1 //===--- AMDGPU.h - Declare AMDGPU target feature support -------*- 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 file declares AMDGPU TargetInfo objects.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #ifndef LLVM_CLANG_LIB_BASIC_TARGETS_AMDGPU_H
14 #define LLVM_CLANG_LIB_BASIC_TARGETS_AMDGPU_H
15 
16 #include "clang/Basic/TargetInfo.h"
17 #include "clang/Basic/TargetOptions.h"
18 #include "llvm/ADT/StringSet.h"
19 #include "llvm/ADT/Triple.h"
20 #include "llvm/Support/Compiler.h"
21 #include "llvm/Support/TargetParser.h"
22 
23 namespace clang {
24 namespace targets {
25 
26 class LLVM_LIBRARY_VISIBILITY AMDGPUTargetInfo final : public TargetInfo {
27 
28   static const Builtin::Info BuiltinInfo[];
29   static const char *const GCCRegNames[];
30 
31   enum AddrSpace {
32     Generic = 0,
33     Global = 1,
34     Local = 3,
35     Constant = 4,
36     Private = 5
37   };
38   static const LangASMap AMDGPUDefIsGenMap;
39   static const LangASMap AMDGPUDefIsPrivMap;
40 
41   llvm::AMDGPU::GPUKind GPUKind;
42   unsigned GPUFeatures;
43 
44   bool hasFP64() const {
45     return getTriple().getArch() == llvm::Triple::amdgcn ||
46            !!(GPUFeatures & llvm::AMDGPU::FEATURE_FP64);
47   }
48 
49   /// Has fast fma f32
50   bool hasFastFMAF() const {
51     return !!(GPUFeatures & llvm::AMDGPU::FEATURE_FAST_FMA_F32);
52   }
53 
54   /// Has fast fma f64
55   bool hasFastFMA() const {
56     return getTriple().getArch() == llvm::Triple::amdgcn;
57   }
58 
59   bool hasFMAF() const {
60     return getTriple().getArch() == llvm::Triple::amdgcn ||
61            !!(GPUFeatures & llvm::AMDGPU::FEATURE_FMA);
62   }
63 
64   bool hasFullRateDenormalsF32() const {
65     return !!(GPUFeatures & llvm::AMDGPU::FEATURE_FAST_DENORMAL_F32);
66   }
67 
68   bool hasLDEXPF() const {
69     return getTriple().getArch() == llvm::Triple::amdgcn ||
70            !!(GPUFeatures & llvm::AMDGPU::FEATURE_LDEXP);
71   }
72 
73   static bool isAMDGCN(const llvm::Triple &TT) {
74     return TT.getArch() == llvm::Triple::amdgcn;
75   }
76 
77   static bool isR600(const llvm::Triple &TT) {
78     return TT.getArch() == llvm::Triple::r600;
79   }
80 
81 public:
82   AMDGPUTargetInfo(const llvm::Triple &Triple, const TargetOptions &Opts);
83 
84   void setAddressSpaceMap(bool DefaultIsPrivate);
85 
86   void adjust(LangOptions &Opts) override;
87 
88   uint64_t getPointerWidthV(unsigned AddrSpace) const override {
89     if (isR600(getTriple()))
90       return 32;
91 
92     if (AddrSpace == Private || AddrSpace == Local)
93       return 32;
94 
95     return 64;
96   }
97 
98   uint64_t getPointerAlignV(unsigned AddrSpace) const override {
99     return getPointerWidthV(AddrSpace);
100   }
101 
102   uint64_t getMaxPointerWidth() const override {
103     return getTriple().getArch() == llvm::Triple::amdgcn ? 64 : 32;
104   }
105 
106   const char *getClobbers() const override { return ""; }
107 
108   ArrayRef<const char *> getGCCRegNames() const override;
109 
110   ArrayRef<TargetInfo::GCCRegAlias> getGCCRegAliases() const override {
111     return None;
112   }
113 
114   /// Accepted register names: (n, m is unsigned integer, n < m)
115   /// v
116   /// s
117   /// {vn}, {v[n]}
118   /// {sn}, {s[n]}
119   /// {S} , where S is a special register name
120   ////{v[n:m]}
121   /// {s[n:m]}
122   bool validateAsmConstraint(const char *&Name,
123                              TargetInfo::ConstraintInfo &Info) const override {
124     static const ::llvm::StringSet<> SpecialRegs({
125         "exec", "vcc", "flat_scratch", "m0", "scc", "tba", "tma",
126         "flat_scratch_lo", "flat_scratch_hi", "vcc_lo", "vcc_hi", "exec_lo",
127         "exec_hi", "tma_lo", "tma_hi", "tba_lo", "tba_hi",
128     });
129 
130     StringRef S(Name);
131     bool HasLeftParen = false;
132     if (S.front() == '{') {
133       HasLeftParen = true;
134       S = S.drop_front();
135     }
136     if (S.empty())
137       return false;
138     if (S.front() != 'v' && S.front() != 's') {
139       if (!HasLeftParen)
140         return false;
141       auto E = S.find('}');
142       if (!SpecialRegs.count(S.substr(0, E)))
143         return false;
144       S = S.drop_front(E + 1);
145       if (!S.empty())
146         return false;
147       // Found {S} where S is a special register.
148       Info.setAllowsRegister();
149       Name = S.data() - 1;
150       return true;
151     }
152     S = S.drop_front();
153     if (!HasLeftParen) {
154       if (!S.empty())
155         return false;
156       // Found s or v.
157       Info.setAllowsRegister();
158       Name = S.data() - 1;
159       return true;
160     }
161     bool HasLeftBracket = false;
162     if (!S.empty() && S.front() == '[') {
163       HasLeftBracket = true;
164       S = S.drop_front();
165     }
166     unsigned long long N;
167     if (S.empty() || consumeUnsignedInteger(S, 10, N))
168       return false;
169     if (!S.empty() && S.front() == ':') {
170       if (!HasLeftBracket)
171         return false;
172       S = S.drop_front();
173       unsigned long long M;
174       if (consumeUnsignedInteger(S, 10, M) || N >= M)
175         return false;
176     }
177     if (HasLeftBracket) {
178       if (S.empty() || S.front() != ']')
179         return false;
180       S = S.drop_front();
181     }
182     if (S.empty() || S.front() != '}')
183       return false;
184     S = S.drop_front();
185     if (!S.empty())
186       return false;
187     // Found {vn}, {sn}, {v[n]}, {s[n]}, {v[n:m]}, or {s[n:m]}.
188     Info.setAllowsRegister();
189     Name = S.data() - 1;
190     return true;
191   }
192 
193   // \p Constraint will be left pointing at the last character of
194   // the constraint.  In practice, it won't be changed unless the
195   // constraint is longer than one character.
196   std::string convertConstraint(const char *&Constraint) const override {
197     const char *Begin = Constraint;
198     TargetInfo::ConstraintInfo Info("", "");
199     if (validateAsmConstraint(Constraint, Info))
200       return std::string(Begin).substr(0, Constraint - Begin + 1);
201 
202     Constraint = Begin;
203     return std::string(1, *Constraint);
204   }
205 
206   bool
207   initFeatureMap(llvm::StringMap<bool> &Features, DiagnosticsEngine &Diags,
208                  StringRef CPU,
209                  const std::vector<std::string> &FeatureVec) const override;
210 
211   void adjustTargetOptions(const CodeGenOptions &CGOpts,
212                            TargetOptions &TargetOpts) const override;
213 
214   ArrayRef<Builtin::Info> getTargetBuiltins() const override;
215 
216   void getTargetDefines(const LangOptions &Opts,
217                         MacroBuilder &Builder) const override;
218 
219   BuiltinVaListKind getBuiltinVaListKind() const override {
220     return TargetInfo::CharPtrBuiltinVaList;
221   }
222 
223   bool isValidCPUName(StringRef Name) const override {
224     if (getTriple().getArch() == llvm::Triple::amdgcn)
225       return llvm::AMDGPU::parseArchAMDGCN(Name) != llvm::AMDGPU::GK_NONE;
226     return llvm::AMDGPU::parseArchR600(Name) != llvm::AMDGPU::GK_NONE;
227   }
228 
229   void fillValidCPUList(SmallVectorImpl<StringRef> &Values) const override;
230 
231   bool setCPU(const std::string &Name) override {
232     if (getTriple().getArch() == llvm::Triple::amdgcn) {
233       GPUKind = llvm::AMDGPU::parseArchAMDGCN(Name);
234       GPUFeatures = llvm::AMDGPU::getArchAttrAMDGCN(GPUKind);
235     } else {
236       GPUKind = llvm::AMDGPU::parseArchR600(Name);
237       GPUFeatures = llvm::AMDGPU::getArchAttrR600(GPUKind);
238     }
239 
240     return GPUKind != llvm::AMDGPU::GK_NONE;
241   }
242 
243   void setSupportedOpenCLOpts() override {
244     auto &Opts = getSupportedOpenCLOpts();
245     Opts.support("cl_clang_storage_class_specifiers");
246     Opts.support("cl_khr_icd");
247 
248     bool IsAMDGCN = isAMDGCN(getTriple());
249 
250     if (hasFP64())
251       Opts.support("cl_khr_fp64");
252 
253     if (IsAMDGCN || GPUKind >= llvm::AMDGPU::GK_CEDAR) {
254       Opts.support("cl_khr_byte_addressable_store");
255       Opts.support("cl_khr_global_int32_base_atomics");
256       Opts.support("cl_khr_global_int32_extended_atomics");
257       Opts.support("cl_khr_local_int32_base_atomics");
258       Opts.support("cl_khr_local_int32_extended_atomics");
259     }
260 
261     if (IsAMDGCN) {
262       Opts.support("cl_khr_fp16");
263       Opts.support("cl_khr_int64_base_atomics");
264       Opts.support("cl_khr_int64_extended_atomics");
265       Opts.support("cl_khr_mipmap_image");
266       Opts.support("cl_khr_mipmap_image_writes");
267       Opts.support("cl_khr_subgroups");
268       Opts.support("cl_khr_3d_image_writes");
269       Opts.support("cl_amd_media_ops");
270       Opts.support("cl_amd_media_ops2");
271     }
272   }
273 
274   LangAS getOpenCLTypeAddrSpace(OpenCLTypeKind TK) const override {
275     switch (TK) {
276     case OCLTK_Image:
277       return LangAS::opencl_constant;
278 
279     case OCLTK_ClkEvent:
280     case OCLTK_Queue:
281     case OCLTK_ReserveID:
282       return LangAS::opencl_global;
283 
284     default:
285       return TargetInfo::getOpenCLTypeAddrSpace(TK);
286     }
287   }
288 
289   LangAS getOpenCLBuiltinAddressSpace(unsigned AS) const override {
290     switch (AS) {
291     case 0:
292       return LangAS::opencl_generic;
293     case 1:
294       return LangAS::opencl_global;
295     case 3:
296       return LangAS::opencl_local;
297     case 4:
298       return LangAS::opencl_constant;
299     case 5:
300       return LangAS::opencl_private;
301     default:
302       return getLangASFromTargetAS(AS);
303     }
304   }
305 
306   LangAS getCUDABuiltinAddressSpace(unsigned AS) const override {
307     return LangAS::Default;
308   }
309 
310   llvm::Optional<LangAS> getConstantAddressSpace() const override {
311     return getLangASFromTargetAS(Constant);
312   }
313 
314   /// \returns Target specific vtbl ptr address space.
315   unsigned getVtblPtrAddressSpace() const override {
316     return static_cast<unsigned>(Constant);
317   }
318 
319   /// \returns If a target requires an address within a target specific address
320   /// space \p AddressSpace to be converted in order to be used, then return the
321   /// corresponding target specific DWARF address space.
322   ///
323   /// \returns Otherwise return None and no conversion will be emitted in the
324   /// DWARF.
325   Optional<unsigned>
326   getDWARFAddressSpace(unsigned AddressSpace) const override {
327     const unsigned DWARF_Private = 1;
328     const unsigned DWARF_Local = 2;
329     if (AddressSpace == Private) {
330       return DWARF_Private;
331     } else if (AddressSpace == Local) {
332       return DWARF_Local;
333     } else {
334       return None;
335     }
336   }
337 
338   CallingConvCheckResult checkCallingConvention(CallingConv CC) const override {
339     switch (CC) {
340     default:
341       return CCCR_Warning;
342     case CC_C:
343     case CC_OpenCLKernel:
344       return CCCR_OK;
345     }
346   }
347 
348   // In amdgcn target the null pointer in global, constant, and generic
349   // address space has value 0 but in private and local address space has
350   // value ~0.
351   uint64_t getNullPointerValue(LangAS AS) const override {
352     return AS == LangAS::opencl_local ? ~0 : 0;
353   }
354 
355   void setAuxTarget(const TargetInfo *Aux) override;
356 };
357 
358 } // namespace targets
359 } // namespace clang
360 
361 #endif // LLVM_CLANG_LIB_BASIC_TARGETS_AMDGPU_H
362