1 //===-- SIRegisterInfo.td - SI Register defs ---------------*- tablegen -*-===//
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
7 //===----------------------------------------------------------------------===//
9 //===----------------------------------------------------------------------===//
10 // Subregister declarations
11 //===----------------------------------------------------------------------===//
13 class Indexes<int N> {
14 list<int> all = [0, 1, 2, 3, 4, 5, 6 , 7,
15 8, 9, 10, 11, 12, 13, 14, 15,
16 16, 17, 18, 19, 20, 21, 22, 23,
17 24, 25, 26, 27, 28, 29, 30, 31];
19 // Returns list of indexes [0..N)
20 list<int> slice = !filter(i, all, !lt(i, N));
23 let Namespace = "AMDGPU" in {
25 def lo16 : SubRegIndex<16, 0>;
26 def hi16 : SubRegIndex<16, 16>;
28 foreach Index = 0...31 in {
29 def sub#Index : SubRegIndex<32, !shl(Index, 5)>;
32 foreach Index = 1...31 in {
33 def sub#Index#_lo16 : ComposedSubRegIndex<!cast<SubRegIndex>(sub#Index), lo16>;
34 def sub#Index#_hi16 : ComposedSubRegIndex<!cast<SubRegIndex>(sub#Index), hi16>;
37 foreach Size = {2...6,8,16} in {
38 foreach Index = Indexes<!sub(33, Size)>.slice in {
39 def !interleave(!foreach(cur, Indexes<Size>.slice, "sub"#!add(cur, Index)),
41 SubRegIndex<!mul(Size, 32), !shl(Index, 5)> {
42 let CoveringSubRegIndices =
43 !foreach(cur, Indexes<Size>.slice,
44 !cast<SubRegIndex>(sub#!add(cur, Index)));
51 //===----------------------------------------------------------------------===//
53 //===----------------------------------------------------------------------===//
55 class getSubRegs<int size> {
56 list<SubRegIndex> ret2 = [sub0, sub1];
57 list<SubRegIndex> ret3 = [sub0, sub1, sub2];
58 list<SubRegIndex> ret4 = [sub0, sub1, sub2, sub3];
59 list<SubRegIndex> ret5 = [sub0, sub1, sub2, sub3, sub4];
60 list<SubRegIndex> ret6 = [sub0, sub1, sub2, sub3, sub4, sub5];
61 list<SubRegIndex> ret7 = [sub0, sub1, sub2, sub3, sub4, sub5, sub6];
62 list<SubRegIndex> ret8 = [sub0, sub1, sub2, sub3, sub4, sub5, sub6, sub7];
63 list<SubRegIndex> ret16 = [sub0, sub1, sub2, sub3,
64 sub4, sub5, sub6, sub7,
65 sub8, sub9, sub10, sub11,
66 sub12, sub13, sub14, sub15];
67 list<SubRegIndex> ret32 = [sub0, sub1, sub2, sub3,
68 sub4, sub5, sub6, sub7,
69 sub8, sub9, sub10, sub11,
70 sub12, sub13, sub14, sub15,
71 sub16, sub17, sub18, sub19,
72 sub20, sub21, sub22, sub23,
73 sub24, sub25, sub26, sub27,
74 sub28, sub29, sub30, sub31];
76 list<SubRegIndex> ret = !if(!eq(size, 2), ret2,
77 !if(!eq(size, 3), ret3,
78 !if(!eq(size, 4), ret4,
79 !if(!eq(size, 5), ret5,
80 !if(!eq(size, 6), ret6,
81 !if(!eq(size, 7), ret7,
82 !if(!eq(size, 8), ret8,
83 !if(!eq(size, 16), ret16,
87 // Generates list of sequential register tuple names.
88 // E.g. RegSeq<3,2,2,"s">.ret -> [ "s[0:1]", "s[2:3]" ]
89 class RegSeqNames<int last_reg, int stride, int size, string prefix,
91 int next = !add(start, stride);
92 int end_reg = !add(start, size, -1);
94 !if(!le(end_reg, last_reg),
95 !listconcat([prefix # "[" # start # ":" # end_reg # "]"],
96 RegSeqNames<last_reg, stride, size, prefix, next>.ret),
100 // Generates list of dags for register tupless.
101 class RegSeqDags<RegisterClass RC, int last_reg, int stride, int size,
103 dag trunc_rc = (trunc RC,
104 !if(!and(!eq(stride, 1), !eq(start, 0)),
105 !sub(!add(last_reg, 2), size),
108 !if(!lt(start, size),
109 !listconcat([(add (decimate (shl trunc_rc, start), stride))],
110 RegSeqDags<RC, last_reg, stride, size, !add(start, 1)>.ret),
114 class SIRegisterTuples<list<SubRegIndex> Indices, RegisterClass RC,
115 int last_reg, int stride, int size, string prefix> :
116 RegisterTuples<Indices,
117 RegSeqDags<RC, last_reg, stride, size>.ret,
118 RegSeqNames<last_reg, stride, size, prefix>.ret>;
120 //===----------------------------------------------------------------------===//
121 // Declarations that describe the SI registers
122 //===----------------------------------------------------------------------===//
123 class SIReg <string n, bits<16> regIdx = 0> :
125 let Namespace = "AMDGPU";
126 let HWEncoding = regIdx;
129 // For register classes that use TSFlags.
130 class SIRegisterClass <string n, list<ValueType> rTypes, int Align, dag rList>
131 : RegisterClass <n, rTypes, Align, rList> {
132 // For vector register classes.
133 field bit HasVGPR = 0;
134 field bit HasAGPR = 0;
136 // For scalar register classes.
137 field bit HasSGPR = 0;
139 // These need to be kept in sync with the enum SIRCFlags.
140 let TSFlags{0} = HasVGPR;
141 let TSFlags{1} = HasAGPR;
142 let TSFlags{2} = HasSGPR;
145 multiclass SIRegLoHi16 <string n, bits<16> regIdx, bit ArtificialHigh = 1,
146 bit HWEncodingHigh = 0> {
147 // There is no special encoding for 16 bit subregs, these are not real
148 // registers but rather operands for instructions preserving other 16 bits
149 // of the result or reading just 16 bits of a 32 bit VGPR.
150 // It is encoded as a corresponding 32 bit register.
151 // Non-VGPR register classes use it as we need to have matching subregisters
152 // to move instructions and data between ALUs.
153 def _LO16 : SIReg<n#".l", regIdx> {
154 let HWEncoding{8} = HWEncodingHigh;
156 def _HI16 : SIReg<!if(ArtificialHigh, "", n#".h"), regIdx> {
157 let isArtificial = ArtificialHigh;
158 let HWEncoding{8} = HWEncodingHigh;
160 def "" : RegisterWithSubRegs<n, [!cast<Register>(NAME#"_LO16"),
161 !cast<Register>(NAME#"_HI16")]> {
162 let Namespace = "AMDGPU";
163 let SubRegIndices = [lo16, hi16];
164 let CoveredBySubRegs = !not(ArtificialHigh);
165 let HWEncoding = regIdx;
166 let HWEncoding{8} = HWEncodingHigh;
171 defm VCC_LO : SIRegLoHi16<"vcc_lo", 106>;
172 defm VCC_HI : SIRegLoHi16<"vcc_hi", 107>;
174 // Pseudo-registers: Used as placeholders during isel and immediately
175 // replaced, never seeing the verifier.
176 def PRIVATE_RSRC_REG : SIReg<"private_rsrc", 0>;
177 def FP_REG : SIReg<"fp", 0>;
178 def SP_REG : SIReg<"sp", 0>;
180 // Pseudo-register to represent the program-counter DWARF register.
181 def PC_REG : SIReg<"pc", 0>, DwarfRegNum<[16, 16]> {
182 // There is no physical register corresponding to a "program counter", but
183 // we need to encode the concept in debug information in order to represent
184 // things like the return value in unwind information.
185 let isArtificial = 1;
188 // VCC for 64-bit instructions
189 def VCC : RegisterWithSubRegs<"vcc", [VCC_LO, VCC_HI]> {
190 let Namespace = "AMDGPU";
191 let SubRegIndices = [sub0, sub1];
192 let HWEncoding = 106;
195 defm EXEC_LO : SIRegLoHi16<"exec_lo", 126>, DwarfRegNum<[1, 1]>;
196 defm EXEC_HI : SIRegLoHi16<"exec_hi", 127>;
198 def EXEC : RegisterWithSubRegs<"exec", [EXEC_LO, EXEC_HI]>, DwarfRegNum<[17, 1]> {
199 let Namespace = "AMDGPU";
200 let SubRegIndices = [sub0, sub1];
201 let HWEncoding = 126;
204 // 32-bit real registers, for MC only.
205 // May be used with both 32-bit and 64-bit operands.
206 defm SRC_VCCZ : SIRegLoHi16<"src_vccz", 251>;
207 defm SRC_EXECZ : SIRegLoHi16<"src_execz", 252>;
208 defm SRC_SCC : SIRegLoHi16<"src_scc", 253>;
210 // 1-bit pseudo register, for codegen only.
211 // Should never be emitted.
212 def SCC : SIReg<"scc">;
214 defm M0 : SIRegLoHi16 <"m0", 124>;
215 defm SGPR_NULL : SIRegLoHi16 <"null", 125>;
217 defm SRC_SHARED_BASE : SIRegLoHi16<"src_shared_base", 235>;
218 defm SRC_SHARED_LIMIT : SIRegLoHi16<"src_shared_limit", 236>;
219 defm SRC_PRIVATE_BASE : SIRegLoHi16<"src_private_base", 237>;
220 defm SRC_PRIVATE_LIMIT : SIRegLoHi16<"src_private_limit", 238>;
221 defm SRC_POPS_EXITING_WAVE_ID : SIRegLoHi16<"src_pops_exiting_wave_id", 239>;
224 def MODE : SIReg <"mode", 0>;
226 def LDS_DIRECT : SIReg <"src_lds_direct", 254> {
227 // There is no physical register corresponding to this. This is an
228 // encoding value in a source field, which will ultimately trigger a
230 let isArtificial = 1;
233 defm XNACK_MASK_LO : SIRegLoHi16<"xnack_mask_lo", 104>;
234 defm XNACK_MASK_HI : SIRegLoHi16<"xnack_mask_hi", 105>;
237 RegisterWithSubRegs<"xnack_mask", [XNACK_MASK_LO, XNACK_MASK_HI]> {
238 let Namespace = "AMDGPU";
239 let SubRegIndices = [sub0, sub1];
240 let HWEncoding = 104;
243 // Trap handler registers
244 defm TBA_LO : SIRegLoHi16<"tba_lo", 108>;
245 defm TBA_HI : SIRegLoHi16<"tba_hi", 109>;
247 def TBA : RegisterWithSubRegs<"tba", [TBA_LO, TBA_HI]> {
248 let Namespace = "AMDGPU";
249 let SubRegIndices = [sub0, sub1];
250 let HWEncoding = 108;
253 defm TMA_LO : SIRegLoHi16<"tma_lo", 110>;
254 defm TMA_HI : SIRegLoHi16<"tma_hi", 111>;
256 def TMA : RegisterWithSubRegs<"tma", [TMA_LO, TMA_HI]> {
257 let Namespace = "AMDGPU";
258 let SubRegIndices = [sub0, sub1];
259 let HWEncoding = 110;
262 foreach Index = 0...15 in {
263 defm TTMP#Index#_vi : SIRegLoHi16<"ttmp"#Index, !add(112, Index)>;
264 defm TTMP#Index#_gfx9plus : SIRegLoHi16<"ttmp"#Index, !add(108, Index)>;
265 defm TTMP#Index : SIRegLoHi16<"ttmp"#Index, 0>;
268 multiclass FLAT_SCR_LOHI_m <string n, bits<16> ci_e, bits<16> vi_e> {
269 defm _ci : SIRegLoHi16<n, ci_e>;
270 defm _vi : SIRegLoHi16<n, vi_e>;
271 defm "" : SIRegLoHi16<n, 0>;
274 class FlatReg <Register lo, Register hi, bits<16> encoding> :
275 RegisterWithSubRegs<"flat_scratch", [lo, hi]> {
276 let Namespace = "AMDGPU";
277 let SubRegIndices = [sub0, sub1];
278 let HWEncoding = encoding;
281 defm FLAT_SCR_LO : FLAT_SCR_LOHI_m<"flat_scratch_lo", 104, 102>; // Offset in units of 256-bytes.
282 defm FLAT_SCR_HI : FLAT_SCR_LOHI_m<"flat_scratch_hi", 105, 103>; // Size is the per-thread scratch size, in bytes.
284 def FLAT_SCR_ci : FlatReg<FLAT_SCR_LO_ci, FLAT_SCR_HI_ci, 104>;
285 def FLAT_SCR_vi : FlatReg<FLAT_SCR_LO_vi, FLAT_SCR_HI_vi, 102>;
286 def FLAT_SCR : FlatReg<FLAT_SCR_LO, FLAT_SCR_HI, 0>;
289 foreach Index = 0...105 in {
291 SIRegLoHi16 <"s"#Index, Index>,
292 DwarfRegNum<[!if(!le(Index, 63), !add(Index, 32), !add(Index, 1024)),
293 !if(!le(Index, 63), !add(Index, 32), !add(Index, 1024))]>;
297 foreach Index = 0...255 in {
299 SIRegLoHi16 <"v"#Index, Index, 0, 1>,
300 DwarfRegNum<[!add(Index, 2560), !add(Index, 1536)]>;
304 foreach Index = 0...255 in {
306 SIRegLoHi16 <"a"#Index, Index, 1, 1>,
307 DwarfRegNum<[!add(Index, 3072), !add(Index, 2048)]>;
310 //===----------------------------------------------------------------------===//
311 // Groupings using register classes and tuples
312 //===----------------------------------------------------------------------===//
314 def SCC_CLASS : SIRegisterClass<"AMDGPU", [i1], 1, (add SCC)> {
316 let isAllocatable = 0;
320 def M0_CLASS : SIRegisterClass<"AMDGPU", [i32], 32, (add M0)> {
322 let isAllocatable = 0;
326 def M0_CLASS_LO16 : SIRegisterClass<"AMDGPU", [i16, f16], 16, (add M0_LO16)> {
329 let isAllocatable = 0;
333 // TODO: Do we need to set DwarfRegAlias on register tuples?
335 def SGPR_LO16 : SIRegisterClass<"AMDGPU", [i16, f16], 16,
336 (add (sequence "SGPR%u_LO16", 0, 105))> {
337 let AllocationPriority = 9;
339 let GeneratePressureSet = 0;
343 def SGPR_HI16 : SIRegisterClass<"AMDGPU", [i16, f16], 16,
344 (add (sequence "SGPR%u_HI16", 0, 105))> {
345 let isAllocatable = 0;
347 let GeneratePressureSet = 0;
351 // SGPR 32-bit registers
352 def SGPR_32 : SIRegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
353 (add (sequence "SGPR%u", 0, 105))> {
354 // Give all SGPR classes higher priority than VGPR classes, because
355 // we want to spill SGPRs to VGPRs.
356 let AllocationPriority = 9;
357 let GeneratePressureSet = 0;
361 // SGPR 64-bit registers
362 def SGPR_64Regs : SIRegisterTuples<getSubRegs<2>.ret, SGPR_32, 105, 2, 2, "s">;
364 // SGPR 96-bit registers. No operations use these, but for symmetry with 96-bit VGPRs.
365 def SGPR_96Regs : SIRegisterTuples<getSubRegs<3>.ret, SGPR_32, 105, 3, 3, "s">;
367 // SGPR 128-bit registers
368 def SGPR_128Regs : SIRegisterTuples<getSubRegs<4>.ret, SGPR_32, 105, 4, 4, "s">;
370 // SGPR 160-bit registers. No operations use these, but for symmetry with 160-bit VGPRs.
371 def SGPR_160Regs : SIRegisterTuples<getSubRegs<5>.ret, SGPR_32, 105, 4, 5, "s">;
373 // SGPR 192-bit registers. No operations use these, but for symmetry with 192-bit VGPRs.
374 def SGPR_192Regs : SIRegisterTuples<getSubRegs<6>.ret, SGPR_32, 105, 4, 6, "s">;
376 // SGPR 224-bit registers. No operations use these, but for symmetry with 224-bit VGPRs.
377 def SGPR_224Regs : SIRegisterTuples<getSubRegs<7>.ret, SGPR_32, 105, 4, 7, "s">;
379 // SGPR 256-bit registers
380 def SGPR_256Regs : SIRegisterTuples<getSubRegs<8>.ret, SGPR_32, 105, 4, 8, "s">;
382 // SGPR 512-bit registers
383 def SGPR_512Regs : SIRegisterTuples<getSubRegs<16>.ret, SGPR_32, 105, 4, 16, "s">;
385 // SGPR 1024-bit registers
386 def SGPR_1024Regs : SIRegisterTuples<getSubRegs<32>.ret, SGPR_32, 105, 4, 32, "s">;
388 // Trap handler TMP 32-bit registers
389 def TTMP_32 : SIRegisterClass<"AMDGPU", [i32, f32, v2i16, v2f16], 32,
390 (add (sequence "TTMP%u", 0, 15))> {
391 let isAllocatable = 0;
395 // Trap handler TMP 16-bit registers
396 def TTMP_LO16 : SIRegisterClass<"AMDGPU", [i16, f16], 16,
397 (add (sequence "TTMP%u_LO16", 0, 15))> {
399 let isAllocatable = 0;
403 // Trap handler TMP 64-bit registers
404 def TTMP_64Regs : SIRegisterTuples<getSubRegs<2>.ret, TTMP_32, 15, 2, 2, "ttmp">;
406 // Trap handler TMP 96-bit registers
407 def TTMP_96Regs : SIRegisterTuples<getSubRegs<3>.ret, TTMP_32, 15, 3, 3, "ttmp">;
409 // Trap handler TMP 128-bit registers
410 def TTMP_128Regs : SIRegisterTuples<getSubRegs<4>.ret, TTMP_32, 15, 4, 4, "ttmp">;
412 // Trap handler TMP 160-bit registers
413 def TTMP_160Regs : SIRegisterTuples<getSubRegs<5>.ret, TTMP_32, 15, 4, 5, "ttmp">;
415 // Trap handler TMP 192-bit registers
416 def TTMP_192Regs : SIRegisterTuples<getSubRegs<6>.ret, TTMP_32, 15, 4, 6, "ttmp">;
418 // Trap handler TMP 224-bit registers
419 def TTMP_224Regs : SIRegisterTuples<getSubRegs<7>.ret, TTMP_32, 15, 4, 7, "ttmp">;
421 // Trap handler TMP 256-bit registers
422 def TTMP_256Regs : SIRegisterTuples<getSubRegs<8>.ret, TTMP_32, 15, 4, 8, "ttmp">;
424 // Trap handler TMP 512-bit registers
425 def TTMP_512Regs : SIRegisterTuples<getSubRegs<16>.ret, TTMP_32, 15, 4, 16, "ttmp">;
427 class TmpRegTuplesBase<int index, int size,
428 list<Register> subRegs,
429 list<SubRegIndex> indices = getSubRegs<size>.ret,
430 int index1 = !add(index, size, -1),
431 string name = "ttmp["#index#":"#index1#"]"> :
432 RegisterWithSubRegs<name, subRegs> {
433 let HWEncoding = subRegs[0].HWEncoding;
434 let SubRegIndices = indices;
437 class TmpRegTuples<string tgt,
440 int index1 = !add(index0, 1),
441 int index2 = !add(index0, !if(!eq(size, 2), 1, 2)),
442 int index3 = !add(index0, !if(!eq(size, 2), 1, 3)),
443 int index4 = !add(index0, !if(!eq(size, 8), 4, 1)),
444 int index5 = !add(index0, !if(!eq(size, 8), 5, 1)),
445 int index6 = !add(index0, !if(!eq(size, 8), 6, 1)),
446 int index7 = !add(index0, !if(!eq(size, 8), 7, 1)),
447 Register r0 = !cast<Register>("TTMP"#index0#tgt),
448 Register r1 = !cast<Register>("TTMP"#index1#tgt),
449 Register r2 = !cast<Register>("TTMP"#index2#tgt),
450 Register r3 = !cast<Register>("TTMP"#index3#tgt),
451 Register r4 = !cast<Register>("TTMP"#index4#tgt),
452 Register r5 = !cast<Register>("TTMP"#index5#tgt),
453 Register r6 = !cast<Register>("TTMP"#index6#tgt),
454 Register r7 = !cast<Register>("TTMP"#index7#tgt)> :
455 TmpRegTuplesBase<index0, size,
456 !if(!eq(size, 2), [r0, r1],
457 !if(!eq(size, 4), [r0, r1, r2, r3],
458 [r0, r1, r2, r3, r4, r5, r6, r7])),
459 getSubRegs<size>.ret>;
461 foreach Index = {0, 2, 4, 6, 8, 10, 12, 14} in {
462 def TTMP#Index#_TTMP#!add(Index,1)#_vi : TmpRegTuples<"_vi", 2, Index>;
463 def TTMP#Index#_TTMP#!add(Index,1)#_gfx9plus : TmpRegTuples<"_gfx9plus", 2, Index>;
466 foreach Index = {0, 4, 8, 12} in {
467 def TTMP#Index#_TTMP#!add(Index,1)#
469 _TTMP#!add(Index,3)#_vi : TmpRegTuples<"_vi", 4, Index>;
470 def TTMP#Index#_TTMP#!add(Index,1)#
472 _TTMP#!add(Index,3)#_gfx9plus : TmpRegTuples<"_gfx9plus", 4, Index>;
475 foreach Index = {0, 4, 8} in {
476 def TTMP#Index#_TTMP#!add(Index,1)#
482 _TTMP#!add(Index,7)#_vi : TmpRegTuples<"_vi", 8, Index>;
483 def TTMP#Index#_TTMP#!add(Index,1)#
489 _TTMP#!add(Index,7)#_gfx9plus : TmpRegTuples<"_gfx9plus", 8, Index>;
492 def TTMP0_TTMP1_TTMP2_TTMP3_TTMP4_TTMP5_TTMP6_TTMP7_TTMP8_TTMP9_TTMP10_TTMP11_TTMP12_TTMP13_TTMP14_TTMP15_vi :
493 TmpRegTuplesBase<0, 16,
494 [TTMP0_vi, TTMP1_vi, TTMP2_vi, TTMP3_vi,
495 TTMP4_vi, TTMP5_vi, TTMP6_vi, TTMP7_vi,
496 TTMP8_vi, TTMP9_vi, TTMP10_vi, TTMP11_vi,
497 TTMP12_vi, TTMP13_vi, TTMP14_vi, TTMP15_vi]>;
499 def TTMP0_TTMP1_TTMP2_TTMP3_TTMP4_TTMP5_TTMP6_TTMP7_TTMP8_TTMP9_TTMP10_TTMP11_TTMP12_TTMP13_TTMP14_TTMP15_gfx9plus :
500 TmpRegTuplesBase<0, 16,
501 [TTMP0_gfx9plus, TTMP1_gfx9plus, TTMP2_gfx9plus, TTMP3_gfx9plus,
502 TTMP4_gfx9plus, TTMP5_gfx9plus, TTMP6_gfx9plus, TTMP7_gfx9plus,
503 TTMP8_gfx9plus, TTMP9_gfx9plus, TTMP10_gfx9plus, TTMP11_gfx9plus,
504 TTMP12_gfx9plus, TTMP13_gfx9plus, TTMP14_gfx9plus, TTMP15_gfx9plus]>;
506 class RegisterTypes<list<ValueType> reg_types> {
507 list<ValueType> types = reg_types;
510 def Reg16Types : RegisterTypes<[i16, f16]>;
511 def Reg32Types : RegisterTypes<[i32, f32, v2i16, v2f16, p2, p3, p5, p6]>;
514 def VGPR_LO16 : SIRegisterClass<"AMDGPU", Reg16Types.types, 16,
515 (add (sequence "VGPR%u_LO16", 0, 255))> {
516 let AllocationPriority = 1;
518 let GeneratePressureSet = 0;
521 def VGPR_HI16 : SIRegisterClass<"AMDGPU", Reg16Types.types, 16,
522 (add (sequence "VGPR%u_HI16", 0, 255))> {
523 let AllocationPriority = 1;
525 let GeneratePressureSet = 0;
528 // VGPR 32-bit registers
529 // i16/f16 only on VI+
530 def VGPR_32 : SIRegisterClass<"AMDGPU", !listconcat(Reg32Types.types, Reg16Types.types), 32,
531 (add (sequence "VGPR%u", 0, 255))> {
532 let AllocationPriority = 1;
538 // VGPR 64-bit registers
539 def VGPR_64 : SIRegisterTuples<getSubRegs<2>.ret, VGPR_32, 255, 1, 2, "v">;
541 // VGPR 96-bit registers
542 def VGPR_96 : SIRegisterTuples<getSubRegs<3>.ret, VGPR_32, 255, 1, 3, "v">;
544 // VGPR 128-bit registers
545 def VGPR_128 : SIRegisterTuples<getSubRegs<4>.ret, VGPR_32, 255, 1, 4, "v">;
547 // VGPR 160-bit registers
548 def VGPR_160 : SIRegisterTuples<getSubRegs<5>.ret, VGPR_32, 255, 1, 5, "v">;
550 // VGPR 192-bit registers
551 def VGPR_192 : SIRegisterTuples<getSubRegs<6>.ret, VGPR_32, 255, 1, 6, "v">;
553 // VGPR 224-bit registers
554 def VGPR_224 : SIRegisterTuples<getSubRegs<7>.ret, VGPR_32, 255, 1, 7, "v">;
556 // VGPR 256-bit registers
557 def VGPR_256 : SIRegisterTuples<getSubRegs<8>.ret, VGPR_32, 255, 1, 8, "v">;
559 // VGPR 512-bit registers
560 def VGPR_512 : SIRegisterTuples<getSubRegs<16>.ret, VGPR_32, 255, 1, 16, "v">;
562 // VGPR 1024-bit registers
563 def VGPR_1024 : SIRegisterTuples<getSubRegs<32>.ret, VGPR_32, 255, 1, 32, "v">;
566 def AGPR_LO16 : SIRegisterClass<"AMDGPU", Reg16Types.types, 16,
567 (add (sequence "AGPR%u_LO16", 0, 255))> {
568 let isAllocatable = 0;
570 let GeneratePressureSet = 0;
573 // AccVGPR 32-bit registers
574 def AGPR_32 : SIRegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
575 (add (sequence "AGPR%u", 0, 255))> {
576 let AllocationPriority = 1;
582 // AGPR 64-bit registers
583 def AGPR_64 : SIRegisterTuples<getSubRegs<2>.ret, AGPR_32, 255, 1, 2, "a">;
585 // AGPR 96-bit registers
586 def AGPR_96 : SIRegisterTuples<getSubRegs<3>.ret, AGPR_32, 255, 1, 3, "a">;
588 // AGPR 128-bit registers
589 def AGPR_128 : SIRegisterTuples<getSubRegs<4>.ret, AGPR_32, 255, 1, 4, "a">;
591 // AGPR 160-bit registers
592 def AGPR_160 : SIRegisterTuples<getSubRegs<5>.ret, AGPR_32, 255, 1, 5, "a">;
594 // AGPR 192-bit registers
595 def AGPR_192 : SIRegisterTuples<getSubRegs<6>.ret, AGPR_32, 255, 1, 6, "a">;
597 // AGPR 224-bit registers
598 def AGPR_224 : SIRegisterTuples<getSubRegs<7>.ret, AGPR_32, 255, 1, 7, "a">;
600 // AGPR 256-bit registers
601 def AGPR_256 : SIRegisterTuples<getSubRegs<8>.ret, AGPR_32, 255, 1, 8, "a">;
603 // AGPR 512-bit registers
604 def AGPR_512 : SIRegisterTuples<getSubRegs<16>.ret, AGPR_32, 255, 1, 16, "a">;
606 // AGPR 1024-bit registers
607 def AGPR_1024 : SIRegisterTuples<getSubRegs<32>.ret, AGPR_32, 255, 1, 32, "a">;
609 //===----------------------------------------------------------------------===//
610 // Register classes used as source and destination
611 //===----------------------------------------------------------------------===//
613 def Pseudo_SReg_32 : SIRegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
614 (add FP_REG, SP_REG)> {
615 let isAllocatable = 0;
620 def Pseudo_SReg_128 : SIRegisterClass<"AMDGPU", [v4i32, v2i64, v2f64, v8i16, v8f16], 32,
621 (add PRIVATE_RSRC_REG)> {
622 let isAllocatable = 0;
627 def LDS_DIRECT_CLASS : RegisterClass<"AMDGPU", [i32], 32,
629 let isAllocatable = 0;
633 let GeneratePressureSet = 0, HasSGPR = 1 in {
634 // Subset of SReg_32 without M0 for SMRD instructions and alike.
635 // See comments in SIInstructions.td for more info.
636 def SReg_32_XM0_XEXEC : SIRegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16, i1], 32,
637 (add SGPR_32, VCC_LO, VCC_HI, FLAT_SCR_LO, FLAT_SCR_HI, XNACK_MASK_LO, XNACK_MASK_HI,
638 SGPR_NULL, TTMP_32, TMA_LO, TMA_HI, TBA_LO, TBA_HI, SRC_SHARED_BASE, SRC_SHARED_LIMIT,
639 SRC_PRIVATE_BASE, SRC_PRIVATE_LIMIT, SRC_POPS_EXITING_WAVE_ID,
640 SRC_VCCZ, SRC_EXECZ, SRC_SCC)> {
641 let AllocationPriority = 10;
644 def SReg_LO16_XM0_XEXEC : SIRegisterClass<"AMDGPU", [i16, f16], 16,
645 (add SGPR_LO16, VCC_LO_LO16, VCC_HI_LO16, FLAT_SCR_LO_LO16, FLAT_SCR_HI_LO16,
646 XNACK_MASK_LO_LO16, XNACK_MASK_HI_LO16, SGPR_NULL_LO16, TTMP_LO16, TMA_LO_LO16,
647 TMA_HI_LO16, TBA_LO_LO16, TBA_HI_LO16, SRC_SHARED_BASE_LO16,
648 SRC_SHARED_LIMIT_LO16, SRC_PRIVATE_BASE_LO16, SRC_PRIVATE_LIMIT_LO16,
649 SRC_POPS_EXITING_WAVE_ID_LO16, SRC_VCCZ_LO16, SRC_EXECZ_LO16, SRC_SCC_LO16)> {
651 let AllocationPriority = 10;
654 def SReg_32_XEXEC_HI : SIRegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16, i1], 32,
655 (add SReg_32_XM0_XEXEC, EXEC_LO, M0_CLASS)> {
656 let AllocationPriority = 10;
659 def SReg_LO16_XEXEC_HI : SIRegisterClass<"AMDGPU", [i16, f16], 16,
660 (add SReg_LO16_XM0_XEXEC, EXEC_LO_LO16, M0_CLASS_LO16)> {
662 let AllocationPriority = 10;
665 def SReg_32_XM0 : SIRegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16, i1], 32,
666 (add SReg_32_XM0_XEXEC, EXEC_LO, EXEC_HI)> {
667 let AllocationPriority = 10;
670 def SReg_LO16_XM0 : SIRegisterClass<"AMDGPU", [i16, f16], 16,
671 (add SReg_LO16_XM0_XEXEC, EXEC_LO_LO16, EXEC_HI_LO16)> {
673 let AllocationPriority = 10;
676 def SReg_LO16 : SIRegisterClass<"AMDGPU", [i16, f16], 16,
677 (add SGPR_LO16, SReg_LO16_XM0, M0_CLASS_LO16, EXEC_LO_LO16, EXEC_HI_LO16, SReg_LO16_XEXEC_HI)> {
679 let AllocationPriority = 10;
681 } // End GeneratePressureSet = 0
683 // Register class for all scalar registers (SGPRs + Special Registers)
684 def SReg_32 : SIRegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16, i1], 32,
685 (add SReg_32_XM0, M0_CLASS, EXEC_LO, EXEC_HI, SReg_32_XEXEC_HI)> {
686 let AllocationPriority = 10;
690 let GeneratePressureSet = 0 in {
691 def SRegOrLds_32 : SIRegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
692 (add SReg_32, LDS_DIRECT_CLASS)> {
693 let isAllocatable = 0;
697 def SGPR_64 : SIRegisterClass<"AMDGPU", [v2i32, i64, v2f32, f64, v4i16, v4f16], 32,
700 let AllocationPriority = 11;
704 // CCR (call clobbered registers) SGPR 64-bit registers
705 def CCR_SGPR_64 : SIRegisterClass<"AMDGPU", SGPR_64.RegTypes, 32,
706 (add (trunc SGPR_64, 16))> {
707 let CopyCost = SGPR_64.CopyCost;
708 let AllocationPriority = SGPR_64.AllocationPriority;
712 // Call clobbered 64-bit SGPRs for AMDGPU_Gfx CC
713 def Gfx_CCR_SGPR_64 : SIRegisterClass<"AMDGPU", SGPR_64.RegTypes, 32,
714 (add (trunc (shl SGPR_64, 15), 1), // s[30:31]
715 (trunc (shl SGPR_64, 18), 14))> { // s[36:37]-s[s62:63]
716 let CopyCost = SGPR_64.CopyCost;
717 let AllocationPriority = SGPR_64.AllocationPriority;
721 def TTMP_64 : SIRegisterClass<"AMDGPU", [v2i32, i64, f64, v4i16, v4f16], 32,
723 let isAllocatable = 0;
727 def SReg_64_XEXEC : SIRegisterClass<"AMDGPU", [v2i32, i64, v2f32, f64, i1, v4i16, v4f16], 32,
728 (add SGPR_64, VCC, FLAT_SCR, XNACK_MASK, TTMP_64, TBA, TMA)> {
730 let AllocationPriority = 13;
734 def SReg_64 : SIRegisterClass<"AMDGPU", [v2i32, i64, v2f32, f64, i1, v4i16, v4f16], 32,
735 (add SReg_64_XEXEC, EXEC)> {
737 let AllocationPriority = 13;
741 def SReg_1_XEXEC : SIRegisterClass<"AMDGPU", [i1], 32,
742 (add SReg_64_XEXEC, SReg_32_XM0_XEXEC)> {
744 let isAllocatable = 0;
748 def SReg_1 : SIRegisterClass<"AMDGPU", [i1], 32,
749 (add SReg_1_XEXEC, EXEC, EXEC_LO)> {
751 let isAllocatable = 0;
755 multiclass SRegClass<int numRegs, int priority,
756 list<ValueType> regTypes,
757 SIRegisterTuples regList,
758 SIRegisterTuples ttmpList = regList,
759 int copyCost = !sra(!add(numRegs, 1), 1)> {
760 defvar hasTTMP = !ne(regList, ttmpList);
761 defvar suffix = !cast<string>(!mul(numRegs, 32));
762 defvar sgprName = !strconcat("SGPR_", suffix);
763 defvar ttmpName = !strconcat("TTMP_", suffix);
765 let AllocationPriority = priority, CopyCost = copyCost, HasSGPR = 1 in {
766 def "" # sgprName : SIRegisterClass<"AMDGPU", regTypes, 32, (add regList)> {
770 def "" # ttmpName : SIRegisterClass<"AMDGPU", regTypes, 32, (add ttmpList)> {
771 let isAllocatable = 0;
776 SIRegisterClass<"AMDGPU", regTypes, 32,
777 !con(!dag(add, [!cast<RegisterClass>(sgprName)], ["sgpr"]),
779 !dag(add, [!cast<RegisterClass>(ttmpName)], ["ttmp"]),
781 let isAllocatable = 0;
786 defm "" : SRegClass<3, 14, [v3i32, v3f32], SGPR_96Regs, TTMP_96Regs>;
787 defm "" : SRegClass<4, 15, [v4i32, v4f32, v2i64, v2f64, v8i16, v8f16], SGPR_128Regs, TTMP_128Regs>;
788 defm "" : SRegClass<5, 16, [v5i32, v5f32], SGPR_160Regs, TTMP_160Regs>;
789 defm "" : SRegClass<6, 17, [v6i32, v6f32, v3i64, v3f64], SGPR_192Regs, TTMP_192Regs>;
790 defm "" : SRegClass<7, 18, [v7i32, v7f32], SGPR_224Regs, TTMP_224Regs>;
791 defm "" : SRegClass<8, 19, [v8i32, v8f32, v4i64, v4f64], SGPR_256Regs, TTMP_256Regs>;
792 defm "" : SRegClass<16, 20, [v16i32, v16f32, v8i64, v8f64], SGPR_512Regs, TTMP_512Regs>;
793 defm "" : SRegClass<32, 21, [v32i32, v32f32, v16i64, v16f64], SGPR_1024Regs>;
795 def VRegOrLds_32 : SIRegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
796 (add VGPR_32, LDS_DIRECT_CLASS)> {
797 let isAllocatable = 0;
801 // Register class for all vector registers (VGPRs + Interpolation Registers)
802 class VRegClassBase<int numRegs, list<ValueType> regTypes, dag regList> :
803 SIRegisterClass<"AMDGPU", regTypes, 32, regList> {
804 let Size = !mul(numRegs, 32);
806 // Requires n v_mov_b32 to copy
807 let CopyCost = numRegs;
808 let AllocationPriority = numRegs;
809 let Weight = numRegs;
812 // Define a register tuple class, along with one requiring an even
813 // aligned base register.
814 multiclass VRegClass<int numRegs, list<ValueType> regTypes, dag regList> {
816 // Define the regular class.
817 def "" : VRegClassBase<numRegs, regTypes, regList>;
819 // Define 2-aligned variant
820 def _Align2 : VRegClassBase<numRegs, regTypes, (decimate regList, 2)>;
824 defm VReg_64 : VRegClass<2, [i64, f64, v2i32, v2f32, v4f16, v4i16, p0, p1, p4],
826 defm VReg_96 : VRegClass<3, [v3i32, v3f32], (add VGPR_96)>;
827 defm VReg_128 : VRegClass<4, [v4i32, v4f32, v2i64, v2f64, v8i16, v8f16], (add VGPR_128)>;
828 defm VReg_160 : VRegClass<5, [v5i32, v5f32], (add VGPR_160)>;
830 defm VReg_192 : VRegClass<6, [v6i32, v6f32, v3i64, v3f64], (add VGPR_192)>;
831 defm VReg_224 : VRegClass<7, [v7i32, v7f32], (add VGPR_224)>;
832 defm VReg_256 : VRegClass<8, [v8i32, v8f32, v4i64, v4f64], (add VGPR_256)>;
833 defm VReg_512 : VRegClass<16, [v16i32, v16f32, v8i64, v8f64], (add VGPR_512)>;
834 defm VReg_1024 : VRegClass<32, [v32i32, v32f32, v16i64, v16f64], (add VGPR_1024)>;
836 multiclass ARegClass<int numRegs, list<ValueType> regTypes, dag regList> {
837 let CopyCost = !add(numRegs, numRegs, 1), HasAGPR = 1 in {
838 // Define the regular class.
839 def "" : VRegClassBase<numRegs, regTypes, regList>;
841 // Define 2-aligned variant
842 def _Align2 : VRegClassBase<numRegs, regTypes, (decimate regList, 2)>;
846 defm AReg_64 : ARegClass<2, [i64, f64, v2i32, v2f32, v4f16, v4i16],
848 defm AReg_96 : ARegClass<3, [v3i32, v3f32], (add AGPR_96)>;
849 defm AReg_128 : ARegClass<4, [v4i32, v4f32, v2i64, v2f64, v8i16, v8f16], (add AGPR_128)>;
850 defm AReg_160 : ARegClass<5, [v5i32, v5f32], (add AGPR_160)>;
851 defm AReg_192 : ARegClass<6, [v6i32, v6f32, v3i64, v3f64], (add AGPR_192)>;
852 defm AReg_224 : ARegClass<7, [v7i32, v7f32], (add AGPR_224)>;
853 defm AReg_256 : ARegClass<8, [v8i32, v8f32, v4i64, v4f64], (add AGPR_256)>;
854 defm AReg_512 : ARegClass<16, [v16i32, v16f32, v8i64, v8f64], (add AGPR_512)>;
855 defm AReg_1024 : ARegClass<32, [v32i32, v32f32, v16i64, v16f64], (add AGPR_1024)>;
857 } // End GeneratePressureSet = 0
859 // This is not a real register. This is just to have a register to add
860 // to VReg_1 that does not alias any real register that would
861 // introduce inferred register classes.
862 def ARTIFICIAL_VGPR : SIReg <"invalid vgpr", 0> {
863 let isArtificial = 1;
866 let GeneratePressureSet = 0 in {
867 // FIXME: Should specify an empty set for this. No register should
868 // ever be allocated using VReg_1. This is a hack for SelectionDAG
869 // that should always be lowered by SILowerI1Copies. TableGen crashes
870 // on an empty register set, but also sorts register classes based on
871 // the number of registerss in them. Add only one register so this is
872 // sorted to the end and not preferred over VGPR_32.
873 def VReg_1 : SIRegisterClass<"AMDGPU", [i1], 32, (add ARTIFICIAL_VGPR)> {
878 def VS_32 : SIRegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
879 (add VGPR_32, SReg_32, LDS_DIRECT_CLASS)> {
880 let isAllocatable = 0;
885 def VS_64 : SIRegisterClass<"AMDGPU", [i64, f64, v2f32], 32, (add VReg_64, SReg_64)> {
886 let isAllocatable = 0;
891 def AV_32 : SIRegisterClass<"AMDGPU", VGPR_32.RegTypes, 32, (add VGPR_32, AGPR_32)> {
895 } // End GeneratePressureSet = 0
897 // Define a register tuple class, along with one requiring an even
898 // aligned base register.
899 multiclass AVRegClass<int numRegs, list<ValueType> regTypes,
900 dag vregList, dag aregList> {
901 let HasVGPR = 1, HasAGPR = 1 in {
902 // Define the regular class.
903 def "" : VRegClassBase<numRegs, regTypes, (add vregList, aregList)>;
905 // Define 2-aligned variant
906 def _Align2 : VRegClassBase<numRegs, regTypes,
907 (add (decimate vregList, 2),
908 (decimate aregList, 2))>;
912 defm AV_64 : AVRegClass<2, VReg_64.RegTypes, (add VGPR_64), (add AGPR_64)>;
913 defm AV_96 : AVRegClass<3, VReg_96.RegTypes, (add VGPR_96), (add AGPR_96)>;
914 defm AV_128 : AVRegClass<4, VReg_128.RegTypes, (add VGPR_128), (add AGPR_128)>;
915 defm AV_160 : AVRegClass<5, VReg_160.RegTypes, (add VGPR_160), (add AGPR_160)>;
916 defm AV_192 : AVRegClass<6, VReg_160.RegTypes, (add VGPR_192), (add AGPR_192)>;
917 defm AV_224 : AVRegClass<7, VReg_160.RegTypes, (add VGPR_224), (add AGPR_224)>;
918 defm AV_256 : AVRegClass<8, VReg_160.RegTypes, (add VGPR_256), (add AGPR_256)>;
919 defm AV_512 : AVRegClass<16, VReg_160.RegTypes, (add VGPR_512), (add AGPR_512)>;
920 defm AV_1024 : AVRegClass<32, VReg_160.RegTypes, (add VGPR_1024), (add AGPR_1024)>;
922 //===----------------------------------------------------------------------===//
924 //===----------------------------------------------------------------------===//
926 class RegImmMatcher<string name> : AsmOperandClass {
928 let RenderMethod = "addRegOrImmOperands";
931 multiclass SIRegOperand32 <string rc, string MatchName, string opType,
932 string rc_suffix = "_32"> {
933 let OperandNamespace = "AMDGPU" in {
934 def _b16 : RegisterOperand<!cast<RegisterClass>(rc#rc_suffix)> {
935 let OperandType = opType#"_INT16";
936 let ParserMatchClass = RegImmMatcher<MatchName#"B16">;
937 let DecoderMethod = "decodeOperand_VSrc16";
940 def _f16 : RegisterOperand<!cast<RegisterClass>(rc#rc_suffix)> {
941 let OperandType = opType#"_FP16";
942 let ParserMatchClass = RegImmMatcher<MatchName#"F16">;
943 let DecoderMethod = "decodeOperand_" # rc # "_16";
946 def _b32 : RegisterOperand<!cast<RegisterClass>(rc#rc_suffix)> {
947 let OperandType = opType#"_INT32";
948 let ParserMatchClass = RegImmMatcher<MatchName#"B32">;
949 let DecoderMethod = "decodeOperand_" # rc # rc_suffix;
952 def _f32 : RegisterOperand<!cast<RegisterClass>(rc#rc_suffix)> {
953 let OperandType = opType#"_FP32";
954 let ParserMatchClass = RegImmMatcher<MatchName#"F32">;
955 let DecoderMethod = "decodeOperand_" # rc # rc_suffix;
958 def _v2b16 : RegisterOperand<!cast<RegisterClass>(rc#rc_suffix)> {
959 let OperandType = opType#"_V2INT16";
960 let ParserMatchClass = RegImmMatcher<MatchName#"V2B16">;
961 let DecoderMethod = "decodeOperand_VSrcV216";
964 def _v2f16 : RegisterOperand<!cast<RegisterClass>(rc#rc_suffix)> {
965 let OperandType = opType#"_V2FP16";
966 let ParserMatchClass = RegImmMatcher<MatchName#"V2F16">;
967 let DecoderMethod = "decodeOperand_VSrcV216";
972 multiclass SIRegOperand64 <string rc, string MatchName, string opType,
973 string rc_suffix = "_64", bit Vectors = 1> {
974 let OperandNamespace = "AMDGPU" in {
975 def _b64 : RegisterOperand<!cast<RegisterClass>(rc#rc_suffix)> {
976 let OperandType = opType#"_INT64";
977 let ParserMatchClass = RegImmMatcher<MatchName#"B64">;
980 def _f64 : RegisterOperand<!cast<RegisterClass>(rc#rc_suffix)> {
981 let OperandType = opType#"_FP64";
982 let ParserMatchClass = RegImmMatcher<MatchName#"F64">;
986 def _v2f32 : RegisterOperand<!cast<RegisterClass>(rc#rc_suffix)> {
987 let OperandType = opType#"_V2FP32";
988 let ParserMatchClass = RegImmMatcher<MatchName#"V2FP32">;
989 let DecoderMethod = "decodeOperand_VSrcV232";
992 def _v2b32 : RegisterOperand<!cast<RegisterClass>(rc#rc_suffix)> {
993 let OperandType = opType#"_V2INT32";
994 let ParserMatchClass = RegImmMatcher<MatchName#"V2INT32">;
995 let DecoderMethod = "decodeOperand_VSrcV232";
1000 multiclass SIRegOperand <string rc, string MatchName, string opType> :
1001 SIRegOperand32<rc, MatchName, opType>,
1002 SIRegOperand64<rc, MatchName, opType>;
1004 // FIXME: 64-bit sources can sometimes use 32-bit constants.
1005 multiclass RegImmOperand <string rc, string MatchName>
1006 : SIRegOperand<rc, MatchName, "OPERAND_REG_IMM">;
1008 multiclass RegInlineOperand <string rc, string MatchName>
1009 : SIRegOperand<rc, MatchName, "OPERAND_REG_INLINE_C">;
1011 multiclass RegInlineOperand32 <string rc, string MatchName,
1012 string rc_suffix = "_32">
1013 : SIRegOperand32<rc, MatchName, "OPERAND_REG_INLINE_C", rc_suffix>;
1015 multiclass RegInlineOperand64 <string rc, string MatchName,
1016 string rc_suffix = "_64">
1017 : SIRegOperand64<rc, MatchName, "OPERAND_REG_INLINE_C", rc_suffix>;
1019 multiclass RegInlineOperandAC <string rc, string MatchName,
1020 string rc_suffix = "_32">
1021 : SIRegOperand32<rc, MatchName, "OPERAND_REG_INLINE_AC", rc_suffix>;
1023 multiclass RegInlineOperandAC64 <string rc, string MatchName,
1024 string rc_suffix = "_64">
1025 : SIRegOperand64<rc, MatchName, "OPERAND_REG_INLINE_AC", rc_suffix, 0>;
1027 //===----------------------------------------------------------------------===//
1028 // SSrc_* Operands with an SGPR or a 32-bit immediate
1029 //===----------------------------------------------------------------------===//
1031 defm SSrc : RegImmOperand<"SReg", "SSrc">;
1033 def SSrcOrLds_b32 : RegisterOperand<SRegOrLds_32> {
1034 let OperandNamespace = "AMDGPU";
1035 let OperandType = "OPERAND_REG_IMM_INT32";
1036 let ParserMatchClass = RegImmMatcher<"SSrcOrLdsB32">;
1039 //===----------------------------------------------------------------------===//
1040 // SCSrc_* Operands with an SGPR or a inline constant
1041 //===----------------------------------------------------------------------===//
1043 defm SCSrc : RegInlineOperand<"SReg", "SCSrc"> ;
1045 //===----------------------------------------------------------------------===//
1046 // VSrc_* Operands with an SGPR, VGPR or a 32-bit immediate
1047 //===----------------------------------------------------------------------===//
1049 defm VSrc : RegImmOperand<"VS", "VSrc">;
1051 def VSrc_128 : RegisterOperand<VReg_128> {
1052 let DecoderMethod = "DecodeVS_128RegisterClass";
1055 //===----------------------------------------------------------------------===//
1056 // VSrc_*_Deferred Operands with an SGPR, VGPR or a 32-bit immediate for use
1058 //===----------------------------------------------------------------------===//
1060 multiclass SIRegOperand32_Deferred <string rc, string MatchName, string opType,
1061 string rc_suffix = "_32"> {
1062 let OperandNamespace = "AMDGPU" in {
1063 def _f16_Deferred : RegisterOperand<!cast<RegisterClass>(rc#rc_suffix)> {
1064 let OperandType = opType#"_FP16_DEFERRED";
1065 let ParserMatchClass = RegImmMatcher<MatchName#"F16">;
1066 let DecoderMethod = "decodeOperand_" # rc # "_16_Deferred";
1069 def _f32_Deferred : RegisterOperand<!cast<RegisterClass>(rc#rc_suffix)> {
1070 let OperandType = opType#"_FP32_DEFERRED";
1071 let ParserMatchClass = RegImmMatcher<MatchName#"F32">;
1072 let DecoderMethod = "decodeOperand_" # rc # "_32_Deferred";
1077 defm VSrc : SIRegOperand32_Deferred<"VS", "VSrc", "OPERAND_REG_IMM">;
1079 //===----------------------------------------------------------------------===//
1080 // VRegSrc_* Operands with a VGPR
1081 //===----------------------------------------------------------------------===//
1083 // This is for operands with the enum(9), VSrc encoding restriction,
1084 // but only allows VGPRs.
1085 def VRegSrc_32 : RegisterOperand<VGPR_32> {
1086 //let ParserMatchClass = RegImmMatcher<"VRegSrc32">;
1087 let DecoderMethod = "DecodeVS_32RegisterClass";
1090 //===----------------------------------------------------------------------===//
1091 // ASrc_* Operands with an AccVGPR
1092 //===----------------------------------------------------------------------===//
1094 def ARegSrc_32 : RegisterOperand<AGPR_32> {
1095 let DecoderMethod = "DecodeAGPR_32RegisterClass";
1096 let EncoderMethod = "getAVOperandEncoding";
1099 //===----------------------------------------------------------------------===//
1100 // VCSrc_* Operands with an SGPR, VGPR or an inline constant
1101 //===----------------------------------------------------------------------===//
1103 defm VCSrc : RegInlineOperand<"VS", "VCSrc">;
1105 //===----------------------------------------------------------------------===//
1106 // VISrc_* Operands with a VGPR or an inline constant
1107 //===----------------------------------------------------------------------===//
1109 defm VISrc : RegInlineOperand32<"VGPR", "VISrc">;
1110 let DecoderMethod = "decodeOperand_VReg_64" in
1111 defm VISrc_64 : RegInlineOperand64<"VReg", "VISrc_64", "_64">;
1112 defm VISrc_128 : RegInlineOperandAC<"VReg", "VISrc_128", "_128">;
1113 let DecoderMethod = "decodeOperand_VReg_256" in
1114 defm VISrc_256 : RegInlineOperand64<"VReg", "VISrc_256", "_256">;
1115 defm VISrc_512 : RegInlineOperandAC<"VReg", "VISrc_512", "_512">;
1116 defm VISrc_1024 : RegInlineOperandAC<"VReg", "VISrc_1024", "_1024">;
1118 //===----------------------------------------------------------------------===//
1119 // AVSrc_* Operands with an AGPR or VGPR
1120 //===----------------------------------------------------------------------===//
1122 def AVSrc_32 : RegisterOperand<AV_32> {
1123 let DecoderMethod = "DecodeAV_32RegisterClass";
1124 let EncoderMethod = "getAVOperandEncoding";
1127 def AVSrc_64 : RegisterOperand<AV_64> {
1128 let DecoderMethod = "DecodeAV_64RegisterClass";
1129 let EncoderMethod = "getAVOperandEncoding";
1132 def AVLdSt_32 : RegisterOperand<AV_32> {
1133 let DecoderMethod = "DecodeAVLdSt_32RegisterClass";
1134 let EncoderMethod = "getAVOperandEncoding";
1137 def AVLdSt_64 : RegisterOperand<AV_64> {
1138 let DecoderMethod = "DecodeAVLdSt_64RegisterClass";
1139 let EncoderMethod = "getAVOperandEncoding";
1142 def AVLdSt_96 : RegisterOperand<AV_96> {
1143 let DecoderMethod = "DecodeAVLdSt_96RegisterClass";
1144 let EncoderMethod = "getAVOperandEncoding";
1147 def AVLdSt_128 : RegisterOperand<AV_128> {
1148 let DecoderMethod = "DecodeAVLdSt_128RegisterClass";
1149 let EncoderMethod = "getAVOperandEncoding";
1152 def AVLdSt_160 : RegisterOperand<AV_160> {
1153 let DecoderMethod = "DecodeAVLdSt_160RegisterClass";
1154 let EncoderMethod = "getAVOperandEncoding";
1157 //===----------------------------------------------------------------------===//
1158 // ACSrc_* Operands with an AGPR or an inline constant
1159 //===----------------------------------------------------------------------===//
1161 defm AISrc : RegInlineOperandAC<"AGPR", "AISrc">;
1162 defm AISrc_128 : RegInlineOperandAC<"AReg", "AISrc_128", "_128">;
1163 defm AISrc_512 : RegInlineOperandAC<"AReg", "AISrc_512", "_512">;
1164 defm AISrc_1024 : RegInlineOperandAC<"AReg", "AISrc_1024", "_1024">;
1166 let DecoderMethod = "decodeOperand_AReg_64" in
1167 defm AISrc_64 : RegInlineOperandAC64<"AReg", "AISrc_64", "_64">;
1168 let DecoderMethod = "decodeOperand_AReg_256" in
1169 defm AISrc_256 : RegInlineOperandAC64<"AReg", "AISrc_256", "_256">;