDGPU::SGPR_32RegClass.getNumRegs());
38 }
39 
emitPrologue(MachineFunction & MF,MachineBasicBlock & MBB) const40 void SIFrameLowering::emitPrologue(MachineFunction &MF,
41                                    MachineBasicBlock &MBB) const {
42   // Emit debugger prologue if "amdgpu-debugger-emit-prologue" attribute was
43   // specified.
44   const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
45   if (ST.debuggerEmitPrologue())
46     emitDebuggerPrologue(MF, MBB);
47 
48   if (!MF.getFrameInfo()->hasStackObjects())
49     return;
50 
51   assert(&MF.front() == &MBB && "Shrink-wrapping not yet supported");
52 
53   SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
54 
55   // If we only have SGPR spills, we won't actually be using scratch memory
56   // since these spill to VGPRs.
57   //
58   // FIXME: We should be cleaning up these unused SGPR spill frame indices
59   // somewhere.
60   if (hasOnlySGPRSpills(MFI, MF.getFrameInfo()))
61     return;
62 
63   const SIInstrInfo *TII = ST.getInstrInfo();
64   const SIRegisterInfo *TRI = &TII->getRegisterInfo();
65   MachineRegisterInfo &MRI = MF.getRegInfo();
66   MachineBasicBlock::iterator I = MBB.begin();
67 
68   // We need to insert initialization of the scratch resource descriptor.
69   unsigned ScratchRsrcReg = MFI->getScratchRSrcReg();
70   assert(ScratchRsrcReg != AMDGPU::NoRegister);
71 
72   unsigned ScratchWaveOffsetReg = MFI->getScratchWaveOffsetReg();
73   assert(ScratchWaveOffsetReg != AMDGPU::NoRegister);
74 
75   unsigned PreloadedScratchWaveOffsetReg = TRI->getPreloadedValue(
76     MF, SIRegisterInfo::PRIVATE_SEGMENT_WAVE_BYTE_OFFSET);
77 
78   unsigned PreloadedPrivateBufferReg = AMDGPU::NoRegister;
79   if (ST.isAmdHsaOS()) {
80     PreloadedPrivateBufferReg = TRI->getPreloadedValue(
81       MF, SIRegisterInfo::PRIVATE_SEGMENT_BUFFER);
82   }
83 
84   if (MFI->hasFlatScratchInit()) {
85     // We don't need this if we only have spills since there is no user facing
86     // scratch.
87 
88     // TODO: If we know we don't have flat instructions earlier, we can omit
89     // this from the input registers.
90     //
91     // TODO: We only need to know if we access scratch space through a flat
92     // pointer. Because we only detect if flat instructions are used at all,
93     // this will be used more often than necessary on VI.
94 
95     // Debug location must be unknown since the first debug location is used to
96     // determine the end of the prologue.
97     DebugLoc DL;
98 
99     unsigned FlatScratchInitReg
100       = TRI->getPreloadedValue(MF, SIRegisterInfo::FLAT_SCRATCH_INIT);
101 
102     MRI.addLiveIn(FlatScratchInitReg);
103     MBB.addLiveIn(FlatScratchInitReg);
104 
105     // Copy the size in bytes.
106     unsigned FlatScrInitHi = TRI->getSubReg(FlatScratchInitReg, AMDGPU::sub1);
107     BuildMI(MBB, I, DL, TII->get(AMDGPU::S_MOV_B32), AMDGPU::FLAT_SCR_LO)
108       .addReg(FlatScrInitHi, RegState::Kill);
109 
110     unsigned FlatScrInitLo = TRI->getSubReg(FlatScratchInitReg, AMDGPU::sub0);
111 
112     // Add wave offset in bytes to private base offset.
113     // See comment in AMDKernelCodeT.h for enable_sgpr_flat_scratch_init.
114     BuildMI(MBB, I, DL, TII->get(AMDGPU::S_ADD_U32), FlatScrInitLo)
115       .addReg(FlatScrInitLo)
116       .addReg(ScratchWaveOffsetReg);
117 
118     // Convert offset to 256-byte units.
119     BuildMI(MBB, I, DL, TII->get(AMDGPU::S_LSHR_B32), AMDGPU::FLAT_SCR_HI)
120       .addReg(FlatScrInitLo, RegState::Kill)
121       .addImm(8);
122   }
123 
124   // If we reserved the original input registers, we don't need to copy to the
125   // reserved registers.
126   if (ScratchRsrcReg == PreloadedPrivateBufferReg) {
127     // We should always reserve these 5 registers at the same time.
128     assert(ScratchWaveOffsetReg == PreloadedScratchWaveOffsetReg &&
129            "scratch wave offset and private segment buffer inconsistent");
130     return;
131   }
132 
133 
134   // We added live-ins during argument lowering, but since they were not used
135   // they were deleted. We're adding the uses now, so add them back.
136   MRI.addLiveIn(PreloadedScratchWaveOffsetReg);
137   MBB.addLiveIn(PreloadedScratchWaveOffsetReg);
138 
139   if (ST.isAmdHsaOS()) {
140     MRI.addLiveIn(PreloadedPrivateBufferReg);
141     MBB.addLiveIn(PreloadedPrivateBufferReg);
142   }
143 
144   if (!ST.hasSGPRInitBug()) {
145     // We reserved the last registers for this. Shift it down to the end of those
146     // which were actually used.
147     //
148     // FIXME: It might be safer to use a pseudoregister before replacement.
149 
150     // FIXME: We should be able to eliminate unused input registers. We only
151     // cannot do this for the resources required for scratch access. For now we
152     // skip over user SGPRs and may leave unused holes.
153 
154     // We find the resource first because it has an alignment requirement.
155     if (ScratchRsrcReg == TRI->reservedPrivateSegmentBufferReg(MF)) {
156       MachineRegisterInfo &MRI = MF.getRegInfo();
157 
158       unsigned NumPreloaded = MFI->getNumPreloadedSGPRs() / 4;
159       // Skip the last 2 elements because the last one is reserved for VCC, and
160       // this is the 2nd to last element already.
161       for (MCPhysReg Reg : getAllSGPR128().drop_back(2).slice(NumPreloaded)) {
162         // Pick the first unallocated one. Make sure we don't clobber the other
163         // reserved input we needed.
164         if (!MRI.isPhysRegUsed(Reg)) {
165           assert(MRI.isAllocatable(Reg));
166           MRI.replaceRegWith(ScratchRsrcReg, Reg);
167           ScratchRsrcReg = Reg;
168           MFI->setScratchRSrcReg(ScratchRsrcReg);
169           break;
170         }
171       }
172     }
173 
174     if (ScratchWaveOffsetReg == TRI->reservedPrivateSegmentWaveByteOffsetReg(MF)) {
175       MachineRegisterInfo &MRI = MF.getRegInfo();
176       unsigned NumPreloaded = MFI->getNumPreloadedSGPRs();
177 
178       // We need to drop register from the end of the list that we cannot use
179       // for the scratch wave offset.
180       // + 2 s102 and s103 do not exist on VI.
181       // + 2 for vcc
182       // + 2 for xnack_mask
183       // + 2 for flat_scratch
184       // + 4 for registers reserved for scratch resource register
185       // + 1 for register reserved for scratch wave offset.  (By exluding this
186       //     register from the list to consider, it means that when this
187       //     register is being used for the scratch wave offset and there
188       //     are no other free SGPRs, then the value will stay in this register.
189       // ----
190       //  13
191       for (MCPhysReg Reg : getAllSGPRs().drop_back(13).slice(NumPreloaded)) {
192         // Pick the first unallocated SGPR. Be careful not to pick an alias of the
193         // scratch descriptor, since we haven’t added its uses yet.
194         if (!MRI.isPhysRegUsed(Reg)) {
195           if (!MRI.isAllocatable(Reg) ||
196               TRI->isSubRegisterEq(ScratchRsrcReg, Reg))
197             continue;
198 
199           MRI.replaceRegWith(ScratchWaveOffsetReg, Reg);
200           ScratchWaveOffsetReg = Reg;
201           MFI->setScratchWaveOffsetReg(ScratchWaveOffsetReg);
202           break;
203         }
204       }
205     }
206   }
207 
208 
209   assert(!TRI->isSubRegister(ScratchRsrcReg, ScratchWaveOffsetReg));
210 
211   const MCInstrDesc &SMovB32 = TII->get(AMDGPU::S_MOV_B32);
212   DebugLoc DL;
213 
214   if (PreloadedScratchWaveOffsetReg != ScratchWaveOffsetReg) {
215     // Make sure we emit the copy for the offset first. We may have chosen to copy
216     // the buffer resource into a register that aliases the input offset register.
217     BuildMI(MBB, I, DL, SMovB32, ScratchWaveOffsetReg)
218       .addReg(PreloadedScratchWaveOffsetReg, RegState::Kill);
219   }
220 
221   if (ST.isAmdHsaOS()) {
222     // Insert copies from argument register.
223     assert(
224       !TRI->isSubRegisterEq(PreloadedPrivateBufferReg, ScratchRsrcReg) &&
225       !TRI->isSubRegisterEq(PreloadedPrivateBufferReg, ScratchWaveOffsetReg));
226 
227     unsigned Rsrc01 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub0_sub1);
228     unsigned Rsrc23 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub2_sub3);
229 
230     unsigned Lo = TRI->getSubReg(PreloadedPrivateBufferReg, AMDGPU::sub0_sub1);
231     unsigned Hi = TRI->getSubReg(PreloadedPrivateBufferReg, AMDGPU::sub2_sub3);
232 
233     const MCInstrDesc &SMovB64 = TII->get(AMDGPU::S_MOV_B64);
234 
235     BuildMI(MBB, I, DL, SMovB64, Rsrc01)
236       .addReg(Lo, RegState::Kill);
237     BuildMI(MBB, I, DL, SMovB64, Rsrc23)
238       .addReg(Hi, RegState::Kill);
239   } else {
240     unsigned Rsrc0 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub0);
241     unsigned Rsrc1 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub1);
242     unsigned Rsrc2 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub2);
243     unsigned Rsrc3 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub3);
244 
245     // Use relocations to get the pointer, and setup the other bits manually.
246     uint64_t Rsrc23 = TII->getScratchRsrcWords23();
247     BuildMI(MBB, I, DL, SMovB32, Rsrc0)
248       .addExternalSymbol("SCRATCH_RSRC_DWORD0")
249       .addReg(ScratchRsrcReg, RegState::ImplicitDefine);
250 
251     BuildMI(MBB, I, DL, SMovB32, Rsrc1)
252       .addExternalSymbol("SCRATCH_RSRC_DWORD1")
253       .addReg(ScratchRsrcReg, RegState::ImplicitDefine);
254 
255     BuildMI(MBB, I, DL, SMovB32, Rsrc2)
256       .addImm(Rsrc23 & 0xffffffff)
257       .addReg(ScratchRsrcReg, RegState::ImplicitDefine);
258 
259     BuildMI(MBB, I, DL, SMovB32, Rsrc3)
260       .addImm(Rsrc23 >> 32)
261       .addReg(ScratchRsrcReg, RegState::ImplicitDefine);
262   }
263 
264   // Make the register selected live throughout the function.
265   for (MachineBasicBlock &OtherBB : MF) {
266     if (&OtherBB == &MBB)
267       continue;
268 
269     OtherBB.addLiveIn(ScratchRsrcReg);
270     OtherBB.addLiveIn(ScratchWaveOffsetReg);
271   }
272 }
273 
emitEpilogue(MachineFunction & MF,MachineBasicBlock & MBB) const274 void SIFrameLowering::emitEpilogue(MachineFunction &MF,
275                                    MachineBasicBlock &MBB) const {
276 
277 }
278 
processFunctionBeforeFrameFinalized(MachineFunction & MF,RegScavenger * RS) const279 void SIFrameLowering::processFunctionBeforeFrameFinalized(
280   MachineFunction &MF,
281   RegScavenger *RS) const {
282   MachineFrameInfo *MFI = MF.getFrameInfo();
283 
284   if (!MFI->hasStackObjects())
285     return;
286 
287   bool MayNeedScavengingEmergencySlot = MFI->hasStackObjects();
288 
289   assert((RS || !MayNeedScavengingEmergencySlot) &&
290          "RegScavenger required if spilling");
291 
292   if (MayNeedScavengingEmergencySlot) {
293     int ScavengeFI = MFI->CreateSpillStackObject(
294       AMDGPU::SGPR_32RegClass.getSize(),
295       AMDGPU::SGPR_32RegClass.getAlignment());
296     RS->addScavengingFrameIndex(ScavengeFI);
297   }
298 }
299 
emitDebuggerPrologue(MachineFunction & MF,MachineBasicBlock & MBB) const300 void SIFrameLowering::emitDebuggerPrologue(MachineFunction &MF,
301                                            MachineBasicBlock &MBB) const {
302   const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
303   const SIInstrInfo *TII = ST.getInstrInfo();
304   const SIRegisterInfo *TRI = &TII->getRegisterInfo();
305   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
306 
307   MachineBasicBlock::iterator I = MBB.begin();
308   DebugLoc DL;
309 
310   // For each dimension:
311   for (unsigned i = 0; i < 3; ++i) {
312     // Get work group ID SGPR, and make it live-in again.
313     unsigned WorkGroupIDSGPR = MFI->getWorkGroupIDSGPR(i);
314     MF.getRegInfo().addLiveIn(WorkGroupIDSGPR);
315     MBB.addLiveIn(WorkGroupIDSGPR);
316 
317     // Since SGPRs are spilled into VGPRs, copy work group ID SGPR to VGPR in
318     // order to spill it to scratch.
319     unsigned WorkGroupIDVGPR =
320       MF.getRegInfo().createVirtualRegister(&AMDGPU::VGPR_32RegClass);
321     BuildMI(MBB, I, DL, TII->get(AMDGPU::V_MOV_B32_e32), WorkGroupIDVGPR)
322       .addReg(WorkGroupIDSGPR);
323 
324     // Spill work group ID.
325     int WorkGroupIDObjectIdx = MFI->getDebuggerWorkGroupIDStackObjectIndex(i);
326     TII->storeRegToStackSlot(MBB, I, WorkGroupIDVGPR, false,
327       WorkGroupIDObjectIdx, &AMDGPU::VGPR_32RegClass, TRI);
328 
329     // Get work item ID VGPR, and make it live-in again.
330     unsigned WorkItemIDVGPR = MFI->getWorkItemIDVGPR(i);
331     MF.getRegInfo().addLiveIn(WorkItemIDVGPR);
332     MBB.addLiveIn(WorkItemIDVGPR);
333 
334     // Spill work item ID.
335     int WorkItemIDObjectIdx = MFI->getDebuggerWorkItemIDStackObjectIndex(i);
336     TII->storeRegToStackSlot(MBB, I, WorkItemIDVGPR, false,
337       WorkItemIDObjectIdx, &AMDGPU::VGPR_32RegClass, TRI);
338   }
339 }
340