blob: 928b5d2d5d3971415402e23e57085d4bc4498e87 [file] [log] [blame]
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +00001//===-- BUFInstructions.td - Buffer Instruction Defintions ----------------===//
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
10def MUBUFAddr32 : ComplexPattern<i64, 9, "SelectMUBUFAddr32">;
11def MUBUFAddr64 : ComplexPattern<i64, 7, "SelectMUBUFAddr64">;
12def MUBUFAddr64Atomic : ComplexPattern<i64, 5, "SelectMUBUFAddr64">;
13
14def MUBUFScratch : ComplexPattern<i64, 4, "SelectMUBUFScratch">;
15def MUBUFOffset : ComplexPattern<i64, 6, "SelectMUBUFOffset">;
16def MUBUFOffsetNoGLC : ComplexPattern<i64, 3, "SelectMUBUFOffset">;
17def MUBUFOffsetAtomic : ComplexPattern<i64, 4, "SelectMUBUFOffset">;
18def MUBUFIntrinsicOffset : ComplexPattern<i32, 2, "SelectMUBUFIntrinsicOffset">;
19def MUBUFIntrinsicVOffset : ComplexPattern<i32, 3, "SelectMUBUFIntrinsicVOffset">;
20
21class MubufLoad <SDPatternOperator op> : PatFrag <
22 (ops node:$ptr), (op node:$ptr), [{
23 auto const AS = cast<MemSDNode>(N)->getAddressSpace();
24 return AS == AMDGPUAS::GLOBAL_ADDRESS ||
25 AS == AMDGPUAS::CONSTANT_ADDRESS;
26}]>;
27
28def mubuf_load : MubufLoad <load>;
29def mubuf_az_extloadi8 : MubufLoad <az_extloadi8>;
30def mubuf_sextloadi8 : MubufLoad <sextloadi8>;
31def mubuf_az_extloadi16 : MubufLoad <az_extloadi16>;
32def mubuf_sextloadi16 : MubufLoad <sextloadi16>;
33def mubuf_load_atomic : MubufLoad <atomic_load>;
34
35def BUFAddrKind {
36 int Offset = 0;
37 int OffEn = 1;
38 int IdxEn = 2;
39 int BothEn = 3;
40 int Addr64 = 4;
41}
42
43class getAddrName<int addrKind> {
44 string ret =
45 !if(!eq(addrKind, BUFAddrKind.Offset), "offset",
46 !if(!eq(addrKind, BUFAddrKind.OffEn), "offen",
47 !if(!eq(addrKind, BUFAddrKind.IdxEn), "idxen",
48 !if(!eq(addrKind, BUFAddrKind.BothEn), "bothen",
49 !if(!eq(addrKind, BUFAddrKind.Addr64), "addr64",
50 "")))));
51}
52
53class MUBUFAddr64Table <bit is_addr64, string suffix = ""> {
54 bit IsAddr64 = is_addr64;
55 string OpName = NAME # suffix;
56}
57
58//===----------------------------------------------------------------------===//
59// MTBUF classes
60//===----------------------------------------------------------------------===//
61
62class MTBUF_Pseudo <string opName, dag outs, dag ins,
63 string asmOps, list<dag> pattern=[]> :
64 InstSI<outs, ins, "", pattern>,
65 SIMCInstr<opName, SIEncodingFamily.NONE> {
66
67 let isPseudo = 1;
68 let isCodeGenOnly = 1;
Matt Arsenault10c17ca2016-10-06 10:13:23 +000069 let Size = 8;
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +000070 let UseNamedOperandTable = 1;
71
72 string Mnemonic = opName;
73 string AsmOperands = asmOps;
74
75 let VM_CNT = 1;
76 let EXP_CNT = 1;
77 let MTBUF = 1;
78 let Uses = [EXEC];
79
80 let hasSideEffects = 0;
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +000081 let SchedRW = [WriteVMEM];
82}
83
Valery Pykhtinfbf2d932016-09-23 21:21:21 +000084class MTBUF_Real <MTBUF_Pseudo ps> :
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +000085 InstSI <ps.OutOperandList, ps.InOperandList, ps.Mnemonic # ps.AsmOperands, []>,
86 Enc64 {
87
88 let isPseudo = 0;
89 let isCodeGenOnly = 0;
90
91 // copy relevant pseudo op flags
92 let SubtargetPredicate = ps.SubtargetPredicate;
93 let AsmMatchConverter = ps.AsmMatchConverter;
94 let Constraints = ps.Constraints;
95 let DisableEncoding = ps.DisableEncoding;
96 let TSFlags = ps.TSFlags;
97
98 bits<8> vdata;
99 bits<12> offset;
100 bits<1> offen;
101 bits<1> idxen;
102 bits<1> glc;
103 bits<1> addr64;
104 bits<4> dfmt;
105 bits<3> nfmt;
106 bits<8> vaddr;
107 bits<7> srsrc;
108 bits<1> slc;
109 bits<1> tfe;
110 bits<8> soffset;
111
112 let Inst{11-0} = offset;
113 let Inst{12} = offen;
114 let Inst{13} = idxen;
115 let Inst{14} = glc;
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000116 let Inst{22-19} = dfmt;
117 let Inst{25-23} = nfmt;
118 let Inst{31-26} = 0x3a; //encoding
119 let Inst{39-32} = vaddr;
120 let Inst{47-40} = vdata;
121 let Inst{52-48} = srsrc{6-2};
122 let Inst{54} = slc;
123 let Inst{55} = tfe;
124 let Inst{63-56} = soffset;
125}
126
127class MTBUF_Load_Pseudo <string opName, RegisterClass regClass> : MTBUF_Pseudo <
128 opName, (outs regClass:$dst),
129 (ins u16imm:$offset, i1imm:$offen, i1imm:$idxen, i1imm:$glc, i1imm:$addr64,
130 i8imm:$dfmt, i8imm:$nfmt, VGPR_32:$vaddr, SReg_128:$srsrc,
131 i1imm:$slc, i1imm:$tfe, SCSrc_b32:$soffset),
132 " $dst, $offset, $offen, $idxen, $glc, $addr64, $dfmt,"#
Matt Arsenaultf3dd8632016-11-01 00:55:14 +0000133 " $nfmt, $vaddr, $srsrc, $slc, $tfe, $soffset"> {
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000134 let mayLoad = 1;
135 let mayStore = 0;
136}
137
138class MTBUF_Store_Pseudo <string opName, RegisterClass regClass> : MTBUF_Pseudo <
139 opName, (outs),
140 (ins regClass:$vdata, u16imm:$offset, i1imm:$offen, i1imm:$idxen, i1imm:$glc,
141 i1imm:$addr64, i8imm:$dfmt, i8imm:$nfmt, VGPR_32:$vaddr,
142 SReg_128:$srsrc, i1imm:$slc, i1imm:$tfe, SCSrc_b32:$soffset),
143 " $vdata, $offset, $offen, $idxen, $glc, $addr64, $dfmt,"#
Matt Arsenaultf3dd8632016-11-01 00:55:14 +0000144 " $nfmt, $vaddr, $srsrc, $slc, $tfe, $soffset"> {
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000145 let mayLoad = 0;
146 let mayStore = 1;
147}
148
149//===----------------------------------------------------------------------===//
150// MUBUF classes
151//===----------------------------------------------------------------------===//
152
153class MUBUF_Pseudo <string opName, dag outs, dag ins,
154 string asmOps, list<dag> pattern=[]> :
155 InstSI<outs, ins, "", pattern>,
156 SIMCInstr<opName, SIEncodingFamily.NONE> {
157
158 let isPseudo = 1;
159 let isCodeGenOnly = 1;
Matt Arsenault10c17ca2016-10-06 10:13:23 +0000160 let Size = 8;
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000161 let UseNamedOperandTable = 1;
162
163 string Mnemonic = opName;
164 string AsmOperands = asmOps;
165
166 let VM_CNT = 1;
167 let EXP_CNT = 1;
168 let MUBUF = 1;
169 let Uses = [EXEC];
170 let hasSideEffects = 0;
171 let SchedRW = [WriteVMEM];
172
173 let AsmMatchConverter = "cvtMubuf";
174
175 bits<1> offen = 0;
176 bits<1> idxen = 0;
177 bits<1> addr64 = 0;
178 bits<1> has_vdata = 1;
179 bits<1> has_vaddr = 1;
180 bits<1> has_glc = 1;
181 bits<1> glc_value = 0; // the value for glc if no such operand
182 bits<1> has_srsrc = 1;
183 bits<1> has_soffset = 1;
184 bits<1> has_offset = 1;
185 bits<1> has_slc = 1;
186 bits<1> has_tfe = 1;
187}
188
189class MUBUF_Real <bits<7> op, MUBUF_Pseudo ps> :
190 InstSI <ps.OutOperandList, ps.InOperandList, ps.Mnemonic # ps.AsmOperands, []> {
191
192 let isPseudo = 0;
193 let isCodeGenOnly = 0;
194
195 // copy relevant pseudo op flags
196 let SubtargetPredicate = ps.SubtargetPredicate;
197 let AsmMatchConverter = ps.AsmMatchConverter;
198 let Constraints = ps.Constraints;
199 let DisableEncoding = ps.DisableEncoding;
200 let TSFlags = ps.TSFlags;
201
202 bits<12> offset;
203 bits<1> glc;
204 bits<1> lds = 0;
205 bits<8> vaddr;
206 bits<8> vdata;
207 bits<7> srsrc;
208 bits<1> slc;
209 bits<1> tfe;
210 bits<8> soffset;
211}
212
213
214// For cache invalidation instructions.
215class MUBUF_Invalidate <string opName, SDPatternOperator node> :
216 MUBUF_Pseudo<opName, (outs), (ins), "", [(node)]> {
217
218 let AsmMatchConverter = "";
219
220 let hasSideEffects = 1;
221 let mayStore = 1;
222
223 // Set everything to 0.
224 let offen = 0;
225 let idxen = 0;
226 let addr64 = 0;
227 let has_vdata = 0;
228 let has_vaddr = 0;
229 let has_glc = 0;
230 let glc_value = 0;
231 let has_srsrc = 0;
232 let has_soffset = 0;
233 let has_offset = 0;
234 let has_slc = 0;
235 let has_tfe = 0;
236}
237
238class getMUBUFInsDA<list<RegisterClass> vdataList,
239 list<RegisterClass> vaddrList=[]> {
240 RegisterClass vdataClass = !if(!empty(vdataList), ?, !head(vdataList));
241 RegisterClass vaddrClass = !if(!empty(vaddrList), ?, !head(vaddrList));
242 dag InsNoData = !if(!empty(vaddrList),
243 (ins SReg_128:$srsrc, SCSrc_b32:$soffset,
Matt Arsenault4b6a6cc2016-10-28 21:55:08 +0000244 offset:$offset, GLC:$glc, slc:$slc, tfe:$tfe),
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000245 (ins vaddrClass:$vaddr, SReg_128:$srsrc, SCSrc_b32:$soffset,
Matt Arsenault4b6a6cc2016-10-28 21:55:08 +0000246 offset:$offset, GLC:$glc, slc:$slc, tfe:$tfe)
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000247 );
248 dag InsData = !if(!empty(vaddrList),
249 (ins vdataClass:$vdata, SReg_128:$srsrc,
Matt Arsenault4b6a6cc2016-10-28 21:55:08 +0000250 SCSrc_b32:$soffset, offset:$offset, GLC:$glc, slc:$slc, tfe:$tfe),
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000251 (ins vdataClass:$vdata, vaddrClass:$vaddr, SReg_128:$srsrc,
Matt Arsenault4b6a6cc2016-10-28 21:55:08 +0000252 SCSrc_b32:$soffset, offset:$offset, GLC:$glc, slc:$slc, tfe:$tfe)
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000253 );
254 dag ret = !if(!empty(vdataList), InsNoData, InsData);
255}
256
257class getMUBUFIns<int addrKind, list<RegisterClass> vdataList=[]> {
258 dag ret =
259 !if(!eq(addrKind, BUFAddrKind.Offset), getMUBUFInsDA<vdataList>.ret,
260 !if(!eq(addrKind, BUFAddrKind.OffEn), getMUBUFInsDA<vdataList, [VGPR_32]>.ret,
261 !if(!eq(addrKind, BUFAddrKind.IdxEn), getMUBUFInsDA<vdataList, [VGPR_32]>.ret,
262 !if(!eq(addrKind, BUFAddrKind.BothEn), getMUBUFInsDA<vdataList, [VReg_64]>.ret,
263 !if(!eq(addrKind, BUFAddrKind.Addr64), getMUBUFInsDA<vdataList, [VReg_64]>.ret,
264 (ins))))));
265}
266
267class getMUBUFAsmOps<int addrKind> {
268 string Pfx =
269 !if(!eq(addrKind, BUFAddrKind.Offset), "off, $srsrc, $soffset",
270 !if(!eq(addrKind, BUFAddrKind.OffEn), "$vaddr, $srsrc, $soffset offen",
271 !if(!eq(addrKind, BUFAddrKind.IdxEn), "$vaddr, $srsrc, $soffset idxen",
272 !if(!eq(addrKind, BUFAddrKind.BothEn), "$vaddr, $srsrc, $soffset idxen offen",
273 !if(!eq(addrKind, BUFAddrKind.Addr64), "$vaddr, $srsrc, $soffset addr64",
274 "")))));
275 string ret = Pfx # "$offset";
276}
277
Matt Arsenaultf3dd8632016-11-01 00:55:14 +0000278class MUBUF_SetupAddr<int addrKind> {
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000279 bits<1> offen = !if(!eq(addrKind, BUFAddrKind.OffEn), 1,
280 !if(!eq(addrKind, BUFAddrKind.BothEn), 1 , 0));
281
282 bits<1> idxen = !if(!eq(addrKind, BUFAddrKind.IdxEn), 1,
283 !if(!eq(addrKind, BUFAddrKind.BothEn), 1 , 0));
284
285 bits<1> addr64 = !if(!eq(addrKind, BUFAddrKind.Addr64), 1, 0);
286
287 bits<1> has_vaddr = !if(!eq(addrKind, BUFAddrKind.Offset), 0, 1);
288}
289
290class MUBUF_Load_Pseudo <string opName,
291 int addrKind,
292 RegisterClass vdataClass,
293 list<dag> pattern=[],
294 // Workaround bug bz30254
295 int addrKindCopy = addrKind>
296 : MUBUF_Pseudo<opName,
297 (outs vdataClass:$vdata),
298 getMUBUFIns<addrKindCopy>.ret,
299 " $vdata, " # getMUBUFAsmOps<addrKindCopy>.ret # "$glc$slc$tfe",
300 pattern>,
301 MUBUF_SetupAddr<addrKindCopy> {
302 let PseudoInstr = opName # "_" # getAddrName<addrKindCopy>.ret;
303 let mayLoad = 1;
304 let mayStore = 0;
305}
306
307// FIXME: tfe can't be an operand because it requires a separate
308// opcode because it needs an N+1 register class dest register.
309multiclass MUBUF_Pseudo_Loads<string opName, RegisterClass vdataClass,
310 ValueType load_vt = i32,
311 SDPatternOperator ld = null_frag> {
312
313 def _OFFSET : MUBUF_Load_Pseudo <opName, BUFAddrKind.Offset, vdataClass,
314 [(set load_vt:$vdata,
315 (ld (MUBUFOffset v4i32:$srsrc, i32:$soffset, i16:$offset, i1:$glc, i1:$slc, i1:$tfe)))]>,
316 MUBUFAddr64Table<0>;
317
318 def _ADDR64 : MUBUF_Load_Pseudo <opName, BUFAddrKind.Addr64, vdataClass,
319 [(set load_vt:$vdata,
320 (ld (MUBUFAddr64 v4i32:$srsrc, i64:$vaddr, i32:$soffset, i16:$offset, i1:$glc, i1:$slc, i1:$tfe)))]>,
321 MUBUFAddr64Table<1>;
322
323 def _OFFEN : MUBUF_Load_Pseudo <opName, BUFAddrKind.OffEn, vdataClass>;
324 def _IDXEN : MUBUF_Load_Pseudo <opName, BUFAddrKind.IdxEn, vdataClass>;
325 def _BOTHEN : MUBUF_Load_Pseudo <opName, BUFAddrKind.BothEn, vdataClass>;
326
327 let DisableWQM = 1 in {
328 def _OFFSET_exact : MUBUF_Load_Pseudo <opName, BUFAddrKind.Offset, vdataClass>;
329 def _OFFEN_exact : MUBUF_Load_Pseudo <opName, BUFAddrKind.OffEn, vdataClass>;
330 def _IDXEN_exact : MUBUF_Load_Pseudo <opName, BUFAddrKind.IdxEn, vdataClass>;
331 def _BOTHEN_exact : MUBUF_Load_Pseudo <opName, BUFAddrKind.BothEn, vdataClass>;
332 }
333}
334
335class MUBUF_Store_Pseudo <string opName,
336 int addrKind,
337 RegisterClass vdataClass,
338 list<dag> pattern=[],
339 // Workaround bug bz30254
340 int addrKindCopy = addrKind,
341 RegisterClass vdataClassCopy = vdataClass>
342 : MUBUF_Pseudo<opName,
343 (outs),
344 getMUBUFIns<addrKindCopy, [vdataClassCopy]>.ret,
345 " $vdata, " # getMUBUFAsmOps<addrKindCopy>.ret # "$glc$slc$tfe",
346 pattern>,
347 MUBUF_SetupAddr<addrKindCopy> {
348 let PseudoInstr = opName # "_" # getAddrName<addrKindCopy>.ret;
349 let mayLoad = 0;
350 let mayStore = 1;
351}
352
353multiclass MUBUF_Pseudo_Stores<string opName, RegisterClass vdataClass,
354 ValueType store_vt = i32,
355 SDPatternOperator st = null_frag> {
356
357 def _OFFSET : MUBUF_Store_Pseudo <opName, BUFAddrKind.Offset, vdataClass,
358 [(st store_vt:$vdata, (MUBUFOffset v4i32:$srsrc, i32:$soffset,
359 i16:$offset, i1:$glc, i1:$slc, i1:$tfe))]>,
360 MUBUFAddr64Table<0>;
361
362 def _ADDR64 : MUBUF_Store_Pseudo <opName, BUFAddrKind.Addr64, vdataClass,
363 [(st store_vt:$vdata, (MUBUFAddr64 v4i32:$srsrc, i64:$vaddr, i32:$soffset,
364 i16:$offset, i1:$glc, i1:$slc, i1:$tfe))]>,
365 MUBUFAddr64Table<1>;
366
367 def _OFFEN : MUBUF_Store_Pseudo <opName, BUFAddrKind.OffEn, vdataClass>;
368 def _IDXEN : MUBUF_Store_Pseudo <opName, BUFAddrKind.IdxEn, vdataClass>;
369 def _BOTHEN : MUBUF_Store_Pseudo <opName, BUFAddrKind.BothEn, vdataClass>;
370
371 let DisableWQM = 1 in {
372 def _OFFSET_exact : MUBUF_Store_Pseudo <opName, BUFAddrKind.Offset, vdataClass>;
373 def _OFFEN_exact : MUBUF_Store_Pseudo <opName, BUFAddrKind.OffEn, vdataClass>;
374 def _IDXEN_exact : MUBUF_Store_Pseudo <opName, BUFAddrKind.IdxEn, vdataClass>;
375 def _BOTHEN_exact : MUBUF_Store_Pseudo <opName, BUFAddrKind.BothEn, vdataClass>;
376 }
377}
378
379
380class getMUBUFAtomicInsDA<RegisterClass vdataClass, bit vdata_in,
381 list<RegisterClass> vaddrList=[]> {
382 RegisterClass vaddrClass = !if(!empty(vaddrList), ?, !head(vaddrList));
383 dag ret = !if(vdata_in,
384 !if(!empty(vaddrList),
385 (ins vdataClass:$vdata_in,
386 SReg_128:$srsrc, SCSrc_b32:$soffset, offset:$offset, slc:$slc),
387 (ins vdataClass:$vdata_in, vaddrClass:$vaddr,
388 SReg_128:$srsrc, SCSrc_b32:$soffset, offset:$offset, slc:$slc)
389 ),
390 !if(!empty(vaddrList),
391 (ins vdataClass:$vdata,
392 SReg_128:$srsrc, SCSrc_b32:$soffset, offset:$offset, slc:$slc),
393 (ins vdataClass:$vdata, vaddrClass:$vaddr,
394 SReg_128:$srsrc, SCSrc_b32:$soffset, offset:$offset, slc:$slc)
395 ));
396}
397
398class getMUBUFAtomicIns<int addrKind,
399 RegisterClass vdataClass,
400 bit vdata_in,
401 // Workaround bug bz30254
402 RegisterClass vdataClassCopy=vdataClass> {
403 dag ret =
404 !if(!eq(addrKind, BUFAddrKind.Offset),
405 getMUBUFAtomicInsDA<vdataClassCopy, vdata_in>.ret,
406 !if(!eq(addrKind, BUFAddrKind.OffEn),
407 getMUBUFAtomicInsDA<vdataClassCopy, vdata_in, [VGPR_32]>.ret,
408 !if(!eq(addrKind, BUFAddrKind.IdxEn),
409 getMUBUFAtomicInsDA<vdataClassCopy, vdata_in, [VGPR_32]>.ret,
410 !if(!eq(addrKind, BUFAddrKind.BothEn),
411 getMUBUFAtomicInsDA<vdataClassCopy, vdata_in, [VReg_64]>.ret,
412 !if(!eq(addrKind, BUFAddrKind.Addr64),
413 getMUBUFAtomicInsDA<vdataClassCopy, vdata_in, [VReg_64]>.ret,
414 (ins))))));
415}
416
417class MUBUF_Atomic_Pseudo<string opName,
418 int addrKind,
419 dag outs,
420 dag ins,
421 string asmOps,
422 list<dag> pattern=[],
423 // Workaround bug bz30254
424 int addrKindCopy = addrKind>
425 : MUBUF_Pseudo<opName, outs, ins, asmOps, pattern>,
426 MUBUF_SetupAddr<addrKindCopy> {
427 let mayStore = 1;
428 let mayLoad = 1;
429 let hasPostISelHook = 1;
430 let hasSideEffects = 1;
431 let DisableWQM = 1;
432 let has_glc = 0;
433 let has_tfe = 0;
434}
435
436class MUBUF_AtomicNoRet_Pseudo<string opName, int addrKind,
437 RegisterClass vdataClass,
438 list<dag> pattern=[],
439 // Workaround bug bz30254
440 int addrKindCopy = addrKind,
441 RegisterClass vdataClassCopy = vdataClass>
442 : MUBUF_Atomic_Pseudo<opName, addrKindCopy,
443 (outs),
444 getMUBUFAtomicIns<addrKindCopy, vdataClassCopy, 0>.ret,
445 " $vdata, " # getMUBUFAsmOps<addrKindCopy>.ret # "$slc",
446 pattern>,
447 AtomicNoRet<opName # "_" # getAddrName<addrKindCopy>.ret, 0> {
448 let PseudoInstr = opName # "_" # getAddrName<addrKindCopy>.ret;
449 let glc_value = 0;
450 let AsmMatchConverter = "cvtMubufAtomic";
451}
452
453class MUBUF_AtomicRet_Pseudo<string opName, int addrKind,
454 RegisterClass vdataClass,
455 list<dag> pattern=[],
456 // Workaround bug bz30254
457 int addrKindCopy = addrKind,
458 RegisterClass vdataClassCopy = vdataClass>
459 : MUBUF_Atomic_Pseudo<opName, addrKindCopy,
460 (outs vdataClassCopy:$vdata),
461 getMUBUFAtomicIns<addrKindCopy, vdataClassCopy, 1>.ret,
462 " $vdata, " # getMUBUFAsmOps<addrKindCopy>.ret # " glc$slc",
463 pattern>,
464 AtomicNoRet<opName # "_" # getAddrName<addrKindCopy>.ret, 1> {
465 let PseudoInstr = opName # "_rtn_" # getAddrName<addrKindCopy>.ret;
466 let glc_value = 1;
467 let Constraints = "$vdata = $vdata_in";
468 let DisableEncoding = "$vdata_in";
469 let AsmMatchConverter = "cvtMubufAtomicReturn";
470}
471
472multiclass MUBUF_Pseudo_Atomics <string opName,
473 RegisterClass vdataClass,
474 ValueType vdataType,
475 SDPatternOperator atomic> {
476
477 def _OFFSET : MUBUF_AtomicNoRet_Pseudo <opName, BUFAddrKind.Offset, vdataClass>,
478 MUBUFAddr64Table <0>;
479 def _ADDR64 : MUBUF_AtomicNoRet_Pseudo <opName, BUFAddrKind.Addr64, vdataClass>,
480 MUBUFAddr64Table <1>;
481 def _OFFEN : MUBUF_AtomicNoRet_Pseudo <opName, BUFAddrKind.OffEn, vdataClass>;
482 def _IDXEN : MUBUF_AtomicNoRet_Pseudo <opName, BUFAddrKind.IdxEn, vdataClass>;
483 def _BOTHEN : MUBUF_AtomicNoRet_Pseudo <opName, BUFAddrKind.BothEn, vdataClass>;
484
485 def _RTN_OFFSET : MUBUF_AtomicRet_Pseudo <opName, BUFAddrKind.Offset, vdataClass,
486 [(set vdataType:$vdata,
487 (atomic (MUBUFOffsetAtomic v4i32:$srsrc, i32:$soffset, i16:$offset, i1:$slc),
488 vdataType:$vdata_in))]>,
489 MUBUFAddr64Table <0, "_RTN">;
490
491 def _RTN_ADDR64 : MUBUF_AtomicRet_Pseudo <opName, BUFAddrKind.Addr64, vdataClass,
492 [(set vdataType:$vdata,
493 (atomic (MUBUFAddr64Atomic v4i32:$srsrc, i64:$vaddr, i32:$soffset, i16:$offset, i1:$slc),
494 vdataType:$vdata_in))]>,
495 MUBUFAddr64Table <1, "_RTN">;
496
497 def _RTN_OFFEN : MUBUF_AtomicRet_Pseudo <opName, BUFAddrKind.OffEn, vdataClass>;
498 def _RTN_IDXEN : MUBUF_AtomicRet_Pseudo <opName, BUFAddrKind.IdxEn, vdataClass>;
499 def _RTN_BOTHEN : MUBUF_AtomicRet_Pseudo <opName, BUFAddrKind.BothEn, vdataClass>;
500}
501
502
503//===----------------------------------------------------------------------===//
504// MUBUF Instructions
505//===----------------------------------------------------------------------===//
506
507let SubtargetPredicate = isGCN in {
508
509defm BUFFER_LOAD_FORMAT_X : MUBUF_Pseudo_Loads <
510 "buffer_load_format_x", VGPR_32
511>;
512defm BUFFER_LOAD_FORMAT_XY : MUBUF_Pseudo_Loads <
513 "buffer_load_format_xy", VReg_64
514>;
515defm BUFFER_LOAD_FORMAT_XYZ : MUBUF_Pseudo_Loads <
516 "buffer_load_format_xyz", VReg_96
517>;
518defm BUFFER_LOAD_FORMAT_XYZW : MUBUF_Pseudo_Loads <
519 "buffer_load_format_xyzw", VReg_128
520>;
521defm BUFFER_STORE_FORMAT_X : MUBUF_Pseudo_Stores <
522 "buffer_store_format_x", VGPR_32
523>;
524defm BUFFER_STORE_FORMAT_XY : MUBUF_Pseudo_Stores <
525 "buffer_store_format_xy", VReg_64
526>;
527defm BUFFER_STORE_FORMAT_XYZ : MUBUF_Pseudo_Stores <
528 "buffer_store_format_xyz", VReg_96
529>;
530defm BUFFER_STORE_FORMAT_XYZW : MUBUF_Pseudo_Stores <
531 "buffer_store_format_xyzw", VReg_128
532>;
533defm BUFFER_LOAD_UBYTE : MUBUF_Pseudo_Loads <
534 "buffer_load_ubyte", VGPR_32, i32, mubuf_az_extloadi8
535>;
536defm BUFFER_LOAD_SBYTE : MUBUF_Pseudo_Loads <
537 "buffer_load_sbyte", VGPR_32, i32, mubuf_sextloadi8
538>;
539defm BUFFER_LOAD_USHORT : MUBUF_Pseudo_Loads <
540 "buffer_load_ushort", VGPR_32, i32, mubuf_az_extloadi16
541>;
542defm BUFFER_LOAD_SSHORT : MUBUF_Pseudo_Loads <
543 "buffer_load_sshort", VGPR_32, i32, mubuf_sextloadi16
544>;
545defm BUFFER_LOAD_DWORD : MUBUF_Pseudo_Loads <
546 "buffer_load_dword", VGPR_32, i32, mubuf_load
547>;
548defm BUFFER_LOAD_DWORDX2 : MUBUF_Pseudo_Loads <
549 "buffer_load_dwordx2", VReg_64, v2i32, mubuf_load
550>;
Artem Tamazov73f1ab22016-10-07 15:53:16 +0000551defm BUFFER_LOAD_DWORDX3 : MUBUF_Pseudo_Loads <
552 "buffer_load_dwordx3", VReg_96, untyped, mubuf_load
553>;
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000554defm BUFFER_LOAD_DWORDX4 : MUBUF_Pseudo_Loads <
555 "buffer_load_dwordx4", VReg_128, v4i32, mubuf_load
556>;
557defm BUFFER_STORE_BYTE : MUBUF_Pseudo_Stores <
558 "buffer_store_byte", VGPR_32, i32, truncstorei8_global
559>;
560defm BUFFER_STORE_SHORT : MUBUF_Pseudo_Stores <
561 "buffer_store_short", VGPR_32, i32, truncstorei16_global
562>;
563defm BUFFER_STORE_DWORD : MUBUF_Pseudo_Stores <
564 "buffer_store_dword", VGPR_32, i32, global_store
565>;
566defm BUFFER_STORE_DWORDX2 : MUBUF_Pseudo_Stores <
567 "buffer_store_dwordx2", VReg_64, v2i32, global_store
568>;
Artem Tamazov73f1ab22016-10-07 15:53:16 +0000569defm BUFFER_STORE_DWORDX3 : MUBUF_Pseudo_Stores <
570 "buffer_store_dwordx3", VReg_96, untyped, global_store
571>;
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000572defm BUFFER_STORE_DWORDX4 : MUBUF_Pseudo_Stores <
573 "buffer_store_dwordx4", VReg_128, v4i32, global_store
574>;
575defm BUFFER_ATOMIC_SWAP : MUBUF_Pseudo_Atomics <
576 "buffer_atomic_swap", VGPR_32, i32, atomic_swap_global
577>;
578defm BUFFER_ATOMIC_CMPSWAP : MUBUF_Pseudo_Atomics <
579 "buffer_atomic_cmpswap", VReg_64, v2i32, null_frag
580>;
581defm BUFFER_ATOMIC_ADD : MUBUF_Pseudo_Atomics <
582 "buffer_atomic_add", VGPR_32, i32, atomic_add_global
583>;
584defm BUFFER_ATOMIC_SUB : MUBUF_Pseudo_Atomics <
585 "buffer_atomic_sub", VGPR_32, i32, atomic_sub_global
586>;
587defm BUFFER_ATOMIC_SMIN : MUBUF_Pseudo_Atomics <
588 "buffer_atomic_smin", VGPR_32, i32, atomic_min_global
589>;
590defm BUFFER_ATOMIC_UMIN : MUBUF_Pseudo_Atomics <
591 "buffer_atomic_umin", VGPR_32, i32, atomic_umin_global
592>;
593defm BUFFER_ATOMIC_SMAX : MUBUF_Pseudo_Atomics <
594 "buffer_atomic_smax", VGPR_32, i32, atomic_max_global
595>;
596defm BUFFER_ATOMIC_UMAX : MUBUF_Pseudo_Atomics <
597 "buffer_atomic_umax", VGPR_32, i32, atomic_umax_global
598>;
599defm BUFFER_ATOMIC_AND : MUBUF_Pseudo_Atomics <
600 "buffer_atomic_and", VGPR_32, i32, atomic_and_global
601>;
602defm BUFFER_ATOMIC_OR : MUBUF_Pseudo_Atomics <
603 "buffer_atomic_or", VGPR_32, i32, atomic_or_global
604>;
605defm BUFFER_ATOMIC_XOR : MUBUF_Pseudo_Atomics <
606 "buffer_atomic_xor", VGPR_32, i32, atomic_xor_global
607>;
608defm BUFFER_ATOMIC_INC : MUBUF_Pseudo_Atomics <
609 "buffer_atomic_inc", VGPR_32, i32, atomic_inc_global
610>;
611defm BUFFER_ATOMIC_DEC : MUBUF_Pseudo_Atomics <
612 "buffer_atomic_dec", VGPR_32, i32, atomic_dec_global
613>;
614defm BUFFER_ATOMIC_SWAP_X2 : MUBUF_Pseudo_Atomics <
615 "buffer_atomic_swap_x2", VReg_64, i64, atomic_swap_global
616>;
617defm BUFFER_ATOMIC_CMPSWAP_X2 : MUBUF_Pseudo_Atomics <
618 "buffer_atomic_cmpswap_x2", VReg_128, v2i64, null_frag
619>;
620defm BUFFER_ATOMIC_ADD_X2 : MUBUF_Pseudo_Atomics <
621 "buffer_atomic_add_x2", VReg_64, i64, atomic_add_global
622>;
623defm BUFFER_ATOMIC_SUB_X2 : MUBUF_Pseudo_Atomics <
624 "buffer_atomic_sub_x2", VReg_64, i64, atomic_sub_global
625>;
626defm BUFFER_ATOMIC_SMIN_X2 : MUBUF_Pseudo_Atomics <
627 "buffer_atomic_smin_x2", VReg_64, i64, atomic_min_global
628>;
629defm BUFFER_ATOMIC_UMIN_X2 : MUBUF_Pseudo_Atomics <
630 "buffer_atomic_umin_x2", VReg_64, i64, atomic_umin_global
631>;
632defm BUFFER_ATOMIC_SMAX_X2 : MUBUF_Pseudo_Atomics <
633 "buffer_atomic_smax_x2", VReg_64, i64, atomic_max_global
634>;
635defm BUFFER_ATOMIC_UMAX_X2 : MUBUF_Pseudo_Atomics <
636 "buffer_atomic_umax_x2", VReg_64, i64, atomic_umax_global
637>;
638defm BUFFER_ATOMIC_AND_X2 : MUBUF_Pseudo_Atomics <
639 "buffer_atomic_and_x2", VReg_64, i64, atomic_and_global
640>;
641defm BUFFER_ATOMIC_OR_X2 : MUBUF_Pseudo_Atomics <
642 "buffer_atomic_or_x2", VReg_64, i64, atomic_or_global
643>;
644defm BUFFER_ATOMIC_XOR_X2 : MUBUF_Pseudo_Atomics <
645 "buffer_atomic_xor_x2", VReg_64, i64, atomic_xor_global
646>;
647defm BUFFER_ATOMIC_INC_X2 : MUBUF_Pseudo_Atomics <
648 "buffer_atomic_inc_x2", VReg_64, i64, atomic_inc_global
649>;
650defm BUFFER_ATOMIC_DEC_X2 : MUBUF_Pseudo_Atomics <
651 "buffer_atomic_dec_x2", VReg_64, i64, atomic_dec_global
652>;
653
654let SubtargetPredicate = isSI in { // isn't on CI & VI
655/*
656defm BUFFER_ATOMIC_RSUB : MUBUF_Pseudo_Atomics <"buffer_atomic_rsub">;
657defm BUFFER_ATOMIC_FCMPSWAP : MUBUF_Pseudo_Atomics <"buffer_atomic_fcmpswap">;
658defm BUFFER_ATOMIC_FMIN : MUBUF_Pseudo_Atomics <"buffer_atomic_fmin">;
659defm BUFFER_ATOMIC_FMAX : MUBUF_Pseudo_Atomics <"buffer_atomic_fmax">;
660defm BUFFER_ATOMIC_RSUB_X2 : MUBUF_Pseudo_Atomics <"buffer_atomic_rsub_x2">;
661defm BUFFER_ATOMIC_FCMPSWAP_X2 : MUBUF_Pseudo_Atomics <"buffer_atomic_fcmpswap_x2">;
662defm BUFFER_ATOMIC_FMIN_X2 : MUBUF_Pseudo_Atomics <"buffer_atomic_fmin_x2">;
663defm BUFFER_ATOMIC_FMAX_X2 : MUBUF_Pseudo_Atomics <"buffer_atomic_fmax_x2">;
664*/
665
666def BUFFER_WBINVL1_SC : MUBUF_Invalidate <"buffer_wbinvl1_sc",
667 int_amdgcn_buffer_wbinvl1_sc>;
668}
669
670def BUFFER_WBINVL1 : MUBUF_Invalidate <"buffer_wbinvl1",
671 int_amdgcn_buffer_wbinvl1>;
672
673//===----------------------------------------------------------------------===//
674// MTBUF Instructions
675//===----------------------------------------------------------------------===//
676
677//def TBUFFER_LOAD_FORMAT_X : MTBUF_ <0, "tbuffer_load_format_x", []>;
678//def TBUFFER_LOAD_FORMAT_XY : MTBUF_ <1, "tbuffer_load_format_xy", []>;
679//def TBUFFER_LOAD_FORMAT_XYZ : MTBUF_ <2, "tbuffer_load_format_xyz", []>;
680def TBUFFER_LOAD_FORMAT_XYZW : MTBUF_Load_Pseudo <"tbuffer_load_format_xyzw", VReg_128>;
681def TBUFFER_STORE_FORMAT_X : MTBUF_Store_Pseudo <"tbuffer_store_format_x", VGPR_32>;
682def TBUFFER_STORE_FORMAT_XY : MTBUF_Store_Pseudo <"tbuffer_store_format_xy", VReg_64>;
683def TBUFFER_STORE_FORMAT_XYZ : MTBUF_Store_Pseudo <"tbuffer_store_format_xyz", VReg_128>;
684def TBUFFER_STORE_FORMAT_XYZW : MTBUF_Store_Pseudo <"tbuffer_store_format_xyzw", VReg_128>;
685
686} // End let SubtargetPredicate = isGCN
687
688let SubtargetPredicate = isCIVI in {
689
690//===----------------------------------------------------------------------===//
691// Instruction definitions for CI and newer.
692//===----------------------------------------------------------------------===//
693// Remaining instructions:
694// BUFFER_LOAD_DWORDX3
695// BUFFER_STORE_DWORDX3
696
697def BUFFER_WBINVL1_VOL : MUBUF_Invalidate <"buffer_wbinvl1_vol",
698 int_amdgcn_buffer_wbinvl1_vol>;
699
700} // End let SubtargetPredicate = isCIVI
701
702//===----------------------------------------------------------------------===//
703// MUBUF Patterns
704//===----------------------------------------------------------------------===//
705
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000706let Predicates = [isGCN] in {
707
708// int_SI_vs_load_input
709def : Pat<
710 (SIload_input v4i32:$tlst, imm:$attr_offset, i32:$buf_idx_vgpr),
Tom Stellard115a6152016-11-10 16:02:37 +0000711 (BUFFER_LOAD_FORMAT_XYZW_IDXEN $buf_idx_vgpr, $tlst, (i32 0), imm:$attr_offset, 0, 0, 0)
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000712>;
713
714// Offset in an 32-bit VGPR
715def : Pat <
716 (SIload_constant v4i32:$sbase, i32:$voff),
Tom Stellard115a6152016-11-10 16:02:37 +0000717 (BUFFER_LOAD_DWORD_OFFEN $voff, $sbase, (i32 0), 0, 0, 0, 0)
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000718>;
719
720
721//===----------------------------------------------------------------------===//
722// buffer_load/store_format patterns
723//===----------------------------------------------------------------------===//
724
725multiclass MUBUF_LoadIntrinsicPat<SDPatternOperator name, ValueType vt,
726 string opcode> {
727 def : Pat<
728 (vt (name v4i32:$rsrc, 0,
729 (MUBUFIntrinsicOffset i32:$soffset, i16:$offset),
730 imm:$glc, imm:$slc)),
731 (!cast<MUBUF_Pseudo>(opcode # _OFFSET) $rsrc, $soffset, (as_i16imm $offset),
732 (as_i1imm $glc), (as_i1imm $slc), 0)
733 >;
734
735 def : Pat<
736 (vt (name v4i32:$rsrc, i32:$vindex,
737 (MUBUFIntrinsicOffset i32:$soffset, i16:$offset),
738 imm:$glc, imm:$slc)),
739 (!cast<MUBUF_Pseudo>(opcode # _IDXEN) $vindex, $rsrc, $soffset, (as_i16imm $offset),
740 (as_i1imm $glc), (as_i1imm $slc), 0)
741 >;
742
743 def : Pat<
744 (vt (name v4i32:$rsrc, 0,
745 (MUBUFIntrinsicVOffset i32:$soffset, i16:$offset, i32:$voffset),
746 imm:$glc, imm:$slc)),
747 (!cast<MUBUF_Pseudo>(opcode # _OFFEN) $voffset, $rsrc, $soffset, (as_i16imm $offset),
748 (as_i1imm $glc), (as_i1imm $slc), 0)
749 >;
750
751 def : Pat<
752 (vt (name v4i32:$rsrc, i32:$vindex,
753 (MUBUFIntrinsicVOffset i32:$soffset, i16:$offset, i32:$voffset),
754 imm:$glc, imm:$slc)),
755 (!cast<MUBUF_Pseudo>(opcode # _BOTHEN)
756 (REG_SEQUENCE VReg_64, $vindex, sub0, $voffset, sub1),
757 $rsrc, $soffset, (as_i16imm $offset),
758 (as_i1imm $glc), (as_i1imm $slc), 0)
759 >;
760}
761
762defm : MUBUF_LoadIntrinsicPat<int_amdgcn_buffer_load_format, f32, "BUFFER_LOAD_FORMAT_X">;
763defm : MUBUF_LoadIntrinsicPat<int_amdgcn_buffer_load_format, v2f32, "BUFFER_LOAD_FORMAT_XY">;
764defm : MUBUF_LoadIntrinsicPat<int_amdgcn_buffer_load_format, v4f32, "BUFFER_LOAD_FORMAT_XYZW">;
765defm : MUBUF_LoadIntrinsicPat<int_amdgcn_buffer_load, f32, "BUFFER_LOAD_DWORD">;
766defm : MUBUF_LoadIntrinsicPat<int_amdgcn_buffer_load, v2f32, "BUFFER_LOAD_DWORDX2">;
767defm : MUBUF_LoadIntrinsicPat<int_amdgcn_buffer_load, v4f32, "BUFFER_LOAD_DWORDX4">;
768
769multiclass MUBUF_StoreIntrinsicPat<SDPatternOperator name, ValueType vt,
770 string opcode> {
771 def : Pat<
772 (name vt:$vdata, v4i32:$rsrc, 0,
773 (MUBUFIntrinsicOffset i32:$soffset, i16:$offset),
774 imm:$glc, imm:$slc),
775 (!cast<MUBUF_Pseudo>(opcode # _OFFSET_exact) $vdata, $rsrc, $soffset, (as_i16imm $offset),
776 (as_i1imm $glc), (as_i1imm $slc), 0)
777 >;
778
779 def : Pat<
780 (name vt:$vdata, v4i32:$rsrc, i32:$vindex,
781 (MUBUFIntrinsicOffset i32:$soffset, i16:$offset),
782 imm:$glc, imm:$slc),
783 (!cast<MUBUF_Pseudo>(opcode # _IDXEN_exact) $vdata, $vindex, $rsrc, $soffset,
784 (as_i16imm $offset), (as_i1imm $glc),
785 (as_i1imm $slc), 0)
786 >;
787
788 def : Pat<
789 (name vt:$vdata, v4i32:$rsrc, 0,
790 (MUBUFIntrinsicVOffset i32:$soffset, i16:$offset, i32:$voffset),
791 imm:$glc, imm:$slc),
792 (!cast<MUBUF_Pseudo>(opcode # _OFFEN_exact) $vdata, $voffset, $rsrc, $soffset,
793 (as_i16imm $offset), (as_i1imm $glc),
794 (as_i1imm $slc), 0)
795 >;
796
797 def : Pat<
798 (name vt:$vdata, v4i32:$rsrc, i32:$vindex,
799 (MUBUFIntrinsicVOffset i32:$soffset, i16:$offset, i32:$voffset),
800 imm:$glc, imm:$slc),
801 (!cast<MUBUF_Pseudo>(opcode # _BOTHEN_exact)
802 $vdata,
803 (REG_SEQUENCE VReg_64, $vindex, sub0, $voffset, sub1),
804 $rsrc, $soffset, (as_i16imm $offset),
805 (as_i1imm $glc), (as_i1imm $slc), 0)
806 >;
807}
808
809defm : MUBUF_StoreIntrinsicPat<int_amdgcn_buffer_store_format, f32, "BUFFER_STORE_FORMAT_X">;
810defm : MUBUF_StoreIntrinsicPat<int_amdgcn_buffer_store_format, v2f32, "BUFFER_STORE_FORMAT_XY">;
811defm : MUBUF_StoreIntrinsicPat<int_amdgcn_buffer_store_format, v4f32, "BUFFER_STORE_FORMAT_XYZW">;
812defm : MUBUF_StoreIntrinsicPat<int_amdgcn_buffer_store, f32, "BUFFER_STORE_DWORD">;
813defm : MUBUF_StoreIntrinsicPat<int_amdgcn_buffer_store, v2f32, "BUFFER_STORE_DWORDX2">;
814defm : MUBUF_StoreIntrinsicPat<int_amdgcn_buffer_store, v4f32, "BUFFER_STORE_DWORDX4">;
815
816//===----------------------------------------------------------------------===//
817// buffer_atomic patterns
818//===----------------------------------------------------------------------===//
819
820multiclass BufferAtomicPatterns<SDPatternOperator name, string opcode> {
821 def : Pat<
822 (name i32:$vdata_in, v4i32:$rsrc, 0,
823 (MUBUFIntrinsicOffset i32:$soffset, i16:$offset),
824 imm:$slc),
825 (!cast<MUBUF_Pseudo>(opcode # _RTN_OFFSET) $vdata_in, $rsrc, $soffset,
826 (as_i16imm $offset), (as_i1imm $slc))
827 >;
828
829 def : Pat<
830 (name i32:$vdata_in, v4i32:$rsrc, i32:$vindex,
831 (MUBUFIntrinsicOffset i32:$soffset, i16:$offset),
832 imm:$slc),
833 (!cast<MUBUF_Pseudo>(opcode # _RTN_IDXEN) $vdata_in, $vindex, $rsrc, $soffset,
834 (as_i16imm $offset), (as_i1imm $slc))
835 >;
836
837 def : Pat<
838 (name i32:$vdata_in, v4i32:$rsrc, 0,
839 (MUBUFIntrinsicVOffset i32:$soffset, i16:$offset, i32:$voffset),
840 imm:$slc),
841 (!cast<MUBUF_Pseudo>(opcode # _RTN_OFFEN) $vdata_in, $voffset, $rsrc, $soffset,
842 (as_i16imm $offset), (as_i1imm $slc))
843 >;
844
845 def : Pat<
846 (name i32:$vdata_in, v4i32:$rsrc, i32:$vindex,
847 (MUBUFIntrinsicVOffset i32:$soffset, i16:$offset, i32:$voffset),
848 imm:$slc),
849 (!cast<MUBUF_Pseudo>(opcode # _RTN_BOTHEN)
850 $vdata_in,
851 (REG_SEQUENCE VReg_64, $vindex, sub0, $voffset, sub1),
852 $rsrc, $soffset, (as_i16imm $offset), (as_i1imm $slc))
853 >;
854}
855
856defm : BufferAtomicPatterns<int_amdgcn_buffer_atomic_swap, "BUFFER_ATOMIC_SWAP">;
857defm : BufferAtomicPatterns<int_amdgcn_buffer_atomic_add, "BUFFER_ATOMIC_ADD">;
858defm : BufferAtomicPatterns<int_amdgcn_buffer_atomic_sub, "BUFFER_ATOMIC_SUB">;
859defm : BufferAtomicPatterns<int_amdgcn_buffer_atomic_smin, "BUFFER_ATOMIC_SMIN">;
860defm : BufferAtomicPatterns<int_amdgcn_buffer_atomic_umin, "BUFFER_ATOMIC_UMIN">;
861defm : BufferAtomicPatterns<int_amdgcn_buffer_atomic_smax, "BUFFER_ATOMIC_SMAX">;
862defm : BufferAtomicPatterns<int_amdgcn_buffer_atomic_umax, "BUFFER_ATOMIC_UMAX">;
863defm : BufferAtomicPatterns<int_amdgcn_buffer_atomic_and, "BUFFER_ATOMIC_AND">;
864defm : BufferAtomicPatterns<int_amdgcn_buffer_atomic_or, "BUFFER_ATOMIC_OR">;
865defm : BufferAtomicPatterns<int_amdgcn_buffer_atomic_xor, "BUFFER_ATOMIC_XOR">;
866
867def : Pat<
868 (int_amdgcn_buffer_atomic_cmpswap
869 i32:$data, i32:$cmp, v4i32:$rsrc, 0,
870 (MUBUFIntrinsicOffset i32:$soffset, i16:$offset),
871 imm:$slc),
872 (EXTRACT_SUBREG
873 (BUFFER_ATOMIC_CMPSWAP_RTN_OFFSET
874 (REG_SEQUENCE VReg_64, $data, sub0, $cmp, sub1),
875 $rsrc, $soffset, (as_i16imm $offset), (as_i1imm $slc)),
876 sub0)
877>;
878
879def : Pat<
880 (int_amdgcn_buffer_atomic_cmpswap
881 i32:$data, i32:$cmp, v4i32:$rsrc, i32:$vindex,
882 (MUBUFIntrinsicOffset i32:$soffset, i16:$offset),
883 imm:$slc),
884 (EXTRACT_SUBREG
885 (BUFFER_ATOMIC_CMPSWAP_RTN_IDXEN
886 (REG_SEQUENCE VReg_64, $data, sub0, $cmp, sub1),
887 $vindex, $rsrc, $soffset, (as_i16imm $offset), (as_i1imm $slc)),
888 sub0)
889>;
890
891def : Pat<
892 (int_amdgcn_buffer_atomic_cmpswap
893 i32:$data, i32:$cmp, v4i32:$rsrc, 0,
894 (MUBUFIntrinsicVOffset i32:$soffset, i16:$offset, i32:$voffset),
895 imm:$slc),
896 (EXTRACT_SUBREG
897 (BUFFER_ATOMIC_CMPSWAP_RTN_OFFEN
898 (REG_SEQUENCE VReg_64, $data, sub0, $cmp, sub1),
899 $voffset, $rsrc, $soffset, (as_i16imm $offset), (as_i1imm $slc)),
900 sub0)
901>;
902
903def : Pat<
904 (int_amdgcn_buffer_atomic_cmpswap
905 i32:$data, i32:$cmp, v4i32:$rsrc, i32:$vindex,
906 (MUBUFIntrinsicVOffset i32:$soffset, i16:$offset, i32:$voffset),
907 imm:$slc),
908 (EXTRACT_SUBREG
909 (BUFFER_ATOMIC_CMPSWAP_RTN_BOTHEN
910 (REG_SEQUENCE VReg_64, $data, sub0, $cmp, sub1),
911 (REG_SEQUENCE VReg_64, $vindex, sub0, $voffset, sub1),
912 $rsrc, $soffset, (as_i16imm $offset), (as_i1imm $slc)),
913 sub0)
914>;
915
916
Tom Stellard115a6152016-11-10 16:02:37 +0000917class MUBUFLoad_PatternADDR64 <MUBUF_Pseudo Instr_ADDR64, ValueType vt,
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000918 PatFrag constant_ld> : Pat <
919 (vt (constant_ld (MUBUFAddr64 v4i32:$srsrc, i64:$vaddr, i32:$soffset,
920 i16:$offset, i1:$glc, i1:$slc, i1:$tfe))),
921 (Instr_ADDR64 $vaddr, $srsrc, $soffset, $offset, $glc, $slc, $tfe)
922 >;
923
924multiclass MUBUFLoad_Atomic_Pattern <MUBUF_Pseudo Instr_ADDR64, MUBUF_Pseudo Instr_OFFSET,
925 ValueType vt, PatFrag atomic_ld> {
926 def : Pat <
927 (vt (atomic_ld (MUBUFAddr64 v4i32:$srsrc, i64:$vaddr, i32:$soffset,
928 i16:$offset, i1:$slc))),
929 (Instr_ADDR64 $vaddr, $srsrc, $soffset, $offset, 1, $slc, 0)
930 >;
931
932 def : Pat <
933 (vt (atomic_ld (MUBUFOffsetNoGLC v4i32:$rsrc, i32:$soffset, i16:$offset))),
934 (Instr_OFFSET $rsrc, $soffset, (as_i16imm $offset), 1, 0, 0)
935 >;
936}
937
938let Predicates = [isSICI] in {
Tom Stellard115a6152016-11-10 16:02:37 +0000939def : MUBUFLoad_PatternADDR64 <BUFFER_LOAD_SBYTE_ADDR64, i32, sextloadi8_constant>;
940def : MUBUFLoad_PatternADDR64 <BUFFER_LOAD_UBYTE_ADDR64, i32, az_extloadi8_constant>;
941def : MUBUFLoad_PatternADDR64 <BUFFER_LOAD_SSHORT_ADDR64, i32, sextloadi16_constant>;
942def : MUBUFLoad_PatternADDR64 <BUFFER_LOAD_USHORT_ADDR64, i32, az_extloadi16_constant>;
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000943
944defm : MUBUFLoad_Atomic_Pattern <BUFFER_LOAD_DWORD_ADDR64, BUFFER_LOAD_DWORD_OFFSET, i32, mubuf_load_atomic>;
945defm : MUBUFLoad_Atomic_Pattern <BUFFER_LOAD_DWORDX2_ADDR64, BUFFER_LOAD_DWORDX2_OFFSET, i64, mubuf_load_atomic>;
946} // End Predicates = [isSICI]
947
Tom Stellard115a6152016-11-10 16:02:37 +0000948multiclass MUBUFLoad_Pattern <MUBUF_Pseudo Instr_OFFSET, ValueType vt,
949 PatFrag ld> {
950
951 def : Pat <
952 (vt (ld (MUBUFOffset v4i32:$srsrc, i32:$soffset,
953 i16:$offset, i1:$glc, i1:$slc, i1:$tfe))),
954 (Instr_OFFSET $srsrc, $soffset, $offset, $glc, $slc, $tfe)
955 >;
956}
957
958let Predicates = [Has16BitInsts] in {
959
960defm : MUBUFLoad_Pattern <BUFFER_LOAD_SBYTE_OFFSET, i16, sextloadi8_constant>;
961defm : MUBUFLoad_Pattern <BUFFER_LOAD_UBYTE_OFFSET, i16, az_extloadi8_constant>;
962defm : MUBUFLoad_Pattern <BUFFER_LOAD_SBYTE_OFFSET, i16, mubuf_sextloadi8>;
963defm : MUBUFLoad_Pattern <BUFFER_LOAD_UBYTE_OFFSET, i16, mubuf_az_extloadi8>;
964
965} // End Predicates = [Has16BitInsts]
966
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000967class MUBUFScratchLoadPat <MUBUF_Pseudo Instr, ValueType vt, PatFrag ld> : Pat <
968 (vt (ld (MUBUFScratch v4i32:$srsrc, i32:$vaddr,
969 i32:$soffset, u16imm:$offset))),
970 (Instr $vaddr, $srsrc, $soffset, $offset, 0, 0, 0)
971>;
972
973def : MUBUFScratchLoadPat <BUFFER_LOAD_SBYTE_OFFEN, i32, sextloadi8_private>;
974def : MUBUFScratchLoadPat <BUFFER_LOAD_UBYTE_OFFEN, i32, extloadi8_private>;
Tom Stellard115a6152016-11-10 16:02:37 +0000975def : MUBUFScratchLoadPat <BUFFER_LOAD_SBYTE_OFFEN, i16, sextloadi8_private>;
976def : MUBUFScratchLoadPat <BUFFER_LOAD_UBYTE_OFFEN, i16, extloadi8_private>;
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +0000977def : MUBUFScratchLoadPat <BUFFER_LOAD_SSHORT_OFFEN, i32, sextloadi16_private>;
978def : MUBUFScratchLoadPat <BUFFER_LOAD_USHORT_OFFEN, i32, extloadi16_private>;
979def : MUBUFScratchLoadPat <BUFFER_LOAD_DWORD_OFFEN, i32, load_private>;
980def : MUBUFScratchLoadPat <BUFFER_LOAD_DWORDX2_OFFEN, v2i32, load_private>;
981def : MUBUFScratchLoadPat <BUFFER_LOAD_DWORDX4_OFFEN, v4i32, load_private>;
982
983// BUFFER_LOAD_DWORD*, addr64=0
984multiclass MUBUF_Load_Dword <ValueType vt,
985 MUBUF_Pseudo offset,
986 MUBUF_Pseudo offen,
987 MUBUF_Pseudo idxen,
988 MUBUF_Pseudo bothen> {
989
990 def : Pat <
991 (vt (int_SI_buffer_load_dword v4i32:$rsrc, (i32 imm), i32:$soffset,
992 imm:$offset, 0, 0, imm:$glc, imm:$slc,
993 imm:$tfe)),
994 (offset $rsrc, $soffset, (as_i16imm $offset), (as_i1imm $glc),
995 (as_i1imm $slc), (as_i1imm $tfe))
996 >;
997
998 def : Pat <
999 (vt (int_SI_buffer_load_dword v4i32:$rsrc, i32:$vaddr, i32:$soffset,
1000 imm:$offset, 1, 0, imm:$glc, imm:$slc,
1001 imm:$tfe)),
1002 (offen $vaddr, $rsrc, $soffset, (as_i16imm $offset), (as_i1imm $glc), (as_i1imm $slc),
1003 (as_i1imm $tfe))
1004 >;
1005
1006 def : Pat <
1007 (vt (int_SI_buffer_load_dword v4i32:$rsrc, i32:$vaddr, i32:$soffset,
1008 imm:$offset, 0, 1, imm:$glc, imm:$slc,
1009 imm:$tfe)),
1010 (idxen $vaddr, $rsrc, $soffset, (as_i16imm $offset), (as_i1imm $glc),
1011 (as_i1imm $slc), (as_i1imm $tfe))
1012 >;
1013
1014 def : Pat <
1015 (vt (int_SI_buffer_load_dword v4i32:$rsrc, v2i32:$vaddr, i32:$soffset,
1016 imm:$offset, 1, 1, imm:$glc, imm:$slc,
1017 imm:$tfe)),
1018 (bothen $vaddr, $rsrc, $soffset, (as_i16imm $offset), (as_i1imm $glc), (as_i1imm $slc),
1019 (as_i1imm $tfe))
1020 >;
1021}
1022
1023defm : MUBUF_Load_Dword <i32, BUFFER_LOAD_DWORD_OFFSET, BUFFER_LOAD_DWORD_OFFEN,
1024 BUFFER_LOAD_DWORD_IDXEN, BUFFER_LOAD_DWORD_BOTHEN>;
1025defm : MUBUF_Load_Dword <v2i32, BUFFER_LOAD_DWORDX2_OFFSET, BUFFER_LOAD_DWORDX2_OFFEN,
1026 BUFFER_LOAD_DWORDX2_IDXEN, BUFFER_LOAD_DWORDX2_BOTHEN>;
1027defm : MUBUF_Load_Dword <v4i32, BUFFER_LOAD_DWORDX4_OFFSET, BUFFER_LOAD_DWORDX4_OFFEN,
1028 BUFFER_LOAD_DWORDX4_IDXEN, BUFFER_LOAD_DWORDX4_BOTHEN>;
1029
1030multiclass MUBUFStore_Atomic_Pattern <MUBUF_Pseudo Instr_ADDR64, MUBUF_Pseudo Instr_OFFSET,
1031 ValueType vt, PatFrag atomic_st> {
1032 // Store follows atomic op convention so address is forst
1033 def : Pat <
1034 (atomic_st (MUBUFAddr64 v4i32:$srsrc, i64:$vaddr, i32:$soffset,
1035 i16:$offset, i1:$slc), vt:$val),
1036 (Instr_ADDR64 $val, $vaddr, $srsrc, $soffset, $offset, 1, $slc, 0)
1037 >;
1038
1039 def : Pat <
1040 (atomic_st (MUBUFOffsetNoGLC v4i32:$rsrc, i32:$soffset, i16:$offset), vt:$val),
1041 (Instr_OFFSET $val, $rsrc, $soffset, (as_i16imm $offset), 1, 0, 0)
1042 >;
1043}
1044let Predicates = [isSICI] in {
1045defm : MUBUFStore_Atomic_Pattern <BUFFER_STORE_DWORD_ADDR64, BUFFER_STORE_DWORD_OFFSET, i32, global_store_atomic>;
1046defm : MUBUFStore_Atomic_Pattern <BUFFER_STORE_DWORDX2_ADDR64, BUFFER_STORE_DWORDX2_OFFSET, i64, global_store_atomic>;
1047} // End Predicates = [isSICI]
1048
Tom Stellard115a6152016-11-10 16:02:37 +00001049
1050multiclass MUBUFStore_Pattern <MUBUF_Pseudo Instr_OFFSET, ValueType vt,
1051 PatFrag st> {
1052
1053 def : Pat <
1054 (st vt:$vdata, (MUBUFOffset v4i32:$srsrc, i32:$soffset,
1055 i16:$offset, i1:$glc, i1:$slc, i1:$tfe)),
1056 (Instr_OFFSET $vdata, $srsrc, $soffset, $offset, $glc, $slc, $tfe)
1057 >;
1058}
1059
1060defm : MUBUFStore_Pattern <BUFFER_STORE_BYTE_OFFSET, i16, truncstorei8_global>;
1061defm : MUBUFStore_Pattern <BUFFER_STORE_SHORT_OFFSET, i16, global_store>;
1062
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +00001063class MUBUFScratchStorePat <MUBUF_Pseudo Instr, ValueType vt, PatFrag st> : Pat <
1064 (st vt:$value, (MUBUFScratch v4i32:$srsrc, i32:$vaddr, i32:$soffset,
1065 u16imm:$offset)),
1066 (Instr $value, $vaddr, $srsrc, $soffset, $offset, 0, 0, 0)
1067>;
1068
1069def : MUBUFScratchStorePat <BUFFER_STORE_BYTE_OFFEN, i32, truncstorei8_private>;
1070def : MUBUFScratchStorePat <BUFFER_STORE_SHORT_OFFEN, i32, truncstorei16_private>;
Tom Stellard115a6152016-11-10 16:02:37 +00001071def : MUBUFScratchStorePat <BUFFER_STORE_BYTE_OFFEN, i16, truncstorei8_private>;
1072def : MUBUFScratchStorePat <BUFFER_STORE_SHORT_OFFEN, i16, store_private>;
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +00001073def : MUBUFScratchStorePat <BUFFER_STORE_DWORD_OFFEN, i32, store_private>;
1074def : MUBUFScratchStorePat <BUFFER_STORE_DWORDX2_OFFEN, v2i32, store_private>;
1075def : MUBUFScratchStorePat <BUFFER_STORE_DWORDX4_OFFEN, v4i32, store_private>;
1076
1077//===----------------------------------------------------------------------===//
1078// MTBUF Patterns
1079//===----------------------------------------------------------------------===//
1080
1081// TBUFFER_STORE_FORMAT_*, addr64=0
1082class MTBUF_StoreResource <ValueType vt, int num_channels, MTBUF_Pseudo opcode> : Pat<
1083 (SItbuffer_store v4i32:$rsrc, vt:$vdata, num_channels, i32:$vaddr,
1084 i32:$soffset, imm:$inst_offset, imm:$dfmt,
1085 imm:$nfmt, imm:$offen, imm:$idxen,
1086 imm:$glc, imm:$slc, imm:$tfe),
1087 (opcode
1088 $vdata, (as_i16imm $inst_offset), (as_i1imm $offen), (as_i1imm $idxen),
1089 (as_i1imm $glc), 0, (as_i8imm $dfmt), (as_i8imm $nfmt), $vaddr, $rsrc,
1090 (as_i1imm $slc), (as_i1imm $tfe), $soffset)
1091>;
1092
1093def : MTBUF_StoreResource <i32, 1, TBUFFER_STORE_FORMAT_X>;
1094def : MTBUF_StoreResource <v2i32, 2, TBUFFER_STORE_FORMAT_XY>;
1095def : MTBUF_StoreResource <v4i32, 3, TBUFFER_STORE_FORMAT_XYZ>;
1096def : MTBUF_StoreResource <v4i32, 4, TBUFFER_STORE_FORMAT_XYZW>;
1097
1098} // End let Predicates = [isGCN]
1099
1100//===----------------------------------------------------------------------===//
1101// Target instructions, move to the appropriate target TD file
1102//===----------------------------------------------------------------------===//
1103
1104//===----------------------------------------------------------------------===//
1105// SI
1106//===----------------------------------------------------------------------===//
1107
1108class MUBUF_Real_si <bits<7> op, MUBUF_Pseudo ps> :
1109 MUBUF_Real<op, ps>,
1110 Enc64,
1111 SIMCInstr<ps.PseudoInstr, SIEncodingFamily.SI> {
1112 let AssemblerPredicate=isSICI;
1113 let DecoderNamespace="SICI";
1114
1115 let Inst{11-0} = !if(ps.has_offset, offset, ?);
1116 let Inst{12} = ps.offen;
1117 let Inst{13} = ps.idxen;
1118 let Inst{14} = !if(ps.has_glc, glc, ps.glc_value);
1119 let Inst{15} = ps.addr64;
1120 let Inst{16} = lds;
1121 let Inst{24-18} = op;
1122 let Inst{31-26} = 0x38; //encoding
1123 let Inst{39-32} = !if(ps.has_vaddr, vaddr, ?);
1124 let Inst{47-40} = !if(ps.has_vdata, vdata, ?);
1125 let Inst{52-48} = !if(ps.has_srsrc, srsrc{6-2}, ?);
1126 let Inst{54} = !if(ps.has_slc, slc, ?);
1127 let Inst{55} = !if(ps.has_tfe, tfe, ?);
1128 let Inst{63-56} = !if(ps.has_soffset, soffset, ?);
1129}
1130
1131multiclass MUBUF_Real_AllAddr_si<bits<7> op> {
1132 def _OFFSET_si : MUBUF_Real_si <op, !cast<MUBUF_Pseudo>(NAME#"_OFFSET")>;
1133 def _ADDR64_si : MUBUF_Real_si <op, !cast<MUBUF_Pseudo>(NAME#"_ADDR64")>;
1134 def _OFFEN_si : MUBUF_Real_si <op, !cast<MUBUF_Pseudo>(NAME#"_OFFEN")>;
1135 def _IDXEN_si : MUBUF_Real_si <op, !cast<MUBUF_Pseudo>(NAME#"_IDXEN")>;
1136 def _BOTHEN_si : MUBUF_Real_si <op, !cast<MUBUF_Pseudo>(NAME#"_BOTHEN")>;
1137}
1138
1139multiclass MUBUF_Real_Atomic_si<bits<7> op> : MUBUF_Real_AllAddr_si<op> {
1140 def _RTN_OFFSET_si : MUBUF_Real_si <op, !cast<MUBUF_Pseudo>(NAME#"_RTN_OFFSET")>;
1141 def _RTN_ADDR64_si : MUBUF_Real_si <op, !cast<MUBUF_Pseudo>(NAME#"_RTN_ADDR64")>;
1142 def _RTN_OFFEN_si : MUBUF_Real_si <op, !cast<MUBUF_Pseudo>(NAME#"_RTN_OFFEN")>;
1143 def _RTN_IDXEN_si : MUBUF_Real_si <op, !cast<MUBUF_Pseudo>(NAME#"_RTN_IDXEN")>;
1144 def _RTN_BOTHEN_si : MUBUF_Real_si <op, !cast<MUBUF_Pseudo>(NAME#"_RTN_BOTHEN")>;
1145}
1146
1147defm BUFFER_LOAD_FORMAT_X : MUBUF_Real_AllAddr_si <0x00>;
1148defm BUFFER_LOAD_FORMAT_XY : MUBUF_Real_AllAddr_si <0x01>;
1149defm BUFFER_LOAD_FORMAT_XYZ : MUBUF_Real_AllAddr_si <0x02>;
1150defm BUFFER_LOAD_FORMAT_XYZW : MUBUF_Real_AllAddr_si <0x03>;
1151defm BUFFER_STORE_FORMAT_X : MUBUF_Real_AllAddr_si <0x04>;
1152defm BUFFER_STORE_FORMAT_XY : MUBUF_Real_AllAddr_si <0x05>;
1153defm BUFFER_STORE_FORMAT_XYZ : MUBUF_Real_AllAddr_si <0x06>;
1154defm BUFFER_STORE_FORMAT_XYZW : MUBUF_Real_AllAddr_si <0x07>;
1155defm BUFFER_LOAD_UBYTE : MUBUF_Real_AllAddr_si <0x08>;
1156defm BUFFER_LOAD_SBYTE : MUBUF_Real_AllAddr_si <0x09>;
1157defm BUFFER_LOAD_USHORT : MUBUF_Real_AllAddr_si <0x0a>;
1158defm BUFFER_LOAD_SSHORT : MUBUF_Real_AllAddr_si <0x0b>;
1159defm BUFFER_LOAD_DWORD : MUBUF_Real_AllAddr_si <0x0c>;
1160defm BUFFER_LOAD_DWORDX2 : MUBUF_Real_AllAddr_si <0x0d>;
1161defm BUFFER_LOAD_DWORDX4 : MUBUF_Real_AllAddr_si <0x0e>;
Artem Tamazov73f1ab22016-10-07 15:53:16 +00001162defm BUFFER_LOAD_DWORDX3 : MUBUF_Real_AllAddr_si <0x0f>;
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +00001163defm BUFFER_STORE_BYTE : MUBUF_Real_AllAddr_si <0x18>;
1164defm BUFFER_STORE_SHORT : MUBUF_Real_AllAddr_si <0x1a>;
1165defm BUFFER_STORE_DWORD : MUBUF_Real_AllAddr_si <0x1c>;
1166defm BUFFER_STORE_DWORDX2 : MUBUF_Real_AllAddr_si <0x1d>;
1167defm BUFFER_STORE_DWORDX4 : MUBUF_Real_AllAddr_si <0x1e>;
Artem Tamazov73f1ab22016-10-07 15:53:16 +00001168defm BUFFER_STORE_DWORDX3 : MUBUF_Real_AllAddr_si <0x1f>;
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +00001169
1170defm BUFFER_ATOMIC_SWAP : MUBUF_Real_Atomic_si <0x30>;
1171defm BUFFER_ATOMIC_CMPSWAP : MUBUF_Real_Atomic_si <0x31>;
1172defm BUFFER_ATOMIC_ADD : MUBUF_Real_Atomic_si <0x32>;
1173defm BUFFER_ATOMIC_SUB : MUBUF_Real_Atomic_si <0x33>;
1174//defm BUFFER_ATOMIC_RSUB : MUBUF_Real_Atomic_si <0x34>; // isn't on CI & VI
1175defm BUFFER_ATOMIC_SMIN : MUBUF_Real_Atomic_si <0x35>;
1176defm BUFFER_ATOMIC_UMIN : MUBUF_Real_Atomic_si <0x36>;
1177defm BUFFER_ATOMIC_SMAX : MUBUF_Real_Atomic_si <0x37>;
1178defm BUFFER_ATOMIC_UMAX : MUBUF_Real_Atomic_si <0x38>;
1179defm BUFFER_ATOMIC_AND : MUBUF_Real_Atomic_si <0x39>;
1180defm BUFFER_ATOMIC_OR : MUBUF_Real_Atomic_si <0x3a>;
1181defm BUFFER_ATOMIC_XOR : MUBUF_Real_Atomic_si <0x3b>;
1182defm BUFFER_ATOMIC_INC : MUBUF_Real_Atomic_si <0x3c>;
1183defm BUFFER_ATOMIC_DEC : MUBUF_Real_Atomic_si <0x3d>;
1184
1185//defm BUFFER_ATOMIC_FCMPSWAP : MUBUF_Real_Atomic_si <0x3e>; // isn't on VI
1186//defm BUFFER_ATOMIC_FMIN : MUBUF_Real_Atomic_si <0x3f>; // isn't on VI
1187//defm BUFFER_ATOMIC_FMAX : MUBUF_Real_Atomic_si <0x40>; // isn't on VI
1188defm BUFFER_ATOMIC_SWAP_X2 : MUBUF_Real_Atomic_si <0x50>;
1189defm BUFFER_ATOMIC_CMPSWAP_X2 : MUBUF_Real_Atomic_si <0x51>;
1190defm BUFFER_ATOMIC_ADD_X2 : MUBUF_Real_Atomic_si <0x52>;
1191defm BUFFER_ATOMIC_SUB_X2 : MUBUF_Real_Atomic_si <0x53>;
1192//defm BUFFER_ATOMIC_RSUB_X2 : MUBUF_Real_Atomic_si <0x54>; // isn't on CI & VI
1193defm BUFFER_ATOMIC_SMIN_X2 : MUBUF_Real_Atomic_si <0x55>;
1194defm BUFFER_ATOMIC_UMIN_X2 : MUBUF_Real_Atomic_si <0x56>;
1195defm BUFFER_ATOMIC_SMAX_X2 : MUBUF_Real_Atomic_si <0x57>;
1196defm BUFFER_ATOMIC_UMAX_X2 : MUBUF_Real_Atomic_si <0x58>;
1197defm BUFFER_ATOMIC_AND_X2 : MUBUF_Real_Atomic_si <0x59>;
1198defm BUFFER_ATOMIC_OR_X2 : MUBUF_Real_Atomic_si <0x5a>;
1199defm BUFFER_ATOMIC_XOR_X2 : MUBUF_Real_Atomic_si <0x5b>;
1200defm BUFFER_ATOMIC_INC_X2 : MUBUF_Real_Atomic_si <0x5c>;
1201defm BUFFER_ATOMIC_DEC_X2 : MUBUF_Real_Atomic_si <0x5d>;
Tom Stellardb133fbb2016-10-27 23:05:31 +00001202// FIXME: Need to handle hazard for BUFFER_ATOMIC_FCMPSWAP_X2 on CI.
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +00001203//defm BUFFER_ATOMIC_FCMPSWAP_X2 : MUBUF_Real_Atomic_si <0x5e">; // isn't on VI
1204//defm BUFFER_ATOMIC_FMIN_X2 : MUBUF_Real_Atomic_si <0x5f>; // isn't on VI
1205//defm BUFFER_ATOMIC_FMAX_X2 : MUBUF_Real_Atomic_si <0x60>; // isn't on VI
1206
1207def BUFFER_WBINVL1_SC_si : MUBUF_Real_si <0x70, BUFFER_WBINVL1_SC>;
1208def BUFFER_WBINVL1_si : MUBUF_Real_si <0x71, BUFFER_WBINVL1>;
1209
1210class MTBUF_Real_si <bits<3> op, MTBUF_Pseudo ps> :
Valery Pykhtinfbf2d932016-09-23 21:21:21 +00001211 MTBUF_Real<ps>,
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +00001212 SIMCInstr<ps.PseudoInstr, SIEncodingFamily.SI> {
1213 let AssemblerPredicate=isSICI;
1214 let DecoderNamespace="SICI";
Valery Pykhtinfbf2d932016-09-23 21:21:21 +00001215
1216 bits<1> addr64;
1217 let Inst{15} = addr64;
1218 let Inst{18-16} = op;
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +00001219}
1220
1221def TBUFFER_LOAD_FORMAT_XYZW_si : MTBUF_Real_si <3, TBUFFER_LOAD_FORMAT_XYZW>;
1222def TBUFFER_STORE_FORMAT_X_si : MTBUF_Real_si <4, TBUFFER_STORE_FORMAT_X>;
1223def TBUFFER_STORE_FORMAT_XY_si : MTBUF_Real_si <5, TBUFFER_STORE_FORMAT_XY>;
1224def TBUFFER_STORE_FORMAT_XYZ_si : MTBUF_Real_si <6, TBUFFER_STORE_FORMAT_XYZ>;
1225def TBUFFER_STORE_FORMAT_XYZW_si : MTBUF_Real_si <7, TBUFFER_STORE_FORMAT_XYZW>;
1226
1227
1228//===----------------------------------------------------------------------===//
1229// CI
1230//===----------------------------------------------------------------------===//
1231
1232class MUBUF_Real_ci <bits<7> op, MUBUF_Pseudo ps> :
1233 MUBUF_Real_si<op, ps> {
1234 let AssemblerPredicate=isCIOnly;
1235 let DecoderNamespace="CI";
1236}
1237
1238def BUFFER_WBINVL1_VOL_ci : MUBUF_Real_ci <0x70, BUFFER_WBINVL1_VOL>;
1239
1240
1241//===----------------------------------------------------------------------===//
1242// VI
1243//===----------------------------------------------------------------------===//
1244
1245class MUBUF_Real_vi <bits<7> op, MUBUF_Pseudo ps> :
1246 MUBUF_Real<op, ps>,
1247 Enc64,
1248 SIMCInstr<ps.PseudoInstr, SIEncodingFamily.VI> {
1249 let AssemblerPredicate=isVI;
1250 let DecoderNamespace="VI";
1251
1252 let Inst{11-0} = !if(ps.has_offset, offset, ?);
1253 let Inst{12} = ps.offen;
1254 let Inst{13} = ps.idxen;
1255 let Inst{14} = !if(ps.has_glc, glc, ps.glc_value);
1256 let Inst{16} = lds;
1257 let Inst{17} = !if(ps.has_slc, slc, ?);
1258 let Inst{24-18} = op;
1259 let Inst{31-26} = 0x38; //encoding
1260 let Inst{39-32} = !if(ps.has_vaddr, vaddr, ?);
1261 let Inst{47-40} = !if(ps.has_vdata, vdata, ?);
1262 let Inst{52-48} = !if(ps.has_srsrc, srsrc{6-2}, ?);
1263 let Inst{55} = !if(ps.has_tfe, tfe, ?);
1264 let Inst{63-56} = !if(ps.has_soffset, soffset, ?);
1265}
1266
1267multiclass MUBUF_Real_AllAddr_vi<bits<7> op> {
1268 def _OFFSET_vi : MUBUF_Real_vi <op, !cast<MUBUF_Pseudo>(NAME#"_OFFSET")>;
1269 def _OFFEN_vi : MUBUF_Real_vi <op, !cast<MUBUF_Pseudo>(NAME#"_OFFEN")>;
1270 def _IDXEN_vi : MUBUF_Real_vi <op, !cast<MUBUF_Pseudo>(NAME#"_IDXEN")>;
1271 def _BOTHEN_vi : MUBUF_Real_vi <op, !cast<MUBUF_Pseudo>(NAME#"_BOTHEN")>;
1272}
1273
1274multiclass MUBUF_Real_Atomic_vi<bits<7> op> :
1275 MUBUF_Real_AllAddr_vi<op> {
1276 def _RTN_OFFSET_vi : MUBUF_Real_vi <op, !cast<MUBUF_Pseudo>(NAME#"_RTN_OFFSET")>;
1277 def _RTN_OFFEN_vi : MUBUF_Real_vi <op, !cast<MUBUF_Pseudo>(NAME#"_RTN_OFFEN")>;
1278 def _RTN_IDXEN_vi : MUBUF_Real_vi <op, !cast<MUBUF_Pseudo>(NAME#"_RTN_IDXEN")>;
1279 def _RTN_BOTHEN_vi : MUBUF_Real_vi <op, !cast<MUBUF_Pseudo>(NAME#"_RTN_BOTHEN")>;
1280}
1281
1282defm BUFFER_LOAD_FORMAT_X : MUBUF_Real_AllAddr_vi <0x00>;
1283defm BUFFER_LOAD_FORMAT_XY : MUBUF_Real_AllAddr_vi <0x01>;
1284defm BUFFER_LOAD_FORMAT_XYZ : MUBUF_Real_AllAddr_vi <0x02>;
1285defm BUFFER_LOAD_FORMAT_XYZW : MUBUF_Real_AllAddr_vi <0x03>;
1286defm BUFFER_STORE_FORMAT_X : MUBUF_Real_AllAddr_vi <0x04>;
1287defm BUFFER_STORE_FORMAT_XY : MUBUF_Real_AllAddr_vi <0x05>;
1288defm BUFFER_STORE_FORMAT_XYZ : MUBUF_Real_AllAddr_vi <0x06>;
1289defm BUFFER_STORE_FORMAT_XYZW : MUBUF_Real_AllAddr_vi <0x07>;
1290defm BUFFER_LOAD_UBYTE : MUBUF_Real_AllAddr_vi <0x10>;
1291defm BUFFER_LOAD_SBYTE : MUBUF_Real_AllAddr_vi <0x11>;
1292defm BUFFER_LOAD_USHORT : MUBUF_Real_AllAddr_vi <0x12>;
1293defm BUFFER_LOAD_SSHORT : MUBUF_Real_AllAddr_vi <0x13>;
1294defm BUFFER_LOAD_DWORD : MUBUF_Real_AllAddr_vi <0x14>;
1295defm BUFFER_LOAD_DWORDX2 : MUBUF_Real_AllAddr_vi <0x15>;
Artem Tamazov73f1ab22016-10-07 15:53:16 +00001296defm BUFFER_LOAD_DWORDX3 : MUBUF_Real_AllAddr_vi <0x16>;
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +00001297defm BUFFER_LOAD_DWORDX4 : MUBUF_Real_AllAddr_vi <0x17>;
1298defm BUFFER_STORE_BYTE : MUBUF_Real_AllAddr_vi <0x18>;
1299defm BUFFER_STORE_SHORT : MUBUF_Real_AllAddr_vi <0x1a>;
1300defm BUFFER_STORE_DWORD : MUBUF_Real_AllAddr_vi <0x1c>;
1301defm BUFFER_STORE_DWORDX2 : MUBUF_Real_AllAddr_vi <0x1d>;
Artem Tamazov73f1ab22016-10-07 15:53:16 +00001302defm BUFFER_STORE_DWORDX3 : MUBUF_Real_AllAddr_vi <0x1e>;
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +00001303defm BUFFER_STORE_DWORDX4 : MUBUF_Real_AllAddr_vi <0x1f>;
1304
1305defm BUFFER_ATOMIC_SWAP : MUBUF_Real_Atomic_vi <0x40>;
1306defm BUFFER_ATOMIC_CMPSWAP : MUBUF_Real_Atomic_vi <0x41>;
1307defm BUFFER_ATOMIC_ADD : MUBUF_Real_Atomic_vi <0x42>;
1308defm BUFFER_ATOMIC_SUB : MUBUF_Real_Atomic_vi <0x43>;
1309defm BUFFER_ATOMIC_SMIN : MUBUF_Real_Atomic_vi <0x44>;
1310defm BUFFER_ATOMIC_UMIN : MUBUF_Real_Atomic_vi <0x45>;
1311defm BUFFER_ATOMIC_SMAX : MUBUF_Real_Atomic_vi <0x46>;
1312defm BUFFER_ATOMIC_UMAX : MUBUF_Real_Atomic_vi <0x47>;
1313defm BUFFER_ATOMIC_AND : MUBUF_Real_Atomic_vi <0x48>;
1314defm BUFFER_ATOMIC_OR : MUBUF_Real_Atomic_vi <0x49>;
1315defm BUFFER_ATOMIC_XOR : MUBUF_Real_Atomic_vi <0x4a>;
1316defm BUFFER_ATOMIC_INC : MUBUF_Real_Atomic_vi <0x4b>;
1317defm BUFFER_ATOMIC_DEC : MUBUF_Real_Atomic_vi <0x4c>;
1318
1319defm BUFFER_ATOMIC_SWAP_X2 : MUBUF_Real_Atomic_vi <0x60>;
1320defm BUFFER_ATOMIC_CMPSWAP_X2 : MUBUF_Real_Atomic_vi <0x61>;
1321defm BUFFER_ATOMIC_ADD_X2 : MUBUF_Real_Atomic_vi <0x62>;
1322defm BUFFER_ATOMIC_SUB_X2 : MUBUF_Real_Atomic_vi <0x63>;
1323defm BUFFER_ATOMIC_SMIN_X2 : MUBUF_Real_Atomic_vi <0x64>;
1324defm BUFFER_ATOMIC_UMIN_X2 : MUBUF_Real_Atomic_vi <0x65>;
1325defm BUFFER_ATOMIC_SMAX_X2 : MUBUF_Real_Atomic_vi <0x66>;
1326defm BUFFER_ATOMIC_UMAX_X2 : MUBUF_Real_Atomic_vi <0x67>;
1327defm BUFFER_ATOMIC_AND_X2 : MUBUF_Real_Atomic_vi <0x68>;
1328defm BUFFER_ATOMIC_OR_X2 : MUBUF_Real_Atomic_vi <0x69>;
1329defm BUFFER_ATOMIC_XOR_X2 : MUBUF_Real_Atomic_vi <0x6a>;
1330defm BUFFER_ATOMIC_INC_X2 : MUBUF_Real_Atomic_vi <0x6b>;
1331defm BUFFER_ATOMIC_DEC_X2 : MUBUF_Real_Atomic_vi <0x6c>;
1332
1333def BUFFER_WBINVL1_vi : MUBUF_Real_vi <0x3e, BUFFER_WBINVL1>;
1334def BUFFER_WBINVL1_VOL_vi : MUBUF_Real_vi <0x3f, BUFFER_WBINVL1_VOL>;
1335
Valery Pykhtinfbf2d932016-09-23 21:21:21 +00001336class MTBUF_Real_vi <bits<4> op, MTBUF_Pseudo ps> :
1337 MTBUF_Real<ps>,
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +00001338 SIMCInstr<ps.PseudoInstr, SIEncodingFamily.VI> {
1339 let AssemblerPredicate=isVI;
1340 let DecoderNamespace="VI";
Valery Pykhtinfbf2d932016-09-23 21:21:21 +00001341
1342 let Inst{18-15} = op;
Valery Pykhtinb66e5eb2016-09-10 13:09:16 +00001343}
1344
1345def TBUFFER_LOAD_FORMAT_XYZW_vi : MTBUF_Real_vi <3, TBUFFER_LOAD_FORMAT_XYZW>;
1346def TBUFFER_STORE_FORMAT_X_vi : MTBUF_Real_vi <4, TBUFFER_STORE_FORMAT_X>;
1347def TBUFFER_STORE_FORMAT_XY_vi : MTBUF_Real_vi <5, TBUFFER_STORE_FORMAT_XY>;
1348def TBUFFER_STORE_FORMAT_XYZ_vi : MTBUF_Real_vi <6, TBUFFER_STORE_FORMAT_XYZ>;
1349def TBUFFER_STORE_FORMAT_XYZW_vi : MTBUF_Real_vi <7, TBUFFER_STORE_FORMAT_XYZW>;
1350