xref: /freebsd/contrib/llvm-project/clang/lib/Basic/Targets/AMDGPU.h (revision 700637cbb5e582861067a11aaca4d053546871d2)
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/TargetID.h"
17 #include "clang/Basic/TargetInfo.h"
18 #include "clang/Basic/TargetOptions.h"
19 #include "llvm/ADT/StringSet.h"
20 #include "llvm/Support/AMDGPUAddrSpace.h"
21 #include "llvm/Support/Compiler.h"
22 #include "llvm/TargetParser/TargetParser.h"
23 #include "llvm/TargetParser/Triple.h"
24 #include <optional>
25 
26 namespace clang {
27 namespace targets {
28 
29 class LLVM_LIBRARY_VISIBILITY AMDGPUTargetInfo final : public TargetInfo {
30 
31   static const char *const GCCRegNames[];
32 
33   static const LangASMap AMDGPUDefIsGenMap;
34   static const LangASMap AMDGPUDefIsPrivMap;
35 
36   llvm::AMDGPU::GPUKind GPUKind;
37   unsigned GPUFeatures;
38   unsigned WavefrontSize;
39 
40   /// Whether to use cumode or WGP mode. True for cumode. False for WGP mode.
41   bool CUMode;
42 
43   /// Whether having image instructions.
44   bool HasImage = false;
45 
46   /// Target ID is device name followed by optional feature name postfixed
47   /// by plus or minus sign delimitted by colon, e.g. gfx908:xnack+:sramecc-.
48   /// If the target ID contains feature+, map it to true.
49   /// If the target ID contains feature-, map it to false.
50   /// If the target ID does not contain a feature (default), do not map it.
51   llvm::StringMap<bool> OffloadArchFeatures;
52   std::string TargetID;
53 
hasFP64()54   bool hasFP64() const {
55     return getTriple().isAMDGCN() ||
56            !!(GPUFeatures & llvm::AMDGPU::FEATURE_FP64);
57   }
58 
59   /// Has fast fma f32
hasFastFMAF()60   bool hasFastFMAF() const {
61     return !!(GPUFeatures & llvm::AMDGPU::FEATURE_FAST_FMA_F32);
62   }
63 
64   /// Has fast fma f64
hasFastFMA()65   bool hasFastFMA() const { return getTriple().isAMDGCN(); }
66 
hasFMAF()67   bool hasFMAF() const {
68     return getTriple().isAMDGCN() ||
69            !!(GPUFeatures & llvm::AMDGPU::FEATURE_FMA);
70   }
71 
hasFullRateDenormalsF32()72   bool hasFullRateDenormalsF32() const {
73     return !!(GPUFeatures & llvm::AMDGPU::FEATURE_FAST_DENORMAL_F32);
74   }
75 
hasLDEXPF()76   bool hasLDEXPF() const {
77     return getTriple().isAMDGCN() ||
78            !!(GPUFeatures & llvm::AMDGPU::FEATURE_LDEXP);
79   }
80 
isAMDGCN(const llvm::Triple & TT)81   static bool isAMDGCN(const llvm::Triple &TT) { return TT.isAMDGCN(); }
82 
isR600(const llvm::Triple & TT)83   static bool isR600(const llvm::Triple &TT) {
84     return TT.getArch() == llvm::Triple::r600;
85   }
86 
87 public:
88   AMDGPUTargetInfo(const llvm::Triple &Triple, const TargetOptions &Opts);
89 
90   void setAddressSpaceMap(bool DefaultIsPrivate);
91 
92   void adjust(DiagnosticsEngine &Diags, LangOptions &Opts,
93               const TargetInfo *Aux) override;
94 
getPointerWidthV(LangAS AS)95   uint64_t getPointerWidthV(LangAS AS) const override {
96     if (isR600(getTriple()))
97       return 32;
98     unsigned TargetAS = getTargetAddressSpace(AS);
99 
100     if (TargetAS == llvm::AMDGPUAS::PRIVATE_ADDRESS ||
101         TargetAS == llvm::AMDGPUAS::LOCAL_ADDRESS)
102       return 32;
103 
104     return 64;
105   }
106 
getPointerAlignV(LangAS AddrSpace)107   uint64_t getPointerAlignV(LangAS AddrSpace) const override {
108     return getPointerWidthV(AddrSpace);
109   }
110 
isAddressSpaceSupersetOf(LangAS A,LangAS B)111   virtual bool isAddressSpaceSupersetOf(LangAS A, LangAS B) const override {
112     // The flat address space AS(0) is a superset of all the other address
113     // spaces used by the backend target.
114     return A == B ||
115            ((A == LangAS::Default ||
116              (isTargetAddressSpace(A) &&
117               toTargetAddressSpace(A) == llvm::AMDGPUAS::FLAT_ADDRESS)) &&
118             isTargetAddressSpace(B) &&
119             toTargetAddressSpace(B) >= llvm::AMDGPUAS::FLAT_ADDRESS &&
120             toTargetAddressSpace(B) <= llvm::AMDGPUAS::PRIVATE_ADDRESS &&
121             toTargetAddressSpace(B) != llvm::AMDGPUAS::REGION_ADDRESS);
122   }
123 
getMaxPointerWidth()124   uint64_t getMaxPointerWidth() const override {
125     return getTriple().isAMDGCN() ? 64 : 32;
126   }
127 
hasBFloat16Type()128   bool hasBFloat16Type() const override { return isAMDGCN(getTriple()); }
129 
getClobbers()130   std::string_view getClobbers() const override { return ""; }
131 
132   ArrayRef<const char *> getGCCRegNames() const override;
133 
getGCCRegAliases()134   ArrayRef<TargetInfo::GCCRegAlias> getGCCRegAliases() const override {
135     return {};
136   }
137 
138   /// Accepted register names: (n, m is unsigned integer, n < m)
139   /// v
140   /// s
141   /// a
142   /// {vn}, {v[n]}
143   /// {sn}, {s[n]}
144   /// {an}, {a[n]}
145   /// {S} , where S is a special register name
146   ////{v[n:m]}
147   /// {s[n:m]}
148   /// {a[n:m]}
validateAsmConstraint(const char * & Name,TargetInfo::ConstraintInfo & Info)149   bool validateAsmConstraint(const char *&Name,
150                              TargetInfo::ConstraintInfo &Info) const override {
151     static const ::llvm::StringSet<> SpecialRegs({
152         "exec", "vcc", "flat_scratch", "m0", "scc", "tba", "tma",
153         "flat_scratch_lo", "flat_scratch_hi", "vcc_lo", "vcc_hi", "exec_lo",
154         "exec_hi", "tma_lo", "tma_hi", "tba_lo", "tba_hi",
155     });
156 
157     switch (*Name) {
158     case 'I':
159       Info.setRequiresImmediate(-16, 64);
160       return true;
161     case 'J':
162       Info.setRequiresImmediate(-32768, 32767);
163       return true;
164     case 'A':
165     case 'B':
166     case 'C':
167       Info.setRequiresImmediate();
168       return true;
169     default:
170       break;
171     }
172 
173     StringRef S(Name);
174 
175     if (S == "DA" || S == "DB") {
176       Name++;
177       Info.setRequiresImmediate();
178       return true;
179     }
180 
181     bool HasLeftParen = S.consume_front("{");
182     if (S.empty())
183       return false;
184     if (S.front() != 'v' && S.front() != 's' && S.front() != 'a') {
185       if (!HasLeftParen)
186         return false;
187       auto E = S.find('}');
188       if (!SpecialRegs.count(S.substr(0, E)))
189         return false;
190       S = S.drop_front(E + 1);
191       if (!S.empty())
192         return false;
193       // Found {S} where S is a special register.
194       Info.setAllowsRegister();
195       Name = S.data() - 1;
196       return true;
197     }
198     S = S.drop_front();
199     if (!HasLeftParen) {
200       if (!S.empty())
201         return false;
202       // Found s, v or a.
203       Info.setAllowsRegister();
204       Name = S.data() - 1;
205       return true;
206     }
207     bool HasLeftBracket = S.consume_front("[");
208     unsigned long long N;
209     if (S.empty() || consumeUnsignedInteger(S, 10, N))
210       return false;
211     if (S.consume_front(":")) {
212       if (!HasLeftBracket)
213         return false;
214       unsigned long long M;
215       if (consumeUnsignedInteger(S, 10, M) || N >= M)
216         return false;
217     }
218     if (HasLeftBracket) {
219       if (!S.consume_front("]"))
220         return false;
221     }
222     if (!S.consume_front("}"))
223       return false;
224     if (!S.empty())
225       return false;
226     // Found {vn}, {sn}, {an}, {v[n]}, {s[n]}, {a[n]}, {v[n:m]}, {s[n:m]}
227     // or {a[n:m]}.
228     Info.setAllowsRegister();
229     Name = S.data() - 1;
230     return true;
231   }
232 
233   // \p Constraint will be left pointing at the last character of
234   // the constraint.  In practice, it won't be changed unless the
235   // constraint is longer than one character.
convertConstraint(const char * & Constraint)236   std::string convertConstraint(const char *&Constraint) const override {
237 
238     StringRef S(Constraint);
239     if (S == "DA" || S == "DB") {
240       return std::string("^") + std::string(Constraint++, 2);
241     }
242 
243     const char *Begin = Constraint;
244     TargetInfo::ConstraintInfo Info("", "");
245     if (validateAsmConstraint(Constraint, Info))
246       return std::string(Begin).substr(0, Constraint - Begin + 1);
247 
248     Constraint = Begin;
249     return std::string(1, *Constraint);
250   }
251 
252   bool
253   initFeatureMap(llvm::StringMap<bool> &Features, DiagnosticsEngine &Diags,
254                  StringRef CPU,
255                  const std::vector<std::string> &FeatureVec) const override;
256 
257   llvm::SmallVector<Builtin::InfosShard> getTargetBuiltins() const override;
258 
useFP16ConversionIntrinsics()259   bool useFP16ConversionIntrinsics() const override { return false; }
260 
261   void getTargetDefines(const LangOptions &Opts,
262                         MacroBuilder &Builder) const override;
263 
getBuiltinVaListKind()264   BuiltinVaListKind getBuiltinVaListKind() const override {
265     return TargetInfo::CharPtrBuiltinVaList;
266   }
267 
isValidCPUName(StringRef Name)268   bool isValidCPUName(StringRef Name) const override {
269     if (getTriple().isAMDGCN())
270       return llvm::AMDGPU::parseArchAMDGCN(Name) != llvm::AMDGPU::GK_NONE;
271     return llvm::AMDGPU::parseArchR600(Name) != llvm::AMDGPU::GK_NONE;
272   }
273 
274   void fillValidCPUList(SmallVectorImpl<StringRef> &Values) const override;
275 
setCPU(const std::string & Name)276   bool setCPU(const std::string &Name) override {
277     if (getTriple().isAMDGCN()) {
278       GPUKind = llvm::AMDGPU::parseArchAMDGCN(Name);
279       GPUFeatures = llvm::AMDGPU::getArchAttrAMDGCN(GPUKind);
280     } else {
281       GPUKind = llvm::AMDGPU::parseArchR600(Name);
282       GPUFeatures = llvm::AMDGPU::getArchAttrR600(GPUKind);
283     }
284 
285     return GPUKind != llvm::AMDGPU::GK_NONE;
286   }
287 
setSupportedOpenCLOpts()288   void setSupportedOpenCLOpts() override {
289     auto &Opts = getSupportedOpenCLOpts();
290     Opts["cl_clang_storage_class_specifiers"] = true;
291     Opts["__cl_clang_variadic_functions"] = true;
292     Opts["__cl_clang_function_pointers"] = true;
293     Opts["__cl_clang_non_portable_kernel_param_types"] = true;
294     Opts["__cl_clang_bitfields"] = true;
295 
296     bool IsAMDGCN = isAMDGCN(getTriple());
297 
298     Opts["cl_khr_fp64"] = hasFP64();
299     Opts["__opencl_c_fp64"] = hasFP64();
300 
301     if (IsAMDGCN || GPUKind >= llvm::AMDGPU::GK_CEDAR) {
302       Opts["cl_khr_byte_addressable_store"] = true;
303       Opts["cl_khr_global_int32_base_atomics"] = true;
304       Opts["cl_khr_global_int32_extended_atomics"] = true;
305       Opts["cl_khr_local_int32_base_atomics"] = true;
306       Opts["cl_khr_local_int32_extended_atomics"] = true;
307     }
308 
309     if (IsAMDGCN) {
310       Opts["cl_khr_fp16"] = true;
311       Opts["cl_khr_int64_base_atomics"] = true;
312       Opts["cl_khr_int64_extended_atomics"] = true;
313       Opts["cl_khr_mipmap_image"] = true;
314       Opts["cl_khr_mipmap_image_writes"] = true;
315       Opts["cl_khr_subgroups"] = true;
316       Opts["cl_amd_media_ops"] = true;
317       Opts["cl_amd_media_ops2"] = true;
318 
319       Opts["__opencl_c_images"] = true;
320       Opts["__opencl_c_3d_image_writes"] = true;
321       Opts["cl_khr_3d_image_writes"] = true;
322 
323       Opts["__opencl_c_generic_address_space"] =
324           GPUKind >= llvm::AMDGPU::GK_GFX700;
325     }
326   }
327 
getOpenCLTypeAddrSpace(OpenCLTypeKind TK)328   LangAS getOpenCLTypeAddrSpace(OpenCLTypeKind TK) const override {
329     switch (TK) {
330     case OCLTK_Image:
331       return LangAS::opencl_constant;
332 
333     case OCLTK_ClkEvent:
334     case OCLTK_Queue:
335     case OCLTK_ReserveID:
336       return LangAS::opencl_global;
337 
338     default:
339       return TargetInfo::getOpenCLTypeAddrSpace(TK);
340     }
341   }
342 
getOpenCLBuiltinAddressSpace(unsigned AS)343   LangAS getOpenCLBuiltinAddressSpace(unsigned AS) const override {
344     switch (AS) {
345     case 0:
346       return LangAS::opencl_generic;
347     case 1:
348       return LangAS::opencl_global;
349     case 3:
350       return LangAS::opencl_local;
351     case 4:
352       return LangAS::opencl_constant;
353     case 5:
354       return LangAS::opencl_private;
355     default:
356       return getLangASFromTargetAS(AS);
357     }
358   }
359 
getCUDABuiltinAddressSpace(unsigned AS)360   LangAS getCUDABuiltinAddressSpace(unsigned AS) const override {
361     switch (AS) {
362     case 0:
363       return LangAS::Default;
364     case 1:
365       return LangAS::cuda_device;
366     case 3:
367       return LangAS::cuda_shared;
368     case 4:
369       return LangAS::cuda_constant;
370     default:
371       return getLangASFromTargetAS(AS);
372     }
373   }
374 
getConstantAddressSpace()375   std::optional<LangAS> getConstantAddressSpace() const override {
376     return getLangASFromTargetAS(llvm::AMDGPUAS::CONSTANT_ADDRESS);
377   }
378 
getGridValue()379   const llvm::omp::GV &getGridValue() const override {
380     switch (WavefrontSize) {
381     case 32:
382       return llvm::omp::getAMDGPUGridValues<32>();
383     case 64:
384       return llvm::omp::getAMDGPUGridValues<64>();
385     default:
386       llvm_unreachable("getGridValue not implemented for this wavesize");
387     }
388   }
389 
390   /// \returns Target specific vtbl ptr address space.
getVtblPtrAddressSpace()391   unsigned getVtblPtrAddressSpace() const override {
392     return static_cast<unsigned>(llvm::AMDGPUAS::CONSTANT_ADDRESS);
393   }
394 
395   /// \returns If a target requires an address within a target specific address
396   /// space \p AddressSpace to be converted in order to be used, then return the
397   /// corresponding target specific DWARF address space.
398   ///
399   /// \returns Otherwise return std::nullopt and no conversion will be emitted
400   /// in the DWARF.
401   std::optional<unsigned>
getDWARFAddressSpace(unsigned AddressSpace)402   getDWARFAddressSpace(unsigned AddressSpace) const override {
403     const unsigned DWARF_Private = 1;
404     const unsigned DWARF_Local = 2;
405     if (AddressSpace == llvm::AMDGPUAS::PRIVATE_ADDRESS) {
406       return DWARF_Private;
407     } else if (AddressSpace == llvm::AMDGPUAS::LOCAL_ADDRESS) {
408       return DWARF_Local;
409     } else {
410       return std::nullopt;
411     }
412   }
413 
checkCallingConvention(CallingConv CC)414   CallingConvCheckResult checkCallingConvention(CallingConv CC) const override {
415     switch (CC) {
416     default:
417       return CCCR_Warning;
418     case CC_C:
419     case CC_DeviceKernel:
420       return CCCR_OK;
421     }
422   }
423 
424   // In amdgcn target the null pointer in global, constant, and generic
425   // address space has value 0 but in private and local address space has
426   // value ~0.
getNullPointerValue(LangAS AS)427   uint64_t getNullPointerValue(LangAS AS) const override {
428     // FIXME: Also should handle region.
429     return (AS == LangAS::opencl_local || AS == LangAS::opencl_private ||
430             AS == LangAS::sycl_local || AS == LangAS::sycl_private)
431                ? ~0
432                : 0;
433   }
434 
435   void setAuxTarget(const TargetInfo *Aux) override;
436 
hasBitIntType()437   bool hasBitIntType() const override { return true; }
438 
439   // Record offload arch features since they are needed for defining the
440   // pre-defined macros.
handleTargetFeatures(std::vector<std::string> & Features,DiagnosticsEngine & Diags)441   bool handleTargetFeatures(std::vector<std::string> &Features,
442                             DiagnosticsEngine &Diags) override {
443     HasFullBFloat16 = true;
444     auto TargetIDFeatures =
445         getAllPossibleTargetIDFeatures(getTriple(), getArchNameAMDGCN(GPUKind));
446     for (const auto &F : Features) {
447       assert(F.front() == '+' || F.front() == '-');
448       if (F == "+wavefrontsize64")
449         WavefrontSize = 64;
450       else if (F == "+cumode")
451         CUMode = true;
452       else if (F == "-cumode")
453         CUMode = false;
454       else if (F == "+image-insts")
455         HasImage = true;
456       bool IsOn = F.front() == '+';
457       StringRef Name = StringRef(F).drop_front();
458       if (!llvm::is_contained(TargetIDFeatures, Name))
459         continue;
460       assert(!OffloadArchFeatures.contains(Name));
461       OffloadArchFeatures[Name] = IsOn;
462     }
463     return true;
464   }
465 
getTargetID()466   std::optional<std::string> getTargetID() const override {
467     if (!isAMDGCN(getTriple()))
468       return std::nullopt;
469     // When -target-cpu is not set, we assume generic code that it is valid
470     // for all GPU and use an empty string as target ID to represent that.
471     if (GPUKind == llvm::AMDGPU::GK_NONE)
472       return std::string("");
473     return getCanonicalTargetID(getArchNameAMDGCN(GPUKind),
474                                 OffloadArchFeatures);
475   }
476 
hasHIPImageSupport()477   bool hasHIPImageSupport() const override { return HasImage; }
478 
hardwareInterferenceSizes()479   std::pair<unsigned, unsigned> hardwareInterferenceSizes() const override {
480     // This is imprecise as the value can vary between 64, 128 (even 256!) bytes
481     // depending on the level of cache and the target architecture. We select
482     // the size that corresponds to the largest L1 cache line for all
483     // architectures.
484     return std::make_pair(128, 128);
485   }
486 };
487 
488 } // namespace targets
489 } // namespace clang
490 
491 #endif // LLVM_CLANG_LIB_BASIC_TARGETS_AMDGPU_H
492