1 //===-- SIMachineFunctionInfo.cpp -------- SI Machine Function Info -------===//
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 "SIMachineFunctionInfo.h"
11 #include "AMDGPUSubtarget.h"
12 #include "SIInstrInfo.h"
13 #include "llvm/CodeGen/MachineFrameInfo.h"
14 #include "llvm/CodeGen/MachineInstrBuilder.h"
15 #include "llvm/CodeGen/MachineRegisterInfo.h"
16 #include "llvm/IR/Function.h"
17 #include "llvm/IR/LLVMContext.h"
18 
19 #define MAX_LANES 64
20 
21 using namespace llvm;
22 
23 static cl::opt<bool> EnableSpillSGPRToVGPR(
24   "amdgpu-spill-sgpr-to-vgpr",
25   cl::desc("Enable spilling VGPRs to SGPRs"),
26   cl::ReallyHidden,
27   cl::init(true));
28 
29 SIMachineFunctionInfo::SIMachineFunctionInfo(const MachineFunction &MF)
30   : AMDGPUMachineFunction(MF),
31     TIDReg(AMDGPU::NoRegister),
32     ScratchRSrcReg(AMDGPU::NoRegister),
33     ScratchWaveOffsetReg(AMDGPU::NoRegister),
34     PrivateSegmentBufferUserSGPR(AMDGPU::NoRegister),
35     DispatchPtrUserSGPR(AMDGPU::NoRegister),
36     QueuePtrUserSGPR(AMDGPU::NoRegister),
37     KernargSegmentPtrUserSGPR(AMDGPU::NoRegister),
38     DispatchIDUserSGPR(AMDGPU::NoRegister),
39     FlatScratchInitUserSGPR(AMDGPU::NoRegister),
40     PrivateSegmentSizeUserSGPR(AMDGPU::NoRegister),
41     GridWorkGroupCountXUserSGPR(AMDGPU::NoRegister),
42     GridWorkGroupCountYUserSGPR(AMDGPU::NoRegister),
43     GridWorkGroupCountZUserSGPR(AMDGPU::NoRegister),
44     WorkGroupIDXSystemSGPR(AMDGPU::NoRegister),
45     WorkGroupIDYSystemSGPR(AMDGPU::NoRegister),
46     WorkGroupIDZSystemSGPR(AMDGPU::NoRegister),
47     WorkGroupInfoSystemSGPR(AMDGPU::NoRegister),
48     PrivateSegmentWaveByteOffsetSystemSGPR(AMDGPU::NoRegister),
49     PSInputAddr(0),
50     ReturnsVoid(true),
51     MaximumWorkGroupSize(0),
52     DebuggerReservedVGPRCount(0),
53     DebuggerWorkGroupIDStackObjectIndices({{0, 0, 0}}),
54     DebuggerWorkItemIDStackObjectIndices({{0, 0, 0}}),
55     LDSWaveSpillSize(0),
56     PSInputEna(0),
57     NumUserSGPRs(0),
58     NumSystemSGPRs(0),
59     HasSpilledSGPRs(false),
60     HasSpilledVGPRs(false),
61     HasNonSpillStackObjects(false),
62     NumSpilledSGPRs(0),
63     NumSpilledVGPRs(0),
64     PrivateSegmentBuffer(false),
65     DispatchPtr(false),
66     QueuePtr(false),
67     KernargSegmentPtr(false),
68     DispatchID(false),
69     FlatScratchInit(false),
70     GridWorkgroupCountX(false),
71     GridWorkgroupCountY(false),
72     GridWorkgroupCountZ(false),
73     WorkGroupIDX(false),
74     WorkGroupIDY(false),
75     WorkGroupIDZ(false),
76     WorkGroupInfo(false),
77     PrivateSegmentWaveByteOffset(false),
78     WorkItemIDX(false),
79     WorkItemIDY(false),
80     WorkItemIDZ(false) {
81   const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
82   const Function *F = MF.getFunction();
83 
84   PSInputAddr = AMDGPU::getInitialPSInputAddr(*F);
85 
86   const MachineFrameInfo &FrameInfo = MF.getFrameInfo();
87 
88   if (!AMDGPU::isShader(F->getCallingConv())) {
89     KernargSegmentPtr = true;
90     WorkGroupIDX = true;
91     WorkItemIDX = true;
92   }
93 
94   if (F->hasFnAttribute("amdgpu-work-group-id-y") || ST.debuggerEmitPrologue())
95     WorkGroupIDY = true;
96 
97   if (F->hasFnAttribute("amdgpu-work-group-id-z") || ST.debuggerEmitPrologue())
98     WorkGroupIDZ = true;
99 
100   if (F->hasFnAttribute("amdgpu-work-item-id-y") || ST.debuggerEmitPrologue())
101     WorkItemIDY = true;
102 
103   if (F->hasFnAttribute("amdgpu-work-item-id-z") || ST.debuggerEmitPrologue())
104     WorkItemIDZ = true;
105 
106   // X, XY, and XYZ are the only supported combinations, so make sure Y is
107   // enabled if Z is.
108   if (WorkItemIDZ)
109     WorkItemIDY = true;
110 
111   bool MaySpill = ST.isVGPRSpillingEnabled(*F);
112   bool HasStackObjects = FrameInfo.hasStackObjects();
113 
114   if (HasStackObjects || MaySpill)
115     PrivateSegmentWaveByteOffset = true;
116 
117   if (ST.isAmdHsaOS()) {
118     if (HasStackObjects || MaySpill)
119       PrivateSegmentBuffer = true;
120 
121     if (F->hasFnAttribute("amdgpu-dispatch-ptr"))
122       DispatchPtr = true;
123 
124     if (F->hasFnAttribute("amdgpu-queue-ptr"))
125       QueuePtr = true;
126 
127     if (F->hasFnAttribute("amdgpu-dispatch-id"))
128       DispatchID = true;
129   }
130 
131   // We don't need to worry about accessing spills with flat instructions.
132   // TODO: On VI where we must use flat for global, we should be able to omit
133   // this if it is never used for generic access.
134   if (HasStackObjects && ST.getGeneration() >= SISubtarget::SEA_ISLANDS &&
135       ST.isAmdHsaOS())
136     FlatScratchInit = true;
137 
138   if (AMDGPU::isCompute(F->getCallingConv()))
139     MaximumWorkGroupSize = AMDGPU::getMaximumWorkGroupSize(*F);
140   else
141     MaximumWorkGroupSize = ST.getWavefrontSize();
142 
143   if (ST.debuggerReserveRegs())
144     DebuggerReservedVGPRCount = 4;
145 }
146 
147 unsigned SIMachineFunctionInfo::addPrivateSegmentBuffer(
148   const SIRegisterInfo &TRI) {
149   PrivateSegmentBufferUserSGPR = TRI.getMatchingSuperReg(
150     getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_128RegClass);
151   NumUserSGPRs += 4;
152   return PrivateSegmentBufferUserSGPR;
153 }
154 
155 unsigned SIMachineFunctionInfo::addDispatchPtr(const SIRegisterInfo &TRI) {
156   DispatchPtrUserSGPR = TRI.getMatchingSuperReg(
157     getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass);
158   NumUserSGPRs += 2;
159   return DispatchPtrUserSGPR;
160 }
161 
162 unsigned SIMachineFunctionInfo::addQueuePtr(const SIRegisterInfo &TRI) {
163   QueuePtrUserSGPR = TRI.getMatchingSuperReg(
164     getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass);
165   NumUserSGPRs += 2;
166   return QueuePtrUserSGPR;
167 }
168 
169 unsigned SIMachineFunctionInfo::addKernargSegmentPtr(const SIRegisterInfo &TRI) {
170   KernargSegmentPtrUserSGPR = TRI.getMatchingSuperReg(
171     getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass);
172   NumUserSGPRs += 2;
173   return KernargSegmentPtrUserSGPR;
174 }
175 
176 unsigned SIMachineFunctionInfo::addDispatchID(const SIRegisterInfo &TRI) {
177   DispatchIDUserSGPR = TRI.getMatchingSuperReg(
178     getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass);
179   NumUserSGPRs += 2;
180   return DispatchIDUserSGPR;
181 }
182 
183 unsigned SIMachineFunctionInfo::addFlatScratchInit(const SIRegisterInfo &TRI) {
184   FlatScratchInitUserSGPR = TRI.getMatchingSuperReg(
185     getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass);
186   NumUserSGPRs += 2;
187   return FlatScratchInitUserSGPR;
188 }
189 
190 SIMachineFunctionInfo::SpilledReg SIMachineFunctionInfo::getSpilledReg (
191                                                        MachineFunction *MF,
192                                                        unsigned FrameIndex,
193                                                        unsigned SubIdx) {
194   if (!EnableSpillSGPRToVGPR)
195     return SpilledReg();
196 
197   const SISubtarget &ST = MF->getSubtarget<SISubtarget>();
198   const SIRegisterInfo *TRI = ST.getRegisterInfo();
199 
200   MachineFrameInfo &FrameInfo = MF->getFrameInfo();
201   MachineRegisterInfo &MRI = MF->getRegInfo();
202   int64_t Offset = FrameInfo.getObjectOffset(FrameIndex);
203   Offset += SubIdx * 4;
204 
205   unsigned LaneVGPRIdx = Offset / (64 * 4);
206   unsigned Lane = (Offset / 4) % 64;
207 
208   struct SpilledReg Spill;
209   Spill.Lane = Lane;
210 
211   if (!LaneVGPRs.count(LaneVGPRIdx)) {
212     unsigned LaneVGPR = TRI->findUnusedRegister(MRI, &AMDGPU::VGPR_32RegClass,
213                                                 *MF);
214 
215     if (LaneVGPR == AMDGPU::NoRegister)
216       // We have no VGPRs left for spilling SGPRs.
217       return Spill;
218 
219     LaneVGPRs[LaneVGPRIdx] = LaneVGPR;
220 
221     // Add this register as live-in to all blocks to avoid machine verifer
222     // complaining about use of an undefined physical register.
223     for (MachineFunction::iterator BI = MF->begin(), BE = MF->end();
224          BI != BE; ++BI) {
225       BI->addLiveIn(LaneVGPR);
226     }
227   }
228 
229   Spill.VGPR = LaneVGPRs[LaneVGPRIdx];
230   return Spill;
231 }
232 
233 unsigned SIMachineFunctionInfo::getMaximumWorkGroupSize(
234                                               const MachineFunction &MF) const {
235   return MaximumWorkGroupSize;
236 }
237