1 //===--- NVPTX.cpp - Implement NVPTX target feature support ---------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements NVPTX TargetInfo objects.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "NVPTX.h"
15 #include "Targets.h"
16 #include "clang/Basic/Builtins.h"
17 #include "clang/Basic/MacroBuilder.h"
18 #include "clang/Basic/TargetBuiltins.h"
19 #include "llvm/ADT/StringSwitch.h"
20 
21 using namespace clang;
22 using namespace clang::targets;
23 
24 const Builtin::Info NVPTXTargetInfo::BuiltinInfo[] = {
25 #define BUILTIN(ID, TYPE, ATTRS)                                               \
26   {#ID, TYPE, ATTRS, nullptr, ALL_LANGUAGES, nullptr},
27 #define LIBBUILTIN(ID, TYPE, ATTRS, HEADER)                                    \
28   {#ID, TYPE, ATTRS, HEADER, ALL_LANGUAGES, nullptr},
29 #define TARGET_BUILTIN(ID, TYPE, ATTRS, FEATURE)                               \
30   {#ID, TYPE, ATTRS, nullptr, ALL_LANGUAGES, FEATURE},
31 #include "clang/Basic/BuiltinsNVPTX.def"
32 };
33 
34 const char *const NVPTXTargetInfo::GCCRegNames[] = {"r0"};
35 
36 NVPTXTargetInfo::NVPTXTargetInfo(const llvm::Triple &Triple,
37                                  const TargetOptions &Opts,
38                                  unsigned TargetPointerWidth)
39     : TargetInfo(Triple) {
40   assert((TargetPointerWidth == 32 || TargetPointerWidth == 64) &&
41          "NVPTX only supports 32- and 64-bit modes.");
42 
43   PTXVersion = 32;
44   for (const StringRef Feature : Opts.FeaturesAsWritten) {
45     if (!Feature.startswith("+ptx"))
46       continue;
47     PTXVersion = llvm::StringSwitch<unsigned>(Feature)
48                      .Case("+ptx61", 61)
49                      .Case("+ptx60", 60)
50                      .Case("+ptx50", 50)
51                      .Case("+ptx43", 43)
52                      .Case("+ptx42", 42)
53                      .Case("+ptx41", 41)
54                      .Case("+ptx40", 40)
55                      .Case("+ptx32", 32)
56                      .Default(32);
57   }
58 
59   TLSSupported = false;
60   VLASupported = false;
61   AddrSpaceMap = &NVPTXAddrSpaceMap;
62   UseAddrSpaceMapMangling = true;
63 
64   // Define available target features
65   // These must be defined in sorted order!
66   NoAsmVariants = true;
67   GPU = CudaArch::SM_20;
68 
69   if (TargetPointerWidth == 32)
70     resetDataLayout("e-p:32:32-i64:64-i128:128-v16:16-v32:32-n16:32:64");
71   else
72     resetDataLayout("e-i64:64-i128:128-v16:16-v32:32-n16:32:64");
73 
74   // If possible, get a TargetInfo for our host triple, so we can match its
75   // types.
76   llvm::Triple HostTriple(Opts.HostTriple);
77   if (!HostTriple.isNVPTX())
78     HostTarget.reset(AllocateTarget(llvm::Triple(Opts.HostTriple), Opts));
79 
80   // If no host target, make some guesses about the data layout and return.
81   if (!HostTarget) {
82     LongWidth = LongAlign = TargetPointerWidth;
83     PointerWidth = PointerAlign = TargetPointerWidth;
84     switch (TargetPointerWidth) {
85     case 32:
86       SizeType = TargetInfo::UnsignedInt;
87       PtrDiffType = TargetInfo::SignedInt;
88       IntPtrType = TargetInfo::SignedInt;
89       break;
90     case 64:
91       SizeType = TargetInfo::UnsignedLong;
92       PtrDiffType = TargetInfo::SignedLong;
93       IntPtrType = TargetInfo::SignedLong;
94       break;
95     default:
96       llvm_unreachable("TargetPointerWidth must be 32 or 64");
97     }
98     return;
99   }
100 
101   // Copy properties from host target.
102   PointerWidth = HostTarget->getPointerWidth(/* AddrSpace = */ 0);
103   PointerAlign = HostTarget->getPointerAlign(/* AddrSpace = */ 0);
104   BoolWidth = HostTarget->getBoolWidth();
105   BoolAlign = HostTarget->getBoolAlign();
106   IntWidth = HostTarget->getIntWidth();
107   IntAlign = HostTarget->getIntAlign();
108   HalfWidth = HostTarget->getHalfWidth();
109   HalfAlign = HostTarget->getHalfAlign();
110   FloatWidth = HostTarget->getFloatWidth();
111   FloatAlign = HostTarget->getFloatAlign();
112   DoubleWidth = HostTarget->getDoubleWidth();
113   DoubleAlign = HostTarget->getDoubleAlign();
114   LongWidth = HostTarget->getLongWidth();
115   LongAlign = HostTarget->getLongAlign();
116   LongLongWidth = HostTarget->getLongLongWidth();
117   LongLongAlign = HostTarget->getLongLongAlign();
118   MinGlobalAlign = HostTarget->getMinGlobalAlign();
119   NewAlign = HostTarget->getNewAlign();
120   DefaultAlignForAttributeAligned =
121       HostTarget->getDefaultAlignForAttributeAligned();
122   SizeType = HostTarget->getSizeType();
123   IntMaxType = HostTarget->getIntMaxType();
124   PtrDiffType = HostTarget->getPtrDiffType(/* AddrSpace = */ 0);
125   IntPtrType = HostTarget->getIntPtrType();
126   WCharType = HostTarget->getWCharType();
127   WIntType = HostTarget->getWIntType();
128   Char16Type = HostTarget->getChar16Type();
129   Char32Type = HostTarget->getChar32Type();
130   Int64Type = HostTarget->getInt64Type();
131   SigAtomicType = HostTarget->getSigAtomicType();
132   ProcessIDType = HostTarget->getProcessIDType();
133 
134   UseBitFieldTypeAlignment = HostTarget->useBitFieldTypeAlignment();
135   UseZeroLengthBitfieldAlignment = HostTarget->useZeroLengthBitfieldAlignment();
136   UseExplicitBitFieldAlignment = HostTarget->useExplicitBitFieldAlignment();
137   ZeroLengthBitfieldBoundary = HostTarget->getZeroLengthBitfieldBoundary();
138 
139   // This is a bit of a lie, but it controls __GCC_ATOMIC_XXX_LOCK_FREE, and
140   // we need those macros to be identical on host and device, because (among
141   // other things) they affect which standard library classes are defined, and
142   // we need all classes to be defined on both the host and device.
143   MaxAtomicInlineWidth = HostTarget->getMaxAtomicInlineWidth();
144 
145   // Properties intentionally not copied from host:
146   // - LargeArrayMinWidth, LargeArrayAlign: Not visible across the
147   //   host/device boundary.
148   // - SuitableAlign: Not visible across the host/device boundary, and may
149   //   correctly be different on host/device, e.g. if host has wider vector
150   //   types than device.
151   // - LongDoubleWidth, LongDoubleAlign: nvptx's long double type is the same
152   //   as its double type, but that's not necessarily true on the host.
153   //   TODO: nvcc emits a warning when using long double on device; we should
154   //   do the same.
155 }
156 
157 ArrayRef<const char *> NVPTXTargetInfo::getGCCRegNames() const {
158   return llvm::makeArrayRef(GCCRegNames);
159 }
160 
161 bool NVPTXTargetInfo::hasFeature(StringRef Feature) const {
162   return llvm::StringSwitch<bool>(Feature)
163       .Cases("ptx", "nvptx", true)
164       .Default(false);
165 }
166 
167 void NVPTXTargetInfo::getTargetDefines(const LangOptions &Opts,
168                                        MacroBuilder &Builder) const {
169   Builder.defineMacro("__PTX__");
170   Builder.defineMacro("__NVPTX__");
171   if (Opts.CUDAIsDevice) {
172     // Set __CUDA_ARCH__ for the GPU specified.
173     std::string CUDAArchCode = [this] {
174       switch (GPU) {
175       case CudaArch::GFX600:
176       case CudaArch::GFX601:
177       case CudaArch::GFX700:
178       case CudaArch::GFX701:
179       case CudaArch::GFX702:
180       case CudaArch::GFX703:
181       case CudaArch::GFX704:
182       case CudaArch::GFX801:
183       case CudaArch::GFX802:
184       case CudaArch::GFX803:
185       case CudaArch::GFX810:
186       case CudaArch::GFX900:
187       case CudaArch::GFX902:
188       case CudaArch::LAST:
189         break;
190       case CudaArch::UNKNOWN:
191         assert(false && "No GPU arch when compiling CUDA device code.");
192         return "";
193       case CudaArch::SM_20:
194         return "200";
195       case CudaArch::SM_21:
196         return "210";
197       case CudaArch::SM_30:
198         return "300";
199       case CudaArch::SM_32:
200         return "320";
201       case CudaArch::SM_35:
202         return "350";
203       case CudaArch::SM_37:
204         return "370";
205       case CudaArch::SM_50:
206         return "500";
207       case CudaArch::SM_52:
208         return "520";
209       case CudaArch::SM_53:
210         return "530";
211       case CudaArch::SM_60:
212         return "600";
213       case CudaArch::SM_61:
214         return "610";
215       case CudaArch::SM_62:
216         return "620";
217       case CudaArch::SM_70:
218         return "700";
219       case CudaArch::SM_72:
220         return "720";
221       }
222       llvm_unreachable("unhandled CudaArch");
223     }();
224     Builder.defineMacro("__CUDA_ARCH__", CUDAArchCode);
225   }
226 }
227 
228 ArrayRef<Builtin::Info> NVPTXTargetInfo::getTargetBuiltins() const {
229   return llvm::makeArrayRef(BuiltinInfo, clang::NVPTX::LastTSBuiltin -
230                                              Builtin::FirstTSBuiltin);
231 }
232