1//===----- CGHLSLRuntime.h - Interface to HLSL Runtimes -----*- 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// This provides an abstract class for HLSL code generation. Concrete
10// subclasses of this implement code generation for specific HLSL
11// runtime libraries.
12//
13//===----------------------------------------------------------------------===//
14
15#ifndef LLVM_CLANG_LIB_CODEGEN_CGHLSLRUNTIME_H
16#define LLVM_CLANG_LIB_CODEGEN_CGHLSLRUNTIME_H
17
18#include "Address.h"
19#include "clang/AST/Attr.h"
20#include "clang/AST/Decl.h"
21#include "clang/Basic/Builtins.h"
22#include "clang/Basic/HLSLRuntime.h"
23#include "llvm/ADT/DenseMap.h"
24#include "llvm/ADT/SmallVector.h"
25#include "llvm/ADT/StringRef.h"
26#include "llvm/Frontend/HLSL/HLSLResource.h"
27#include "llvm/IR/IRBuilder.h"
28#include "llvm/IR/Intrinsics.h"
29#include "llvm/IR/IntrinsicsDirectX.h"
30#include "llvm/IR/IntrinsicsSPIRV.h"
31
32#include <optional>
33#include <vector>
34
35// A function generator macro for picking the right intrinsic
36// for the target backend
37#define GENERATE_HLSL_INTRINSIC_FUNCTION(FunctionName, IntrinsicPostfix) \
38 llvm::Intrinsic::ID get##FunctionName##Intrinsic() { \
39 llvm::Triple::ArchType Arch = getArch(); \
40 switch (Arch) { \
41 case llvm::Triple::dxil: \
42 return llvm::Intrinsic::dx_##IntrinsicPostfix; \
43 case llvm::Triple::spirv: \
44 return llvm::Intrinsic::spv_##IntrinsicPostfix; \
45 default: \
46 llvm_unreachable("Intrinsic " #IntrinsicPostfix \
47 " not supported by target architecture"); \
48 } \
49 }
50
51using ResourceClass = llvm::dxil::ResourceClass;
52
53namespace llvm {
54class GlobalVariable;
55class Function;
56class StructType;
57class Metadata;
58} // namespace llvm
59
60namespace clang {
61class NamedDecl;
62class VarDecl;
63class ParmVarDecl;
64class InitListExpr;
65class HLSLBufferDecl;
66class HLSLRootSignatureDecl;
67class HLSLVkBindingAttr;
68class HLSLResourceBindingAttr;
69class Type;
70class RecordType;
71class DeclContext;
72class HLSLPackOffsetAttr;
73class ArraySubscriptExpr;
74
75class FunctionDecl;
76
77namespace CodeGen {
78
79class CodeGenModule;
80class CodeGenFunction;
81class LValue;
82
83class CGHLSLOffsetInfo {
84 SmallVector<uint32_t> Offsets;
85
86public:
87 static const uint32_t Unspecified = ~0U;
88
89 /// Iterates over all declarations in the HLSL buffer and based on the
90 /// packoffset or register(c#) annotations it fills outs the Offsets vector
91 /// with the user-specified layout offsets. The buffer offsets can be
92 /// specified 2 ways: 1. declarations in cbuffer {} block can have a
93 /// packoffset annotation (translates to HLSLPackOffsetAttr) 2. default
94 /// constant buffer declarations at global scope can have register(c#)
95 /// annotations (translates to HLSLResourceBindingAttr with RegisterType::C)
96 /// It is not guaranteed that all declarations in a buffer have an annotation.
97 /// For those where it is not specified a `~0U` value is added to the Offsets
98 /// vector. In the final layout these declarations will be placed at the end
99 /// of the HLSL buffer after all of the elements with specified offset.
100 static CGHLSLOffsetInfo fromDecl(const HLSLBufferDecl &BufDecl);
101
102 /// Comparison function for offsets received from `operator[]` suitable for
103 /// use in a `stable_sort`. This will order implicit bindings after explicit
104 /// offsets.
105 static bool compareOffsets(uint32_t LHS, uint32_t RHS) { return LHS < RHS; }
106
107 /// Get the given offset, or `~0U` if there is no offset for the member.
108 uint32_t operator[](size_t I) const {
109 if (Offsets.empty())
110 return Unspecified;
111 return Offsets[I];
112 }
113
114 bool empty() const { return Offsets.empty(); }
115};
116
117class CGHLSLRuntime {
118public:
119 //===----------------------------------------------------------------------===//
120 // Start of reserved area for HLSL intrinsic getters.
121 //===----------------------------------------------------------------------===//
122
123 GENERATE_HLSL_INTRINSIC_FUNCTION(All, all)
124 GENERATE_HLSL_INTRINSIC_FUNCTION(Any, any)
125 GENERATE_HLSL_INTRINSIC_FUNCTION(Cross, cross)
126 GENERATE_HLSL_INTRINSIC_FUNCTION(Degrees, degrees)
127 GENERATE_HLSL_INTRINSIC_FUNCTION(Frac, frac)
128 GENERATE_HLSL_INTRINSIC_FUNCTION(FlattenedThreadIdInGroup,
129 flattened_thread_id_in_group)
130 GENERATE_HLSL_INTRINSIC_FUNCTION(IsInf, isinf)
131 GENERATE_HLSL_INTRINSIC_FUNCTION(IsNaN, isnan)
132 GENERATE_HLSL_INTRINSIC_FUNCTION(Lerp, lerp)
133 GENERATE_HLSL_INTRINSIC_FUNCTION(Normalize, normalize)
134 GENERATE_HLSL_INTRINSIC_FUNCTION(Rsqrt, rsqrt)
135 GENERATE_HLSL_INTRINSIC_FUNCTION(Saturate, saturate)
136 GENERATE_HLSL_INTRINSIC_FUNCTION(Sign, sign)
137 GENERATE_HLSL_INTRINSIC_FUNCTION(Step, step)
138 GENERATE_HLSL_INTRINSIC_FUNCTION(Radians, radians)
139 GENERATE_HLSL_INTRINSIC_FUNCTION(ThreadId, thread_id)
140 GENERATE_HLSL_INTRINSIC_FUNCTION(GroupThreadId, thread_id_in_group)
141 GENERATE_HLSL_INTRINSIC_FUNCTION(GroupId, group_id)
142 GENERATE_HLSL_INTRINSIC_FUNCTION(FDot, fdot)
143 GENERATE_HLSL_INTRINSIC_FUNCTION(SDot, sdot)
144 GENERATE_HLSL_INTRINSIC_FUNCTION(UDot, udot)
145 GENERATE_HLSL_INTRINSIC_FUNCTION(Dot4AddI8Packed, dot4add_i8packed)
146 GENERATE_HLSL_INTRINSIC_FUNCTION(Dot4AddU8Packed, dot4add_u8packed)
147 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveAllTrue, wave_all)
148 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveAnyTrue, wave_any)
149 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveMax, wave_reduce_max)
150 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveUMax, wave_reduce_umax)
151 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveMin, wave_reduce_min)
152 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveUMin, wave_reduce_umin)
153 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveCountBits, wave_active_countbits)
154 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveIsFirstLane, wave_is_first_lane)
155 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveGetLaneCount, wave_get_lane_count)
156 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveReadLaneAt, wave_readlane)
157 GENERATE_HLSL_INTRINSIC_FUNCTION(FirstBitUHigh, firstbituhigh)
158 GENERATE_HLSL_INTRINSIC_FUNCTION(FirstBitSHigh, firstbitshigh)
159 GENERATE_HLSL_INTRINSIC_FUNCTION(FirstBitLow, firstbitlow)
160 GENERATE_HLSL_INTRINSIC_FUNCTION(NClamp, nclamp)
161 GENERATE_HLSL_INTRINSIC_FUNCTION(SClamp, sclamp)
162 GENERATE_HLSL_INTRINSIC_FUNCTION(UClamp, uclamp)
163
164 GENERATE_HLSL_INTRINSIC_FUNCTION(CreateResourceGetPointer,
165 resource_getpointer)
166 GENERATE_HLSL_INTRINSIC_FUNCTION(Sample, resource_sample)
167 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleClamp, resource_sample_clamp)
168 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleBias, resource_samplebias)
169 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleBiasClamp, resource_samplebias_clamp)
170 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleGrad, resource_samplegrad)
171 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleGradClamp, resource_samplegrad_clamp)
172 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleLevel, resource_samplelevel)
173 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleCmp, resource_samplecmp)
174 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleCmpClamp, resource_samplecmp_clamp)
175 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleCmpLevelZero,
176 resource_samplecmplevelzero)
177 GENERATE_HLSL_INTRINSIC_FUNCTION(CreateHandleFromBinding,
178 resource_handlefrombinding)
179 GENERATE_HLSL_INTRINSIC_FUNCTION(CreateHandleFromImplicitBinding,
180 resource_handlefromimplicitbinding)
181 GENERATE_HLSL_INTRINSIC_FUNCTION(NonUniformResourceIndex,
182 resource_nonuniformindex)
183 GENERATE_HLSL_INTRINSIC_FUNCTION(BufferUpdateCounter, resource_updatecounter)
184 GENERATE_HLSL_INTRINSIC_FUNCTION(GroupMemoryBarrierWithGroupSync,
185 group_memory_barrier_with_group_sync)
186 GENERATE_HLSL_INTRINSIC_FUNCTION(GetDimensionsX, resource_getdimensions_x)
187 GENERATE_HLSL_INTRINSIC_FUNCTION(DdxCoarse, ddx_coarse)
188 GENERATE_HLSL_INTRINSIC_FUNCTION(DdyCoarse, ddy_coarse)
189 GENERATE_HLSL_INTRINSIC_FUNCTION(DdxFine, ddx_fine)
190 GENERATE_HLSL_INTRINSIC_FUNCTION(DdyFine, ddy_fine)
191
192 //===----------------------------------------------------------------------===//
193 // End of reserved area for HLSL intrinsic getters.
194 //===----------------------------------------------------------------------===//
195
196protected:
197 CodeGenModule &CGM;
198
199 llvm::Value *emitSystemSemanticLoad(llvm::IRBuilder<> &B,
200 const FunctionDecl *FD, llvm::Type *Type,
201 const clang::DeclaratorDecl *Decl,
202 HLSLAppliedSemanticAttr *Semantic,
203 std::optional<unsigned> Index);
204
205 void emitSystemSemanticStore(llvm::IRBuilder<> &B, llvm::Value *Source,
206 const clang::DeclaratorDecl *Decl,
207 HLSLAppliedSemanticAttr *Semantic,
208 std::optional<unsigned> Index);
209
210 llvm::Value *handleScalarSemanticLoad(llvm::IRBuilder<> &B,
211 const FunctionDecl *FD,
212 llvm::Type *Type,
213 const clang::DeclaratorDecl *Decl,
214 HLSLAppliedSemanticAttr *Semantic);
215
216 void handleScalarSemanticStore(llvm::IRBuilder<> &B, const FunctionDecl *FD,
217 llvm::Value *Source,
218 const clang::DeclaratorDecl *Decl,
219 HLSLAppliedSemanticAttr *Semantic);
220
221 std::pair<llvm::Value *, specific_attr_iterator<HLSLAppliedSemanticAttr>>
222 handleStructSemanticLoad(
223 llvm::IRBuilder<> &B, const FunctionDecl *FD, llvm::Type *Type,
224 const clang::DeclaratorDecl *Decl,
225 specific_attr_iterator<HLSLAppliedSemanticAttr> begin,
226 specific_attr_iterator<HLSLAppliedSemanticAttr> end);
227
228 specific_attr_iterator<HLSLAppliedSemanticAttr> handleStructSemanticStore(
229 llvm::IRBuilder<> &B, const FunctionDecl *FD, llvm::Value *Source,
230 const clang::DeclaratorDecl *Decl,
231 specific_attr_iterator<HLSLAppliedSemanticAttr> AttrBegin,
232 specific_attr_iterator<HLSLAppliedSemanticAttr> AttrEnd);
233
234 std::pair<llvm::Value *, specific_attr_iterator<HLSLAppliedSemanticAttr>>
235 handleSemanticLoad(llvm::IRBuilder<> &B, const FunctionDecl *FD,
236 llvm::Type *Type, const clang::DeclaratorDecl *Decl,
237 specific_attr_iterator<HLSLAppliedSemanticAttr> begin,
238 specific_attr_iterator<HLSLAppliedSemanticAttr> end);
239
240 specific_attr_iterator<HLSLAppliedSemanticAttr>
241 handleSemanticStore(llvm::IRBuilder<> &B, const FunctionDecl *FD,
242 llvm::Value *Source, const clang::DeclaratorDecl *Decl,
243 specific_attr_iterator<HLSLAppliedSemanticAttr> AttrBegin,
244 specific_attr_iterator<HLSLAppliedSemanticAttr> AttrEnd);
245
246public:
247 CGHLSLRuntime(CodeGenModule &CGM) : CGM(CGM) {}
248 virtual ~CGHLSLRuntime() {}
249
250 llvm::Type *convertHLSLSpecificType(const Type *T,
251 const CGHLSLOffsetInfo &OffsetInfo);
252 llvm::Type *convertHLSLSpecificType(const Type *T) {
253 return convertHLSLSpecificType(T, OffsetInfo: CGHLSLOffsetInfo());
254 }
255
256 void generateGlobalCtorDtorCalls();
257
258 void addBuffer(const HLSLBufferDecl *D);
259 void addRootSignature(const HLSLRootSignatureDecl *D);
260 void finishCodeGen();
261
262 void setHLSLEntryAttributes(const FunctionDecl *FD, llvm::Function *Fn);
263
264 void emitEntryFunction(const FunctionDecl *FD, llvm::Function *Fn);
265 void setHLSLFunctionAttributes(const FunctionDecl *FD, llvm::Function *Fn);
266 void handleGlobalVarDefinition(const VarDecl *VD, llvm::GlobalVariable *Var);
267
268 llvm::Instruction *getConvergenceToken(llvm::BasicBlock &BB);
269
270 llvm::StructType *getHLSLBufferLayoutType(const RecordType *LayoutStructTy);
271 void addHLSLBufferLayoutType(const RecordType *LayoutStructTy,
272 llvm::StructType *LayoutTy);
273 void emitInitListOpaqueValues(CodeGenFunction &CGF, InitListExpr *E);
274
275 std::optional<LValue>
276 emitResourceArraySubscriptExpr(const ArraySubscriptExpr *E,
277 CodeGenFunction &CGF);
278 bool emitResourceArrayCopy(LValue &LHS, Expr *RHSExpr, CodeGenFunction &CGF);
279
280 std::optional<LValue> emitBufferArraySubscriptExpr(
281 const ArraySubscriptExpr *E, CodeGenFunction &CGF,
282 llvm::function_ref<llvm::Value *(bool Promote)> EmitIdxAfterBase);
283
284 RawAddress createBufferMatrixTempAddress(const LValue &LV, SourceLocation Loc,
285 CodeGenFunction &CGF);
286
287 bool emitBufferCopy(CodeGenFunction &CGF, Address DestPtr, Address SrcPtr,
288 QualType CType);
289
290 LValue emitBufferMemberExpr(CodeGenFunction &CGF, const MemberExpr *E);
291
292private:
293 void emitBufferGlobalsAndMetadata(const HLSLBufferDecl *BufDecl,
294 llvm::GlobalVariable *BufGV,
295 const CGHLSLOffsetInfo &OffsetInfo);
296 void initializeBufferFromBinding(const HLSLBufferDecl *BufDecl,
297 llvm::GlobalVariable *GV);
298 void initializeBufferFromBinding(const HLSLBufferDecl *BufDecl,
299 llvm::GlobalVariable *GV,
300 HLSLResourceBindingAttr *RBA);
301
302 llvm::Value *emitSPIRVUserSemanticLoad(llvm::IRBuilder<> &B, llvm::Type *Type,
303 const clang::DeclaratorDecl *Decl,
304 HLSLAppliedSemanticAttr *Semantic,
305 std::optional<unsigned> Index);
306 llvm::Value *emitDXILUserSemanticLoad(llvm::IRBuilder<> &B, llvm::Type *Type,
307 HLSLAppliedSemanticAttr *Semantic,
308 std::optional<unsigned> Index);
309 llvm::Value *emitUserSemanticLoad(llvm::IRBuilder<> &B, llvm::Type *Type,
310 const clang::DeclaratorDecl *Decl,
311 HLSLAppliedSemanticAttr *Semantic,
312 std::optional<unsigned> Index);
313
314 void emitSPIRVUserSemanticStore(llvm::IRBuilder<> &B, llvm::Value *Source,
315 const clang::DeclaratorDecl *Decl,
316 HLSLAppliedSemanticAttr *Semantic,
317 std::optional<unsigned> Index);
318 void emitDXILUserSemanticStore(llvm::IRBuilder<> &B, llvm::Value *Source,
319 HLSLAppliedSemanticAttr *Semantic,
320 std::optional<unsigned> Index);
321 void emitUserSemanticStore(llvm::IRBuilder<> &B, llvm::Value *Source,
322 const clang::DeclaratorDecl *Decl,
323 HLSLAppliedSemanticAttr *Semantic,
324 std::optional<unsigned> Index);
325
326 llvm::Triple::ArchType getArch();
327
328 llvm::DenseMap<const clang::RecordType *, llvm::StructType *> LayoutTypes;
329 unsigned SPIRVLastAssignedInputSemanticLocation = 0;
330 unsigned SPIRVLastAssignedOutputSemanticLocation = 0;
331};
332
333} // namespace CodeGen
334} // namespace clang
335
336#endif
337