1//===-- FLATInstructions.td - FLAT Instruction Definitions ----------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8
9def FlatOffset : ComplexPattern<iPTR, 2, "SelectFlatOffset", [], [SDNPWantRoot], -10>;
10def GlobalOffset : ComplexPattern<iPTR, 2, "SelectGlobalOffset", [], [SDNPWantRoot], -10>;
11def ScratchOffset : ComplexPattern<iPTR, 2, "SelectScratchOffset", [], [SDNPWantRoot], -10>;
12
13def GlobalSAddr : ComplexPattern<iPTR, 3, "SelectGlobalSAddr", [], [SDNPWantRoot], -10>;
14def ScratchSAddr : ComplexPattern<iPTR, 2, "SelectScratchSAddr", [], [SDNPWantRoot], -10>;
15def ScratchSVAddr : ComplexPattern<iPTR, 3, "SelectScratchSVAddr", [], [SDNPWantRoot], -10>;
16
17//===----------------------------------------------------------------------===//
18// FLAT classes
19//===----------------------------------------------------------------------===//
20
21class FLAT_Pseudo<string opName, dag outs, dag ins,
22                  string asmOps, list<dag> pattern=[]> :
23  InstSI<outs, ins, "", pattern>,
24  SIMCInstr<opName, SIEncodingFamily.NONE> {
25
26  let isPseudo = 1;
27  let isCodeGenOnly = 1;
28
29  let FLAT = 1;
30
31  let UseNamedOperandTable = 1;
32  let hasSideEffects = 0;
33  let SchedRW = [WriteVMEM];
34
35  string Mnemonic = opName;
36  string AsmOperands = asmOps;
37
38  bits<1> is_flat_global = 0;
39  bits<1> is_flat_scratch = 0;
40
41  bits<1> has_vdst = 1;
42
43  // We need to distinguish having saddr and enabling saddr because
44  // saddr is only valid for scratch and global instructions. Pre-gfx9
45  // these bits were reserved, so we also don't necessarily want to
46  // set these bits to the disabled value for the original flat
47  // segment instructions.
48  bits<1> has_saddr = 0;
49  bits<1> enabled_saddr = 0;
50  bits<7> saddr_value = 0;
51  bits<1> has_vaddr = 1;
52
53  bits<1> has_data = 1;
54  bits<1> has_glc  = 1;
55  bits<1> glcValue = 0;
56  bits<1> has_dlc  = 1;
57  bits<1> dlcValue = 0;
58  bits<1> has_sccb  = 1;
59  bits<1> sccbValue = 0;
60  bits<1> has_sve  = 0; // Scratch VGPR Enable
61  bits<1> sve = 0;
62
63  let SubtargetPredicate = !if(is_flat_global, HasFlatGlobalInsts,
64    !if(is_flat_scratch, HasFlatScratchInsts, HasFlatAddressSpace));
65
66  // TODO: M0 if it could possibly access LDS (before gfx9? only)?
67  let Uses = !if(is_flat_global, [EXEC], [EXEC, FLAT_SCR]);
68
69  // Internally, FLAT instruction are executed as both an LDS and a
70  // Buffer instruction; so, they increment both VM_CNT and LGKM_CNT
71  // and are not considered done until both have been decremented.
72  let VM_CNT = 1;
73  let LGKM_CNT = !not(!or(is_flat_global, is_flat_scratch));
74
75  let FlatGlobal = is_flat_global;
76
77  let FlatScratch = is_flat_scratch;
78}
79
80class FLAT_Real <bits<7> op, FLAT_Pseudo ps> :
81  InstSI <ps.OutOperandList, ps.InOperandList, ps.Mnemonic # ps.AsmOperands, []>,
82  Enc64 {
83
84  let isPseudo = 0;
85  let isCodeGenOnly = 0;
86
87  let FLAT = 1;
88
89  // copy relevant pseudo op flags
90  let SubtargetPredicate   = ps.SubtargetPredicate;
91  let AsmMatchConverter    = ps.AsmMatchConverter;
92  let OtherPredicates      = ps.OtherPredicates;
93  let TSFlags              = ps.TSFlags;
94  let UseNamedOperandTable = ps.UseNamedOperandTable;
95  let SchedRW              = ps.SchedRW;
96  let mayLoad              = ps.mayLoad;
97  let mayStore             = ps.mayStore;
98  let IsAtomicRet          = ps.IsAtomicRet;
99  let IsAtomicNoRet        = ps.IsAtomicNoRet;
100  let VM_CNT               = ps.VM_CNT;
101  let LGKM_CNT             = ps.LGKM_CNT;
102  let VALU                 = ps.VALU;
103
104  // encoding fields
105  bits<8> vaddr;
106  bits<10> vdata;
107  bits<7> saddr;
108  bits<10> vdst;
109
110  bits<5> cpol;
111
112  // Only valid on gfx9
113  bits<1> lds = 0; // XXX - What does this actually do?
114
115  // Segment, 00=flat, 01=scratch, 10=global, 11=reserved
116  bits<2> seg = !if(ps.is_flat_global, 0b10,
117                  !if(ps.is_flat_scratch, 0b01, 0));
118
119  // Signed offset. Highest bit ignored for flat and treated as 12-bit
120  // unsigned for flat accesses.
121  bits<13> offset;
122  // GFX90A+ only: instruction uses AccVGPR for data
123  bits<1> acc = !if(ps.has_vdst, vdst{9}, !if(ps.has_data, vdata{9}, 0));
124
125  // We don't use tfe right now, and it was removed in gfx9.
126  bits<1> tfe = 0;
127
128  // Only valid on GFX9+
129  let Inst{12-0} = offset;
130  let Inst{13} = !if(ps.has_sve, ps.sve, lds);
131  let Inst{15-14} = seg;
132
133  let Inst{16}    = !if(ps.has_glc, cpol{CPolBit.GLC}, ps.glcValue);
134  let Inst{17}    = cpol{CPolBit.SLC};
135  let Inst{24-18} = op;
136  let Inst{31-26} = 0x37; // Encoding.
137  let Inst{39-32} = !if(ps.has_vaddr, vaddr, ?);
138  let Inst{47-40} = !if(ps.has_data, vdata{7-0}, ?);
139  let Inst{54-48} = !if(ps.has_saddr, !if(ps.enabled_saddr, saddr, 0x7f), 0);
140
141  // 54-48 is reserved.
142  let Inst{55}    = acc; // nv on GFX9+, TFE before. AccVGPR for data on GFX90A.
143  let Inst{63-56} = !if(ps.has_vdst, vdst{7-0}, ?);
144}
145
146class GlobalSaddrTable <bit is_saddr, string Name = ""> {
147  bit IsSaddr = is_saddr;
148  string SaddrOp = Name;
149}
150
151// TODO: Is exec allowed for saddr? The disabled value 0x7f is the
152// same encoding value as exec_hi, so it isn't possible to use that if
153// saddr is 32-bit (which isn't handled here yet).
154class FLAT_Load_Pseudo <string opName, RegisterClass regClass,
155  bit HasTiedOutput = 0,
156  bit HasSaddr = 0, bit EnableSaddr = 0,
157  RegisterOperand vdata_op = getLdStRegisterOperand<regClass>.ret> : FLAT_Pseudo<
158  opName,
159  (outs vdata_op:$vdst),
160  !con(
161    !con(
162      !if(EnableSaddr,
163        (ins SReg_64:$saddr, VGPR_32:$vaddr),
164        (ins VReg_64:$vaddr)),
165        (ins flat_offset:$offset)),
166        // FIXME: Operands with default values do not work with following non-optional operands.
167        !if(HasTiedOutput, (ins CPol:$cpol, vdata_op:$vdst_in),
168                           (ins CPol_0:$cpol))),
169  " $vdst, $vaddr"#!if(HasSaddr, !if(EnableSaddr, ", $saddr", ", off"), "")#"$offset$cpol"> {
170  let has_data = 0;
171  let mayLoad = 1;
172  let has_saddr = HasSaddr;
173  let enabled_saddr = EnableSaddr;
174  let PseudoInstr = opName#!if(!and(HasSaddr, EnableSaddr), "_SADDR", "");
175  let maybeAtomic = 1;
176
177  let Constraints = !if(HasTiedOutput, "$vdst = $vdst_in", "");
178  let DisableEncoding = !if(HasTiedOutput, "$vdst_in", "");
179}
180
181class FLAT_Store_Pseudo <string opName, RegisterClass vdataClass,
182  bit HasSaddr = 0, bit EnableSaddr = 0> : FLAT_Pseudo<
183  opName,
184  (outs),
185  !con(
186    !if(EnableSaddr,
187      (ins VGPR_32:$vaddr, getLdStRegisterOperand<vdataClass>.ret:$vdata, SReg_64:$saddr),
188      (ins VReg_64:$vaddr, getLdStRegisterOperand<vdataClass>.ret:$vdata)),
189      (ins flat_offset:$offset, CPol_0:$cpol)),
190  " $vaddr, $vdata"#!if(HasSaddr, !if(EnableSaddr, ", $saddr", ", off"), "")#"$offset$cpol"> {
191  let mayLoad  = 0;
192  let mayStore = 1;
193  let has_vdst = 0;
194  let has_saddr = HasSaddr;
195  let enabled_saddr = EnableSaddr;
196  let PseudoInstr = opName#!if(!and(HasSaddr, EnableSaddr), "_SADDR", "");
197  let maybeAtomic = 1;
198}
199
200multiclass FLAT_Global_Load_Pseudo<string opName, RegisterClass regClass, bit HasTiedInput = 0> {
201  let is_flat_global = 1, SubtargetPredicate = HasFlatGlobalInsts in {
202    def "" : FLAT_Load_Pseudo<opName, regClass, HasTiedInput, 1>,
203      GlobalSaddrTable<0, opName>;
204    def _SADDR : FLAT_Load_Pseudo<opName, regClass, HasTiedInput, 1, 1>,
205      GlobalSaddrTable<1, opName>;
206  }
207}
208
209class FLAT_Global_Load_AddTid_Pseudo <string opName, RegisterClass regClass,
210  bit HasTiedOutput = 0, bit EnableSaddr = 0> : FLAT_Pseudo<
211  opName,
212  (outs regClass:$vdst),
213  !con(!if(EnableSaddr, (ins SReg_64:$saddr), (ins)),
214    (ins flat_offset:$offset, CPol_0:$cpol),
215    !if(HasTiedOutput, (ins regClass:$vdst_in), (ins))),
216  " $vdst, "#!if(EnableSaddr, "$saddr", "off")#"$offset$cpol"> {
217  let is_flat_global = 1;
218  let has_data = 0;
219  let mayLoad = 1;
220  let has_vaddr = 0;
221  let has_saddr = 1;
222  let enabled_saddr = EnableSaddr;
223  let maybeAtomic = 1;
224  let PseudoInstr = opName#!if(EnableSaddr, "_SADDR", "");
225
226  let Constraints = !if(HasTiedOutput, "$vdst = $vdst_in", "");
227  let DisableEncoding = !if(HasTiedOutput, "$vdst_in", "");
228}
229
230multiclass FLAT_Global_Load_AddTid_Pseudo<string opName, RegisterClass regClass,
231  bit HasTiedOutput = 0> {
232  def "" : FLAT_Global_Load_AddTid_Pseudo<opName, regClass, HasTiedOutput>,
233    GlobalSaddrTable<0, opName>;
234  def _SADDR : FLAT_Global_Load_AddTid_Pseudo<opName, regClass, HasTiedOutput, 1>,
235    GlobalSaddrTable<1, opName>;
236}
237
238multiclass FLAT_Global_Store_Pseudo<string opName, RegisterClass regClass> {
239  let is_flat_global = 1, SubtargetPredicate = HasFlatGlobalInsts in {
240    def "" : FLAT_Store_Pseudo<opName, regClass, 1>,
241      GlobalSaddrTable<0, opName>;
242    def _SADDR : FLAT_Store_Pseudo<opName, regClass, 1, 1>,
243      GlobalSaddrTable<1, opName>;
244  }
245}
246
247class FLAT_Global_Load_LDS_Pseudo <string opName, bit EnableSaddr = 0> : FLAT_Pseudo<
248  opName,
249  (outs ),
250  !con(
251      !if(EnableSaddr, (ins SReg_64:$saddr, VGPR_32:$vaddr), (ins VReg_64:$vaddr)),
252      (ins flat_offset:$offset, CPol_0:$cpol)),
253  " $vaddr"#!if(EnableSaddr, ", $saddr", ", off")#"$offset$cpol"> {
254  let LGKM_CNT = 1;
255  let is_flat_global = 1;
256  let has_data = 0;
257  let has_vdst = 0;
258  let mayLoad = 1;
259  let mayStore = 1;
260  let has_saddr = 1;
261  let enabled_saddr = EnableSaddr;
262  let VALU = 1;
263  let PseudoInstr = opName#!if(EnableSaddr, "_SADDR", "");
264  let Uses = [M0, EXEC];
265  let SchedRW = [WriteVMEM, WriteLDS];
266}
267
268multiclass FLAT_Global_Load_LDS_Pseudo<string opName> {
269  def ""     : FLAT_Global_Load_LDS_Pseudo<opName>,
270    GlobalSaddrTable<0, opName>;
271  def _SADDR : FLAT_Global_Load_LDS_Pseudo<opName, 1>,
272    GlobalSaddrTable<1, opName>;
273}
274
275class FLAT_Global_Store_AddTid_Pseudo <string opName, RegisterClass vdataClass,
276  bit EnableSaddr = 0> : FLAT_Pseudo<
277  opName,
278  (outs),
279  !con(!if(EnableSaddr, (ins vdataClass:$vdata, SReg_64:$saddr), (ins vdataClass:$vdata)),
280    (ins flat_offset:$offset, CPol:$cpol)),
281  " $vdata, "#!if(EnableSaddr, "$saddr", "off")#"$offset$cpol"> {
282  let is_flat_global = 1;
283  let mayLoad  = 0;
284  let mayStore = 1;
285  let has_vdst = 0;
286  let has_vaddr = 0;
287  let has_saddr = 1;
288  let enabled_saddr = EnableSaddr;
289  let maybeAtomic = 1;
290  let PseudoInstr = opName#!if(EnableSaddr, "_SADDR", "");
291}
292
293multiclass FLAT_Global_Store_AddTid_Pseudo<string opName, RegisterClass regClass> {
294  def "" : FLAT_Global_Store_AddTid_Pseudo<opName, regClass>,
295    GlobalSaddrTable<0, opName>;
296  def _SADDR : FLAT_Global_Store_AddTid_Pseudo<opName, regClass, 1>,
297    GlobalSaddrTable<1, opName>;
298}
299
300class FlatScratchInst <string sv_op, string mode> {
301  string SVOp = sv_op;
302  string Mode = mode;
303}
304
305class FLAT_Scratch_Load_Pseudo <string opName, RegisterClass regClass,
306  bit HasTiedOutput = 0,
307  bit EnableSaddr = 0,
308  bit EnableSVE = 0,
309  bit EnableVaddr = !or(EnableSVE, !not(EnableSaddr))>
310  : FLAT_Pseudo<
311  opName,
312  (outs getLdStRegisterOperand<regClass>.ret:$vdst),
313  !con(
314    !if(EnableSVE,
315        (ins VGPR_32:$vaddr, SReg_32_XEXEC_HI:$saddr, flat_offset:$offset),
316        !if(EnableSaddr,
317          (ins SReg_32_XEXEC_HI:$saddr, flat_offset:$offset),
318          !if(EnableVaddr,
319            (ins VGPR_32:$vaddr, flat_offset:$offset),
320            (ins flat_offset:$offset)))),
321     !if(HasTiedOutput, (ins CPol:$cpol, getLdStRegisterOperand<regClass>.ret:$vdst_in),
322                        (ins CPol_0:$cpol))),
323  " $vdst, "#!if(EnableVaddr, "$vaddr, ", "off, ")#!if(EnableSaddr, "$saddr", "off")#"$offset$cpol"> {
324  let has_data = 0;
325  let mayLoad = 1;
326  let has_saddr = 1;
327  let enabled_saddr = EnableSaddr;
328  let has_vaddr = EnableVaddr;
329  let has_sve = EnableSVE;
330  let sve = EnableVaddr;
331  let PseudoInstr = opName#!if(EnableSVE, "_SVS", !if(EnableSaddr, "_SADDR", !if(EnableVaddr, "", "_ST")));
332  let maybeAtomic = 1;
333
334  let Constraints = !if(HasTiedOutput, "$vdst = $vdst_in", "");
335  let DisableEncoding = !if(HasTiedOutput, "$vdst_in", "");
336}
337
338class FLAT_Scratch_Store_Pseudo <string opName, RegisterClass vdataClass, bit EnableSaddr = 0,
339  bit EnableSVE = 0,
340  bit EnableVaddr = !or(EnableSVE, !not(EnableSaddr)),
341  RegisterOperand vdata_op = getLdStRegisterOperand<vdataClass>.ret> : FLAT_Pseudo<
342  opName,
343  (outs),
344  !if(EnableSVE,
345    (ins vdata_op:$vdata, VGPR_32:$vaddr, SReg_32_XEXEC_HI:$saddr, flat_offset:$offset, CPol_0:$cpol),
346    !if(EnableSaddr,
347      (ins vdata_op:$vdata, SReg_32_XEXEC_HI:$saddr, flat_offset:$offset, CPol_0:$cpol),
348      !if(EnableVaddr,
349        (ins vdata_op:$vdata, VGPR_32:$vaddr, flat_offset:$offset, CPol_0:$cpol),
350        (ins vdata_op:$vdata, flat_offset:$offset, CPol_0:$cpol)))),
351  " "#!if(EnableVaddr, "$vaddr", "off")#", $vdata, "#!if(EnableSaddr, "$saddr", "off")#"$offset$cpol"> {
352  let mayLoad  = 0;
353  let mayStore = 1;
354  let has_vdst = 0;
355  let has_saddr = 1;
356  let enabled_saddr = EnableSaddr;
357  let has_vaddr = EnableVaddr;
358  let has_sve = EnableSVE;
359  let sve = EnableVaddr;
360  let PseudoInstr = opName#!if(EnableSVE, "_SVS", !if(EnableSaddr, "_SADDR", !if(EnableVaddr, "", "_ST")));
361  let maybeAtomic = 1;
362}
363
364multiclass FLAT_Scratch_Load_Pseudo<string opName, RegisterClass regClass, bit HasTiedOutput = 0> {
365  let is_flat_scratch = 1 in {
366    def "" : FLAT_Scratch_Load_Pseudo<opName, regClass, HasTiedOutput>,
367             FlatScratchInst<opName, "SV">;
368    def _SADDR : FLAT_Scratch_Load_Pseudo<opName, regClass, HasTiedOutput, 1>,
369                 FlatScratchInst<opName, "SS">;
370
371    let SubtargetPredicate = HasFlatScratchSVSMode in
372    def _SVS : FLAT_Scratch_Load_Pseudo<opName, regClass, HasTiedOutput, 1, 1>,
373               FlatScratchInst<opName, "SVS">;
374
375    let SubtargetPredicate = HasFlatScratchSTMode in
376    def _ST  : FLAT_Scratch_Load_Pseudo<opName, regClass, HasTiedOutput, 0, 0, 0>,
377               FlatScratchInst<opName, "ST">;
378  }
379}
380
381multiclass FLAT_Scratch_Store_Pseudo<string opName, RegisterClass regClass> {
382  let is_flat_scratch = 1 in {
383    def "" : FLAT_Scratch_Store_Pseudo<opName, regClass>,
384             FlatScratchInst<opName, "SV">;
385    def _SADDR : FLAT_Scratch_Store_Pseudo<opName, regClass, 1>,
386                 FlatScratchInst<opName, "SS">;
387
388    let SubtargetPredicate = HasFlatScratchSVSMode in
389    def _SVS : FLAT_Scratch_Store_Pseudo<opName, regClass, 1, 1>,
390               FlatScratchInst<opName, "SVS">;
391
392    let SubtargetPredicate = HasFlatScratchSTMode in
393    def _ST  : FLAT_Scratch_Store_Pseudo<opName, regClass, 0, 0, 0>,
394               FlatScratchInst<opName, "ST">;
395  }
396}
397
398class FLAT_Scratch_Load_LDS_Pseudo <string opName, bit EnableSaddr = 0,
399  bit EnableSVE = 0,
400  bit EnableVaddr = !or(EnableSVE, !not(EnableSaddr))> : FLAT_Pseudo<
401  opName,
402  (outs ),
403  !if(EnableSVE,
404    (ins VGPR_32:$vaddr, SReg_32_XEXEC_HI:$saddr, flat_offset:$offset, CPol:$cpol),
405    !if(EnableSaddr,
406      (ins SReg_32_XEXEC_HI:$saddr, flat_offset:$offset, CPol:$cpol),
407      !if(EnableVaddr,
408        (ins VGPR_32:$vaddr, flat_offset:$offset, CPol:$cpol),
409        (ins flat_offset:$offset, CPol:$cpol)))),
410  " "#!if(EnableVaddr, "$vaddr, ", "off, ")#!if(EnableSaddr, "$saddr", "off")#"$offset$cpol"> {
411
412  let LGKM_CNT = 1;
413  let is_flat_scratch = 1;
414  let has_data = 0;
415  let has_vdst = 0;
416  let mayLoad = 1;
417  let mayStore = 1;
418  let has_saddr = 1;
419  let enabled_saddr = EnableSaddr;
420  let has_vaddr = EnableVaddr;
421  let has_sve = EnableSVE;
422  let sve = EnableVaddr;
423  let VALU = 1;
424  let PseudoInstr = opName#!if(EnableSVE, "_SVS", !if(EnableSaddr, "_SADDR", !if(EnableVaddr, "", "_ST")));
425  let Uses = [M0, EXEC];
426  let SchedRW = [WriteVMEM, WriteLDS];
427}
428
429multiclass FLAT_Scratch_Load_LDS_Pseudo<string opName> {
430  def ""     : FLAT_Scratch_Load_LDS_Pseudo<opName>,
431               FlatScratchInst<opName, "SV">;
432  def _SADDR : FLAT_Scratch_Load_LDS_Pseudo<opName, 1>,
433               FlatScratchInst<opName, "SS">;
434  def _SVS   : FLAT_Scratch_Load_LDS_Pseudo<opName, 1, 1>,
435               FlatScratchInst<opName, "SVS">;
436  def _ST    : FLAT_Scratch_Load_LDS_Pseudo<opName, 0, 0, 0>,
437               FlatScratchInst<opName, "ST">;
438}
439
440class FLAT_AtomicNoRet_Pseudo<string opName, dag outs, dag ins,
441                               string asm, list<dag> pattern = []> :
442  FLAT_Pseudo<opName, outs, ins, asm, pattern> {
443    let mayLoad = 1;
444    let mayStore = 1;
445    let has_glc  = 0;
446    let glcValue = 0;
447    let has_vdst = 0;
448    let has_sccb  = 1;
449    let sccbValue = 0;
450    let maybeAtomic = 1;
451    let IsAtomicNoRet = 1;
452}
453
454class FLAT_AtomicRet_Pseudo<string opName, dag outs, dag ins,
455                            string asm, list<dag> pattern = []>
456  : FLAT_AtomicNoRet_Pseudo<opName, outs, ins, asm, pattern> {
457  let hasPostISelHook = 1;
458  let has_vdst = 1;
459  let glcValue = 1;
460  let sccbValue = 0;
461  let IsAtomicNoRet = 0;
462  let IsAtomicRet = 1;
463  let PseudoInstr = NAME # "_RTN";
464}
465
466multiclass FLAT_Atomic_Pseudo<
467  string opName,
468  RegisterClass vdst_rc,
469  ValueType vt,
470  ValueType data_vt = vt,
471  RegisterClass data_rc = vdst_rc,
472  bit isFP = isFloatType<data_vt>.ret,
473  RegisterOperand data_op = getLdStRegisterOperand<data_rc>.ret> {
474  def "" : FLAT_AtomicNoRet_Pseudo <opName,
475    (outs),
476    (ins VReg_64:$vaddr, data_op:$vdata, flat_offset:$offset, CPol_0:$cpol),
477    " $vaddr, $vdata$offset$cpol">,
478    GlobalSaddrTable<0, opName>,
479    AtomicNoRet <opName, 0> {
480    let PseudoInstr = NAME;
481    let FPAtomic = isFP;
482    let AddedComplexity = -1; // Prefer global atomics if available
483  }
484
485  def _RTN : FLAT_AtomicRet_Pseudo <opName,
486    (outs getLdStRegisterOperand<vdst_rc>.ret:$vdst),
487    (ins VReg_64:$vaddr, data_op:$vdata, flat_offset:$offset, CPol_GLC1:$cpol),
488    " $vdst, $vaddr, $vdata$offset$cpol">,
489    GlobalSaddrTable<0, opName#"_rtn">,
490    AtomicNoRet <opName, 1> {
491    let FPAtomic = isFP;
492    let AddedComplexity = -1; // Prefer global atomics if available
493  }
494}
495
496multiclass FLAT_Global_Atomic_Pseudo_NO_RTN<
497  string opName,
498  RegisterClass vdst_rc,
499  ValueType vt,
500  ValueType data_vt = vt,
501  RegisterClass data_rc = vdst_rc,
502  bit isFP = isFloatType<data_vt>.ret,
503  RegisterOperand data_op = getLdStRegisterOperand<data_rc>.ret> {
504
505  def "" : FLAT_AtomicNoRet_Pseudo <opName,
506    (outs),
507    (ins VReg_64:$vaddr, data_op:$vdata, flat_offset:$offset, CPol_0:$cpol),
508    " $vaddr, $vdata, off$offset$cpol">,
509    GlobalSaddrTable<0, opName>,
510    AtomicNoRet <opName, 0> {
511    let has_saddr = 1;
512    let PseudoInstr = NAME;
513    let FPAtomic = isFP;
514  }
515
516  def _SADDR : FLAT_AtomicNoRet_Pseudo <opName,
517    (outs),
518    (ins VGPR_32:$vaddr, data_op:$vdata, SReg_64:$saddr, flat_offset:$offset, CPol_0:$cpol),
519    " $vaddr, $vdata, $saddr$offset$cpol">,
520    GlobalSaddrTable<1, opName>,
521    AtomicNoRet <opName#"_saddr", 0> {
522    let has_saddr = 1;
523    let enabled_saddr = 1;
524    let PseudoInstr = NAME#"_SADDR";
525    let FPAtomic = isFP;
526  }
527}
528
529multiclass FLAT_Global_Atomic_Pseudo_RTN<
530  string opName,
531  RegisterClass vdst_rc,
532  ValueType vt,
533  ValueType data_vt = vt,
534  RegisterClass data_rc = vdst_rc,
535  bit isFP = isFloatType<data_vt>.ret,
536  RegisterOperand data_op = getLdStRegisterOperand<data_rc>.ret,
537  RegisterOperand vdst_op = getLdStRegisterOperand<vdst_rc>.ret> {
538
539  def _RTN : FLAT_AtomicRet_Pseudo <opName,
540    (outs vdst_op:$vdst),
541      (ins VReg_64:$vaddr, data_op:$vdata, flat_offset:$offset, CPol_GLC1:$cpol),
542    " $vdst, $vaddr, $vdata, off$offset$cpol">,
543    GlobalSaddrTable<0, opName#"_rtn">,
544    AtomicNoRet <opName, 1> {
545    let has_saddr = 1;
546    let FPAtomic = isFP;
547  }
548
549  def _SADDR_RTN : FLAT_AtomicRet_Pseudo <opName,
550    (outs vdst_op:$vdst),
551      (ins VGPR_32:$vaddr, data_op:$vdata, SReg_64:$saddr, flat_offset:$offset, CPol_GLC1:$cpol),
552    " $vdst, $vaddr, $vdata, $saddr$offset$cpol">,
553    GlobalSaddrTable<1, opName#"_rtn">,
554    AtomicNoRet <opName#"_saddr", 1> {
555     let has_saddr = 1;
556     let enabled_saddr = 1;
557     let PseudoInstr = NAME#"_SADDR_RTN";
558     let FPAtomic = isFP;
559  }
560}
561
562multiclass FLAT_Global_Atomic_Pseudo<
563  string opName,
564  RegisterClass vdst_rc,
565  ValueType vt,
566  ValueType data_vt = vt,
567  RegisterClass data_rc = vdst_rc> {
568  let is_flat_global = 1, SubtargetPredicate = HasFlatGlobalInsts in {
569    defm "" : FLAT_Global_Atomic_Pseudo_NO_RTN<opName, vdst_rc, vt, data_vt, data_rc>;
570    defm "" : FLAT_Global_Atomic_Pseudo_RTN<opName, vdst_rc, vt, data_vt, data_rc>;
571  }
572}
573
574//===----------------------------------------------------------------------===//
575// Flat Instructions
576//===----------------------------------------------------------------------===//
577
578def FLAT_LOAD_UBYTE    : FLAT_Load_Pseudo <"flat_load_ubyte", VGPR_32>;
579def FLAT_LOAD_SBYTE    : FLAT_Load_Pseudo <"flat_load_sbyte", VGPR_32>;
580def FLAT_LOAD_USHORT   : FLAT_Load_Pseudo <"flat_load_ushort", VGPR_32>;
581def FLAT_LOAD_SSHORT   : FLAT_Load_Pseudo <"flat_load_sshort", VGPR_32>;
582def FLAT_LOAD_DWORD    : FLAT_Load_Pseudo <"flat_load_dword", VGPR_32>;
583def FLAT_LOAD_DWORDX2  : FLAT_Load_Pseudo <"flat_load_dwordx2", VReg_64>;
584def FLAT_LOAD_DWORDX4  : FLAT_Load_Pseudo <"flat_load_dwordx4", VReg_128>;
585def FLAT_LOAD_DWORDX3  : FLAT_Load_Pseudo <"flat_load_dwordx3", VReg_96>;
586
587def FLAT_STORE_BYTE    : FLAT_Store_Pseudo <"flat_store_byte", VGPR_32>;
588def FLAT_STORE_SHORT   : FLAT_Store_Pseudo <"flat_store_short", VGPR_32>;
589def FLAT_STORE_DWORD   : FLAT_Store_Pseudo <"flat_store_dword", VGPR_32>;
590def FLAT_STORE_DWORDX2 : FLAT_Store_Pseudo <"flat_store_dwordx2", VReg_64>;
591def FLAT_STORE_DWORDX4 : FLAT_Store_Pseudo <"flat_store_dwordx4", VReg_128>;
592def FLAT_STORE_DWORDX3 : FLAT_Store_Pseudo <"flat_store_dwordx3", VReg_96>;
593
594let SubtargetPredicate = HasD16LoadStore in {
595def FLAT_LOAD_UBYTE_D16     : FLAT_Load_Pseudo <"flat_load_ubyte_d16", VGPR_32, 1>;
596def FLAT_LOAD_UBYTE_D16_HI  : FLAT_Load_Pseudo <"flat_load_ubyte_d16_hi", VGPR_32, 1>;
597def FLAT_LOAD_SBYTE_D16     : FLAT_Load_Pseudo <"flat_load_sbyte_d16", VGPR_32, 1>;
598def FLAT_LOAD_SBYTE_D16_HI  : FLAT_Load_Pseudo <"flat_load_sbyte_d16_hi", VGPR_32, 1>;
599def FLAT_LOAD_SHORT_D16     : FLAT_Load_Pseudo <"flat_load_short_d16", VGPR_32, 1>;
600def FLAT_LOAD_SHORT_D16_HI  : FLAT_Load_Pseudo <"flat_load_short_d16_hi", VGPR_32, 1>;
601
602def FLAT_STORE_BYTE_D16_HI  : FLAT_Store_Pseudo <"flat_store_byte_d16_hi", VGPR_32>;
603def FLAT_STORE_SHORT_D16_HI : FLAT_Store_Pseudo <"flat_store_short_d16_hi", VGPR_32>;
604}
605
606defm FLAT_ATOMIC_CMPSWAP    : FLAT_Atomic_Pseudo <"flat_atomic_cmpswap",
607                                VGPR_32, i32, v2i32, VReg_64>;
608
609defm FLAT_ATOMIC_CMPSWAP_X2 : FLAT_Atomic_Pseudo <"flat_atomic_cmpswap_x2",
610                                VReg_64, i64, v2i64, VReg_128>;
611
612defm FLAT_ATOMIC_SWAP       : FLAT_Atomic_Pseudo <"flat_atomic_swap",
613                                VGPR_32, i32>;
614
615defm FLAT_ATOMIC_SWAP_X2    : FLAT_Atomic_Pseudo <"flat_atomic_swap_x2",
616                                VReg_64, i64>;
617
618defm FLAT_ATOMIC_ADD        : FLAT_Atomic_Pseudo <"flat_atomic_add",
619                                VGPR_32, i32>;
620
621defm FLAT_ATOMIC_SUB        : FLAT_Atomic_Pseudo <"flat_atomic_sub",
622                                VGPR_32, i32>;
623
624defm FLAT_ATOMIC_SMIN       : FLAT_Atomic_Pseudo <"flat_atomic_smin",
625                                VGPR_32, i32>;
626
627defm FLAT_ATOMIC_UMIN       : FLAT_Atomic_Pseudo <"flat_atomic_umin",
628                                VGPR_32, i32>;
629
630defm FLAT_ATOMIC_SMAX       : FLAT_Atomic_Pseudo <"flat_atomic_smax",
631                                VGPR_32, i32>;
632
633defm FLAT_ATOMIC_UMAX       : FLAT_Atomic_Pseudo <"flat_atomic_umax",
634                                VGPR_32, i32>;
635
636defm FLAT_ATOMIC_AND        : FLAT_Atomic_Pseudo <"flat_atomic_and",
637                                VGPR_32, i32>;
638
639defm FLAT_ATOMIC_OR         : FLAT_Atomic_Pseudo <"flat_atomic_or",
640                                VGPR_32, i32>;
641
642defm FLAT_ATOMIC_XOR        : FLAT_Atomic_Pseudo <"flat_atomic_xor",
643                                VGPR_32, i32>;
644
645defm FLAT_ATOMIC_INC        : FLAT_Atomic_Pseudo <"flat_atomic_inc",
646                                VGPR_32, i32>;
647
648defm FLAT_ATOMIC_DEC        : FLAT_Atomic_Pseudo <"flat_atomic_dec",
649                                VGPR_32, i32>;
650
651defm FLAT_ATOMIC_ADD_X2     : FLAT_Atomic_Pseudo <"flat_atomic_add_x2",
652                                VReg_64, i64>;
653
654defm FLAT_ATOMIC_SUB_X2     : FLAT_Atomic_Pseudo <"flat_atomic_sub_x2",
655                                VReg_64, i64>;
656
657defm FLAT_ATOMIC_SMIN_X2    : FLAT_Atomic_Pseudo <"flat_atomic_smin_x2",
658                                VReg_64, i64>;
659
660defm FLAT_ATOMIC_UMIN_X2    : FLAT_Atomic_Pseudo <"flat_atomic_umin_x2",
661                                VReg_64, i64>;
662
663defm FLAT_ATOMIC_SMAX_X2    : FLAT_Atomic_Pseudo <"flat_atomic_smax_x2",
664                                VReg_64, i64>;
665
666defm FLAT_ATOMIC_UMAX_X2    : FLAT_Atomic_Pseudo <"flat_atomic_umax_x2",
667                                VReg_64, i64>;
668
669defm FLAT_ATOMIC_AND_X2     : FLAT_Atomic_Pseudo <"flat_atomic_and_x2",
670                                VReg_64, i64>;
671
672defm FLAT_ATOMIC_OR_X2      : FLAT_Atomic_Pseudo <"flat_atomic_or_x2",
673                                VReg_64, i64>;
674
675defm FLAT_ATOMIC_XOR_X2     : FLAT_Atomic_Pseudo <"flat_atomic_xor_x2",
676                                VReg_64, i64>;
677
678defm FLAT_ATOMIC_INC_X2     : FLAT_Atomic_Pseudo <"flat_atomic_inc_x2",
679                                VReg_64, i64>;
680
681defm FLAT_ATOMIC_DEC_X2     : FLAT_Atomic_Pseudo <"flat_atomic_dec_x2",
682                                VReg_64, i64>;
683
684// GFX7-, GFX10-only flat instructions.
685let SubtargetPredicate = isGFX7GFX10 in {
686
687defm FLAT_ATOMIC_FCMPSWAP    : FLAT_Atomic_Pseudo <"flat_atomic_fcmpswap",
688                                VGPR_32, f32, v2f32, VReg_64>;
689
690defm FLAT_ATOMIC_FCMPSWAP_X2 : FLAT_Atomic_Pseudo <"flat_atomic_fcmpswap_x2",
691                                VReg_64, f64, v2f64, VReg_128>;
692
693defm FLAT_ATOMIC_FMIN        : FLAT_Atomic_Pseudo <"flat_atomic_fmin",
694                                VGPR_32, f32>;
695
696defm FLAT_ATOMIC_FMAX        : FLAT_Atomic_Pseudo <"flat_atomic_fmax",
697                                VGPR_32, f32>;
698
699defm FLAT_ATOMIC_FMIN_X2     : FLAT_Atomic_Pseudo <"flat_atomic_fmin_x2",
700                                VReg_64, f64>;
701
702defm FLAT_ATOMIC_FMAX_X2     : FLAT_Atomic_Pseudo <"flat_atomic_fmax_x2",
703                                VReg_64, f64>;
704
705} // End SubtargetPredicate = isGFX7GFX10
706
707let SubtargetPredicate = isGFX90APlus in {
708  defm FLAT_ATOMIC_ADD_F64   : FLAT_Atomic_Pseudo<"flat_atomic_add_f64", VReg_64, f64>;
709  defm FLAT_ATOMIC_MIN_F64   : FLAT_Atomic_Pseudo<"flat_atomic_min_f64", VReg_64, f64>;
710  defm FLAT_ATOMIC_MAX_F64   : FLAT_Atomic_Pseudo<"flat_atomic_max_f64", VReg_64, f64>;
711  defm GLOBAL_ATOMIC_ADD_F64 : FLAT_Global_Atomic_Pseudo<"global_atomic_add_f64", VReg_64, f64>;
712  defm GLOBAL_ATOMIC_MIN_F64 : FLAT_Global_Atomic_Pseudo<"global_atomic_min_f64", VReg_64, f64>;
713  defm GLOBAL_ATOMIC_MAX_F64 : FLAT_Global_Atomic_Pseudo<"global_atomic_max_f64", VReg_64, f64>;
714} // End SubtargetPredicate = isGFX90APlus
715
716let SubtargetPredicate = isGFX940Plus in {
717  defm FLAT_ATOMIC_ADD_F32       : FLAT_Atomic_Pseudo<"flat_atomic_add_f32",     VGPR_32, f32>;
718  defm FLAT_ATOMIC_PK_ADD_F16    : FLAT_Atomic_Pseudo<"flat_atomic_pk_add_f16",  VGPR_32, v2f16>;
719  defm FLAT_ATOMIC_PK_ADD_BF16   : FLAT_Atomic_Pseudo<"flat_atomic_pk_add_bf16", VGPR_32, v2f16>;
720  defm GLOBAL_ATOMIC_PK_ADD_BF16 : FLAT_Global_Atomic_Pseudo<"global_atomic_pk_add_bf16", VGPR_32, v2f16>;
721} // End SubtargetPredicate = isGFX940Plus
722
723defm GLOBAL_LOAD_UBYTE    : FLAT_Global_Load_Pseudo <"global_load_ubyte", VGPR_32>;
724defm GLOBAL_LOAD_SBYTE    : FLAT_Global_Load_Pseudo <"global_load_sbyte", VGPR_32>;
725defm GLOBAL_LOAD_USHORT   : FLAT_Global_Load_Pseudo <"global_load_ushort", VGPR_32>;
726defm GLOBAL_LOAD_SSHORT   : FLAT_Global_Load_Pseudo <"global_load_sshort", VGPR_32>;
727defm GLOBAL_LOAD_DWORD    : FLAT_Global_Load_Pseudo <"global_load_dword", VGPR_32>;
728defm GLOBAL_LOAD_DWORDX2  : FLAT_Global_Load_Pseudo <"global_load_dwordx2", VReg_64>;
729defm GLOBAL_LOAD_DWORDX3  : FLAT_Global_Load_Pseudo <"global_load_dwordx3", VReg_96>;
730defm GLOBAL_LOAD_DWORDX4  : FLAT_Global_Load_Pseudo <"global_load_dwordx4", VReg_128>;
731
732defm GLOBAL_LOAD_UBYTE_D16    : FLAT_Global_Load_Pseudo <"global_load_ubyte_d16", VGPR_32, 1>;
733defm GLOBAL_LOAD_UBYTE_D16_HI : FLAT_Global_Load_Pseudo <"global_load_ubyte_d16_hi", VGPR_32, 1>;
734defm GLOBAL_LOAD_SBYTE_D16    : FLAT_Global_Load_Pseudo <"global_load_sbyte_d16", VGPR_32, 1>;
735defm GLOBAL_LOAD_SBYTE_D16_HI : FLAT_Global_Load_Pseudo <"global_load_sbyte_d16_hi", VGPR_32, 1>;
736defm GLOBAL_LOAD_SHORT_D16    : FLAT_Global_Load_Pseudo <"global_load_short_d16", VGPR_32, 1>;
737defm GLOBAL_LOAD_SHORT_D16_HI : FLAT_Global_Load_Pseudo <"global_load_short_d16_hi", VGPR_32, 1>;
738let OtherPredicates = [HasGFX10_BEncoding] in
739defm GLOBAL_LOAD_DWORD_ADDTID : FLAT_Global_Load_AddTid_Pseudo <"global_load_dword_addtid", VGPR_32>;
740
741defm GLOBAL_STORE_BYTE    : FLAT_Global_Store_Pseudo <"global_store_byte", VGPR_32>;
742defm GLOBAL_STORE_SHORT   : FLAT_Global_Store_Pseudo <"global_store_short", VGPR_32>;
743defm GLOBAL_STORE_DWORD   : FLAT_Global_Store_Pseudo <"global_store_dword", VGPR_32>;
744defm GLOBAL_STORE_DWORDX2 : FLAT_Global_Store_Pseudo <"global_store_dwordx2", VReg_64>;
745defm GLOBAL_STORE_DWORDX3 : FLAT_Global_Store_Pseudo <"global_store_dwordx3", VReg_96>;
746defm GLOBAL_STORE_DWORDX4 : FLAT_Global_Store_Pseudo <"global_store_dwordx4", VReg_128>;
747let OtherPredicates = [HasGFX10_BEncoding] in
748defm GLOBAL_STORE_DWORD_ADDTID : FLAT_Global_Store_AddTid_Pseudo <"global_store_dword_addtid", VGPR_32>;
749
750defm GLOBAL_STORE_BYTE_D16_HI  : FLAT_Global_Store_Pseudo <"global_store_byte_d16_hi", VGPR_32>;
751defm GLOBAL_STORE_SHORT_D16_HI : FLAT_Global_Store_Pseudo <"global_store_short_d16_hi", VGPR_32>;
752
753let is_flat_global = 1 in {
754defm GLOBAL_ATOMIC_CMPSWAP : FLAT_Global_Atomic_Pseudo <"global_atomic_cmpswap",
755                               VGPR_32, i32, v2i32, VReg_64>;
756
757defm GLOBAL_ATOMIC_CMPSWAP_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_cmpswap_x2",
758                                  VReg_64, i64, v2i64, VReg_128>;
759
760defm GLOBAL_ATOMIC_SWAP : FLAT_Global_Atomic_Pseudo <"global_atomic_swap",
761                             VGPR_32, i32>;
762
763defm GLOBAL_ATOMIC_SWAP_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_swap_x2",
764                                VReg_64, i64>;
765
766defm GLOBAL_ATOMIC_ADD : FLAT_Global_Atomic_Pseudo <"global_atomic_add",
767                           VGPR_32, i32>;
768
769defm GLOBAL_ATOMIC_SUB : FLAT_Global_Atomic_Pseudo <"global_atomic_sub",
770                           VGPR_32, i32>;
771
772defm GLOBAL_ATOMIC_SMIN : FLAT_Global_Atomic_Pseudo <"global_atomic_smin",
773                            VGPR_32, i32>;
774
775defm GLOBAL_ATOMIC_UMIN : FLAT_Global_Atomic_Pseudo <"global_atomic_umin",
776                            VGPR_32, i32>;
777
778defm GLOBAL_ATOMIC_SMAX : FLAT_Global_Atomic_Pseudo <"global_atomic_smax",
779                            VGPR_32, i32>;
780
781defm GLOBAL_ATOMIC_UMAX : FLAT_Global_Atomic_Pseudo <"global_atomic_umax",
782                            VGPR_32, i32>;
783
784defm GLOBAL_ATOMIC_AND : FLAT_Global_Atomic_Pseudo <"global_atomic_and",
785                           VGPR_32, i32>;
786
787defm GLOBAL_ATOMIC_OR : FLAT_Global_Atomic_Pseudo <"global_atomic_or",
788                          VGPR_32, i32>;
789
790defm GLOBAL_ATOMIC_XOR : FLAT_Global_Atomic_Pseudo <"global_atomic_xor",
791                           VGPR_32, i32>;
792
793defm GLOBAL_ATOMIC_INC : FLAT_Global_Atomic_Pseudo <"global_atomic_inc",
794                           VGPR_32, i32>;
795
796defm GLOBAL_ATOMIC_DEC : FLAT_Global_Atomic_Pseudo <"global_atomic_dec",
797                           VGPR_32, i32>;
798
799defm GLOBAL_ATOMIC_ADD_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_add_x2",
800                              VReg_64, i64>;
801
802defm GLOBAL_ATOMIC_SUB_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_sub_x2",
803                              VReg_64, i64>;
804
805defm GLOBAL_ATOMIC_SMIN_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_smin_x2",
806                               VReg_64, i64>;
807
808defm GLOBAL_ATOMIC_UMIN_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_umin_x2",
809                               VReg_64, i64>;
810
811defm GLOBAL_ATOMIC_SMAX_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_smax_x2",
812                               VReg_64, i64>;
813
814defm GLOBAL_ATOMIC_UMAX_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_umax_x2",
815                               VReg_64, i64>;
816
817defm GLOBAL_ATOMIC_AND_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_and_x2",
818                              VReg_64, i64>;
819
820defm GLOBAL_ATOMIC_OR_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_or_x2",
821                             VReg_64, i64>;
822
823defm GLOBAL_ATOMIC_XOR_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_xor_x2",
824                              VReg_64, i64>;
825
826defm GLOBAL_ATOMIC_INC_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_inc_x2",
827                              VReg_64, i64>;
828
829defm GLOBAL_ATOMIC_DEC_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_dec_x2",
830                              VReg_64, i64>;
831
832let SubtargetPredicate = HasGFX10_BEncoding in
833defm GLOBAL_ATOMIC_CSUB : FLAT_Global_Atomic_Pseudo_RTN <"global_atomic_csub",
834                              VGPR_32, i32>;
835
836let SubtargetPredicate = isGFX940Plus in {
837
838defm GLOBAL_LOAD_LDS_UBYTE  : FLAT_Global_Load_LDS_Pseudo <"global_load_lds_ubyte">;
839defm GLOBAL_LOAD_LDS_SBYTE  : FLAT_Global_Load_LDS_Pseudo <"global_load_lds_sbyte">;
840defm GLOBAL_LOAD_LDS_USHORT : FLAT_Global_Load_LDS_Pseudo <"global_load_lds_ushort">;
841defm GLOBAL_LOAD_LDS_SSHORT : FLAT_Global_Load_LDS_Pseudo <"global_load_lds_sshort">;
842defm GLOBAL_LOAD_LDS_DWORD  : FLAT_Global_Load_LDS_Pseudo <"global_load_lds_dword">;
843
844} // End let SubtargetPredicate = isGFX940Plus
845} // End is_flat_global = 1
846
847
848
849let SubtargetPredicate = HasFlatScratchInsts in {
850defm SCRATCH_LOAD_UBYTE    : FLAT_Scratch_Load_Pseudo <"scratch_load_ubyte", VGPR_32>;
851defm SCRATCH_LOAD_SBYTE    : FLAT_Scratch_Load_Pseudo <"scratch_load_sbyte", VGPR_32>;
852defm SCRATCH_LOAD_USHORT   : FLAT_Scratch_Load_Pseudo <"scratch_load_ushort", VGPR_32>;
853defm SCRATCH_LOAD_SSHORT   : FLAT_Scratch_Load_Pseudo <"scratch_load_sshort", VGPR_32>;
854defm SCRATCH_LOAD_DWORD    : FLAT_Scratch_Load_Pseudo <"scratch_load_dword", VGPR_32>;
855defm SCRATCH_LOAD_DWORDX2  : FLAT_Scratch_Load_Pseudo <"scratch_load_dwordx2", VReg_64>;
856defm SCRATCH_LOAD_DWORDX3  : FLAT_Scratch_Load_Pseudo <"scratch_load_dwordx3", VReg_96>;
857defm SCRATCH_LOAD_DWORDX4  : FLAT_Scratch_Load_Pseudo <"scratch_load_dwordx4", VReg_128>;
858
859defm SCRATCH_LOAD_UBYTE_D16    : FLAT_Scratch_Load_Pseudo <"scratch_load_ubyte_d16", VGPR_32, 1>;
860defm SCRATCH_LOAD_UBYTE_D16_HI : FLAT_Scratch_Load_Pseudo <"scratch_load_ubyte_d16_hi", VGPR_32, 1>;
861defm SCRATCH_LOAD_SBYTE_D16    : FLAT_Scratch_Load_Pseudo <"scratch_load_sbyte_d16", VGPR_32, 1>;
862defm SCRATCH_LOAD_SBYTE_D16_HI : FLAT_Scratch_Load_Pseudo <"scratch_load_sbyte_d16_hi", VGPR_32, 1>;
863defm SCRATCH_LOAD_SHORT_D16    : FLAT_Scratch_Load_Pseudo <"scratch_load_short_d16", VGPR_32, 1>;
864defm SCRATCH_LOAD_SHORT_D16_HI : FLAT_Scratch_Load_Pseudo <"scratch_load_short_d16_hi", VGPR_32, 1>;
865
866defm SCRATCH_STORE_BYTE    : FLAT_Scratch_Store_Pseudo <"scratch_store_byte", VGPR_32>;
867defm SCRATCH_STORE_SHORT   : FLAT_Scratch_Store_Pseudo <"scratch_store_short", VGPR_32>;
868defm SCRATCH_STORE_DWORD   : FLAT_Scratch_Store_Pseudo <"scratch_store_dword", VGPR_32>;
869defm SCRATCH_STORE_DWORDX2 : FLAT_Scratch_Store_Pseudo <"scratch_store_dwordx2", VReg_64>;
870defm SCRATCH_STORE_DWORDX3 : FLAT_Scratch_Store_Pseudo <"scratch_store_dwordx3", VReg_96>;
871defm SCRATCH_STORE_DWORDX4 : FLAT_Scratch_Store_Pseudo <"scratch_store_dwordx4", VReg_128>;
872
873defm SCRATCH_STORE_BYTE_D16_HI : FLAT_Scratch_Store_Pseudo <"scratch_store_byte_d16_hi", VGPR_32>;
874defm SCRATCH_STORE_SHORT_D16_HI : FLAT_Scratch_Store_Pseudo <"scratch_store_short_d16_hi", VGPR_32>;
875
876let SubtargetPredicate = isGFX940Plus in {
877
878defm SCRATCH_LOAD_LDS_UBYTE  : FLAT_Scratch_Load_LDS_Pseudo <"scratch_load_lds_ubyte">;
879defm SCRATCH_LOAD_LDS_SBYTE  : FLAT_Scratch_Load_LDS_Pseudo <"scratch_load_lds_sbyte">;
880defm SCRATCH_LOAD_LDS_USHORT : FLAT_Scratch_Load_LDS_Pseudo <"scratch_load_lds_ushort">;
881defm SCRATCH_LOAD_LDS_SSHORT : FLAT_Scratch_Load_LDS_Pseudo <"scratch_load_lds_sshort">;
882defm SCRATCH_LOAD_LDS_DWORD  : FLAT_Scratch_Load_LDS_Pseudo <"scratch_load_lds_dword">;
883
884} // End let SubtargetPredicate = isGFX940Plus
885
886} // End SubtargetPredicate = HasFlatScratchInsts
887
888let SubtargetPredicate = isGFX10Plus, is_flat_global = 1 in {
889  defm GLOBAL_ATOMIC_FCMPSWAP :
890    FLAT_Global_Atomic_Pseudo<"global_atomic_fcmpswap", VGPR_32, f32, v2f32, VReg_64>;
891  defm GLOBAL_ATOMIC_FMIN :
892    FLAT_Global_Atomic_Pseudo<"global_atomic_fmin", VGPR_32, f32>;
893  defm GLOBAL_ATOMIC_FMAX :
894    FLAT_Global_Atomic_Pseudo<"global_atomic_fmax", VGPR_32, f32>;
895  defm GLOBAL_ATOMIC_FCMPSWAP_X2 :
896    FLAT_Global_Atomic_Pseudo<"global_atomic_fcmpswap_x2", VReg_64, f64, v2f64, VReg_128>;
897  defm GLOBAL_ATOMIC_FMIN_X2 :
898    FLAT_Global_Atomic_Pseudo<"global_atomic_fmin_x2", VReg_64, f64>;
899  defm GLOBAL_ATOMIC_FMAX_X2 :
900    FLAT_Global_Atomic_Pseudo<"global_atomic_fmax_x2", VReg_64, f64>;
901} // End SubtargetPredicate = isGFX10Plus, is_flat_global = 1
902
903let is_flat_global = 1 in {
904let OtherPredicates = [HasAtomicFaddNoRtnInsts] in
905  defm GLOBAL_ATOMIC_ADD_F32 : FLAT_Global_Atomic_Pseudo_NO_RTN <
906    "global_atomic_add_f32", VGPR_32, f32
907  >;
908let OtherPredicates = [HasAtomicPkFaddNoRtnInsts] in
909  defm GLOBAL_ATOMIC_PK_ADD_F16 : FLAT_Global_Atomic_Pseudo_NO_RTN <
910    "global_atomic_pk_add_f16", VGPR_32, v2f16
911  >;
912let OtherPredicates = [HasAtomicFaddRtnInsts] in
913  defm GLOBAL_ATOMIC_ADD_F32 : FLAT_Global_Atomic_Pseudo_RTN <
914    "global_atomic_add_f32", VGPR_32, f32
915  >;
916let OtherPredicates = [isGFX90APlus] in
917  defm GLOBAL_ATOMIC_PK_ADD_F16 : FLAT_Global_Atomic_Pseudo_RTN <
918    "global_atomic_pk_add_f16", VGPR_32, v2f16
919  >;
920} // End is_flat_global = 1
921
922//===----------------------------------------------------------------------===//
923// Flat Patterns
924//===----------------------------------------------------------------------===//
925
926// Patterns for global loads with no offset.
927class FlatLoadPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
928  (vt (node (FlatOffset i64:$vaddr, i16:$offset))),
929  (inst $vaddr, $offset)
930>;
931
932class FlatLoadPat_D16 <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
933  (node (FlatOffset (i64 VReg_64:$vaddr), i16:$offset), vt:$in),
934  (inst $vaddr, $offset, 0, $in)
935>;
936
937class FlatSignedLoadPat_D16 <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
938  (node (GlobalOffset (i64 VReg_64:$vaddr), i16:$offset), vt:$in),
939  (inst $vaddr, $offset, 0, $in)
940>;
941
942class GlobalLoadSaddrPat_D16 <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
943  (vt (node (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i16:$offset), vt:$in)),
944  (inst $saddr, $voffset, $offset, 0, $in)
945>;
946
947class FlatLoadSignedPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
948  (vt (node (GlobalOffset (i64 VReg_64:$vaddr), i16:$offset))),
949  (inst $vaddr, $offset)
950>;
951
952class GlobalLoadSaddrPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
953  (vt (node (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i16:$offset))),
954  (inst $saddr, $voffset, $offset, 0)
955>;
956
957class GlobalStoreSaddrPat <FLAT_Pseudo inst, SDPatternOperator node,
958                           ValueType vt> : GCNPat <
959  (node vt:$data, (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i16:$offset)),
960  (inst $voffset, getVregSrcForVT<vt>.ret:$data, $saddr, $offset)
961>;
962
963class GlobalAtomicStoreSaddrPat <FLAT_Pseudo inst, SDPatternOperator node,
964                                 ValueType vt> : GCNPat <
965  (node (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i16:$offset), vt:$data),
966  (inst $voffset, getVregSrcForVT<vt>.ret:$data, $saddr, $offset)
967>;
968
969class GlobalAtomicSaddrPat <FLAT_Pseudo inst, SDPatternOperator node,
970                            ValueType vt, ValueType data_vt = vt> : GCNPat <
971  (vt (node (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i16:$offset), data_vt:$data)),
972  (inst $voffset, getVregSrcForVT<data_vt>.ret:$data, $saddr, $offset)
973>;
974
975class GlobalAtomicNoRtnSaddrPat <FLAT_Pseudo inst, SDPatternOperator node,
976                                 ValueType vt> : GCNPat <
977  (node (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i16:$offset), vt:$data),
978  (inst $voffset, getVregSrcForVT<vt>.ret:$data, $saddr, $offset)
979>;
980
981class FlatStorePat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
982  (node vt:$data, (FlatOffset i64:$vaddr, i16:$offset)),
983  (inst $vaddr, getVregSrcForVT<vt>.ret:$data, $offset)
984>;
985
986class FlatStoreSignedPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
987  (node vt:$data, (GlobalOffset i64:$vaddr, i16:$offset)),
988  (inst $vaddr, getVregSrcForVT<vt>.ret:$data, $offset)
989>;
990
991class FlatStoreAtomicPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
992  // atomic store follows atomic binop convention so the address comes
993  // first.
994  (node (FlatOffset i64:$vaddr, i16:$offset), vt:$data),
995  (inst $vaddr, getVregSrcForVT<vt>.ret:$data, $offset)
996>;
997
998class FlatStoreSignedAtomicPat <FLAT_Pseudo inst, SDPatternOperator node,
999                                ValueType vt, ValueType data_vt = vt> : GCNPat <
1000  // atomic store follows atomic binop convention so the address comes
1001  // first.
1002  (node (GlobalOffset i64:$vaddr, i16:$offset), data_vt:$data),
1003  (inst $vaddr, getVregSrcForVT<data_vt>.ret:$data, $offset)
1004>;
1005
1006class FlatAtomicPatNoRtn <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
1007  (node (FlatOffset i64:$vaddr, i16:$offset), vt:$data),
1008  (inst VReg_64:$vaddr, getVregSrcForVT<vt>.ret:$data, $offset)
1009>;
1010
1011multiclass FlatAtomicPat <string inst, string node, ValueType vt,
1012                          ValueType data_vt = vt> {
1013  defvar rtnNode = !cast<PatFrags>(node#"_ret_"#vt.Size);
1014  defvar noRtnNode = !cast<PatFrags>(node#"_noret_"#vt.Size);
1015
1016  def : GCNPat <(vt (rtnNode (FlatOffset i64:$vaddr, i16:$offset), data_vt:$data)),
1017    (!cast<FLAT_Pseudo>(inst#"_RTN") VReg_64:$vaddr, getVregSrcForVT<data_vt>.ret:$data, $offset)>;
1018
1019  def : GCNPat <(vt (noRtnNode (FlatOffset i64:$vaddr, i16:$offset), data_vt:$data)),
1020    (!cast<FLAT_Pseudo>(inst) VReg_64:$vaddr, getVregSrcForVT<data_vt>.ret:$data, $offset)>;
1021}
1022
1023multiclass FlatSignedAtomicPat <string inst, string node, ValueType vt,
1024                                ValueType data_vt = vt, bit isIntr = 0> {
1025  defvar rtnNode = !cast<PatFrags>(node # "_ret" # !if(isIntr, "", "_" # vt.Size));
1026  defvar noRtnNode = !cast<PatFrags>(node # "_noret" # !if(isIntr, "", "_" # vt.Size));
1027
1028  def : GCNPat <(vt (rtnNode (GlobalOffset i64:$vaddr, i16:$offset), data_vt:$data)),
1029    (!cast<FLAT_Pseudo>(inst#"_RTN") VReg_64:$vaddr, getVregSrcForVT<data_vt>.ret:$data, $offset)>;
1030
1031  def : GCNPat <(vt (noRtnNode (GlobalOffset i64:$vaddr, i16:$offset), data_vt:$data)),
1032    (!cast<FLAT_Pseudo>(inst) VReg_64:$vaddr, getVregSrcForVT<data_vt>.ret:$data, $offset)>;
1033}
1034
1035multiclass FlatSignedAtomicIntrPat <string inst, string node, ValueType vt,
1036                                    ValueType data_vt = vt> {
1037  defm : FlatSignedAtomicPat<inst, node, vt, data_vt, /* isIntr */ 1>;
1038}
1039
1040class FlatSignedAtomicPatNoRtn <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
1041  (node (GlobalOffset i64:$vaddr, i16:$offset), vt:$data),
1042  (inst VReg_64:$vaddr, getVregSrcForVT<vt>.ret:$data, $offset)
1043>;
1044
1045class FlatSignedAtomicPatRtn <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt,
1046                              ValueType data_vt = vt> : GCNPat <
1047  (vt (node (GlobalOffset i64:$vaddr, i16:$offset), data_vt:$data)),
1048  (inst VReg_64:$vaddr, getVregSrcForVT<data_vt>.ret:$data, $offset)
1049>;
1050
1051class ScratchLoadSignedPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
1052  (vt (node (ScratchOffset (i32 VGPR_32:$vaddr), i16:$offset))),
1053  (inst $vaddr, $offset)
1054>;
1055
1056class ScratchLoadSignedPat_D16 <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
1057  (node (ScratchOffset (i32 VGPR_32:$vaddr), i16:$offset), vt:$in),
1058  (inst $vaddr, $offset, 0, $in)
1059>;
1060
1061class ScratchStoreSignedPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
1062  (node vt:$data, (ScratchOffset (i32 VGPR_32:$vaddr), i16:$offset)),
1063  (inst getVregSrcForVT<vt>.ret:$data, $vaddr, $offset)
1064>;
1065
1066class ScratchLoadSaddrPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
1067  (vt (node (ScratchSAddr (i32 SGPR_32:$saddr), i16:$offset))),
1068  (inst $saddr, $offset)
1069>;
1070
1071class ScratchLoadSaddrPat_D16 <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
1072  (vt (node (ScratchSAddr (i32 SGPR_32:$saddr), i16:$offset), vt:$in)),
1073  (inst $saddr, $offset, 0, $in)
1074>;
1075
1076class ScratchStoreSaddrPat <FLAT_Pseudo inst, SDPatternOperator node,
1077                            ValueType vt> : GCNPat <
1078  (node vt:$data, (ScratchSAddr (i32 SGPR_32:$saddr), i16:$offset)),
1079  (inst getVregSrcForVT<vt>.ret:$data, $saddr, $offset)
1080>;
1081
1082class ScratchLoadSVaddrPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
1083  (vt (node (ScratchSVAddr (i32 VGPR_32:$vaddr), (i32 SGPR_32:$saddr), i16:$offset))),
1084  (inst $vaddr, $saddr, $offset, 0)
1085>;
1086
1087class ScratchStoreSVaddrPat <FLAT_Pseudo inst, SDPatternOperator node,
1088                             ValueType vt> : GCNPat <
1089  (node vt:$data, (ScratchSVAddr (i32 VGPR_32:$vaddr), (i32 SGPR_32:$saddr), i16:$offset)),
1090  (inst getVregSrcForVT<vt>.ret:$data, $vaddr, $saddr, $offset)
1091>;
1092
1093class ScratchLoadSVaddrPat_D16 <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
1094  (vt (node (ScratchSVAddr (i32 VGPR_32:$vaddr), (i32 SGPR_32:$saddr), i16:$offset), vt:$in)),
1095  (inst $vaddr, $saddr, $offset, 0, $in)
1096>;
1097
1098let OtherPredicates = [HasFlatAddressSpace] in {
1099
1100def : FlatLoadPat <FLAT_LOAD_UBYTE, atomic_load_8_flat, i32>;
1101def : FlatLoadPat <FLAT_LOAD_UBYTE, atomic_load_8_flat, i16>;
1102def : FlatLoadPat <FLAT_LOAD_USHORT, atomic_load_16_flat, i32>;
1103def : FlatLoadPat <FLAT_LOAD_USHORT, atomic_load_16_flat, i16>;
1104def : FlatLoadPat <FLAT_LOAD_UBYTE, extloadi8_flat, i32>;
1105def : FlatLoadPat <FLAT_LOAD_UBYTE, zextloadi8_flat, i32>;
1106def : FlatLoadPat <FLAT_LOAD_SBYTE, sextloadi8_flat, i32>;
1107def : FlatLoadPat <FLAT_LOAD_UBYTE, extloadi8_flat, i16>;
1108def : FlatLoadPat <FLAT_LOAD_UBYTE, zextloadi8_flat, i16>;
1109def : FlatLoadPat <FLAT_LOAD_SBYTE, sextloadi8_flat, i16>;
1110def : FlatLoadPat <FLAT_LOAD_USHORT, extloadi16_flat, i32>;
1111def : FlatLoadPat <FLAT_LOAD_USHORT, zextloadi16_flat, i32>;
1112def : FlatLoadPat <FLAT_LOAD_USHORT, load_flat, i16>;
1113def : FlatLoadPat <FLAT_LOAD_SSHORT, sextloadi16_flat, i32>;
1114def : FlatLoadPat <FLAT_LOAD_DWORDX3, load_flat, v3i32>;
1115
1116def : FlatLoadPat <FLAT_LOAD_DWORD, atomic_load_32_flat, i32>;
1117def : FlatLoadPat <FLAT_LOAD_DWORDX2, atomic_load_64_flat, i64>;
1118
1119def : FlatStorePat <FLAT_STORE_BYTE, truncstorei8_flat, i32>;
1120def : FlatStorePat <FLAT_STORE_SHORT, truncstorei16_flat, i32>;
1121
1122foreach vt = Reg32Types.types in {
1123def : FlatLoadPat <FLAT_LOAD_DWORD, load_flat, vt>;
1124def : FlatStorePat <FLAT_STORE_DWORD, store_flat, vt>;
1125}
1126
1127foreach vt = VReg_64.RegTypes in {
1128def : FlatStorePat <FLAT_STORE_DWORDX2, store_flat, vt>;
1129def : FlatLoadPat <FLAT_LOAD_DWORDX2, load_flat, vt>;
1130}
1131
1132def : FlatStorePat <FLAT_STORE_DWORDX3, store_flat, v3i32>;
1133
1134foreach vt = VReg_128.RegTypes in {
1135def : FlatLoadPat <FLAT_LOAD_DWORDX4, load_flat, vt>;
1136def : FlatStorePat <FLAT_STORE_DWORDX4, store_flat, vt>;
1137}
1138
1139def : FlatStoreAtomicPat <FLAT_STORE_DWORD, atomic_store_32_flat, i32>;
1140def : FlatStoreAtomicPat <FLAT_STORE_DWORDX2, atomic_store_64_flat, i64>;
1141def : FlatStoreAtomicPat <FLAT_STORE_BYTE, atomic_store_8_flat, i32>;
1142def : FlatStoreAtomicPat <FLAT_STORE_BYTE, atomic_store_8_flat, i16>;
1143def : FlatStoreAtomicPat <FLAT_STORE_SHORT, atomic_store_16_flat, i32>;
1144def : FlatStoreAtomicPat <FLAT_STORE_SHORT, atomic_store_16_flat, i16>;
1145
1146foreach as = [ "flat", "global" ] in {
1147defm : FlatAtomicPat <"FLAT_ATOMIC_ADD", "atomic_load_add_"#as, i32>;
1148defm : FlatAtomicPat <"FLAT_ATOMIC_SUB", "atomic_load_sub_"#as, i32>;
1149defm : FlatAtomicPat <"FLAT_ATOMIC_INC", "atomic_inc_"#as, i32>;
1150defm : FlatAtomicPat <"FLAT_ATOMIC_DEC", "atomic_dec_"#as, i32>;
1151defm : FlatAtomicPat <"FLAT_ATOMIC_AND", "atomic_load_and_"#as, i32>;
1152defm : FlatAtomicPat <"FLAT_ATOMIC_SMAX", "atomic_load_max_"#as, i32>;
1153defm : FlatAtomicPat <"FLAT_ATOMIC_UMAX", "atomic_load_umax_"#as, i32>;
1154defm : FlatAtomicPat <"FLAT_ATOMIC_SMIN", "atomic_load_min_"#as, i32>;
1155defm : FlatAtomicPat <"FLAT_ATOMIC_UMIN", "atomic_load_umin_"#as, i32>;
1156defm : FlatAtomicPat <"FLAT_ATOMIC_OR", "atomic_load_or_"#as, i32>;
1157defm : FlatAtomicPat <"FLAT_ATOMIC_SWAP", "atomic_swap_"#as, i32>;
1158defm : FlatAtomicPat <"FLAT_ATOMIC_CMPSWAP", "AMDGPUatomic_cmp_swap_"#as, i32, v2i32>;
1159defm : FlatAtomicPat <"FLAT_ATOMIC_XOR", "atomic_load_xor_"#as, i32>;
1160
1161defm : FlatAtomicPat <"FLAT_ATOMIC_ADD_X2", "atomic_load_add_"#as, i64>;
1162defm : FlatAtomicPat <"FLAT_ATOMIC_SUB_X2", "atomic_load_sub_"#as, i64>;
1163defm : FlatAtomicPat <"FLAT_ATOMIC_INC_X2", "atomic_inc_"#as, i64>;
1164defm : FlatAtomicPat <"FLAT_ATOMIC_DEC_X2", "atomic_dec_"#as, i64>;
1165defm : FlatAtomicPat <"FLAT_ATOMIC_AND_X2", "atomic_load_and_"#as, i64>;
1166defm : FlatAtomicPat <"FLAT_ATOMIC_SMAX_X2", "atomic_load_max_"#as, i64>;
1167defm : FlatAtomicPat <"FLAT_ATOMIC_UMAX_X2", "atomic_load_umax_"#as, i64>;
1168defm : FlatAtomicPat <"FLAT_ATOMIC_SMIN_X2", "atomic_load_min_"#as, i64>;
1169defm : FlatAtomicPat <"FLAT_ATOMIC_UMIN_X2", "atomic_load_umin_"#as, i64>;
1170defm : FlatAtomicPat <"FLAT_ATOMIC_OR_X2", "atomic_load_or_"#as, i64>;
1171defm : FlatAtomicPat <"FLAT_ATOMIC_SWAP_X2", "atomic_swap_"#as, i64>;
1172defm : FlatAtomicPat <"FLAT_ATOMIC_CMPSWAP_X2", "AMDGPUatomic_cmp_swap_"#as, i64, v2i64>;
1173defm : FlatAtomicPat <"FLAT_ATOMIC_XOR_X2", "atomic_load_xor_"#as, i64>;
1174} // end foreach as
1175
1176def : FlatStorePat <FLAT_STORE_BYTE, truncstorei8_flat, i16>;
1177def : FlatStorePat <FLAT_STORE_SHORT, store_flat, i16>;
1178
1179let OtherPredicates = [HasD16LoadStore] in {
1180def : FlatStorePat <FLAT_STORE_SHORT_D16_HI, truncstorei16_hi16_flat, i32>;
1181def : FlatStorePat <FLAT_STORE_BYTE_D16_HI, truncstorei8_hi16_flat, i32>;
1182}
1183
1184let OtherPredicates = [D16PreservesUnusedBits] in {
1185def : FlatLoadPat_D16 <FLAT_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_flat, v2i16>;
1186def : FlatLoadPat_D16 <FLAT_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_flat, v2f16>;
1187def : FlatLoadPat_D16 <FLAT_LOAD_SBYTE_D16_HI, sextloadi8_d16_hi_flat, v2i16>;
1188def : FlatLoadPat_D16 <FLAT_LOAD_SBYTE_D16_HI, sextloadi8_d16_hi_flat, v2f16>;
1189def : FlatLoadPat_D16 <FLAT_LOAD_SHORT_D16_HI, load_d16_hi_flat, v2i16>;
1190def : FlatLoadPat_D16 <FLAT_LOAD_SHORT_D16_HI, load_d16_hi_flat, v2f16>;
1191
1192def : FlatLoadPat_D16 <FLAT_LOAD_UBYTE_D16, az_extloadi8_d16_lo_flat, v2i16>;
1193def : FlatLoadPat_D16 <FLAT_LOAD_UBYTE_D16, az_extloadi8_d16_lo_flat, v2f16>;
1194def : FlatLoadPat_D16 <FLAT_LOAD_SBYTE_D16, sextloadi8_d16_lo_flat, v2i16>;
1195def : FlatLoadPat_D16 <FLAT_LOAD_SBYTE_D16, sextloadi8_d16_lo_flat, v2f16>;
1196def : FlatLoadPat_D16 <FLAT_LOAD_SHORT_D16, load_d16_lo_flat, v2i16>;
1197def : FlatLoadPat_D16 <FLAT_LOAD_SHORT_D16, load_d16_lo_flat, v2f16>;
1198}
1199
1200} // End OtherPredicates = [HasFlatAddressSpace]
1201
1202
1203multiclass GlobalFLATLoadPats<FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> {
1204  def : FlatLoadSignedPat <inst, node, vt> {
1205    let AddedComplexity = 10;
1206  }
1207
1208  def : GlobalLoadSaddrPat<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SADDR"), node, vt> {
1209    let AddedComplexity = 11;
1210  }
1211}
1212
1213multiclass GlobalFLATLoadPats_D16<FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> {
1214  def : FlatSignedLoadPat_D16 <inst, node, vt> {
1215    let AddedComplexity = 10;
1216  }
1217
1218  def : GlobalLoadSaddrPat_D16<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SADDR"), node, vt> {
1219    let AddedComplexity = 11;
1220  }
1221}
1222
1223multiclass GlobalFLATStorePats<FLAT_Pseudo inst, SDPatternOperator node,
1224                               ValueType vt> {
1225  def : FlatStoreSignedPat <inst, node, vt> {
1226    let AddedComplexity = 10;
1227  }
1228
1229  def : GlobalStoreSaddrPat<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SADDR"), node, vt> {
1230    let AddedComplexity = 11;
1231  }
1232}
1233
1234// Deal with swapped operands for atomic_store vs. regular store
1235multiclass GlobalFLATAtomicStorePats<FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> {
1236  def : FlatStoreSignedAtomicPat <inst, node, vt> {
1237    let AddedComplexity = 10;
1238  }
1239
1240  def : GlobalAtomicStoreSaddrPat<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SADDR"), node, vt> {
1241    let AddedComplexity = 11;
1242  }
1243}
1244
1245multiclass GlobalFLATAtomicPatsRtn<string nortn_inst_name, SDPatternOperator node,
1246                                   ValueType vt, ValueType data_vt = vt> {
1247  def : FlatSignedAtomicPatRtn <!cast<FLAT_Pseudo>(nortn_inst_name#"_RTN"), node, vt, data_vt> {
1248    let AddedComplexity = 10;
1249  }
1250
1251  def : GlobalAtomicSaddrPat<!cast<FLAT_Pseudo>(nortn_inst_name#"_SADDR_RTN"), node, vt, data_vt> {
1252    let AddedComplexity = 11;
1253  }
1254}
1255
1256multiclass GlobalFLATAtomicPats<string inst, string node, ValueType vt,
1257                                ValueType data_vt = vt, bit isIntr = 0> {
1258  defvar rtnNode = !cast<PatFrags>(node # "_ret" # !if(isIntr, "", "_" # vt.Size));
1259  defvar noRtnNode = !cast<PatFrags>(node # "_noret" # !if(isIntr, "", "_" # vt.Size));
1260
1261  let AddedComplexity = 10 in {
1262    defm : FlatSignedAtomicPat <inst, node, vt, data_vt, isIntr>;
1263  }
1264
1265  let AddedComplexity = 11 in {
1266    def : GlobalAtomicSaddrPat<!cast<FLAT_Pseudo>(inst#"_SADDR"), noRtnNode, vt, data_vt>;
1267    def : GlobalAtomicSaddrPat<!cast<FLAT_Pseudo>(inst#"_SADDR_RTN"), rtnNode, vt, data_vt>;
1268  }
1269}
1270
1271multiclass GlobalFLATAtomicIntrPats<string inst, string node, ValueType vt,
1272                                    ValueType data_vt = vt> {
1273  defm : GlobalFLATAtomicPats<inst, node, vt, data_vt, /* isIntr */ 1>;
1274}
1275
1276multiclass GlobalFLATNoRtnAtomicPats<FLAT_Pseudo inst, SDPatternOperator node,
1277                                     ValueType vt> {
1278  def : FlatSignedAtomicPatNoRtn <inst, node, vt> {
1279    let AddedComplexity = 10;
1280  }
1281
1282  def : GlobalAtomicNoRtnSaddrPat<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SADDR"), node, vt> {
1283    let AddedComplexity = 11;
1284  }
1285}
1286
1287multiclass ScratchFLATLoadPats<FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> {
1288  def : ScratchLoadSignedPat <inst, node, vt> {
1289    let AddedComplexity = 25;
1290  }
1291
1292  def : ScratchLoadSaddrPat<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SADDR"), node, vt> {
1293    let AddedComplexity = 26;
1294  }
1295
1296  def : ScratchLoadSVaddrPat<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SVS"), node, vt> {
1297    let SubtargetPredicate = HasFlatScratchSVSMode;
1298    let AddedComplexity = 27;
1299  }
1300}
1301
1302multiclass ScratchFLATStorePats<FLAT_Pseudo inst, SDPatternOperator node,
1303                               ValueType vt> {
1304  def : ScratchStoreSignedPat <inst, node, vt> {
1305    let AddedComplexity = 25;
1306  }
1307
1308  def : ScratchStoreSaddrPat<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SADDR"), node, vt> {
1309    let AddedComplexity = 26;
1310  }
1311
1312  def : ScratchStoreSVaddrPat<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SVS"), node, vt> {
1313    let SubtargetPredicate = HasFlatScratchSVSMode;
1314    let AddedComplexity = 27;
1315  }
1316}
1317
1318multiclass ScratchFLATLoadPats_D16<FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> {
1319  def : ScratchLoadSignedPat_D16 <inst, node, vt> {
1320    let AddedComplexity = 25;
1321  }
1322
1323  def : ScratchLoadSaddrPat_D16<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SADDR"), node, vt> {
1324    let AddedComplexity = 26;
1325  }
1326
1327  def : ScratchLoadSVaddrPat_D16 <!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SVS"), node, vt> {
1328    let SubtargetPredicate = HasFlatScratchSVSMode;
1329    let AddedComplexity = 27;
1330  }
1331}
1332
1333let OtherPredicates = [HasFlatGlobalInsts] in {
1334
1335defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, atomic_load_8_global, i32>;
1336defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, atomic_load_8_global, i16>;
1337defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, atomic_load_16_global, i32>;
1338defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, atomic_load_16_global, i16>;
1339defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, extloadi8_global, i32>;
1340defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, zextloadi8_global, i32>;
1341defm : GlobalFLATLoadPats <GLOBAL_LOAD_SBYTE, sextloadi8_global, i32>;
1342defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, extloadi8_global, i16>;
1343defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, zextloadi8_global, i16>;
1344defm : GlobalFLATLoadPats <GLOBAL_LOAD_SBYTE, sextloadi8_global, i16>;
1345defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, extloadi16_global, i32>;
1346defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, zextloadi16_global, i32>;
1347defm : GlobalFLATLoadPats <GLOBAL_LOAD_SSHORT, sextloadi16_global, i32>;
1348defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, load_global, i16>;
1349
1350foreach vt = Reg32Types.types in {
1351defm : GlobalFLATLoadPats <GLOBAL_LOAD_DWORD, load_global, vt>;
1352defm : GlobalFLATStorePats <GLOBAL_STORE_DWORD, store_global, vt>;
1353}
1354
1355foreach vt = VReg_64.RegTypes in {
1356defm : GlobalFLATLoadPats <GLOBAL_LOAD_DWORDX2, load_global, vt>;
1357defm : GlobalFLATStorePats <GLOBAL_STORE_DWORDX2, store_global, vt>;
1358}
1359
1360defm : GlobalFLATLoadPats <GLOBAL_LOAD_DWORDX3, load_global, v3i32>;
1361
1362foreach vt = VReg_128.RegTypes in {
1363defm : GlobalFLATLoadPats <GLOBAL_LOAD_DWORDX4, load_global, vt>;
1364defm : GlobalFLATStorePats <GLOBAL_STORE_DWORDX4, store_global, vt>;
1365}
1366
1367// There is no distinction for atomic load lowering during selection;
1368// the memory legalizer will set the cache bits and insert the
1369// appropriate waits.
1370defm : GlobalFLATLoadPats <GLOBAL_LOAD_DWORD, atomic_load_32_global, i32>;
1371defm : GlobalFLATLoadPats <GLOBAL_LOAD_DWORDX2, atomic_load_64_global, i64>;
1372
1373defm : GlobalFLATStorePats <GLOBAL_STORE_BYTE, truncstorei8_global, i32>;
1374defm : GlobalFLATStorePats <GLOBAL_STORE_BYTE, truncstorei8_global, i16>;
1375defm : GlobalFLATStorePats <GLOBAL_STORE_SHORT, truncstorei16_global, i32>;
1376defm : GlobalFLATStorePats <GLOBAL_STORE_SHORT, store_global, i16>;
1377defm : GlobalFLATStorePats <GLOBAL_STORE_DWORDX3, store_global, v3i32>;
1378
1379let OtherPredicates = [HasD16LoadStore] in {
1380defm : GlobalFLATStorePats <GLOBAL_STORE_SHORT_D16_HI, truncstorei16_hi16_global, i32>;
1381defm : GlobalFLATStorePats <GLOBAL_STORE_BYTE_D16_HI, truncstorei8_hi16_global, i32>;
1382}
1383
1384let OtherPredicates = [D16PreservesUnusedBits] in {
1385defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_global, v2i16>;
1386defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_global, v2f16>;
1387defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SBYTE_D16_HI, sextloadi8_d16_hi_global, v2i16>;
1388defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SBYTE_D16_HI, sextloadi8_d16_hi_global, v2f16>;
1389defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SHORT_D16_HI, load_d16_hi_global, v2i16>;
1390defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SHORT_D16_HI, load_d16_hi_global, v2f16>;
1391
1392defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_UBYTE_D16, az_extloadi8_d16_lo_global, v2i16>;
1393defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_UBYTE_D16, az_extloadi8_d16_lo_global, v2f16>;
1394defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SBYTE_D16, sextloadi8_d16_lo_global, v2i16>;
1395defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SBYTE_D16, sextloadi8_d16_lo_global, v2f16>;
1396defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SHORT_D16, load_d16_lo_global, v2i16>;
1397defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SHORT_D16, load_d16_lo_global, v2f16>;
1398}
1399
1400defm : GlobalFLATAtomicStorePats <GLOBAL_STORE_BYTE, atomic_store_8_global, i32>;
1401defm : GlobalFLATAtomicStorePats <GLOBAL_STORE_BYTE, atomic_store_8_global, i16>;
1402defm : GlobalFLATAtomicStorePats <GLOBAL_STORE_SHORT, atomic_store_16_global, i32>;
1403defm : GlobalFLATAtomicStorePats <GLOBAL_STORE_SHORT, atomic_store_16_global, i16>;
1404defm : GlobalFLATAtomicStorePats <GLOBAL_STORE_DWORD, atomic_store_32_global, i32>;
1405defm : GlobalFLATAtomicStorePats <GLOBAL_STORE_DWORDX2, atomic_store_64_global, i64>;
1406
1407defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_ADD", "atomic_load_add_global", i32>;
1408defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SUB", "atomic_load_sub_global", i32>;
1409defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_INC", "atomic_inc_global", i32>;
1410defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_DEC", "atomic_dec_global", i32>;
1411defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_AND", "atomic_load_and_global", i32>;
1412defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SMAX", "atomic_load_max_global", i32>;
1413defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_UMAX", "atomic_load_umax_global", i32>;
1414defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SMIN", "atomic_load_min_global", i32>;
1415defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_UMIN", "atomic_load_umin_global", i32>;
1416defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_OR", "atomic_load_or_global", i32>;
1417defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SWAP", "atomic_swap_global", i32>;
1418defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_CMPSWAP", "AMDGPUatomic_cmp_swap_global", i32, v2i32>;
1419defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_XOR", "atomic_load_xor_global", i32>;
1420defm : GlobalFLATAtomicPatsRtn <"GLOBAL_ATOMIC_CSUB", int_amdgcn_global_atomic_csub, i32>;
1421
1422defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_ADD_X2", "atomic_load_add_global", i64>;
1423defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SUB_X2", "atomic_load_sub_global", i64>;
1424defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_INC_X2", "atomic_inc_global", i64>;
1425defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_DEC_X2", "atomic_dec_global", i64>;
1426defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_AND_X2", "atomic_load_and_global", i64>;
1427defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SMAX_X2", "atomic_load_max_global", i64>;
1428defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_UMAX_X2", "atomic_load_umax_global", i64>;
1429defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SMIN_X2", "atomic_load_min_global", i64>;
1430defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_UMIN_X2", "atomic_load_umin_global", i64>;
1431defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_OR_X2", "atomic_load_or_global", i64>;
1432defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SWAP_X2", "atomic_swap_global", i64>;
1433defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_CMPSWAP_X2", "AMDGPUatomic_cmp_swap_global", i64, v2i64>;
1434defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_XOR_X2", "atomic_load_xor_global", i64>;
1435
1436let OtherPredicates = [isGFX10Plus] in {
1437defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_FMIN", "atomic_load_fmin_global", f32>;
1438defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_FMAX", "atomic_load_fmax_global", f32>;
1439defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_FMIN_X2", "atomic_load_fmin_global", f64>;
1440defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_FMAX_X2", "atomic_load_fmax_global", f64>;
1441defm : GlobalFLATAtomicIntrPats <"GLOBAL_ATOMIC_FMIN", "int_amdgcn_global_atomic_fmin", f32>;
1442defm : GlobalFLATAtomicIntrPats <"GLOBAL_ATOMIC_FMAX", "int_amdgcn_global_atomic_fmax", f32>;
1443defm : GlobalFLATAtomicIntrPats <"GLOBAL_ATOMIC_FMIN_X2", "int_amdgcn_global_atomic_fmin", f64>;
1444defm : GlobalFLATAtomicIntrPats <"GLOBAL_ATOMIC_FMAX_X2", "int_amdgcn_global_atomic_fmax", f64>;
1445}
1446
1447let OtherPredicates = [HasAtomicFaddNoRtnInsts] in
1448defm : GlobalFLATNoRtnAtomicPats <GLOBAL_ATOMIC_ADD_F32,    atomic_load_fadd_global_noret_32, f32>;
1449let OtherPredicates = [HasAtomicPkFaddNoRtnInsts] in
1450defm : GlobalFLATNoRtnAtomicPats <GLOBAL_ATOMIC_PK_ADD_F16, atomic_load_fadd_v2f16_global_noret_32, v2f16>;
1451
1452let OtherPredicates = [isGFX90APlus] in {
1453defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_ADD_F32", "atomic_load_fadd_global", f32>;
1454defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_PK_ADD_F16", "atomic_load_fadd_v2f16_global", v2f16>;
1455defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_ADD_F64", "atomic_load_fadd_global", f64>;
1456defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_MIN_F64", "atomic_load_fmin_global", f64>;
1457defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_MAX_F64", "atomic_load_fmax_global", f64>;
1458defm : GlobalFLATAtomicIntrPats <"GLOBAL_ATOMIC_ADD_F32", "int_amdgcn_global_atomic_fadd", f32>;
1459defm : GlobalFLATAtomicIntrPats <"GLOBAL_ATOMIC_ADD_F64", "int_amdgcn_global_atomic_fadd", f64>;
1460defm : GlobalFLATAtomicIntrPats <"GLOBAL_ATOMIC_PK_ADD_F16", "int_amdgcn_global_atomic_fadd", v2f16>;
1461defm : GlobalFLATAtomicIntrPats <"GLOBAL_ATOMIC_MIN_F64", "int_amdgcn_global_atomic_fmin", f64>;
1462defm : GlobalFLATAtomicIntrPats <"GLOBAL_ATOMIC_MAX_F64", "int_amdgcn_global_atomic_fmax", f64>;
1463defm : FlatSignedAtomicPat <"FLAT_ATOMIC_ADD_F64", "atomic_load_fadd_flat", f64>;
1464defm : FlatSignedAtomicPat <"FLAT_ATOMIC_MIN_F64", "atomic_load_fmin_flat", f64>;
1465defm : FlatSignedAtomicPat <"FLAT_ATOMIC_MAX_F64", "atomic_load_fmax_flat", f64>;
1466defm : FlatSignedAtomicIntrPat <"FLAT_ATOMIC_ADD_F64", "int_amdgcn_flat_atomic_fadd", f64>;
1467defm : FlatSignedAtomicIntrPat <"FLAT_ATOMIC_MIN_F64", "int_amdgcn_flat_atomic_fmin", f64>;
1468defm : FlatSignedAtomicIntrPat <"FLAT_ATOMIC_MAX_F64", "int_amdgcn_flat_atomic_fmax", f64>;
1469}
1470
1471let OtherPredicates = [isGFX940Plus] in {
1472defm : FlatSignedAtomicPat <"FLAT_ATOMIC_ADD_F32",    "atomic_load_fadd_flat",       f32>;
1473defm : FlatSignedAtomicPat <"FLAT_ATOMIC_PK_ADD_F16", "atomic_load_fadd_v2f16_flat", v2f16>;
1474defm : FlatSignedAtomicIntrPat <"FLAT_ATOMIC_ADD_F32",     "int_amdgcn_flat_atomic_fadd",        f32>;
1475defm : FlatSignedAtomicIntrPat <"FLAT_ATOMIC_PK_ADD_F16",  "int_amdgcn_flat_atomic_fadd",        v2f16>;
1476defm : FlatSignedAtomicIntrPat <"FLAT_ATOMIC_PK_ADD_BF16", "int_amdgcn_flat_atomic_fadd_v2bf16", v2i16>;
1477defm : GlobalFLATAtomicIntrPats <"GLOBAL_ATOMIC_PK_ADD_BF16", "int_amdgcn_global_atomic_fadd_v2bf16", v2i16>;
1478}
1479
1480} // End OtherPredicates = [HasFlatGlobalInsts], AddedComplexity = 10
1481
1482let OtherPredicates = [HasFlatScratchInsts, EnableFlatScratch] in {
1483
1484defm : ScratchFLATLoadPats <SCRATCH_LOAD_UBYTE, extloadi8_private, i32>;
1485defm : ScratchFLATLoadPats <SCRATCH_LOAD_UBYTE, zextloadi8_private, i32>;
1486defm : ScratchFLATLoadPats <SCRATCH_LOAD_SBYTE, sextloadi8_private, i32>;
1487defm : ScratchFLATLoadPats <SCRATCH_LOAD_UBYTE, extloadi8_private, i16>;
1488defm : ScratchFLATLoadPats <SCRATCH_LOAD_UBYTE, zextloadi8_private, i16>;
1489defm : ScratchFLATLoadPats <SCRATCH_LOAD_SBYTE, sextloadi8_private, i16>;
1490defm : ScratchFLATLoadPats <SCRATCH_LOAD_USHORT, extloadi16_private, i32>;
1491defm : ScratchFLATLoadPats <SCRATCH_LOAD_USHORT, zextloadi16_private, i32>;
1492defm : ScratchFLATLoadPats <SCRATCH_LOAD_SSHORT, sextloadi16_private, i32>;
1493defm : ScratchFLATLoadPats <SCRATCH_LOAD_USHORT, load_private, i16>;
1494
1495foreach vt = Reg32Types.types in {
1496defm : ScratchFLATLoadPats <SCRATCH_LOAD_DWORD, load_private, vt>;
1497defm : ScratchFLATStorePats <SCRATCH_STORE_DWORD, store_private, vt>;
1498}
1499
1500foreach vt = VReg_64.RegTypes in {
1501defm : ScratchFLATLoadPats <SCRATCH_LOAD_DWORDX2, load_private, vt>;
1502defm : ScratchFLATStorePats <SCRATCH_STORE_DWORDX2, store_private, vt>;
1503}
1504
1505defm : ScratchFLATLoadPats <SCRATCH_LOAD_DWORDX3, load_private, v3i32>;
1506
1507foreach vt = VReg_128.RegTypes in {
1508defm : ScratchFLATLoadPats <SCRATCH_LOAD_DWORDX4, load_private, vt>;
1509defm : ScratchFLATStorePats <SCRATCH_STORE_DWORDX4, store_private, vt>;
1510}
1511
1512defm : ScratchFLATStorePats <SCRATCH_STORE_BYTE, truncstorei8_private, i32>;
1513defm : ScratchFLATStorePats <SCRATCH_STORE_BYTE, truncstorei8_private, i16>;
1514defm : ScratchFLATStorePats <SCRATCH_STORE_SHORT, truncstorei16_private, i32>;
1515defm : ScratchFLATStorePats <SCRATCH_STORE_SHORT, store_private, i16>;
1516defm : ScratchFLATStorePats <SCRATCH_STORE_DWORDX3, store_private, v3i32>;
1517
1518let OtherPredicates = [HasD16LoadStore, HasFlatScratchInsts, EnableFlatScratch] in {
1519defm : ScratchFLATStorePats <SCRATCH_STORE_SHORT_D16_HI, truncstorei16_hi16_private, i32>;
1520defm : ScratchFLATStorePats <SCRATCH_STORE_BYTE_D16_HI, truncstorei8_hi16_private, i32>;
1521}
1522
1523let OtherPredicates = [D16PreservesUnusedBits, HasFlatScratchInsts, EnableFlatScratch] in {
1524defm : ScratchFLATLoadPats_D16 <SCRATCH_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_private, v2i16>;
1525defm : ScratchFLATLoadPats_D16 <SCRATCH_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_private, v2f16>;
1526defm : ScratchFLATLoadPats_D16 <SCRATCH_LOAD_SBYTE_D16_HI, sextloadi8_d16_hi_private, v2i16>;
1527defm : ScratchFLATLoadPats_D16 <SCRATCH_LOAD_SBYTE_D16_HI, sextloadi8_d16_hi_private, v2f16>;
1528defm : ScratchFLATLoadPats_D16 <SCRATCH_LOAD_SHORT_D16_HI, load_d16_hi_private, v2i16>;
1529defm : ScratchFLATLoadPats_D16 <SCRATCH_LOAD_SHORT_D16_HI, load_d16_hi_private, v2f16>;
1530
1531defm : ScratchFLATLoadPats_D16 <SCRATCH_LOAD_UBYTE_D16, az_extloadi8_d16_lo_private, v2i16>;
1532defm : ScratchFLATLoadPats_D16 <SCRATCH_LOAD_UBYTE_D16, az_extloadi8_d16_lo_private, v2f16>;
1533defm : ScratchFLATLoadPats_D16 <SCRATCH_LOAD_SBYTE_D16, sextloadi8_d16_lo_private, v2i16>;
1534defm : ScratchFLATLoadPats_D16 <SCRATCH_LOAD_SBYTE_D16, sextloadi8_d16_lo_private, v2f16>;
1535defm : ScratchFLATLoadPats_D16 <SCRATCH_LOAD_SHORT_D16, load_d16_lo_private, v2i16>;
1536defm : ScratchFLATLoadPats_D16 <SCRATCH_LOAD_SHORT_D16, load_d16_lo_private, v2f16>;
1537}
1538
1539} // End OtherPredicates = [HasFlatScratchInsts,EnableFlatScratch]
1540
1541//===----------------------------------------------------------------------===//
1542// Target
1543//===----------------------------------------------------------------------===//
1544
1545//===----------------------------------------------------------------------===//
1546// CI
1547//===----------------------------------------------------------------------===//
1548
1549class FLAT_Real_ci <bits<7> op, FLAT_Pseudo ps> :
1550  FLAT_Real <op, ps>,
1551  SIMCInstr <ps.PseudoInstr, SIEncodingFamily.SI> {
1552  let AssemblerPredicate = isGFX7Only;
1553  let DecoderNamespace="GFX7";
1554}
1555
1556def FLAT_LOAD_UBYTE_ci         : FLAT_Real_ci <0x8,  FLAT_LOAD_UBYTE>;
1557def FLAT_LOAD_SBYTE_ci         : FLAT_Real_ci <0x9,  FLAT_LOAD_SBYTE>;
1558def FLAT_LOAD_USHORT_ci        : FLAT_Real_ci <0xa,  FLAT_LOAD_USHORT>;
1559def FLAT_LOAD_SSHORT_ci        : FLAT_Real_ci <0xb,  FLAT_LOAD_SSHORT>;
1560def FLAT_LOAD_DWORD_ci         : FLAT_Real_ci <0xc,  FLAT_LOAD_DWORD>;
1561def FLAT_LOAD_DWORDX2_ci       : FLAT_Real_ci <0xd,  FLAT_LOAD_DWORDX2>;
1562def FLAT_LOAD_DWORDX4_ci       : FLAT_Real_ci <0xe,  FLAT_LOAD_DWORDX4>;
1563def FLAT_LOAD_DWORDX3_ci       : FLAT_Real_ci <0xf,  FLAT_LOAD_DWORDX3>;
1564
1565def FLAT_STORE_BYTE_ci         : FLAT_Real_ci <0x18, FLAT_STORE_BYTE>;
1566def FLAT_STORE_SHORT_ci        : FLAT_Real_ci <0x1a, FLAT_STORE_SHORT>;
1567def FLAT_STORE_DWORD_ci        : FLAT_Real_ci <0x1c, FLAT_STORE_DWORD>;
1568def FLAT_STORE_DWORDX2_ci      : FLAT_Real_ci <0x1d, FLAT_STORE_DWORDX2>;
1569def FLAT_STORE_DWORDX4_ci      : FLAT_Real_ci <0x1e, FLAT_STORE_DWORDX4>;
1570def FLAT_STORE_DWORDX3_ci      : FLAT_Real_ci <0x1f, FLAT_STORE_DWORDX3>;
1571
1572multiclass FLAT_Real_Atomics_ci <bits<7> op, FLAT_Pseudo ps> {
1573  def _ci     : FLAT_Real_ci<op, !cast<FLAT_Pseudo>(ps.PseudoInstr)>;
1574  def _RTN_ci : FLAT_Real_ci<op, !cast<FLAT_Pseudo>(ps.PseudoInstr # "_RTN")>;
1575}
1576
1577defm FLAT_ATOMIC_SWAP          : FLAT_Real_Atomics_ci <0x30, FLAT_ATOMIC_SWAP>;
1578defm FLAT_ATOMIC_CMPSWAP       : FLAT_Real_Atomics_ci <0x31, FLAT_ATOMIC_CMPSWAP>;
1579defm FLAT_ATOMIC_ADD           : FLAT_Real_Atomics_ci <0x32, FLAT_ATOMIC_ADD>;
1580defm FLAT_ATOMIC_SUB           : FLAT_Real_Atomics_ci <0x33, FLAT_ATOMIC_SUB>;
1581defm FLAT_ATOMIC_SMIN          : FLAT_Real_Atomics_ci <0x35, FLAT_ATOMIC_SMIN>;
1582defm FLAT_ATOMIC_UMIN          : FLAT_Real_Atomics_ci <0x36, FLAT_ATOMIC_UMIN>;
1583defm FLAT_ATOMIC_SMAX          : FLAT_Real_Atomics_ci <0x37, FLAT_ATOMIC_SMAX>;
1584defm FLAT_ATOMIC_UMAX          : FLAT_Real_Atomics_ci <0x38, FLAT_ATOMIC_UMAX>;
1585defm FLAT_ATOMIC_AND           : FLAT_Real_Atomics_ci <0x39, FLAT_ATOMIC_AND>;
1586defm FLAT_ATOMIC_OR            : FLAT_Real_Atomics_ci <0x3a, FLAT_ATOMIC_OR>;
1587defm FLAT_ATOMIC_XOR           : FLAT_Real_Atomics_ci <0x3b, FLAT_ATOMIC_XOR>;
1588defm FLAT_ATOMIC_INC           : FLAT_Real_Atomics_ci <0x3c, FLAT_ATOMIC_INC>;
1589defm FLAT_ATOMIC_DEC           : FLAT_Real_Atomics_ci <0x3d, FLAT_ATOMIC_DEC>;
1590defm FLAT_ATOMIC_SWAP_X2       : FLAT_Real_Atomics_ci <0x50, FLAT_ATOMIC_SWAP_X2>;
1591defm FLAT_ATOMIC_CMPSWAP_X2    : FLAT_Real_Atomics_ci <0x51, FLAT_ATOMIC_CMPSWAP_X2>;
1592defm FLAT_ATOMIC_ADD_X2        : FLAT_Real_Atomics_ci <0x52, FLAT_ATOMIC_ADD_X2>;
1593defm FLAT_ATOMIC_SUB_X2        : FLAT_Real_Atomics_ci <0x53, FLAT_ATOMIC_SUB_X2>;
1594defm FLAT_ATOMIC_SMIN_X2       : FLAT_Real_Atomics_ci <0x55, FLAT_ATOMIC_SMIN_X2>;
1595defm FLAT_ATOMIC_UMIN_X2       : FLAT_Real_Atomics_ci <0x56, FLAT_ATOMIC_UMIN_X2>;
1596defm FLAT_ATOMIC_SMAX_X2       : FLAT_Real_Atomics_ci <0x57, FLAT_ATOMIC_SMAX_X2>;
1597defm FLAT_ATOMIC_UMAX_X2       : FLAT_Real_Atomics_ci <0x58, FLAT_ATOMIC_UMAX_X2>;
1598defm FLAT_ATOMIC_AND_X2        : FLAT_Real_Atomics_ci <0x59, FLAT_ATOMIC_AND_X2>;
1599defm FLAT_ATOMIC_OR_X2         : FLAT_Real_Atomics_ci <0x5a, FLAT_ATOMIC_OR_X2>;
1600defm FLAT_ATOMIC_XOR_X2        : FLAT_Real_Atomics_ci <0x5b, FLAT_ATOMIC_XOR_X2>;
1601defm FLAT_ATOMIC_INC_X2        : FLAT_Real_Atomics_ci <0x5c, FLAT_ATOMIC_INC_X2>;
1602defm FLAT_ATOMIC_DEC_X2        : FLAT_Real_Atomics_ci <0x5d, FLAT_ATOMIC_DEC_X2>;
1603
1604// CI Only flat instructions
1605defm FLAT_ATOMIC_FCMPSWAP      : FLAT_Real_Atomics_ci <0x3e, FLAT_ATOMIC_FCMPSWAP>;
1606defm FLAT_ATOMIC_FMIN          : FLAT_Real_Atomics_ci <0x3f, FLAT_ATOMIC_FMIN>;
1607defm FLAT_ATOMIC_FMAX          : FLAT_Real_Atomics_ci <0x40, FLAT_ATOMIC_FMAX>;
1608defm FLAT_ATOMIC_FCMPSWAP_X2   : FLAT_Real_Atomics_ci <0x5e, FLAT_ATOMIC_FCMPSWAP_X2>;
1609defm FLAT_ATOMIC_FMIN_X2       : FLAT_Real_Atomics_ci <0x5f, FLAT_ATOMIC_FMIN_X2>;
1610defm FLAT_ATOMIC_FMAX_X2       : FLAT_Real_Atomics_ci <0x60, FLAT_ATOMIC_FMAX_X2>;
1611
1612
1613//===----------------------------------------------------------------------===//
1614// VI
1615//===----------------------------------------------------------------------===//
1616
1617class FLAT_Real_vi <bits<7> op, FLAT_Pseudo ps, bit has_sccb = ps.has_sccb> :
1618  FLAT_Real <op, ps>,
1619  SIMCInstr <ps.PseudoInstr, SIEncodingFamily.VI> {
1620  let AssemblerPredicate = isGFX8GFX9;
1621  let DecoderNamespace = "GFX8";
1622
1623  let Inst{25} = !if(has_sccb, cpol{CPolBit.SCC}, ps.sccbValue);
1624  let AsmString = ps.Mnemonic #
1625                  !subst("$sccb", !if(has_sccb, "$sccb",""), ps.AsmOperands);
1626}
1627
1628multiclass FLAT_Real_AllAddr_vi<bits<7> op,
1629  bit has_sccb = !cast<FLAT_Pseudo>(NAME).has_sccb> {
1630  def _vi : FLAT_Real_vi<op, !cast<FLAT_Pseudo>(NAME), has_sccb>;
1631  def _SADDR_vi : FLAT_Real_vi<op, !cast<FLAT_Pseudo>(NAME#"_SADDR"), has_sccb>;
1632}
1633
1634class FLAT_Real_gfx940 <bits<7> op, FLAT_Pseudo ps> :
1635  FLAT_Real <op, ps>,
1636  SIMCInstr <ps.PseudoInstr, SIEncodingFamily.GFX940> {
1637  let AssemblerPredicate = isGFX940Plus;
1638  let DecoderNamespace = "GFX9";
1639  let Inst{13} = ps.sve;
1640  let Inst{25} = !if(ps.has_sccb, cpol{CPolBit.SCC}, ps.sccbValue);
1641}
1642
1643multiclass FLAT_Real_AllAddr_SVE_vi<bits<7> op> {
1644  def _vi : FLAT_Real_vi<op, !cast<FLAT_Pseudo>(NAME)> {
1645    let AssemblerPredicate = isGFX8GFX9NotGFX940;
1646    let OtherPredicates = [isGFX8GFX9NotGFX940];
1647  }
1648  def _SADDR_vi : FLAT_Real_vi<op, !cast<FLAT_Pseudo>(NAME#"_SADDR")> {
1649    let DecoderNamespace = "GFX9";
1650  }
1651  let AssemblerPredicate = isGFX940Plus, SubtargetPredicate = isGFX940Plus in {
1652    def _VE_gfx940  : FLAT_Real_gfx940<op, !cast<FLAT_Pseudo>(NAME)>;
1653    def _SVS_gfx940 : FLAT_Real_gfx940<op, !cast<FLAT_Pseudo>(NAME#"_SVS")>;
1654    def _ST_gfx940  : FLAT_Real_gfx940<op, !cast<FLAT_Pseudo>(NAME#"_ST")>;
1655  }
1656}
1657
1658def FLAT_LOAD_UBYTE_vi         : FLAT_Real_vi <0x10, FLAT_LOAD_UBYTE>;
1659def FLAT_LOAD_SBYTE_vi         : FLAT_Real_vi <0x11, FLAT_LOAD_SBYTE>;
1660def FLAT_LOAD_USHORT_vi        : FLAT_Real_vi <0x12, FLAT_LOAD_USHORT>;
1661def FLAT_LOAD_SSHORT_vi        : FLAT_Real_vi <0x13, FLAT_LOAD_SSHORT>;
1662def FLAT_LOAD_DWORD_vi         : FLAT_Real_vi <0x14, FLAT_LOAD_DWORD>;
1663def FLAT_LOAD_DWORDX2_vi       : FLAT_Real_vi <0x15, FLAT_LOAD_DWORDX2>;
1664def FLAT_LOAD_DWORDX4_vi       : FLAT_Real_vi <0x17, FLAT_LOAD_DWORDX4>;
1665def FLAT_LOAD_DWORDX3_vi       : FLAT_Real_vi <0x16, FLAT_LOAD_DWORDX3>;
1666
1667def FLAT_STORE_BYTE_vi         : FLAT_Real_vi <0x18, FLAT_STORE_BYTE>;
1668def FLAT_STORE_BYTE_D16_HI_vi  : FLAT_Real_vi <0x19, FLAT_STORE_BYTE_D16_HI>;
1669def FLAT_STORE_SHORT_vi        : FLAT_Real_vi <0x1a, FLAT_STORE_SHORT>;
1670def FLAT_STORE_SHORT_D16_HI_vi : FLAT_Real_vi <0x1b, FLAT_STORE_SHORT_D16_HI>;
1671def FLAT_STORE_DWORD_vi        : FLAT_Real_vi <0x1c, FLAT_STORE_DWORD>;
1672def FLAT_STORE_DWORDX2_vi      : FLAT_Real_vi <0x1d, FLAT_STORE_DWORDX2>;
1673def FLAT_STORE_DWORDX4_vi      : FLAT_Real_vi <0x1f, FLAT_STORE_DWORDX4>;
1674def FLAT_STORE_DWORDX3_vi      : FLAT_Real_vi <0x1e, FLAT_STORE_DWORDX3>;
1675
1676def FLAT_LOAD_UBYTE_D16_vi    : FLAT_Real_vi <0x20, FLAT_LOAD_UBYTE_D16>;
1677def FLAT_LOAD_UBYTE_D16_HI_vi : FLAT_Real_vi <0x21, FLAT_LOAD_UBYTE_D16_HI>;
1678def FLAT_LOAD_SBYTE_D16_vi    : FLAT_Real_vi <0x22, FLAT_LOAD_SBYTE_D16>;
1679def FLAT_LOAD_SBYTE_D16_HI_vi : FLAT_Real_vi <0x23, FLAT_LOAD_SBYTE_D16_HI>;
1680def FLAT_LOAD_SHORT_D16_vi    : FLAT_Real_vi <0x24, FLAT_LOAD_SHORT_D16>;
1681def FLAT_LOAD_SHORT_D16_HI_vi : FLAT_Real_vi <0x25, FLAT_LOAD_SHORT_D16_HI>;
1682
1683multiclass FLAT_Real_Atomics_vi <bits<7> op, FLAT_Pseudo ps,
1684  bit has_sccb = !cast<FLAT_Pseudo>(NAME).has_sccb> {
1685  def _vi     : FLAT_Real_vi<op, !cast<FLAT_Pseudo>(ps.PseudoInstr), has_sccb>;
1686  def _RTN_vi : FLAT_Real_vi<op, !cast<FLAT_Pseudo>(ps.PseudoInstr # "_RTN"), has_sccb>;
1687}
1688
1689multiclass FLAT_Global_Real_Atomics_vi<bits<7> op,
1690  bit has_sccb = !cast<FLAT_Pseudo>(NAME).has_sccb> :
1691  FLAT_Real_AllAddr_vi<op, has_sccb> {
1692  def _RTN_vi  : FLAT_Real_vi <op, !cast<FLAT_Pseudo>(NAME#"_RTN"), has_sccb>;
1693  def _SADDR_RTN_vi : FLAT_Real_vi <op, !cast<FLAT_Pseudo>(NAME#"_SADDR_RTN"), has_sccb>;
1694}
1695
1696
1697defm FLAT_ATOMIC_SWAP       : FLAT_Real_Atomics_vi <0x40, FLAT_ATOMIC_SWAP>;
1698defm FLAT_ATOMIC_CMPSWAP    : FLAT_Real_Atomics_vi <0x41, FLAT_ATOMIC_CMPSWAP>;
1699defm FLAT_ATOMIC_ADD        : FLAT_Real_Atomics_vi <0x42, FLAT_ATOMIC_ADD>;
1700defm FLAT_ATOMIC_SUB        : FLAT_Real_Atomics_vi <0x43, FLAT_ATOMIC_SUB>;
1701defm FLAT_ATOMIC_SMIN       : FLAT_Real_Atomics_vi <0x44, FLAT_ATOMIC_SMIN>;
1702defm FLAT_ATOMIC_UMIN       : FLAT_Real_Atomics_vi <0x45, FLAT_ATOMIC_UMIN>;
1703defm FLAT_ATOMIC_SMAX       : FLAT_Real_Atomics_vi <0x46, FLAT_ATOMIC_SMAX>;
1704defm FLAT_ATOMIC_UMAX       : FLAT_Real_Atomics_vi <0x47, FLAT_ATOMIC_UMAX>;
1705defm FLAT_ATOMIC_AND        : FLAT_Real_Atomics_vi <0x48, FLAT_ATOMIC_AND>;
1706defm FLAT_ATOMIC_OR         : FLAT_Real_Atomics_vi <0x49, FLAT_ATOMIC_OR>;
1707defm FLAT_ATOMIC_XOR        : FLAT_Real_Atomics_vi <0x4a, FLAT_ATOMIC_XOR>;
1708defm FLAT_ATOMIC_INC        : FLAT_Real_Atomics_vi <0x4b, FLAT_ATOMIC_INC>;
1709defm FLAT_ATOMIC_DEC        : FLAT_Real_Atomics_vi <0x4c, FLAT_ATOMIC_DEC>;
1710defm FLAT_ATOMIC_SWAP_X2    : FLAT_Real_Atomics_vi <0x60, FLAT_ATOMIC_SWAP_X2>;
1711defm FLAT_ATOMIC_CMPSWAP_X2 : FLAT_Real_Atomics_vi <0x61, FLAT_ATOMIC_CMPSWAP_X2>;
1712defm FLAT_ATOMIC_ADD_X2     : FLAT_Real_Atomics_vi <0x62, FLAT_ATOMIC_ADD_X2>;
1713defm FLAT_ATOMIC_SUB_X2     : FLAT_Real_Atomics_vi <0x63, FLAT_ATOMIC_SUB_X2>;
1714defm FLAT_ATOMIC_SMIN_X2    : FLAT_Real_Atomics_vi <0x64, FLAT_ATOMIC_SMIN_X2>;
1715defm FLAT_ATOMIC_UMIN_X2    : FLAT_Real_Atomics_vi <0x65, FLAT_ATOMIC_UMIN_X2>;
1716defm FLAT_ATOMIC_SMAX_X2    : FLAT_Real_Atomics_vi <0x66, FLAT_ATOMIC_SMAX_X2>;
1717defm FLAT_ATOMIC_UMAX_X2    : FLAT_Real_Atomics_vi <0x67, FLAT_ATOMIC_UMAX_X2>;
1718defm FLAT_ATOMIC_AND_X2     : FLAT_Real_Atomics_vi <0x68, FLAT_ATOMIC_AND_X2>;
1719defm FLAT_ATOMIC_OR_X2      : FLAT_Real_Atomics_vi <0x69, FLAT_ATOMIC_OR_X2>;
1720defm FLAT_ATOMIC_XOR_X2     : FLAT_Real_Atomics_vi <0x6a, FLAT_ATOMIC_XOR_X2>;
1721defm FLAT_ATOMIC_INC_X2     : FLAT_Real_Atomics_vi <0x6b, FLAT_ATOMIC_INC_X2>;
1722defm FLAT_ATOMIC_DEC_X2     : FLAT_Real_Atomics_vi <0x6c, FLAT_ATOMIC_DEC_X2>;
1723
1724defm GLOBAL_LOAD_UBYTE : FLAT_Real_AllAddr_vi <0x10>;
1725defm GLOBAL_LOAD_SBYTE : FLAT_Real_AllAddr_vi <0x11>;
1726defm GLOBAL_LOAD_USHORT : FLAT_Real_AllAddr_vi <0x12>;
1727defm GLOBAL_LOAD_SSHORT : FLAT_Real_AllAddr_vi <0x13>;
1728defm GLOBAL_LOAD_DWORD : FLAT_Real_AllAddr_vi <0x14>;
1729defm GLOBAL_LOAD_DWORDX2 : FLAT_Real_AllAddr_vi <0x15>;
1730defm GLOBAL_LOAD_DWORDX3 : FLAT_Real_AllAddr_vi <0x16>;
1731defm GLOBAL_LOAD_DWORDX4 : FLAT_Real_AllAddr_vi <0x17>;
1732
1733defm GLOBAL_LOAD_UBYTE_D16    : FLAT_Real_AllAddr_vi <0x20>;
1734defm GLOBAL_LOAD_UBYTE_D16_HI : FLAT_Real_AllAddr_vi <0x21>;
1735defm GLOBAL_LOAD_SBYTE_D16    : FLAT_Real_AllAddr_vi <0x22>;
1736defm GLOBAL_LOAD_SBYTE_D16_HI : FLAT_Real_AllAddr_vi <0x23>;
1737defm GLOBAL_LOAD_SHORT_D16    : FLAT_Real_AllAddr_vi <0x24>;
1738defm GLOBAL_LOAD_SHORT_D16_HI : FLAT_Real_AllAddr_vi <0x25>;
1739
1740defm GLOBAL_STORE_BYTE : FLAT_Real_AllAddr_vi <0x18>;
1741defm GLOBAL_STORE_BYTE_D16_HI : FLAT_Real_AllAddr_vi <0x19>;
1742defm GLOBAL_STORE_SHORT : FLAT_Real_AllAddr_vi <0x1a>;
1743defm GLOBAL_STORE_SHORT_D16_HI : FLAT_Real_AllAddr_vi <0x1b>;
1744defm GLOBAL_STORE_DWORD : FLAT_Real_AllAddr_vi <0x1c>;
1745defm GLOBAL_STORE_DWORDX2 : FLAT_Real_AllAddr_vi <0x1d>;
1746defm GLOBAL_STORE_DWORDX3 : FLAT_Real_AllAddr_vi <0x1e>;
1747defm GLOBAL_STORE_DWORDX4 : FLAT_Real_AllAddr_vi <0x1f>;
1748
1749let AssemblerPredicate = isGFX940Plus in {
1750defm GLOBAL_LOAD_LDS_UBYTE  : FLAT_Real_AllAddr_vi <0x026>;
1751defm GLOBAL_LOAD_LDS_SBYTE  : FLAT_Real_AllAddr_vi <0x027>;
1752defm GLOBAL_LOAD_LDS_USHORT : FLAT_Real_AllAddr_vi <0x028>;
1753defm GLOBAL_LOAD_LDS_SSHORT : FLAT_Real_AllAddr_vi <0x029>;
1754defm GLOBAL_LOAD_LDS_DWORD  : FLAT_Real_AllAddr_vi <0x02a>;
1755} // End let AssemblerPredicate = isGFX940Plus
1756
1757defm GLOBAL_ATOMIC_SWAP       : FLAT_Global_Real_Atomics_vi <0x40>;
1758defm GLOBAL_ATOMIC_CMPSWAP    : FLAT_Global_Real_Atomics_vi <0x41>;
1759defm GLOBAL_ATOMIC_ADD        : FLAT_Global_Real_Atomics_vi <0x42>;
1760defm GLOBAL_ATOMIC_SUB        : FLAT_Global_Real_Atomics_vi <0x43>;
1761defm GLOBAL_ATOMIC_SMIN       : FLAT_Global_Real_Atomics_vi <0x44>;
1762defm GLOBAL_ATOMIC_UMIN       : FLAT_Global_Real_Atomics_vi <0x45>;
1763defm GLOBAL_ATOMIC_SMAX       : FLAT_Global_Real_Atomics_vi <0x46>;
1764defm GLOBAL_ATOMIC_UMAX       : FLAT_Global_Real_Atomics_vi <0x47>;
1765defm GLOBAL_ATOMIC_AND        : FLAT_Global_Real_Atomics_vi <0x48>;
1766defm GLOBAL_ATOMIC_OR         : FLAT_Global_Real_Atomics_vi <0x49>;
1767defm GLOBAL_ATOMIC_XOR        : FLAT_Global_Real_Atomics_vi <0x4a>;
1768defm GLOBAL_ATOMIC_INC        : FLAT_Global_Real_Atomics_vi <0x4b>;
1769defm GLOBAL_ATOMIC_DEC        : FLAT_Global_Real_Atomics_vi <0x4c>;
1770defm GLOBAL_ATOMIC_SWAP_X2    : FLAT_Global_Real_Atomics_vi <0x60>;
1771defm GLOBAL_ATOMIC_CMPSWAP_X2 : FLAT_Global_Real_Atomics_vi <0x61>;
1772defm GLOBAL_ATOMIC_ADD_X2     : FLAT_Global_Real_Atomics_vi <0x62>;
1773defm GLOBAL_ATOMIC_SUB_X2     : FLAT_Global_Real_Atomics_vi <0x63>;
1774defm GLOBAL_ATOMIC_SMIN_X2    : FLAT_Global_Real_Atomics_vi <0x64>;
1775defm GLOBAL_ATOMIC_UMIN_X2    : FLAT_Global_Real_Atomics_vi <0x65>;
1776defm GLOBAL_ATOMIC_SMAX_X2    : FLAT_Global_Real_Atomics_vi <0x66>;
1777defm GLOBAL_ATOMIC_UMAX_X2    : FLAT_Global_Real_Atomics_vi <0x67>;
1778defm GLOBAL_ATOMIC_AND_X2     : FLAT_Global_Real_Atomics_vi <0x68>;
1779defm GLOBAL_ATOMIC_OR_X2      : FLAT_Global_Real_Atomics_vi <0x69>;
1780defm GLOBAL_ATOMIC_XOR_X2     : FLAT_Global_Real_Atomics_vi <0x6a>;
1781defm GLOBAL_ATOMIC_INC_X2     : FLAT_Global_Real_Atomics_vi <0x6b>;
1782defm GLOBAL_ATOMIC_DEC_X2     : FLAT_Global_Real_Atomics_vi <0x6c>;
1783
1784let AssemblerPredicate = isGFX940Plus in {
1785defm SCRATCH_LOAD_LDS_UBYTE  : FLAT_Real_AllAddr_SVE_vi <0x026>;
1786defm SCRATCH_LOAD_LDS_SBYTE  : FLAT_Real_AllAddr_SVE_vi <0x027>;
1787defm SCRATCH_LOAD_LDS_USHORT : FLAT_Real_AllAddr_SVE_vi <0x028>;
1788defm SCRATCH_LOAD_LDS_SSHORT : FLAT_Real_AllAddr_SVE_vi <0x029>;
1789defm SCRATCH_LOAD_LDS_DWORD  : FLAT_Real_AllAddr_SVE_vi <0x02a>;
1790} // End let AssemblerPredicate = isGFX940Plus
1791
1792defm SCRATCH_LOAD_UBYTE         : FLAT_Real_AllAddr_SVE_vi <0x10>;
1793defm SCRATCH_LOAD_SBYTE         : FLAT_Real_AllAddr_SVE_vi <0x11>;
1794defm SCRATCH_LOAD_USHORT        : FLAT_Real_AllAddr_SVE_vi <0x12>;
1795defm SCRATCH_LOAD_SSHORT        : FLAT_Real_AllAddr_SVE_vi <0x13>;
1796defm SCRATCH_LOAD_DWORD         : FLAT_Real_AllAddr_SVE_vi <0x14>;
1797defm SCRATCH_LOAD_DWORDX2       : FLAT_Real_AllAddr_SVE_vi <0x15>;
1798defm SCRATCH_LOAD_DWORDX3       : FLAT_Real_AllAddr_SVE_vi <0x16>;
1799defm SCRATCH_LOAD_DWORDX4       : FLAT_Real_AllAddr_SVE_vi <0x17>;
1800defm SCRATCH_STORE_BYTE         : FLAT_Real_AllAddr_SVE_vi <0x18>;
1801defm SCRATCH_STORE_BYTE_D16_HI  : FLAT_Real_AllAddr_SVE_vi <0x19>;
1802defm SCRATCH_LOAD_UBYTE_D16     : FLAT_Real_AllAddr_SVE_vi <0x20>;
1803defm SCRATCH_LOAD_UBYTE_D16_HI  : FLAT_Real_AllAddr_SVE_vi <0x21>;
1804defm SCRATCH_LOAD_SBYTE_D16     : FLAT_Real_AllAddr_SVE_vi <0x22>;
1805defm SCRATCH_LOAD_SBYTE_D16_HI  : FLAT_Real_AllAddr_SVE_vi <0x23>;
1806defm SCRATCH_LOAD_SHORT_D16     : FLAT_Real_AllAddr_SVE_vi <0x24>;
1807defm SCRATCH_LOAD_SHORT_D16_HI  : FLAT_Real_AllAddr_SVE_vi <0x25>;
1808defm SCRATCH_STORE_SHORT        : FLAT_Real_AllAddr_SVE_vi <0x1a>;
1809defm SCRATCH_STORE_SHORT_D16_HI : FLAT_Real_AllAddr_SVE_vi <0x1b>;
1810defm SCRATCH_STORE_DWORD        : FLAT_Real_AllAddr_SVE_vi <0x1c>;
1811defm SCRATCH_STORE_DWORDX2      : FLAT_Real_AllAddr_SVE_vi <0x1d>;
1812defm SCRATCH_STORE_DWORDX3      : FLAT_Real_AllAddr_SVE_vi <0x1e>;
1813defm SCRATCH_STORE_DWORDX4      : FLAT_Real_AllAddr_SVE_vi <0x1f>;
1814
1815let SubtargetPredicate = isGFX8GFX9NotGFX940 in {
1816  // These instructions are encoded differently on gfx90* and gfx940.
1817  defm GLOBAL_ATOMIC_ADD_F32    : FLAT_Global_Real_Atomics_vi <0x04d, 0>;
1818  defm GLOBAL_ATOMIC_PK_ADD_F16 : FLAT_Global_Real_Atomics_vi <0x04e, 0>;
1819}
1820
1821let SubtargetPredicate = isGFX90AOnly in {
1822  defm FLAT_ATOMIC_ADD_F64   : FLAT_Real_Atomics_vi<0x4f, FLAT_ATOMIC_ADD_F64, 0>;
1823  defm FLAT_ATOMIC_MIN_F64   : FLAT_Real_Atomics_vi<0x50, FLAT_ATOMIC_MIN_F64, 0>;
1824  defm FLAT_ATOMIC_MAX_F64   : FLAT_Real_Atomics_vi<0x51, FLAT_ATOMIC_MAX_F64, 0>;
1825  defm GLOBAL_ATOMIC_ADD_F64 : FLAT_Global_Real_Atomics_vi<0x4f, 0>;
1826  defm GLOBAL_ATOMIC_MIN_F64 : FLAT_Global_Real_Atomics_vi<0x50, 0>;
1827  defm GLOBAL_ATOMIC_MAX_F64 : FLAT_Global_Real_Atomics_vi<0x51, 0>;
1828} // End SubtargetPredicate = isGFX90AOnly
1829
1830multiclass FLAT_Real_AllAddr_gfx940<bits<7> op> {
1831  def _gfx940       : FLAT_Real_gfx940<op, !cast<FLAT_Pseudo>(NAME)>;
1832  def _SADDR_gfx940 : FLAT_Real_gfx940<op, !cast<FLAT_Pseudo>(NAME#"_SADDR")>;
1833}
1834
1835multiclass FLAT_Real_Atomics_gfx940 <bits<7> op, FLAT_Pseudo ps> {
1836  def _gfx940     : FLAT_Real_gfx940<op, !cast<FLAT_Pseudo>(ps.PseudoInstr)>;
1837  def _RTN_gfx940 : FLAT_Real_gfx940<op, !cast<FLAT_Pseudo>(ps.PseudoInstr # "_RTN")>;
1838}
1839
1840multiclass FLAT_Global_Real_Atomics_gfx940<bits<7> op> :
1841  FLAT_Real_AllAddr_gfx940<op> {
1842  def _RTN_gfx940       : FLAT_Real_gfx940 <op, !cast<FLAT_Pseudo>(NAME#"_RTN")>;
1843  def _SADDR_RTN_gfx940 : FLAT_Real_gfx940 <op, !cast<FLAT_Pseudo>(NAME#"_SADDR_RTN")>;
1844}
1845
1846let SubtargetPredicate = isGFX940Plus in {
1847  // These instructions are encoded differently on gfx90* and gfx940.
1848  defm GLOBAL_ATOMIC_ADD_F32     : FLAT_Global_Real_Atomics_gfx940 <0x04d>;
1849  defm GLOBAL_ATOMIC_PK_ADD_F16  : FLAT_Global_Real_Atomics_gfx940 <0x04e>;
1850
1851  defm FLAT_ATOMIC_ADD_F64       : FLAT_Real_Atomics_gfx940<0x4f, FLAT_ATOMIC_ADD_F64>;
1852  defm FLAT_ATOMIC_MIN_F64       : FLAT_Real_Atomics_gfx940<0x50, FLAT_ATOMIC_MIN_F64>;
1853  defm FLAT_ATOMIC_MAX_F64       : FLAT_Real_Atomics_gfx940<0x51, FLAT_ATOMIC_MAX_F64>;
1854  defm GLOBAL_ATOMIC_ADD_F64     : FLAT_Global_Real_Atomics_gfx940<0x4f>;
1855  defm GLOBAL_ATOMIC_MIN_F64     : FLAT_Global_Real_Atomics_gfx940<0x50>;
1856  defm GLOBAL_ATOMIC_MAX_F64     : FLAT_Global_Real_Atomics_gfx940<0x51>;
1857  defm FLAT_ATOMIC_ADD_F32       : FLAT_Real_Atomics_vi<0x4d, FLAT_ATOMIC_ADD_F32>;
1858  defm FLAT_ATOMIC_PK_ADD_F16    : FLAT_Real_Atomics_vi<0x4e, FLAT_ATOMIC_PK_ADD_F16>;
1859  defm FLAT_ATOMIC_PK_ADD_BF16   : FLAT_Real_Atomics_vi<0x52, FLAT_ATOMIC_PK_ADD_BF16>;
1860  defm GLOBAL_ATOMIC_PK_ADD_BF16 : FLAT_Global_Real_Atomics_vi<0x52>;
1861} // End SubtargetPredicate = isGFX940Plus
1862
1863//===----------------------------------------------------------------------===//
1864// GFX10.
1865//===----------------------------------------------------------------------===//
1866
1867class FLAT_Real_gfx10<bits<7> op, FLAT_Pseudo ps> :
1868    FLAT_Real<op, ps>, SIMCInstr<ps.PseudoInstr, SIEncodingFamily.GFX10> {
1869  let AssemblerPredicate = isGFX10Plus;
1870  let DecoderNamespace = "GFX10";
1871
1872  let Inst{11-0}  = offset{11-0};
1873  let Inst{12}    = !if(ps.has_dlc, cpol{CPolBit.DLC}, ps.dlcValue);
1874  let Inst{54-48} = !if(ps.has_saddr, !if(ps.enabled_saddr, saddr, 0x7d), 0x7d);
1875  let Inst{55}    = 0;
1876}
1877
1878
1879multiclass FLAT_Real_Base_gfx10<bits<7> op> {
1880  def _gfx10 :
1881    FLAT_Real_gfx10<op, !cast<FLAT_Pseudo>(NAME)>;
1882}
1883
1884multiclass FLAT_Real_RTN_gfx10<bits<7> op> {
1885  def _RTN_gfx10 :
1886    FLAT_Real_gfx10<op, !cast<FLAT_Pseudo>(NAME#"_RTN")>;
1887}
1888
1889multiclass FLAT_Real_SADDR_gfx10<bits<7> op> {
1890  def _SADDR_gfx10 :
1891    FLAT_Real_gfx10<op, !cast<FLAT_Pseudo>(NAME#"_SADDR")>;
1892}
1893
1894multiclass FLAT_Real_SADDR_RTN_gfx10<bits<7> op> {
1895  def _SADDR_RTN_gfx10 :
1896    FLAT_Real_gfx10<op, !cast<FLAT_Pseudo>(NAME#"_SADDR_RTN")>;
1897}
1898
1899multiclass FLAT_Real_ST_gfx10<bits<7> op> {
1900  def _ST_gfx10 :
1901    FLAT_Real_gfx10<op, !cast<FLAT_Pseudo>(NAME#"_ST")> {
1902      let Inst{54-48} = !cast<int>(EXEC_HI.HWEncoding);
1903      let OtherPredicates = [HasFlatScratchSTMode];
1904    }
1905}
1906
1907multiclass FLAT_Real_AllAddr_gfx10<bits<7> op> :
1908  FLAT_Real_Base_gfx10<op>,
1909  FLAT_Real_SADDR_gfx10<op>;
1910
1911multiclass FLAT_Real_Atomics_gfx10<bits<7> op> :
1912  FLAT_Real_Base_gfx10<op>,
1913  FLAT_Real_RTN_gfx10<op>;
1914
1915multiclass FLAT_Real_GlblAtomics_gfx10<bits<7> op> :
1916  FLAT_Real_AllAddr_gfx10<op>,
1917  FLAT_Real_RTN_gfx10<op>,
1918  FLAT_Real_SADDR_RTN_gfx10<op>;
1919
1920multiclass FLAT_Real_GlblAtomics_RTN_gfx10<bits<7> op> :
1921  FLAT_Real_RTN_gfx10<op>,
1922  FLAT_Real_SADDR_RTN_gfx10<op>;
1923
1924multiclass FLAT_Real_ScratchAllAddr_gfx10<bits<7> op> :
1925  FLAT_Real_Base_gfx10<op>,
1926  FLAT_Real_SADDR_gfx10<op>,
1927  FLAT_Real_ST_gfx10<op>;
1928
1929// ENC_FLAT.
1930defm FLAT_LOAD_UBYTE            : FLAT_Real_Base_gfx10<0x008>;
1931defm FLAT_LOAD_SBYTE            : FLAT_Real_Base_gfx10<0x009>;
1932defm FLAT_LOAD_USHORT           : FLAT_Real_Base_gfx10<0x00a>;
1933defm FLAT_LOAD_SSHORT           : FLAT_Real_Base_gfx10<0x00b>;
1934defm FLAT_LOAD_DWORD            : FLAT_Real_Base_gfx10<0x00c>;
1935defm FLAT_LOAD_DWORDX2          : FLAT_Real_Base_gfx10<0x00d>;
1936defm FLAT_LOAD_DWORDX4          : FLAT_Real_Base_gfx10<0x00e>;
1937defm FLAT_LOAD_DWORDX3          : FLAT_Real_Base_gfx10<0x00f>;
1938defm FLAT_STORE_BYTE            : FLAT_Real_Base_gfx10<0x018>;
1939defm FLAT_STORE_BYTE_D16_HI     : FLAT_Real_Base_gfx10<0x019>;
1940defm FLAT_STORE_SHORT           : FLAT_Real_Base_gfx10<0x01a>;
1941defm FLAT_STORE_SHORT_D16_HI    : FLAT_Real_Base_gfx10<0x01b>;
1942defm FLAT_STORE_DWORD           : FLAT_Real_Base_gfx10<0x01c>;
1943defm FLAT_STORE_DWORDX2         : FLAT_Real_Base_gfx10<0x01d>;
1944defm FLAT_STORE_DWORDX4         : FLAT_Real_Base_gfx10<0x01e>;
1945defm FLAT_STORE_DWORDX3         : FLAT_Real_Base_gfx10<0x01f>;
1946defm FLAT_LOAD_UBYTE_D16        : FLAT_Real_Base_gfx10<0x020>;
1947defm FLAT_LOAD_UBYTE_D16_HI     : FLAT_Real_Base_gfx10<0x021>;
1948defm FLAT_LOAD_SBYTE_D16        : FLAT_Real_Base_gfx10<0x022>;
1949defm FLAT_LOAD_SBYTE_D16_HI     : FLAT_Real_Base_gfx10<0x023>;
1950defm FLAT_LOAD_SHORT_D16        : FLAT_Real_Base_gfx10<0x024>;
1951defm FLAT_LOAD_SHORT_D16_HI     : FLAT_Real_Base_gfx10<0x025>;
1952defm FLAT_ATOMIC_SWAP           : FLAT_Real_Atomics_gfx10<0x030>;
1953defm FLAT_ATOMIC_CMPSWAP        : FLAT_Real_Atomics_gfx10<0x031>;
1954defm FLAT_ATOMIC_ADD            : FLAT_Real_Atomics_gfx10<0x032>;
1955defm FLAT_ATOMIC_SUB            : FLAT_Real_Atomics_gfx10<0x033>;
1956defm FLAT_ATOMIC_SMIN           : FLAT_Real_Atomics_gfx10<0x035>;
1957defm FLAT_ATOMIC_UMIN           : FLAT_Real_Atomics_gfx10<0x036>;
1958defm FLAT_ATOMIC_SMAX           : FLAT_Real_Atomics_gfx10<0x037>;
1959defm FLAT_ATOMIC_UMAX           : FLAT_Real_Atomics_gfx10<0x038>;
1960defm FLAT_ATOMIC_AND            : FLAT_Real_Atomics_gfx10<0x039>;
1961defm FLAT_ATOMIC_OR             : FLAT_Real_Atomics_gfx10<0x03a>;
1962defm FLAT_ATOMIC_XOR            : FLAT_Real_Atomics_gfx10<0x03b>;
1963defm FLAT_ATOMIC_INC            : FLAT_Real_Atomics_gfx10<0x03c>;
1964defm FLAT_ATOMIC_DEC            : FLAT_Real_Atomics_gfx10<0x03d>;
1965defm FLAT_ATOMIC_FCMPSWAP       : FLAT_Real_Atomics_gfx10<0x03e>;
1966defm FLAT_ATOMIC_FMIN           : FLAT_Real_Atomics_gfx10<0x03f>;
1967defm FLAT_ATOMIC_FMAX           : FLAT_Real_Atomics_gfx10<0x040>;
1968defm FLAT_ATOMIC_SWAP_X2        : FLAT_Real_Atomics_gfx10<0x050>;
1969defm FLAT_ATOMIC_CMPSWAP_X2     : FLAT_Real_Atomics_gfx10<0x051>;
1970defm FLAT_ATOMIC_ADD_X2         : FLAT_Real_Atomics_gfx10<0x052>;
1971defm FLAT_ATOMIC_SUB_X2         : FLAT_Real_Atomics_gfx10<0x053>;
1972defm FLAT_ATOMIC_SMIN_X2        : FLAT_Real_Atomics_gfx10<0x055>;
1973defm FLAT_ATOMIC_UMIN_X2        : FLAT_Real_Atomics_gfx10<0x056>;
1974defm FLAT_ATOMIC_SMAX_X2        : FLAT_Real_Atomics_gfx10<0x057>;
1975defm FLAT_ATOMIC_UMAX_X2        : FLAT_Real_Atomics_gfx10<0x058>;
1976defm FLAT_ATOMIC_AND_X2         : FLAT_Real_Atomics_gfx10<0x059>;
1977defm FLAT_ATOMIC_OR_X2          : FLAT_Real_Atomics_gfx10<0x05a>;
1978defm FLAT_ATOMIC_XOR_X2         : FLAT_Real_Atomics_gfx10<0x05b>;
1979defm FLAT_ATOMIC_INC_X2         : FLAT_Real_Atomics_gfx10<0x05c>;
1980defm FLAT_ATOMIC_DEC_X2         : FLAT_Real_Atomics_gfx10<0x05d>;
1981defm FLAT_ATOMIC_FCMPSWAP_X2    : FLAT_Real_Atomics_gfx10<0x05e>;
1982defm FLAT_ATOMIC_FMIN_X2        : FLAT_Real_Atomics_gfx10<0x05f>;
1983defm FLAT_ATOMIC_FMAX_X2        : FLAT_Real_Atomics_gfx10<0x060>;
1984
1985
1986// ENC_FLAT_GLBL.
1987defm GLOBAL_LOAD_UBYTE          : FLAT_Real_AllAddr_gfx10<0x008>;
1988defm GLOBAL_LOAD_SBYTE          : FLAT_Real_AllAddr_gfx10<0x009>;
1989defm GLOBAL_LOAD_USHORT         : FLAT_Real_AllAddr_gfx10<0x00a>;
1990defm GLOBAL_LOAD_SSHORT         : FLAT_Real_AllAddr_gfx10<0x00b>;
1991defm GLOBAL_LOAD_DWORD          : FLAT_Real_AllAddr_gfx10<0x00c>;
1992defm GLOBAL_LOAD_DWORDX2        : FLAT_Real_AllAddr_gfx10<0x00d>;
1993defm GLOBAL_LOAD_DWORDX4        : FLAT_Real_AllAddr_gfx10<0x00e>;
1994defm GLOBAL_LOAD_DWORDX3        : FLAT_Real_AllAddr_gfx10<0x00f>;
1995defm GLOBAL_STORE_BYTE          : FLAT_Real_AllAddr_gfx10<0x018>;
1996defm GLOBAL_STORE_BYTE_D16_HI   : FLAT_Real_AllAddr_gfx10<0x019>;
1997defm GLOBAL_STORE_SHORT         : FLAT_Real_AllAddr_gfx10<0x01a>;
1998defm GLOBAL_STORE_SHORT_D16_HI  : FLAT_Real_AllAddr_gfx10<0x01b>;
1999defm GLOBAL_STORE_DWORD         : FLAT_Real_AllAddr_gfx10<0x01c>;
2000defm GLOBAL_STORE_DWORDX2       : FLAT_Real_AllAddr_gfx10<0x01d>;
2001defm GLOBAL_STORE_DWORDX4       : FLAT_Real_AllAddr_gfx10<0x01e>;
2002defm GLOBAL_STORE_DWORDX3       : FLAT_Real_AllAddr_gfx10<0x01f>;
2003defm GLOBAL_LOAD_UBYTE_D16      : FLAT_Real_AllAddr_gfx10<0x020>;
2004defm GLOBAL_LOAD_UBYTE_D16_HI   : FLAT_Real_AllAddr_gfx10<0x021>;
2005defm GLOBAL_LOAD_SBYTE_D16      : FLAT_Real_AllAddr_gfx10<0x022>;
2006defm GLOBAL_LOAD_SBYTE_D16_HI   : FLAT_Real_AllAddr_gfx10<0x023>;
2007defm GLOBAL_LOAD_SHORT_D16      : FLAT_Real_AllAddr_gfx10<0x024>;
2008defm GLOBAL_LOAD_SHORT_D16_HI   : FLAT_Real_AllAddr_gfx10<0x025>;
2009defm GLOBAL_ATOMIC_SWAP         : FLAT_Real_GlblAtomics_gfx10<0x030>;
2010defm GLOBAL_ATOMIC_CMPSWAP      : FLAT_Real_GlblAtomics_gfx10<0x031>;
2011defm GLOBAL_ATOMIC_ADD          : FLAT_Real_GlblAtomics_gfx10<0x032>;
2012defm GLOBAL_ATOMIC_SUB          : FLAT_Real_GlblAtomics_gfx10<0x033>;
2013defm GLOBAL_ATOMIC_CSUB         : FLAT_Real_GlblAtomics_RTN_gfx10<0x034>;
2014defm GLOBAL_ATOMIC_SMIN         : FLAT_Real_GlblAtomics_gfx10<0x035>;
2015defm GLOBAL_ATOMIC_UMIN         : FLAT_Real_GlblAtomics_gfx10<0x036>;
2016defm GLOBAL_ATOMIC_SMAX         : FLAT_Real_GlblAtomics_gfx10<0x037>;
2017defm GLOBAL_ATOMIC_UMAX         : FLAT_Real_GlblAtomics_gfx10<0x038>;
2018defm GLOBAL_ATOMIC_AND          : FLAT_Real_GlblAtomics_gfx10<0x039>;
2019defm GLOBAL_ATOMIC_OR           : FLAT_Real_GlblAtomics_gfx10<0x03a>;
2020defm GLOBAL_ATOMIC_XOR          : FLAT_Real_GlblAtomics_gfx10<0x03b>;
2021defm GLOBAL_ATOMIC_INC          : FLAT_Real_GlblAtomics_gfx10<0x03c>;
2022defm GLOBAL_ATOMIC_DEC          : FLAT_Real_GlblAtomics_gfx10<0x03d>;
2023defm GLOBAL_ATOMIC_FCMPSWAP     : FLAT_Real_GlblAtomics_gfx10<0x03e>;
2024defm GLOBAL_ATOMIC_FMIN         : FLAT_Real_GlblAtomics_gfx10<0x03f>;
2025defm GLOBAL_ATOMIC_FMAX         : FLAT_Real_GlblAtomics_gfx10<0x040>;
2026defm GLOBAL_ATOMIC_SWAP_X2      : FLAT_Real_GlblAtomics_gfx10<0x050>;
2027defm GLOBAL_ATOMIC_CMPSWAP_X2   : FLAT_Real_GlblAtomics_gfx10<0x051>;
2028defm GLOBAL_ATOMIC_ADD_X2       : FLAT_Real_GlblAtomics_gfx10<0x052>;
2029defm GLOBAL_ATOMIC_SUB_X2       : FLAT_Real_GlblAtomics_gfx10<0x053>;
2030defm GLOBAL_ATOMIC_SMIN_X2      : FLAT_Real_GlblAtomics_gfx10<0x055>;
2031defm GLOBAL_ATOMIC_UMIN_X2      : FLAT_Real_GlblAtomics_gfx10<0x056>;
2032defm GLOBAL_ATOMIC_SMAX_X2      : FLAT_Real_GlblAtomics_gfx10<0x057>;
2033defm GLOBAL_ATOMIC_UMAX_X2      : FLAT_Real_GlblAtomics_gfx10<0x058>;
2034defm GLOBAL_ATOMIC_AND_X2       : FLAT_Real_GlblAtomics_gfx10<0x059>;
2035defm GLOBAL_ATOMIC_OR_X2        : FLAT_Real_GlblAtomics_gfx10<0x05a>;
2036defm GLOBAL_ATOMIC_XOR_X2       : FLAT_Real_GlblAtomics_gfx10<0x05b>;
2037defm GLOBAL_ATOMIC_INC_X2       : FLAT_Real_GlblAtomics_gfx10<0x05c>;
2038defm GLOBAL_ATOMIC_DEC_X2       : FLAT_Real_GlblAtomics_gfx10<0x05d>;
2039defm GLOBAL_ATOMIC_FCMPSWAP_X2  : FLAT_Real_GlblAtomics_gfx10<0x05e>;
2040defm GLOBAL_ATOMIC_FMIN_X2      : FLAT_Real_GlblAtomics_gfx10<0x05f>;
2041defm GLOBAL_ATOMIC_FMAX_X2      : FLAT_Real_GlblAtomics_gfx10<0x060>;
2042defm GLOBAL_LOAD_DWORD_ADDTID   : FLAT_Real_AllAddr_gfx10<0x016>;
2043defm GLOBAL_STORE_DWORD_ADDTID  : FLAT_Real_AllAddr_gfx10<0x017>;
2044
2045// ENC_FLAT_SCRATCH.
2046defm SCRATCH_LOAD_UBYTE         : FLAT_Real_ScratchAllAddr_gfx10<0x008>;
2047defm SCRATCH_LOAD_SBYTE         : FLAT_Real_ScratchAllAddr_gfx10<0x009>;
2048defm SCRATCH_LOAD_USHORT        : FLAT_Real_ScratchAllAddr_gfx10<0x00a>;
2049defm SCRATCH_LOAD_SSHORT        : FLAT_Real_ScratchAllAddr_gfx10<0x00b>;
2050defm SCRATCH_LOAD_DWORD         : FLAT_Real_ScratchAllAddr_gfx10<0x00c>;
2051defm SCRATCH_LOAD_DWORDX2       : FLAT_Real_ScratchAllAddr_gfx10<0x00d>;
2052defm SCRATCH_LOAD_DWORDX4       : FLAT_Real_ScratchAllAddr_gfx10<0x00e>;
2053defm SCRATCH_LOAD_DWORDX3       : FLAT_Real_ScratchAllAddr_gfx10<0x00f>;
2054defm SCRATCH_STORE_BYTE         : FLAT_Real_ScratchAllAddr_gfx10<0x018>;
2055defm SCRATCH_STORE_BYTE_D16_HI  : FLAT_Real_ScratchAllAddr_gfx10<0x019>;
2056defm SCRATCH_STORE_SHORT        : FLAT_Real_ScratchAllAddr_gfx10<0x01a>;
2057defm SCRATCH_STORE_SHORT_D16_HI : FLAT_Real_ScratchAllAddr_gfx10<0x01b>;
2058defm SCRATCH_STORE_DWORD        : FLAT_Real_ScratchAllAddr_gfx10<0x01c>;
2059defm SCRATCH_STORE_DWORDX2      : FLAT_Real_ScratchAllAddr_gfx10<0x01d>;
2060defm SCRATCH_STORE_DWORDX4      : FLAT_Real_ScratchAllAddr_gfx10<0x01e>;
2061defm SCRATCH_STORE_DWORDX3      : FLAT_Real_ScratchAllAddr_gfx10<0x01f>;
2062defm SCRATCH_LOAD_UBYTE_D16     : FLAT_Real_ScratchAllAddr_gfx10<0x020>;
2063defm SCRATCH_LOAD_UBYTE_D16_HI  : FLAT_Real_ScratchAllAddr_gfx10<0x021>;
2064defm SCRATCH_LOAD_SBYTE_D16     : FLAT_Real_ScratchAllAddr_gfx10<0x022>;
2065defm SCRATCH_LOAD_SBYTE_D16_HI  : FLAT_Real_ScratchAllAddr_gfx10<0x023>;
2066defm SCRATCH_LOAD_SHORT_D16     : FLAT_Real_ScratchAllAddr_gfx10<0x024>;
2067defm SCRATCH_LOAD_SHORT_D16_HI  : FLAT_Real_ScratchAllAddr_gfx10<0x025>;
2068