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(WaveActiveAllEqual, wave_all_equal)
148 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveAllTrue, wave_all)
149 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveAnyTrue, wave_any)
150 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveBitOr, wave_reduce_or)
151 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveBitXor, wave_reduce_xor)
152 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveMax, wave_reduce_max)
153 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveUMax, wave_reduce_umax)
154 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveMin, wave_reduce_min)
155 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveUMin, wave_reduce_umin)
156 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveActiveCountBits, wave_active_countbits)
157 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveIsFirstLane, wave_is_first_lane)
158 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveGetLaneCount, wave_get_lane_count)
159 GENERATE_HLSL_INTRINSIC_FUNCTION(WaveReadLaneAt, wave_readlane)
160 GENERATE_HLSL_INTRINSIC_FUNCTION(FirstBitUHigh, firstbituhigh)
161 GENERATE_HLSL_INTRINSIC_FUNCTION(FirstBitSHigh, firstbitshigh)
162 GENERATE_HLSL_INTRINSIC_FUNCTION(FirstBitLow, firstbitlow)
163 GENERATE_HLSL_INTRINSIC_FUNCTION(NClamp, nclamp)
164 GENERATE_HLSL_INTRINSIC_FUNCTION(SClamp, sclamp)
165 GENERATE_HLSL_INTRINSIC_FUNCTION(UClamp, uclamp)
166
167 GENERATE_HLSL_INTRINSIC_FUNCTION(CreateResourceGetPointer,
168 resource_getpointer)
169 GENERATE_HLSL_INTRINSIC_FUNCTION(Sample, resource_sample)
170 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleClamp, resource_sample_clamp)
171 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleBias, resource_samplebias)
172 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleBiasClamp, resource_samplebias_clamp)
173 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleGrad, resource_samplegrad)
174 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleGradClamp, resource_samplegrad_clamp)
175 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleLevel, resource_samplelevel)
176 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleCmp, resource_samplecmp)
177 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleCmpClamp, resource_samplecmp_clamp)
178 GENERATE_HLSL_INTRINSIC_FUNCTION(SampleCmpLevelZero,
179 resource_samplecmplevelzero)
180 GENERATE_HLSL_INTRINSIC_FUNCTION(Gather, resource_gather)
181 GENERATE_HLSL_INTRINSIC_FUNCTION(GatherCmp, resource_gather_cmp)
182 GENERATE_HLSL_INTRINSIC_FUNCTION(CreateHandleFromBinding,
183 resource_handlefrombinding)
184 GENERATE_HLSL_INTRINSIC_FUNCTION(CreateHandleFromImplicitBinding,
185 resource_handlefromimplicitbinding)
186 GENERATE_HLSL_INTRINSIC_FUNCTION(NonUniformResourceIndex,
187 resource_nonuniformindex)
188 GENERATE_HLSL_INTRINSIC_FUNCTION(BufferUpdateCounter, resource_updatecounter)
189 GENERATE_HLSL_INTRINSIC_FUNCTION(GroupMemoryBarrierWithGroupSync,
190 group_memory_barrier_with_group_sync)
191 GENERATE_HLSL_INTRINSIC_FUNCTION(GetDimensionsX, resource_getdimensions_x)
192 GENERATE_HLSL_INTRINSIC_FUNCTION(LoadLevel, resource_load_level)
193 GENERATE_HLSL_INTRINSIC_FUNCTION(DdxCoarse, ddx_coarse)
194 GENERATE_HLSL_INTRINSIC_FUNCTION(DdyCoarse, ddy_coarse)
195 GENERATE_HLSL_INTRINSIC_FUNCTION(DdxFine, ddx_fine)
196 GENERATE_HLSL_INTRINSIC_FUNCTION(DdyFine, ddy_fine)
197
198 //===----------------------------------------------------------------------===//
199 // End of reserved area for HLSL intrinsic getters.
200 //===----------------------------------------------------------------------===//
201
202protected:
203 CodeGenModule &CGM;
204
205 llvm::Value *emitSystemSemanticLoad(llvm::IRBuilder<> &B,
206 const FunctionDecl *FD, llvm::Type *Type,
207 const clang::DeclaratorDecl *Decl,
208 HLSLAppliedSemanticAttr *Semantic,
209 std::optional<unsigned> Index);
210
211 void emitSystemSemanticStore(llvm::IRBuilder<> &B, llvm::Value *Source,
212 const clang::DeclaratorDecl *Decl,
213 HLSLAppliedSemanticAttr *Semantic,
214 std::optional<unsigned> Index);
215
216 llvm::Value *handleScalarSemanticLoad(llvm::IRBuilder<> &B,
217 const FunctionDecl *FD,
218 llvm::Type *Type,
219 const clang::DeclaratorDecl *Decl,
220 HLSLAppliedSemanticAttr *Semantic);
221
222 void handleScalarSemanticStore(llvm::IRBuilder<> &B, const FunctionDecl *FD,
223 llvm::Value *Source,
224 const clang::DeclaratorDecl *Decl,
225 HLSLAppliedSemanticAttr *Semantic);
226
227 std::pair<llvm::Value *, specific_attr_iterator<HLSLAppliedSemanticAttr>>
228 handleStructSemanticLoad(
229 llvm::IRBuilder<> &B, const FunctionDecl *FD, llvm::Type *Type,
230 const clang::DeclaratorDecl *Decl,
231 specific_attr_iterator<HLSLAppliedSemanticAttr> begin,
232 specific_attr_iterator<HLSLAppliedSemanticAttr> end);
233
234 specific_attr_iterator<HLSLAppliedSemanticAttr> handleStructSemanticStore(
235 llvm::IRBuilder<> &B, const FunctionDecl *FD, llvm::Value *Source,
236 const clang::DeclaratorDecl *Decl,
237 specific_attr_iterator<HLSLAppliedSemanticAttr> AttrBegin,
238 specific_attr_iterator<HLSLAppliedSemanticAttr> AttrEnd);
239
240 std::pair<llvm::Value *, specific_attr_iterator<HLSLAppliedSemanticAttr>>
241 handleSemanticLoad(llvm::IRBuilder<> &B, const FunctionDecl *FD,
242 llvm::Type *Type, const clang::DeclaratorDecl *Decl,
243 specific_attr_iterator<HLSLAppliedSemanticAttr> begin,
244 specific_attr_iterator<HLSLAppliedSemanticAttr> end);
245
246 specific_attr_iterator<HLSLAppliedSemanticAttr>
247 handleSemanticStore(llvm::IRBuilder<> &B, const FunctionDecl *FD,
248 llvm::Value *Source, const clang::DeclaratorDecl *Decl,
249 specific_attr_iterator<HLSLAppliedSemanticAttr> AttrBegin,
250 specific_attr_iterator<HLSLAppliedSemanticAttr> AttrEnd);
251
252public:
253 CGHLSLRuntime(CodeGenModule &CGM) : CGM(CGM) {}
254 virtual ~CGHLSLRuntime() {}
255
256 llvm::Type *convertHLSLSpecificType(const Type *T,
257 const CGHLSLOffsetInfo &OffsetInfo);
258 llvm::Type *convertHLSLSpecificType(const Type *T) {
259 return convertHLSLSpecificType(T, OffsetInfo: CGHLSLOffsetInfo());
260 }
261
262 void generateGlobalCtorDtorCalls();
263
264 void addBuffer(const HLSLBufferDecl *D);
265 void addRootSignature(const HLSLRootSignatureDecl *D);
266 void finishCodeGen();
267
268 void setHLSLEntryAttributes(const FunctionDecl *FD, llvm::Function *Fn);
269
270 void emitEntryFunction(const FunctionDecl *FD, llvm::Function *Fn);
271 void setHLSLFunctionAttributes(const FunctionDecl *FD, llvm::Function *Fn);
272 void handleGlobalVarDefinition(const VarDecl *VD, llvm::GlobalVariable *Var);
273
274 llvm::Instruction *getConvergenceToken(llvm::BasicBlock &BB);
275
276 llvm::StructType *getHLSLBufferLayoutType(const RecordType *LayoutStructTy);
277 void addHLSLBufferLayoutType(const RecordType *LayoutStructTy,
278 llvm::StructType *LayoutTy);
279 void emitInitListOpaqueValues(CodeGenFunction &CGF, InitListExpr *E);
280
281 std::optional<LValue>
282 emitResourceArraySubscriptExpr(const ArraySubscriptExpr *E,
283 CodeGenFunction &CGF);
284 bool emitResourceArrayCopy(LValue &LHS, Expr *RHSExpr, CodeGenFunction &CGF);
285
286 std::optional<LValue> emitBufferArraySubscriptExpr(
287 const ArraySubscriptExpr *E, CodeGenFunction &CGF,
288 llvm::function_ref<llvm::Value *(bool Promote)> EmitIdxAfterBase);
289
290 RawAddress createBufferMatrixTempAddress(const LValue &LV, SourceLocation Loc,
291 CodeGenFunction &CGF);
292
293 bool emitBufferCopy(CodeGenFunction &CGF, Address DestPtr, Address SrcPtr,
294 QualType CType);
295
296 LValue emitBufferMemberExpr(CodeGenFunction &CGF, const MemberExpr *E);
297
298private:
299 void emitBufferGlobalsAndMetadata(const HLSLBufferDecl *BufDecl,
300 llvm::GlobalVariable *BufGV,
301 const CGHLSLOffsetInfo &OffsetInfo);
302 void initializeBufferFromBinding(const HLSLBufferDecl *BufDecl,
303 llvm::GlobalVariable *GV);
304 void initializeBufferFromBinding(const HLSLBufferDecl *BufDecl,
305 llvm::GlobalVariable *GV,
306 HLSLResourceBindingAttr *RBA);
307
308 llvm::Value *emitSPIRVUserSemanticLoad(llvm::IRBuilder<> &B, llvm::Type *Type,
309 const clang::DeclaratorDecl *Decl,
310 HLSLAppliedSemanticAttr *Semantic,
311 std::optional<unsigned> Index);
312 llvm::Value *emitDXILUserSemanticLoad(llvm::IRBuilder<> &B, llvm::Type *Type,
313 HLSLAppliedSemanticAttr *Semantic,
314 std::optional<unsigned> Index);
315 llvm::Value *emitUserSemanticLoad(llvm::IRBuilder<> &B, llvm::Type *Type,
316 const clang::DeclaratorDecl *Decl,
317 HLSLAppliedSemanticAttr *Semantic,
318 std::optional<unsigned> Index);
319
320 void emitSPIRVUserSemanticStore(llvm::IRBuilder<> &B, llvm::Value *Source,
321 const clang::DeclaratorDecl *Decl,
322 HLSLAppliedSemanticAttr *Semantic,
323 std::optional<unsigned> Index);
324 void emitDXILUserSemanticStore(llvm::IRBuilder<> &B, llvm::Value *Source,
325 HLSLAppliedSemanticAttr *Semantic,
326 std::optional<unsigned> Index);
327 void emitUserSemanticStore(llvm::IRBuilder<> &B, llvm::Value *Source,
328 const clang::DeclaratorDecl *Decl,
329 HLSLAppliedSemanticAttr *Semantic,
330 std::optional<unsigned> Index);
331
332 llvm::Triple::ArchType getArch();
333
334 llvm::DenseMap<const clang::RecordType *, llvm::StructType *> LayoutTypes;
335 unsigned SPIRVLastAssignedInputSemanticLocation = 0;
336 unsigned SPIRVLastAssignedOutputSemanticLocation = 0;
337};
338
339} // namespace CodeGen
340} // namespace clang
341
342#endif
343