blob: 28887ea4a4977aeec7e1456af013033a4f90d895 [file] [log] [blame]
Valery Pykhtin902db312016-08-01 14:21:30 +00001//===-- DSInstructions.td - DS 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
10class DS_Pseudo <string opName, dag outs, dag ins, string asmOps, list<dag> pattern=[]> :
11 InstSI <outs, ins, "", pattern>,
12 SIMCInstr <opName, SIEncodingFamily.NONE> {
13
14 let SubtargetPredicate = isGCN;
15
16 let LGKM_CNT = 1;
17 let DS = 1;
Matt Arsenault10c17ca2016-10-06 10:13:23 +000018 let Size = 8;
Valery Pykhtin902db312016-08-01 14:21:30 +000019 let UseNamedOperandTable = 1;
Valery Pykhtin902db312016-08-01 14:21:30 +000020
21 // Most instruction load and store data, so set this as the default.
22 let mayLoad = 1;
23 let mayStore = 1;
24
25 let hasSideEffects = 0;
26 let SchedRW = [WriteLDS];
27
28 let isPseudo = 1;
29 let isCodeGenOnly = 1;
30
31 let AsmMatchConverter = "cvtDS";
32
33 string Mnemonic = opName;
34 string AsmOperands = asmOps;
35
36 // Well these bits a kind of hack because it would be more natural
37 // to test "outs" and "ins" dags for the presence of particular operands
38 bits<1> has_vdst = 1;
39 bits<1> has_addr = 1;
40 bits<1> has_data0 = 1;
41 bits<1> has_data1 = 1;
42
43 bits<1> has_offset = 1; // has "offset" that should be split to offset0,1
44 bits<1> has_offset0 = 1;
45 bits<1> has_offset1 = 1;
46
47 bits<1> has_gds = 1;
48 bits<1> gdsValue = 0; // if has_gds == 0 set gds to this value
Matt Arsenault10c472d2017-11-15 01:34:06 +000049
50 bits<1> has_m0_read = 1;
51
52 let Uses = !if(has_m0_read, [M0, EXEC], [EXEC]);
Valery Pykhtin902db312016-08-01 14:21:30 +000053}
54
55class DS_Real <DS_Pseudo ds> :
56 InstSI <ds.OutOperandList, ds.InOperandList, ds.Mnemonic # " " # ds.AsmOperands, []>,
57 Enc64 {
58
59 let isPseudo = 0;
60 let isCodeGenOnly = 0;
61
62 // copy relevant pseudo op flags
63 let SubtargetPredicate = ds.SubtargetPredicate;
64 let AsmMatchConverter = ds.AsmMatchConverter;
65
66 // encoding fields
67 bits<8> vdst;
68 bits<1> gds;
69 bits<8> addr;
70 bits<8> data0;
71 bits<8> data1;
72 bits<8> offset0;
73 bits<8> offset1;
74
75 bits<16> offset;
76 let offset0 = !if(ds.has_offset, offset{7-0}, ?);
77 let offset1 = !if(ds.has_offset, offset{15-8}, ?);
78}
79
80
81// DS Pseudo instructions
82
83class DS_1A1D_NORET<string opName, RegisterClass rc = VGPR_32>
84: DS_Pseudo<opName,
85 (outs),
86 (ins VGPR_32:$addr, rc:$data0, offset:$offset, gds:$gds),
Matt Arsenault10c472d2017-11-15 01:34:06 +000087 "$addr, $data0$offset$gds"> {
Valery Pykhtin902db312016-08-01 14:21:30 +000088
89 let has_data1 = 0;
90 let has_vdst = 0;
91}
92
Matt Arsenault10c472d2017-11-15 01:34:06 +000093multiclass DS_1A1D_NORET_mc<string opName, RegisterClass rc = VGPR_32> {
94 def "" : DS_1A1D_NORET<opName, rc>,
95 AtomicNoRet<opName, 0>;
96
97 let has_m0_read = 0 in {
98 def _gfx9 : DS_1A1D_NORET<opName, rc>,
99 AtomicNoRet<opName#"_gfx9", 0>;
100 }
101}
102
Matt Arsenaultf3dd8632016-11-01 00:55:14 +0000103class DS_1A2D_NORET<string opName, RegisterClass rc = VGPR_32>
Valery Pykhtin902db312016-08-01 14:21:30 +0000104: DS_Pseudo<opName,
105 (outs),
106 (ins VGPR_32:$addr, rc:$data0, rc:$data1, offset:$offset, gds:$gds),
Matt Arsenault10c472d2017-11-15 01:34:06 +0000107 "$addr, $data0, $data1"#"$offset"#"$gds"> {
Valery Pykhtin902db312016-08-01 14:21:30 +0000108
109 let has_vdst = 0;
110}
111
Matt Arsenault10c472d2017-11-15 01:34:06 +0000112multiclass DS_1A2D_NORET_mc<string opName, RegisterClass rc = VGPR_32> {
113 def "" : DS_1A2D_NORET<opName, rc>,
114 AtomicNoRet<opName, 0>;
115
116 let has_m0_read = 0 in {
117 def _gfx9 : DS_1A2D_NORET<opName, rc>,
118 AtomicNoRet<opName#"_gfx9", 0>;
119 }
120}
121
Valery Pykhtin902db312016-08-01 14:21:30 +0000122class DS_1A2D_Off8_NORET <string opName, RegisterClass rc = VGPR_32>
123: DS_Pseudo<opName,
124 (outs),
125 (ins VGPR_32:$addr, rc:$data0, rc:$data1,
126 offset0:$offset0, offset1:$offset1, gds:$gds),
127 "$addr, $data0, $data1$offset0$offset1$gds"> {
128
129 let has_vdst = 0;
130 let has_offset = 0;
131 let AsmMatchConverter = "cvtDSOffset01";
132}
133
Matt Arsenault10c472d2017-11-15 01:34:06 +0000134multiclass DS_1A2D_Off8_NORET_mc <string opName, RegisterClass rc = VGPR_32> {
135 def "" : DS_1A2D_Off8_NORET<opName, rc>;
136
137 let has_m0_read = 0 in {
138 def _gfx9 : DS_1A2D_Off8_NORET<opName, rc>;
139 }
140}
141
Valery Pykhtin902db312016-08-01 14:21:30 +0000142class DS_1A1D_RET <string opName, RegisterClass rc = VGPR_32>
143: DS_Pseudo<opName,
144 (outs rc:$vdst),
145 (ins VGPR_32:$addr, rc:$data0, offset:$offset, gds:$gds),
146 "$vdst, $addr, $data0$offset$gds"> {
147
148 let hasPostISelHook = 1;
149 let has_data1 = 0;
150}
151
Matt Arsenault10c472d2017-11-15 01:34:06 +0000152multiclass DS_1A1D_RET_mc <string opName, RegisterClass rc = VGPR_32,
153 string NoRetOp = ""> {
154 def "" : DS_1A1D_RET<opName, rc>,
155 AtomicNoRet<NoRetOp, !if(!eq(NoRetOp, ""), 0, 1)>;
156
157 let has_m0_read = 0 in {
158 def _gfx9 : DS_1A1D_RET<opName, rc>,
159 AtomicNoRet<!if(!eq(NoRetOp, ""), "", NoRetOp#"_gfx9"),
160 !if(!eq(NoRetOp, ""), 0, 1)>;
161 }
162}
163
Valery Pykhtin902db312016-08-01 14:21:30 +0000164class DS_1A2D_RET<string opName,
Matt Arsenaultf3dd8632016-11-01 00:55:14 +0000165 RegisterClass rc = VGPR_32,
Valery Pykhtin902db312016-08-01 14:21:30 +0000166 RegisterClass src = rc>
167: DS_Pseudo<opName,
168 (outs rc:$vdst),
169 (ins VGPR_32:$addr, src:$data0, src:$data1, offset:$offset, gds:$gds),
170 "$vdst, $addr, $data0, $data1$offset$gds"> {
171
172 let hasPostISelHook = 1;
173}
174
Matt Arsenault10c472d2017-11-15 01:34:06 +0000175multiclass DS_1A2D_RET_mc<string opName,
176 RegisterClass rc = VGPR_32,
177 string NoRetOp = "",
178 RegisterClass src = rc> {
179 def "" : DS_1A2D_RET<opName, rc, src>,
180 AtomicNoRet<NoRetOp, !if(!eq(NoRetOp, ""), 0, 1)>;
181
182 let has_m0_read = 0 in {
183 def _gfx9 : DS_1A2D_RET<opName, rc, src>,
184 AtomicNoRet<NoRetOp#"_gfx9", !if(!eq(NoRetOp, ""), 0, 1)>;
185 }
186}
187
Dmitry Preobrazhensky7184c442017-04-12 14:29:45 +0000188class DS_1A2D_Off8_RET<string opName,
189 RegisterClass rc = VGPR_32,
190 RegisterClass src = rc>
191: DS_Pseudo<opName,
192 (outs rc:$vdst),
193 (ins VGPR_32:$addr, src:$data0, src:$data1, offset0:$offset0, offset1:$offset1, gds:$gds),
194 "$vdst, $addr, $data0, $data1$offset0$offset1$gds"> {
195
196 let has_offset = 0;
197 let AsmMatchConverter = "cvtDSOffset01";
198
199 let hasPostISelHook = 1;
200}
201
Matt Arsenault10c472d2017-11-15 01:34:06 +0000202multiclass DS_1A2D_Off8_RET_mc<string opName,
203 RegisterClass rc = VGPR_32,
204 RegisterClass src = rc> {
205 def "" : DS_1A2D_Off8_RET<opName, rc, src>;
206
207 let has_m0_read = 0 in {
208 def _gfx9 : DS_1A2D_Off8_RET<opName, rc, src>;
209 }
210}
211
212
Matt Arsenaultb81495d2017-09-20 05:01:53 +0000213class DS_1A_RET<string opName, RegisterClass rc = VGPR_32, bit HasTiedOutput = 0, Operand ofs = offset>
Valery Pykhtin902db312016-08-01 14:21:30 +0000214: DS_Pseudo<opName,
215 (outs rc:$vdst),
Matt Arsenaultb81495d2017-09-20 05:01:53 +0000216 !if(HasTiedOutput,
217 (ins VGPR_32:$addr, ofs:$offset, gds:$gds, rc:$vdst_in),
218 (ins VGPR_32:$addr, ofs:$offset, gds:$gds)),
Valery Pykhtin902db312016-08-01 14:21:30 +0000219 "$vdst, $addr$offset$gds"> {
Matt Arsenaultb81495d2017-09-20 05:01:53 +0000220 let Constraints = !if(HasTiedOutput, "$vdst = $vdst_in", "");
221 let DisableEncoding = !if(HasTiedOutput, "$vdst_in", "");
Valery Pykhtin902db312016-08-01 14:21:30 +0000222 let has_data0 = 0;
223 let has_data1 = 0;
224}
225
Matt Arsenault10c472d2017-11-15 01:34:06 +0000226multiclass DS_1A_RET_mc<string opName, RegisterClass rc = VGPR_32, bit HasTiedOutput = 0, Operand ofs = offset> {
227 def "" : DS_1A_RET<opName, rc, HasTiedOutput, ofs>;
228
229 let has_m0_read = 0 in {
230 def _gfx9 : DS_1A_RET<opName, rc, HasTiedOutput, ofs>;
231 }
232}
233
Matt Arsenaultb81495d2017-09-20 05:01:53 +0000234class DS_1A_RET_Tied<string opName, RegisterClass rc = VGPR_32> :
235 DS_1A_RET<opName, rc, 1>;
236
Valery Pykhtin902db312016-08-01 14:21:30 +0000237class DS_1A_Off8_RET <string opName, RegisterClass rc = VGPR_32>
238: DS_Pseudo<opName,
239 (outs rc:$vdst),
240 (ins VGPR_32:$addr, offset0:$offset0, offset1:$offset1, gds:$gds),
241 "$vdst, $addr$offset0$offset1$gds"> {
242
243 let has_offset = 0;
244 let has_data0 = 0;
245 let has_data1 = 0;
246 let AsmMatchConverter = "cvtDSOffset01";
247}
248
Matt Arsenault10c472d2017-11-15 01:34:06 +0000249multiclass DS_1A_Off8_RET_mc <string opName, RegisterClass rc = VGPR_32> {
250 def "" : DS_1A_Off8_RET<opName, rc>;
251
252 let has_m0_read = 0 in {
253 def _gfx9 : DS_1A_Off8_RET<opName, rc>;
254 }
255}
256
Valery Pykhtin902db312016-08-01 14:21:30 +0000257class DS_1A_RET_GDS <string opName> : DS_Pseudo<opName,
258 (outs VGPR_32:$vdst),
259 (ins VGPR_32:$addr, offset:$offset),
260 "$vdst, $addr$offset gds"> {
261
262 let has_data0 = 0;
263 let has_data1 = 0;
264 let has_gds = 0;
265 let gdsValue = 1;
Artem Tamazov43b61562017-02-03 12:47:30 +0000266 let AsmMatchConverter = "cvtDSGds";
Valery Pykhtin902db312016-08-01 14:21:30 +0000267}
268
269class DS_0A_RET <string opName> : DS_Pseudo<opName,
270 (outs VGPR_32:$vdst),
271 (ins offset:$offset, gds:$gds),
272 "$vdst$offset$gds"> {
273
274 let mayLoad = 1;
275 let mayStore = 1;
276
277 let has_addr = 0;
278 let has_data0 = 0;
279 let has_data1 = 0;
280}
281
282class DS_1A <string opName> : DS_Pseudo<opName,
283 (outs),
284 (ins VGPR_32:$addr, offset:$offset, gds:$gds),
285 "$addr$offset$gds"> {
286
287 let mayLoad = 1;
288 let mayStore = 1;
289
290 let has_vdst = 0;
291 let has_data0 = 0;
292 let has_data1 = 0;
293}
294
Matt Arsenault10c472d2017-11-15 01:34:06 +0000295multiclass DS_1A_mc <string opName> {
296 def "" : DS_1A<opName>;
297
298 let has_m0_read = 0 in {
299 def _gfx9 : DS_1A<opName>;
300 }
301}
302
303
Dmitry Preobrazhenskye5147242017-04-07 13:07:13 +0000304class DS_GWS <string opName, dag ins, string asmOps>
305: DS_Pseudo<opName, (outs), ins, asmOps> {
Valery Pykhtin902db312016-08-01 14:21:30 +0000306
Dmitry Preobrazhenskye5147242017-04-07 13:07:13 +0000307 let has_vdst = 0;
308 let has_addr = 0;
309 let has_data0 = 0;
310 let has_data1 = 0;
Valery Pykhtin902db312016-08-01 14:21:30 +0000311
Dmitry Preobrazhenskye5147242017-04-07 13:07:13 +0000312 let has_gds = 0;
313 let gdsValue = 1;
314 let AsmMatchConverter = "cvtDSGds";
315}
316
317class DS_GWS_0D <string opName>
318: DS_GWS<opName,
319 (ins offset:$offset, gds:$gds), "$offset gds">;
320
321class DS_GWS_1D <string opName>
322: DS_GWS<opName,
323 (ins VGPR_32:$data0, offset:$offset, gds:$gds), "$data0$offset gds"> {
324
325 let has_data0 = 1;
Valery Pykhtin902db312016-08-01 14:21:30 +0000326}
327
Matt Arsenault78124982017-02-28 20:15:46 +0000328class DS_VOID <string opName> : DS_Pseudo<opName,
329 (outs), (ins), ""> {
330 let mayLoad = 0;
331 let mayStore = 0;
332 let hasSideEffects = 1;
333 let UseNamedOperandTable = 0;
334 let AsmMatchConverter = "";
335
336 let has_vdst = 0;
337 let has_addr = 0;
338 let has_data0 = 0;
339 let has_data1 = 0;
340 let has_offset = 0;
341 let has_offset0 = 0;
342 let has_offset1 = 0;
343 let has_gds = 0;
344}
345
Valery Pykhtin902db312016-08-01 14:21:30 +0000346class DS_1A1D_PERMUTE <string opName, SDPatternOperator node = null_frag>
347: DS_Pseudo<opName,
348 (outs VGPR_32:$vdst),
349 (ins VGPR_32:$addr, VGPR_32:$data0, offset:$offset),
350 "$vdst, $addr, $data0$offset",
351 [(set i32:$vdst,
352 (node (DS1Addr1Offset i32:$addr, i16:$offset), i32:$data0))] > {
353
354 let mayLoad = 0;
355 let mayStore = 0;
356 let isConvergent = 1;
357
358 let has_data1 = 0;
359 let has_gds = 0;
360}
361
Matt Arsenault10c472d2017-11-15 01:34:06 +0000362defm DS_ADD_U32 : DS_1A1D_NORET_mc<"ds_add_u32">;
363defm DS_SUB_U32 : DS_1A1D_NORET_mc<"ds_sub_u32">;
364defm DS_RSUB_U32 : DS_1A1D_NORET_mc<"ds_rsub_u32">;
365defm DS_INC_U32 : DS_1A1D_NORET_mc<"ds_inc_u32">;
366defm DS_DEC_U32 : DS_1A1D_NORET_mc<"ds_dec_u32">;
367defm DS_MIN_I32 : DS_1A1D_NORET_mc<"ds_min_i32">;
368defm DS_MAX_I32 : DS_1A1D_NORET_mc<"ds_max_i32">;
369defm DS_MIN_U32 : DS_1A1D_NORET_mc<"ds_min_u32">;
370defm DS_MAX_U32 : DS_1A1D_NORET_mc<"ds_max_u32">;
371defm DS_AND_B32 : DS_1A1D_NORET_mc<"ds_and_b32">;
372defm DS_OR_B32 : DS_1A1D_NORET_mc<"ds_or_b32">;
373defm DS_XOR_B32 : DS_1A1D_NORET_mc<"ds_xor_b32">;
374defm DS_ADD_F32 : DS_1A1D_NORET_mc<"ds_add_f32">;
375defm DS_MIN_F32 : DS_1A1D_NORET_mc<"ds_min_f32">;
376defm DS_MAX_F32 : DS_1A1D_NORET_mc<"ds_max_f32">;
Valery Pykhtin902db312016-08-01 14:21:30 +0000377
378let mayLoad = 0 in {
Matt Arsenault10c472d2017-11-15 01:34:06 +0000379defm DS_WRITE_B8 : DS_1A1D_NORET_mc<"ds_write_b8">;
380defm DS_WRITE_B16 : DS_1A1D_NORET_mc<"ds_write_b16">;
381defm DS_WRITE_B32 : DS_1A1D_NORET_mc<"ds_write_b32">;
382defm DS_WRITE2_B32 : DS_1A2D_Off8_NORET_mc<"ds_write2_b32">;
383defm DS_WRITE2ST64_B32: DS_1A2D_Off8_NORET_mc<"ds_write2st64_b32">;
384
385
386let has_m0_read = 0 in {
Matt Arsenaulted6e8f02017-09-01 18:36:06 +0000387
388let SubtargetPredicate = HasD16LoadStore in {
389def DS_WRITE_B8_D16_HI : DS_1A1D_NORET<"ds_write_b8_d16_hi">;
390def DS_WRITE_B16_D16_HI : DS_1A1D_NORET<"ds_write_b16_d16_hi">;
391}
392
Matt Arsenaultefa1d652017-09-01 18:38:02 +0000393let SubtargetPredicate = HasDSAddTid in {
394def DS_WRITE_ADDTID_B32 : DS_1A1D_NORET<"ds_write_addtid_b32">;
395}
396
Matt Arsenault10c472d2017-11-15 01:34:06 +0000397} // End has_m0_read = 0
398} // End mayLoad = 0
Valery Pykhtin902db312016-08-01 14:21:30 +0000399
Matt Arsenault10c472d2017-11-15 01:34:06 +0000400defm DS_MSKOR_B32 : DS_1A2D_NORET_mc<"ds_mskor_b32">;
401defm DS_CMPST_B32 : DS_1A2D_NORET_mc<"ds_cmpst_b32">;
402defm DS_CMPST_F32 : DS_1A2D_NORET_mc<"ds_cmpst_f32">;
Valery Pykhtin902db312016-08-01 14:21:30 +0000403
Matt Arsenault10c472d2017-11-15 01:34:06 +0000404defm DS_ADD_U64 : DS_1A1D_NORET_mc<"ds_add_u64", VReg_64>;
405defm DS_SUB_U64 : DS_1A1D_NORET_mc<"ds_sub_u64", VReg_64>;
406defm DS_RSUB_U64 : DS_1A1D_NORET_mc<"ds_rsub_u64", VReg_64>;
407defm DS_INC_U64 : DS_1A1D_NORET_mc<"ds_inc_u64", VReg_64>;
408defm DS_DEC_U64 : DS_1A1D_NORET_mc<"ds_dec_u64", VReg_64>;
409defm DS_MIN_I64 : DS_1A1D_NORET_mc<"ds_min_i64", VReg_64>;
410defm DS_MAX_I64 : DS_1A1D_NORET_mc<"ds_max_i64", VReg_64>;
411defm DS_MIN_U64 : DS_1A1D_NORET_mc<"ds_min_u64", VReg_64>;
412defm DS_MAX_U64 : DS_1A1D_NORET_mc<"ds_max_u64", VReg_64>;
413defm DS_AND_B64 : DS_1A1D_NORET_mc<"ds_and_b64", VReg_64>;
414defm DS_OR_B64 : DS_1A1D_NORET_mc<"ds_or_b64", VReg_64>;
415defm DS_XOR_B64 : DS_1A1D_NORET_mc<"ds_xor_b64", VReg_64>;
416defm DS_MSKOR_B64 : DS_1A2D_NORET_mc<"ds_mskor_b64", VReg_64>;
Valery Pykhtin902db312016-08-01 14:21:30 +0000417let mayLoad = 0 in {
Matt Arsenault10c472d2017-11-15 01:34:06 +0000418defm DS_WRITE_B64 : DS_1A1D_NORET_mc<"ds_write_b64", VReg_64>;
419defm DS_WRITE2_B64 : DS_1A2D_Off8_NORET_mc<"ds_write2_b64", VReg_64>;
420defm DS_WRITE2ST64_B64: DS_1A2D_Off8_NORET_mc<"ds_write2st64_b64", VReg_64>;
Valery Pykhtin902db312016-08-01 14:21:30 +0000421}
Matt Arsenault10c472d2017-11-15 01:34:06 +0000422defm DS_CMPST_B64 : DS_1A2D_NORET_mc<"ds_cmpst_b64", VReg_64>;
423defm DS_CMPST_F64 : DS_1A2D_NORET_mc<"ds_cmpst_f64", VReg_64>;
424defm DS_MIN_F64 : DS_1A1D_NORET_mc<"ds_min_f64", VReg_64>;
425defm DS_MAX_F64 : DS_1A1D_NORET_mc<"ds_max_f64", VReg_64>;
Valery Pykhtin902db312016-08-01 14:21:30 +0000426
Matt Arsenault10c472d2017-11-15 01:34:06 +0000427defm DS_ADD_RTN_U32 : DS_1A1D_RET_mc<"ds_add_rtn_u32", VGPR_32, "ds_add_u32">;
428defm DS_ADD_RTN_F32 : DS_1A1D_RET_mc<"ds_add_rtn_f32", VGPR_32, "ds_add_f32">;
429defm DS_SUB_RTN_U32 : DS_1A1D_RET_mc<"ds_sub_rtn_u32", VGPR_32, "ds_sub_u32">;
430defm DS_RSUB_RTN_U32 : DS_1A1D_RET_mc<"ds_rsub_rtn_u32", VGPR_32, "ds_rsub_u32">;
431defm DS_INC_RTN_U32 : DS_1A1D_RET_mc<"ds_inc_rtn_u32", VGPR_32, "ds_inc_u32">;
432defm DS_DEC_RTN_U32 : DS_1A1D_RET_mc<"ds_dec_rtn_u32", VGPR_32, "ds_dec_u32">;
433defm DS_MIN_RTN_I32 : DS_1A1D_RET_mc<"ds_min_rtn_i32", VGPR_32, "ds_min_i32">;
434defm DS_MAX_RTN_I32 : DS_1A1D_RET_mc<"ds_max_rtn_i32", VGPR_32, "ds_max_i32">;
435defm DS_MIN_RTN_U32 : DS_1A1D_RET_mc<"ds_min_rtn_u32", VGPR_32, "ds_min_u32">;
436defm DS_MAX_RTN_U32 : DS_1A1D_RET_mc<"ds_max_rtn_u32", VGPR_32, "ds_max_u32">;
437defm DS_AND_RTN_B32 : DS_1A1D_RET_mc<"ds_and_rtn_b32", VGPR_32, "ds_and_b32">;
438defm DS_OR_RTN_B32 : DS_1A1D_RET_mc<"ds_or_rtn_b32", VGPR_32, "ds_or_b32">;
439defm DS_XOR_RTN_B32 : DS_1A1D_RET_mc<"ds_xor_rtn_b32", VGPR_32, "ds_xor_b32">;
440defm DS_MSKOR_RTN_B32 : DS_1A2D_RET_mc<"ds_mskor_rtn_b32", VGPR_32, "ds_mskor_b32">;
441defm DS_CMPST_RTN_B32 : DS_1A2D_RET_mc<"ds_cmpst_rtn_b32", VGPR_32, "ds_cmpst_b32">;
442defm DS_CMPST_RTN_F32 : DS_1A2D_RET_mc<"ds_cmpst_rtn_f32", VGPR_32, "ds_cmpst_f32">;
Daniil Fukalovd5fca552018-01-17 14:05:05 +0000443defm DS_MIN_RTN_F32 : DS_1A1D_RET_mc<"ds_min_rtn_f32", VGPR_32, "ds_min_f32">;
Matt Arsenault10c472d2017-11-15 01:34:06 +0000444defm DS_MAX_RTN_F32 : DS_1A1D_RET_mc<"ds_max_rtn_f32", VGPR_32, "ds_max_f32">;
Valery Pykhtin902db312016-08-01 14:21:30 +0000445
Matt Arsenault10c472d2017-11-15 01:34:06 +0000446defm DS_WRXCHG_RTN_B32 : DS_1A1D_RET_mc<"ds_wrxchg_rtn_b32">;
447defm DS_WRXCHG2_RTN_B32 : DS_1A2D_Off8_RET_mc<"ds_wrxchg2_rtn_b32", VReg_64, VGPR_32>;
448defm DS_WRXCHG2ST64_RTN_B32 : DS_1A2D_Off8_RET_mc<"ds_wrxchg2st64_rtn_b32", VReg_64, VGPR_32>;
Valery Pykhtin902db312016-08-01 14:21:30 +0000449
Matt Arsenault10c472d2017-11-15 01:34:06 +0000450defm DS_ADD_RTN_U64 : DS_1A1D_RET_mc<"ds_add_rtn_u64", VReg_64, "ds_add_u64">;
451defm DS_SUB_RTN_U64 : DS_1A1D_RET_mc<"ds_sub_rtn_u64", VReg_64, "ds_sub_u64">;
452defm DS_RSUB_RTN_U64 : DS_1A1D_RET_mc<"ds_rsub_rtn_u64", VReg_64, "ds_rsub_u64">;
453defm DS_INC_RTN_U64 : DS_1A1D_RET_mc<"ds_inc_rtn_u64", VReg_64, "ds_inc_u64">;
454defm DS_DEC_RTN_U64 : DS_1A1D_RET_mc<"ds_dec_rtn_u64", VReg_64, "ds_dec_u64">;
455defm DS_MIN_RTN_I64 : DS_1A1D_RET_mc<"ds_min_rtn_i64", VReg_64, "ds_min_i64">;
456defm DS_MAX_RTN_I64 : DS_1A1D_RET_mc<"ds_max_rtn_i64", VReg_64, "ds_max_i64">;
457defm DS_MIN_RTN_U64 : DS_1A1D_RET_mc<"ds_min_rtn_u64", VReg_64, "ds_min_u64">;
458defm DS_MAX_RTN_U64 : DS_1A1D_RET_mc<"ds_max_rtn_u64", VReg_64, "ds_max_u64">;
459defm DS_AND_RTN_B64 : DS_1A1D_RET_mc<"ds_and_rtn_b64", VReg_64, "ds_and_b64">;
460defm DS_OR_RTN_B64 : DS_1A1D_RET_mc<"ds_or_rtn_b64", VReg_64, "ds_or_b64">;
461defm DS_XOR_RTN_B64 : DS_1A1D_RET_mc<"ds_xor_rtn_b64", VReg_64, "ds_xor_b64">;
462defm DS_MSKOR_RTN_B64 : DS_1A2D_RET_mc<"ds_mskor_rtn_b64", VReg_64, "ds_mskor_b64">;
463defm DS_CMPST_RTN_B64 : DS_1A2D_RET_mc<"ds_cmpst_rtn_b64", VReg_64, "ds_cmpst_b64">;
464defm DS_CMPST_RTN_F64 : DS_1A2D_RET_mc<"ds_cmpst_rtn_f64", VReg_64, "ds_cmpst_f64">;
465defm DS_MIN_RTN_F64 : DS_1A1D_RET_mc<"ds_min_rtn_f64", VReg_64, "ds_min_f64">;
466defm DS_MAX_RTN_F64 : DS_1A1D_RET_mc<"ds_max_rtn_f64", VReg_64, "ds_max_f64">;
Valery Pykhtin902db312016-08-01 14:21:30 +0000467
Matt Arsenault10c472d2017-11-15 01:34:06 +0000468defm DS_WRXCHG_RTN_B64 : DS_1A1D_RET_mc<"ds_wrxchg_rtn_b64", VReg_64>;
469defm DS_WRXCHG2_RTN_B64 : DS_1A2D_Off8_RET_mc<"ds_wrxchg2_rtn_b64", VReg_128, VReg_64>;
470defm DS_WRXCHG2ST64_RTN_B64 : DS_1A2D_Off8_RET_mc<"ds_wrxchg2st64_rtn_b64", VReg_128, VReg_64>;
Valery Pykhtin902db312016-08-01 14:21:30 +0000471
Dmitry Preobrazhenskye5147242017-04-07 13:07:13 +0000472def DS_GWS_INIT : DS_GWS_1D<"ds_gws_init">;
473def DS_GWS_SEMA_V : DS_GWS_0D<"ds_gws_sema_v">;
474def DS_GWS_SEMA_BR : DS_GWS_1D<"ds_gws_sema_br">;
475def DS_GWS_SEMA_P : DS_GWS_0D<"ds_gws_sema_p">;
476def DS_GWS_BARRIER : DS_GWS_1D<"ds_gws_barrier">;
Valery Pykhtin902db312016-08-01 14:21:30 +0000477
478def DS_ADD_SRC2_U32 : DS_1A<"ds_add_src2_u32">;
479def DS_SUB_SRC2_U32 : DS_1A<"ds_sub_src2_u32">;
480def DS_RSUB_SRC2_U32 : DS_1A<"ds_rsub_src2_u32">;
481def DS_INC_SRC2_U32 : DS_1A<"ds_inc_src2_u32">;
482def DS_DEC_SRC2_U32 : DS_1A<"ds_dec_src2_u32">;
483def DS_MIN_SRC2_I32 : DS_1A<"ds_min_src2_i32">;
484def DS_MAX_SRC2_I32 : DS_1A<"ds_max_src2_i32">;
485def DS_MIN_SRC2_U32 : DS_1A<"ds_min_src2_u32">;
486def DS_MAX_SRC2_U32 : DS_1A<"ds_max_src2_u32">;
Dmitry Preobrazhenskye5147242017-04-07 13:07:13 +0000487def DS_AND_SRC2_B32 : DS_1A<"ds_and_src2_b32">;
Valery Pykhtin902db312016-08-01 14:21:30 +0000488def DS_OR_SRC2_B32 : DS_1A<"ds_or_src2_b32">;
489def DS_XOR_SRC2_B32 : DS_1A<"ds_xor_src2_b32">;
490def DS_MIN_SRC2_F32 : DS_1A<"ds_min_src2_f32">;
491def DS_MAX_SRC2_F32 : DS_1A<"ds_max_src2_f32">;
492
493def DS_ADD_SRC2_U64 : DS_1A<"ds_add_src2_u64">;
494def DS_SUB_SRC2_U64 : DS_1A<"ds_sub_src2_u64">;
495def DS_RSUB_SRC2_U64 : DS_1A<"ds_rsub_src2_u64">;
496def DS_INC_SRC2_U64 : DS_1A<"ds_inc_src2_u64">;
497def DS_DEC_SRC2_U64 : DS_1A<"ds_dec_src2_u64">;
498def DS_MIN_SRC2_I64 : DS_1A<"ds_min_src2_i64">;
499def DS_MAX_SRC2_I64 : DS_1A<"ds_max_src2_i64">;
500def DS_MIN_SRC2_U64 : DS_1A<"ds_min_src2_u64">;
501def DS_MAX_SRC2_U64 : DS_1A<"ds_max_src2_u64">;
502def DS_AND_SRC2_B64 : DS_1A<"ds_and_src2_b64">;
503def DS_OR_SRC2_B64 : DS_1A<"ds_or_src2_b64">;
504def DS_XOR_SRC2_B64 : DS_1A<"ds_xor_src2_b64">;
505def DS_MIN_SRC2_F64 : DS_1A<"ds_min_src2_f64">;
506def DS_MAX_SRC2_F64 : DS_1A<"ds_max_src2_f64">;
507
Dmitry Preobrazhenskye6ef0992017-04-14 12:28:07 +0000508def DS_WRITE_SRC2_B32 : DS_1A<"ds_write_src2_b32">;
509def DS_WRITE_SRC2_B64 : DS_1A<"ds_write_src2_b64">;
Valery Pykhtin902db312016-08-01 14:21:30 +0000510
511let Uses = [EXEC], mayLoad = 0, mayStore = 0, isConvergent = 1 in {
Matt Arsenaultb81495d2017-09-20 05:01:53 +0000512def DS_SWIZZLE_B32 : DS_1A_RET <"ds_swizzle_b32", VGPR_32, 0, SwizzleImm>;
Valery Pykhtin902db312016-08-01 14:21:30 +0000513}
514
515let mayStore = 0 in {
Matt Arsenault10c472d2017-11-15 01:34:06 +0000516defm DS_READ_I8 : DS_1A_RET_mc<"ds_read_i8">;
517defm DS_READ_U8 : DS_1A_RET_mc<"ds_read_u8">;
518defm DS_READ_I16 : DS_1A_RET_mc<"ds_read_i16">;
519defm DS_READ_U16 : DS_1A_RET_mc<"ds_read_u16">;
520defm DS_READ_B32 : DS_1A_RET_mc<"ds_read_b32">;
521defm DS_READ_B64 : DS_1A_RET_mc<"ds_read_b64", VReg_64>;
Valery Pykhtin902db312016-08-01 14:21:30 +0000522
Matt Arsenault10c472d2017-11-15 01:34:06 +0000523defm DS_READ2_B32 : DS_1A_Off8_RET_mc<"ds_read2_b32", VReg_64>;
524defm DS_READ2ST64_B32: DS_1A_Off8_RET_mc<"ds_read2st64_b32", VReg_64>;
Valery Pykhtin902db312016-08-01 14:21:30 +0000525
Matt Arsenault10c472d2017-11-15 01:34:06 +0000526defm DS_READ2_B64 : DS_1A_Off8_RET_mc<"ds_read2_b64", VReg_128>;
527defm DS_READ2ST64_B64: DS_1A_Off8_RET_mc<"ds_read2st64_b64", VReg_128>;
Matt Arsenaulted6e8f02017-09-01 18:36:06 +0000528
Matt Arsenault10c472d2017-11-15 01:34:06 +0000529let has_m0_read = 0 in {
Matt Arsenaulted6e8f02017-09-01 18:36:06 +0000530let SubtargetPredicate = HasD16LoadStore in {
Matt Arsenaultb81495d2017-09-20 05:01:53 +0000531def DS_READ_U8_D16 : DS_1A_RET_Tied<"ds_read_u8_d16">;
532def DS_READ_U8_D16_HI : DS_1A_RET_Tied<"ds_read_u8_d16_hi">;
533def DS_READ_I8_D16 : DS_1A_RET_Tied<"ds_read_i8_d16">;
534def DS_READ_I8_D16_HI : DS_1A_RET_Tied<"ds_read_i8_d16_hi">;
535def DS_READ_U16_D16 : DS_1A_RET_Tied<"ds_read_u16_d16">;
536def DS_READ_U16_D16_HI : DS_1A_RET_Tied<"ds_read_u16_d16_hi">;
Matt Arsenaulted6e8f02017-09-01 18:36:06 +0000537}
Matt Arsenaultefa1d652017-09-01 18:38:02 +0000538
539let SubtargetPredicate = HasDSAddTid in {
540def DS_READ_ADDTID_B32 : DS_1A_RET<"ds_read_addtid_b32">;
541}
Matt Arsenault10c472d2017-11-15 01:34:06 +0000542} // End has_m0_read = 0
Valery Pykhtin902db312016-08-01 14:21:30 +0000543}
544
Valery Pykhtin902db312016-08-01 14:21:30 +0000545def DS_CONSUME : DS_0A_RET<"ds_consume">;
546def DS_APPEND : DS_0A_RET<"ds_append">;
547def DS_ORDERED_COUNT : DS_1A_RET_GDS<"ds_ordered_count">;
Valery Pykhtin902db312016-08-01 14:21:30 +0000548
549//===----------------------------------------------------------------------===//
550// Instruction definitions for CI and newer.
551//===----------------------------------------------------------------------===//
Valery Pykhtin902db312016-08-01 14:21:30 +0000552
553let SubtargetPredicate = isCIVI in {
554
Matt Arsenault10c472d2017-11-15 01:34:06 +0000555defm DS_WRAP_RTN_B32 : DS_1A2D_RET_mc<"ds_wrap_rtn_b32", VGPR_32>;
556defm DS_CONDXCHG32_RTN_B64 : DS_1A1D_RET_mc<"ds_condxchg32_rtn_b64", VReg_64>;
Dmitry Preobrazhenskye5147242017-04-07 13:07:13 +0000557
558def DS_GWS_SEMA_RELEASE_ALL : DS_GWS_0D<"ds_gws_sema_release_all">;
Valery Pykhtin902db312016-08-01 14:21:30 +0000559
Matt Arsenaultdedc5442017-02-28 20:15:43 +0000560let mayStore = 0 in {
Matt Arsenault10c472d2017-11-15 01:34:06 +0000561defm DS_READ_B96 : DS_1A_RET_mc<"ds_read_b96", VReg_96>;
562defm DS_READ_B128: DS_1A_RET_mc<"ds_read_b128", VReg_128>;
Matt Arsenaultdedc5442017-02-28 20:15:43 +0000563} // End mayStore = 0
564
565let mayLoad = 0 in {
Matt Arsenault10c472d2017-11-15 01:34:06 +0000566defm DS_WRITE_B96 : DS_1A1D_NORET_mc<"ds_write_b96", VReg_96>;
567defm DS_WRITE_B128 : DS_1A1D_NORET_mc<"ds_write_b128", VReg_128>;
Matt Arsenaultdedc5442017-02-28 20:15:43 +0000568} // End mayLoad = 0
569
Matt Arsenault78124982017-02-28 20:15:46 +0000570def DS_NOP : DS_VOID<"ds_nop">;
Matt Arsenaultdedc5442017-02-28 20:15:43 +0000571
Valery Pykhtin902db312016-08-01 14:21:30 +0000572} // let SubtargetPredicate = isCIVI
573
574//===----------------------------------------------------------------------===//
575// Instruction definitions for VI and newer.
576//===----------------------------------------------------------------------===//
577
578let SubtargetPredicate = isVI in {
579
580let Uses = [EXEC] in {
581def DS_PERMUTE_B32 : DS_1A1D_PERMUTE <"ds_permute_b32",
582 int_amdgcn_ds_permute>;
583def DS_BPERMUTE_B32 : DS_1A1D_PERMUTE <"ds_bpermute_b32",
584 int_amdgcn_ds_bpermute>;
585}
586
Dmitry Preobrazhensky622bde82018-03-28 16:21:56 +0000587def DS_ADD_SRC2_F32 : DS_1A<"ds_add_src2_f32">;
588
Valery Pykhtin902db312016-08-01 14:21:30 +0000589} // let SubtargetPredicate = isVI
590
591//===----------------------------------------------------------------------===//
592// DS Patterns
593//===----------------------------------------------------------------------===//
594
Matt Arsenault90c75932017-10-03 00:06:41 +0000595def : GCNPat <
Valery Pykhtin902db312016-08-01 14:21:30 +0000596 (int_amdgcn_ds_swizzle i32:$src, imm:$offset16),
597 (DS_SWIZZLE_B32 $src, (as_i16imm $offset16), (i1 0))
598>;
599
Matt Arsenault90c75932017-10-03 00:06:41 +0000600class DSReadPat <DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat <
Valery Pykhtin902db312016-08-01 14:21:30 +0000601 (vt (frag (DS1Addr1Offset i32:$ptr, i32:$offset))),
602 (inst $ptr, (as_i16imm $offset), (i1 0))
603>;
604
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000605multiclass DSReadPat_mc<DS_Pseudo inst, ValueType vt, string frag> {
606
607 let OtherPredicates = [LDSRequiresM0Init] in {
608 def : DSReadPat<inst, vt, !cast<PatFrag>(frag#"_m0")>;
609 }
610
611 let OtherPredicates = [NotLDSRequiresM0Init] in {
Nicolai Haehnle40b140f2018-02-22 15:25:11 +0000612 def : DSReadPat<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt, !cast<PatFrag>(frag)>;
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000613 }
614}
615
616
Matt Arsenaultb81495d2017-09-20 05:01:53 +0000617multiclass DSReadPat_Hi16 <DS_Pseudo inst, PatFrag frag, ValueType vt = i16> {
Matt Arsenault90c75932017-10-03 00:06:41 +0000618 def : GCNPat <
Matt Arsenaultb81495d2017-09-20 05:01:53 +0000619 (build_vector vt:$lo, (vt (frag (DS1Addr1Offset i32:$ptr, i32:$offset)))),
620 (v2i16 (inst $ptr, (as_i16imm $offset), (i1 0), $lo))
621 >;
622
Matt Arsenault90c75932017-10-03 00:06:41 +0000623 def : GCNPat <
Matt Arsenaultb81495d2017-09-20 05:01:53 +0000624 (build_vector f16:$lo, (f16 (bitconvert (vt (frag (DS1Addr1Offset i32:$ptr, i32:$offset)))))),
625 (v2f16 (inst $ptr, (as_i16imm $offset), (i1 0), $lo))
626 >;
627}
628
Matt Arsenaulte1cd4822017-11-13 00:22:09 +0000629multiclass DSReadPat_Lo16 <DS_Pseudo inst, PatFrag frag, ValueType vt = i16> {
630 def : GCNPat <
631 (build_vector (vt (frag (DS1Addr1Offset i32:$ptr, i32:$offset))), (vt (Hi16Elt vt:$hi))),
632 (v2i16 (inst $ptr, (as_i16imm $offset), 0, $hi))
633 >;
634
635 def : GCNPat <
636 (build_vector (f16 (bitconvert (vt (frag (DS1Addr1Offset i32:$ptr, i32:$offset))))), (f16 (Hi16Elt f16:$hi))),
637 (v2f16 (inst $ptr, (as_i16imm $offset), 0, $hi))
638 >;
639}
640
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000641defm : DSReadPat_mc <DS_READ_I8, i32, "sextloadi8_local">;
642defm : DSReadPat_mc <DS_READ_U8, i32, "az_extloadi8_local">;
643defm : DSReadPat_mc <DS_READ_I8, i16, "sextloadi8_local">;
644defm : DSReadPat_mc <DS_READ_U8, i16, "az_extloadi8_local">;
645defm : DSReadPat_mc <DS_READ_I16, i32, "sextloadi16_local">;
646defm : DSReadPat_mc <DS_READ_I16, i32, "sextloadi16_local">;
647defm : DSReadPat_mc <DS_READ_U16, i32, "az_extloadi16_local">;
648defm : DSReadPat_mc <DS_READ_U16, i16, "load_local">;
649defm : DSReadPat_mc <DS_READ_B32, i32, "load_local">;
Valery Pykhtin902db312016-08-01 14:21:30 +0000650
651let AddedComplexity = 100 in {
652
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000653defm : DSReadPat_mc <DS_READ_B64, v2i32, "load_align8_local">;
Farhana Aleena7cb3112018-03-09 17:41:39 +0000654defm : DSReadPat_mc <DS_READ_B128, v4i32, "load_align16_local">;
Valery Pykhtin902db312016-08-01 14:21:30 +0000655
656} // End AddedComplexity = 100
657
Konstantin Zhuravlyovc2c2eb72018-05-04 20:06:57 +0000658let OtherPredicates = [D16PreservesUnusedBits] in {
Matt Arsenaultb81495d2017-09-20 05:01:53 +0000659let AddedComplexity = 100 in {
660defm : DSReadPat_Hi16<DS_READ_U16_D16_HI, load_local>;
661defm : DSReadPat_Hi16<DS_READ_U8_D16_HI, az_extloadi8_local>;
662defm : DSReadPat_Hi16<DS_READ_I8_D16_HI, sextloadi8_local>;
Matt Arsenaulte1cd4822017-11-13 00:22:09 +0000663
664defm : DSReadPat_Lo16<DS_READ_U16_D16, load_local>;
665defm : DSReadPat_Lo16<DS_READ_U8_D16, az_extloadi8_local>;
666defm : DSReadPat_Lo16<DS_READ_I8_D16, sextloadi8_local>;
667
Matt Arsenaultb81495d2017-09-20 05:01:53 +0000668}
669}
670
Matt Arsenault90c75932017-10-03 00:06:41 +0000671class DSWritePat <DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat <
Valery Pykhtin902db312016-08-01 14:21:30 +0000672 (frag vt:$value, (DS1Addr1Offset i32:$ptr, i32:$offset)),
673 (inst $ptr, $value, (as_i16imm $offset), (i1 0))
674>;
675
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000676multiclass DSWritePat_mc <DS_Pseudo inst, ValueType vt, string frag> {
677 let OtherPredicates = [LDSRequiresM0Init] in {
678 def : DSWritePat<inst, vt, !cast<PatFrag>(frag#"_m0")>;
679 }
680
681 let OtherPredicates = [NotLDSRequiresM0Init] in {
Nicolai Haehnle40b140f2018-02-22 15:25:11 +0000682 def : DSWritePat<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt, !cast<PatFrag>(frag)>;
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000683 }
684}
685
686defm : DSWritePat_mc <DS_WRITE_B8, i32, "truncstorei8_local">;
687defm : DSWritePat_mc <DS_WRITE_B16, i32, "truncstorei16_local">;
688defm : DSWritePat_mc <DS_WRITE_B8, i16, "truncstorei8_local">;
689defm : DSWritePat_mc <DS_WRITE_B16, i16, "store_local">;
690defm : DSWritePat_mc <DS_WRITE_B32, i32, "store_local">;
Valery Pykhtin902db312016-08-01 14:21:30 +0000691
Konstantin Zhuravlyovc2c2eb72018-05-04 20:06:57 +0000692let OtherPredicates = [D16PreservesUnusedBits] in {
Matt Arsenaultfcc213f2017-09-20 03:20:09 +0000693def : DSWritePat <DS_WRITE_B16_D16_HI, i32, store_local_hi16>;
694def : DSWritePat <DS_WRITE_B8_D16_HI, i32, truncstorei8_local_hi16>;
695}
696
Valery Pykhtin902db312016-08-01 14:21:30 +0000697
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000698class DS64Bit4ByteAlignedReadPat<DS_Pseudo inst, PatFrag frag> : GCNPat <
699 (v2i32 (frag (DS64Bit4ByteAligned i32:$ptr, i8:$offset0, i8:$offset1))),
700 (inst $ptr, $offset0, $offset1, (i1 0))
Valery Pykhtin902db312016-08-01 14:21:30 +0000701>;
702
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000703class DS64Bit4ByteAlignedWritePat<DS_Pseudo inst, PatFrag frag> : GCNPat<
704 (frag v2i32:$value, (DS64Bit4ByteAligned i32:$ptr, i8:$offset0, i8:$offset1)),
705 (inst $ptr, (i32 (EXTRACT_SUBREG $value, sub0)),
706 (i32 (EXTRACT_SUBREG $value, sub1)), $offset0, $offset1,
707 (i1 0))
708>;
709
710let OtherPredicates = [LDSRequiresM0Init] in {
711def : DS64Bit4ByteAlignedReadPat<DS_READ2_B32, load_local_m0>;
712def : DS64Bit4ByteAlignedWritePat<DS_WRITE2_B32, store_local_m0>;
713}
714
715let OtherPredicates = [NotLDSRequiresM0Init] in {
716def : DS64Bit4ByteAlignedReadPat<DS_READ2_B32_gfx9, load_local>;
717def : DS64Bit4ByteAlignedWritePat<DS_WRITE2_B32_gfx9, store_local>;
718}
719
720
721let AddedComplexity = 100 in {
722
723defm : DSWritePat_mc <DS_WRITE_B64, v2i32, "store_align8_local">;
Farhana Aleenc6c9dc82018-03-16 18:12:00 +0000724defm : DSWritePat_mc <DS_WRITE_B128, v4i32, "store_align16_local">;
725
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000726} // End AddedComplexity = 100
Matt Arsenault90c75932017-10-03 00:06:41 +0000727class DSAtomicRetPat<DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat <
Valery Pykhtin902db312016-08-01 14:21:30 +0000728 (frag (DS1Addr1Offset i32:$ptr, i32:$offset), vt:$value),
729 (inst $ptr, $value, (as_i16imm $offset), (i1 0))
730>;
731
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000732multiclass DSAtomicRetPat_mc<DS_Pseudo inst, ValueType vt, string frag> {
733 let OtherPredicates = [LDSRequiresM0Init] in {
734 def : DSAtomicRetPat<inst, vt, !cast<PatFrag>(frag#"_m0")>;
735 }
736
737 let OtherPredicates = [NotLDSRequiresM0Init] in {
Nicolai Haehnle40b140f2018-02-22 15:25:11 +0000738 def : DSAtomicRetPat<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt,
739 !cast<PatFrag>(frag)>;
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000740 }
741}
742
743
744
Matt Arsenault90c75932017-10-03 00:06:41 +0000745class DSAtomicCmpXChg<DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat <
Valery Pykhtin902db312016-08-01 14:21:30 +0000746 (frag (DS1Addr1Offset i32:$ptr, i32:$offset), vt:$cmp, vt:$swap),
747 (inst $ptr, $cmp, $swap, (as_i16imm $offset), (i1 0))
748>;
749
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000750multiclass DSAtomicCmpXChg_mc<DS_Pseudo inst, ValueType vt, string frag> {
751 let OtherPredicates = [LDSRequiresM0Init] in {
752 def : DSAtomicCmpXChg<inst, vt, !cast<PatFrag>(frag#"_m0")>;
753 }
754
755 let OtherPredicates = [NotLDSRequiresM0Init] in {
Nicolai Haehnle40b140f2018-02-22 15:25:11 +0000756 def : DSAtomicCmpXChg<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt,
757 !cast<PatFrag>(frag)>;
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000758 }
759}
760
761
Valery Pykhtin902db312016-08-01 14:21:30 +0000762
763// 32-bit atomics.
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000764defm : DSAtomicRetPat_mc<DS_WRXCHG_RTN_B32, i32, "atomic_swap_local">;
765defm : DSAtomicRetPat_mc<DS_ADD_RTN_U32, i32, "atomic_load_add_local">;
766defm : DSAtomicRetPat_mc<DS_SUB_RTN_U32, i32, "atomic_load_sub_local">;
767defm : DSAtomicRetPat_mc<DS_INC_RTN_U32, i32, "atomic_inc_local">;
768defm : DSAtomicRetPat_mc<DS_DEC_RTN_U32, i32, "atomic_dec_local">;
769defm : DSAtomicRetPat_mc<DS_AND_RTN_B32, i32, "atomic_load_and_local">;
770defm : DSAtomicRetPat_mc<DS_OR_RTN_B32, i32, "atomic_load_or_local">;
771defm : DSAtomicRetPat_mc<DS_XOR_RTN_B32, i32, "atomic_load_xor_local">;
772defm : DSAtomicRetPat_mc<DS_MIN_RTN_I32, i32, "atomic_load_min_local">;
773defm : DSAtomicRetPat_mc<DS_MAX_RTN_I32, i32, "atomic_load_max_local">;
774defm : DSAtomicRetPat_mc<DS_MIN_RTN_U32, i32, "atomic_load_umin_local">;
775defm : DSAtomicRetPat_mc<DS_MAX_RTN_U32, i32, "atomic_load_umax_local">;
776defm : DSAtomicCmpXChg_mc<DS_CMPST_RTN_B32, i32, "atomic_cmp_swap_local">;
Daniil Fukalovd5fca552018-01-17 14:05:05 +0000777defm : DSAtomicRetPat_mc<DS_MIN_RTN_F32, f32, "atomic_load_fmin_local">;
778defm : DSAtomicRetPat_mc<DS_MAX_RTN_F32, f32, "atomic_load_fmax_local">;
779defm : DSAtomicRetPat_mc<DS_ADD_RTN_F32, f32, "atomic_load_fadd_local">;
Valery Pykhtin902db312016-08-01 14:21:30 +0000780
781// 64-bit atomics.
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000782defm : DSAtomicRetPat_mc<DS_WRXCHG_RTN_B64, i64, "atomic_swap_local">;
783defm : DSAtomicRetPat_mc<DS_ADD_RTN_U64, i64, "atomic_load_add_local">;
784defm : DSAtomicRetPat_mc<DS_SUB_RTN_U64, i64, "atomic_load_sub_local">;
785defm : DSAtomicRetPat_mc<DS_INC_RTN_U64, i64, "atomic_inc_local">;
786defm : DSAtomicRetPat_mc<DS_DEC_RTN_U64, i64, "atomic_dec_local">;
787defm : DSAtomicRetPat_mc<DS_AND_RTN_B64, i64, "atomic_load_and_local">;
788defm : DSAtomicRetPat_mc<DS_OR_RTN_B64, i64, "atomic_load_or_local">;
789defm : DSAtomicRetPat_mc<DS_XOR_RTN_B64, i64, "atomic_load_xor_local">;
790defm : DSAtomicRetPat_mc<DS_MIN_RTN_I64, i64, "atomic_load_min_local">;
791defm : DSAtomicRetPat_mc<DS_MAX_RTN_I64, i64, "atomic_load_max_local">;
792defm : DSAtomicRetPat_mc<DS_MIN_RTN_U64, i64, "atomic_load_umin_local">;
793defm : DSAtomicRetPat_mc<DS_MAX_RTN_U64, i64, "atomic_load_umax_local">;
Valery Pykhtin902db312016-08-01 14:21:30 +0000794
Matt Arsenault3f71c0e2017-11-29 00:55:57 +0000795defm : DSAtomicCmpXChg_mc<DS_CMPST_RTN_B64, i64, "atomic_cmp_swap_local">;
Valery Pykhtin902db312016-08-01 14:21:30 +0000796
Valery Pykhtin902db312016-08-01 14:21:30 +0000797//===----------------------------------------------------------------------===//
798// Real instructions
799//===----------------------------------------------------------------------===//
800
801//===----------------------------------------------------------------------===//
802// SIInstructions.td
803//===----------------------------------------------------------------------===//
804
805class DS_Real_si <bits<8> op, DS_Pseudo ds> :
806 DS_Real <ds>,
807 SIMCInstr <ds.Mnemonic, SIEncodingFamily.SI> {
808 let AssemblerPredicates=[isSICI];
809 let DecoderNamespace="SICI";
810
811 // encoding
812 let Inst{7-0} = !if(ds.has_offset0, offset0, 0);
813 let Inst{15-8} = !if(ds.has_offset1, offset1, 0);
814 let Inst{17} = !if(ds.has_gds, gds, ds.gdsValue);
815 let Inst{25-18} = op;
816 let Inst{31-26} = 0x36; // ds prefix
817 let Inst{39-32} = !if(ds.has_addr, addr, 0);
818 let Inst{47-40} = !if(ds.has_data0, data0, 0);
819 let Inst{55-48} = !if(ds.has_data1, data1, 0);
820 let Inst{63-56} = !if(ds.has_vdst, vdst, 0);
821}
822
823def DS_ADD_U32_si : DS_Real_si<0x0, DS_ADD_U32>;
824def DS_SUB_U32_si : DS_Real_si<0x1, DS_SUB_U32>;
825def DS_RSUB_U32_si : DS_Real_si<0x2, DS_RSUB_U32>;
826def DS_INC_U32_si : DS_Real_si<0x3, DS_INC_U32>;
827def DS_DEC_U32_si : DS_Real_si<0x4, DS_DEC_U32>;
828def DS_MIN_I32_si : DS_Real_si<0x5, DS_MIN_I32>;
829def DS_MAX_I32_si : DS_Real_si<0x6, DS_MAX_I32>;
830def DS_MIN_U32_si : DS_Real_si<0x7, DS_MIN_U32>;
831def DS_MAX_U32_si : DS_Real_si<0x8, DS_MAX_U32>;
832def DS_AND_B32_si : DS_Real_si<0x9, DS_AND_B32>;
833def DS_OR_B32_si : DS_Real_si<0xa, DS_OR_B32>;
834def DS_XOR_B32_si : DS_Real_si<0xb, DS_XOR_B32>;
835def DS_MSKOR_B32_si : DS_Real_si<0xc, DS_MSKOR_B32>;
836def DS_WRITE_B32_si : DS_Real_si<0xd, DS_WRITE_B32>;
837def DS_WRITE2_B32_si : DS_Real_si<0xe, DS_WRITE2_B32>;
838def DS_WRITE2ST64_B32_si : DS_Real_si<0xf, DS_WRITE2ST64_B32>;
839def DS_CMPST_B32_si : DS_Real_si<0x10, DS_CMPST_B32>;
840def DS_CMPST_F32_si : DS_Real_si<0x11, DS_CMPST_F32>;
841def DS_MIN_F32_si : DS_Real_si<0x12, DS_MIN_F32>;
842def DS_MAX_F32_si : DS_Real_si<0x13, DS_MAX_F32>;
Matt Arsenault78124982017-02-28 20:15:46 +0000843def DS_NOP_si : DS_Real_si<0x14, DS_NOP>;
Valery Pykhtin902db312016-08-01 14:21:30 +0000844def DS_GWS_INIT_si : DS_Real_si<0x19, DS_GWS_INIT>;
845def DS_GWS_SEMA_V_si : DS_Real_si<0x1a, DS_GWS_SEMA_V>;
846def DS_GWS_SEMA_BR_si : DS_Real_si<0x1b, DS_GWS_SEMA_BR>;
847def DS_GWS_SEMA_P_si : DS_Real_si<0x1c, DS_GWS_SEMA_P>;
848def DS_GWS_BARRIER_si : DS_Real_si<0x1d, DS_GWS_BARRIER>;
849def DS_WRITE_B8_si : DS_Real_si<0x1e, DS_WRITE_B8>;
850def DS_WRITE_B16_si : DS_Real_si<0x1f, DS_WRITE_B16>;
851def DS_ADD_RTN_U32_si : DS_Real_si<0x20, DS_ADD_RTN_U32>;
852def DS_SUB_RTN_U32_si : DS_Real_si<0x21, DS_SUB_RTN_U32>;
853def DS_RSUB_RTN_U32_si : DS_Real_si<0x22, DS_RSUB_RTN_U32>;
854def DS_INC_RTN_U32_si : DS_Real_si<0x23, DS_INC_RTN_U32>;
855def DS_DEC_RTN_U32_si : DS_Real_si<0x24, DS_DEC_RTN_U32>;
856def DS_MIN_RTN_I32_si : DS_Real_si<0x25, DS_MIN_RTN_I32>;
857def DS_MAX_RTN_I32_si : DS_Real_si<0x26, DS_MAX_RTN_I32>;
858def DS_MIN_RTN_U32_si : DS_Real_si<0x27, DS_MIN_RTN_U32>;
859def DS_MAX_RTN_U32_si : DS_Real_si<0x28, DS_MAX_RTN_U32>;
860def DS_AND_RTN_B32_si : DS_Real_si<0x29, DS_AND_RTN_B32>;
861def DS_OR_RTN_B32_si : DS_Real_si<0x2a, DS_OR_RTN_B32>;
862def DS_XOR_RTN_B32_si : DS_Real_si<0x2b, DS_XOR_RTN_B32>;
863def DS_MSKOR_RTN_B32_si : DS_Real_si<0x2c, DS_MSKOR_RTN_B32>;
864def DS_WRXCHG_RTN_B32_si : DS_Real_si<0x2d, DS_WRXCHG_RTN_B32>;
865def DS_WRXCHG2_RTN_B32_si : DS_Real_si<0x2e, DS_WRXCHG2_RTN_B32>;
866def DS_WRXCHG2ST64_RTN_B32_si : DS_Real_si<0x2f, DS_WRXCHG2ST64_RTN_B32>;
867def DS_CMPST_RTN_B32_si : DS_Real_si<0x30, DS_CMPST_RTN_B32>;
868def DS_CMPST_RTN_F32_si : DS_Real_si<0x31, DS_CMPST_RTN_F32>;
869def DS_MIN_RTN_F32_si : DS_Real_si<0x32, DS_MIN_RTN_F32>;
870def DS_MAX_RTN_F32_si : DS_Real_si<0x33, DS_MAX_RTN_F32>;
871
Dmitry Preobrazhenskye5147242017-04-07 13:07:13 +0000872// These instruction are CI/VI only
873def DS_WRAP_RTN_B32_si : DS_Real_si<0x34, DS_WRAP_RTN_B32>;
874def DS_CONDXCHG32_RTN_B64_si : DS_Real_si<0x7e, DS_CONDXCHG32_RTN_B64>;
875def DS_GWS_SEMA_RELEASE_ALL_si : DS_Real_si<0x18, DS_GWS_SEMA_RELEASE_ALL>;
Valery Pykhtin902db312016-08-01 14:21:30 +0000876
877def DS_SWIZZLE_B32_si : DS_Real_si<0x35, DS_SWIZZLE_B32>;
878def DS_READ_B32_si : DS_Real_si<0x36, DS_READ_B32>;
879def DS_READ2_B32_si : DS_Real_si<0x37, DS_READ2_B32>;
880def DS_READ2ST64_B32_si : DS_Real_si<0x38, DS_READ2ST64_B32>;
881def DS_READ_I8_si : DS_Real_si<0x39, DS_READ_I8>;
882def DS_READ_U8_si : DS_Real_si<0x3a, DS_READ_U8>;
883def DS_READ_I16_si : DS_Real_si<0x3b, DS_READ_I16>;
884def DS_READ_U16_si : DS_Real_si<0x3c, DS_READ_U16>;
885def DS_CONSUME_si : DS_Real_si<0x3d, DS_CONSUME>;
886def DS_APPEND_si : DS_Real_si<0x3e, DS_APPEND>;
887def DS_ORDERED_COUNT_si : DS_Real_si<0x3f, DS_ORDERED_COUNT>;
888def DS_ADD_U64_si : DS_Real_si<0x40, DS_ADD_U64>;
889def DS_SUB_U64_si : DS_Real_si<0x41, DS_SUB_U64>;
890def DS_RSUB_U64_si : DS_Real_si<0x42, DS_RSUB_U64>;
891def DS_INC_U64_si : DS_Real_si<0x43, DS_INC_U64>;
892def DS_DEC_U64_si : DS_Real_si<0x44, DS_DEC_U64>;
893def DS_MIN_I64_si : DS_Real_si<0x45, DS_MIN_I64>;
894def DS_MAX_I64_si : DS_Real_si<0x46, DS_MAX_I64>;
895def DS_MIN_U64_si : DS_Real_si<0x47, DS_MIN_U64>;
896def DS_MAX_U64_si : DS_Real_si<0x48, DS_MAX_U64>;
897def DS_AND_B64_si : DS_Real_si<0x49, DS_AND_B64>;
898def DS_OR_B64_si : DS_Real_si<0x4a, DS_OR_B64>;
899def DS_XOR_B64_si : DS_Real_si<0x4b, DS_XOR_B64>;
900def DS_MSKOR_B64_si : DS_Real_si<0x4c, DS_MSKOR_B64>;
901def DS_WRITE_B64_si : DS_Real_si<0x4d, DS_WRITE_B64>;
902def DS_WRITE2_B64_si : DS_Real_si<0x4E, DS_WRITE2_B64>;
903def DS_WRITE2ST64_B64_si : DS_Real_si<0x4f, DS_WRITE2ST64_B64>;
904def DS_CMPST_B64_si : DS_Real_si<0x50, DS_CMPST_B64>;
905def DS_CMPST_F64_si : DS_Real_si<0x51, DS_CMPST_F64>;
906def DS_MIN_F64_si : DS_Real_si<0x52, DS_MIN_F64>;
907def DS_MAX_F64_si : DS_Real_si<0x53, DS_MAX_F64>;
908
909def DS_ADD_RTN_U64_si : DS_Real_si<0x60, DS_ADD_RTN_U64>;
910def DS_SUB_RTN_U64_si : DS_Real_si<0x61, DS_SUB_RTN_U64>;
911def DS_RSUB_RTN_U64_si : DS_Real_si<0x62, DS_RSUB_RTN_U64>;
912def DS_INC_RTN_U64_si : DS_Real_si<0x63, DS_INC_RTN_U64>;
913def DS_DEC_RTN_U64_si : DS_Real_si<0x64, DS_DEC_RTN_U64>;
914def DS_MIN_RTN_I64_si : DS_Real_si<0x65, DS_MIN_RTN_I64>;
915def DS_MAX_RTN_I64_si : DS_Real_si<0x66, DS_MAX_RTN_I64>;
916def DS_MIN_RTN_U64_si : DS_Real_si<0x67, DS_MIN_RTN_U64>;
917def DS_MAX_RTN_U64_si : DS_Real_si<0x68, DS_MAX_RTN_U64>;
918def DS_AND_RTN_B64_si : DS_Real_si<0x69, DS_AND_RTN_B64>;
919def DS_OR_RTN_B64_si : DS_Real_si<0x6a, DS_OR_RTN_B64>;
920def DS_XOR_RTN_B64_si : DS_Real_si<0x6b, DS_XOR_RTN_B64>;
921def DS_MSKOR_RTN_B64_si : DS_Real_si<0x6c, DS_MSKOR_RTN_B64>;
922def DS_WRXCHG_RTN_B64_si : DS_Real_si<0x6d, DS_WRXCHG_RTN_B64>;
923def DS_WRXCHG2_RTN_B64_si : DS_Real_si<0x6e, DS_WRXCHG2_RTN_B64>;
924def DS_WRXCHG2ST64_RTN_B64_si : DS_Real_si<0x6f, DS_WRXCHG2ST64_RTN_B64>;
925def DS_CMPST_RTN_B64_si : DS_Real_si<0x70, DS_CMPST_RTN_B64>;
926def DS_CMPST_RTN_F64_si : DS_Real_si<0x71, DS_CMPST_RTN_F64>;
927def DS_MIN_RTN_F64_si : DS_Real_si<0x72, DS_MIN_RTN_F64>;
928def DS_MAX_RTN_F64_si : DS_Real_si<0x73, DS_MAX_RTN_F64>;
929
930def DS_READ_B64_si : DS_Real_si<0x76, DS_READ_B64>;
931def DS_READ2_B64_si : DS_Real_si<0x77, DS_READ2_B64>;
932def DS_READ2ST64_B64_si : DS_Real_si<0x78, DS_READ2ST64_B64>;
933
934def DS_ADD_SRC2_U32_si : DS_Real_si<0x80, DS_ADD_SRC2_U32>;
935def DS_SUB_SRC2_U32_si : DS_Real_si<0x81, DS_SUB_SRC2_U32>;
936def DS_RSUB_SRC2_U32_si : DS_Real_si<0x82, DS_RSUB_SRC2_U32>;
937def DS_INC_SRC2_U32_si : DS_Real_si<0x83, DS_INC_SRC2_U32>;
938def DS_DEC_SRC2_U32_si : DS_Real_si<0x84, DS_DEC_SRC2_U32>;
939def DS_MIN_SRC2_I32_si : DS_Real_si<0x85, DS_MIN_SRC2_I32>;
940def DS_MAX_SRC2_I32_si : DS_Real_si<0x86, DS_MAX_SRC2_I32>;
941def DS_MIN_SRC2_U32_si : DS_Real_si<0x87, DS_MIN_SRC2_U32>;
942def DS_MAX_SRC2_U32_si : DS_Real_si<0x88, DS_MAX_SRC2_U32>;
943def DS_AND_SRC2_B32_si : DS_Real_si<0x89, DS_AND_SRC2_B32>;
944def DS_OR_SRC2_B32_si : DS_Real_si<0x8a, DS_OR_SRC2_B32>;
945def DS_XOR_SRC2_B32_si : DS_Real_si<0x8b, DS_XOR_SRC2_B32>;
946def DS_WRITE_SRC2_B32_si : DS_Real_si<0x8d, DS_WRITE_SRC2_B32>;
947
948def DS_MIN_SRC2_F32_si : DS_Real_si<0x92, DS_MIN_SRC2_F32>;
949def DS_MAX_SRC2_F32_si : DS_Real_si<0x93, DS_MAX_SRC2_F32>;
950
951def DS_ADD_SRC2_U64_si : DS_Real_si<0xc0, DS_ADD_SRC2_U64>;
952def DS_SUB_SRC2_U64_si : DS_Real_si<0xc1, DS_SUB_SRC2_U64>;
953def DS_RSUB_SRC2_U64_si : DS_Real_si<0xc2, DS_RSUB_SRC2_U64>;
954def DS_INC_SRC2_U64_si : DS_Real_si<0xc3, DS_INC_SRC2_U64>;
955def DS_DEC_SRC2_U64_si : DS_Real_si<0xc4, DS_DEC_SRC2_U64>;
956def DS_MIN_SRC2_I64_si : DS_Real_si<0xc5, DS_MIN_SRC2_I64>;
957def DS_MAX_SRC2_I64_si : DS_Real_si<0xc6, DS_MAX_SRC2_I64>;
958def DS_MIN_SRC2_U64_si : DS_Real_si<0xc7, DS_MIN_SRC2_U64>;
959def DS_MAX_SRC2_U64_si : DS_Real_si<0xc8, DS_MAX_SRC2_U64>;
960def DS_AND_SRC2_B64_si : DS_Real_si<0xc9, DS_AND_SRC2_B64>;
961def DS_OR_SRC2_B64_si : DS_Real_si<0xca, DS_OR_SRC2_B64>;
962def DS_XOR_SRC2_B64_si : DS_Real_si<0xcb, DS_XOR_SRC2_B64>;
963def DS_WRITE_SRC2_B64_si : DS_Real_si<0xcd, DS_WRITE_SRC2_B64>;
964
965def DS_MIN_SRC2_F64_si : DS_Real_si<0xd2, DS_MIN_SRC2_F64>;
966def DS_MAX_SRC2_F64_si : DS_Real_si<0xd3, DS_MAX_SRC2_F64>;
Matt Arsenaultdedc5442017-02-28 20:15:43 +0000967def DS_WRITE_B96_si : DS_Real_si<0xde, DS_WRITE_B96>;
968def DS_WRITE_B128_si : DS_Real_si<0xdf, DS_WRITE_B128>;
969def DS_READ_B96_si : DS_Real_si<0xfe, DS_READ_B96>;
970def DS_READ_B128_si : DS_Real_si<0xff, DS_READ_B128>;
Valery Pykhtin902db312016-08-01 14:21:30 +0000971
972//===----------------------------------------------------------------------===//
973// VIInstructions.td
974//===----------------------------------------------------------------------===//
975
976class DS_Real_vi <bits<8> op, DS_Pseudo ds> :
977 DS_Real <ds>,
978 SIMCInstr <ds.Mnemonic, SIEncodingFamily.VI> {
979 let AssemblerPredicates = [isVI];
980 let DecoderNamespace="VI";
981
Matt Arsenaultf3dd8632016-11-01 00:55:14 +0000982 // encoding
Valery Pykhtin902db312016-08-01 14:21:30 +0000983 let Inst{7-0} = !if(ds.has_offset0, offset0, 0);
984 let Inst{15-8} = !if(ds.has_offset1, offset1, 0);
985 let Inst{16} = !if(ds.has_gds, gds, ds.gdsValue);
986 let Inst{24-17} = op;
987 let Inst{31-26} = 0x36; // ds prefix
988 let Inst{39-32} = !if(ds.has_addr, addr, 0);
989 let Inst{47-40} = !if(ds.has_data0, data0, 0);
990 let Inst{55-48} = !if(ds.has_data1, data1, 0);
991 let Inst{63-56} = !if(ds.has_vdst, vdst, 0);
992}
993
994def DS_ADD_U32_vi : DS_Real_vi<0x0, DS_ADD_U32>;
995def DS_SUB_U32_vi : DS_Real_vi<0x1, DS_SUB_U32>;
996def DS_RSUB_U32_vi : DS_Real_vi<0x2, DS_RSUB_U32>;
997def DS_INC_U32_vi : DS_Real_vi<0x3, DS_INC_U32>;
998def DS_DEC_U32_vi : DS_Real_vi<0x4, DS_DEC_U32>;
999def DS_MIN_I32_vi : DS_Real_vi<0x5, DS_MIN_I32>;
1000def DS_MAX_I32_vi : DS_Real_vi<0x6, DS_MAX_I32>;
1001def DS_MIN_U32_vi : DS_Real_vi<0x7, DS_MIN_U32>;
1002def DS_MAX_U32_vi : DS_Real_vi<0x8, DS_MAX_U32>;
1003def DS_AND_B32_vi : DS_Real_vi<0x9, DS_AND_B32>;
1004def DS_OR_B32_vi : DS_Real_vi<0xa, DS_OR_B32>;
1005def DS_XOR_B32_vi : DS_Real_vi<0xb, DS_XOR_B32>;
1006def DS_MSKOR_B32_vi : DS_Real_vi<0xc, DS_MSKOR_B32>;
1007def DS_WRITE_B32_vi : DS_Real_vi<0xd, DS_WRITE_B32>;
1008def DS_WRITE2_B32_vi : DS_Real_vi<0xe, DS_WRITE2_B32>;
1009def DS_WRITE2ST64_B32_vi : DS_Real_vi<0xf, DS_WRITE2ST64_B32>;
1010def DS_CMPST_B32_vi : DS_Real_vi<0x10, DS_CMPST_B32>;
1011def DS_CMPST_F32_vi : DS_Real_vi<0x11, DS_CMPST_F32>;
1012def DS_MIN_F32_vi : DS_Real_vi<0x12, DS_MIN_F32>;
1013def DS_MAX_F32_vi : DS_Real_vi<0x13, DS_MAX_F32>;
Matt Arsenault78124982017-02-28 20:15:46 +00001014def DS_NOP_vi : DS_Real_vi<0x14, DS_NOP>;
Artem Tamazov2e217b82016-09-21 16:35:44 +00001015def DS_ADD_F32_vi : DS_Real_vi<0x15, DS_ADD_F32>;
Dmitry Preobrazhenskye5147242017-04-07 13:07:13 +00001016def DS_GWS_INIT_vi : DS_Real_vi<0x99, DS_GWS_INIT>;
1017def DS_GWS_SEMA_V_vi : DS_Real_vi<0x9a, DS_GWS_SEMA_V>;
1018def DS_GWS_SEMA_BR_vi : DS_Real_vi<0x9b, DS_GWS_SEMA_BR>;
1019def DS_GWS_SEMA_P_vi : DS_Real_vi<0x9c, DS_GWS_SEMA_P>;
1020def DS_GWS_BARRIER_vi : DS_Real_vi<0x9d, DS_GWS_BARRIER>;
Matt Arsenaultefa1d652017-09-01 18:38:02 +00001021def DS_WRITE_ADDTID_B32_vi : DS_Real_vi<0x1d, DS_WRITE_ADDTID_B32>;
Valery Pykhtin902db312016-08-01 14:21:30 +00001022def DS_WRITE_B8_vi : DS_Real_vi<0x1e, DS_WRITE_B8>;
1023def DS_WRITE_B16_vi : DS_Real_vi<0x1f, DS_WRITE_B16>;
1024def DS_ADD_RTN_U32_vi : DS_Real_vi<0x20, DS_ADD_RTN_U32>;
1025def DS_SUB_RTN_U32_vi : DS_Real_vi<0x21, DS_SUB_RTN_U32>;
1026def DS_RSUB_RTN_U32_vi : DS_Real_vi<0x22, DS_RSUB_RTN_U32>;
1027def DS_INC_RTN_U32_vi : DS_Real_vi<0x23, DS_INC_RTN_U32>;
1028def DS_DEC_RTN_U32_vi : DS_Real_vi<0x24, DS_DEC_RTN_U32>;
1029def DS_MIN_RTN_I32_vi : DS_Real_vi<0x25, DS_MIN_RTN_I32>;
1030def DS_MAX_RTN_I32_vi : DS_Real_vi<0x26, DS_MAX_RTN_I32>;
1031def DS_MIN_RTN_U32_vi : DS_Real_vi<0x27, DS_MIN_RTN_U32>;
1032def DS_MAX_RTN_U32_vi : DS_Real_vi<0x28, DS_MAX_RTN_U32>;
1033def DS_AND_RTN_B32_vi : DS_Real_vi<0x29, DS_AND_RTN_B32>;
1034def DS_OR_RTN_B32_vi : DS_Real_vi<0x2a, DS_OR_RTN_B32>;
1035def DS_XOR_RTN_B32_vi : DS_Real_vi<0x2b, DS_XOR_RTN_B32>;
1036def DS_MSKOR_RTN_B32_vi : DS_Real_vi<0x2c, DS_MSKOR_RTN_B32>;
1037def DS_WRXCHG_RTN_B32_vi : DS_Real_vi<0x2d, DS_WRXCHG_RTN_B32>;
1038def DS_WRXCHG2_RTN_B32_vi : DS_Real_vi<0x2e, DS_WRXCHG2_RTN_B32>;
1039def DS_WRXCHG2ST64_RTN_B32_vi : DS_Real_vi<0x2f, DS_WRXCHG2ST64_RTN_B32>;
1040def DS_CMPST_RTN_B32_vi : DS_Real_vi<0x30, DS_CMPST_RTN_B32>;
1041def DS_CMPST_RTN_F32_vi : DS_Real_vi<0x31, DS_CMPST_RTN_F32>;
1042def DS_MIN_RTN_F32_vi : DS_Real_vi<0x32, DS_MIN_RTN_F32>;
1043def DS_MAX_RTN_F32_vi : DS_Real_vi<0x33, DS_MAX_RTN_F32>;
Dmitry Preobrazhenskye5147242017-04-07 13:07:13 +00001044def DS_WRAP_RTN_B32_vi : DS_Real_vi<0x34, DS_WRAP_RTN_B32>;
Artem Tamazov2e217b82016-09-21 16:35:44 +00001045def DS_ADD_RTN_F32_vi : DS_Real_vi<0x35, DS_ADD_RTN_F32>;
Valery Pykhtin902db312016-08-01 14:21:30 +00001046def DS_READ_B32_vi : DS_Real_vi<0x36, DS_READ_B32>;
1047def DS_READ2_B32_vi : DS_Real_vi<0x37, DS_READ2_B32>;
1048def DS_READ2ST64_B32_vi : DS_Real_vi<0x38, DS_READ2ST64_B32>;
1049def DS_READ_I8_vi : DS_Real_vi<0x39, DS_READ_I8>;
1050def DS_READ_U8_vi : DS_Real_vi<0x3a, DS_READ_U8>;
1051def DS_READ_I16_vi : DS_Real_vi<0x3b, DS_READ_I16>;
1052def DS_READ_U16_vi : DS_Real_vi<0x3c, DS_READ_U16>;
Matt Arsenaultefa1d652017-09-01 18:38:02 +00001053def DS_READ_ADDTID_B32_vi : DS_Real_vi<0xb6, DS_READ_ADDTID_B32>;
Dmitry Preobrazhenskye5147242017-04-07 13:07:13 +00001054def DS_CONSUME_vi : DS_Real_vi<0xbd, DS_CONSUME>;
1055def DS_APPEND_vi : DS_Real_vi<0xbe, DS_APPEND>;
1056def DS_ORDERED_COUNT_vi : DS_Real_vi<0xbf, DS_ORDERED_COUNT>;
Valery Pykhtin902db312016-08-01 14:21:30 +00001057def DS_SWIZZLE_B32_vi : DS_Real_vi<0x3d, DS_SWIZZLE_B32>;
1058def DS_PERMUTE_B32_vi : DS_Real_vi<0x3e, DS_PERMUTE_B32>;
1059def DS_BPERMUTE_B32_vi : DS_Real_vi<0x3f, DS_BPERMUTE_B32>;
1060
1061def DS_ADD_U64_vi : DS_Real_vi<0x40, DS_ADD_U64>;
1062def DS_SUB_U64_vi : DS_Real_vi<0x41, DS_SUB_U64>;
1063def DS_RSUB_U64_vi : DS_Real_vi<0x42, DS_RSUB_U64>;
1064def DS_INC_U64_vi : DS_Real_vi<0x43, DS_INC_U64>;
1065def DS_DEC_U64_vi : DS_Real_vi<0x44, DS_DEC_U64>;
1066def DS_MIN_I64_vi : DS_Real_vi<0x45, DS_MIN_I64>;
1067def DS_MAX_I64_vi : DS_Real_vi<0x46, DS_MAX_I64>;
1068def DS_MIN_U64_vi : DS_Real_vi<0x47, DS_MIN_U64>;
1069def DS_MAX_U64_vi : DS_Real_vi<0x48, DS_MAX_U64>;
1070def DS_AND_B64_vi : DS_Real_vi<0x49, DS_AND_B64>;
1071def DS_OR_B64_vi : DS_Real_vi<0x4a, DS_OR_B64>;
1072def DS_XOR_B64_vi : DS_Real_vi<0x4b, DS_XOR_B64>;
1073def DS_MSKOR_B64_vi : DS_Real_vi<0x4c, DS_MSKOR_B64>;
1074def DS_WRITE_B64_vi : DS_Real_vi<0x4d, DS_WRITE_B64>;
1075def DS_WRITE2_B64_vi : DS_Real_vi<0x4E, DS_WRITE2_B64>;
1076def DS_WRITE2ST64_B64_vi : DS_Real_vi<0x4f, DS_WRITE2ST64_B64>;
1077def DS_CMPST_B64_vi : DS_Real_vi<0x50, DS_CMPST_B64>;
1078def DS_CMPST_F64_vi : DS_Real_vi<0x51, DS_CMPST_F64>;
1079def DS_MIN_F64_vi : DS_Real_vi<0x52, DS_MIN_F64>;
1080def DS_MAX_F64_vi : DS_Real_vi<0x53, DS_MAX_F64>;
1081
Matt Arsenaulted6e8f02017-09-01 18:36:06 +00001082def DS_WRITE_B8_D16_HI_vi : DS_Real_vi<0x54, DS_WRITE_B8_D16_HI>;
1083def DS_WRITE_B16_D16_HI_vi : DS_Real_vi<0x55, DS_WRITE_B16_D16_HI>;
1084
1085def DS_READ_U8_D16_vi : DS_Real_vi<0x56, DS_READ_U8_D16>;
1086def DS_READ_U8_D16_HI_vi : DS_Real_vi<0x57, DS_READ_U8_D16_HI>;
1087def DS_READ_I8_D16_vi : DS_Real_vi<0x58, DS_READ_I8_D16>;
1088def DS_READ_I8_D16_HI_vi : DS_Real_vi<0x59, DS_READ_I8_D16_HI>;
1089def DS_READ_U16_D16_vi : DS_Real_vi<0x5a, DS_READ_U16_D16>;
1090def DS_READ_U16_D16_HI_vi : DS_Real_vi<0x5b, DS_READ_U16_D16_HI>;
1091
Valery Pykhtin902db312016-08-01 14:21:30 +00001092def DS_ADD_RTN_U64_vi : DS_Real_vi<0x60, DS_ADD_RTN_U64>;
1093def DS_SUB_RTN_U64_vi : DS_Real_vi<0x61, DS_SUB_RTN_U64>;
1094def DS_RSUB_RTN_U64_vi : DS_Real_vi<0x62, DS_RSUB_RTN_U64>;
1095def DS_INC_RTN_U64_vi : DS_Real_vi<0x63, DS_INC_RTN_U64>;
1096def DS_DEC_RTN_U64_vi : DS_Real_vi<0x64, DS_DEC_RTN_U64>;
1097def DS_MIN_RTN_I64_vi : DS_Real_vi<0x65, DS_MIN_RTN_I64>;
1098def DS_MAX_RTN_I64_vi : DS_Real_vi<0x66, DS_MAX_RTN_I64>;
1099def DS_MIN_RTN_U64_vi : DS_Real_vi<0x67, DS_MIN_RTN_U64>;
1100def DS_MAX_RTN_U64_vi : DS_Real_vi<0x68, DS_MAX_RTN_U64>;
1101def DS_AND_RTN_B64_vi : DS_Real_vi<0x69, DS_AND_RTN_B64>;
1102def DS_OR_RTN_B64_vi : DS_Real_vi<0x6a, DS_OR_RTN_B64>;
1103def DS_XOR_RTN_B64_vi : DS_Real_vi<0x6b, DS_XOR_RTN_B64>;
1104def DS_MSKOR_RTN_B64_vi : DS_Real_vi<0x6c, DS_MSKOR_RTN_B64>;
1105def DS_WRXCHG_RTN_B64_vi : DS_Real_vi<0x6d, DS_WRXCHG_RTN_B64>;
1106def DS_WRXCHG2_RTN_B64_vi : DS_Real_vi<0x6e, DS_WRXCHG2_RTN_B64>;
1107def DS_WRXCHG2ST64_RTN_B64_vi : DS_Real_vi<0x6f, DS_WRXCHG2ST64_RTN_B64>;
Dmitry Preobrazhenskye5147242017-04-07 13:07:13 +00001108def DS_CONDXCHG32_RTN_B64_vi : DS_Real_vi<0x7e, DS_CONDXCHG32_RTN_B64>;
1109def DS_GWS_SEMA_RELEASE_ALL_vi : DS_Real_vi<0x98, DS_GWS_SEMA_RELEASE_ALL>;
Valery Pykhtin902db312016-08-01 14:21:30 +00001110def DS_CMPST_RTN_B64_vi : DS_Real_vi<0x70, DS_CMPST_RTN_B64>;
1111def DS_CMPST_RTN_F64_vi : DS_Real_vi<0x71, DS_CMPST_RTN_F64>;
1112def DS_MIN_RTN_F64_vi : DS_Real_vi<0x72, DS_MIN_RTN_F64>;
1113def DS_MAX_RTN_F64_vi : DS_Real_vi<0x73, DS_MAX_RTN_F64>;
1114
1115def DS_READ_B64_vi : DS_Real_vi<0x76, DS_READ_B64>;
1116def DS_READ2_B64_vi : DS_Real_vi<0x77, DS_READ2_B64>;
1117def DS_READ2ST64_B64_vi : DS_Real_vi<0x78, DS_READ2ST64_B64>;
1118
1119def DS_ADD_SRC2_U32_vi : DS_Real_vi<0x80, DS_ADD_SRC2_U32>;
1120def DS_SUB_SRC2_U32_vi : DS_Real_vi<0x81, DS_SUB_SRC2_U32>;
1121def DS_RSUB_SRC2_U32_vi : DS_Real_vi<0x82, DS_RSUB_SRC2_U32>;
1122def DS_INC_SRC2_U32_vi : DS_Real_vi<0x83, DS_INC_SRC2_U32>;
1123def DS_DEC_SRC2_U32_vi : DS_Real_vi<0x84, DS_DEC_SRC2_U32>;
1124def DS_MIN_SRC2_I32_vi : DS_Real_vi<0x85, DS_MIN_SRC2_I32>;
1125def DS_MAX_SRC2_I32_vi : DS_Real_vi<0x86, DS_MAX_SRC2_I32>;
1126def DS_MIN_SRC2_U32_vi : DS_Real_vi<0x87, DS_MIN_SRC2_U32>;
1127def DS_MAX_SRC2_U32_vi : DS_Real_vi<0x88, DS_MAX_SRC2_U32>;
1128def DS_AND_SRC2_B32_vi : DS_Real_vi<0x89, DS_AND_SRC2_B32>;
1129def DS_OR_SRC2_B32_vi : DS_Real_vi<0x8a, DS_OR_SRC2_B32>;
1130def DS_XOR_SRC2_B32_vi : DS_Real_vi<0x8b, DS_XOR_SRC2_B32>;
1131def DS_WRITE_SRC2_B32_vi : DS_Real_vi<0x8d, DS_WRITE_SRC2_B32>;
1132def DS_MIN_SRC2_F32_vi : DS_Real_vi<0x92, DS_MIN_SRC2_F32>;
1133def DS_MAX_SRC2_F32_vi : DS_Real_vi<0x93, DS_MAX_SRC2_F32>;
Dmitry Preobrazhensky622bde82018-03-28 16:21:56 +00001134def DS_ADD_SRC2_F32_vi : DS_Real_vi<0x95, DS_ADD_SRC2_F32>;
Valery Pykhtin902db312016-08-01 14:21:30 +00001135def DS_ADD_SRC2_U64_vi : DS_Real_vi<0xc0, DS_ADD_SRC2_U64>;
1136def DS_SUB_SRC2_U64_vi : DS_Real_vi<0xc1, DS_SUB_SRC2_U64>;
1137def DS_RSUB_SRC2_U64_vi : DS_Real_vi<0xc2, DS_RSUB_SRC2_U64>;
1138def DS_INC_SRC2_U64_vi : DS_Real_vi<0xc3, DS_INC_SRC2_U64>;
1139def DS_DEC_SRC2_U64_vi : DS_Real_vi<0xc4, DS_DEC_SRC2_U64>;
1140def DS_MIN_SRC2_I64_vi : DS_Real_vi<0xc5, DS_MIN_SRC2_I64>;
1141def DS_MAX_SRC2_I64_vi : DS_Real_vi<0xc6, DS_MAX_SRC2_I64>;
1142def DS_MIN_SRC2_U64_vi : DS_Real_vi<0xc7, DS_MIN_SRC2_U64>;
1143def DS_MAX_SRC2_U64_vi : DS_Real_vi<0xc8, DS_MAX_SRC2_U64>;
1144def DS_AND_SRC2_B64_vi : DS_Real_vi<0xc9, DS_AND_SRC2_B64>;
1145def DS_OR_SRC2_B64_vi : DS_Real_vi<0xca, DS_OR_SRC2_B64>;
1146def DS_XOR_SRC2_B64_vi : DS_Real_vi<0xcb, DS_XOR_SRC2_B64>;
1147def DS_WRITE_SRC2_B64_vi : DS_Real_vi<0xcd, DS_WRITE_SRC2_B64>;
1148def DS_MIN_SRC2_F64_vi : DS_Real_vi<0xd2, DS_MIN_SRC2_F64>;
1149def DS_MAX_SRC2_F64_vi : DS_Real_vi<0xd3, DS_MAX_SRC2_F64>;
Matt Arsenaultdedc5442017-02-28 20:15:43 +00001150def DS_WRITE_B96_vi : DS_Real_vi<0xde, DS_WRITE_B96>;
1151def DS_WRITE_B128_vi : DS_Real_vi<0xdf, DS_WRITE_B128>;
1152def DS_READ_B96_vi : DS_Real_vi<0xfe, DS_READ_B96>;
1153def DS_READ_B128_vi : DS_Real_vi<0xff, DS_READ_B128>;