5913bf7f773f4e497573ba90ad88c96275daa180
[oota-llvm.git] / lib / Target / AMDGPU / AMDGPU.td
1 //===-- AMDGPU.td - AMDGPU Tablegen files ------------------*- tablegen -*-===//
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 include "llvm/Target/Target.td"
11
12 //===----------------------------------------------------------------------===//
13 // Subtarget Features
14 //===----------------------------------------------------------------------===//
15
16 // Debugging Features
17
18 def FeatureDumpCode : SubtargetFeature <"DumpCode",
19         "DumpCode",
20         "true",
21         "Dump MachineInstrs in the CodeEmitter">;
22
23 def FeatureDumpCodeLower : SubtargetFeature <"dumpcode",
24         "DumpCode",
25         "true",
26         "Dump MachineInstrs in the CodeEmitter">;
27
28 def FeatureIRStructurizer : SubtargetFeature <"disable-irstructurizer",
29         "EnableIRStructurizer",
30         "false",
31         "Disable IR Structurizer">;
32
33 def FeaturePromoteAlloca : SubtargetFeature <"promote-alloca",
34         "EnablePromoteAlloca",
35         "true",
36         "Enable promote alloca pass">;
37
38 // Target features
39
40 def FeatureIfCvt : SubtargetFeature <"disable-ifcvt",
41         "EnableIfCvt",
42         "false",
43         "Disable the if conversion pass">;
44
45 def FeatureFP64 : SubtargetFeature<"fp64",
46         "FP64",
47         "true",
48         "Enable double precision operations">;
49
50 def FeatureFP64Denormals : SubtargetFeature<"fp64-denormals",
51         "FP64Denormals",
52         "true",
53         "Enable double precision denormal handling",
54         [FeatureFP64]>;
55
56 def FeatureFastFMAF32 : SubtargetFeature<"fast-fmaf",
57         "FastFMAF32",
58         "true",
59         "Assuming f32 fma is at least as fast as mul + add",
60         []>;
61
62 // Some instructions do not support denormals despite this flag. Using
63 // fp32 denormals also causes instructions to run at the double
64 // precision rate for the device.
65 def FeatureFP32Denormals : SubtargetFeature<"fp32-denormals",
66         "FP32Denormals",
67         "true",
68         "Enable single precision denormal handling">;
69
70 def Feature64BitPtr : SubtargetFeature<"64BitPtr",
71         "Is64bit",
72         "true",
73         "Specify if 64-bit addressing should be used">;
74
75 def FeatureR600ALUInst : SubtargetFeature<"R600ALUInst",
76         "R600ALUInst",
77         "false",
78         "Older version of ALU instructions encoding">;
79
80 def FeatureVertexCache : SubtargetFeature<"HasVertexCache",
81         "HasVertexCache",
82         "true",
83         "Specify use of dedicated vertex cache">;
84
85 def FeatureCaymanISA : SubtargetFeature<"caymanISA",
86         "CaymanISA",
87         "true",
88         "Use Cayman ISA">;
89
90 def FeatureCFALUBug : SubtargetFeature<"cfalubug",
91         "CFALUBug",
92         "true",
93         "GPU has CF_ALU bug">;
94
95 // XXX - This should probably be removed once enabled by default
96 def FeatureEnableLoadStoreOpt : SubtargetFeature <"load-store-opt",
97         "EnableLoadStoreOpt",
98         "true",
99         "Enable SI load/store optimizer pass">;
100
101 // Performance debugging feature. Allow using DS instruction immediate
102 // offsets even if the base pointer can't be proven to be base. On SI,
103 // base pointer values that won't give the same result as a 16-bit add
104 // are not safe to fold, but this will override the conservative test
105 // for the base pointer.
106 def FeatureEnableUnsafeDSOffsetFolding : SubtargetFeature <"unsafe-ds-offset-folding",
107         "EnableUnsafeDSOffsetFolding",
108         "true",
109         "Force using DS instruction immediate offsets on SI">;
110
111 def FeatureFlatForGlobal : SubtargetFeature<"flat-for-global",
112         "FlatForGlobal",
113         "true",
114         "Force to generate flat instruction for global">;
115
116 def FeatureFlatAddressSpace : SubtargetFeature<"flat-address-space",
117         "FlatAddressSpace",
118         "true",
119         "Support flat address space">;
120
121 def FeatureVGPRSpilling : SubtargetFeature<"vgpr-spilling",
122         "EnableVGPRSpilling",
123         "true",
124         "Enable spilling of VGPRs to scratch memory">;
125
126 def FeatureSGPRInitBug : SubtargetFeature<"sgpr-init-bug",
127         "SGPRInitBug",
128         "true",
129         "VI SGPR initilization bug requiring a fixed SGPR allocation size">;
130
131 def FeatureEnableHugeScratchBuffer : SubtargetFeature<"huge-scratch-buffer",
132         "EnableHugeScratchBuffer",
133         "true",
134         "Enable scratch buffer sizes greater than 128 GB">;
135
136 class SubtargetFeatureFetchLimit <string Value> :
137                           SubtargetFeature <"fetch"#Value,
138         "TexVTXClauseSize",
139         Value,
140         "Limit the maximum number of fetches in a clause to "#Value>;
141
142 def FeatureFetchLimit8 : SubtargetFeatureFetchLimit <"8">;
143 def FeatureFetchLimit16 : SubtargetFeatureFetchLimit <"16">;
144
145 class SubtargetFeatureWavefrontSize <int Value> : SubtargetFeature<
146         "wavefrontsize"#Value,
147         "WavefrontSize",
148         !cast<string>(Value),
149         "The number of threads per wavefront">;
150
151 def FeatureWavefrontSize16 : SubtargetFeatureWavefrontSize<16>;
152 def FeatureWavefrontSize32 : SubtargetFeatureWavefrontSize<32>;
153 def FeatureWavefrontSize64 : SubtargetFeatureWavefrontSize<64>;
154
155 class SubtargetFeatureLDSBankCount <int Value> : SubtargetFeature <
156       "ldsbankcount"#Value,
157       "LDSBankCount",
158       !cast<string>(Value),
159       "The number of LDS banks per compute unit.">;
160
161 def FeatureLDSBankCount16 : SubtargetFeatureLDSBankCount<16>;
162 def FeatureLDSBankCount32 : SubtargetFeatureLDSBankCount<32>;
163
164 class SubtargetFeatureISAVersion <int Major, int Minor, int Stepping>
165                                  : SubtargetFeature <
166       "isaver"#Major#"."#Minor#"."#Stepping,
167       "IsaVersion",
168       "ISAVersion"#Major#"_"#Minor#"_"#Stepping,
169       "Instruction set version number"
170 >;
171
172 def FeatureISAVersion7_0_0 : SubtargetFeatureISAVersion <7,0,0>;
173 def FeatureISAVersion7_0_1 : SubtargetFeatureISAVersion <7,0,1>;
174 def FeatureISAVersion8_0_0 : SubtargetFeatureISAVersion <8,0,0>;
175 def FeatureISAVersion8_0_1 : SubtargetFeatureISAVersion <8,0,1>;
176
177 class SubtargetFeatureLocalMemorySize <int Value> : SubtargetFeature<
178         "localmemorysize"#Value,
179         "LocalMemorySize",
180         !cast<string>(Value),
181         "The size of local memory in bytes">;
182
183 def FeatureGCN : SubtargetFeature<"gcn",
184         "IsGCN",
185         "true",
186         "GCN or newer GPU">;
187
188 def FeatureGCN1Encoding : SubtargetFeature<"gcn1-encoding",
189         "GCN1Encoding",
190         "true",
191         "Encoding format for SI and CI">;
192
193 def FeatureGCN3Encoding : SubtargetFeature<"gcn3-encoding",
194         "GCN3Encoding",
195         "true",
196         "Encoding format for VI">;
197
198 def FeatureCIInsts : SubtargetFeature<"ci-insts",
199         "CIInsts",
200         "true",
201         "Additional intstructions for CI+">;
202
203 // Dummy feature used to disable assembler instructions.
204 def FeatureDisable : SubtargetFeature<"",
205                                       "FeatureDisable","true",
206                                       "Dummy feature to disable assembler"
207                                       " instructions">;
208
209 class SubtargetFeatureGeneration <string Value,
210                                   list<SubtargetFeature> Implies> :
211         SubtargetFeature <Value, "Gen", "AMDGPUSubtarget::"#Value,
212                           Value#" GPU generation", Implies>;
213
214 def FeatureLocalMemorySize0 : SubtargetFeatureLocalMemorySize<0>;
215 def FeatureLocalMemorySize32768 : SubtargetFeatureLocalMemorySize<32768>;
216 def FeatureLocalMemorySize65536 : SubtargetFeatureLocalMemorySize<65536>;
217
218 def FeatureR600 : SubtargetFeatureGeneration<"R600",
219         [FeatureR600ALUInst, FeatureFetchLimit8, FeatureLocalMemorySize0]>;
220
221 def FeatureR700 : SubtargetFeatureGeneration<"R700",
222         [FeatureFetchLimit16, FeatureLocalMemorySize0]>;
223
224 def FeatureEvergreen : SubtargetFeatureGeneration<"EVERGREEN",
225         [FeatureFetchLimit16, FeatureLocalMemorySize32768]>;
226
227 def FeatureNorthernIslands : SubtargetFeatureGeneration<"NORTHERN_ISLANDS",
228         [FeatureFetchLimit16, FeatureWavefrontSize64,
229          FeatureLocalMemorySize32768]
230 >;
231
232 def FeatureSouthernIslands : SubtargetFeatureGeneration<"SOUTHERN_ISLANDS",
233         [Feature64BitPtr, FeatureFP64, FeatureLocalMemorySize32768,
234          FeatureWavefrontSize64, FeatureGCN, FeatureGCN1Encoding,
235          FeatureLDSBankCount32]>;
236
237 def FeatureSeaIslands : SubtargetFeatureGeneration<"SEA_ISLANDS",
238         [Feature64BitPtr, FeatureFP64, FeatureLocalMemorySize65536,
239          FeatureWavefrontSize64, FeatureGCN, FeatureFlatAddressSpace,
240          FeatureGCN1Encoding, FeatureCIInsts]>;
241
242 def FeatureVolcanicIslands : SubtargetFeatureGeneration<"VOLCANIC_ISLANDS",
243         [Feature64BitPtr, FeatureFP64, FeatureLocalMemorySize65536,
244          FeatureWavefrontSize64, FeatureFlatAddressSpace, FeatureGCN,
245          FeatureGCN3Encoding, FeatureCIInsts, FeatureLDSBankCount32]>;
246
247 //===----------------------------------------------------------------------===//
248
249 def AMDGPUInstrInfo : InstrInfo {
250   let guessInstructionProperties = 1;
251   let noNamedPositionallyEncodedOperands = 1;
252 }
253
254 def AMDGPUAsmParser : AsmParser {
255   // Some of the R600 registers have the same name, so this crashes.
256   // For example T0_XYZW and T0_XY both have the asm name T0.
257   let ShouldEmitMatchRegisterName = 0;
258 }
259
260 def AMDGPU : Target {
261   // Pull in Instruction Info:
262   let InstructionSet = AMDGPUInstrInfo;
263   let AssemblyParsers = [AMDGPUAsmParser];
264 }
265
266 // Dummy Instruction itineraries for pseudo instructions
267 def ALU_NULL : FuncUnit;
268 def NullALU : InstrItinClass;
269
270 //===----------------------------------------------------------------------===//
271 // Predicate helper class
272 //===----------------------------------------------------------------------===//
273
274 def TruePredicate : Predicate<"true">;
275 def isSICI : Predicate<
276   "Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS ||"
277   "Subtarget->getGeneration() == AMDGPUSubtarget::SEA_ISLANDS"
278 >, AssemblerPredicate<"FeatureGCN1Encoding">;
279
280 class PredicateControl {
281   Predicate SubtargetPredicate;
282   Predicate SIAssemblerPredicate = isSICI;
283   list<Predicate> AssemblerPredicates = [];
284   Predicate AssemblerPredicate = TruePredicate;
285   list<Predicate> OtherPredicates = [];
286   list<Predicate> Predicates = !listconcat([SubtargetPredicate, AssemblerPredicate],
287                                             AssemblerPredicates,
288                                             OtherPredicates);
289 }
290
291 // Include AMDGPU TD files
292 include "R600Schedule.td"
293 include "SISchedule.td"
294 include "Processors.td"
295 include "AMDGPUInstrInfo.td"
296 include "AMDGPUIntrinsics.td"
297 include "AMDGPURegisterInfo.td"
298 include "AMDGPUInstructions.td"
299 include "AMDGPUCallingConv.td"