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
getArgumentTypeAlign(const Argument & Arg,const DataLayout & DL)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::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 // HSAMetadataStreamerV4
53 //===----------------------------------------------------------------------===//
54
dump(StringRef HSAMetadataString) const55 void MetadataStreamerMsgPackV4::dump(StringRef HSAMetadataString) const {
56 errs() << "AMDGPU HSA Metadata:\n" << HSAMetadataString << '\n';
57 }
58
verify(StringRef HSAMetadataString) const59 void MetadataStreamerMsgPackV4::verify(StringRef HSAMetadataString) const {
60 errs() << "AMDGPU HSA Metadata Parser Test: ";
61
62 msgpack::Document FromHSAMetadataString;
63
64 if (!FromHSAMetadataString.fromYAML(HSAMetadataString)) {
65 errs() << "FAIL\n";
66 return;
67 }
68
69 std::string ToHSAMetadataString;
70 raw_string_ostream StrOS(ToHSAMetadataString);
71 FromHSAMetadataString.toYAML(StrOS);
72
73 errs() << (HSAMetadataString == StrOS.str() ? "PASS" : "FAIL") << '\n';
74 if (HSAMetadataString != ToHSAMetadataString) {
75 errs() << "Original input: " << HSAMetadataString << '\n'
76 << "Produced output: " << StrOS.str() << '\n';
77 }
78 }
79
80 std::optional<StringRef>
getAccessQualifier(StringRef AccQual) const81 MetadataStreamerMsgPackV4::getAccessQualifier(StringRef AccQual) const {
82 return StringSwitch<std::optional<StringRef>>(AccQual)
83 .Case("read_only", StringRef("read_only"))
84 .Case("write_only", StringRef("write_only"))
85 .Case("read_write", StringRef("read_write"))
86 .Default(std::nullopt);
87 }
88
getAddressSpaceQualifier(unsigned AddressSpace) const89 std::optional<StringRef> MetadataStreamerMsgPackV4::getAddressSpaceQualifier(
90 unsigned AddressSpace) const {
91 switch (AddressSpace) {
92 case AMDGPUAS::PRIVATE_ADDRESS:
93 return StringRef("private");
94 case AMDGPUAS::GLOBAL_ADDRESS:
95 return StringRef("global");
96 case AMDGPUAS::CONSTANT_ADDRESS:
97 return StringRef("constant");
98 case AMDGPUAS::LOCAL_ADDRESS:
99 return StringRef("local");
100 case AMDGPUAS::FLAT_ADDRESS:
101 return StringRef("generic");
102 case AMDGPUAS::REGION_ADDRESS:
103 return StringRef("region");
104 default:
105 return std::nullopt;
106 }
107 }
108
109 StringRef
getValueKind(Type * Ty,StringRef TypeQual,StringRef BaseTypeName) const110 MetadataStreamerMsgPackV4::getValueKind(Type *Ty, StringRef TypeQual,
111 StringRef BaseTypeName) const {
112 if (TypeQual.contains("pipe"))
113 return "pipe";
114
115 return StringSwitch<StringRef>(BaseTypeName)
116 .Case("image1d_t", "image")
117 .Case("image1d_array_t", "image")
118 .Case("image1d_buffer_t", "image")
119 .Case("image2d_t", "image")
120 .Case("image2d_array_t", "image")
121 .Case("image2d_array_depth_t", "image")
122 .Case("image2d_array_msaa_t", "image")
123 .Case("image2d_array_msaa_depth_t", "image")
124 .Case("image2d_depth_t", "image")
125 .Case("image2d_msaa_t", "image")
126 .Case("image2d_msaa_depth_t", "image")
127 .Case("image3d_t", "image")
128 .Case("sampler_t", "sampler")
129 .Case("queue_t", "queue")
130 .Default(isa<PointerType>(Ty)
131 ? (Ty->getPointerAddressSpace() == AMDGPUAS::LOCAL_ADDRESS
132 ? "dynamic_shared_pointer"
133 : "global_buffer")
134 : "by_value");
135 }
136
getTypeName(Type * Ty,bool Signed) const137 std::string MetadataStreamerMsgPackV4::getTypeName(Type *Ty,
138 bool Signed) const {
139 switch (Ty->getTypeID()) {
140 case Type::IntegerTyID: {
141 if (!Signed)
142 return (Twine('u') + getTypeName(Ty, true)).str();
143
144 auto BitWidth = Ty->getIntegerBitWidth();
145 switch (BitWidth) {
146 case 8:
147 return "char";
148 case 16:
149 return "short";
150 case 32:
151 return "int";
152 case 64:
153 return "long";
154 default:
155 return (Twine('i') + Twine(BitWidth)).str();
156 }
157 }
158 case Type::HalfTyID:
159 return "half";
160 case Type::FloatTyID:
161 return "float";
162 case Type::DoubleTyID:
163 return "double";
164 case Type::FixedVectorTyID: {
165 auto VecTy = cast<FixedVectorType>(Ty);
166 auto ElTy = VecTy->getElementType();
167 auto NumElements = VecTy->getNumElements();
168 return (Twine(getTypeName(ElTy, Signed)) + Twine(NumElements)).str();
169 }
170 default:
171 return "unknown";
172 }
173 }
174
175 msgpack::ArrayDocNode
getWorkGroupDimensions(MDNode * Node) const176 MetadataStreamerMsgPackV4::getWorkGroupDimensions(MDNode *Node) const {
177 auto Dims = HSAMetadataDoc->getArrayNode();
178 if (Node->getNumOperands() != 3)
179 return Dims;
180
181 for (auto &Op : Node->operands())
182 Dims.push_back(Dims.getDocument()->getNode(
183 uint64_t(mdconst::extract<ConstantInt>(Op)->getZExtValue())));
184 return Dims;
185 }
186
emitVersion()187 void MetadataStreamerMsgPackV4::emitVersion() {
188 auto Version = HSAMetadataDoc->getArrayNode();
189 Version.push_back(Version.getDocument()->getNode(VersionMajorV4));
190 Version.push_back(Version.getDocument()->getNode(VersionMinorV4));
191 getRootMetadata("amdhsa.version") = Version;
192 }
193
emitTargetID(const IsaInfo::AMDGPUTargetID & TargetID)194 void MetadataStreamerMsgPackV4::emitTargetID(
195 const IsaInfo::AMDGPUTargetID &TargetID) {
196 getRootMetadata("amdhsa.target") =
197 HSAMetadataDoc->getNode(TargetID.toString(), /*Copy=*/true);
198 }
199
emitPrintf(const Module & Mod)200 void MetadataStreamerMsgPackV4::emitPrintf(const Module &Mod) {
201 auto Node = Mod.getNamedMetadata("llvm.printf.fmts");
202 if (!Node)
203 return;
204
205 auto Printf = HSAMetadataDoc->getArrayNode();
206 for (auto *Op : Node->operands())
207 if (Op->getNumOperands())
208 Printf.push_back(Printf.getDocument()->getNode(
209 cast<MDString>(Op->getOperand(0))->getString(), /*Copy=*/true));
210 getRootMetadata("amdhsa.printf") = Printf;
211 }
212
emitKernelLanguage(const Function & Func,msgpack::MapDocNode Kern)213 void MetadataStreamerMsgPackV4::emitKernelLanguage(const Function &Func,
214 msgpack::MapDocNode Kern) {
215 // TODO: What about other languages?
216 auto Node = Func.getParent()->getNamedMetadata("opencl.ocl.version");
217 if (!Node || !Node->getNumOperands())
218 return;
219 auto Op0 = Node->getOperand(0);
220 if (Op0->getNumOperands() <= 1)
221 return;
222
223 Kern[".language"] = Kern.getDocument()->getNode("OpenCL C");
224 auto LanguageVersion = Kern.getDocument()->getArrayNode();
225 LanguageVersion.push_back(Kern.getDocument()->getNode(
226 mdconst::extract<ConstantInt>(Op0->getOperand(0))->getZExtValue()));
227 LanguageVersion.push_back(Kern.getDocument()->getNode(
228 mdconst::extract<ConstantInt>(Op0->getOperand(1))->getZExtValue()));
229 Kern[".language_version"] = LanguageVersion;
230 }
231
emitKernelAttrs(const Function & Func,msgpack::MapDocNode Kern)232 void MetadataStreamerMsgPackV4::emitKernelAttrs(const Function &Func,
233 msgpack::MapDocNode Kern) {
234
235 if (auto Node = Func.getMetadata("reqd_work_group_size"))
236 Kern[".reqd_workgroup_size"] = getWorkGroupDimensions(Node);
237 if (auto Node = Func.getMetadata("work_group_size_hint"))
238 Kern[".workgroup_size_hint"] = getWorkGroupDimensions(Node);
239 if (auto Node = Func.getMetadata("vec_type_hint")) {
240 Kern[".vec_type_hint"] = Kern.getDocument()->getNode(
241 getTypeName(
242 cast<ValueAsMetadata>(Node->getOperand(0))->getType(),
243 mdconst::extract<ConstantInt>(Node->getOperand(1))->getZExtValue()),
244 /*Copy=*/true);
245 }
246 if (Func.hasFnAttribute("runtime-handle")) {
247 Kern[".device_enqueue_symbol"] = Kern.getDocument()->getNode(
248 Func.getFnAttribute("runtime-handle").getValueAsString().str(),
249 /*Copy=*/true);
250 }
251 if (Func.hasFnAttribute("device-init"))
252 Kern[".kind"] = Kern.getDocument()->getNode("init");
253 else if (Func.hasFnAttribute("device-fini"))
254 Kern[".kind"] = Kern.getDocument()->getNode("fini");
255 }
256
emitKernelArgs(const MachineFunction & MF,msgpack::MapDocNode Kern)257 void MetadataStreamerMsgPackV4::emitKernelArgs(const MachineFunction &MF,
258 msgpack::MapDocNode Kern) {
259 auto &Func = MF.getFunction();
260 unsigned Offset = 0;
261 auto Args = HSAMetadataDoc->getArrayNode();
262 for (auto &Arg : Func.args())
263 emitKernelArg(Arg, Offset, Args);
264
265 emitHiddenKernelArgs(MF, Offset, Args);
266
267 Kern[".args"] = Args;
268 }
269
emitKernelArg(const Argument & Arg,unsigned & Offset,msgpack::ArrayDocNode Args)270 void MetadataStreamerMsgPackV4::emitKernelArg(const Argument &Arg,
271 unsigned &Offset,
272 msgpack::ArrayDocNode Args) {
273 auto Func = Arg.getParent();
274 auto ArgNo = Arg.getArgNo();
275 const MDNode *Node;
276
277 StringRef Name;
278 Node = Func->getMetadata("kernel_arg_name");
279 if (Node && ArgNo < Node->getNumOperands())
280 Name = cast<MDString>(Node->getOperand(ArgNo))->getString();
281 else if (Arg.hasName())
282 Name = Arg.getName();
283
284 StringRef TypeName;
285 Node = Func->getMetadata("kernel_arg_type");
286 if (Node && ArgNo < Node->getNumOperands())
287 TypeName = cast<MDString>(Node->getOperand(ArgNo))->getString();
288
289 StringRef BaseTypeName;
290 Node = Func->getMetadata("kernel_arg_base_type");
291 if (Node && ArgNo < Node->getNumOperands())
292 BaseTypeName = cast<MDString>(Node->getOperand(ArgNo))->getString();
293
294 StringRef ActAccQual;
295 // Do we really need NoAlias check here?
296 if (Arg.getType()->isPointerTy() && Arg.hasNoAliasAttr()) {
297 if (Arg.onlyReadsMemory())
298 ActAccQual = "read_only";
299 else if (Arg.hasAttribute(Attribute::WriteOnly))
300 ActAccQual = "write_only";
301 }
302
303 StringRef AccQual;
304 Node = Func->getMetadata("kernel_arg_access_qual");
305 if (Node && ArgNo < Node->getNumOperands())
306 AccQual = cast<MDString>(Node->getOperand(ArgNo))->getString();
307
308 StringRef TypeQual;
309 Node = Func->getMetadata("kernel_arg_type_qual");
310 if (Node && ArgNo < Node->getNumOperands())
311 TypeQual = cast<MDString>(Node->getOperand(ArgNo))->getString();
312
313 const DataLayout &DL = Func->getParent()->getDataLayout();
314
315 MaybeAlign PointeeAlign;
316 Type *Ty = Arg.hasByRefAttr() ? Arg.getParamByRefType() : Arg.getType();
317
318 // FIXME: Need to distinguish in memory alignment from pointer alignment.
319 if (auto PtrTy = dyn_cast<PointerType>(Ty)) {
320 if (PtrTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS)
321 PointeeAlign = Arg.getParamAlign().valueOrOne();
322 }
323
324 // There's no distinction between byval aggregates and raw aggregates.
325 Type *ArgTy;
326 Align ArgAlign;
327 std::tie(ArgTy, ArgAlign) = getArgumentTypeAlign(Arg, DL);
328
329 emitKernelArg(DL, ArgTy, ArgAlign,
330 getValueKind(ArgTy, TypeQual, BaseTypeName), Offset, Args,
331 PointeeAlign, Name, TypeName, BaseTypeName, ActAccQual,
332 AccQual, TypeQual);
333 }
334
emitKernelArg(const DataLayout & DL,Type * Ty,Align Alignment,StringRef ValueKind,unsigned & Offset,msgpack::ArrayDocNode Args,MaybeAlign PointeeAlign,StringRef Name,StringRef TypeName,StringRef BaseTypeName,StringRef ActAccQual,StringRef AccQual,StringRef TypeQual)335 void MetadataStreamerMsgPackV4::emitKernelArg(
336 const DataLayout &DL, Type *Ty, Align Alignment, StringRef ValueKind,
337 unsigned &Offset, msgpack::ArrayDocNode Args, MaybeAlign PointeeAlign,
338 StringRef Name, StringRef TypeName, StringRef BaseTypeName,
339 StringRef ActAccQual, StringRef AccQual, StringRef TypeQual) {
340 auto Arg = Args.getDocument()->getMapNode();
341
342 if (!Name.empty())
343 Arg[".name"] = Arg.getDocument()->getNode(Name, /*Copy=*/true);
344 if (!TypeName.empty())
345 Arg[".type_name"] = Arg.getDocument()->getNode(TypeName, /*Copy=*/true);
346 auto Size = DL.getTypeAllocSize(Ty);
347 Arg[".size"] = Arg.getDocument()->getNode(Size);
348 Offset = alignTo(Offset, Alignment);
349 Arg[".offset"] = Arg.getDocument()->getNode(Offset);
350 Offset += Size;
351 Arg[".value_kind"] = Arg.getDocument()->getNode(ValueKind, /*Copy=*/true);
352 if (PointeeAlign)
353 Arg[".pointee_align"] = Arg.getDocument()->getNode(PointeeAlign->value());
354
355 if (auto PtrTy = dyn_cast<PointerType>(Ty))
356 if (auto Qualifier = getAddressSpaceQualifier(PtrTy->getAddressSpace()))
357 // Limiting address space to emit only for a certain ValueKind.
358 if (ValueKind == "global_buffer" || ValueKind == "dynamic_shared_pointer")
359 Arg[".address_space"] = Arg.getDocument()->getNode(*Qualifier,
360 /*Copy=*/true);
361
362 if (auto AQ = getAccessQualifier(AccQual))
363 Arg[".access"] = Arg.getDocument()->getNode(*AQ, /*Copy=*/true);
364
365 if (auto AAQ = getAccessQualifier(ActAccQual))
366 Arg[".actual_access"] = Arg.getDocument()->getNode(*AAQ, /*Copy=*/true);
367
368 SmallVector<StringRef, 1> SplitTypeQuals;
369 TypeQual.split(SplitTypeQuals, " ", -1, false);
370 for (StringRef Key : SplitTypeQuals) {
371 if (Key == "const")
372 Arg[".is_const"] = Arg.getDocument()->getNode(true);
373 else if (Key == "restrict")
374 Arg[".is_restrict"] = Arg.getDocument()->getNode(true);
375 else if (Key == "volatile")
376 Arg[".is_volatile"] = Arg.getDocument()->getNode(true);
377 else if (Key == "pipe")
378 Arg[".is_pipe"] = Arg.getDocument()->getNode(true);
379 }
380
381 Args.push_back(Arg);
382 }
383
emitHiddenKernelArgs(const MachineFunction & MF,unsigned & Offset,msgpack::ArrayDocNode Args)384 void MetadataStreamerMsgPackV4::emitHiddenKernelArgs(
385 const MachineFunction &MF, unsigned &Offset, msgpack::ArrayDocNode Args) {
386 auto &Func = MF.getFunction();
387 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
388
389 unsigned HiddenArgNumBytes = ST.getImplicitArgNumBytes(Func);
390 if (!HiddenArgNumBytes)
391 return;
392
393 const Module *M = Func.getParent();
394 auto &DL = M->getDataLayout();
395 auto Int64Ty = Type::getInt64Ty(Func.getContext());
396
397 Offset = alignTo(Offset, ST.getAlignmentForImplicitArgPtr());
398
399 if (HiddenArgNumBytes >= 8)
400 emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_x", Offset,
401 Args);
402 if (HiddenArgNumBytes >= 16)
403 emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_y", Offset,
404 Args);
405 if (HiddenArgNumBytes >= 24)
406 emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_z", Offset,
407 Args);
408
409 auto Int8PtrTy =
410 PointerType::get(Func.getContext(), AMDGPUAS::GLOBAL_ADDRESS);
411
412 if (HiddenArgNumBytes >= 32) {
413 // We forbid the use of features requiring hostcall when compiling OpenCL
414 // before code object V5, which makes the mutual exclusion between the
415 // "printf buffer" and "hostcall buffer" here sound.
416 if (M->getNamedMetadata("llvm.printf.fmts"))
417 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_printf_buffer", Offset,
418 Args);
419 else if (!Func.hasFnAttribute("amdgpu-no-hostcall-ptr"))
420 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_hostcall_buffer", Offset,
421 Args);
422 else
423 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_none", Offset, Args);
424 }
425
426 // Emit "default queue" and "completion action" arguments if enqueue kernel is
427 // used, otherwise emit dummy "none" arguments.
428 if (HiddenArgNumBytes >= 40) {
429 if (!Func.hasFnAttribute("amdgpu-no-default-queue")) {
430 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_default_queue", Offset,
431 Args);
432 } else {
433 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_none", Offset, Args);
434 }
435 }
436
437 if (HiddenArgNumBytes >= 48) {
438 if (!Func.hasFnAttribute("amdgpu-no-completion-action")) {
439 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_completion_action", Offset,
440 Args);
441 } else {
442 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_none", Offset, Args);
443 }
444 }
445
446 // Emit the pointer argument for multi-grid object.
447 if (HiddenArgNumBytes >= 56) {
448 if (!Func.hasFnAttribute("amdgpu-no-multigrid-sync-arg")) {
449 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_multigrid_sync_arg", Offset,
450 Args);
451 } else {
452 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_none", Offset, Args);
453 }
454 }
455 }
456
457 msgpack::MapDocNode
getHSAKernelProps(const MachineFunction & MF,const SIProgramInfo & ProgramInfo,unsigned CodeObjectVersion) const458 MetadataStreamerMsgPackV4::getHSAKernelProps(const MachineFunction &MF,
459 const SIProgramInfo &ProgramInfo,
460 unsigned CodeObjectVersion) const {
461 const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>();
462 const SIMachineFunctionInfo &MFI = *MF.getInfo<SIMachineFunctionInfo>();
463 const Function &F = MF.getFunction();
464
465 auto Kern = HSAMetadataDoc->getMapNode();
466
467 Align MaxKernArgAlign;
468 Kern[".kernarg_segment_size"] = Kern.getDocument()->getNode(
469 STM.getKernArgSegmentSize(F, MaxKernArgAlign));
470 Kern[".group_segment_fixed_size"] =
471 Kern.getDocument()->getNode(ProgramInfo.LDSSize);
472 Kern[".private_segment_fixed_size"] =
473 Kern.getDocument()->getNode(ProgramInfo.ScratchSize);
474 if (CodeObjectVersion >= AMDGPU::AMDHSA_COV5)
475 Kern[".uses_dynamic_stack"] =
476 Kern.getDocument()->getNode(ProgramInfo.DynamicCallStack);
477
478 if (CodeObjectVersion >= AMDGPU::AMDHSA_COV5 && STM.supportsWGP())
479 Kern[".workgroup_processor_mode"] =
480 Kern.getDocument()->getNode(ProgramInfo.WgpMode);
481
482 // FIXME: The metadata treats the minimum as 16?
483 Kern[".kernarg_segment_align"] =
484 Kern.getDocument()->getNode(std::max(Align(4), MaxKernArgAlign).value());
485 Kern[".wavefront_size"] =
486 Kern.getDocument()->getNode(STM.getWavefrontSize());
487 Kern[".sgpr_count"] = Kern.getDocument()->getNode(ProgramInfo.NumSGPR);
488 Kern[".vgpr_count"] = Kern.getDocument()->getNode(ProgramInfo.NumVGPR);
489
490 // Only add AGPR count to metadata for supported devices
491 if (STM.hasMAIInsts()) {
492 Kern[".agpr_count"] = Kern.getDocument()->getNode(ProgramInfo.NumAccVGPR);
493 }
494
495 Kern[".max_flat_workgroup_size"] =
496 Kern.getDocument()->getNode(MFI.getMaxFlatWorkGroupSize());
497 Kern[".sgpr_spill_count"] =
498 Kern.getDocument()->getNode(MFI.getNumSpilledSGPRs());
499 Kern[".vgpr_spill_count"] =
500 Kern.getDocument()->getNode(MFI.getNumSpilledVGPRs());
501
502 return Kern;
503 }
504
emitTo(AMDGPUTargetStreamer & TargetStreamer)505 bool MetadataStreamerMsgPackV4::emitTo(AMDGPUTargetStreamer &TargetStreamer) {
506 return TargetStreamer.EmitHSAMetadata(*HSAMetadataDoc, true);
507 }
508
begin(const Module & Mod,const IsaInfo::AMDGPUTargetID & TargetID)509 void MetadataStreamerMsgPackV4::begin(const Module &Mod,
510 const IsaInfo::AMDGPUTargetID &TargetID) {
511 emitVersion();
512 emitTargetID(TargetID);
513 emitPrintf(Mod);
514 getRootMetadata("amdhsa.kernels") = HSAMetadataDoc->getArrayNode();
515 }
516
end()517 void MetadataStreamerMsgPackV4::end() {
518 std::string HSAMetadataString;
519 raw_string_ostream StrOS(HSAMetadataString);
520 HSAMetadataDoc->toYAML(StrOS);
521
522 if (DumpHSAMetadata)
523 dump(StrOS.str());
524 if (VerifyHSAMetadata)
525 verify(StrOS.str());
526 }
527
emitKernel(const MachineFunction & MF,const SIProgramInfo & ProgramInfo)528 void MetadataStreamerMsgPackV4::emitKernel(const MachineFunction &MF,
529 const SIProgramInfo &ProgramInfo) {
530 auto &Func = MF.getFunction();
531 if (Func.getCallingConv() != CallingConv::AMDGPU_KERNEL &&
532 Func.getCallingConv() != CallingConv::SPIR_KERNEL)
533 return;
534
535 auto CodeObjectVersion =
536 AMDGPU::getAMDHSACodeObjectVersion(*Func.getParent());
537 auto Kern = getHSAKernelProps(MF, ProgramInfo, CodeObjectVersion);
538
539 auto Kernels =
540 getRootMetadata("amdhsa.kernels").getArray(/*Convert=*/true);
541
542 {
543 Kern[".name"] = Kern.getDocument()->getNode(Func.getName());
544 Kern[".symbol"] = Kern.getDocument()->getNode(
545 (Twine(Func.getName()) + Twine(".kd")).str(), /*Copy=*/true);
546 emitKernelLanguage(Func, Kern);
547 emitKernelAttrs(Func, Kern);
548 emitKernelArgs(MF, Kern);
549 }
550
551 Kernels.push_back(Kern);
552 }
553
554 //===----------------------------------------------------------------------===//
555 // HSAMetadataStreamerV5
556 //===----------------------------------------------------------------------===//
557
emitVersion()558 void MetadataStreamerMsgPackV5::emitVersion() {
559 auto Version = HSAMetadataDoc->getArrayNode();
560 Version.push_back(Version.getDocument()->getNode(VersionMajorV5));
561 Version.push_back(Version.getDocument()->getNode(VersionMinorV5));
562 getRootMetadata("amdhsa.version") = Version;
563 }
564
emitHiddenKernelArgs(const MachineFunction & MF,unsigned & Offset,msgpack::ArrayDocNode Args)565 void MetadataStreamerMsgPackV5::emitHiddenKernelArgs(
566 const MachineFunction &MF, unsigned &Offset, msgpack::ArrayDocNode Args) {
567 auto &Func = MF.getFunction();
568 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
569
570 // No implicit kernel argument is used.
571 if (ST.getImplicitArgNumBytes(Func) == 0)
572 return;
573
574 const Module *M = Func.getParent();
575 auto &DL = M->getDataLayout();
576 const SIMachineFunctionInfo &MFI = *MF.getInfo<SIMachineFunctionInfo>();
577
578 auto Int64Ty = Type::getInt64Ty(Func.getContext());
579 auto Int32Ty = Type::getInt32Ty(Func.getContext());
580 auto Int16Ty = Type::getInt16Ty(Func.getContext());
581
582 Offset = alignTo(Offset, ST.getAlignmentForImplicitArgPtr());
583 emitKernelArg(DL, Int32Ty, Align(4), "hidden_block_count_x", Offset, Args);
584 emitKernelArg(DL, Int32Ty, Align(4), "hidden_block_count_y", Offset, Args);
585 emitKernelArg(DL, Int32Ty, Align(4), "hidden_block_count_z", Offset, Args);
586
587 emitKernelArg(DL, Int16Ty, Align(2), "hidden_group_size_x", Offset, Args);
588 emitKernelArg(DL, Int16Ty, Align(2), "hidden_group_size_y", Offset, Args);
589 emitKernelArg(DL, Int16Ty, Align(2), "hidden_group_size_z", Offset, Args);
590
591 emitKernelArg(DL, Int16Ty, Align(2), "hidden_remainder_x", Offset, Args);
592 emitKernelArg(DL, Int16Ty, Align(2), "hidden_remainder_y", Offset, Args);
593 emitKernelArg(DL, Int16Ty, Align(2), "hidden_remainder_z", Offset, Args);
594
595 // Reserved for hidden_tool_correlation_id.
596 Offset += 8;
597
598 Offset += 8; // Reserved.
599
600 emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_x", Offset, Args);
601 emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_y", Offset, Args);
602 emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_z", Offset, Args);
603
604 emitKernelArg(DL, Int16Ty, Align(2), "hidden_grid_dims", Offset, Args);
605
606 Offset += 6; // Reserved.
607 auto Int8PtrTy =
608 PointerType::get(Func.getContext(), AMDGPUAS::GLOBAL_ADDRESS);
609
610 if (M->getNamedMetadata("llvm.printf.fmts")) {
611 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_printf_buffer", Offset,
612 Args);
613 } else {
614 Offset += 8; // Skipped.
615 }
616
617 if (!Func.hasFnAttribute("amdgpu-no-hostcall-ptr")) {
618 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_hostcall_buffer", Offset,
619 Args);
620 } else {
621 Offset += 8; // Skipped.
622 }
623
624 if (!Func.hasFnAttribute("amdgpu-no-multigrid-sync-arg")) {
625 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_multigrid_sync_arg", Offset,
626 Args);
627 } else {
628 Offset += 8; // Skipped.
629 }
630
631 if (!Func.hasFnAttribute("amdgpu-no-heap-ptr"))
632 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_heap_v1", Offset, Args);
633 else
634 Offset += 8; // Skipped.
635
636 if (!Func.hasFnAttribute("amdgpu-no-default-queue")) {
637 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_default_queue", Offset,
638 Args);
639 } else {
640 Offset += 8; // Skipped.
641 }
642
643 if (!Func.hasFnAttribute("amdgpu-no-completion-action")) {
644 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_completion_action", Offset,
645 Args);
646 } else {
647 Offset += 8; // Skipped.
648 }
649
650 // Emit argument for hidden dynamic lds size
651 if (MFI.isDynamicLDSUsed()) {
652 emitKernelArg(DL, Int32Ty, Align(4), "hidden_dynamic_lds_size", Offset,
653 Args);
654 } else {
655 Offset += 4; // skipped
656 }
657
658 Offset += 68; // Reserved.
659
660 // hidden_private_base and hidden_shared_base are only when the subtarget has
661 // ApertureRegs.
662 if (!ST.hasApertureRegs()) {
663 emitKernelArg(DL, Int32Ty, Align(4), "hidden_private_base", Offset, Args);
664 emitKernelArg(DL, Int32Ty, Align(4), "hidden_shared_base", Offset, Args);
665 } else {
666 Offset += 8; // Skipped.
667 }
668
669 if (MFI.getUserSGPRInfo().hasQueuePtr())
670 emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_queue_ptr", Offset, Args);
671 }
672
emitKernelAttrs(const Function & Func,msgpack::MapDocNode Kern)673 void MetadataStreamerMsgPackV5::emitKernelAttrs(const Function &Func,
674 msgpack::MapDocNode Kern) {
675 MetadataStreamerMsgPackV4::emitKernelAttrs(Func, Kern);
676
677 if (Func.getFnAttribute("uniform-work-group-size").getValueAsBool())
678 Kern[".uniform_work_group_size"] = Kern.getDocument()->getNode(1);
679 }
680
681
682 } // end namespace HSAMD
683 } // end namespace AMDGPU
684 } // end namespace llvm
685