xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp (revision 02e9120893770924227138ba49df1edb3896112a)
1 //===-- AMDGPUTargetStreamer.cpp - Mips Target Streamer Methods -----------===//
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 provides AMDGPU specific target streamer methods.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "AMDGPUTargetStreamer.h"
14 #include "AMDGPUPTNote.h"
15 #include "AMDKernelCodeT.h"
16 #include "Utils/AMDGPUBaseInfo.h"
17 #include "Utils/AMDKernelCodeTUtils.h"
18 #include "llvm/BinaryFormat/AMDGPUMetadataVerifier.h"
19 #include "llvm/BinaryFormat/ELF.h"
20 #include "llvm/MC/MCAssembler.h"
21 #include "llvm/MC/MCContext.h"
22 #include "llvm/MC/MCELFStreamer.h"
23 #include "llvm/MC/MCSectionELF.h"
24 #include "llvm/MC/MCSubtargetInfo.h"
25 #include "llvm/Support/AMDGPUMetadata.h"
26 #include "llvm/Support/AMDHSAKernelDescriptor.h"
27 #include "llvm/Support/Casting.h"
28 #include "llvm/Support/FormattedStream.h"
29 #include "llvm/TargetParser/TargetParser.h"
30 
31 using namespace llvm;
32 using namespace llvm::AMDGPU;
33 
34 //===----------------------------------------------------------------------===//
35 // AMDGPUTargetStreamer
36 //===----------------------------------------------------------------------===//
37 
38 static void convertIsaVersionV2(uint32_t &Major, uint32_t &Minor,
39                                 uint32_t &Stepping, bool Sramecc, bool Xnack) {
40   if (Major == 9 && Minor == 0) {
41     switch (Stepping) {
42       case 0:
43       case 2:
44       case 4:
45       case 6:
46         if (Xnack)
47           Stepping++;
48     }
49   }
50 }
51 
52 bool AMDGPUTargetStreamer::EmitHSAMetadataV2(StringRef HSAMetadataString) {
53   HSAMD::Metadata HSAMetadata;
54   if (HSAMD::fromString(HSAMetadataString, HSAMetadata))
55     return false;
56   return EmitHSAMetadata(HSAMetadata);
57 }
58 
59 bool AMDGPUTargetStreamer::EmitHSAMetadataV3(StringRef HSAMetadataString) {
60   msgpack::Document HSAMetadataDoc;
61   if (!HSAMetadataDoc.fromYAML(HSAMetadataString))
62     return false;
63   return EmitHSAMetadata(HSAMetadataDoc, false);
64 }
65 
66 StringRef AMDGPUTargetStreamer::getArchNameFromElfMach(unsigned ElfMach) {
67   AMDGPU::GPUKind AK;
68 
69   switch (ElfMach) {
70   default: llvm_unreachable("Unhandled ELF::EF_AMDGPU type");
71   case ELF::EF_AMDGPU_MACH_R600_R600:      AK = GK_R600;    break;
72   case ELF::EF_AMDGPU_MACH_R600_R630:      AK = GK_R630;    break;
73   case ELF::EF_AMDGPU_MACH_R600_RS880:     AK = GK_RS880;   break;
74   case ELF::EF_AMDGPU_MACH_R600_RV670:     AK = GK_RV670;   break;
75   case ELF::EF_AMDGPU_MACH_R600_RV710:     AK = GK_RV710;   break;
76   case ELF::EF_AMDGPU_MACH_R600_RV730:     AK = GK_RV730;   break;
77   case ELF::EF_AMDGPU_MACH_R600_RV770:     AK = GK_RV770;   break;
78   case ELF::EF_AMDGPU_MACH_R600_CEDAR:     AK = GK_CEDAR;   break;
79   case ELF::EF_AMDGPU_MACH_R600_CYPRESS:   AK = GK_CYPRESS; break;
80   case ELF::EF_AMDGPU_MACH_R600_JUNIPER:   AK = GK_JUNIPER; break;
81   case ELF::EF_AMDGPU_MACH_R600_REDWOOD:   AK = GK_REDWOOD; break;
82   case ELF::EF_AMDGPU_MACH_R600_SUMO:      AK = GK_SUMO;    break;
83   case ELF::EF_AMDGPU_MACH_R600_BARTS:     AK = GK_BARTS;   break;
84   case ELF::EF_AMDGPU_MACH_R600_CAICOS:    AK = GK_CAICOS;  break;
85   case ELF::EF_AMDGPU_MACH_R600_CAYMAN:    AK = GK_CAYMAN;  break;
86   case ELF::EF_AMDGPU_MACH_R600_TURKS:     AK = GK_TURKS;   break;
87   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX600:  AK = GK_GFX600;  break;
88   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX601:  AK = GK_GFX601;  break;
89   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX602:  AK = GK_GFX602;  break;
90   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX700:  AK = GK_GFX700;  break;
91   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX701:  AK = GK_GFX701;  break;
92   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX702:  AK = GK_GFX702;  break;
93   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX703:  AK = GK_GFX703;  break;
94   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX704:  AK = GK_GFX704;  break;
95   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX705:  AK = GK_GFX705;  break;
96   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX801:  AK = GK_GFX801;  break;
97   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX802:  AK = GK_GFX802;  break;
98   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX803:  AK = GK_GFX803;  break;
99   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX805:  AK = GK_GFX805;  break;
100   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX810:  AK = GK_GFX810;  break;
101   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX900:  AK = GK_GFX900;  break;
102   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX902:  AK = GK_GFX902;  break;
103   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX904:  AK = GK_GFX904;  break;
104   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX906:  AK = GK_GFX906;  break;
105   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX908:  AK = GK_GFX908;  break;
106   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX909:  AK = GK_GFX909;  break;
107   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX90A:  AK = GK_GFX90A;  break;
108   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX90C:  AK = GK_GFX90C;  break;
109   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX940:  AK = GK_GFX940;  break;
110   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX941:  AK = GK_GFX941;  break;
111   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX942:  AK = GK_GFX942;  break;
112   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1010: AK = GK_GFX1010; break;
113   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1011: AK = GK_GFX1011; break;
114   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1012: AK = GK_GFX1012; break;
115   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1013: AK = GK_GFX1013; break;
116   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1030: AK = GK_GFX1030; break;
117   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1031: AK = GK_GFX1031; break;
118   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1032: AK = GK_GFX1032; break;
119   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1033: AK = GK_GFX1033; break;
120   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1034: AK = GK_GFX1034; break;
121   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1035: AK = GK_GFX1035; break;
122   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1036: AK = GK_GFX1036; break;
123   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1100: AK = GK_GFX1100; break;
124   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1101: AK = GK_GFX1101; break;
125   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1102: AK = GK_GFX1102; break;
126   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1103: AK = GK_GFX1103; break;
127   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1150: AK = GK_GFX1150; break;
128   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1151: AK = GK_GFX1151; break;
129   case ELF::EF_AMDGPU_MACH_NONE:           AK = GK_NONE;    break;
130   }
131 
132   StringRef GPUName = getArchNameAMDGCN(AK);
133   if (GPUName != "")
134     return GPUName;
135   return getArchNameR600(AK);
136 }
137 
138 unsigned AMDGPUTargetStreamer::getElfMach(StringRef GPU) {
139   AMDGPU::GPUKind AK = parseArchAMDGCN(GPU);
140   if (AK == AMDGPU::GPUKind::GK_NONE)
141     AK = parseArchR600(GPU);
142 
143   switch (AK) {
144   case GK_R600:    return ELF::EF_AMDGPU_MACH_R600_R600;
145   case GK_R630:    return ELF::EF_AMDGPU_MACH_R600_R630;
146   case GK_RS880:   return ELF::EF_AMDGPU_MACH_R600_RS880;
147   case GK_RV670:   return ELF::EF_AMDGPU_MACH_R600_RV670;
148   case GK_RV710:   return ELF::EF_AMDGPU_MACH_R600_RV710;
149   case GK_RV730:   return ELF::EF_AMDGPU_MACH_R600_RV730;
150   case GK_RV770:   return ELF::EF_AMDGPU_MACH_R600_RV770;
151   case GK_CEDAR:   return ELF::EF_AMDGPU_MACH_R600_CEDAR;
152   case GK_CYPRESS: return ELF::EF_AMDGPU_MACH_R600_CYPRESS;
153   case GK_JUNIPER: return ELF::EF_AMDGPU_MACH_R600_JUNIPER;
154   case GK_REDWOOD: return ELF::EF_AMDGPU_MACH_R600_REDWOOD;
155   case GK_SUMO:    return ELF::EF_AMDGPU_MACH_R600_SUMO;
156   case GK_BARTS:   return ELF::EF_AMDGPU_MACH_R600_BARTS;
157   case GK_CAICOS:  return ELF::EF_AMDGPU_MACH_R600_CAICOS;
158   case GK_CAYMAN:  return ELF::EF_AMDGPU_MACH_R600_CAYMAN;
159   case GK_TURKS:   return ELF::EF_AMDGPU_MACH_R600_TURKS;
160   case GK_GFX600:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX600;
161   case GK_GFX601:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX601;
162   case GK_GFX602:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX602;
163   case GK_GFX700:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX700;
164   case GK_GFX701:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX701;
165   case GK_GFX702:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX702;
166   case GK_GFX703:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX703;
167   case GK_GFX704:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX704;
168   case GK_GFX705:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX705;
169   case GK_GFX801:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX801;
170   case GK_GFX802:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX802;
171   case GK_GFX803:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX803;
172   case GK_GFX805:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX805;
173   case GK_GFX810:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX810;
174   case GK_GFX900:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX900;
175   case GK_GFX902:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX902;
176   case GK_GFX904:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX904;
177   case GK_GFX906:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX906;
178   case GK_GFX908:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX908;
179   case GK_GFX909:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX909;
180   case GK_GFX90A:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX90A;
181   case GK_GFX90C:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX90C;
182   case GK_GFX940:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX940;
183   case GK_GFX941:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX941;
184   case GK_GFX942:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX942;
185   case GK_GFX1010: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1010;
186   case GK_GFX1011: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1011;
187   case GK_GFX1012: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1012;
188   case GK_GFX1013: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1013;
189   case GK_GFX1030: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1030;
190   case GK_GFX1031: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1031;
191   case GK_GFX1032: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1032;
192   case GK_GFX1033: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1033;
193   case GK_GFX1034: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1034;
194   case GK_GFX1035: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1035;
195   case GK_GFX1036: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1036;
196   case GK_GFX1100: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1100;
197   case GK_GFX1101: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1101;
198   case GK_GFX1102: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1102;
199   case GK_GFX1103: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1103;
200   case GK_GFX1150: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1150;
201   case GK_GFX1151: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1151;
202   case GK_NONE:    return ELF::EF_AMDGPU_MACH_NONE;
203   }
204 
205   llvm_unreachable("unknown GPU");
206 }
207 
208 //===----------------------------------------------------------------------===//
209 // AMDGPUTargetAsmStreamer
210 //===----------------------------------------------------------------------===//
211 
212 AMDGPUTargetAsmStreamer::AMDGPUTargetAsmStreamer(MCStreamer &S,
213                                                  formatted_raw_ostream &OS)
214     : AMDGPUTargetStreamer(S), OS(OS) { }
215 
216 // A hook for emitting stuff at the end.
217 // We use it for emitting the accumulated PAL metadata as directives.
218 // The PAL metadata is reset after it is emitted.
219 void AMDGPUTargetAsmStreamer::finish() {
220   std::string S;
221   getPALMetadata()->toString(S);
222   OS << S;
223 
224   // Reset the pal metadata so its data will not affect a compilation that
225   // reuses this object.
226   getPALMetadata()->reset();
227 }
228 
229 void AMDGPUTargetAsmStreamer::EmitDirectiveAMDGCNTarget() {
230   OS << "\t.amdgcn_target \"" << getTargetID()->toString() << "\"\n";
231 }
232 
233 void AMDGPUTargetAsmStreamer::EmitDirectiveHSACodeObjectVersion(
234     uint32_t Major, uint32_t Minor) {
235   OS << "\t.hsa_code_object_version " <<
236         Twine(Major) << "," << Twine(Minor) << '\n';
237 }
238 
239 void
240 AMDGPUTargetAsmStreamer::EmitDirectiveHSACodeObjectISAV2(uint32_t Major,
241                                                          uint32_t Minor,
242                                                          uint32_t Stepping,
243                                                          StringRef VendorName,
244                                                          StringRef ArchName) {
245   convertIsaVersionV2(Major, Minor, Stepping, TargetID->isSramEccOnOrAny(), TargetID->isXnackOnOrAny());
246   OS << "\t.hsa_code_object_isa " << Twine(Major) << "," << Twine(Minor) << ","
247      << Twine(Stepping) << ",\"" << VendorName << "\",\"" << ArchName << "\"\n";
248 }
249 
250 void
251 AMDGPUTargetAsmStreamer::EmitAMDKernelCodeT(const amd_kernel_code_t &Header) {
252   OS << "\t.amd_kernel_code_t\n";
253   dumpAmdKernelCode(&Header, OS, "\t\t");
254   OS << "\t.end_amd_kernel_code_t\n";
255 }
256 
257 void AMDGPUTargetAsmStreamer::EmitAMDGPUSymbolType(StringRef SymbolName,
258                                                    unsigned Type) {
259   switch (Type) {
260     default: llvm_unreachable("Invalid AMDGPU symbol type");
261     case ELF::STT_AMDGPU_HSA_KERNEL:
262       OS << "\t.amdgpu_hsa_kernel " << SymbolName << '\n' ;
263       break;
264   }
265 }
266 
267 void AMDGPUTargetAsmStreamer::emitAMDGPULDS(MCSymbol *Symbol, unsigned Size,
268                                             Align Alignment) {
269   OS << "\t.amdgpu_lds " << Symbol->getName() << ", " << Size << ", "
270      << Alignment.value() << '\n';
271 }
272 
273 bool AMDGPUTargetAsmStreamer::EmitISAVersion() {
274   OS << "\t.amd_amdgpu_isa \"" << getTargetID()->toString() << "\"\n";
275   return true;
276 }
277 
278 bool AMDGPUTargetAsmStreamer::EmitHSAMetadata(
279     const AMDGPU::HSAMD::Metadata &HSAMetadata) {
280   std::string HSAMetadataString;
281   if (HSAMD::toString(HSAMetadata, HSAMetadataString))
282     return false;
283 
284   OS << '\t' << HSAMD::AssemblerDirectiveBegin << '\n';
285   OS << HSAMetadataString << '\n';
286   OS << '\t' << HSAMD::AssemblerDirectiveEnd << '\n';
287   return true;
288 }
289 
290 bool AMDGPUTargetAsmStreamer::EmitHSAMetadata(
291     msgpack::Document &HSAMetadataDoc, bool Strict) {
292   HSAMD::V3::MetadataVerifier Verifier(Strict);
293   if (!Verifier.verify(HSAMetadataDoc.getRoot()))
294     return false;
295 
296   std::string HSAMetadataString;
297   raw_string_ostream StrOS(HSAMetadataString);
298   HSAMetadataDoc.toYAML(StrOS);
299 
300   OS << '\t' << HSAMD::V3::AssemblerDirectiveBegin << '\n';
301   OS << StrOS.str() << '\n';
302   OS << '\t' << HSAMD::V3::AssemblerDirectiveEnd << '\n';
303   return true;
304 }
305 
306 bool AMDGPUTargetAsmStreamer::EmitCodeEnd(const MCSubtargetInfo &STI) {
307   const uint32_t Encoded_s_code_end = 0xbf9f0000;
308   const uint32_t Encoded_s_nop = 0xbf800000;
309   uint32_t Encoded_pad = Encoded_s_code_end;
310 
311   // Instruction cache line size in bytes.
312   const unsigned Log2CacheLineSize = AMDGPU::isGFX11Plus(STI) ? 7 : 6;
313   const unsigned CacheLineSize = 1u << Log2CacheLineSize;
314 
315   // Extra padding amount in bytes to support prefetch mode 3.
316   unsigned FillSize = 3 * CacheLineSize;
317 
318   if (AMDGPU::isGFX90A(STI)) {
319     Encoded_pad = Encoded_s_nop;
320     FillSize = 16 * CacheLineSize;
321   }
322 
323   OS << "\t.p2alignl " << Log2CacheLineSize << ", " << Encoded_pad << '\n';
324   OS << "\t.fill " << (FillSize / 4) << ", 4, " << Encoded_pad << '\n';
325   return true;
326 }
327 
328 void AMDGPUTargetAsmStreamer::EmitAmdhsaKernelDescriptor(
329     const MCSubtargetInfo &STI, StringRef KernelName,
330     const amdhsa::kernel_descriptor_t &KD, uint64_t NextVGPR, uint64_t NextSGPR,
331     bool ReserveVCC, bool ReserveFlatScr, unsigned CodeObjectVersion) {
332   IsaVersion IVersion = getIsaVersion(STI.getCPU());
333 
334   OS << "\t.amdhsa_kernel " << KernelName << '\n';
335 
336 #define PRINT_FIELD(STREAM, DIRECTIVE, KERNEL_DESC, MEMBER_NAME, FIELD_NAME)   \
337   STREAM << "\t\t" << DIRECTIVE << " "                                         \
338          << AMDHSA_BITS_GET(KERNEL_DESC.MEMBER_NAME, FIELD_NAME) << '\n';
339 
340   OS << "\t\t.amdhsa_group_segment_fixed_size " << KD.group_segment_fixed_size
341      << '\n';
342   OS << "\t\t.amdhsa_private_segment_fixed_size "
343      << KD.private_segment_fixed_size << '\n';
344   OS << "\t\t.amdhsa_kernarg_size " << KD.kernarg_size << '\n';
345 
346   PRINT_FIELD(OS, ".amdhsa_user_sgpr_count", KD,
347               compute_pgm_rsrc2,
348               amdhsa::COMPUTE_PGM_RSRC2_USER_SGPR_COUNT);
349 
350   if (!hasArchitectedFlatScratch(STI))
351     PRINT_FIELD(
352         OS, ".amdhsa_user_sgpr_private_segment_buffer", KD,
353         kernel_code_properties,
354         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER);
355   PRINT_FIELD(OS, ".amdhsa_user_sgpr_dispatch_ptr", KD,
356               kernel_code_properties,
357               amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR);
358   PRINT_FIELD(OS, ".amdhsa_user_sgpr_queue_ptr", KD,
359               kernel_code_properties,
360               amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR);
361   PRINT_FIELD(OS, ".amdhsa_user_sgpr_kernarg_segment_ptr", KD,
362               kernel_code_properties,
363               amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR);
364   PRINT_FIELD(OS, ".amdhsa_user_sgpr_dispatch_id", KD,
365               kernel_code_properties,
366               amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID);
367   if (!hasArchitectedFlatScratch(STI))
368     PRINT_FIELD(OS, ".amdhsa_user_sgpr_flat_scratch_init", KD,
369                 kernel_code_properties,
370                 amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT);
371   PRINT_FIELD(OS, ".amdhsa_user_sgpr_private_segment_size", KD,
372               kernel_code_properties,
373               amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE);
374   if (IVersion.Major >= 10)
375     PRINT_FIELD(OS, ".amdhsa_wavefront_size32", KD,
376                 kernel_code_properties,
377                 amdhsa::KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32);
378   if (CodeObjectVersion >= AMDGPU::AMDHSA_COV5)
379     PRINT_FIELD(OS, ".amdhsa_uses_dynamic_stack", KD, kernel_code_properties,
380                 amdhsa::KERNEL_CODE_PROPERTY_USES_DYNAMIC_STACK);
381   PRINT_FIELD(OS,
382               (hasArchitectedFlatScratch(STI)
383                    ? ".amdhsa_enable_private_segment"
384                    : ".amdhsa_system_sgpr_private_segment_wavefront_offset"),
385               KD, compute_pgm_rsrc2,
386               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_PRIVATE_SEGMENT);
387   PRINT_FIELD(OS, ".amdhsa_system_sgpr_workgroup_id_x", KD,
388               compute_pgm_rsrc2,
389               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_X);
390   PRINT_FIELD(OS, ".amdhsa_system_sgpr_workgroup_id_y", KD,
391               compute_pgm_rsrc2,
392               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Y);
393   PRINT_FIELD(OS, ".amdhsa_system_sgpr_workgroup_id_z", KD,
394               compute_pgm_rsrc2,
395               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Z);
396   PRINT_FIELD(OS, ".amdhsa_system_sgpr_workgroup_info", KD,
397               compute_pgm_rsrc2,
398               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_INFO);
399   PRINT_FIELD(OS, ".amdhsa_system_vgpr_workitem_id", KD,
400               compute_pgm_rsrc2,
401               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_VGPR_WORKITEM_ID);
402 
403   // These directives are required.
404   OS << "\t\t.amdhsa_next_free_vgpr " << NextVGPR << '\n';
405   OS << "\t\t.amdhsa_next_free_sgpr " << NextSGPR << '\n';
406 
407   if (AMDGPU::isGFX90A(STI))
408     OS << "\t\t.amdhsa_accum_offset " <<
409       (AMDHSA_BITS_GET(KD.compute_pgm_rsrc3,
410                        amdhsa::COMPUTE_PGM_RSRC3_GFX90A_ACCUM_OFFSET) + 1) * 4
411       << '\n';
412 
413   if (!ReserveVCC)
414     OS << "\t\t.amdhsa_reserve_vcc " << ReserveVCC << '\n';
415   if (IVersion.Major >= 7 && !ReserveFlatScr && !hasArchitectedFlatScratch(STI))
416     OS << "\t\t.amdhsa_reserve_flat_scratch " << ReserveFlatScr << '\n';
417 
418   switch (CodeObjectVersion) {
419   default:
420     break;
421   case AMDGPU::AMDHSA_COV2:
422     break;
423   case AMDGPU::AMDHSA_COV3:
424   case AMDGPU::AMDHSA_COV4:
425   case AMDGPU::AMDHSA_COV5:
426     if (getTargetID()->isXnackSupported())
427       OS << "\t\t.amdhsa_reserve_xnack_mask " << getTargetID()->isXnackOnOrAny() << '\n';
428     break;
429   }
430 
431   PRINT_FIELD(OS, ".amdhsa_float_round_mode_32", KD,
432               compute_pgm_rsrc1,
433               amdhsa::COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_32);
434   PRINT_FIELD(OS, ".amdhsa_float_round_mode_16_64", KD,
435               compute_pgm_rsrc1,
436               amdhsa::COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_16_64);
437   PRINT_FIELD(OS, ".amdhsa_float_denorm_mode_32", KD,
438               compute_pgm_rsrc1,
439               amdhsa::COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_32);
440   PRINT_FIELD(OS, ".amdhsa_float_denorm_mode_16_64", KD,
441               compute_pgm_rsrc1,
442               amdhsa::COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_16_64);
443   PRINT_FIELD(OS, ".amdhsa_dx10_clamp", KD,
444               compute_pgm_rsrc1,
445               amdhsa::COMPUTE_PGM_RSRC1_ENABLE_DX10_CLAMP);
446   PRINT_FIELD(OS, ".amdhsa_ieee_mode", KD,
447               compute_pgm_rsrc1,
448               amdhsa::COMPUTE_PGM_RSRC1_ENABLE_IEEE_MODE);
449   if (IVersion.Major >= 9)
450     PRINT_FIELD(OS, ".amdhsa_fp16_overflow", KD,
451                 compute_pgm_rsrc1,
452                 amdhsa::COMPUTE_PGM_RSRC1_FP16_OVFL);
453   if (AMDGPU::isGFX90A(STI))
454     PRINT_FIELD(OS, ".amdhsa_tg_split", KD,
455                 compute_pgm_rsrc3,
456                 amdhsa::COMPUTE_PGM_RSRC3_GFX90A_TG_SPLIT);
457   if (IVersion.Major >= 10) {
458     PRINT_FIELD(OS, ".amdhsa_workgroup_processor_mode", KD,
459                 compute_pgm_rsrc1,
460                 amdhsa::COMPUTE_PGM_RSRC1_WGP_MODE);
461     PRINT_FIELD(OS, ".amdhsa_memory_ordered", KD,
462                 compute_pgm_rsrc1,
463                 amdhsa::COMPUTE_PGM_RSRC1_MEM_ORDERED);
464     PRINT_FIELD(OS, ".amdhsa_forward_progress", KD,
465                 compute_pgm_rsrc1,
466                 amdhsa::COMPUTE_PGM_RSRC1_FWD_PROGRESS);
467     PRINT_FIELD(OS, ".amdhsa_shared_vgpr_count", KD, compute_pgm_rsrc3,
468                 amdhsa::COMPUTE_PGM_RSRC3_GFX10_PLUS_SHARED_VGPR_COUNT);
469   }
470   PRINT_FIELD(
471       OS, ".amdhsa_exception_fp_ieee_invalid_op", KD,
472       compute_pgm_rsrc2,
473       amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INVALID_OPERATION);
474   PRINT_FIELD(OS, ".amdhsa_exception_fp_denorm_src", KD,
475               compute_pgm_rsrc2,
476               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_FP_DENORMAL_SOURCE);
477   PRINT_FIELD(
478       OS, ".amdhsa_exception_fp_ieee_div_zero", KD,
479       compute_pgm_rsrc2,
480       amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_DIVISION_BY_ZERO);
481   PRINT_FIELD(OS, ".amdhsa_exception_fp_ieee_overflow", KD,
482               compute_pgm_rsrc2,
483               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_OVERFLOW);
484   PRINT_FIELD(OS, ".amdhsa_exception_fp_ieee_underflow", KD,
485               compute_pgm_rsrc2,
486               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_UNDERFLOW);
487   PRINT_FIELD(OS, ".amdhsa_exception_fp_ieee_inexact", KD,
488               compute_pgm_rsrc2,
489               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INEXACT);
490   PRINT_FIELD(OS, ".amdhsa_exception_int_div_zero", KD,
491               compute_pgm_rsrc2,
492               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_INT_DIVIDE_BY_ZERO);
493 #undef PRINT_FIELD
494 
495   OS << "\t.end_amdhsa_kernel\n";
496 }
497 
498 //===----------------------------------------------------------------------===//
499 // AMDGPUTargetELFStreamer
500 //===----------------------------------------------------------------------===//
501 
502 AMDGPUTargetELFStreamer::AMDGPUTargetELFStreamer(MCStreamer &S,
503                                                  const MCSubtargetInfo &STI)
504     : AMDGPUTargetStreamer(S), STI(STI), Streamer(S) {}
505 
506 MCELFStreamer &AMDGPUTargetELFStreamer::getStreamer() {
507   return static_cast<MCELFStreamer &>(Streamer);
508 }
509 
510 // A hook for emitting stuff at the end.
511 // We use it for emitting the accumulated PAL metadata as a .note record.
512 // The PAL metadata is reset after it is emitted.
513 void AMDGPUTargetELFStreamer::finish() {
514   MCAssembler &MCA = getStreamer().getAssembler();
515   MCA.setELFHeaderEFlags(getEFlags());
516 
517   std::string Blob;
518   const char *Vendor = getPALMetadata()->getVendor();
519   unsigned Type = getPALMetadata()->getType();
520   getPALMetadata()->toBlob(Type, Blob);
521   if (Blob.empty())
522     return;
523   EmitNote(Vendor, MCConstantExpr::create(Blob.size(), getContext()), Type,
524            [&](MCELFStreamer &OS) { OS.emitBytes(Blob); });
525 
526   // Reset the pal metadata so its data will not affect a compilation that
527   // reuses this object.
528   getPALMetadata()->reset();
529 }
530 
531 void AMDGPUTargetELFStreamer::EmitNote(
532     StringRef Name, const MCExpr *DescSZ, unsigned NoteType,
533     function_ref<void(MCELFStreamer &)> EmitDesc) {
534   auto &S = getStreamer();
535   auto &Context = S.getContext();
536 
537   auto NameSZ = Name.size() + 1;
538 
539   unsigned NoteFlags = 0;
540   // TODO Apparently, this is currently needed for OpenCL as mentioned in
541   // https://reviews.llvm.org/D74995
542   if (STI.getTargetTriple().getOS() == Triple::AMDHSA)
543     NoteFlags = ELF::SHF_ALLOC;
544 
545   S.pushSection();
546   S.switchSection(
547       Context.getELFSection(ElfNote::SectionName, ELF::SHT_NOTE, NoteFlags));
548   S.emitInt32(NameSZ);                                        // namesz
549   S.emitValue(DescSZ, 4);                                     // descz
550   S.emitInt32(NoteType);                                      // type
551   S.emitBytes(Name);                                          // name
552   S.emitValueToAlignment(Align(4), 0, 1, 0);                  // padding 0
553   EmitDesc(S);                                                // desc
554   S.emitValueToAlignment(Align(4), 0, 1, 0);                  // padding 0
555   S.popSection();
556 }
557 
558 unsigned AMDGPUTargetELFStreamer::getEFlags() {
559   switch (STI.getTargetTriple().getArch()) {
560   default:
561     llvm_unreachable("Unsupported Arch");
562   case Triple::r600:
563     return getEFlagsR600();
564   case Triple::amdgcn:
565     return getEFlagsAMDGCN();
566   }
567 }
568 
569 unsigned AMDGPUTargetELFStreamer::getEFlagsR600() {
570   assert(STI.getTargetTriple().getArch() == Triple::r600);
571 
572   return getElfMach(STI.getCPU());
573 }
574 
575 unsigned AMDGPUTargetELFStreamer::getEFlagsAMDGCN() {
576   assert(STI.getTargetTriple().getArch() == Triple::amdgcn);
577 
578   switch (STI.getTargetTriple().getOS()) {
579   default:
580     // TODO: Why are some tests have "mingw" listed as OS?
581     // llvm_unreachable("Unsupported OS");
582   case Triple::UnknownOS:
583     return getEFlagsUnknownOS();
584   case Triple::AMDHSA:
585     return getEFlagsAMDHSA();
586   case Triple::AMDPAL:
587     return getEFlagsAMDPAL();
588   case Triple::Mesa3D:
589     return getEFlagsMesa3D();
590   }
591 }
592 
593 unsigned AMDGPUTargetELFStreamer::getEFlagsUnknownOS() {
594   // TODO: Why are some tests have "mingw" listed as OS?
595   // assert(STI.getTargetTriple().getOS() == Triple::UnknownOS);
596 
597   return getEFlagsV3();
598 }
599 
600 unsigned AMDGPUTargetELFStreamer::getEFlagsAMDHSA() {
601   assert(STI.getTargetTriple().getOS() == Triple::AMDHSA);
602 
603   if (std::optional<uint8_t> HsaAbiVer = getHsaAbiVersion(&STI)) {
604     switch (*HsaAbiVer) {
605     case ELF::ELFABIVERSION_AMDGPU_HSA_V2:
606     case ELF::ELFABIVERSION_AMDGPU_HSA_V3:
607       return getEFlagsV3();
608     case ELF::ELFABIVERSION_AMDGPU_HSA_V4:
609     case ELF::ELFABIVERSION_AMDGPU_HSA_V5:
610       return getEFlagsV4();
611     }
612   }
613 
614   llvm_unreachable("HSA OS ABI Version identification must be defined");
615 }
616 
617 unsigned AMDGPUTargetELFStreamer::getEFlagsAMDPAL() {
618   assert(STI.getTargetTriple().getOS() == Triple::AMDPAL);
619 
620   return getEFlagsV3();
621 }
622 
623 unsigned AMDGPUTargetELFStreamer::getEFlagsMesa3D() {
624   assert(STI.getTargetTriple().getOS() == Triple::Mesa3D);
625 
626   return getEFlagsV3();
627 }
628 
629 unsigned AMDGPUTargetELFStreamer::getEFlagsV3() {
630   unsigned EFlagsV3 = 0;
631 
632   // mach.
633   EFlagsV3 |= getElfMach(STI.getCPU());
634 
635   // xnack.
636   if (getTargetID()->isXnackOnOrAny())
637     EFlagsV3 |= ELF::EF_AMDGPU_FEATURE_XNACK_V3;
638   // sramecc.
639   if (getTargetID()->isSramEccOnOrAny())
640     EFlagsV3 |= ELF::EF_AMDGPU_FEATURE_SRAMECC_V3;
641 
642   return EFlagsV3;
643 }
644 
645 unsigned AMDGPUTargetELFStreamer::getEFlagsV4() {
646   unsigned EFlagsV4 = 0;
647 
648   // mach.
649   EFlagsV4 |= getElfMach(STI.getCPU());
650 
651   // xnack.
652   switch (getTargetID()->getXnackSetting()) {
653   case AMDGPU::IsaInfo::TargetIDSetting::Unsupported:
654     EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_XNACK_UNSUPPORTED_V4;
655     break;
656   case AMDGPU::IsaInfo::TargetIDSetting::Any:
657     EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_XNACK_ANY_V4;
658     break;
659   case AMDGPU::IsaInfo::TargetIDSetting::Off:
660     EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_XNACK_OFF_V4;
661     break;
662   case AMDGPU::IsaInfo::TargetIDSetting::On:
663     EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_XNACK_ON_V4;
664     break;
665   }
666   // sramecc.
667   switch (getTargetID()->getSramEccSetting()) {
668   case AMDGPU::IsaInfo::TargetIDSetting::Unsupported:
669     EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_SRAMECC_UNSUPPORTED_V4;
670     break;
671   case AMDGPU::IsaInfo::TargetIDSetting::Any:
672     EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_SRAMECC_ANY_V4;
673     break;
674   case AMDGPU::IsaInfo::TargetIDSetting::Off:
675     EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_SRAMECC_OFF_V4;
676     break;
677   case AMDGPU::IsaInfo::TargetIDSetting::On:
678     EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_SRAMECC_ON_V4;
679     break;
680   }
681 
682   return EFlagsV4;
683 }
684 
685 void AMDGPUTargetELFStreamer::EmitDirectiveAMDGCNTarget() {}
686 
687 void AMDGPUTargetELFStreamer::EmitDirectiveHSACodeObjectVersion(
688     uint32_t Major, uint32_t Minor) {
689 
690   EmitNote(ElfNote::NoteNameV2, MCConstantExpr::create(8, getContext()),
691            ELF::NT_AMD_HSA_CODE_OBJECT_VERSION, [&](MCELFStreamer &OS) {
692              OS.emitInt32(Major);
693              OS.emitInt32(Minor);
694            });
695 }
696 
697 void
698 AMDGPUTargetELFStreamer::EmitDirectiveHSACodeObjectISAV2(uint32_t Major,
699                                                          uint32_t Minor,
700                                                          uint32_t Stepping,
701                                                          StringRef VendorName,
702                                                          StringRef ArchName) {
703   uint16_t VendorNameSize = VendorName.size() + 1;
704   uint16_t ArchNameSize = ArchName.size() + 1;
705 
706   unsigned DescSZ = sizeof(VendorNameSize) + sizeof(ArchNameSize) +
707     sizeof(Major) + sizeof(Minor) + sizeof(Stepping) +
708     VendorNameSize + ArchNameSize;
709 
710   convertIsaVersionV2(Major, Minor, Stepping, TargetID->isSramEccOnOrAny(), TargetID->isXnackOnOrAny());
711   EmitNote(ElfNote::NoteNameV2, MCConstantExpr::create(DescSZ, getContext()),
712            ELF::NT_AMD_HSA_ISA_VERSION, [&](MCELFStreamer &OS) {
713              OS.emitInt16(VendorNameSize);
714              OS.emitInt16(ArchNameSize);
715              OS.emitInt32(Major);
716              OS.emitInt32(Minor);
717              OS.emitInt32(Stepping);
718              OS.emitBytes(VendorName);
719              OS.emitInt8(0); // NULL terminate VendorName
720              OS.emitBytes(ArchName);
721              OS.emitInt8(0); // NULL terminate ArchName
722            });
723 }
724 
725 void
726 AMDGPUTargetELFStreamer::EmitAMDKernelCodeT(const amd_kernel_code_t &Header) {
727 
728   MCStreamer &OS = getStreamer();
729   OS.pushSection();
730   OS.emitBytes(StringRef((const char*)&Header, sizeof(Header)));
731   OS.popSection();
732 }
733 
734 void AMDGPUTargetELFStreamer::EmitAMDGPUSymbolType(StringRef SymbolName,
735                                                    unsigned Type) {
736   MCSymbolELF *Symbol = cast<MCSymbolELF>(
737       getStreamer().getContext().getOrCreateSymbol(SymbolName));
738   Symbol->setType(Type);
739 }
740 
741 void AMDGPUTargetELFStreamer::emitAMDGPULDS(MCSymbol *Symbol, unsigned Size,
742                                             Align Alignment) {
743   MCSymbolELF *SymbolELF = cast<MCSymbolELF>(Symbol);
744   SymbolELF->setType(ELF::STT_OBJECT);
745 
746   if (!SymbolELF->isBindingSet()) {
747     SymbolELF->setBinding(ELF::STB_GLOBAL);
748     SymbolELF->setExternal(true);
749   }
750 
751   if (SymbolELF->declareCommon(Size, Alignment, true)) {
752     report_fatal_error("Symbol: " + Symbol->getName() +
753                        " redeclared as different type");
754   }
755 
756   SymbolELF->setIndex(ELF::SHN_AMDGPU_LDS);
757   SymbolELF->setSize(MCConstantExpr::create(Size, getContext()));
758 }
759 
760 bool AMDGPUTargetELFStreamer::EmitISAVersion() {
761   // Create two labels to mark the beginning and end of the desc field
762   // and a MCExpr to calculate the size of the desc field.
763   auto &Context = getContext();
764   auto *DescBegin = Context.createTempSymbol();
765   auto *DescEnd = Context.createTempSymbol();
766   auto *DescSZ = MCBinaryExpr::createSub(
767     MCSymbolRefExpr::create(DescEnd, Context),
768     MCSymbolRefExpr::create(DescBegin, Context), Context);
769 
770   EmitNote(ElfNote::NoteNameV2, DescSZ, ELF::NT_AMD_HSA_ISA_NAME,
771            [&](MCELFStreamer &OS) {
772              OS.emitLabel(DescBegin);
773              OS.emitBytes(getTargetID()->toString());
774              OS.emitLabel(DescEnd);
775            });
776   return true;
777 }
778 
779 bool AMDGPUTargetELFStreamer::EmitHSAMetadata(msgpack::Document &HSAMetadataDoc,
780                                               bool Strict) {
781   HSAMD::V3::MetadataVerifier Verifier(Strict);
782   if (!Verifier.verify(HSAMetadataDoc.getRoot()))
783     return false;
784 
785   std::string HSAMetadataString;
786   HSAMetadataDoc.writeToBlob(HSAMetadataString);
787 
788   // Create two labels to mark the beginning and end of the desc field
789   // and a MCExpr to calculate the size of the desc field.
790   auto &Context = getContext();
791   auto *DescBegin = Context.createTempSymbol();
792   auto *DescEnd = Context.createTempSymbol();
793   auto *DescSZ = MCBinaryExpr::createSub(
794       MCSymbolRefExpr::create(DescEnd, Context),
795       MCSymbolRefExpr::create(DescBegin, Context), Context);
796 
797   EmitNote(ElfNote::NoteNameV3, DescSZ, ELF::NT_AMDGPU_METADATA,
798            [&](MCELFStreamer &OS) {
799              OS.emitLabel(DescBegin);
800              OS.emitBytes(HSAMetadataString);
801              OS.emitLabel(DescEnd);
802            });
803   return true;
804 }
805 
806 bool AMDGPUTargetELFStreamer::EmitHSAMetadata(
807     const AMDGPU::HSAMD::Metadata &HSAMetadata) {
808   std::string HSAMetadataString;
809   if (HSAMD::toString(HSAMetadata, HSAMetadataString))
810     return false;
811 
812   // Create two labels to mark the beginning and end of the desc field
813   // and a MCExpr to calculate the size of the desc field.
814   auto &Context = getContext();
815   auto *DescBegin = Context.createTempSymbol();
816   auto *DescEnd = Context.createTempSymbol();
817   auto *DescSZ = MCBinaryExpr::createSub(
818     MCSymbolRefExpr::create(DescEnd, Context),
819     MCSymbolRefExpr::create(DescBegin, Context), Context);
820 
821   EmitNote(ElfNote::NoteNameV2, DescSZ, ELF::NT_AMD_HSA_METADATA,
822            [&](MCELFStreamer &OS) {
823              OS.emitLabel(DescBegin);
824              OS.emitBytes(HSAMetadataString);
825              OS.emitLabel(DescEnd);
826            });
827   return true;
828 }
829 
830 bool AMDGPUTargetELFStreamer::EmitCodeEnd(const MCSubtargetInfo &STI) {
831   const uint32_t Encoded_s_code_end = 0xbf9f0000;
832   const uint32_t Encoded_s_nop = 0xbf800000;
833   uint32_t Encoded_pad = Encoded_s_code_end;
834 
835   // Instruction cache line size in bytes.
836   const unsigned Log2CacheLineSize = AMDGPU::isGFX11Plus(STI) ? 7 : 6;
837   const unsigned CacheLineSize = 1u << Log2CacheLineSize;
838 
839   // Extra padding amount in bytes to support prefetch mode 3.
840   unsigned FillSize = 3 * CacheLineSize;
841 
842   if (AMDGPU::isGFX90A(STI)) {
843     Encoded_pad = Encoded_s_nop;
844     FillSize = 16 * CacheLineSize;
845   }
846 
847   MCStreamer &OS = getStreamer();
848   OS.pushSection();
849   OS.emitValueToAlignment(Align(CacheLineSize), Encoded_pad, 4);
850   for (unsigned I = 0; I < FillSize; I += 4)
851     OS.emitInt32(Encoded_pad);
852   OS.popSection();
853   return true;
854 }
855 
856 void AMDGPUTargetELFStreamer::EmitAmdhsaKernelDescriptor(
857     const MCSubtargetInfo &STI, StringRef KernelName,
858     const amdhsa::kernel_descriptor_t &KernelDescriptor, uint64_t NextVGPR,
859     uint64_t NextSGPR, bool ReserveVCC, bool ReserveFlatScr,
860     unsigned CodeObjectVersion) {
861   auto &Streamer = getStreamer();
862   auto &Context = Streamer.getContext();
863 
864   MCSymbolELF *KernelCodeSymbol = cast<MCSymbolELF>(
865       Context.getOrCreateSymbol(Twine(KernelName)));
866   MCSymbolELF *KernelDescriptorSymbol = cast<MCSymbolELF>(
867       Context.getOrCreateSymbol(Twine(KernelName) + Twine(".kd")));
868 
869   // Copy kernel descriptor symbol's binding, other and visibility from the
870   // kernel code symbol.
871   KernelDescriptorSymbol->setBinding(KernelCodeSymbol->getBinding());
872   KernelDescriptorSymbol->setOther(KernelCodeSymbol->getOther());
873   KernelDescriptorSymbol->setVisibility(KernelCodeSymbol->getVisibility());
874   // Kernel descriptor symbol's type and size are fixed.
875   KernelDescriptorSymbol->setType(ELF::STT_OBJECT);
876   KernelDescriptorSymbol->setSize(
877       MCConstantExpr::create(sizeof(KernelDescriptor), Context));
878 
879   // The visibility of the kernel code symbol must be protected or less to allow
880   // static relocations from the kernel descriptor to be used.
881   if (KernelCodeSymbol->getVisibility() == ELF::STV_DEFAULT)
882     KernelCodeSymbol->setVisibility(ELF::STV_PROTECTED);
883 
884   Streamer.emitLabel(KernelDescriptorSymbol);
885   Streamer.emitInt32(KernelDescriptor.group_segment_fixed_size);
886   Streamer.emitInt32(KernelDescriptor.private_segment_fixed_size);
887   Streamer.emitInt32(KernelDescriptor.kernarg_size);
888 
889   for (uint8_t Res : KernelDescriptor.reserved0)
890     Streamer.emitInt8(Res);
891 
892   // FIXME: Remove the use of VK_AMDGPU_REL64 in the expression below. The
893   // expression being created is:
894   //   (start of kernel code) - (start of kernel descriptor)
895   // It implies R_AMDGPU_REL64, but ends up being R_AMDGPU_ABS64.
896   Streamer.emitValue(MCBinaryExpr::createSub(
897       MCSymbolRefExpr::create(
898           KernelCodeSymbol, MCSymbolRefExpr::VK_AMDGPU_REL64, Context),
899       MCSymbolRefExpr::create(
900           KernelDescriptorSymbol, MCSymbolRefExpr::VK_None, Context),
901       Context),
902       sizeof(KernelDescriptor.kernel_code_entry_byte_offset));
903   for (uint8_t Res : KernelDescriptor.reserved1)
904     Streamer.emitInt8(Res);
905   Streamer.emitInt32(KernelDescriptor.compute_pgm_rsrc3);
906   Streamer.emitInt32(KernelDescriptor.compute_pgm_rsrc1);
907   Streamer.emitInt32(KernelDescriptor.compute_pgm_rsrc2);
908   Streamer.emitInt16(KernelDescriptor.kernel_code_properties);
909   for (uint8_t Res : KernelDescriptor.reserved2)
910     Streamer.emitInt8(Res);
911 }
912