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 using namespace llvm;
23 
24 static std::pair<Type *, Align> getArgumentTypeAlign(const Argument &Arg,
25                                                      const DataLayout &DL) {
26   Type *Ty = Arg.getType();
27   MaybeAlign ArgAlign;
28   if (Arg.hasByRefAttr()) {
29     Ty = Arg.getParamByRefType();
30     ArgAlign = Arg.getParamAlign();
31   }
32 
33   if (!ArgAlign)
34     ArgAlign = DL.getABITypeAlign(Ty);
35 
36   return std::make_pair(Ty, *ArgAlign);
37 }
38 
39 namespace llvm {
40 
41 static cl::opt<bool> DumpHSAMetadata(
42     "amdgpu-dump-hsa-metadata",
43     cl::desc("Dump AMDGPU HSA Metadata"));
44 static cl::opt<bool> VerifyHSAMetadata(
45     "amdgpu-verify-hsa-metadata",
46     cl::desc("Verify AMDGPU HSA Metadata"));
47 
48 namespace AMDGPU {
49 namespace HSAMD {
50 
51 //===----------------------------------------------------------------------===//
52 // HSAMetadataStreamerV2
53 //===----------------------------------------------------------------------===//
54 void MetadataStreamerV2::dump(StringRef HSAMetadataString) const {
55   errs() << "AMDGPU HSA Metadata:\n" << HSAMetadataString << '\n';
56 }
57 
58 void MetadataStreamerV2::verify(StringRef HSAMetadataString) const {
59   errs() << "AMDGPU HSA Metadata Parser Test: ";
60 
61   HSAMD::Metadata FromHSAMetadataString;
62   if (fromString(HSAMetadataString, FromHSAMetadataString)) {
63     errs() << "FAIL\n";
64     return;
65   }
66 
67   std::string ToHSAMetadataString;
68   if (toString(FromHSAMetadataString, ToHSAMetadataString)) {
69     errs() << "FAIL\n";
70     return;
71   }
72 
73   errs() << (HSAMetadataString == ToHSAMetadataString ? "PASS" : "FAIL")
74          << '\n';
75   if (HSAMetadataString != ToHSAMetadataString) {
76     errs() << "Original input: " << HSAMetadataString << '\n'
77            << "Produced output: " << ToHSAMetadataString << '\n';
78   }
79 }
80 
81 AccessQualifier
82 MetadataStreamerV2::getAccessQualifier(StringRef AccQual) const {
83   if (AccQual.empty())
84     return AccessQualifier::Unknown;
85 
86   return StringSwitch<AccessQualifier>(AccQual)
87              .Case("read_only",  AccessQualifier::ReadOnly)
88              .Case("write_only", AccessQualifier::WriteOnly)
89              .Case("read_write", AccessQualifier::ReadWrite)
90              .Default(AccessQualifier::Default);
91 }
92 
93 AddressSpaceQualifier
94 MetadataStreamerV2::getAddressSpaceQualifier(
95     unsigned AddressSpace) const {
96   switch (AddressSpace) {
97   case AMDGPUAS::PRIVATE_ADDRESS:
98     return AddressSpaceQualifier::Private;
99   case AMDGPUAS::GLOBAL_ADDRESS:
100     return AddressSpaceQualifier::Global;
101   case AMDGPUAS::CONSTANT_ADDRESS:
102     return AddressSpaceQualifier::Constant;
103   case AMDGPUAS::LOCAL_ADDRESS:
104     return AddressSpaceQualifier::Local;
105   case AMDGPUAS::FLAT_ADDRESS:
106     return AddressSpaceQualifier::Generic;
107   case AMDGPUAS::REGION_ADDRESS:
108     return AddressSpaceQualifier::Region;
109   default:
110     return AddressSpaceQualifier::Unknown;
111   }
112 }
113 
114 ValueKind MetadataStreamerV2::getValueKind(Type *Ty, StringRef TypeQual,
115                                            StringRef BaseTypeName) const {
116   if (TypeQual.contains("pipe"))
117     return ValueKind::Pipe;
118 
119   return StringSwitch<ValueKind>(BaseTypeName)
120              .Case("image1d_t", ValueKind::Image)
121              .Case("image1d_array_t", ValueKind::Image)
122              .Case("image1d_buffer_t", ValueKind::Image)
123              .Case("image2d_t", ValueKind::Image)
124              .Case("image2d_array_t", ValueKind::Image)
125              .Case("image2d_array_depth_t", ValueKind::Image)
126              .Case("image2d_array_msaa_t", ValueKind::Image)
127              .Case("image2d_array_msaa_depth_t", ValueKind::Image)
128              .Case("image2d_depth_t", ValueKind::Image)
129              .Case("image2d_msaa_t", ValueKind::Image)
130              .Case("image2d_msaa_depth_t", ValueKind::Image)
131              .Case("image3d_t", ValueKind::Image)
132              .Case("sampler_t", ValueKind::Sampler)
133              .Case("queue_t", ValueKind::Queue)
134              .Default(isa<PointerType>(Ty) ?
135                           (Ty->getPointerAddressSpace() ==
136                            AMDGPUAS::LOCAL_ADDRESS ?
137                            ValueKind::DynamicSharedPointer :
138                            ValueKind::GlobalBuffer) :
139                       ValueKind::ByValue);
140 }
141 
142 std::string MetadataStreamerV2::getTypeName(Type *Ty, bool Signed) const {
143   switch (Ty->getTypeID()) {
144   case Type::IntegerTyID: {
145     if (!Signed)
146       return (Twine('u') + getTypeName(Ty, true)).str();
147 
148     auto BitWidth = Ty->getIntegerBitWidth();
149     switch (BitWidth) {
150     case 8:
151       return "char";
152     case 16:
153       return "short";
154     case 32:
155       return "int";
156     case 64:
157       return "long";
158     default:
159       return (Twine('i') + Twine(BitWidth)).str();
160     }
161   }
162   case Type::HalfTyID:
163     return "half";
164   case Type::FloatTyID:
165     return "float";
166   case Type::DoubleTyID:
167     return "double";
168   case Type::FixedVectorTyID: {
169     auto VecTy = cast<FixedVectorType>(Ty);
170     auto ElTy = VecTy->getElementType();
171     auto NumElements = VecTy->getNumElements();
172     return (Twine(getTypeName(ElTy, Signed)) + Twine(NumElements)).str();
173   }
174   default:
175     return "unknown";
176   }
177 }
178 
179 std::vector<uint32_t>
180 MetadataStreamerV2::getWorkGroupDimensions(MDNode *Node) const {
181   std::vector<uint32_t> Dims;
182   if (Node->getNumOperands() != 3)
183     return Dims;
184 
185   for (auto &Op : Node->operands())
186     Dims.push_back(mdconst::extract<ConstantInt>(Op)->getZExtValue());
187   return Dims;
188 }
189 
190 Kernel::CodeProps::Metadata
191 MetadataStreamerV2::getHSACodeProps(const MachineFunction &MF,
192                                     const SIProgramInfo &ProgramInfo) const {
193   const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>();
194   const SIMachineFunctionInfo &MFI = *MF.getInfo<SIMachineFunctionInfo>();
195   HSAMD::Kernel::CodeProps::Metadata HSACodeProps;
196   const Function &F = MF.getFunction();
197 
198   assert(F.getCallingConv() == CallingConv::AMDGPU_KERNEL ||
199          F.getCallingConv() == CallingConv::SPIR_KERNEL);
200 
201   Align MaxKernArgAlign;
202   HSACodeProps.mKernargSegmentSize = STM.getKernArgSegmentSize(F,
203                                                                MaxKernArgAlign);
204   HSACodeProps.mKernargSegmentAlign =
205     std::max(MaxKernArgAlign, Align(4)).value();
206 
207   HSACodeProps.mGroupSegmentFixedSize = ProgramInfo.LDSSize;
208   HSACodeProps.mPrivateSegmentFixedSize = ProgramInfo.ScratchSize;
209   HSACodeProps.mWavefrontSize = STM.getWavefrontSize();
210   HSACodeProps.mNumSGPRs = ProgramInfo.NumSGPR;
211   HSACodeProps.mNumVGPRs = ProgramInfo.NumVGPR;
212   HSACodeProps.mMaxFlatWorkGroupSize = MFI.getMaxFlatWorkGroupSize();
213   HSACodeProps.mIsDynamicCallStack = ProgramInfo.DynamicCallStack;
214   HSACodeProps.mIsXNACKEnabled = STM.isXNACKEnabled();
215   HSACodeProps.mNumSpilledSGPRs = MFI.getNumSpilledSGPRs();
216   HSACodeProps.mNumSpilledVGPRs = MFI.getNumSpilledVGPRs();
217 
218   return HSACodeProps;
219 }
220 
221 Kernel::DebugProps::Metadata
222 MetadataStreamerV2::getHSADebugProps(const MachineFunction &MF,
223                                      const SIProgramInfo &ProgramInfo) const {
224   return HSAMD::Kernel::DebugProps::Metadata();
225 }
226 
227 void MetadataStreamerV2::emitVersion() {
228   auto &Version = HSAMetadata.mVersion;
229 
230   Version.push_back(VersionMajorV2);
231   Version.push_back(VersionMinorV2);
232 }
233 
234 void MetadataStreamerV2::emitPrintf(const Module &Mod) {
235   auto &Printf = HSAMetadata.mPrintf;
236 
237   auto Node = Mod.getNamedMetadata("llvm.printf.fmts");
238   if (!Node)
239     return;
240 
241   for (auto Op : Node->operands())
242     if (Op->getNumOperands())
243       Printf.push_back(
244           std::string(cast<MDString>(Op->getOperand(0))->getString()));
245 }
246 
247 void MetadataStreamerV2::emitKernelLanguage(const Function &Func) {
248   auto &Kernel = HSAMetadata.mKernels.back();
249 
250   // TODO: What about other languages?
251   auto Node = Func.getParent()->getNamedMetadata("opencl.ocl.version");
252   if (!Node || !Node->getNumOperands())
253     return;
254   auto Op0 = Node->getOperand(0);
255   if (Op0->getNumOperands() <= 1)
256     return;
257 
258   Kernel.mLanguage = "OpenCL C";
259   Kernel.mLanguageVersion.push_back(
260       mdconst::extract<ConstantInt>(Op0->getOperand(0))->getZExtValue());
261   Kernel.mLanguageVersion.push_back(
262       mdconst::extract<ConstantInt>(Op0->getOperand(1))->getZExtValue());
263 }
264 
265 void MetadataStreamerV2::emitKernelAttrs(const Function &Func) {
266   auto &Attrs = HSAMetadata.mKernels.back().mAttrs;
267 
268   if (auto Node = Func.getMetadata("reqd_work_group_size"))
269     Attrs.mReqdWorkGroupSize = getWorkGroupDimensions(Node);
270   if (auto Node = Func.getMetadata("work_group_size_hint"))
271     Attrs.mWorkGroupSizeHint = getWorkGroupDimensions(Node);
272   if (auto Node = Func.getMetadata("vec_type_hint")) {
273     Attrs.mVecTypeHint = getTypeName(
274         cast<ValueAsMetadata>(Node->getOperand(0))->getType(),
275         mdconst::extract<ConstantInt>(Node->getOperand(1))->getZExtValue());
276   }
277   if (Func.hasFnAttribute("runtime-handle")) {
278     Attrs.mRuntimeHandle =
279         Func.getFnAttribute("runtime-handle").getValueAsString().str();
280   }
281 }
282 
283 void MetadataStreamerV2::emitKernelArgs(const Function &Func,
284                                         const GCNSubtarget &ST) {
285   for (auto &Arg : Func.args())
286     emitKernelArg(Arg);
287 
288   emitHiddenKernelArgs(Func, ST);
289 }
290 
291 void MetadataStreamerV2::emitKernelArg(const Argument &Arg) {
292   auto Func = Arg.getParent();
293   auto ArgNo = Arg.getArgNo();
294   const MDNode *Node;
295 
296   StringRef Name;
297   Node = Func->getMetadata("kernel_arg_name");
298   if (Node && ArgNo < Node->getNumOperands())
299     Name = cast<MDString>(Node->getOperand(ArgNo))->getString();
300   else if (Arg.hasName())
301     Name = Arg.getName();
302 
303   StringRef TypeName;
304   Node = Func->getMetadata("kernel_arg_type");
305   if (Node && ArgNo < Node->getNumOperands())
306     TypeName = cast<MDString>(Node->getOperand(ArgNo))->getString();
307 
308   StringRef BaseTypeName;
309   Node = Func->getMetadata("kernel_arg_base_type");
310   if (Node && ArgNo < Node->getNumOperands())
311     BaseTypeName = cast<MDString>(Node->getOperand(ArgNo))->getString();
312 
313   StringRef AccQual;
314   if (Arg.getType()->isPointerTy() && Arg.onlyReadsMemory() &&
315       Arg.hasNoAliasAttr()) {
316     AccQual = "read_only";
317   } else {
318     Node = Func->getMetadata("kernel_arg_access_qual");
319     if (Node && ArgNo < Node->getNumOperands())
320       AccQual = cast<MDString>(Node->getOperand(ArgNo))->getString();
321   }
322 
323   StringRef TypeQual;
324   Node = Func->getMetadata("kernel_arg_type_qual");
325   if (Node && ArgNo < Node->getNumOperands())
326     TypeQual = cast<MDString>(Node->getOperand(ArgNo))->getString();
327 
328   const DataLayout &DL = Func->getParent()->getDataLayout();
329 
330   MaybeAlign PointeeAlign;
331   if (auto PtrTy = dyn_cast<PointerType>(Arg.getType())) {
332     if (PtrTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) {
333       // FIXME: Should report this for all address spaces
334       PointeeAlign = Arg.getParamAlign().valueOrOne();
335     }
336   }
337 
338   Type *ArgTy;
339   Align ArgAlign;
340   std::tie(ArgTy, ArgAlign) = getArgumentTypeAlign(Arg, DL);
341 
342   emitKernelArg(DL, ArgTy, ArgAlign,
343                 getValueKind(ArgTy, TypeQual, BaseTypeName), PointeeAlign, Name,
344                 TypeName, BaseTypeName, AccQual, TypeQual);
345 }
346 
347 void MetadataStreamerV2::emitKernelArg(const DataLayout &DL, Type *Ty,
348                                        Align Alignment, ValueKind ValueKind,
349                                        MaybeAlign PointeeAlign, StringRef Name,
350                                        StringRef TypeName,
351                                        StringRef BaseTypeName,
352                                        StringRef AccQual, StringRef TypeQual) {
353   HSAMetadata.mKernels.back().mArgs.push_back(Kernel::Arg::Metadata());
354   auto &Arg = HSAMetadata.mKernels.back().mArgs.back();
355 
356   Arg.mName = std::string(Name);
357   Arg.mTypeName = std::string(TypeName);
358   Arg.mSize = DL.getTypeAllocSize(Ty);
359   Arg.mAlign = Alignment.value();
360   Arg.mValueKind = ValueKind;
361   Arg.mPointeeAlign = PointeeAlign ? PointeeAlign->value() : 0;
362 
363   if (auto PtrTy = dyn_cast<PointerType>(Ty))
364     Arg.mAddrSpaceQual = getAddressSpaceQualifier(PtrTy->getAddressSpace());
365 
366   Arg.mAccQual = getAccessQualifier(AccQual);
367 
368   // TODO: Emit Arg.mActualAccQual.
369 
370   SmallVector<StringRef, 1> SplitTypeQuals;
371   TypeQual.split(SplitTypeQuals, " ", -1, false);
372   for (StringRef Key : SplitTypeQuals) {
373     auto P = StringSwitch<bool*>(Key)
374                  .Case("const",    &Arg.mIsConst)
375                  .Case("restrict", &Arg.mIsRestrict)
376                  .Case("volatile", &Arg.mIsVolatile)
377                  .Case("pipe",     &Arg.mIsPipe)
378                  .Default(nullptr);
379     if (P)
380       *P = true;
381   }
382 }
383 
384 void MetadataStreamerV2::emitHiddenKernelArgs(const Function &Func,
385                                               const GCNSubtarget &ST) {
386   unsigned HiddenArgNumBytes = ST.getImplicitArgNumBytes(Func);
387   if (!HiddenArgNumBytes)
388     return;
389 
390   auto &DL = Func.getParent()->getDataLayout();
391   auto Int64Ty = Type::getInt64Ty(Func.getContext());
392 
393   if (HiddenArgNumBytes >= 8)
394     emitKernelArg(DL, Int64Ty, Align(8), ValueKind::HiddenGlobalOffsetX);
395   if (HiddenArgNumBytes >= 16)
396     emitKernelArg(DL, Int64Ty, Align(8), ValueKind::HiddenGlobalOffsetY);
397   if (HiddenArgNumBytes >= 24)
398     emitKernelArg(DL, Int64Ty, Align(8), ValueKind::HiddenGlobalOffsetZ);
399 
400   auto Int8PtrTy = Type::getInt8PtrTy(Func.getContext(),
401                                       AMDGPUAS::GLOBAL_ADDRESS);
402 
403   if (HiddenArgNumBytes >= 32) {
404     // We forbid the use of features requiring hostcall when compiling OpenCL
405     // before code object V5, which makes the mutual exclusion between the
406     // "printf buffer" and "hostcall buffer" here sound.
407     if (Func.getParent()->getNamedMetadata("llvm.printf.fmts"))
408       emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenPrintfBuffer);
409     else if (!Func.hasFnAttribute("amdgpu-no-hostcall-ptr"))
410       emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenHostcallBuffer);
411     else
412       emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenNone);
413   }
414 
415   // Emit "default queue" and "completion action" arguments if enqueue kernel is
416   // used, otherwise emit dummy "none" arguments.
417   if (HiddenArgNumBytes >= 48) {
418     if (Func.hasFnAttribute("calls-enqueue-kernel")) {
419       emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenDefaultQueue);
420       emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenCompletionAction);
421     } else {
422       emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenNone);
423       emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenNone);
424     }
425   }
426 
427   // Emit the pointer argument for multi-grid object.
428   if (HiddenArgNumBytes >= 56)
429     emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenMultiGridSyncArg);
430 }
431 
432 bool MetadataStreamerV2::emitTo(AMDGPUTargetStreamer &TargetStreamer) {
433   return TargetStreamer.EmitHSAMetadata(getHSAMetadata());
434 }
435 
436 void MetadataStreamerV2::begin(const Module &Mod,
437                                const IsaInfo::AMDGPUTargetID &TargetID) {
438   emitVersion();
439   emitPrintf(Mod);
440 }
441 
442 void MetadataStreamerV2::end() {
443   std::string HSAMetadataString;
444   if (toString(HSAMetadata, HSAMetadataString))
445     return;
446 
447   if (DumpHSAMetadata)
448     dump(HSAMetadataString);
449   if (VerifyHSAMetadata)
450     verify(HSAMetadataString);
451 }
452 
453 void MetadataStreamerV2::emitKernel(const MachineFunction &MF,
454                                     const SIProgramInfo &ProgramInfo) {
455   auto &Func = MF.getFunction();
456   if (Func.getCallingConv() != CallingConv::AMDGPU_KERNEL)
457     return;
458 
459   auto CodeProps = getHSACodeProps(MF, ProgramInfo);
460   auto DebugProps = getHSADebugProps(MF, ProgramInfo);
461 
462   HSAMetadata.mKernels.push_back(Kernel::Metadata());
463   auto &Kernel = HSAMetadata.mKernels.back();
464 
465   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
466   Kernel.mName = std::string(Func.getName());
467   Kernel.mSymbolName = (Twine(Func.getName()) + Twine("@kd")).str();
468   emitKernelLanguage(Func);
469   emitKernelAttrs(Func);
470   emitKernelArgs(Func, ST);
471   HSAMetadata.mKernels.back().mCodeProps = CodeProps;
472   HSAMetadata.mKernels.back().mDebugProps = DebugProps;
473 }
474 
475 //===----------------------------------------------------------------------===//
476 // HSAMetadataStreamerV3
477 //===----------------------------------------------------------------------===//
478 
479 void MetadataStreamerV3::dump(StringRef HSAMetadataString) const {
480   errs() << "AMDGPU HSA Metadata:\n" << HSAMetadataString << '\n';
481 }
482 
483 void MetadataStreamerV3::verify(StringRef HSAMetadataString) const {
484   errs() << "AMDGPU HSA Metadata Parser Test: ";
485 
486   msgpack::Document FromHSAMetadataString;
487 
488   if (!FromHSAMetadataString.fromYAML(HSAMetadataString)) {
489     errs() << "FAIL\n";
490     return;
491   }
492 
493   std::string ToHSAMetadataString;
494   raw_string_ostream StrOS(ToHSAMetadataString);
495   FromHSAMetadataString.toYAML(StrOS);
496 
497   errs() << (HSAMetadataString == StrOS.str() ? "PASS" : "FAIL") << '\n';
498   if (HSAMetadataString != ToHSAMetadataString) {
499     errs() << "Original input: " << HSAMetadataString << '\n'
500            << "Produced output: " << StrOS.str() << '\n';
501   }
502 }
503 
504 Optional<StringRef>
505 MetadataStreamerV3::getAccessQualifier(StringRef AccQual) const {
506   return StringSwitch<Optional<StringRef>>(AccQual)
507       .Case("read_only", StringRef("read_only"))
508       .Case("write_only", StringRef("write_only"))
509       .Case("read_write", StringRef("read_write"))
510       .Default(None);
511 }
512 
513 Optional<StringRef>
514 MetadataStreamerV3::getAddressSpaceQualifier(unsigned AddressSpace) const {
515   switch (AddressSpace) {
516   case AMDGPUAS::PRIVATE_ADDRESS:
517     return StringRef("private");
518   case AMDGPUAS::GLOBAL_ADDRESS:
519     return StringRef("global");
520   case AMDGPUAS::CONSTANT_ADDRESS:
521     return StringRef("constant");
522   case AMDGPUAS::LOCAL_ADDRESS:
523     return StringRef("local");
524   case AMDGPUAS::FLAT_ADDRESS:
525     return StringRef("generic");
526   case AMDGPUAS::REGION_ADDRESS:
527     return StringRef("region");
528   default:
529     return None;
530   }
531 }
532 
533 StringRef MetadataStreamerV3::getValueKind(Type *Ty, StringRef TypeQual,
534                                            StringRef BaseTypeName) const {
535   if (TypeQual.contains("pipe"))
536     return "pipe";
537 
538   return StringSwitch<StringRef>(BaseTypeName)
539       .Case("image1d_t", "image")
540       .Case("image1d_array_t", "image")
541       .Case("image1d_buffer_t", "image")
542       .Case("image2d_t", "image")
543       .Case("image2d_array_t", "image")
544       .Case("image2d_array_depth_t", "image")
545       .Case("image2d_array_msaa_t", "image")
546       .Case("image2d_array_msaa_depth_t", "image")
547       .Case("image2d_depth_t", "image")
548       .Case("image2d_msaa_t", "image")
549       .Case("image2d_msaa_depth_t", "image")
550       .Case("image3d_t", "image")
551       .Case("sampler_t", "sampler")
552       .Case("queue_t", "queue")
553       .Default(isa<PointerType>(Ty)
554                    ? (Ty->getPointerAddressSpace() == AMDGPUAS::LOCAL_ADDRESS
555                           ? "dynamic_shared_pointer"
556                           : "global_buffer")
557                    : "by_value");
558 }
559 
560 std::string MetadataStreamerV3::getTypeName(Type *Ty, bool Signed) const {
561   switch (Ty->getTypeID()) {
562   case Type::IntegerTyID: {
563     if (!Signed)
564       return (Twine('u') + getTypeName(Ty, true)).str();
565 
566     auto BitWidth = Ty->getIntegerBitWidth();
567     switch (BitWidth) {
568     case 8:
569       return "char";
570     case 16:
571       return "short";
572     case 32:
573       return "int";
574     case 64:
575       return "long";
576     default:
577       return (Twine('i') + Twine(BitWidth)).str();
578     }
579   }
580   case Type::HalfTyID:
581     return "half";
582   case Type::FloatTyID:
583     return "float";
584   case Type::DoubleTyID:
585     return "double";
586   case Type::FixedVectorTyID: {
587     auto VecTy = cast<FixedVectorType>(Ty);
588     auto ElTy = VecTy->getElementType();
589     auto NumElements = VecTy->getNumElements();
590     return (Twine(getTypeName(ElTy, Signed)) + Twine(NumElements)).str();
591   }
592   default:
593     return "unknown";
594   }
595 }
596 
597 msgpack::ArrayDocNode
598 MetadataStreamerV3::getWorkGroupDimensions(MDNode *Node) const {
599   auto Dims = HSAMetadataDoc->getArrayNode();
600   if (Node->getNumOperands() != 3)
601     return Dims;
602 
603   for (auto &Op : Node->operands())
604     Dims.push_back(Dims.getDocument()->getNode(
605         uint64_t(mdconst::extract<ConstantInt>(Op)->getZExtValue())));
606   return Dims;
607 }
608 
609 void MetadataStreamerV3::emitVersion() {
610   auto Version = HSAMetadataDoc->getArrayNode();
611   Version.push_back(Version.getDocument()->getNode(VersionMajorV3));
612   Version.push_back(Version.getDocument()->getNode(VersionMinorV3));
613   getRootMetadata("amdhsa.version") = Version;
614 }
615 
616 void MetadataStreamerV3::emitPrintf(const Module &Mod) {
617   auto Node = Mod.getNamedMetadata("llvm.printf.fmts");
618   if (!Node)
619     return;
620 
621   auto Printf = HSAMetadataDoc->getArrayNode();
622   for (auto Op : Node->operands())
623     if (Op->getNumOperands())
624       Printf.push_back(Printf.getDocument()->getNode(
625           cast<MDString>(Op->getOperand(0))->getString(), /*Copy=*/true));
626   getRootMetadata("amdhsa.printf") = Printf;
627 }
628 
629 void MetadataStreamerV3::emitKernelLanguage(const Function &Func,
630                                             msgpack::MapDocNode Kern) {
631   // TODO: What about other languages?
632   auto Node = Func.getParent()->getNamedMetadata("opencl.ocl.version");
633   if (!Node || !Node->getNumOperands())
634     return;
635   auto Op0 = Node->getOperand(0);
636   if (Op0->getNumOperands() <= 1)
637     return;
638 
639   Kern[".language"] = Kern.getDocument()->getNode("OpenCL C");
640   auto LanguageVersion = Kern.getDocument()->getArrayNode();
641   LanguageVersion.push_back(Kern.getDocument()->getNode(
642       mdconst::extract<ConstantInt>(Op0->getOperand(0))->getZExtValue()));
643   LanguageVersion.push_back(Kern.getDocument()->getNode(
644       mdconst::extract<ConstantInt>(Op0->getOperand(1))->getZExtValue()));
645   Kern[".language_version"] = LanguageVersion;
646 }
647 
648 void MetadataStreamerV3::emitKernelAttrs(const Function &Func,
649                                          msgpack::MapDocNode Kern) {
650 
651   if (auto Node = Func.getMetadata("reqd_work_group_size"))
652     Kern[".reqd_workgroup_size"] = getWorkGroupDimensions(Node);
653   if (auto Node = Func.getMetadata("work_group_size_hint"))
654     Kern[".workgroup_size_hint"] = getWorkGroupDimensions(Node);
655   if (auto Node = Func.getMetadata("vec_type_hint")) {
656     Kern[".vec_type_hint"] = Kern.getDocument()->getNode(
657         getTypeName(
658             cast<ValueAsMetadata>(Node->getOperand(0))->getType(),
659             mdconst::extract<ConstantInt>(Node->getOperand(1))->getZExtValue()),
660         /*Copy=*/true);
661   }
662   if (Func.hasFnAttribute("runtime-handle")) {
663     Kern[".device_enqueue_symbol"] = Kern.getDocument()->getNode(
664         Func.getFnAttribute("runtime-handle").getValueAsString().str(),
665         /*Copy=*/true);
666   }
667   if (Func.hasFnAttribute("device-init"))
668     Kern[".kind"] = Kern.getDocument()->getNode("init");
669   else if (Func.hasFnAttribute("device-fini"))
670     Kern[".kind"] = Kern.getDocument()->getNode("fini");
671 }
672 
673 void MetadataStreamerV3::emitKernelArgs(const MachineFunction &MF,
674                                         msgpack::MapDocNode Kern) {
675   auto &Func = MF.getFunction();
676   unsigned Offset = 0;
677   auto Args = HSAMetadataDoc->getArrayNode();
678   for (auto &Arg : Func.args())
679     emitKernelArg(Arg, Offset, Args);
680 
681   emitHiddenKernelArgs(MF, Offset, Args);
682 
683   Kern[".args"] = Args;
684 }
685 
686 void MetadataStreamerV3::emitKernelArg(const Argument &Arg, unsigned &Offset,
687                                        msgpack::ArrayDocNode Args) {
688   auto Func = Arg.getParent();
689   auto ArgNo = Arg.getArgNo();
690   const MDNode *Node;
691 
692   StringRef Name;
693   Node = Func->getMetadata("kernel_arg_name");
694   if (Node && ArgNo < Node->getNumOperands())
695     Name = cast<MDString>(Node->getOperand(ArgNo))->getString();
696   else if (Arg.hasName())
697     Name = Arg.getName();
698 
699   StringRef TypeName;
700   Node = Func->getMetadata("kernel_arg_type");
701   if (Node && ArgNo < Node->getNumOperands())
702     TypeName = cast<MDString>(Node->getOperand(ArgNo))->getString();
703 
704   StringRef BaseTypeName;
705   Node = Func->getMetadata("kernel_arg_base_type");
706   if (Node && ArgNo < Node->getNumOperands())
707     BaseTypeName = cast<MDString>(Node->getOperand(ArgNo))->getString();
708 
709   StringRef AccQual;
710   if (Arg.getType()->isPointerTy() && Arg.onlyReadsMemory() &&
711       Arg.hasNoAliasAttr()) {
712     AccQual = "read_only";
713   } else {
714     Node = Func->getMetadata("kernel_arg_access_qual");
715     if (Node && ArgNo < Node->getNumOperands())
716       AccQual = cast<MDString>(Node->getOperand(ArgNo))->getString();
717   }
718 
719   StringRef TypeQual;
720   Node = Func->getMetadata("kernel_arg_type_qual");
721   if (Node && ArgNo < Node->getNumOperands())
722     TypeQual = cast<MDString>(Node->getOperand(ArgNo))->getString();
723 
724   const DataLayout &DL = Func->getParent()->getDataLayout();
725 
726   MaybeAlign PointeeAlign;
727   Type *Ty = Arg.hasByRefAttr() ? Arg.getParamByRefType() : Arg.getType();
728 
729   // FIXME: Need to distinguish in memory alignment from pointer alignment.
730   if (auto PtrTy = dyn_cast<PointerType>(Ty)) {
731     if (PtrTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS)
732       PointeeAlign = Arg.getParamAlign().valueOrOne();
733   }
734 
735   // There's no distinction between byval aggregates and raw aggregates.
736   Type *ArgTy;
737   Align ArgAlign;
738   std::tie(ArgTy, ArgAlign) = getArgumentTypeAlign(Arg, DL);
739 
740   emitKernelArg(DL, ArgTy, ArgAlign,
741                 getValueKind(ArgTy, TypeQual, BaseTypeName), Offset, Args,
742                 PointeeAlign, Name, TypeName, BaseTypeName, AccQual, TypeQual);
743 }
744 
745 void MetadataStreamerV3::emitKernelArg(
746     const DataLayout &DL, Type *Ty, Align Alignment, StringRef ValueKind,
747     unsigned &Offset, msgpack::ArrayDocNode Args, MaybeAlign PointeeAlign,
748     StringRef Name, StringRef TypeName, StringRef BaseTypeName,
749     StringRef AccQual, StringRef TypeQual) {
750   auto Arg = Args.getDocument()->getMapNode();
751 
752   if (!Name.empty())
753     Arg[".name"] = Arg.getDocument()->getNode(Name, /*Copy=*/true);
754   if (!TypeName.empty())
755     Arg[".type_name"] = Arg.getDocument()->getNode(TypeName, /*Copy=*/true);
756   auto Size = DL.getTypeAllocSize(Ty);
757   Arg[".size"] = Arg.getDocument()->getNode(Size);
758   Offset = alignTo(Offset, Alignment);
759   Arg[".offset"] = Arg.getDocument()->getNode(Offset);
760   Offset += Size;
761   Arg[".value_kind"] = Arg.getDocument()->getNode(ValueKind, /*Copy=*/true);
762   if (PointeeAlign)
763     Arg[".pointee_align"] = Arg.getDocument()->getNode(PointeeAlign->value());
764 
765   if (auto PtrTy = dyn_cast<PointerType>(Ty))
766     if (auto Qualifier = getAddressSpaceQualifier(PtrTy->getAddressSpace()))
767       Arg[".address_space"] = Arg.getDocument()->getNode(*Qualifier, /*Copy=*/true);
768 
769   if (auto AQ = getAccessQualifier(AccQual))
770     Arg[".access"] = Arg.getDocument()->getNode(*AQ, /*Copy=*/true);
771 
772   // TODO: Emit Arg[".actual_access"].
773 
774   SmallVector<StringRef, 1> SplitTypeQuals;
775   TypeQual.split(SplitTypeQuals, " ", -1, false);
776   for (StringRef Key : SplitTypeQuals) {
777     if (Key == "const")
778       Arg[".is_const"] = Arg.getDocument()->getNode(true);
779     else if (Key == "restrict")
780       Arg[".is_restrict"] = Arg.getDocument()->getNode(true);
781     else if (Key == "volatile")
782       Arg[".is_volatile"] = Arg.getDocument()->getNode(true);
783     else if (Key == "pipe")
784       Arg[".is_pipe"] = Arg.getDocument()->getNode(true);
785   }
786 
787   Args.push_back(Arg);
788 }
789 
790 void MetadataStreamerV3::emitHiddenKernelArgs(const MachineFunction &MF,
791                                               unsigned &Offset,
792                                               msgpack::ArrayDocNode Args) {
793   auto &Func = MF.getFunction();
794   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
795 
796   unsigned HiddenArgNumBytes = ST.getImplicitArgNumBytes(Func);
797   if (!HiddenArgNumBytes)
798     return;
799 
800   const Module *M = Func.getParent();
801   auto &DL = M->getDataLayout();
802   auto Int64Ty = Type::getInt64Ty(Func.getContext());
803 
804   if (HiddenArgNumBytes >= 8)
805     emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_x", Offset,
806                   Args);
807   if (HiddenArgNumBytes >= 16)
808     emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_y", Offset,
809                   Args);
810   if (HiddenArgNumBytes >= 24)
811     emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_z", Offset,
812                   Args);
813 
814   auto Int8PtrTy =
815       Type::getInt8PtrTy(Func.getContext(), AMDGPUAS::GLOBAL_ADDRESS);
816 
817   if (HiddenArgNumBytes >= 32) {
818     // We forbid the use of features requiring hostcall when compiling OpenCL
819     // before code object V5, which makes the mutual exclusion between the
820     // "printf buffer" and "hostcall buffer" here sound.
821     if (M->getNamedMetadata("llvm.printf.fmts"))
822       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_printf_buffer", Offset,
823                     Args);
824     else if (!Func.hasFnAttribute("amdgpu-no-hostcall-ptr"))
825       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_hostcall_buffer", Offset,
826                     Args);
827     else
828       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_none", Offset, Args);
829   }
830 
831   // Emit "default queue" and "completion action" arguments if enqueue kernel is
832   // used, otherwise emit dummy "none" arguments.
833   if (HiddenArgNumBytes >= 48) {
834     if (Func.hasFnAttribute("calls-enqueue-kernel")) {
835       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_default_queue", Offset,
836                     Args);
837       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_completion_action", Offset,
838                     Args);
839     } else {
840       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_none", Offset, Args);
841       emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_none", Offset, Args);
842     }
843   }
844 
845   // Emit the pointer argument for multi-grid object.
846   if (HiddenArgNumBytes >= 56)
847     emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_multigrid_sync_arg", Offset,
848                   Args);
849 }
850 
851 msgpack::MapDocNode
852 MetadataStreamerV3::getHSAKernelProps(const MachineFunction &MF,
853                                       const SIProgramInfo &ProgramInfo) const {
854   const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>();
855   const SIMachineFunctionInfo &MFI = *MF.getInfo<SIMachineFunctionInfo>();
856   const Function &F = MF.getFunction();
857 
858   auto Kern = HSAMetadataDoc->getMapNode();
859 
860   Align MaxKernArgAlign;
861   Kern[".kernarg_segment_size"] = Kern.getDocument()->getNode(
862       STM.getKernArgSegmentSize(F, MaxKernArgAlign));
863   Kern[".group_segment_fixed_size"] =
864       Kern.getDocument()->getNode(ProgramInfo.LDSSize);
865   Kern[".private_segment_fixed_size"] =
866       Kern.getDocument()->getNode(ProgramInfo.ScratchSize);
867 
868   // FIXME: The metadata treats the minimum as 16?
869   Kern[".kernarg_segment_align"] =
870       Kern.getDocument()->getNode(std::max(Align(4), MaxKernArgAlign).value());
871   Kern[".wavefront_size"] =
872       Kern.getDocument()->getNode(STM.getWavefrontSize());
873   Kern[".sgpr_count"] = Kern.getDocument()->getNode(ProgramInfo.NumSGPR);
874   Kern[".vgpr_count"] = Kern.getDocument()->getNode(ProgramInfo.NumVGPR);
875 
876   // Only add AGPR count to metadata for supported devices
877   if (STM.hasMAIInsts()) {
878     Kern[".agpr_count"] = Kern.getDocument()->getNode(ProgramInfo.NumAccVGPR);
879   }
880 
881   Kern[".max_flat_workgroup_size"] =
882       Kern.getDocument()->getNode(MFI.getMaxFlatWorkGroupSize());
883   Kern[".sgpr_spill_count"] =
884       Kern.getDocument()->getNode(MFI.getNumSpilledSGPRs());
885   Kern[".vgpr_spill_count"] =
886       Kern.getDocument()->getNode(MFI.getNumSpilledVGPRs());
887 
888   return Kern;
889 }
890 
891 bool MetadataStreamerV3::emitTo(AMDGPUTargetStreamer &TargetStreamer) {
892   return TargetStreamer.EmitHSAMetadata(*HSAMetadataDoc, true);
893 }
894 
895 void MetadataStreamerV3::begin(const Module &Mod,
896                                const IsaInfo::AMDGPUTargetID &TargetID) {
897   emitVersion();
898   emitPrintf(Mod);
899   getRootMetadata("amdhsa.kernels") = HSAMetadataDoc->getArrayNode();
900 }
901 
902 void MetadataStreamerV3::end() {
903   std::string HSAMetadataString;
904   raw_string_ostream StrOS(HSAMetadataString);
905   HSAMetadataDoc->toYAML(StrOS);
906 
907   if (DumpHSAMetadata)
908     dump(StrOS.str());
909   if (VerifyHSAMetadata)
910     verify(StrOS.str());
911 }
912 
913 void MetadataStreamerV3::emitKernel(const MachineFunction &MF,
914                                     const SIProgramInfo &ProgramInfo) {
915   auto &Func = MF.getFunction();
916   auto Kern = getHSAKernelProps(MF, ProgramInfo);
917 
918   assert(Func.getCallingConv() == CallingConv::AMDGPU_KERNEL ||
919          Func.getCallingConv() == CallingConv::SPIR_KERNEL);
920 
921   auto Kernels =
922       getRootMetadata("amdhsa.kernels").getArray(/*Convert=*/true);
923 
924   {
925     Kern[".name"] = Kern.getDocument()->getNode(Func.getName());
926     Kern[".symbol"] = Kern.getDocument()->getNode(
927         (Twine(Func.getName()) + Twine(".kd")).str(), /*Copy=*/true);
928     emitKernelLanguage(Func, Kern);
929     emitKernelAttrs(Func, Kern);
930     emitKernelArgs(MF, Kern);
931   }
932 
933   Kernels.push_back(Kern);
934 }
935 
936 //===----------------------------------------------------------------------===//
937 // HSAMetadataStreamerV4
938 //===----------------------------------------------------------------------===//
939 
940 void MetadataStreamerV4::emitVersion() {
941   auto Version = HSAMetadataDoc->getArrayNode();
942   Version.push_back(Version.getDocument()->getNode(VersionMajorV4));
943   Version.push_back(Version.getDocument()->getNode(VersionMinorV4));
944   getRootMetadata("amdhsa.version") = Version;
945 }
946 
947 void MetadataStreamerV4::emitTargetID(const IsaInfo::AMDGPUTargetID &TargetID) {
948   getRootMetadata("amdhsa.target") =
949       HSAMetadataDoc->getNode(TargetID.toString(), /*Copy=*/true);
950 }
951 
952 void MetadataStreamerV4::begin(const Module &Mod,
953                                const IsaInfo::AMDGPUTargetID &TargetID) {
954   emitVersion();
955   emitTargetID(TargetID);
956   emitPrintf(Mod);
957   getRootMetadata("amdhsa.kernels") = HSAMetadataDoc->getArrayNode();
958 }
959 
960 //===----------------------------------------------------------------------===//
961 // HSAMetadataStreamerV5
962 //===----------------------------------------------------------------------===//
963 
964 void MetadataStreamerV5::emitVersion() {
965   auto Version = HSAMetadataDoc->getArrayNode();
966   Version.push_back(Version.getDocument()->getNode(VersionMajorV5));
967   Version.push_back(Version.getDocument()->getNode(VersionMinorV5));
968   getRootMetadata("amdhsa.version") = Version;
969 }
970 
971 void MetadataStreamerV5::emitHiddenKernelArgs(const MachineFunction &MF,
972                                               unsigned &Offset,
973                                               msgpack::ArrayDocNode Args) {
974   auto &Func = MF.getFunction();
975   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
976   const Module *M = Func.getParent();
977   auto &DL = M->getDataLayout();
978   const SIMachineFunctionInfo &MFI = *MF.getInfo<SIMachineFunctionInfo>();
979 
980   auto Int64Ty = Type::getInt64Ty(Func.getContext());
981   auto Int32Ty = Type::getInt32Ty(Func.getContext());
982   auto Int16Ty = Type::getInt16Ty(Func.getContext());
983 
984   emitKernelArg(DL, Int32Ty, Align(4), "hidden_block_count_x", Offset, Args);
985   emitKernelArg(DL, Int32Ty, Align(4), "hidden_block_count_y", Offset, Args);
986   emitKernelArg(DL, Int32Ty, Align(4), "hidden_block_count_z", Offset, Args);
987 
988   emitKernelArg(DL, Int16Ty, Align(2), "hidden_group_size_x", Offset, Args);
989   emitKernelArg(DL, Int16Ty, Align(2), "hidden_group_size_y", Offset, Args);
990   emitKernelArg(DL, Int16Ty, Align(2), "hidden_group_size_z", Offset, Args);
991 
992   emitKernelArg(DL, Int16Ty, Align(2), "hidden_remainder_x", Offset, Args);
993   emitKernelArg(DL, Int16Ty, Align(2), "hidden_remainder_y", Offset, Args);
994   emitKernelArg(DL, Int16Ty, Align(2), "hidden_remainder_z", Offset, Args);
995 
996   // Reserved for hidden_tool_correlation_id.
997   Offset += 8;
998 
999   Offset += 8; // Reserved.
1000 
1001   emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_x", Offset, Args);
1002   emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_y", Offset, Args);
1003   emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_z", Offset, Args);
1004 
1005   emitKernelArg(DL, Int16Ty, Align(2), "hidden_grid_dims", Offset, Args);
1006 
1007   Offset += 6; // Reserved.
1008   auto Int8PtrTy =
1009       Type::getInt8PtrTy(Func.getContext(), AMDGPUAS::GLOBAL_ADDRESS);
1010 
1011   if (M->getNamedMetadata("llvm.printf.fmts")) {
1012     emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_printf_buffer", Offset,
1013                   Args);
1014   } else
1015     Offset += 8; // Skipped.
1016 
1017   if (!Func.hasFnAttribute("amdgpu-no-hostcall-ptr")) {
1018     emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_hostcall_buffer", Offset,
1019                   Args);
1020   } else
1021     Offset += 8; // Skipped.
1022 
1023   emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_multigrid_sync_arg", Offset,
1024                 Args);
1025 
1026   if (!Func.hasFnAttribute("amdgpu-no-heap-ptr"))
1027     emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_heap_v1", Offset, Args);
1028   else
1029     Offset += 8; // Skipped.
1030 
1031   if (Func.hasFnAttribute("calls-enqueue-kernel")) {
1032     emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_default_queue", Offset,
1033                   Args);
1034     emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_completion_action", Offset,
1035                   Args);
1036   } else
1037     Offset += 16; // Skipped.
1038 
1039   Offset += 72; // Reserved.
1040 
1041   // hidden_private_base and hidden_shared_base are only when the subtarget has
1042   // ApertureRegs.
1043   if (!ST.hasApertureRegs()) {
1044     emitKernelArg(DL, Int32Ty, Align(4), "hidden_private_base", Offset, Args);
1045     emitKernelArg(DL, Int32Ty, Align(4), "hidden_shared_base", Offset, Args);
1046   } else
1047     Offset += 8; // Skipped.
1048 
1049   if (MFI.hasQueuePtr())
1050     emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_queue_ptr", Offset, Args);
1051 }
1052 
1053 } // end namespace HSAMD
1054 } // end namespace AMDGPU
1055 } // end namespace llvm
1056