xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp (revision 179219ea046f46927d6478d43431e8b541703539)
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 
34 bool AMDGPUTargetStreamer::EmitHSAMetadataV2(StringRef HSAMetadataString) {
35   HSAMD::Metadata HSAMetadata;
36   if (HSAMD::fromString(HSAMetadataString, HSAMetadata))
37     return false;
38   return EmitHSAMetadata(HSAMetadata);
39 }
40 
41 bool AMDGPUTargetStreamer::EmitHSAMetadataV3(StringRef HSAMetadataString) {
42   msgpack::Document HSAMetadataDoc;
43   if (!HSAMetadataDoc.fromYAML(HSAMetadataString))
44     return false;
45   return EmitHSAMetadata(HSAMetadataDoc, false);
46 }
47 
48 StringRef AMDGPUTargetStreamer::getArchNameFromElfMach(unsigned ElfMach) {
49   AMDGPU::GPUKind AK;
50 
51   switch (ElfMach) {
52   default: llvm_unreachable("Unhandled ELF::EF_AMDGPU type");
53   case ELF::EF_AMDGPU_MACH_R600_R600:      AK = GK_R600;    break;
54   case ELF::EF_AMDGPU_MACH_R600_R630:      AK = GK_R630;    break;
55   case ELF::EF_AMDGPU_MACH_R600_RS880:     AK = GK_RS880;   break;
56   case ELF::EF_AMDGPU_MACH_R600_RV670:     AK = GK_RV670;   break;
57   case ELF::EF_AMDGPU_MACH_R600_RV710:     AK = GK_RV710;   break;
58   case ELF::EF_AMDGPU_MACH_R600_RV730:     AK = GK_RV730;   break;
59   case ELF::EF_AMDGPU_MACH_R600_RV770:     AK = GK_RV770;   break;
60   case ELF::EF_AMDGPU_MACH_R600_CEDAR:     AK = GK_CEDAR;   break;
61   case ELF::EF_AMDGPU_MACH_R600_CYPRESS:   AK = GK_CYPRESS; break;
62   case ELF::EF_AMDGPU_MACH_R600_JUNIPER:   AK = GK_JUNIPER; break;
63   case ELF::EF_AMDGPU_MACH_R600_REDWOOD:   AK = GK_REDWOOD; break;
64   case ELF::EF_AMDGPU_MACH_R600_SUMO:      AK = GK_SUMO;    break;
65   case ELF::EF_AMDGPU_MACH_R600_BARTS:     AK = GK_BARTS;   break;
66   case ELF::EF_AMDGPU_MACH_R600_CAICOS:    AK = GK_CAICOS;  break;
67   case ELF::EF_AMDGPU_MACH_R600_CAYMAN:    AK = GK_CAYMAN;  break;
68   case ELF::EF_AMDGPU_MACH_R600_TURKS:     AK = GK_TURKS;   break;
69   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX600:  AK = GK_GFX600;  break;
70   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX601:  AK = GK_GFX601;  break;
71   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX602:  AK = GK_GFX602;  break;
72   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX700:  AK = GK_GFX700;  break;
73   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX701:  AK = GK_GFX701;  break;
74   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX702:  AK = GK_GFX702;  break;
75   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX703:  AK = GK_GFX703;  break;
76   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX704:  AK = GK_GFX704;  break;
77   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX705:  AK = GK_GFX705;  break;
78   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX801:  AK = GK_GFX801;  break;
79   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX802:  AK = GK_GFX802;  break;
80   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX803:  AK = GK_GFX803;  break;
81   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX805:  AK = GK_GFX805;  break;
82   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX810:  AK = GK_GFX810;  break;
83   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX900:  AK = GK_GFX900;  break;
84   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX902:  AK = GK_GFX902;  break;
85   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX904:  AK = GK_GFX904;  break;
86   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX906:  AK = GK_GFX906;  break;
87   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX908:  AK = GK_GFX908;  break;
88   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX909:  AK = GK_GFX909;  break;
89   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX90C:  AK = GK_GFX90C;  break;
90   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1010: AK = GK_GFX1010; break;
91   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1011: AK = GK_GFX1011; break;
92   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1012: AK = GK_GFX1012; break;
93   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1030: AK = GK_GFX1030; break;
94   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1031: AK = GK_GFX1031; break;
95   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1032: AK = GK_GFX1032; break;
96   case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1033: AK = GK_GFX1033; break;
97   case ELF::EF_AMDGPU_MACH_NONE:           AK = GK_NONE;    break;
98   }
99 
100   StringRef GPUName = getArchNameAMDGCN(AK);
101   if (GPUName != "")
102     return GPUName;
103   return getArchNameR600(AK);
104 }
105 
106 unsigned AMDGPUTargetStreamer::getElfMach(StringRef GPU) {
107   AMDGPU::GPUKind AK = parseArchAMDGCN(GPU);
108   if (AK == AMDGPU::GPUKind::GK_NONE)
109     AK = parseArchR600(GPU);
110 
111   switch (AK) {
112   case GK_R600:    return ELF::EF_AMDGPU_MACH_R600_R600;
113   case GK_R630:    return ELF::EF_AMDGPU_MACH_R600_R630;
114   case GK_RS880:   return ELF::EF_AMDGPU_MACH_R600_RS880;
115   case GK_RV670:   return ELF::EF_AMDGPU_MACH_R600_RV670;
116   case GK_RV710:   return ELF::EF_AMDGPU_MACH_R600_RV710;
117   case GK_RV730:   return ELF::EF_AMDGPU_MACH_R600_RV730;
118   case GK_RV770:   return ELF::EF_AMDGPU_MACH_R600_RV770;
119   case GK_CEDAR:   return ELF::EF_AMDGPU_MACH_R600_CEDAR;
120   case GK_CYPRESS: return ELF::EF_AMDGPU_MACH_R600_CYPRESS;
121   case GK_JUNIPER: return ELF::EF_AMDGPU_MACH_R600_JUNIPER;
122   case GK_REDWOOD: return ELF::EF_AMDGPU_MACH_R600_REDWOOD;
123   case GK_SUMO:    return ELF::EF_AMDGPU_MACH_R600_SUMO;
124   case GK_BARTS:   return ELF::EF_AMDGPU_MACH_R600_BARTS;
125   case GK_CAICOS:  return ELF::EF_AMDGPU_MACH_R600_CAICOS;
126   case GK_CAYMAN:  return ELF::EF_AMDGPU_MACH_R600_CAYMAN;
127   case GK_TURKS:   return ELF::EF_AMDGPU_MACH_R600_TURKS;
128   case GK_GFX600:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX600;
129   case GK_GFX601:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX601;
130   case GK_GFX602:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX602;
131   case GK_GFX700:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX700;
132   case GK_GFX701:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX701;
133   case GK_GFX702:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX702;
134   case GK_GFX703:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX703;
135   case GK_GFX704:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX704;
136   case GK_GFX705:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX705;
137   case GK_GFX801:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX801;
138   case GK_GFX802:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX802;
139   case GK_GFX803:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX803;
140   case GK_GFX805:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX805;
141   case GK_GFX810:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX810;
142   case GK_GFX900:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX900;
143   case GK_GFX902:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX902;
144   case GK_GFX904:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX904;
145   case GK_GFX906:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX906;
146   case GK_GFX908:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX908;
147   case GK_GFX909:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX909;
148   case GK_GFX90C:  return ELF::EF_AMDGPU_MACH_AMDGCN_GFX90C;
149   case GK_GFX1010: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1010;
150   case GK_GFX1011: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1011;
151   case GK_GFX1012: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1012;
152   case GK_GFX1030: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1030;
153   case GK_GFX1031: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1031;
154   case GK_GFX1032: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1032;
155   case GK_GFX1033: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1033;
156   case GK_NONE:    return ELF::EF_AMDGPU_MACH_NONE;
157   }
158 
159   llvm_unreachable("unknown GPU");
160 }
161 
162 //===----------------------------------------------------------------------===//
163 // AMDGPUTargetAsmStreamer
164 //===----------------------------------------------------------------------===//
165 
166 AMDGPUTargetAsmStreamer::AMDGPUTargetAsmStreamer(MCStreamer &S,
167                                                  formatted_raw_ostream &OS)
168     : AMDGPUTargetStreamer(S), OS(OS) { }
169 
170 // A hook for emitting stuff at the end.
171 // We use it for emitting the accumulated PAL metadata as directives.
172 // The PAL metadata is reset after it is emitted.
173 void AMDGPUTargetAsmStreamer::finish() {
174   std::string S;
175   getPALMetadata()->toString(S);
176   OS << S;
177 
178   // Reset the pal metadata so its data will not affect a compilation that
179   // reuses this object.
180   getPALMetadata()->reset();
181 }
182 
183 void AMDGPUTargetAsmStreamer::EmitDirectiveAMDGCNTarget(StringRef Target) {
184   OS << "\t.amdgcn_target \"" << Target << "\"\n";
185 }
186 
187 void AMDGPUTargetAsmStreamer::EmitDirectiveHSACodeObjectVersion(
188     uint32_t Major, uint32_t Minor) {
189   OS << "\t.hsa_code_object_version " <<
190         Twine(Major) << "," << Twine(Minor) << '\n';
191 }
192 
193 void
194 AMDGPUTargetAsmStreamer::EmitDirectiveHSACodeObjectISA(uint32_t Major,
195                                                        uint32_t Minor,
196                                                        uint32_t Stepping,
197                                                        StringRef VendorName,
198                                                        StringRef ArchName) {
199   OS << "\t.hsa_code_object_isa " <<
200         Twine(Major) << "," << Twine(Minor) << "," << Twine(Stepping) <<
201         ",\"" << VendorName << "\",\"" << ArchName << "\"\n";
202 
203 }
204 
205 void
206 AMDGPUTargetAsmStreamer::EmitAMDKernelCodeT(const amd_kernel_code_t &Header) {
207   OS << "\t.amd_kernel_code_t\n";
208   dumpAmdKernelCode(&Header, OS, "\t\t");
209   OS << "\t.end_amd_kernel_code_t\n";
210 }
211 
212 void AMDGPUTargetAsmStreamer::EmitAMDGPUSymbolType(StringRef SymbolName,
213                                                    unsigned Type) {
214   switch (Type) {
215     default: llvm_unreachable("Invalid AMDGPU symbol type");
216     case ELF::STT_AMDGPU_HSA_KERNEL:
217       OS << "\t.amdgpu_hsa_kernel " << SymbolName << '\n' ;
218       break;
219   }
220 }
221 
222 void AMDGPUTargetAsmStreamer::emitAMDGPULDS(MCSymbol *Symbol, unsigned Size,
223                                             Align Alignment) {
224   OS << "\t.amdgpu_lds " << Symbol->getName() << ", " << Size << ", "
225      << Alignment.value() << '\n';
226 }
227 
228 bool AMDGPUTargetAsmStreamer::EmitISAVersion(StringRef IsaVersionString) {
229   OS << "\t.amd_amdgpu_isa \"" << IsaVersionString << "\"\n";
230   return true;
231 }
232 
233 bool AMDGPUTargetAsmStreamer::EmitHSAMetadata(
234     const AMDGPU::HSAMD::Metadata &HSAMetadata) {
235   std::string HSAMetadataString;
236   if (HSAMD::toString(HSAMetadata, HSAMetadataString))
237     return false;
238 
239   OS << '\t' << HSAMD::AssemblerDirectiveBegin << '\n';
240   OS << HSAMetadataString << '\n';
241   OS << '\t' << HSAMD::AssemblerDirectiveEnd << '\n';
242   return true;
243 }
244 
245 bool AMDGPUTargetAsmStreamer::EmitHSAMetadata(
246     msgpack::Document &HSAMetadataDoc, bool Strict) {
247   HSAMD::V3::MetadataVerifier Verifier(Strict);
248   if (!Verifier.verify(HSAMetadataDoc.getRoot()))
249     return false;
250 
251   std::string HSAMetadataString;
252   raw_string_ostream StrOS(HSAMetadataString);
253   HSAMetadataDoc.toYAML(StrOS);
254 
255   OS << '\t' << HSAMD::V3::AssemblerDirectiveBegin << '\n';
256   OS << StrOS.str() << '\n';
257   OS << '\t' << HSAMD::V3::AssemblerDirectiveEnd << '\n';
258   return true;
259 }
260 
261 bool AMDGPUTargetAsmStreamer::EmitCodeEnd() {
262   const uint32_t Encoded_s_code_end = 0xbf9f0000;
263   OS << "\t.p2alignl 6, " << Encoded_s_code_end << '\n';
264   OS << "\t.fill 48, 4, " << Encoded_s_code_end << '\n';
265   return true;
266 }
267 
268 void AMDGPUTargetAsmStreamer::EmitAmdhsaKernelDescriptor(
269     const MCSubtargetInfo &STI, StringRef KernelName,
270     const amdhsa::kernel_descriptor_t &KD, uint64_t NextVGPR, uint64_t NextSGPR,
271     bool ReserveVCC, bool ReserveFlatScr, bool ReserveXNACK) {
272   IsaVersion IVersion = getIsaVersion(STI.getCPU());
273 
274   OS << "\t.amdhsa_kernel " << KernelName << '\n';
275 
276 #define PRINT_FIELD(STREAM, DIRECTIVE, KERNEL_DESC, MEMBER_NAME, FIELD_NAME)   \
277   STREAM << "\t\t" << DIRECTIVE << " "                                         \
278          << AMDHSA_BITS_GET(KERNEL_DESC.MEMBER_NAME, FIELD_NAME) << '\n';
279 
280   OS << "\t\t.amdhsa_group_segment_fixed_size " << KD.group_segment_fixed_size
281      << '\n';
282   OS << "\t\t.amdhsa_private_segment_fixed_size "
283      << KD.private_segment_fixed_size << '\n';
284 
285   PRINT_FIELD(OS, ".amdhsa_user_sgpr_private_segment_buffer", KD,
286               kernel_code_properties,
287               amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER);
288   PRINT_FIELD(OS, ".amdhsa_user_sgpr_dispatch_ptr", KD,
289               kernel_code_properties,
290               amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR);
291   PRINT_FIELD(OS, ".amdhsa_user_sgpr_queue_ptr", KD,
292               kernel_code_properties,
293               amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR);
294   PRINT_FIELD(OS, ".amdhsa_user_sgpr_kernarg_segment_ptr", KD,
295               kernel_code_properties,
296               amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR);
297   PRINT_FIELD(OS, ".amdhsa_user_sgpr_dispatch_id", KD,
298               kernel_code_properties,
299               amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID);
300   PRINT_FIELD(OS, ".amdhsa_user_sgpr_flat_scratch_init", KD,
301               kernel_code_properties,
302               amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT);
303   PRINT_FIELD(OS, ".amdhsa_user_sgpr_private_segment_size", KD,
304               kernel_code_properties,
305               amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE);
306   if (IVersion.Major >= 10)
307     PRINT_FIELD(OS, ".amdhsa_wavefront_size32", KD,
308                 kernel_code_properties,
309                 amdhsa::KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32);
310   PRINT_FIELD(
311       OS, ".amdhsa_system_sgpr_private_segment_wavefront_offset", KD,
312       compute_pgm_rsrc2,
313       amdhsa::COMPUTE_PGM_RSRC2_ENABLE_PRIVATE_SEGMENT);
314   PRINT_FIELD(OS, ".amdhsa_system_sgpr_workgroup_id_x", KD,
315               compute_pgm_rsrc2,
316               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_X);
317   PRINT_FIELD(OS, ".amdhsa_system_sgpr_workgroup_id_y", KD,
318               compute_pgm_rsrc2,
319               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Y);
320   PRINT_FIELD(OS, ".amdhsa_system_sgpr_workgroup_id_z", KD,
321               compute_pgm_rsrc2,
322               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Z);
323   PRINT_FIELD(OS, ".amdhsa_system_sgpr_workgroup_info", KD,
324               compute_pgm_rsrc2,
325               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_INFO);
326   PRINT_FIELD(OS, ".amdhsa_system_vgpr_workitem_id", KD,
327               compute_pgm_rsrc2,
328               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_VGPR_WORKITEM_ID);
329 
330   // These directives are required.
331   OS << "\t\t.amdhsa_next_free_vgpr " << NextVGPR << '\n';
332   OS << "\t\t.amdhsa_next_free_sgpr " << NextSGPR << '\n';
333 
334   if (!ReserveVCC)
335     OS << "\t\t.amdhsa_reserve_vcc " << ReserveVCC << '\n';
336   if (IVersion.Major >= 7 && !ReserveFlatScr)
337     OS << "\t\t.amdhsa_reserve_flat_scratch " << ReserveFlatScr << '\n';
338   if (IVersion.Major >= 8 && ReserveXNACK != hasXNACK(STI))
339     OS << "\t\t.amdhsa_reserve_xnack_mask " << ReserveXNACK << '\n';
340 
341   PRINT_FIELD(OS, ".amdhsa_float_round_mode_32", KD,
342               compute_pgm_rsrc1,
343               amdhsa::COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_32);
344   PRINT_FIELD(OS, ".amdhsa_float_round_mode_16_64", KD,
345               compute_pgm_rsrc1,
346               amdhsa::COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_16_64);
347   PRINT_FIELD(OS, ".amdhsa_float_denorm_mode_32", KD,
348               compute_pgm_rsrc1,
349               amdhsa::COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_32);
350   PRINT_FIELD(OS, ".amdhsa_float_denorm_mode_16_64", KD,
351               compute_pgm_rsrc1,
352               amdhsa::COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_16_64);
353   PRINT_FIELD(OS, ".amdhsa_dx10_clamp", KD,
354               compute_pgm_rsrc1,
355               amdhsa::COMPUTE_PGM_RSRC1_ENABLE_DX10_CLAMP);
356   PRINT_FIELD(OS, ".amdhsa_ieee_mode", KD,
357               compute_pgm_rsrc1,
358               amdhsa::COMPUTE_PGM_RSRC1_ENABLE_IEEE_MODE);
359   if (IVersion.Major >= 9)
360     PRINT_FIELD(OS, ".amdhsa_fp16_overflow", KD,
361                 compute_pgm_rsrc1,
362                 amdhsa::COMPUTE_PGM_RSRC1_FP16_OVFL);
363   if (IVersion.Major >= 10) {
364     PRINT_FIELD(OS, ".amdhsa_workgroup_processor_mode", KD,
365                 compute_pgm_rsrc1,
366                 amdhsa::COMPUTE_PGM_RSRC1_WGP_MODE);
367     PRINT_FIELD(OS, ".amdhsa_memory_ordered", KD,
368                 compute_pgm_rsrc1,
369                 amdhsa::COMPUTE_PGM_RSRC1_MEM_ORDERED);
370     PRINT_FIELD(OS, ".amdhsa_forward_progress", KD,
371                 compute_pgm_rsrc1,
372                 amdhsa::COMPUTE_PGM_RSRC1_FWD_PROGRESS);
373   }
374   PRINT_FIELD(
375       OS, ".amdhsa_exception_fp_ieee_invalid_op", KD,
376       compute_pgm_rsrc2,
377       amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INVALID_OPERATION);
378   PRINT_FIELD(OS, ".amdhsa_exception_fp_denorm_src", KD,
379               compute_pgm_rsrc2,
380               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_FP_DENORMAL_SOURCE);
381   PRINT_FIELD(
382       OS, ".amdhsa_exception_fp_ieee_div_zero", KD,
383       compute_pgm_rsrc2,
384       amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_DIVISION_BY_ZERO);
385   PRINT_FIELD(OS, ".amdhsa_exception_fp_ieee_overflow", KD,
386               compute_pgm_rsrc2,
387               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_OVERFLOW);
388   PRINT_FIELD(OS, ".amdhsa_exception_fp_ieee_underflow", KD,
389               compute_pgm_rsrc2,
390               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_UNDERFLOW);
391   PRINT_FIELD(OS, ".amdhsa_exception_fp_ieee_inexact", KD,
392               compute_pgm_rsrc2,
393               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INEXACT);
394   PRINT_FIELD(OS, ".amdhsa_exception_int_div_zero", KD,
395               compute_pgm_rsrc2,
396               amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_INT_DIVIDE_BY_ZERO);
397 #undef PRINT_FIELD
398 
399   OS << "\t.end_amdhsa_kernel\n";
400 }
401 
402 //===----------------------------------------------------------------------===//
403 // AMDGPUTargetELFStreamer
404 //===----------------------------------------------------------------------===//
405 
406 AMDGPUTargetELFStreamer::AMDGPUTargetELFStreamer(MCStreamer &S,
407                                                  const MCSubtargetInfo &STI)
408     : AMDGPUTargetStreamer(S), Streamer(S), Os(STI.getTargetTriple().getOS()) {
409   MCAssembler &MCA = getStreamer().getAssembler();
410   unsigned EFlags = MCA.getELFHeaderEFlags();
411 
412   EFlags &= ~ELF::EF_AMDGPU_MACH;
413   EFlags |= getElfMach(STI.getCPU());
414 
415   EFlags &= ~ELF::EF_AMDGPU_XNACK;
416   if (AMDGPU::hasXNACK(STI))
417     EFlags |= ELF::EF_AMDGPU_XNACK;
418 
419   EFlags &= ~ELF::EF_AMDGPU_SRAM_ECC;
420   if (AMDGPU::hasSRAMECC(STI))
421     EFlags |= ELF::EF_AMDGPU_SRAM_ECC;
422 
423   MCA.setELFHeaderEFlags(EFlags);
424 }
425 
426 MCELFStreamer &AMDGPUTargetELFStreamer::getStreamer() {
427   return static_cast<MCELFStreamer &>(Streamer);
428 }
429 
430 // A hook for emitting stuff at the end.
431 // We use it for emitting the accumulated PAL metadata as a .note record.
432 // The PAL metadata is reset after it is emitted.
433 void AMDGPUTargetELFStreamer::finish() {
434   std::string Blob;
435   const char *Vendor = getPALMetadata()->getVendor();
436   unsigned Type = getPALMetadata()->getType();
437   getPALMetadata()->toBlob(Type, Blob);
438   if (Blob.empty())
439     return;
440   EmitNote(Vendor, MCConstantExpr::create(Blob.size(), getContext()), Type,
441            [&](MCELFStreamer &OS) { OS.emitBytes(Blob); });
442 
443   // Reset the pal metadata so its data will not affect a compilation that
444   // reuses this object.
445   getPALMetadata()->reset();
446 }
447 
448 void AMDGPUTargetELFStreamer::EmitNote(
449     StringRef Name, const MCExpr *DescSZ, unsigned NoteType,
450     function_ref<void(MCELFStreamer &)> EmitDesc) {
451   auto &S = getStreamer();
452   auto &Context = S.getContext();
453 
454   auto NameSZ = Name.size() + 1;
455 
456   unsigned NoteFlags = 0;
457   // TODO Apparently, this is currently needed for OpenCL as mentioned in
458   // https://reviews.llvm.org/D74995
459   if (Os == Triple::AMDHSA)
460     NoteFlags = ELF::SHF_ALLOC;
461 
462   S.PushSection();
463   S.SwitchSection(
464       Context.getELFSection(ElfNote::SectionName, ELF::SHT_NOTE, NoteFlags));
465   S.emitInt32(NameSZ);                                        // namesz
466   S.emitValue(DescSZ, 4);                                     // descz
467   S.emitInt32(NoteType);                                      // type
468   S.emitBytes(Name);                                          // name
469   S.emitValueToAlignment(4, 0, 1, 0);                         // padding 0
470   EmitDesc(S);                                                // desc
471   S.emitValueToAlignment(4, 0, 1, 0);                         // padding 0
472   S.PopSection();
473 }
474 
475 void AMDGPUTargetELFStreamer::EmitDirectiveAMDGCNTarget(StringRef Target) {}
476 
477 void AMDGPUTargetELFStreamer::EmitDirectiveHSACodeObjectVersion(
478     uint32_t Major, uint32_t Minor) {
479 
480   EmitNote(ElfNote::NoteNameV2, MCConstantExpr::create(8, getContext()),
481            ElfNote::NT_AMDGPU_HSA_CODE_OBJECT_VERSION, [&](MCELFStreamer &OS) {
482              OS.emitInt32(Major);
483              OS.emitInt32(Minor);
484            });
485 }
486 
487 void
488 AMDGPUTargetELFStreamer::EmitDirectiveHSACodeObjectISA(uint32_t Major,
489                                                        uint32_t Minor,
490                                                        uint32_t Stepping,
491                                                        StringRef VendorName,
492                                                        StringRef ArchName) {
493   uint16_t VendorNameSize = VendorName.size() + 1;
494   uint16_t ArchNameSize = ArchName.size() + 1;
495 
496   unsigned DescSZ = sizeof(VendorNameSize) + sizeof(ArchNameSize) +
497     sizeof(Major) + sizeof(Minor) + sizeof(Stepping) +
498     VendorNameSize + ArchNameSize;
499 
500   EmitNote(ElfNote::NoteNameV2, MCConstantExpr::create(DescSZ, getContext()),
501            ElfNote::NT_AMDGPU_HSA_ISA, [&](MCELFStreamer &OS) {
502              OS.emitInt16(VendorNameSize);
503              OS.emitInt16(ArchNameSize);
504              OS.emitInt32(Major);
505              OS.emitInt32(Minor);
506              OS.emitInt32(Stepping);
507              OS.emitBytes(VendorName);
508              OS.emitInt8(0); // NULL terminate VendorName
509              OS.emitBytes(ArchName);
510              OS.emitInt8(0); // NULL terminte ArchName
511            });
512 }
513 
514 void
515 AMDGPUTargetELFStreamer::EmitAMDKernelCodeT(const amd_kernel_code_t &Header) {
516 
517   MCStreamer &OS = getStreamer();
518   OS.PushSection();
519   OS.emitBytes(StringRef((const char*)&Header, sizeof(Header)));
520   OS.PopSection();
521 }
522 
523 void AMDGPUTargetELFStreamer::EmitAMDGPUSymbolType(StringRef SymbolName,
524                                                    unsigned Type) {
525   MCSymbolELF *Symbol = cast<MCSymbolELF>(
526       getStreamer().getContext().getOrCreateSymbol(SymbolName));
527   Symbol->setType(Type);
528 }
529 
530 void AMDGPUTargetELFStreamer::emitAMDGPULDS(MCSymbol *Symbol, unsigned Size,
531                                             Align Alignment) {
532   MCSymbolELF *SymbolELF = cast<MCSymbolELF>(Symbol);
533   SymbolELF->setType(ELF::STT_OBJECT);
534 
535   if (!SymbolELF->isBindingSet()) {
536     SymbolELF->setBinding(ELF::STB_GLOBAL);
537     SymbolELF->setExternal(true);
538   }
539 
540   if (SymbolELF->declareCommon(Size, Alignment.value(), true)) {
541     report_fatal_error("Symbol: " + Symbol->getName() +
542                        " redeclared as different type");
543   }
544 
545   SymbolELF->setIndex(ELF::SHN_AMDGPU_LDS);
546   SymbolELF->setSize(MCConstantExpr::create(Size, getContext()));
547 }
548 
549 bool AMDGPUTargetELFStreamer::EmitISAVersion(StringRef IsaVersionString) {
550   // Create two labels to mark the beginning and end of the desc field
551   // and a MCExpr to calculate the size of the desc field.
552   auto &Context = getContext();
553   auto *DescBegin = Context.createTempSymbol();
554   auto *DescEnd = Context.createTempSymbol();
555   auto *DescSZ = MCBinaryExpr::createSub(
556     MCSymbolRefExpr::create(DescEnd, Context),
557     MCSymbolRefExpr::create(DescBegin, Context), Context);
558 
559   EmitNote(ElfNote::NoteNameV2, DescSZ, ELF::NT_AMD_AMDGPU_ISA,
560            [&](MCELFStreamer &OS) {
561              OS.emitLabel(DescBegin);
562              OS.emitBytes(IsaVersionString);
563              OS.emitLabel(DescEnd);
564            });
565   return true;
566 }
567 
568 bool AMDGPUTargetELFStreamer::EmitHSAMetadata(msgpack::Document &HSAMetadataDoc,
569                                               bool Strict) {
570   HSAMD::V3::MetadataVerifier Verifier(Strict);
571   if (!Verifier.verify(HSAMetadataDoc.getRoot()))
572     return false;
573 
574   std::string HSAMetadataString;
575   HSAMetadataDoc.writeToBlob(HSAMetadataString);
576 
577   // Create two labels to mark the beginning and end of the desc field
578   // and a MCExpr to calculate the size of the desc field.
579   auto &Context = getContext();
580   auto *DescBegin = Context.createTempSymbol();
581   auto *DescEnd = Context.createTempSymbol();
582   auto *DescSZ = MCBinaryExpr::createSub(
583       MCSymbolRefExpr::create(DescEnd, Context),
584       MCSymbolRefExpr::create(DescBegin, Context), Context);
585 
586   EmitNote(ElfNote::NoteNameV3, DescSZ, ELF::NT_AMDGPU_METADATA,
587            [&](MCELFStreamer &OS) {
588              OS.emitLabel(DescBegin);
589              OS.emitBytes(HSAMetadataString);
590              OS.emitLabel(DescEnd);
591            });
592   return true;
593 }
594 
595 bool AMDGPUTargetELFStreamer::EmitHSAMetadata(
596     const AMDGPU::HSAMD::Metadata &HSAMetadata) {
597   std::string HSAMetadataString;
598   if (HSAMD::toString(HSAMetadata, HSAMetadataString))
599     return false;
600 
601   // Create two labels to mark the beginning and end of the desc field
602   // and a MCExpr to calculate the size of the desc field.
603   auto &Context = getContext();
604   auto *DescBegin = Context.createTempSymbol();
605   auto *DescEnd = Context.createTempSymbol();
606   auto *DescSZ = MCBinaryExpr::createSub(
607     MCSymbolRefExpr::create(DescEnd, Context),
608     MCSymbolRefExpr::create(DescBegin, Context), Context);
609 
610   EmitNote(ElfNote::NoteNameV2, DescSZ, ELF::NT_AMD_AMDGPU_HSA_METADATA,
611            [&](MCELFStreamer &OS) {
612              OS.emitLabel(DescBegin);
613              OS.emitBytes(HSAMetadataString);
614              OS.emitLabel(DescEnd);
615            });
616   return true;
617 }
618 
619 bool AMDGPUTargetELFStreamer::EmitCodeEnd() {
620   const uint32_t Encoded_s_code_end = 0xbf9f0000;
621 
622   MCStreamer &OS = getStreamer();
623   OS.PushSection();
624   OS.emitValueToAlignment(64, Encoded_s_code_end, 4);
625   for (unsigned I = 0; I < 48; ++I)
626     OS.emitInt32(Encoded_s_code_end);
627   OS.PopSection();
628   return true;
629 }
630 
631 void AMDGPUTargetELFStreamer::EmitAmdhsaKernelDescriptor(
632     const MCSubtargetInfo &STI, StringRef KernelName,
633     const amdhsa::kernel_descriptor_t &KernelDescriptor, uint64_t NextVGPR,
634     uint64_t NextSGPR, bool ReserveVCC, bool ReserveFlatScr,
635     bool ReserveXNACK) {
636   auto &Streamer = getStreamer();
637   auto &Context = Streamer.getContext();
638 
639   MCSymbolELF *KernelCodeSymbol = cast<MCSymbolELF>(
640       Context.getOrCreateSymbol(Twine(KernelName)));
641   MCSymbolELF *KernelDescriptorSymbol = cast<MCSymbolELF>(
642       Context.getOrCreateSymbol(Twine(KernelName) + Twine(".kd")));
643 
644   // Copy kernel descriptor symbol's binding, other and visibility from the
645   // kernel code symbol.
646   KernelDescriptorSymbol->setBinding(KernelCodeSymbol->getBinding());
647   KernelDescriptorSymbol->setOther(KernelCodeSymbol->getOther());
648   KernelDescriptorSymbol->setVisibility(KernelCodeSymbol->getVisibility());
649   // Kernel descriptor symbol's type and size are fixed.
650   KernelDescriptorSymbol->setType(ELF::STT_OBJECT);
651   KernelDescriptorSymbol->setSize(
652       MCConstantExpr::create(sizeof(KernelDescriptor), Context));
653 
654   // The visibility of the kernel code symbol must be protected or less to allow
655   // static relocations from the kernel descriptor to be used.
656   if (KernelCodeSymbol->getVisibility() == ELF::STV_DEFAULT)
657     KernelCodeSymbol->setVisibility(ELF::STV_PROTECTED);
658 
659   Streamer.emitLabel(KernelDescriptorSymbol);
660   Streamer.emitInt32(KernelDescriptor.group_segment_fixed_size);
661   Streamer.emitInt32(KernelDescriptor.private_segment_fixed_size);
662   for (uint8_t Res : KernelDescriptor.reserved0)
663     Streamer.emitInt8(Res);
664   // FIXME: Remove the use of VK_AMDGPU_REL64 in the expression below. The
665   // expression being created is:
666   //   (start of kernel code) - (start of kernel descriptor)
667   // It implies R_AMDGPU_REL64, but ends up being R_AMDGPU_ABS64.
668   Streamer.emitValue(MCBinaryExpr::createSub(
669       MCSymbolRefExpr::create(
670           KernelCodeSymbol, MCSymbolRefExpr::VK_AMDGPU_REL64, Context),
671       MCSymbolRefExpr::create(
672           KernelDescriptorSymbol, MCSymbolRefExpr::VK_None, Context),
673       Context),
674       sizeof(KernelDescriptor.kernel_code_entry_byte_offset));
675   for (uint8_t Res : KernelDescriptor.reserved1)
676     Streamer.emitInt8(Res);
677   Streamer.emitInt32(KernelDescriptor.compute_pgm_rsrc3);
678   Streamer.emitInt32(KernelDescriptor.compute_pgm_rsrc1);
679   Streamer.emitInt32(KernelDescriptor.compute_pgm_rsrc2);
680   Streamer.emitInt16(KernelDescriptor.kernel_code_properties);
681   for (uint8_t Res : KernelDescriptor.reserved2)
682     Streamer.emitInt8(Res);
683 }
684