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