AMDGPU/SI: Add debugging subtarget feature for DS offsets
[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 FeatureFlatAddressSpace : SubtargetFeature<"flat-address-space",
112         "FlatAddressSpace",
113         "true",
114         "Support flat address space">;
115
116 def FeatureVGPRSpilling : SubtargetFeature<"vgpr-spilling",
117         "EnableVGPRSpilling",
118         "true",
119         "Enable spilling of VGPRs to scratch memory">;
120
121 def FeatureSGPRInitBug : SubtargetFeature<"sgpr-init-bug",
122         "SGPRInitBug",
123         "true",
124         "VI SGPR initilization bug requiring a fixed SGPR allocation size">;
125
126 class SubtargetFeatureFetchLimit <string Value> :
127                           SubtargetFeature <"fetch"#Value,
128         "TexVTXClauseSize",
129         Value,
130         "Limit the maximum number of fetches in a clause to "#Value>;
131
132 def FeatureFetchLimit8 : SubtargetFeatureFetchLimit <"8">;
133 def FeatureFetchLimit16 : SubtargetFeatureFetchLimit <"16">;
134
135 class SubtargetFeatureWavefrontSize <int Value> : SubtargetFeature<
136         "wavefrontsize"#Value,
137         "WavefrontSize",
138         !cast<string>(Value),
139         "The number of threads per wavefront">;
140
141 def FeatureWavefrontSize16 : SubtargetFeatureWavefrontSize<16>;
142 def FeatureWavefrontSize32 : SubtargetFeatureWavefrontSize<32>;
143 def FeatureWavefrontSize64 : SubtargetFeatureWavefrontSize<64>;
144
145 class SubtargetFeatureLDSBankCount <int Value> : SubtargetFeature <
146       "ldsbankcount"#Value,
147       "LDSBankCount",
148       !cast<string>(Value),
149       "The number of LDS banks per compute unit.">;
150
151 def FeatureLDSBankCount16 : SubtargetFeatureLDSBankCount<16>;
152 def FeatureLDSBankCount32 : SubtargetFeatureLDSBankCount<32>;
153
154 class SubtargetFeatureISAVersion <int Major, int Minor, int Stepping>
155                                  : SubtargetFeature <
156       "isaver"#Major#"."#Minor#"."#Stepping,
157       "IsaVersion",
158       "ISAVersion"#Major#"_"#Minor#"_"#Stepping,
159       "Instruction set version number"
160 >;
161
162 def FeatureISAVersion7_0_0 : SubtargetFeatureISAVersion <7,0,0>;
163 def FeatureISAVersion7_0_1 : SubtargetFeatureISAVersion <7,0,1>;
164 def FeatureISAVersion8_0_0 : SubtargetFeatureISAVersion <8,0,0>;
165 def FeatureISAVersion8_0_1 : SubtargetFeatureISAVersion <8,0,1>;
166
167 class SubtargetFeatureLocalMemorySize <int Value> : SubtargetFeature<
168         "localmemorysize"#Value,
169         "LocalMemorySize",
170         !cast<string>(Value),
171         "The size of local memory in bytes">;
172
173 def FeatureGCN : SubtargetFeature<"gcn",
174         "IsGCN",
175         "true",
176         "GCN or newer GPU">;
177
178 def FeatureGCN1Encoding : SubtargetFeature<"gcn1-encoding",
179         "GCN1Encoding",
180         "true",
181         "Encoding format for SI and CI">;
182
183 def FeatureGCN3Encoding : SubtargetFeature<"gcn3-encoding",
184         "GCN3Encoding",
185         "true",
186         "Encoding format for VI">;
187
188 def FeatureCIInsts : SubtargetFeature<"ci-insts",
189         "CIInsts",
190         "true",
191         "Additional intstructions for CI+">;
192
193 // Dummy feature used to disable assembler instructions.
194 def FeatureDisable : SubtargetFeature<"",
195                                       "FeatureDisable","true",
196                                       "Dummy feature to disable assembler"
197                                       " instructions">;
198
199 class SubtargetFeatureGeneration <string Value,
200                                   list<SubtargetFeature> Implies> :
201         SubtargetFeature <Value, "Gen", "AMDGPUSubtarget::"#Value,
202                           Value#" GPU generation", Implies>;
203
204 def FeatureLocalMemorySize0 : SubtargetFeatureLocalMemorySize<0>;
205 def FeatureLocalMemorySize32768 : SubtargetFeatureLocalMemorySize<32768>;
206 def FeatureLocalMemorySize65536 : SubtargetFeatureLocalMemorySize<65536>;
207
208 def FeatureR600 : SubtargetFeatureGeneration<"R600",
209         [FeatureR600ALUInst, FeatureFetchLimit8, FeatureLocalMemorySize0]>;
210
211 def FeatureR700 : SubtargetFeatureGeneration<"R700",
212         [FeatureFetchLimit16, FeatureLocalMemorySize0]>;
213
214 def FeatureEvergreen : SubtargetFeatureGeneration<"EVERGREEN",
215         [FeatureFetchLimit16, FeatureLocalMemorySize32768]>;
216
217 def FeatureNorthernIslands : SubtargetFeatureGeneration<"NORTHERN_ISLANDS",
218         [FeatureFetchLimit16, FeatureWavefrontSize64,
219          FeatureLocalMemorySize32768]
220 >;
221
222 def FeatureSouthernIslands : SubtargetFeatureGeneration<"SOUTHERN_ISLANDS",
223         [Feature64BitPtr, FeatureFP64, FeatureLocalMemorySize32768,
224          FeatureWavefrontSize64, FeatureGCN, FeatureGCN1Encoding,
225          FeatureLDSBankCount32]>;
226
227 def FeatureSeaIslands : SubtargetFeatureGeneration<"SEA_ISLANDS",
228         [Feature64BitPtr, FeatureFP64, FeatureLocalMemorySize65536,
229          FeatureWavefrontSize64, FeatureGCN, FeatureFlatAddressSpace,
230          FeatureGCN1Encoding, FeatureCIInsts]>;
231
232 def FeatureVolcanicIslands : SubtargetFeatureGeneration<"VOLCANIC_ISLANDS",
233         [Feature64BitPtr, FeatureFP64, FeatureLocalMemorySize65536,
234          FeatureWavefrontSize64, FeatureFlatAddressSpace, FeatureGCN,
235          FeatureGCN3Encoding, FeatureCIInsts, FeatureLDSBankCount32]>;
236
237 //===----------------------------------------------------------------------===//
238
239 def AMDGPUInstrInfo : InstrInfo {
240   let guessInstructionProperties = 1;
241   let noNamedPositionallyEncodedOperands = 1;
242 }
243
244 def AMDGPUAsmParser : AsmParser {
245   // Some of the R600 registers have the same name, so this crashes.
246   // For example T0_XYZW and T0_XY both have the asm name T0.
247   let ShouldEmitMatchRegisterName = 0;
248 }
249
250 def AMDGPU : Target {
251   // Pull in Instruction Info:
252   let InstructionSet = AMDGPUInstrInfo;
253   let AssemblyParsers = [AMDGPUAsmParser];
254 }
255
256 // Dummy Instruction itineraries for pseudo instructions
257 def ALU_NULL : FuncUnit;
258 def NullALU : InstrItinClass;
259
260 //===----------------------------------------------------------------------===//
261 // Predicate helper class
262 //===----------------------------------------------------------------------===//
263
264 def TruePredicate : Predicate<"true">;
265 def isSICI : Predicate<
266   "Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS ||"
267   "Subtarget->getGeneration() == AMDGPUSubtarget::SEA_ISLANDS"
268 >, AssemblerPredicate<"FeatureGCN1Encoding">;
269
270 class PredicateControl {
271   Predicate SubtargetPredicate;
272   Predicate SIAssemblerPredicate = isSICI;
273   list<Predicate> AssemblerPredicates = [];
274   Predicate AssemblerPredicate = TruePredicate;
275   list<Predicate> OtherPredicates = [];
276   list<Predicate> Predicates = !listconcat([SubtargetPredicate, AssemblerPredicate],
277                                             AssemblerPredicates,
278                                             OtherPredicates);
279 }
280
281 // Include AMDGPU TD files
282 include "R600Schedule.td"
283 include "SISchedule.td"
284 include "Processors.td"
285 include "AMDGPUInstrInfo.td"
286 include "AMDGPUIntrinsics.td"
287 include "AMDGPURegisterInfo.td"
288 include "AMDGPUInstructions.td"
289 include "AMDGPUCallingConv.td"