1 //===-- SIRegisterInfo.td - SI Register defs ---------------*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 //===----------------------------------------------------------------------===//
11 // Declarations that describe the SI registers
12 //===----------------------------------------------------------------------===//
13 class SIReg <string n, bits<16> regIdx = 0> : Register<n>,
14 DwarfRegNum<[!cast<int>(HWEncoding)]> {
15 let Namespace = "AMDGPU";
17 // This is the not yet the complete register encoding. An additional
18 // bit is set for VGPRs.
19 let HWEncoding = regIdx;
23 def VCC_LO : SIReg<"vcc_lo", 106>;
24 def VCC_HI : SIReg<"vcc_hi", 107>;
26 // Pseudo-registers: Used as placeholders during isel and immediately
27 // replaced, never seeing the verifier.
28 def PRIVATE_RSRC_REG : SIReg<"", 0>;
29 def FP_REG : SIReg<"", 0>;
30 def SP_REG : SIReg<"", 0>;
31 def SCRATCH_WAVE_OFFSET_REG : SIReg<"", 0>;
33 // VCC for 64-bit instructions
34 def VCC : RegisterWithSubRegs<"vcc", [VCC_LO, VCC_HI]>,
35 DwarfRegAlias<VCC_LO> {
36 let Namespace = "AMDGPU";
37 let SubRegIndices = [sub0, sub1];
41 def EXEC_LO : SIReg<"exec_lo", 126>;
42 def EXEC_HI : SIReg<"exec_hi", 127>;
44 def EXEC : RegisterWithSubRegs<"EXEC", [EXEC_LO, EXEC_HI]>,
45 DwarfRegAlias<EXEC_LO> {
46 let Namespace = "AMDGPU";
47 let SubRegIndices = [sub0, sub1];
51 def SCC : SIReg<"scc", 253>;
52 def M0 : SIReg <"m0", 124>;
54 def SRC_SHARED_BASE : SIReg<"src_shared_base", 235>;
55 def SRC_SHARED_LIMIT : SIReg<"src_shared_limit", 236>;
56 def SRC_PRIVATE_BASE : SIReg<"src_private_base", 237>;
57 def SRC_PRIVATE_LIMIT : SIReg<"src_private_limit", 238>;
59 // Trap handler registers
60 def TBA_LO : SIReg<"tba_lo", 108>;
61 def TBA_HI : SIReg<"tba_hi", 109>;
63 def TBA : RegisterWithSubRegs<"tba", [TBA_LO, TBA_HI]>,
64 DwarfRegAlias<TBA_LO> {
65 let Namespace = "AMDGPU";
66 let SubRegIndices = [sub0, sub1];
70 def TMA_LO : SIReg<"tma_lo", 110>;
71 def TMA_HI : SIReg<"tma_hi", 111>;
73 def TMA : RegisterWithSubRegs<"tma", [TMA_LO, TMA_HI]>,
74 DwarfRegAlias<TMA_LO> {
75 let Namespace = "AMDGPU";
76 let SubRegIndices = [sub0, sub1];
80 def TTMP0 : SIReg <"ttmp0", 112>;
81 def TTMP1 : SIReg <"ttmp1", 113>;
82 def TTMP2 : SIReg <"ttmp2", 114>;
83 def TTMP3 : SIReg <"ttmp3", 115>;
84 def TTMP4 : SIReg <"ttmp4", 116>;
85 def TTMP5 : SIReg <"ttmp5", 117>;
86 def TTMP6 : SIReg <"ttmp6", 118>;
87 def TTMP7 : SIReg <"ttmp7", 119>;
88 def TTMP8 : SIReg <"ttmp8", 120>;
89 def TTMP9 : SIReg <"ttmp9", 121>;
90 def TTMP10 : SIReg <"ttmp10", 122>;
91 def TTMP11 : SIReg <"ttmp11", 123>;
93 multiclass FLAT_SCR_LOHI_m <string n, bits<16> ci_e, bits<16> vi_e> {
94 def _ci : SIReg<n, ci_e>;
95 def _vi : SIReg<n, vi_e>;
96 def "" : SIReg<"", 0>;
99 class FlatReg <Register lo, Register hi, bits<16> encoding> :
100 RegisterWithSubRegs<"flat_scratch", [lo, hi]>,
102 let Namespace = "AMDGPU";
103 let SubRegIndices = [sub0, sub1];
104 let HWEncoding = encoding;
107 defm FLAT_SCR_LO : FLAT_SCR_LOHI_m<"flat_scratch_lo", 104, 102>; // Offset in units of 256-bytes.
108 defm FLAT_SCR_HI : FLAT_SCR_LOHI_m<"flat_scratch_hi", 105, 103>; // Size is the per-thread scratch size, in bytes.
110 def FLAT_SCR_ci : FlatReg<FLAT_SCR_LO_ci, FLAT_SCR_HI_ci, 104>;
111 def FLAT_SCR_vi : FlatReg<FLAT_SCR_LO_vi, FLAT_SCR_HI_vi, 102>;
112 def FLAT_SCR : FlatReg<FLAT_SCR_LO, FLAT_SCR_HI, 0>;
115 foreach Index = 0-103 in {
116 def SGPR#Index : SIReg <"SGPR"#Index, Index>;
120 foreach Index = 0-255 in {
121 def VGPR#Index : SIReg <"VGPR"#Index, Index> {
122 let HWEncoding{8} = 1;
126 //===----------------------------------------------------------------------===//
127 // Groupings using register classes and tuples
128 //===----------------------------------------------------------------------===//
130 def SCC_CLASS : RegisterClass<"AMDGPU", [i1], 1, (add SCC)> {
132 let isAllocatable = 0;
135 def M0_CLASS : RegisterClass<"AMDGPU", [i32], 32, (add M0)> {
137 let isAllocatable = 0;
140 // TODO: Do we need to set DwarfRegAlias on register tuples?
142 // SGPR 32-bit registers
143 def SGPR_32 : RegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
144 (add (sequence "SGPR%u", 0, 103))> {
145 // Give all SGPR classes higher priority than VGPR classes, because
146 // we want to spill SGPRs to VGPRs.
147 let AllocationPriority = 7;
150 // SGPR 64-bit registers
151 def SGPR_64Regs : RegisterTuples<[sub0, sub1],
152 [(add (decimate SGPR_32, 2)),
153 (add (decimate (shl SGPR_32, 1), 2))]>;
155 // SGPR 128-bit registers
156 def SGPR_128Regs : RegisterTuples<[sub0, sub1, sub2, sub3],
157 [(add (decimate SGPR_32, 4)),
158 (add (decimate (shl SGPR_32, 1), 4)),
159 (add (decimate (shl SGPR_32, 2), 4)),
160 (add (decimate (shl SGPR_32, 3), 4))]>;
162 // SGPR 256-bit registers
163 def SGPR_256 : RegisterTuples<[sub0, sub1, sub2, sub3, sub4, sub5, sub6, sub7],
164 [(add (decimate SGPR_32, 4)),
165 (add (decimate (shl SGPR_32, 1), 4)),
166 (add (decimate (shl SGPR_32, 2), 4)),
167 (add (decimate (shl SGPR_32, 3), 4)),
168 (add (decimate (shl SGPR_32, 4), 4)),
169 (add (decimate (shl SGPR_32, 5), 4)),
170 (add (decimate (shl SGPR_32, 6), 4)),
171 (add (decimate (shl SGPR_32, 7), 4))]>;
173 // SGPR 512-bit registers
174 def SGPR_512 : RegisterTuples<[sub0, sub1, sub2, sub3, sub4, sub5, sub6, sub7,
175 sub8, sub9, sub10, sub11, sub12, sub13, sub14, sub15],
176 [(add (decimate SGPR_32, 4)),
177 (add (decimate (shl SGPR_32, 1), 4)),
178 (add (decimate (shl SGPR_32, 2), 4)),
179 (add (decimate (shl SGPR_32, 3), 4)),
180 (add (decimate (shl SGPR_32, 4), 4)),
181 (add (decimate (shl SGPR_32, 5), 4)),
182 (add (decimate (shl SGPR_32, 6), 4)),
183 (add (decimate (shl SGPR_32, 7), 4)),
184 (add (decimate (shl SGPR_32, 8), 4)),
185 (add (decimate (shl SGPR_32, 9), 4)),
186 (add (decimate (shl SGPR_32, 10), 4)),
187 (add (decimate (shl SGPR_32, 11), 4)),
188 (add (decimate (shl SGPR_32, 12), 4)),
189 (add (decimate (shl SGPR_32, 13), 4)),
190 (add (decimate (shl SGPR_32, 14), 4)),
191 (add (decimate (shl SGPR_32, 15), 4))]>;
193 // Trap handler TMP 32-bit registers
194 def TTMP_32 : RegisterClass<"AMDGPU", [i32, f32, v2i16, v2f16], 32,
195 (add (sequence "TTMP%u", 0, 11))> {
196 let isAllocatable = 0;
199 // Trap handler TMP 64-bit registers
200 def TTMP_64Regs : RegisterTuples<[sub0, sub1],
201 [(add (decimate TTMP_32, 2)),
202 (add (decimate (shl TTMP_32, 1), 2))]>;
204 // Trap handler TMP 128-bit registers
205 def TTMP_128Regs : RegisterTuples<[sub0, sub1, sub2, sub3],
206 [(add (decimate TTMP_32, 4)),
207 (add (decimate (shl TTMP_32, 1), 4)),
208 (add (decimate (shl TTMP_32, 2), 4)),
209 (add (decimate (shl TTMP_32, 3), 4))]>;
211 // VGPR 32-bit registers
212 // i16/f16 only on VI+
213 def VGPR_32 : RegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
214 (add (sequence "VGPR%u", 0, 255))> {
215 let AllocationPriority = 1;
219 // VGPR 64-bit registers
220 def VGPR_64 : RegisterTuples<[sub0, sub1],
221 [(add (trunc VGPR_32, 255)),
222 (add (shl VGPR_32, 1))]>;
224 // VGPR 96-bit registers
225 def VGPR_96 : RegisterTuples<[sub0, sub1, sub2],
226 [(add (trunc VGPR_32, 254)),
227 (add (shl VGPR_32, 1)),
228 (add (shl VGPR_32, 2))]>;
230 // VGPR 128-bit registers
231 def VGPR_128 : RegisterTuples<[sub0, sub1, sub2, sub3],
232 [(add (trunc VGPR_32, 253)),
233 (add (shl VGPR_32, 1)),
234 (add (shl VGPR_32, 2)),
235 (add (shl VGPR_32, 3))]>;
237 // VGPR 256-bit registers
238 def VGPR_256 : RegisterTuples<[sub0, sub1, sub2, sub3, sub4, sub5, sub6, sub7],
239 [(add (trunc VGPR_32, 249)),
240 (add (shl VGPR_32, 1)),
241 (add (shl VGPR_32, 2)),
242 (add (shl VGPR_32, 3)),
243 (add (shl VGPR_32, 4)),
244 (add (shl VGPR_32, 5)),
245 (add (shl VGPR_32, 6)),
246 (add (shl VGPR_32, 7))]>;
248 // VGPR 512-bit registers
249 def VGPR_512 : RegisterTuples<[sub0, sub1, sub2, sub3, sub4, sub5, sub6, sub7,
250 sub8, sub9, sub10, sub11, sub12, sub13, sub14, sub15],
251 [(add (trunc VGPR_32, 241)),
252 (add (shl VGPR_32, 1)),
253 (add (shl VGPR_32, 2)),
254 (add (shl VGPR_32, 3)),
255 (add (shl VGPR_32, 4)),
256 (add (shl VGPR_32, 5)),
257 (add (shl VGPR_32, 6)),
258 (add (shl VGPR_32, 7)),
259 (add (shl VGPR_32, 8)),
260 (add (shl VGPR_32, 9)),
261 (add (shl VGPR_32, 10)),
262 (add (shl VGPR_32, 11)),
263 (add (shl VGPR_32, 12)),
264 (add (shl VGPR_32, 13)),
265 (add (shl VGPR_32, 14)),
266 (add (shl VGPR_32, 15))]>;
268 //===----------------------------------------------------------------------===//
269 // Register classes used as source and destination
270 //===----------------------------------------------------------------------===//
272 // Subset of SReg_32 without M0 for SMRD instructions and alike.
273 // See comments in SIInstructions.td for more info.
274 def SReg_32_XM0_XEXEC : RegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
275 (add SGPR_32, VCC_LO, VCC_HI, FLAT_SCR_LO, FLAT_SCR_HI,
276 TTMP_32, TMA_LO, TMA_HI, TBA_LO, TBA_HI, SRC_SHARED_BASE, SRC_SHARED_LIMIT,
277 SRC_PRIVATE_BASE, SRC_PRIVATE_LIMIT,
278 FP_REG, SP_REG, SCRATCH_WAVE_OFFSET_REG)> {
279 let AllocationPriority = 7;
282 def SReg_32_XM0 : RegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
283 (add SReg_32_XM0_XEXEC, EXEC_LO, EXEC_HI)> {
284 let AllocationPriority = 7;
287 // Register class for all scalar registers (SGPRs + Special Registers)
288 def SReg_32 : RegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
289 (add SReg_32_XM0, M0_CLASS, EXEC_LO, EXEC_HI)> {
290 let AllocationPriority = 7;
293 def SGPR_64 : RegisterClass<"AMDGPU", [v2i32, i64, f64], 32, (add SGPR_64Regs)> {
295 let AllocationPriority = 8;
298 def TTMP_64 : RegisterClass<"AMDGPU", [v2i32, i64, f64], 32, (add TTMP_64Regs)> {
299 let isAllocatable = 0;
302 def SReg_64_XEXEC : RegisterClass<"AMDGPU", [v2i32, i64, f64, i1], 32,
303 (add SGPR_64, VCC, FLAT_SCR, TTMP_64, TBA, TMA)> {
305 let AllocationPriority = 8;
308 def SReg_64 : RegisterClass<"AMDGPU", [v2i32, i64, f64, i1], 32,
309 (add SReg_64_XEXEC, EXEC)> {
311 let AllocationPriority = 8;
314 // Requires 2 s_mov_b64 to copy
315 let CopyCost = 2 in {
317 def SGPR_128 : RegisterClass<"AMDGPU", [v4i32, v16i8, v2i64], 32, (add SGPR_128Regs)> {
318 let AllocationPriority = 10;
321 def TTMP_128 : RegisterClass<"AMDGPU", [v4i32, v16i8, v2i64], 32, (add TTMP_128Regs)> {
322 let isAllocatable = 0;
325 def SReg_128 : RegisterClass<"AMDGPU", [v4i32, v16i8, v2i64], 32,
326 (add SGPR_128, TTMP_128)> {
327 let AllocationPriority = 10;
330 } // End CopyCost = 2
332 def SReg_256 : RegisterClass<"AMDGPU", [v8i32, v8f32], 32, (add SGPR_256)> {
333 // Requires 4 s_mov_b64 to copy
335 let AllocationPriority = 11;
338 def SReg_512 : RegisterClass<"AMDGPU", [v16i32, v16f32], 32, (add SGPR_512)> {
339 // Requires 8 s_mov_b64 to copy
341 let AllocationPriority = 12;
344 // Register class for all vector registers (VGPRs + Interploation Registers)
345 def VReg_64 : RegisterClass<"AMDGPU", [i64, f64, v2i32, v2f32], 32, (add VGPR_64)> {
348 // Requires 2 v_mov_b32 to copy
350 let AllocationPriority = 2;
353 def VReg_96 : RegisterClass<"AMDGPU", [untyped], 32, (add VGPR_96)> {
356 // Requires 3 v_mov_b32 to copy
358 let AllocationPriority = 3;
361 def VReg_128 : RegisterClass<"AMDGPU", [v4i32, v4f32, v2i64, v2f64], 32, (add VGPR_128)> {
364 // Requires 4 v_mov_b32 to copy
366 let AllocationPriority = 4;
369 def VReg_256 : RegisterClass<"AMDGPU", [v8i32, v8f32], 32, (add VGPR_256)> {
372 let AllocationPriority = 5;
375 def VReg_512 : RegisterClass<"AMDGPU", [v16i32, v16f32], 32, (add VGPR_512)> {
378 let AllocationPriority = 6;
381 def VReg_1 : RegisterClass<"AMDGPU", [i1], 32, (add VGPR_32)> {
385 def VS_32 : RegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
386 (add VGPR_32, SReg_32)> {
387 let isAllocatable = 0;
390 def VS_64 : RegisterClass<"AMDGPU", [i64, f64], 32, (add VReg_64, SReg_64)> {
391 let isAllocatable = 0;
394 //===----------------------------------------------------------------------===//
396 //===----------------------------------------------------------------------===//
398 class RegImmMatcher<string name> : AsmOperandClass {
400 let RenderMethod = "addRegOrImmOperands";
403 multiclass SIRegOperand <string rc, string MatchName, string opType> {
404 let OperandNamespace = "AMDGPU" in {
405 def _b16 : RegisterOperand<!cast<RegisterClass>(rc#"_32")> {
406 let OperandType = opType#"_INT16";
407 let ParserMatchClass = RegImmMatcher<MatchName#"B16">;
408 let DecoderMethod = "decodeOperand_VSrc16";
411 def _f16 : RegisterOperand<!cast<RegisterClass>(rc#"_32")> {
412 let OperandType = opType#"_FP16";
413 let ParserMatchClass = RegImmMatcher<MatchName#"F16">;
414 let DecoderMethod = "decodeOperand_VSrc16";
417 def _b32 : RegisterOperand<!cast<RegisterClass>(rc#"_32")> {
418 let OperandType = opType#"_INT32";
419 let ParserMatchClass = RegImmMatcher<MatchName#"B32">;
422 def _f32 : RegisterOperand<!cast<RegisterClass>(rc#"_32")> {
423 let OperandType = opType#"_FP32";
424 let ParserMatchClass = RegImmMatcher<MatchName#"F32">;
427 def _b64 : RegisterOperand<!cast<RegisterClass>(rc#"_64")> {
428 let OperandType = opType#"_INT64";
429 let ParserMatchClass = RegImmMatcher<MatchName#"B64">;
432 def _f64 : RegisterOperand<!cast<RegisterClass>(rc#"_64")> {
433 let OperandType = opType#"_FP64";
434 let ParserMatchClass = RegImmMatcher<MatchName#"F64">;
437 def _v2b16 : RegisterOperand<!cast<RegisterClass>(rc#"_32")> {
438 let OperandType = opType#"_V2INT16";
439 let ParserMatchClass = RegImmMatcher<MatchName#"V2B16">;
440 let DecoderMethod = "decodeOperand_VSrcV216";
443 def _v2f16 : RegisterOperand<!cast<RegisterClass>(rc#"_32")> {
444 let OperandType = opType#"_V2FP16";
445 let ParserMatchClass = RegImmMatcher<MatchName#"V2F16">;
446 let DecoderMethod = "decodeOperand_VSrcV216";
451 // FIXME: 64-bit sources can sometimes use 32-bit constants.
452 multiclass RegImmOperand <string rc, string MatchName>
453 : SIRegOperand<rc, MatchName, "OPERAND_REG_IMM">;
455 multiclass RegInlineOperand <string rc, string MatchName>
456 : SIRegOperand<rc, MatchName, "OPERAND_REG_INLINE_C">;
458 //===----------------------------------------------------------------------===//
459 // SSrc_* Operands with an SGPR or a 32-bit immediate
460 //===----------------------------------------------------------------------===//
462 defm SSrc : RegImmOperand<"SReg", "SSrc">;
464 //===----------------------------------------------------------------------===//
465 // SCSrc_* Operands with an SGPR or a inline constant
466 //===----------------------------------------------------------------------===//
468 defm SCSrc : RegInlineOperand<"SReg", "SCSrc"> ;
470 //===----------------------------------------------------------------------===//
471 // VSrc_* Operands with an SGPR, VGPR or a 32-bit immediate
472 //===----------------------------------------------------------------------===//
474 defm VSrc : RegImmOperand<"VS", "VSrc">;
476 def VSrc_128 : RegisterOperand<VReg_128> {
477 let DecoderMethod = "DecodeVS_128RegisterClass";
480 //===----------------------------------------------------------------------===//
481 // VSrc_* Operands with an VGPR
482 //===----------------------------------------------------------------------===//
484 // This is for operands with the enum(9), VSrc encoding restriction,
485 // but only allows VGPRs.
486 def VRegSrc_32 : RegisterOperand<VGPR_32> {
487 //let ParserMatchClass = RegImmMatcher<"VRegSrc32">;
488 let DecoderMethod = "DecodeVS_32RegisterClass";
491 //===----------------------------------------------------------------------===//
492 // VCSrc_* Operands with an SGPR, VGPR or an inline constant
493 //===----------------------------------------------------------------------===//
495 defm VCSrc : RegInlineOperand<"VS", "VCSrc">;