[AMDGPU][AsmParser][NFC] Get rid of custom default operand handlers.
[llvm-project.git] / clang / lib / CodeGen / CGHLSLRuntime.cpp
blobe9fa273f21cc8d023a98aa933e47fdaba6f68236
1 //===----- CGHLSLRuntime.cpp - Interface to HLSL Runtimes -----------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This provides an abstract class for HLSL code generation. Concrete
10 // subclasses of this implement code generation for specific HLSL
11 // runtime libraries.
13 //===----------------------------------------------------------------------===//
15 #include "CGHLSLRuntime.h"
16 #include "CGDebugInfo.h"
17 #include "CodeGenModule.h"
18 #include "clang/AST/Decl.h"
19 #include "clang/Basic/TargetOptions.h"
20 #include "llvm/IR/IntrinsicsDirectX.h"
21 #include "llvm/IR/Metadata.h"
22 #include "llvm/IR/Module.h"
23 #include "llvm/Support/FormatVariadic.h"
25 using namespace clang;
26 using namespace CodeGen;
27 using namespace clang::hlsl;
28 using namespace llvm;
30 namespace {
32 void addDxilValVersion(StringRef ValVersionStr, llvm::Module &M) {
33 // The validation of ValVersionStr is done at HLSLToolChain::TranslateArgs.
34 // Assume ValVersionStr is legal here.
35 VersionTuple Version;
36 if (Version.tryParse(ValVersionStr) || Version.getBuild() ||
37 Version.getSubminor() || !Version.getMinor()) {
38 return;
41 uint64_t Major = Version.getMajor();
42 uint64_t Minor = *Version.getMinor();
44 auto &Ctx = M.getContext();
45 IRBuilder<> B(M.getContext());
46 MDNode *Val = MDNode::get(Ctx, {ConstantAsMetadata::get(B.getInt32(Major)),
47 ConstantAsMetadata::get(B.getInt32(Minor))});
48 StringRef DXILValKey = "dx.valver";
49 auto *DXILValMD = M.getOrInsertNamedMetadata(DXILValKey);
50 DXILValMD->addOperand(Val);
52 void addDisableOptimizations(llvm::Module &M) {
53 StringRef Key = "dx.disable_optimizations";
54 M.addModuleFlag(llvm::Module::ModFlagBehavior::Override, Key, 1);
56 // cbuffer will be translated into global variable in special address space.
57 // If translate into C,
58 // cbuffer A {
59 // float a;
60 // float b;
61 // }
62 // float foo() { return a + b; }
64 // will be translated into
66 // struct A {
67 // float a;
68 // float b;
69 // } cbuffer_A __attribute__((address_space(4)));
70 // float foo() { return cbuffer_A.a + cbuffer_A.b; }
72 // layoutBuffer will create the struct A type.
73 // replaceBuffer will replace use of global variable a and b with cbuffer_A.a
74 // and cbuffer_A.b.
76 void layoutBuffer(CGHLSLRuntime::Buffer &Buf, const DataLayout &DL) {
77 if (Buf.Constants.empty())
78 return;
80 std::vector<llvm::Type *> EltTys;
81 for (auto &Const : Buf.Constants) {
82 GlobalVariable *GV = Const.first;
83 Const.second = EltTys.size();
84 llvm::Type *Ty = GV->getValueType();
85 EltTys.emplace_back(Ty);
87 Buf.LayoutStruct = llvm::StructType::get(EltTys[0]->getContext(), EltTys);
90 GlobalVariable *replaceBuffer(CGHLSLRuntime::Buffer &Buf) {
91 // Create global variable for CB.
92 GlobalVariable *CBGV = new GlobalVariable(
93 Buf.LayoutStruct, /*isConstant*/ true,
94 GlobalValue::LinkageTypes::ExternalLinkage, nullptr,
95 llvm::formatv("{0}{1}", Buf.Name, Buf.IsCBuffer ? ".cb." : ".tb."),
96 GlobalValue::NotThreadLocal);
98 IRBuilder<> B(CBGV->getContext());
99 Value *ZeroIdx = B.getInt32(0);
100 // Replace Const use with CB use.
101 for (auto &[GV, Offset] : Buf.Constants) {
102 Value *GEP =
103 B.CreateGEP(Buf.LayoutStruct, CBGV, {ZeroIdx, B.getInt32(Offset)});
105 assert(Buf.LayoutStruct->getElementType(Offset) == GV->getValueType() &&
106 "constant type mismatch");
108 // Replace.
109 GV->replaceAllUsesWith(GEP);
110 // Erase GV.
111 GV->removeDeadConstantUsers();
112 GV->eraseFromParent();
114 return CBGV;
117 } // namespace
119 void CGHLSLRuntime::addConstant(VarDecl *D, Buffer &CB) {
120 if (D->getStorageClass() == SC_Static) {
121 // For static inside cbuffer, take as global static.
122 // Don't add to cbuffer.
123 CGM.EmitGlobal(D);
124 return;
127 auto *GV = cast<GlobalVariable>(CGM.GetAddrOfGlobalVar(D));
128 // Add debug info for constVal.
129 if (CGDebugInfo *DI = CGM.getModuleDebugInfo())
130 if (CGM.getCodeGenOpts().getDebugInfo() >=
131 codegenoptions::DebugInfoKind::LimitedDebugInfo)
132 DI->EmitGlobalVariable(cast<GlobalVariable>(GV), D);
134 // FIXME: support packoffset.
135 // See https://github.com/llvm/llvm-project/issues/57914.
136 uint32_t Offset = 0;
137 bool HasUserOffset = false;
139 unsigned LowerBound = HasUserOffset ? Offset : UINT_MAX;
140 CB.Constants.emplace_back(std::make_pair(GV, LowerBound));
143 void CGHLSLRuntime::addBufferDecls(const DeclContext *DC, Buffer &CB) {
144 for (Decl *it : DC->decls()) {
145 if (auto *ConstDecl = dyn_cast<VarDecl>(it)) {
146 addConstant(ConstDecl, CB);
147 } else if (isa<CXXRecordDecl, EmptyDecl>(it)) {
148 // Nothing to do for this declaration.
149 } else if (isa<FunctionDecl>(it)) {
150 // A function within an cbuffer is effectively a top-level function,
151 // as it only refers to globally scoped declarations.
152 CGM.EmitTopLevelDecl(it);
157 void CGHLSLRuntime::addBuffer(const HLSLBufferDecl *D) {
158 Buffers.emplace_back(Buffer(D));
159 addBufferDecls(D, Buffers.back());
162 void CGHLSLRuntime::finishCodeGen() {
163 auto &TargetOpts = CGM.getTarget().getTargetOpts();
164 llvm::Module &M = CGM.getModule();
165 Triple T(M.getTargetTriple());
166 if (T.getArch() == Triple::ArchType::dxil)
167 addDxilValVersion(TargetOpts.DxilValidatorVersion, M);
169 generateGlobalCtorDtorCalls();
170 if (CGM.getCodeGenOpts().OptimizationLevel == 0)
171 addDisableOptimizations(M);
173 const DataLayout &DL = M.getDataLayout();
175 for (auto &Buf : Buffers) {
176 layoutBuffer(Buf, DL);
177 GlobalVariable *GV = replaceBuffer(Buf);
178 M.insertGlobalVariable(GV);
179 llvm::hlsl::ResourceClass RC = Buf.IsCBuffer
180 ? llvm::hlsl::ResourceClass::CBuffer
181 : llvm::hlsl::ResourceClass::SRV;
182 llvm::hlsl::ResourceKind RK = Buf.IsCBuffer
183 ? llvm::hlsl::ResourceKind::CBuffer
184 : llvm::hlsl::ResourceKind::TBuffer;
185 std::string TyName =
186 Buf.Name.str() + (Buf.IsCBuffer ? ".cb." : ".tb.") + "ty";
187 addBufferResourceAnnotation(GV, TyName, RC, RK, Buf.Binding);
191 CGHLSLRuntime::Buffer::Buffer(const HLSLBufferDecl *D)
192 : Name(D->getName()), IsCBuffer(D->isCBuffer()),
193 Binding(D->getAttr<HLSLResourceBindingAttr>()) {}
195 void CGHLSLRuntime::addBufferResourceAnnotation(llvm::GlobalVariable *GV,
196 llvm::StringRef TyName,
197 llvm::hlsl::ResourceClass RC,
198 llvm::hlsl::ResourceKind RK,
199 BufferResBinding &Binding) {
200 llvm::Module &M = CGM.getModule();
202 NamedMDNode *ResourceMD = nullptr;
203 switch (RC) {
204 case llvm::hlsl::ResourceClass::UAV:
205 ResourceMD = M.getOrInsertNamedMetadata("hlsl.uavs");
206 break;
207 case llvm::hlsl::ResourceClass::SRV:
208 ResourceMD = M.getOrInsertNamedMetadata("hlsl.srvs");
209 break;
210 case llvm::hlsl::ResourceClass::CBuffer:
211 ResourceMD = M.getOrInsertNamedMetadata("hlsl.cbufs");
212 break;
213 default:
214 assert(false && "Unsupported buffer type!");
215 return;
218 assert(ResourceMD != nullptr &&
219 "ResourceMD must have been set by the switch above.");
221 llvm::hlsl::FrontendResource Res(
222 GV, TyName, RK, Binding.Reg.value_or(UINT_MAX), Binding.Space);
223 ResourceMD->addOperand(Res.getMetadata());
226 static llvm::hlsl::ResourceKind
227 castResourceShapeToResourceKind(HLSLResourceAttr::ResourceKind RK) {
228 switch (RK) {
229 case HLSLResourceAttr::ResourceKind::Texture1D:
230 return llvm::hlsl::ResourceKind::Texture1D;
231 case HLSLResourceAttr::ResourceKind::Texture2D:
232 return llvm::hlsl::ResourceKind::Texture2D;
233 case HLSLResourceAttr::ResourceKind::Texture2DMS:
234 return llvm::hlsl::ResourceKind::Texture2DMS;
235 case HLSLResourceAttr::ResourceKind::Texture3D:
236 return llvm::hlsl::ResourceKind::Texture3D;
237 case HLSLResourceAttr::ResourceKind::TextureCube:
238 return llvm::hlsl::ResourceKind::TextureCube;
239 case HLSLResourceAttr::ResourceKind::Texture1DArray:
240 return llvm::hlsl::ResourceKind::Texture1DArray;
241 case HLSLResourceAttr::ResourceKind::Texture2DArray:
242 return llvm::hlsl::ResourceKind::Texture2DArray;
243 case HLSLResourceAttr::ResourceKind::Texture2DMSArray:
244 return llvm::hlsl::ResourceKind::Texture2DMSArray;
245 case HLSLResourceAttr::ResourceKind::TextureCubeArray:
246 return llvm::hlsl::ResourceKind::TextureCubeArray;
247 case HLSLResourceAttr::ResourceKind::TypedBuffer:
248 return llvm::hlsl::ResourceKind::TypedBuffer;
249 case HLSLResourceAttr::ResourceKind::RawBuffer:
250 return llvm::hlsl::ResourceKind::RawBuffer;
251 case HLSLResourceAttr::ResourceKind::StructuredBuffer:
252 return llvm::hlsl::ResourceKind::StructuredBuffer;
253 case HLSLResourceAttr::ResourceKind::CBufferKind:
254 return llvm::hlsl::ResourceKind::CBuffer;
255 case HLSLResourceAttr::ResourceKind::SamplerKind:
256 return llvm::hlsl::ResourceKind::Sampler;
257 case HLSLResourceAttr::ResourceKind::TBuffer:
258 return llvm::hlsl::ResourceKind::TBuffer;
259 case HLSLResourceAttr::ResourceKind::RTAccelerationStructure:
260 return llvm::hlsl::ResourceKind::RTAccelerationStructure;
261 case HLSLResourceAttr::ResourceKind::FeedbackTexture2D:
262 return llvm::hlsl::ResourceKind::FeedbackTexture2D;
263 case HLSLResourceAttr::ResourceKind::FeedbackTexture2DArray:
264 return llvm::hlsl::ResourceKind::FeedbackTexture2DArray;
266 // Make sure to update HLSLResourceAttr::ResourceKind when add new Kind to
267 // hlsl::ResourceKind. Assume FeedbackTexture2DArray is the last enum for
268 // HLSLResourceAttr::ResourceKind.
269 static_assert(
270 static_cast<uint32_t>(
271 HLSLResourceAttr::ResourceKind::FeedbackTexture2DArray) ==
272 (static_cast<uint32_t>(llvm::hlsl::ResourceKind::NumEntries) - 2));
273 llvm_unreachable("all switch cases should be covered");
276 void CGHLSLRuntime::annotateHLSLResource(const VarDecl *D, GlobalVariable *GV) {
277 const Type *Ty = D->getType()->getPointeeOrArrayElementType();
278 if (!Ty)
279 return;
280 const auto *RD = Ty->getAsCXXRecordDecl();
281 if (!RD)
282 return;
283 const auto *Attr = RD->getAttr<HLSLResourceAttr>();
284 if (!Attr)
285 return;
287 HLSLResourceAttr::ResourceClass RC = Attr->getResourceType();
288 llvm::hlsl::ResourceKind RK =
289 castResourceShapeToResourceKind(Attr->getResourceShape());
291 QualType QT(Ty, 0);
292 BufferResBinding Binding(D->getAttr<HLSLResourceBindingAttr>());
293 addBufferResourceAnnotation(GV, QT.getAsString(),
294 static_cast<llvm::hlsl::ResourceClass>(RC), RK,
295 Binding);
298 CGHLSLRuntime::BufferResBinding::BufferResBinding(
299 HLSLResourceBindingAttr *Binding) {
300 if (Binding) {
301 llvm::APInt RegInt(64, 0);
302 Binding->getSlot().substr(1).getAsInteger(10, RegInt);
303 Reg = RegInt.getLimitedValue();
304 llvm::APInt SpaceInt(64, 0);
305 Binding->getSpace().substr(5).getAsInteger(10, SpaceInt);
306 Space = SpaceInt.getLimitedValue();
307 } else {
308 Space = 0;
312 void clang::CodeGen::CGHLSLRuntime::setHLSLEntryAttributes(
313 const FunctionDecl *FD, llvm::Function *Fn) {
314 const auto *ShaderAttr = FD->getAttr<HLSLShaderAttr>();
315 assert(ShaderAttr && "All entry functions must have a HLSLShaderAttr");
316 const StringRef ShaderAttrKindStr = "hlsl.shader";
317 Fn->addFnAttr(ShaderAttrKindStr,
318 ShaderAttr->ConvertShaderTypeToStr(ShaderAttr->getType()));
319 if (HLSLNumThreadsAttr *NumThreadsAttr = FD->getAttr<HLSLNumThreadsAttr>()) {
320 const StringRef NumThreadsKindStr = "hlsl.numthreads";
321 std::string NumThreadsStr =
322 formatv("{0},{1},{2}", NumThreadsAttr->getX(), NumThreadsAttr->getY(),
323 NumThreadsAttr->getZ());
324 Fn->addFnAttr(NumThreadsKindStr, NumThreadsStr);
328 static Value *buildVectorInput(IRBuilder<> &B, Function *F, llvm::Type *Ty) {
329 if (const auto *VT = dyn_cast<FixedVectorType>(Ty)) {
330 Value *Result = PoisonValue::get(Ty);
331 for (unsigned I = 0; I < VT->getNumElements(); ++I) {
332 Value *Elt = B.CreateCall(F, {B.getInt32(I)});
333 Result = B.CreateInsertElement(Result, Elt, I);
335 return Result;
337 return B.CreateCall(F, {B.getInt32(0)});
340 llvm::Value *CGHLSLRuntime::emitInputSemantic(IRBuilder<> &B,
341 const ParmVarDecl &D,
342 llvm::Type *Ty) {
343 assert(D.hasAttrs() && "Entry parameter missing annotation attribute!");
344 if (D.hasAttr<HLSLSV_GroupIndexAttr>()) {
345 llvm::Function *DxGroupIndex =
346 CGM.getIntrinsic(Intrinsic::dx_flattened_thread_id_in_group);
347 return B.CreateCall(FunctionCallee(DxGroupIndex));
349 if (D.hasAttr<HLSLSV_DispatchThreadIDAttr>()) {
350 llvm::Function *DxThreadID = CGM.getIntrinsic(Intrinsic::dx_thread_id);
351 return buildVectorInput(B, DxThreadID, Ty);
353 assert(false && "Unhandled parameter attribute");
354 return nullptr;
357 void CGHLSLRuntime::emitEntryFunction(const FunctionDecl *FD,
358 llvm::Function *Fn) {
359 llvm::Module &M = CGM.getModule();
360 llvm::LLVMContext &Ctx = M.getContext();
361 auto *EntryTy = llvm::FunctionType::get(llvm::Type::getVoidTy(Ctx), false);
362 Function *EntryFn =
363 Function::Create(EntryTy, Function::ExternalLinkage, FD->getName(), &M);
365 // Copy function attributes over, we have no argument or return attributes
366 // that can be valid on the real entry.
367 AttributeList NewAttrs = AttributeList::get(Ctx, AttributeList::FunctionIndex,
368 Fn->getAttributes().getFnAttrs());
369 EntryFn->setAttributes(NewAttrs);
370 setHLSLEntryAttributes(FD, EntryFn);
372 // Set the called function as internal linkage.
373 Fn->setLinkage(GlobalValue::InternalLinkage);
375 BasicBlock *BB = BasicBlock::Create(Ctx, "entry", EntryFn);
376 IRBuilder<> B(BB);
377 llvm::SmallVector<Value *> Args;
378 // FIXME: support struct parameters where semantics are on members.
379 // See: https://github.com/llvm/llvm-project/issues/57874
380 unsigned SRetOffset = 0;
381 for (const auto &Param : Fn->args()) {
382 if (Param.hasStructRetAttr()) {
383 // FIXME: support output.
384 // See: https://github.com/llvm/llvm-project/issues/57874
385 SRetOffset = 1;
386 Args.emplace_back(PoisonValue::get(Param.getType()));
387 continue;
389 const ParmVarDecl *PD = FD->getParamDecl(Param.getArgNo() - SRetOffset);
390 Args.push_back(emitInputSemantic(B, *PD, Param.getType()));
393 CallInst *CI = B.CreateCall(FunctionCallee(Fn), Args);
394 (void)CI;
395 // FIXME: Handle codegen for return type semantics.
396 // See: https://github.com/llvm/llvm-project/issues/57875
397 B.CreateRetVoid();
400 static void gatherFunctions(SmallVectorImpl<Function *> &Fns, llvm::Module &M,
401 bool CtorOrDtor) {
402 const auto *GV =
403 M.getNamedGlobal(CtorOrDtor ? "llvm.global_ctors" : "llvm.global_dtors");
404 if (!GV)
405 return;
406 const auto *CA = dyn_cast<ConstantArray>(GV->getInitializer());
407 if (!CA)
408 return;
409 // The global_ctor array elements are a struct [Priority, Fn *, COMDat].
410 // HLSL neither supports priorities or COMDat values, so we will check those
411 // in an assert but not handle them.
413 llvm::SmallVector<Function *> CtorFns;
414 for (const auto &Ctor : CA->operands()) {
415 if (isa<ConstantAggregateZero>(Ctor))
416 continue;
417 ConstantStruct *CS = cast<ConstantStruct>(Ctor);
419 assert(cast<ConstantInt>(CS->getOperand(0))->getValue() == 65535 &&
420 "HLSL doesn't support setting priority for global ctors.");
421 assert(isa<ConstantPointerNull>(CS->getOperand(2)) &&
422 "HLSL doesn't support COMDat for global ctors.");
423 Fns.push_back(cast<Function>(CS->getOperand(1)));
427 void CGHLSLRuntime::generateGlobalCtorDtorCalls() {
428 llvm::Module &M = CGM.getModule();
429 SmallVector<Function *> CtorFns;
430 SmallVector<Function *> DtorFns;
431 gatherFunctions(CtorFns, M, true);
432 gatherFunctions(DtorFns, M, false);
434 // Insert a call to the global constructor at the beginning of the entry block
435 // to externally exported functions. This is a bit of a hack, but HLSL allows
436 // global constructors, but doesn't support driver initialization of globals.
437 for (auto &F : M.functions()) {
438 if (!F.hasFnAttribute("hlsl.shader"))
439 continue;
440 IRBuilder<> B(&F.getEntryBlock(), F.getEntryBlock().begin());
441 for (auto *Fn : CtorFns)
442 B.CreateCall(FunctionCallee(Fn));
444 // Insert global dtors before the terminator of the last instruction
445 B.SetInsertPoint(F.back().getTerminator());
446 for (auto *Fn : DtorFns)
447 B.CreateCall(FunctionCallee(Fn));
450 // No need to keep global ctors/dtors for non-lib profile after call to
451 // ctors/dtors added for entry.
452 Triple T(M.getTargetTriple());
453 if (T.getEnvironment() != Triple::EnvironmentType::Library) {
454 if (auto *GV = M.getNamedGlobal("llvm.global_ctors"))
455 GV->eraseFromParent();
456 if (auto *GV = M.getNamedGlobal("llvm.global_dtors"))
457 GV->eraseFromParent();