xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUHSAMetadataStreamer.cpp (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a) 
1 //===--- AMDGPUHSAMetadataStreamer.cpp --------------------------*- C++ -*-===//
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 /// \file
10 /// AMDGPU HSA Metadata Streamer.
11 ///
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "AMDGPUHSAMetadataStreamer.h"
16 #include "AMDGPU.h"
17 #include "GCNSubtarget.h"
18 #include "MCTargetDesc/AMDGPUTargetStreamer.h"
19 #include "SIMachineFunctionInfo.h"
20 #include "SIProgramInfo.h"
21 #include "llvm/IR/Module.h"
22 #include "llvm/MC/MCContext.h"
23 #include "llvm/MC/MCExpr.h"
24 using namespace llvm;
25 
26 static std::pair<Type *, Align> getArgumentTypeAlign(const Argument &Arg,
27                                                      const DataLayout &DL) {
28   Type *Ty = Arg.getType();
29   MaybeAlign ArgAlign;
30   if (Arg.hasByRefAttr()) {
31     Ty = Arg.getParamByRefType();
32     ArgAlign = Arg.getParamAlign();
33   }
34 
35   if (!ArgAlign)
36     ArgAlign = DL.getABITypeAlign(Ty);
37 
38   return std::pair(Ty, *ArgAlign);
39 }
40 
41 namespace llvm {
42 
43 static cl::opt<bool> DumpHSAMetadata(
44     "amdgpu-dump-hsa-metadata",
45     cl::desc("Dump AMDGPU HSA Metadata"));
46 static cl::opt<bool> VerifyHSAMetadata(
47     "amdgpu-verify-hsa-metadata",
48     cl::desc("Verify AMDGPU HSA Metadata"));
49 
50 namespace AMDGPU::HSAMD {
51 
52 //===----------------------------------------------------------------------===//
53 // HSAMetadataStreamerV4
54 //===----------------------------------------------------------------------===//
55 
56 void MetadataStreamerMsgPackV4::dump(StringRef HSAMetadataString) const {
57   errs() << "AMDGPU HSA Metadata:\n" << HSAMetadataString << '\n';
58 }
59 
60 void MetadataStreamerMsgPackV4::verify(StringRef HSAMetadataString) const {
61   errs() << "AMDGPU HSA Metadata Parser Test: ";
62 
63   msgpack::Document FromHSAMetadataString;
64 
65   if (!FromHSAMetadataString.fromYAML(HSAMetadataString)) {
66     errs() << "FAIL\n";
67     return;
68   }
69 
70   std::string ToHSAMetadataString;
71   raw_string_ostream StrOS(ToHSAMetadataString);
72   FromHSAMetadataString.toYAML(StrOS);
73 
74   errs() << (HSAMetadataString == StrOS.str() ? "PASS" : "FAIL") << '\n';
75   if (HSAMetadataString != ToHSAMetadataString) {
76     errs() << "Original input: " << HSAMetadataString << '\n'
77            << "Produced output: " << StrOS.str() << '\n';
78   }
79 }
80 
81 std::optional<StringRef>
82 MetadataStreamerMsgPackV4::getAccessQualifier(StringRef AccQual) const {
83   return StringSwitch<std::optional<StringRef>>(AccQual)
84       .Case("read_only", StringRef("read_only"))
85       .Case("write_only", StringRef("write_only"))
86       .Case("read_write", StringRef("read_write"))
87       .Default(std::nullopt);
88 }
89 
90 std::optional<StringRef> MetadataStreamerMsgPackV4::getAddressSpaceQualifier(
91     unsigned AddressSpace) const {
92   switch (AddressSpace) {
93   case AMDGPUAS::PRIVATE_ADDRESS:
94     return StringRef("private");
95   case AMDGPUAS::GLOBAL_ADDRESS:
96     return StringRef("global");
97   case AMDGPUAS::CONSTANT_ADDRESS:
98     return StringRef("constant");
99   case AMDGPUAS::LOCAL_ADDRESS:
100     return StringRef("local");
101   case AMDGPUAS::FLAT_ADDRESS:
102     return StringRef("generic");
103   case AMDGPUAS::REGION_ADDRESS:
104     return StringRef("region");
105   default:
106     return std::nullopt;
107   }
108 }
109 
110 StringRef
111 MetadataStreamerMsgPackV4::getValueKind(Type *Ty, StringRef TypeQual,
112                                         StringRef BaseTypeName) const {
113   if (TypeQual.contains("pipe"))
114     return "pipe";
115 
116   return StringSwitch<StringRef>(BaseTypeName)
117       .Case("image1d_t", "image")
118       .Case("image1d_array_t", "image")
119       .Case("image1d_buffer_t", "image")
120       .Case("image2d_t", "image")
121       .Case("image2d_array_t", "image")
122       .Case("image2d_array_depth_t", "image")
123       .Case("image2d_array_msaa_t", "image")
124       .Case("image2d_array_msaa_depth_t", "image")
125       .Case("image2d_depth_t", "image")
126       .Case("image2d_msaa_t", "image")
127       .Case("image2d_msaa_depth_t", "image")
128       .Case("image3d_t", "image")
129       .Case("sampler_t", "sampler")
130       .Case("queue_t", "queue")
131       .Default(isa<PointerType>(Ty)
132                    ? (Ty->getPointerAddressSpace() == AMDGPUAS::LOCAL_ADDRESS
133                           ? "dynamic_shared_pointer"
134                           : "global_buffer")
135                    : "by_value");
136 }
137 
138 std::string MetadataStreamerMsgPackV4::getTypeName(Type *Ty,
139                                                    bool Signed) const {
140   switch (Ty->getTypeID()) {
141   case Type::IntegerTyID: {
142     if (!Signed)
143       return (Twine('u') + getTypeName(Ty, true)).str();
144 
145     auto BitWidth = Ty->getIntegerBitWidth();
146     switch (BitWidth) {
147     case 8:
148       return "char";
149     case 16:
150       return "short";
151     case 32:
152       return "int";
153     case 64:
154       return "long";
155     default:
156       return (Twine('i') + Twine(BitWidth)).str();
157     }
158   }
159   case Type::HalfTyID:
160     return "half";
161   case Type::FloatTyID:
162     return "float";
163   case Type::DoubleTyID:
164     return "double";
165   case Type::FixedVectorTyID: {
166     auto VecTy = cast<FixedVectorType>(Ty);
167     auto ElTy = VecTy->getElementType();
168     auto NumElements = VecTy->getNumElements();
169     return (Twine(getTypeName(ElTy, Signed)) + Twine(NumElements)).str();
170   }
171   default:
172     return "unknown";
173   }
174 }
175 
176 msgpack::ArrayDocNode
177 MetadataStreamerMsgPackV4::getWorkGroupDimensions(MDNode *Node) const {
178   auto Dims = HSAMetadataDoc->getArrayNode();
179   if (Node->getNumOperands() != 3)
180     return Dims;
181 
182   for (auto &Op : Node->operands())
183     Dims.push_back(Dims.getDocument()->getNode(
184         uint64_t(mdconst::extract<ConstantInt>(Op)->getZExtValue())));
185   return Dims;
186 }
187 
188 void MetadataStreamerMsgPackV4::emitVersion() {
189   auto Version = HSAMetadataDoc->getArrayNode();
190   Version.push_back(Version.getDocument()->getNode(VersionMajorV4));
191   Version.push_back(Version.getDocument()->getNode(VersionMinorV4));
192   getRootMetadata("amdhsa.version") = Version;
193 }
194 
195 void MetadataStreamerMsgPackV4::emitTargetID(
196     const IsaInfo::AMDGPUTargetID &TargetID) {
197   getRootMetadata("amdhsa.target") =
198       HSAMetadataDoc->getNode(TargetID.toString(), /*Copy=*/true);
199 }
200 
201 void MetadataStreamerMsgPackV4::emitPrintf(const Module &Mod) {
202   auto Node = Mod.getNamedMetadata("llvm.printf.fmts");
203   if (!Node)
204     return;
205 
206   auto Printf = HSAMetadataDoc->getArrayNode();
207   for (auto *Op : Node->operands())
208     if (Op->getNumOperands())
209       Printf.push_back(Printf.getDocument()->getNode(
210           cast<MDString>(Op->getOperand(0))->getString(), /*Copy=*/true));
211   getRootMetadata("amdhsa.printf") = Printf;
212 }
213 
214 void MetadataStreamerMsgPackV4::emitKernelLanguage(const Function &Func,
215                                                    msgpack::MapDocNode Kern) {
216   // TODO: What about other languages?
217   auto Node = Func.getParent()->getNamedMetadata("opencl.ocl.version");
218   if (!Node || !Node->getNumOperands())
219     return;
220   auto Op0 = Node->getOperand(0);
221   if (Op0->getNumOperands() <= 1)
222     return;
223 
224   Kern[".language"] = Kern.getDocument()->getNode("OpenCL C");
225   auto LanguageVersion = Kern.getDocument()->getArrayNode();
226   LanguageVersion.push_back(Kern.getDocument()->getNode(
227       mdconst::extract<ConstantInt>(Op0->getOperand(0))->getZExtValue()));
228   LanguageVersion.push_back(Kern.getDocument()->getNode(
229       mdconst::extract<ConstantInt>(Op0->getOperand(1))->getZExtValue()));
230   Kern[".language_version"] = LanguageVersion;
231 }
232 
233 void MetadataStreamerMsgPackV4::emitKernelAttrs(const Function &Func,
234                                                 msgpack::MapDocNode Kern) {
235 
236   if (auto Node = Func.getMetadata("reqd_work_group_size"))
237     Kern[".reqd_workgroup_size"] = getWorkGroupDimensions(Node);
238   if (auto Node = Func.getMetadata("work_group_size_hint"))
239     Kern[".workgroup_size_hint"] = getWorkGroupDimensions(Node);
240   if (auto Node = Func.getMetadata("vec_type_hint")) {
241     Kern[".vec_type_hint"] = Kern.getDocument()->getNode(
242         getTypeName(
243             cast<ValueAsMetadata>(Node->getOperand(0))->getType(),
244             mdconst::extract<ConstantInt>(Node->getOperand(1))->getZExtValue()),
245         /*Copy=*/true);
246   }
247   if (Func.hasFnAttribute("runtime-handle")) {
248     Kern[".device_enqueue_symbol"] = Kern.getDocument()->getNode(
249         Func.getFnAttribute("runtime-handle").getValueAsString().str(),
250         /*Copy=*/true);
251   }
252   if (Func.hasFnAttribute("device-init"))
253     Kern[".kind"] = Kern.getDocument()->getNode("init");
254   else if (Func.hasFnAttribute("device-fini"))
255     Kern[".kind"] = Kern.getDocument()->getNode("fini");
256 }
257 
258 void MetadataStreamerMsgPackV4::emitKernelArgs(const MachineFunction &MF,
259                                                msgpack::MapDocNode Kern) {
260   auto &Func = MF.getFunction();
261   unsigned Offset = 0;
262   auto Args = HSAMetadataDoc->getArrayNode();
263   for (auto &Arg : Func.args())
264     emitKernelArg(Arg, Offset, Args);
265 
266   emitHiddenKernelArgs(MF, Offset, Args);
267 
268   Kern[".args"] = Args;
269 }
270 
271 void MetadataStreamerMsgPackV4::emitKernelArg(const Argument &Arg,
272                                               unsigned &Offset,
273                                               msgpack::ArrayDocNode Args) {
274   auto Func = Arg.getParent();
275   auto ArgNo = Arg.getArgNo();
276   const MDNode *Node;
277 
278   StringRef Name;
279   Node = Func->getMetadata("kernel_arg_name");
280   if (Node && ArgNo < Node->getNumOperands())
281     Name = cast<MDString>(Node->getOperand(ArgNo))->getString();
282   else if (Arg.hasName())
283     Name = Arg.getName();
284 
285   StringRef TypeName;
286   Node = Func->getMetadata("kernel_arg_type");
287   if (Node && ArgNo < Node->getNumOperands())
288     TypeName = cast<MDString>(Node->getOperand(ArgNo))->getString();
289 
290   StringRef BaseTypeName;
291   Node = Func->getMetadata("kernel_arg_base_type");
292   if (Node && ArgNo < Node->getNumOperands())
293     BaseTypeName = cast<MDString>(Node->getOperand(ArgNo))->getString();
294 
295   StringRef ActAccQual;
296   // Do we really need NoAlias check here?
297   if (Arg.getType()->isPointerTy() && Arg.hasNoAliasAttr()) {
298     if (Arg.onlyReadsMemory())
299       ActAccQual = "read_only";
300     else if (Arg.hasAttribute(Attribute::WriteOnly))
301       ActAccQual = "write_only";
302   }
303 
304   StringRef AccQual;
305   Node = Func->getMetadata("kernel_arg_access_qual");
306   if (Node && ArgNo < Node->getNumOperands())
307     AccQual = cast<MDString>(Node->getOperand(ArgNo))->getString();
308 
309   StringRef TypeQual;
310   Node = Func->getMetadata("kernel_arg_type_qual");
311   if (Node && ArgNo < Node->getNumOperands())
312     TypeQual = cast<MDString>(Node->getOperand(ArgNo))->getString();
313 
314   const DataLayout &DL = Func->getDataLayout();
315 
316   MaybeAlign PointeeAlign;
317   Type *Ty = Arg.hasByRefAttr() ? Arg.getParamByRefType() : Arg.getType();
318 
319   // FIXME: Need to distinguish in memory alignment from pointer alignment.
320   if (auto PtrTy = dyn_cast<PointerType>(Ty)) {
321     if (PtrTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS)
322       PointeeAlign = Arg.getParamAlign().valueOrOne();
323   }
324 
325   // There's no distinction between byval aggregates and raw aggregates.
326   Type *ArgTy;
327   Align ArgAlign;
328   std::tie(ArgTy, ArgAlign) = getArgumentTypeAlign(Arg, DL);
329 
330   emitKernelArg(DL, ArgTy, ArgAlign,
331                 getValueKind(ArgTy, TypeQual, BaseTypeName), Offset, Args,
332                 PointeeAlign, Name, TypeName, BaseTypeName, ActAccQual,
333                 AccQual, TypeQual);
334 }
335 
336 void MetadataStreamerMsgPackV4::emitKernelArg(
337     const DataLayout &DL, Type *Ty, Align Alignment, StringRef ValueKind,
338     unsigned &Offset, msgpack::ArrayDocNode Args, MaybeAlign PointeeAlign,
339     StringRef Name, StringRef TypeName, StringRef BaseTypeName,
340     StringRef ActAccQual, StringRef AccQual, StringRef TypeQual) {
341   auto Arg = Args.getDocument()->getMapNode();
342 
343   if (!Name.empty())
344     Arg[".name"] = Arg.getDocument()->getNode(Name, /*Copy=*/true);
345   if (!TypeName.empty())
346     Arg[".type_name"] = Arg.getDocument()->getNode(TypeName, /*Copy=*/true);
347   auto Size = DL.getTypeAllocSize(Ty);
348   Arg[".size"] = Arg.getDocument()->getNode(Size);
349   Offset = alignTo(Offset, Alignment);
350   Arg[".offset"] = Arg.getDocument()->getNode(Offset);
351   Offset += Size;
352   Arg[".value_kind"] = Arg.getDocument()->getNode(ValueKind, /*Copy=*/true);
353   if (PointeeAlign)
354     Arg[".pointee_align"] = Arg.getDocument()->getNode(PointeeAlign->value());
355 
356   if (auto PtrTy = dyn_cast<PointerType>(Ty))
357     if (auto Qualifier = getAddressSpaceQualifier(PtrTy->getAddressSpace()))
358       // Limiting address space to emit only for a certain ValueKind.
359       if (ValueKind == "global_buffer" || ValueKind == "dynamic_shared_pointer")
360         Arg[".address_space"] = Arg.getDocument()->getNode(*Qualifier,
361                                                            /*Copy=*/true);
362 
363   if (auto AQ = getAccessQualifier(AccQual))
364     Arg[".access"] = Arg.getDocument()->getNode(*AQ, /*Copy=*/true);
365 
366   if (auto AAQ = getAccessQualifier(ActAccQual))
367     Arg[".actual_access"] = Arg.getDocument()->getNode(*AAQ, /*Copy=*/true);
368 
369   SmallVector<StringRef, 1> SplitTypeQuals;
370   TypeQual.split(SplitTypeQuals, " ", -1, false);
371   for (StringRef Key : SplitTypeQuals) {
372     if (Key == "const")
373       Arg[".is_const"] = Arg.getDocument()->getNode(true);
374     else if (Key == "restrict")
375       Arg[".is_restrict"] = Arg.getDocument()->getNode(true);
376     else if (Key == "volatile")
377       Arg[".is_volatile"] = Arg.getDocument()->getNode(true);
378     else if (Key == "pipe")
379       Arg[".is_pipe"] = Arg.getDocument()->getNode(true);
380   }
381 
382   Args.push_back(Arg);
383 }
384 
385 void MetadataStreamerMsgPackV4::emitHiddenKernelArgs(
386     const MachineFunction &MF, unsigned &Offset, msgpack::ArrayDocNode Args) {
387   auto &Func = MF.getFunction();
388   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
389 
390   unsigned HiddenArgNumBytes = ST.getImplicitArgNumBytes(Func);
391   if (!HiddenArgNumBytes)
392     return;
393 
394   const Module *M = Func.getParent();
395   auto &DL = M->getDataLayout();
396   auto Int64Ty = Type::getInt64Ty(Func.getContext());
397 
398   Offset = alignTo(Offset, ST.getAlignmentForImplicitArgPtr());
399 
400   if (HiddenArgNumBytes >= 8)
401     emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_x", Offset,
402                   Args);
403   if (HiddenArgNumBytes >= 16)
404     emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_y", Offset,
405                   Args);
406   if (HiddenArgNumBytes >= 24)
407     emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_z", Offset,
408                   Args);
409 
410   auto Int8PtrTy =
411       PointerType::get(Func.getContext(), AMDGPUAS::GLOBAL_ADDRESS);
412 
413   if (HiddenArgNumBytes >= 32) {
414     // We forbid the use of features requiring hostcall when compiling OpenCL
415     // before code object V5, which makes the mutual exclusion between the
416     // "printf buffer" and "hostcall buffer" here sound.
417     if (M->getNamedMetadata("llvm.printf.fmts"))
418       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_printf_buffer", Offset,
419                     Args);
420     else if (!Func.hasFnAttribute("amdgpu-no-hostcall-ptr"))
421       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_hostcall_buffer", Offset,
422                     Args);
423     else
424       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_none", Offset, Args);
425   }
426 
427   // Emit "default queue" and "completion action" arguments if enqueue kernel is
428   // used, otherwise emit dummy "none" arguments.
429   if (HiddenArgNumBytes >= 40) {
430     if (!Func.hasFnAttribute("amdgpu-no-default-queue")) {
431       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_default_queue", Offset,
432                     Args);
433     } else {
434       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_none", Offset, Args);
435     }
436   }
437 
438   if (HiddenArgNumBytes >= 48) {
439     if (!Func.hasFnAttribute("amdgpu-no-completion-action")) {
440       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_completion_action", Offset,
441                     Args);
442     } else {
443       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_none", Offset, Args);
444     }
445   }
446 
447   // Emit the pointer argument for multi-grid object.
448   if (HiddenArgNumBytes >= 56) {
449     if (!Func.hasFnAttribute("amdgpu-no-multigrid-sync-arg")) {
450       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_multigrid_sync_arg", Offset,
451                     Args);
452     } else {
453       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_none", Offset, Args);
454     }
455   }
456 }
457 
458 msgpack::MapDocNode
459 MetadataStreamerMsgPackV4::getHSAKernelProps(const MachineFunction &MF,
460                                              const SIProgramInfo &ProgramInfo,
461                                              unsigned CodeObjectVersion) const {
462   const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>();
463   const SIMachineFunctionInfo &MFI = *MF.getInfo<SIMachineFunctionInfo>();
464   const Function &F = MF.getFunction();
465 
466   auto Kern = HSAMetadataDoc->getMapNode();
467 
468   Align MaxKernArgAlign;
469   Kern[".kernarg_segment_size"] = Kern.getDocument()->getNode(
470       STM.getKernArgSegmentSize(F, MaxKernArgAlign));
471   Kern[".group_segment_fixed_size"] =
472       Kern.getDocument()->getNode(ProgramInfo.LDSSize);
473   DelayedExprs->assignDocNode(Kern[".private_segment_fixed_size"],
474                               msgpack::Type::UInt, ProgramInfo.ScratchSize);
475   if (CodeObjectVersion >= AMDGPU::AMDHSA_COV5) {
476     DelayedExprs->assignDocNode(Kern[".uses_dynamic_stack"],
477                                 msgpack::Type::Boolean,
478                                 ProgramInfo.DynamicCallStack);
479   }
480 
481   if (CodeObjectVersion >= AMDGPU::AMDHSA_COV5 && STM.supportsWGP())
482     Kern[".workgroup_processor_mode"] =
483         Kern.getDocument()->getNode(ProgramInfo.WgpMode);
484 
485   // FIXME: The metadata treats the minimum as 16?
486   Kern[".kernarg_segment_align"] =
487       Kern.getDocument()->getNode(std::max(Align(4), MaxKernArgAlign).value());
488   Kern[".wavefront_size"] =
489       Kern.getDocument()->getNode(STM.getWavefrontSize());
490   DelayedExprs->assignDocNode(Kern[".sgpr_count"], msgpack::Type::UInt,
491                               ProgramInfo.NumSGPR);
492   DelayedExprs->assignDocNode(Kern[".vgpr_count"], msgpack::Type::UInt,
493                               ProgramInfo.NumVGPR);
494 
495   // Only add AGPR count to metadata for supported devices
496   if (STM.hasMAIInsts()) {
497     DelayedExprs->assignDocNode(Kern[".agpr_count"], msgpack::Type::UInt,
498                                 ProgramInfo.NumAccVGPR);
499   }
500 
501   Kern[".max_flat_workgroup_size"] =
502       Kern.getDocument()->getNode(MFI.getMaxFlatWorkGroupSize());
503   unsigned NumWGX = MFI.getMaxNumWorkGroupsX();
504   unsigned NumWGY = MFI.getMaxNumWorkGroupsY();
505   unsigned NumWGZ = MFI.getMaxNumWorkGroupsZ();
506   if (NumWGX != 0 && NumWGY != 0 && NumWGZ != 0) {
507     Kern[".max_num_workgroups_x"] = Kern.getDocument()->getNode(NumWGX);
508     Kern[".max_num_workgroups_y"] = Kern.getDocument()->getNode(NumWGY);
509     Kern[".max_num_workgroups_z"] = Kern.getDocument()->getNode(NumWGZ);
510   }
511   Kern[".sgpr_spill_count"] =
512       Kern.getDocument()->getNode(MFI.getNumSpilledSGPRs());
513   Kern[".vgpr_spill_count"] =
514       Kern.getDocument()->getNode(MFI.getNumSpilledVGPRs());
515 
516   return Kern;
517 }
518 
519 bool MetadataStreamerMsgPackV4::emitTo(AMDGPUTargetStreamer &TargetStreamer) {
520   DelayedExprs->resolveDelayedExpressions();
521   return TargetStreamer.EmitHSAMetadata(*HSAMetadataDoc, true);
522 }
523 
524 void MetadataStreamerMsgPackV4::begin(const Module &Mod,
525                                       const IsaInfo::AMDGPUTargetID &TargetID) {
526   emitVersion();
527   emitTargetID(TargetID);
528   emitPrintf(Mod);
529   getRootMetadata("amdhsa.kernels") = HSAMetadataDoc->getArrayNode();
530   DelayedExprs->clear();
531 }
532 
533 void MetadataStreamerMsgPackV4::end() {
534   DelayedExprs->resolveDelayedExpressions();
535   std::string HSAMetadataString;
536   raw_string_ostream StrOS(HSAMetadataString);
537   HSAMetadataDoc->toYAML(StrOS);
538 
539   if (DumpHSAMetadata)
540     dump(StrOS.str());
541   if (VerifyHSAMetadata)
542     verify(StrOS.str());
543 }
544 
545 void MetadataStreamerMsgPackV4::emitKernel(const MachineFunction &MF,
546                                            const SIProgramInfo &ProgramInfo) {
547   auto &Func = MF.getFunction();
548   if (Func.getCallingConv() != CallingConv::AMDGPU_KERNEL &&
549       Func.getCallingConv() != CallingConv::SPIR_KERNEL)
550     return;
551 
552   auto CodeObjectVersion =
553       AMDGPU::getAMDHSACodeObjectVersion(*Func.getParent());
554   auto Kern = getHSAKernelProps(MF, ProgramInfo, CodeObjectVersion);
555 
556   auto Kernels =
557       getRootMetadata("amdhsa.kernels").getArray(/*Convert=*/true);
558 
559   {
560     Kern[".name"] = Kern.getDocument()->getNode(Func.getName());
561     Kern[".symbol"] = Kern.getDocument()->getNode(
562         (Twine(Func.getName()) + Twine(".kd")).str(), /*Copy=*/true);
563     emitKernelLanguage(Func, Kern);
564     emitKernelAttrs(Func, Kern);
565     emitKernelArgs(MF, Kern);
566   }
567 
568   Kernels.push_back(Kern);
569 }
570 
571 //===----------------------------------------------------------------------===//
572 // HSAMetadataStreamerV5
573 //===----------------------------------------------------------------------===//
574 
575 void MetadataStreamerMsgPackV5::emitVersion() {
576   auto Version = HSAMetadataDoc->getArrayNode();
577   Version.push_back(Version.getDocument()->getNode(VersionMajorV5));
578   Version.push_back(Version.getDocument()->getNode(VersionMinorV5));
579   getRootMetadata("amdhsa.version") = Version;
580 }
581 
582 void MetadataStreamerMsgPackV5::emitHiddenKernelArgs(
583     const MachineFunction &MF, unsigned &Offset, msgpack::ArrayDocNode Args) {
584   auto &Func = MF.getFunction();
585   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
586 
587   // No implicit kernel argument is used.
588   if (ST.getImplicitArgNumBytes(Func) == 0)
589     return;
590 
591   const Module *M = Func.getParent();
592   auto &DL = M->getDataLayout();
593   const SIMachineFunctionInfo &MFI = *MF.getInfo<SIMachineFunctionInfo>();
594 
595   auto Int64Ty = Type::getInt64Ty(Func.getContext());
596   auto Int32Ty = Type::getInt32Ty(Func.getContext());
597   auto Int16Ty = Type::getInt16Ty(Func.getContext());
598 
599   Offset = alignTo(Offset, ST.getAlignmentForImplicitArgPtr());
600   emitKernelArg(DL, Int32Ty, Align(4), "hidden_block_count_x", Offset, Args);
601   emitKernelArg(DL, Int32Ty, Align(4), "hidden_block_count_y", Offset, Args);
602   emitKernelArg(DL, Int32Ty, Align(4), "hidden_block_count_z", Offset, Args);
603 
604   emitKernelArg(DL, Int16Ty, Align(2), "hidden_group_size_x", Offset, Args);
605   emitKernelArg(DL, Int16Ty, Align(2), "hidden_group_size_y", Offset, Args);
606   emitKernelArg(DL, Int16Ty, Align(2), "hidden_group_size_z", Offset, Args);
607 
608   emitKernelArg(DL, Int16Ty, Align(2), "hidden_remainder_x", Offset, Args);
609   emitKernelArg(DL, Int16Ty, Align(2), "hidden_remainder_y", Offset, Args);
610   emitKernelArg(DL, Int16Ty, Align(2), "hidden_remainder_z", Offset, Args);
611 
612   // Reserved for hidden_tool_correlation_id.
613   Offset += 8;
614 
615   Offset += 8; // Reserved.
616 
617   emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_x", Offset, Args);
618   emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_y", Offset, Args);
619   emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_z", Offset, Args);
620 
621   emitKernelArg(DL, Int16Ty, Align(2), "hidden_grid_dims", Offset, Args);
622 
623   Offset += 6; // Reserved.
624   auto Int8PtrTy =
625       PointerType::get(Func.getContext(), AMDGPUAS::GLOBAL_ADDRESS);
626 
627   if (M->getNamedMetadata("llvm.printf.fmts")) {
628     emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_printf_buffer", Offset,
629                   Args);
630   } else {
631     Offset += 8; // Skipped.
632   }
633 
634   if (!Func.hasFnAttribute("amdgpu-no-hostcall-ptr")) {
635     emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_hostcall_buffer", Offset,
636                   Args);
637   } else {
638     Offset += 8; // Skipped.
639   }
640 
641   if (!Func.hasFnAttribute("amdgpu-no-multigrid-sync-arg")) {
642     emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_multigrid_sync_arg", Offset,
643                 Args);
644   } else {
645     Offset += 8; // Skipped.
646   }
647 
648   if (!Func.hasFnAttribute("amdgpu-no-heap-ptr"))
649     emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_heap_v1", Offset, Args);
650   else
651     Offset += 8; // Skipped.
652 
653   if (!Func.hasFnAttribute("amdgpu-no-default-queue")) {
654     emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_default_queue", Offset,
655                   Args);
656   } else {
657     Offset += 8; // Skipped.
658   }
659 
660   if (!Func.hasFnAttribute("amdgpu-no-completion-action")) {
661     emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_completion_action", Offset,
662                   Args);
663   } else {
664     Offset += 8; // Skipped.
665   }
666 
667   // Emit argument for hidden dynamic lds size
668   if (MFI.isDynamicLDSUsed()) {
669     emitKernelArg(DL, Int32Ty, Align(4), "hidden_dynamic_lds_size", Offset,
670                   Args);
671   } else {
672     Offset += 4; // skipped
673   }
674 
675   Offset += 68; // Reserved.
676 
677   // hidden_private_base and hidden_shared_base are only when the subtarget has
678   // ApertureRegs.
679   if (!ST.hasApertureRegs()) {
680     emitKernelArg(DL, Int32Ty, Align(4), "hidden_private_base", Offset, Args);
681     emitKernelArg(DL, Int32Ty, Align(4), "hidden_shared_base", Offset, Args);
682   } else {
683     Offset += 8; // Skipped.
684   }
685 
686   if (MFI.getUserSGPRInfo().hasQueuePtr())
687     emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_queue_ptr", Offset, Args);
688 }
689 
690 void MetadataStreamerMsgPackV5::emitKernelAttrs(const Function &Func,
691                                                 msgpack::MapDocNode Kern) {
692   MetadataStreamerMsgPackV4::emitKernelAttrs(Func, Kern);
693 
694   if (Func.getFnAttribute("uniform-work-group-size").getValueAsBool())
695     Kern[".uniform_work_group_size"] = Kern.getDocument()->getNode(1);
696 }
697 
698 //===----------------------------------------------------------------------===//
699 // HSAMetadataStreamerV6
700 //===----------------------------------------------------------------------===//
701 
702 void MetadataStreamerMsgPackV6::emitVersion() {
703   auto Version = HSAMetadataDoc->getArrayNode();
704   Version.push_back(Version.getDocument()->getNode(VersionMajorV6));
705   Version.push_back(Version.getDocument()->getNode(VersionMinorV6));
706   getRootMetadata("amdhsa.version") = Version;
707 }
708 
709 } // end namespace AMDGPU::HSAMD
710 } // end namespace llvm
711