1//===-- VOP2Instructions.td - Vector Instruction Defintions ---------------===//
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
9//===----------------------------------------------------------------------===//
10// VOP2 Classes
11//===----------------------------------------------------------------------===//
12
13class VOP2e <bits<6> op, VOPProfile P> : Enc32 {
14  bits<8> vdst;
15  bits<9> src0;
16  bits<8> src1;
17
18  let Inst{8-0}   = !if(P.HasSrc0, src0, 0);
19  let Inst{16-9}  = !if(P.HasSrc1, src1, 0);
20  let Inst{24-17} = !if(P.EmitDst, vdst, 0);
21  let Inst{30-25} = op;
22  let Inst{31}    = 0x0; //encoding
23}
24
25class VOP2_MADKe <bits<6> op, VOPProfile P> : Enc64 {
26  bits<8>  vdst;
27  bits<9>  src0;
28  bits<8>  src1;
29  bits<32> imm;
30
31  let Inst{8-0}   = !if(P.HasSrc0, src0, 0);
32  let Inst{16-9}  = !if(P.HasSrc1, src1, 0);
33  let Inst{24-17} = !if(P.EmitDst, vdst, 0);
34  let Inst{30-25} = op;
35  let Inst{31}    = 0x0; // encoding
36  let Inst{63-32} = imm;
37}
38
39class VOP2_SDWAe <bits<6> op, VOPProfile P> : VOP_SDWAe <P> {
40  bits<8> vdst;
41  bits<8> src1;
42
43  let Inst{8-0}   = 0xf9; // sdwa
44  let Inst{16-9}  = !if(P.HasSrc1, src1{7-0}, 0);
45  let Inst{24-17} = !if(P.EmitDst, vdst{7-0}, 0);
46  let Inst{30-25} = op;
47  let Inst{31}    = 0x0; // encoding
48}
49
50class VOP2_SDWA9Ae <bits<6> op, VOPProfile P> : VOP_SDWA9Ae <P> {
51  bits<8> vdst;
52  bits<9> src1;
53
54  let Inst{8-0}   = 0xf9; // sdwa
55  let Inst{16-9}  = !if(P.HasSrc1, src1{7-0}, 0);
56  let Inst{24-17} = !if(P.EmitDst, vdst{7-0}, 0);
57  let Inst{30-25} = op;
58  let Inst{31}    = 0x0; // encoding
59  let Inst{63}    = !if(P.HasSrc1, src1{8}, 0); // src1_sgpr
60}
61
62class VOP2_Pseudo <string opName, VOPProfile P, list<dag> pattern=[], string suffix = "_e32"> :
63  VOP_Pseudo <opName, suffix, P, P.Outs32, P.Ins32, "", pattern> {
64
65  let AsmOperands = P.Asm32;
66
67  let Size = 4;
68  let mayLoad = 0;
69  let mayStore = 0;
70  let hasSideEffects = 0;
71
72  let VOP2 = 1;
73  let VALU = 1;
74  let Uses = [EXEC];
75
76  let AsmVariantName = AMDGPUAsmVariants.Default;
77}
78
79class VOP2_Real <VOP2_Pseudo ps, int EncodingFamily> :
80  InstSI <ps.OutOperandList, ps.InOperandList, ps.Mnemonic # ps.AsmOperands, []>,
81  SIMCInstr <ps.PseudoInstr, EncodingFamily> {
82
83  let isPseudo = 0;
84  let isCodeGenOnly = 0;
85
86  let Constraints     = ps.Constraints;
87  let DisableEncoding = ps.DisableEncoding;
88
89  // copy relevant pseudo op flags
90  let SubtargetPredicate = ps.SubtargetPredicate;
91  let AsmMatchConverter  = ps.AsmMatchConverter;
92  let AsmVariantName     = ps.AsmVariantName;
93  let Constraints        = ps.Constraints;
94  let DisableEncoding    = ps.DisableEncoding;
95  let TSFlags            = ps.TSFlags;
96  let UseNamedOperandTable = ps.UseNamedOperandTable;
97  let Uses                 = ps.Uses;
98  let Defs                 = ps.Defs;
99}
100
101class VOP2_SDWA_Pseudo <string OpName, VOPProfile P, list<dag> pattern=[]> :
102  VOP_SDWA_Pseudo <OpName, P, pattern> {
103  let AsmMatchConverter = "cvtSdwaVOP2";
104}
105
106class VOP2_DPP_Pseudo <string OpName, VOPProfile P, list<dag> pattern=[]> :
107  VOP_DPP_Pseudo <OpName, P, pattern> {
108}
109
110
111class getVOP2Pat64 <SDPatternOperator node, VOPProfile P> : LetDummies {
112  list<dag> ret = !if(P.HasModifiers,
113    [(set P.DstVT:$vdst,
114      (node (P.Src0VT
115              !if(P.HasOMod,
116                  (VOP3Mods0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp, i32:$omod),
117                  (VOP3Mods0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp))),
118            (P.Src1VT (VOP3Mods P.Src1VT:$src1, i32:$src1_modifiers))))],
119    [(set P.DstVT:$vdst, (node P.Src0VT:$src0, P.Src1VT:$src1))]);
120}
121
122multiclass VOP2Inst_e32<string opName,
123                        VOPProfile P,
124                        SDPatternOperator node = null_frag,
125                        string revOp = opName,
126                        bit GFX9Renamed = 0> {
127  let renamedInGFX9 = GFX9Renamed in {
128    def _e32 : VOP2_Pseudo <opName, P, VOPPatOrNull<node,P>.ret>,
129               Commutable_REV<revOp#"_e32", !eq(revOp, opName)>;
130  } // End renamedInGFX9 = GFX9Renamed
131}
132
133multiclass VOP2Inst_e64<string opName,
134                        VOPProfile P,
135                        SDPatternOperator node = null_frag,
136                        string revOp = opName,
137                        bit GFX9Renamed = 0> {
138  let renamedInGFX9 = GFX9Renamed in {
139    def _e64 : VOP3_Pseudo <opName, P, getVOP2Pat64<node, P>.ret>,
140               Commutable_REV<revOp#"_e64", !eq(revOp, opName)>;
141  } // End renamedInGFX9 = GFX9Renamed
142}
143
144multiclass VOP2Inst_sdwa<string opName,
145                         VOPProfile P,
146                         SDPatternOperator node = null_frag,
147                         string revOp = opName,
148                         bit GFX9Renamed = 0> {
149  let renamedInGFX9 = GFX9Renamed in {
150    foreach _ = BoolToList<P.HasExtSDWA>.ret in
151      def _sdwa : VOP2_SDWA_Pseudo <opName, P>;
152  } // End renamedInGFX9 = GFX9Renamed
153}
154
155multiclass VOP2Inst<string opName,
156                    VOPProfile P,
157                    SDPatternOperator node = null_frag,
158                    string revOp = opName,
159                    bit GFX9Renamed = 0> :
160    VOP2Inst_e32<opName, P, node, revOp, GFX9Renamed>,
161    VOP2Inst_e64<opName, P, node, revOp, GFX9Renamed>,
162    VOP2Inst_sdwa<opName, P, node, revOp, GFX9Renamed> {
163  let renamedInGFX9 = GFX9Renamed in {
164    foreach _ = BoolToList<P.HasExtDPP>.ret in
165      def _dpp  : VOP2_DPP_Pseudo <opName, P>;
166  }
167}
168
169multiclass VOP2bInst <string opName,
170                      VOPProfile P,
171                      SDPatternOperator node = null_frag,
172                      string revOp = opName,
173                      bit GFX9Renamed = 0,
174                      bit useSGPRInput = !eq(P.NumSrcArgs, 3)> {
175  let renamedInGFX9 = GFX9Renamed in {
176    let SchedRW = [Write32Bit, WriteSALU] in {
177      let Uses = !if(useSGPRInput, [VCC, EXEC], [EXEC]), Defs = [VCC] in {
178        def _e32 : VOP2_Pseudo <opName, P, VOPPatOrNull<node,P>.ret>,
179                   Commutable_REV<revOp#"_e32", !eq(revOp, opName)> {
180          let usesCustomInserter = !eq(P.NumSrcArgs, 2);
181        }
182
183        foreach _ = BoolToList<P.HasExtSDWA>.ret in
184          def _sdwa  : VOP2_SDWA_Pseudo <opName, P> {
185            let AsmMatchConverter = "cvtSdwaVOP2b";
186          }
187        foreach _ = BoolToList<P.HasExtDPP>.ret in
188          def _dpp  : VOP2_DPP_Pseudo <opName, P>;
189      }
190
191      def _e64 : VOP3_Pseudo <opName, P, getVOP2Pat64<node, P>.ret>,
192                 Commutable_REV<revOp#"_e64", !eq(revOp, opName)>;
193    }
194  }
195}
196
197class VOP2bInstAlias <VOP2_Pseudo ps, Instruction inst,
198                      string OpName, string opnd> :
199  InstAlias <OpName#" "#!subst("vcc", opnd, ps.Pfl.Asm32),
200             (inst ps.Pfl.DstRC:$vdst, ps.Pfl.Src0RC32:$src0,
201                   ps.Pfl.Src1RC32:$src1)>,
202  PredicateControl {
203}
204
205multiclass VOP2bInstAliases<VOP2_Pseudo ps, VOP2_Real inst, string OpName> {
206  let WaveSizePredicate = isWave32 in {
207    def : VOP2bInstAlias<ps, inst, OpName, "vcc_lo">;
208  }
209  let WaveSizePredicate = isWave64 in {
210    def : VOP2bInstAlias<ps, inst, OpName, "vcc">;
211  }
212}
213
214multiclass VOP2eInst <string opName,
215                      VOPProfile P,
216                      SDPatternOperator node = null_frag,
217                      string revOp = opName,
218                      bit useSGPRInput = !eq(P.NumSrcArgs, 3)> {
219
220  let SchedRW = [Write32Bit] in {
221    let Uses = !if(useSGPRInput, [VCC, EXEC], [EXEC]) in {
222      def _e32 : VOP2_Pseudo <opName, P>,
223                 Commutable_REV<revOp#"_e32", !eq(revOp, opName)>;
224
225      foreach _ = BoolToList<P.HasExtSDWA>.ret in
226        def _sdwa : VOP2_SDWA_Pseudo <opName, P> {
227          let AsmMatchConverter = "cvtSdwaVOP2e";
228        }
229
230      foreach _ = BoolToList<P.HasExtDPP>.ret in
231        def _dpp  : VOP2_DPP_Pseudo <opName, P>;
232    }
233
234    def _e64 : VOP3_Pseudo <opName, P, getVOP2Pat64<node, P>.ret>,
235               Commutable_REV<revOp#"_e64", !eq(revOp, opName)>;
236  }
237}
238
239class VOP2eInstAlias <VOP2_Pseudo ps, Instruction inst, string opnd> :
240  InstAlias <ps.OpName#" "#ps.Pfl.Asm32#", "#opnd,
241             (inst ps.Pfl.DstRC:$vdst, ps.Pfl.Src0RC32:$src0,
242                   ps.Pfl.Src1RC32:$src1)>,
243  PredicateControl {
244}
245
246multiclass VOP2eInstAliases<VOP2_Pseudo ps, VOP2_Real inst> {
247  let WaveSizePredicate = isWave32 in {
248    def : VOP2eInstAlias<ps, inst, "vcc_lo">;
249  }
250  let WaveSizePredicate = isWave64 in {
251    def : VOP2eInstAlias<ps, inst, "vcc">;
252  }
253}
254
255class VOP_MADAK <ValueType vt> : VOPProfile <[vt, vt, vt, vt]> {
256  field Operand ImmOpType = !if(!eq(vt.Size, 32), f32kimm, f16kimm);
257  field dag Ins32 = !if(!eq(vt.Size, 32),
258                        (ins VCSrc_f32:$src0, VGPR_32:$src1, ImmOpType:$imm),
259                        (ins VCSrc_f16:$src0, VGPR_32:$src1, ImmOpType:$imm));
260  field bit HasExt = 0;
261
262  // Hack to stop printing _e64
263  let DstRC = RegisterOperand<VGPR_32>;
264  field string Asm32 = " $vdst, $src0, $src1, $imm";
265}
266
267def VOP_MADAK_F16 : VOP_MADAK <f16>;
268def VOP_MADAK_F32 : VOP_MADAK <f32>;
269
270class VOP_MADMK <ValueType vt> : VOPProfile <[vt, vt, vt, vt]> {
271  field Operand ImmOpType = !if(!eq(vt.Size, 32), f32kimm, f16kimm);
272  field dag Ins32 = (ins VCSrc_f32:$src0, ImmOpType:$imm, VGPR_32:$src1);
273  field bit HasExt = 0;
274
275  // Hack to stop printing _e64
276  let DstRC = RegisterOperand<VGPR_32>;
277  field string Asm32 = " $vdst, $src0, $imm, $src1";
278}
279
280def VOP_MADMK_F16 : VOP_MADMK <f16>;
281def VOP_MADMK_F32 : VOP_MADMK <f32>;
282
283// FIXME: Remove src2_modifiers. It isn't used, so is wasting memory
284// and processing time but it makes it easier to convert to mad.
285class VOP_MAC <ValueType vt0, ValueType vt1=vt0> : VOPProfile <[vt0, vt1, vt1, vt0]> {
286  let Ins32 = (ins Src0RC32:$src0, Src1RC32:$src1, VGPR_32:$src2);
287  let Ins64 = getIns64<Src0RC64, Src1RC64, RegisterOperand<VGPR_32>, 3,
288                       0, HasModifiers, HasModifiers, HasOMod,
289                       Src0Mod, Src1Mod, Src2Mod>.ret;
290  let InsDPP = (ins Src0ModDPP:$src0_modifiers, Src0DPP:$src0,
291                    Src1ModDPP:$src1_modifiers, Src1DPP:$src1,
292                    VGPR_32:$src2, // stub argument
293                    dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
294                    bank_mask:$bank_mask, bound_ctrl:$bound_ctrl);
295  let InsDPP16 = !con(InsDPP, (ins FI:$fi));
296
297  let InsDPP8 = (ins Src0ModDPP:$src0_modifiers, Src0DPP:$src0,
298                     Src1ModDPP:$src1_modifiers, Src1DPP:$src1,
299                     VGPR_32:$src2, // stub argument
300                     dpp8:$dpp8, FI:$fi);
301
302  let InsSDWA = (ins Src0ModSDWA:$src0_modifiers, Src0SDWA:$src0,
303                     Src1ModSDWA:$src1_modifiers, Src1SDWA:$src1,
304                     VGPR_32:$src2, // stub argument
305                     clampmod:$clamp, omod:$omod,
306                     dst_sel:$dst_sel, dst_unused:$dst_unused,
307                     src0_sel:$src0_sel, src1_sel:$src1_sel);
308  let Asm32 = getAsm32<1, 2, vt0>.ret;
309  let Asm64 = getAsm64<1, 2, 0, HasModifiers, HasOMod, vt0>.ret;
310  let AsmDPP = getAsmDPP<1, 2, HasModifiers, vt0>.ret;
311  let AsmDPP16 = getAsmDPP16<1, 2, HasModifiers, vt0>.ret;
312  let AsmDPP8 = getAsmDPP8<1, 2, 0, vt0>.ret;
313  let AsmSDWA = getAsmSDWA<1, 2, vt0>.ret;
314  let AsmSDWA9 = getAsmSDWA9<1, 1, 2, vt0>.ret;
315  let HasSrc2 = 0;
316  let HasSrc2Mods = 0;
317
318  let HasExt = 1;
319  let HasExtDPP = 1;
320  let HasExtSDWA = 1;
321  let HasExtSDWA9 = 0;
322  let TieRegDPP = "$src2";
323}
324
325def VOP_MAC_F16 : VOP_MAC <f16>;
326def VOP_MAC_F32 : VOP_MAC <f32>;
327
328class VOP_DOT_ACC<ValueType vt0, ValueType vt1> : VOP_MAC<vt0, vt1> {
329  let HasClamp = 0;
330  let HasExtSDWA = 0;
331  let HasModifiers = 1;
332  let HasOpSel = 0;
333  let IsPacked = 0;
334}
335
336def VOP_DOT_ACC_F32_V2F16 : VOP_DOT_ACC<f32, v2f16> {
337  let Src0ModDPP = FPVRegInputMods;
338  let Src1ModDPP = FPVRegInputMods;
339}
340def VOP_DOT_ACC_I32_I32   : VOP_DOT_ACC<i32, i32>;
341
342// Write out to vcc or arbitrary SGPR.
343def VOP2b_I32_I1_I32_I32 : VOPProfile<[i32, i32, i32, untyped], 0, /*EnableClamp=*/1> {
344  let Asm32 = "$vdst, vcc, $src0, $src1";
345  let Asm64 = "$vdst, $sdst, $src0, $src1$clamp";
346  let AsmSDWA = "$vdst, vcc, $src0_modifiers, $src1_modifiers$clamp $dst_sel $dst_unused $src0_sel $src1_sel";
347  let AsmSDWA9 = "$vdst, vcc, $src0_modifiers, $src1_modifiers$clamp $dst_sel $dst_unused $src0_sel $src1_sel";
348  let AsmDPP = "$vdst, vcc, $src0, $src1 $dpp_ctrl$row_mask$bank_mask$bound_ctrl";
349  let AsmDPP8 = "$vdst, vcc, $src0, $src1 $dpp8$fi";
350  let AsmDPP16 = AsmDPP#"$fi";
351  let Outs32 = (outs DstRC:$vdst);
352  let Outs64 = (outs DstRC:$vdst, VOPDstS64orS32:$sdst);
353}
354
355// Write out to vcc or arbitrary SGPR and read in from vcc or
356// arbitrary SGPR.
357def VOP2b_I32_I1_I32_I32_I1 : VOPProfile<[i32, i32, i32, i1], 0, /*EnableClamp=*/1> {
358  let Asm32 = "$vdst, vcc, $src0, $src1, vcc";
359  let Asm64 = "$vdst, $sdst, $src0, $src1, $src2$clamp";
360  let AsmSDWA = "$vdst, vcc, $src0_modifiers, $src1_modifiers, vcc $clamp $dst_sel $dst_unused $src0_sel $src1_sel";
361  let AsmSDWA9 = "$vdst, vcc, $src0_modifiers, $src1_modifiers, vcc $clamp $dst_sel $dst_unused $src0_sel $src1_sel";
362  let AsmDPP = "$vdst, vcc, $src0, $src1, vcc $dpp_ctrl$row_mask$bank_mask$bound_ctrl";
363  let AsmDPP8 = "$vdst, vcc, $src0, $src1, vcc $dpp8$fi";
364  let AsmDPP16 = AsmDPP#"$fi";
365  let Outs32 = (outs DstRC:$vdst);
366  let Outs64 = (outs DstRC:$vdst, VOPDstS64orS32:$sdst);
367
368  // Suppress src2 implied by type since the 32-bit encoding uses an
369  // implicit VCC use.
370  let Ins32 = (ins Src0RC32:$src0, Src1RC32:$src1);
371
372  let InsSDWA = (ins Src0ModSDWA:$src0_modifiers, Src0SDWA:$src0,
373                     Src1ModSDWA:$src1_modifiers, Src1SDWA:$src1,
374                     clampmod:$clamp,
375                     dst_sel:$dst_sel, dst_unused:$dst_unused,
376                     src0_sel:$src0_sel, src1_sel:$src1_sel);
377
378  let InsDPP = (ins DstRCDPP:$old,
379                    Src0DPP:$src0,
380                    Src1DPP:$src1,
381                    dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
382                    bank_mask:$bank_mask, bound_ctrl:$bound_ctrl);
383  let InsDPP16 = !con(InsDPP, (ins FI:$fi));
384
385  let HasExt = 1;
386  let HasExtDPP = 1;
387  let HasExtSDWA = 1;
388  let HasExtSDWA9 = 1;
389}
390
391// Read in from vcc or arbitrary SGPR.
392def VOP2e_I32_I32_I32_I1 : VOPProfile<[i32, i32, i32, i1], /*EnableF32SrcMods=*/1> {
393  let Asm32 = "$vdst, $src0, $src1";
394  let Asm64 = "$vdst, $src0_modifiers, $src1_modifiers, $src2";
395  let AsmSDWA = "$vdst, $src0_modifiers, $src1_modifiers, vcc $clamp $dst_sel $dst_unused $src0_sel $src1_sel";
396  let AsmSDWA9 = "$vdst, $src0_modifiers, $src1_modifiers, vcc $clamp $dst_sel $dst_unused $src0_sel $src1_sel";
397  let AsmDPP = "$vdst, $src0, $src1, vcc $dpp_ctrl$row_mask$bank_mask$bound_ctrl";
398  let AsmDPP8 = "$vdst, $src0, $src1, vcc $dpp8$fi";
399  let AsmDPP16 = AsmDPP#"$fi";
400
401  let Outs32 = (outs DstRC:$vdst);
402  let Outs64 = (outs DstRC:$vdst);
403
404  // Suppress src2 implied by type since the 32-bit encoding uses an
405  // implicit VCC use.
406  let Ins32 = (ins Src0RC32:$src0, Src1RC32:$src1);
407
408  let InsSDWA = (ins Src0ModSDWA:$src0_modifiers, Src0SDWA:$src0,
409                     Src1ModSDWA:$src1_modifiers, Src1SDWA:$src1,
410                     clampmod:$clamp,
411                     dst_sel:$dst_sel, dst_unused:$dst_unused,
412                     src0_sel:$src0_sel, src1_sel:$src1_sel);
413
414  let InsDPP = (ins DstRCDPP:$old,
415                    Src0ModDPP:$src0_modifiers, Src0DPP:$src0,
416                    Src1ModDPP:$src1_modifiers, Src1DPP:$src1,
417                    dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
418                    bank_mask:$bank_mask, bound_ctrl:$bound_ctrl);
419  let InsDPP16 = !con(InsDPP, (ins FI:$fi));
420
421  let HasExt = 1;
422  let HasExtDPP = 1;
423  let HasExtSDWA = 1;
424  let HasExtSDWA9 = 1;
425}
426
427def VOP_READLANE : VOPProfile<[i32, i32, i32]> {
428  let Outs32 = (outs SReg_32:$vdst);
429  let Outs64 = Outs32;
430  let Ins32 = (ins VRegOrLds_32:$src0, SCSrc_b32:$src1);
431  let Ins64 = Ins32;
432  let Asm32 = " $vdst, $src0, $src1";
433  let Asm64 = Asm32;
434
435  let HasExt = 0;
436  let HasExtDPP = 0;
437  let HasExtSDWA = 0;
438  let HasExtSDWA9 = 0;
439}
440
441def VOP_WRITELANE : VOPProfile<[i32, i32, i32, i32]> {
442  let Outs32 = (outs VGPR_32:$vdst);
443  let Outs64 = Outs32;
444  let Ins32 = (ins SCSrc_b32:$src0, SCSrc_b32:$src1, VGPR_32:$vdst_in);
445  let Ins64 = Ins32;
446  let Asm32 = " $vdst, $src0, $src1";
447  let Asm64 = Asm32;
448  let HasSrc2 = 0;
449  let HasSrc2Mods = 0;
450
451  let HasExt = 0;
452  let HasExtDPP = 0;
453  let HasExtSDWA = 0;
454  let HasExtSDWA9 = 0;
455}
456
457//===----------------------------------------------------------------------===//
458// VOP2 Instructions
459//===----------------------------------------------------------------------===//
460
461defm V_CNDMASK_B32 : VOP2eInst <"v_cndmask_b32", VOP2e_I32_I32_I32_I1>;
462def V_MADMK_F32 : VOP2_Pseudo <"v_madmk_f32", VOP_MADMK_F32, []>;
463
464let isCommutable = 1 in {
465defm V_ADD_F32 : VOP2Inst <"v_add_f32", VOP_F32_F32_F32, fadd>;
466defm V_SUB_F32 : VOP2Inst <"v_sub_f32", VOP_F32_F32_F32, fsub>;
467defm V_SUBREV_F32 : VOP2Inst <"v_subrev_f32", VOP_F32_F32_F32, null_frag, "v_sub_f32">;
468defm V_MUL_LEGACY_F32 : VOP2Inst <"v_mul_legacy_f32", VOP_F32_F32_F32, AMDGPUfmul_legacy>;
469defm V_MUL_F32 : VOP2Inst <"v_mul_f32", VOP_F32_F32_F32, fmul>;
470defm V_MUL_I32_I24 : VOP2Inst <"v_mul_i32_i24", VOP_I32_I32_I32, AMDGPUmul_i24>;
471defm V_MUL_HI_I32_I24 : VOP2Inst <"v_mul_hi_i32_i24", VOP_PAT_GEN<VOP_I32_I32_I32, 2>, AMDGPUmulhi_i24>;
472defm V_MUL_U32_U24 : VOP2Inst <"v_mul_u32_u24", VOP_I32_I32_I32, AMDGPUmul_u24>;
473defm V_MUL_HI_U32_U24 : VOP2Inst <"v_mul_hi_u32_u24", VOP_PAT_GEN<VOP_I32_I32_I32, 2>, AMDGPUmulhi_u24>;
474defm V_MIN_F32 : VOP2Inst <"v_min_f32", VOP_F32_F32_F32, fminnum_like>;
475defm V_MAX_F32 : VOP2Inst <"v_max_f32", VOP_F32_F32_F32, fmaxnum_like>;
476defm V_MIN_I32 : VOP2Inst <"v_min_i32", VOP_PAT_GEN<VOP_I32_I32_I32>, smin>;
477defm V_MAX_I32 : VOP2Inst <"v_max_i32", VOP_PAT_GEN<VOP_I32_I32_I32>, smax>;
478defm V_MIN_U32 : VOP2Inst <"v_min_u32", VOP_PAT_GEN<VOP_I32_I32_I32>, umin>;
479defm V_MAX_U32 : VOP2Inst <"v_max_u32", VOP_PAT_GEN<VOP_I32_I32_I32>, umax>;
480defm V_LSHRREV_B32 : VOP2Inst <"v_lshrrev_b32", VOP_I32_I32_I32, lshr_rev, "v_lshr_b32">;
481defm V_ASHRREV_I32 : VOP2Inst <"v_ashrrev_i32", VOP_I32_I32_I32, ashr_rev, "v_ashr_i32">;
482defm V_LSHLREV_B32 : VOP2Inst <"v_lshlrev_b32", VOP_I32_I32_I32, lshl_rev, "v_lshl_b32">;
483defm V_AND_B32 : VOP2Inst <"v_and_b32", VOP_PAT_GEN<VOP_I32_I32_I32>, and>;
484defm V_OR_B32 : VOP2Inst <"v_or_b32", VOP_PAT_GEN<VOP_I32_I32_I32>, or>;
485defm V_XOR_B32 : VOP2Inst <"v_xor_b32", VOP_PAT_GEN<VOP_I32_I32_I32>, xor>;
486
487let Constraints = "$vdst = $src2", DisableEncoding="$src2",
488    isConvertibleToThreeAddress = 1 in {
489defm V_MAC_F32 : VOP2Inst <"v_mac_f32", VOP_MAC_F32>;
490}
491
492def V_MADAK_F32 : VOP2_Pseudo <"v_madak_f32", VOP_MADAK_F32, []>;
493
494// No patterns so that the scalar instructions are always selected.
495// The scalar versions will be replaced with vector when needed later.
496
497// V_ADD_I32, V_SUB_I32, and V_SUBREV_I32 where renamed to *_U32 in VI,
498// but the VI instructions behave the same as the SI versions.
499defm V_ADD_I32 : VOP2bInst <"v_add_i32", VOP2b_I32_I1_I32_I32, null_frag, "v_add_i32", 1>;
500defm V_SUB_I32 : VOP2bInst <"v_sub_i32", VOP2b_I32_I1_I32_I32, null_frag, "v_sub_i32", 1>;
501defm V_SUBREV_I32 : VOP2bInst <"v_subrev_i32", VOP2b_I32_I1_I32_I32, null_frag, "v_sub_i32", 1>;
502defm V_ADDC_U32 : VOP2bInst <"v_addc_u32", VOP2b_I32_I1_I32_I32_I1, null_frag, "v_addc_u32", 1>;
503defm V_SUBB_U32 : VOP2bInst <"v_subb_u32", VOP2b_I32_I1_I32_I32_I1, null_frag, "v_subb_u32", 1>;
504defm V_SUBBREV_U32 : VOP2bInst <"v_subbrev_u32", VOP2b_I32_I1_I32_I32_I1, null_frag, "v_subb_u32", 1>;
505
506
507let SubtargetPredicate = HasAddNoCarryInsts in {
508defm V_ADD_U32 : VOP2Inst <"v_add_u32", VOP_I32_I32_I32_ARITH, null_frag, "v_add_u32", 1>;
509defm V_SUB_U32 : VOP2Inst <"v_sub_u32", VOP_I32_I32_I32_ARITH, null_frag, "v_sub_u32", 1>;
510defm V_SUBREV_U32 : VOP2Inst <"v_subrev_u32", VOP_I32_I32_I32_ARITH, null_frag, "v_sub_u32", 1>;
511}
512
513} // End isCommutable = 1
514
515// These are special and do not read the exec mask.
516let isConvergent = 1, Uses = []<Register> in {
517def V_READLANE_B32 : VOP2_Pseudo<"v_readlane_b32", VOP_READLANE,
518  [(set i32:$vdst, (int_amdgcn_readlane i32:$src0, i32:$src1))]>;
519
520let Constraints = "$vdst = $vdst_in", DisableEncoding="$vdst_in" in {
521def V_WRITELANE_B32 : VOP2_Pseudo<"v_writelane_b32", VOP_WRITELANE,
522  [(set i32:$vdst, (int_amdgcn_writelane i32:$src0, i32:$src1, i32:$vdst_in))]>;
523} // End $vdst = $vdst_in, DisableEncoding $vdst_in
524} // End isConvergent = 1
525
526defm V_BFM_B32 : VOP2Inst <"v_bfm_b32", VOP_NO_EXT<VOP_I32_I32_I32>>;
527defm V_BCNT_U32_B32 : VOP2Inst <"v_bcnt_u32_b32", VOP_NO_EXT<VOP_I32_I32_I32>, add_ctpop>;
528defm V_MBCNT_LO_U32_B32 : VOP2Inst <"v_mbcnt_lo_u32_b32", VOP_NO_EXT<VOP_I32_I32_I32>, int_amdgcn_mbcnt_lo>;
529defm V_MBCNT_HI_U32_B32 : VOP2Inst <"v_mbcnt_hi_u32_b32", VOP_NO_EXT<VOP_I32_I32_I32>, int_amdgcn_mbcnt_hi>;
530defm V_LDEXP_F32 : VOP2Inst <"v_ldexp_f32", VOP_NO_EXT<VOP_F32_F32_I32>, AMDGPUldexp>;
531defm V_CVT_PKACCUM_U8_F32 : VOP2Inst <"v_cvt_pkaccum_u8_f32", VOP_NO_EXT<VOP_I32_F32_I32>>; // TODO: set "Uses = dst"
532defm V_CVT_PKNORM_I16_F32 : VOP2Inst <"v_cvt_pknorm_i16_f32", VOP_NO_EXT<VOP_V2I16_F32_F32>, AMDGPUpknorm_i16_f32>;
533defm V_CVT_PKNORM_U16_F32 : VOP2Inst <"v_cvt_pknorm_u16_f32", VOP_NO_EXT<VOP_V2I16_F32_F32>, AMDGPUpknorm_u16_f32>;
534defm V_CVT_PKRTZ_F16_F32 : VOP2Inst <"v_cvt_pkrtz_f16_f32", VOP_NO_EXT<VOP_V2F16_F32_F32>, AMDGPUpkrtz_f16_f32>;
535defm V_CVT_PK_U16_U32 : VOP2Inst <"v_cvt_pk_u16_u32", VOP_NO_EXT<VOP_V2I16_I32_I32>, AMDGPUpk_u16_u32>;
536defm V_CVT_PK_I16_I32 : VOP2Inst <"v_cvt_pk_i16_i32", VOP_NO_EXT<VOP_V2I16_I32_I32>, AMDGPUpk_i16_i32>;
537
538
539let SubtargetPredicate = isGFX6GFX7 in {
540defm V_MIN_LEGACY_F32 : VOP2Inst <"v_min_legacy_f32", VOP_F32_F32_F32, AMDGPUfmin_legacy>;
541defm V_MAX_LEGACY_F32 : VOP2Inst <"v_max_legacy_f32", VOP_F32_F32_F32, AMDGPUfmax_legacy>;
542} // End SubtargetPredicate = isGFX6GFX7
543
544let isCommutable = 1 in {
545let SubtargetPredicate = isGFX6GFX7GFX10 in {
546defm V_MAC_LEGACY_F32 : VOP2Inst <"v_mac_legacy_f32", VOP_F32_F32_F32>;
547} // End SubtargetPredicate = isGFX6GFX7GFX10
548let SubtargetPredicate = isGFX6GFX7 in {
549defm V_LSHR_B32 : VOP2Inst <"v_lshr_b32", VOP_PAT_GEN<VOP_I32_I32_I32>, srl>;
550defm V_ASHR_I32 : VOP2Inst <"v_ashr_i32", VOP_PAT_GEN<VOP_I32_I32_I32>, sra>;
551defm V_LSHL_B32 : VOP2Inst <"v_lshl_b32", VOP_PAT_GEN<VOP_I32_I32_I32>, shl>;
552} // End SubtargetPredicate = isGFX6GFX7
553} // End isCommutable = 1
554
555
556class DivergentBinOp<SDPatternOperator Op, VOP_Pseudo Inst> :
557  GCNPat<
558      (getDivergentFrag<Op>.ret Inst.Pfl.Src0VT:$src0, Inst.Pfl.Src1VT:$src1),
559      !if(!cast<Commutable_REV>(Inst).IsOrig,
560        (Inst $src0, $src1),
561        (Inst $src1, $src0)
562      )
563  >;
564
565class DivergentClampingBinOp<SDPatternOperator Op, VOP_Pseudo Inst> :
566  GCNPat<
567      (getDivergentFrag<Op>.ret Inst.Pfl.Src0VT:$src0, Inst.Pfl.Src1VT:$src1),
568      !if(!cast<Commutable_REV>(Inst).IsOrig,
569        (Inst $src0, $src1, 0),
570        (Inst $src1, $src0, 0)
571      )
572  >;
573
574def : DivergentBinOp<srl, V_LSHRREV_B32_e64>;
575def : DivergentBinOp<sra, V_ASHRREV_I32_e64>;
576def : DivergentBinOp<shl, V_LSHLREV_B32_e64>;
577
578let SubtargetPredicate = HasAddNoCarryInsts in {
579  def : DivergentClampingBinOp<add, V_ADD_U32_e64>;
580  def : DivergentClampingBinOp<sub, V_SUB_U32_e64>;
581}
582
583let SubtargetPredicate = isGFX6GFX7GFX8GFX9, Predicates = [isGFX6GFX7GFX8GFX9] in {
584def : DivergentClampingBinOp<add, V_ADD_I32_e64>;
585def : DivergentClampingBinOp<sub, V_SUB_I32_e64>;
586}
587
588def : DivergentBinOp<adde, V_ADDC_U32_e32>;
589def : DivergentBinOp<sube, V_SUBB_U32_e32>;
590
591class divergent_i64_BinOp <SDPatternOperator Op, Instruction Inst> :
592  GCNPat<
593      (getDivergentFrag<Op>.ret i64:$src0, i64:$src1),
594      (REG_SEQUENCE VReg_64,
595        (Inst
596          (i32 (EXTRACT_SUBREG $src0, sub0)),
597          (i32 (EXTRACT_SUBREG $src1, sub0))
598        ), sub0,
599        (Inst
600          (i32 (EXTRACT_SUBREG $src0, sub1)),
601          (i32 (EXTRACT_SUBREG $src1, sub1))
602        ), sub1
603      )
604  >;
605
606def :  divergent_i64_BinOp <and, V_AND_B32_e32>;
607def :  divergent_i64_BinOp <or,  V_OR_B32_e32>;
608def :  divergent_i64_BinOp <xor, V_XOR_B32_e32>;
609
610let SubtargetPredicate = Has16BitInsts in {
611
612let FPDPRounding = 1 in {
613def V_MADMK_F16 : VOP2_Pseudo <"v_madmk_f16", VOP_MADMK_F16, [], "">;
614defm V_LDEXP_F16 : VOP2Inst <"v_ldexp_f16", VOP_F16_F16_I32, AMDGPUldexp>;
615} // End FPDPRounding = 1
616
617defm V_LSHLREV_B16 : VOP2Inst <"v_lshlrev_b16", VOP_I16_I16_I16, lshl_rev>;
618defm V_LSHRREV_B16 : VOP2Inst <"v_lshrrev_b16", VOP_I16_I16_I16, lshr_rev>;
619defm V_ASHRREV_I16 : VOP2Inst <"v_ashrrev_i16", VOP_I16_I16_I16, ashr_rev>;
620
621let isCommutable = 1 in {
622let FPDPRounding = 1 in {
623defm V_ADD_F16 : VOP2Inst <"v_add_f16", VOP_F16_F16_F16, fadd>;
624defm V_SUB_F16 : VOP2Inst <"v_sub_f16", VOP_F16_F16_F16, fsub>;
625defm V_SUBREV_F16 : VOP2Inst <"v_subrev_f16", VOP_F16_F16_F16, null_frag, "v_sub_f16">;
626defm V_MUL_F16 : VOP2Inst <"v_mul_f16", VOP_F16_F16_F16, fmul>;
627def V_MADAK_F16 : VOP2_Pseudo <"v_madak_f16", VOP_MADAK_F16, [], "">;
628} // End FPDPRounding = 1
629defm V_ADD_U16 : VOP2Inst <"v_add_u16", VOP_I16_I16_I16, add>;
630defm V_SUB_U16 : VOP2Inst <"v_sub_u16" , VOP_I16_I16_I16, sub>;
631defm V_SUBREV_U16 : VOP2Inst <"v_subrev_u16", VOP_I16_I16_I16, null_frag, "v_sub_u16">;
632defm V_MUL_LO_U16 : VOP2Inst <"v_mul_lo_u16", VOP_I16_I16_I16, mul>;
633defm V_MAX_F16 : VOP2Inst <"v_max_f16", VOP_F16_F16_F16, fmaxnum_like>;
634defm V_MIN_F16 : VOP2Inst <"v_min_f16", VOP_F16_F16_F16, fminnum_like>;
635defm V_MAX_U16 : VOP2Inst <"v_max_u16", VOP_I16_I16_I16, umax>;
636defm V_MAX_I16 : VOP2Inst <"v_max_i16", VOP_I16_I16_I16, smax>;
637defm V_MIN_U16 : VOP2Inst <"v_min_u16", VOP_I16_I16_I16, umin>;
638defm V_MIN_I16 : VOP2Inst <"v_min_i16", VOP_I16_I16_I16, smin>;
639
640let Constraints = "$vdst = $src2", DisableEncoding="$src2",
641    isConvertibleToThreeAddress = 1 in {
642defm V_MAC_F16 : VOP2Inst <"v_mac_f16", VOP_MAC_F16>;
643}
644} // End isCommutable = 1
645
646} // End SubtargetPredicate = Has16BitInsts
647
648let SubtargetPredicate = HasDLInsts in {
649
650defm V_XNOR_B32 : VOP2Inst <"v_xnor_b32", VOP_I32_I32_I32>;
651
652let Constraints = "$vdst = $src2",
653    DisableEncoding="$src2",
654    isConvertibleToThreeAddress = 1,
655    isCommutable = 1 in {
656defm V_FMAC_F32 : VOP2Inst <"v_fmac_f32", VOP_MAC_F32>;
657}
658
659} // End SubtargetPredicate = HasDLInsts
660
661let Constraints = "$vdst = $src2",
662      DisableEncoding="$src2",
663      isConvertibleToThreeAddress = 1,
664      isCommutable = 1,
665      IsDOT = 1 in {
666  let SubtargetPredicate = HasDot5Insts in
667    defm V_DOT2C_F32_F16 : VOP2Inst<"v_dot2c_f32_f16", VOP_DOT_ACC_F32_V2F16>;
668  let SubtargetPredicate = HasDot6Insts in
669    defm V_DOT4C_I32_I8  : VOP2Inst<"v_dot4c_i32_i8",  VOP_DOT_ACC_I32_I32>;
670
671  let SubtargetPredicate = HasDot4Insts in
672    defm V_DOT2C_I32_I16 : VOP2Inst<"v_dot2c_i32_i16", VOP_DOT_ACC_I32_I32>;
673  let SubtargetPredicate = HasDot3Insts in
674    defm V_DOT8C_I32_I4  : VOP2Inst<"v_dot8c_i32_i4",  VOP_DOT_ACC_I32_I32>;
675}
676
677let AddedComplexity = 30 in {
678  def : GCNPat<
679    (f32 (AMDGPUfdot2 v2f16:$src0, v2f16:$src1, f32:$src2, (i1 DSTCLAMP.NONE))),
680    (f32 (V_DOT2C_F32_F16_e32 $src0, $src1, $src2))
681  > {
682    let SubtargetPredicate = HasDot5Insts;
683  }
684  def : GCNPat<
685    (i32 (int_amdgcn_sdot4 i32:$src0, i32:$src1, i32:$src2, (i1 DSTCLAMP.NONE))),
686    (i32 (V_DOT4C_I32_I8_e32 $src0, $src1, $src2))
687  > {
688    let SubtargetPredicate = HasDot6Insts;
689  }
690  def : GCNPat<
691    (i32 (int_amdgcn_sdot2 v2i16:$src0, v2i16:$src1, i32:$src2, (i1 DSTCLAMP.NONE))),
692    (i32 (V_DOT2C_I32_I16_e32 $src0, $src1, $src2))
693  > {
694    let SubtargetPredicate = HasDot4Insts;
695  }
696  def : GCNPat<
697    (i32 (int_amdgcn_sdot8 i32:$src0, i32:$src1, i32:$src2, (i1 DSTCLAMP.NONE))),
698    (i32 (V_DOT8C_I32_I4_e32 $src0, $src1, $src2))
699  > {
700    let SubtargetPredicate = HasDot3Insts;
701  }
702} // End AddedComplexity = 30
703
704let SubtargetPredicate = isGFX10Plus in {
705
706def V_FMAMK_F32 : VOP2_Pseudo<"v_fmamk_f32", VOP_MADMK_F32, [], "">;
707let FPDPRounding = 1 in
708def V_FMAMK_F16 : VOP2_Pseudo <"v_fmamk_f16", VOP_MADMK_F16, [], "">;
709
710let isCommutable = 1 in {
711def V_FMAAK_F32 : VOP2_Pseudo<"v_fmaak_f32", VOP_MADAK_F32, [], "">;
712let FPDPRounding = 1 in
713def V_FMAAK_F16 : VOP2_Pseudo <"v_fmaak_f16", VOP_MADAK_F16, [], "">;
714} // End isCommutable = 1
715
716let Constraints = "$vdst = $src2",
717    DisableEncoding="$src2",
718    isConvertibleToThreeAddress = 1,
719    isCommutable = 1 in {
720defm V_FMAC_F16 : VOP2Inst <"v_fmac_f16", VOP_MAC_F16>;
721}
722
723} // End SubtargetPredicate = isGFX10Plus
724
725let SubtargetPredicate = HasPkFmacF16Inst in {
726defm V_PK_FMAC_F16 : VOP2Inst<"v_pk_fmac_f16", VOP_V2F16_V2F16_V2F16>;
727} // End SubtargetPredicate = HasPkFmacF16Inst
728
729// Note: 16-bit instructions produce a 0 result in the high 16-bits
730// on GFX8 and GFX9 and preserve high 16 bits on GFX10+
731multiclass Arithmetic_i16_0Hi_Pats <SDPatternOperator op, Instruction inst> {
732
733def : GCNPat<
734  (i32 (zext (op i16:$src0, i16:$src1))),
735  (inst VSrc_b16:$src0, VSrc_b16:$src1)
736>;
737
738def : GCNPat<
739  (i64 (zext (op i16:$src0, i16:$src1))),
740   (REG_SEQUENCE VReg_64,
741     (inst $src0, $src1), sub0,
742     (V_MOV_B32_e32 (i32 0)), sub1)
743>;
744}
745
746class ZExt_i16_i1_Pat <SDNode ext> : GCNPat <
747  (i16 (ext i1:$src)),
748  (V_CNDMASK_B32_e64 (i32 0/*src0mod*/), (i32 0/*src0*/),
749                     (i32 0/*src1mod*/), (i32 1/*src1*/),
750                     $src)
751>;
752
753foreach vt = [i16, v2i16] in {
754def : GCNPat <
755  (and vt:$src0, vt:$src1),
756  (V_AND_B32_e64 VSrc_b32:$src0, VSrc_b32:$src1)
757>;
758
759def : GCNPat <
760  (or vt:$src0, vt:$src1),
761  (V_OR_B32_e64 VSrc_b32:$src0, VSrc_b32:$src1)
762>;
763
764def : GCNPat <
765  (xor vt:$src0, vt:$src1),
766  (V_XOR_B32_e64 VSrc_b32:$src0, VSrc_b32:$src1)
767>;
768}
769
770let Predicates = [Has16BitInsts] in {
771
772// Undo sub x, c -> add x, -c canonicalization since c is more likely
773// an inline immediate than -c.
774// TODO: Also do for 64-bit.
775def : GCNPat<
776  (add i16:$src0, (i16 NegSubInlineConst16:$src1)),
777  (V_SUB_U16_e64 VSrc_b16:$src0, NegSubInlineConst16:$src1)
778>;
779
780
781let Predicates = [Has16BitInsts, isGFX7GFX8GFX9] in {
782
783def : GCNPat<
784  (i32 (zext (add i16:$src0, (i16 NegSubInlineConst16:$src1)))),
785  (V_SUB_U16_e64 VSrc_b16:$src0, NegSubInlineConst16:$src1)
786>;
787
788defm : Arithmetic_i16_0Hi_Pats<add, V_ADD_U16_e64>;
789defm : Arithmetic_i16_0Hi_Pats<mul, V_MUL_LO_U16_e64>;
790defm : Arithmetic_i16_0Hi_Pats<sub, V_SUB_U16_e64>;
791defm : Arithmetic_i16_0Hi_Pats<smin, V_MIN_I16_e64>;
792defm : Arithmetic_i16_0Hi_Pats<smax, V_MAX_I16_e64>;
793defm : Arithmetic_i16_0Hi_Pats<umin, V_MIN_U16_e64>;
794defm : Arithmetic_i16_0Hi_Pats<umax, V_MAX_U16_e64>;
795defm : Arithmetic_i16_0Hi_Pats<lshl_rev, V_LSHLREV_B16_e64>;
796defm : Arithmetic_i16_0Hi_Pats<lshr_rev, V_LSHRREV_B16_e64>;
797defm : Arithmetic_i16_0Hi_Pats<ashr_rev, V_ASHRREV_I16_e64>;
798}  // End Predicates = [Has16BitInsts, isGFX7GFX8GFX9]
799
800def : ZExt_i16_i1_Pat<zext>;
801def : ZExt_i16_i1_Pat<anyext>;
802
803def : GCNPat <
804  (i16 (sext i1:$src)),
805  (V_CNDMASK_B32_e64 /*src0mod*/(i32 0), /*src0*/(i32 0),
806                     /*src1mod*/(i32 0), /*src1*/(i32 -1), $src)
807>;
808
809} // End Predicates = [Has16BitInsts]
810
811
812//===----------------------------------------------------------------------===//
813// Target-specific instruction encodings.
814//===----------------------------------------------------------------------===//
815
816class VOP2_DPP<bits<6> op, VOP2_DPP_Pseudo ps,
817               string opName = ps.OpName, VOPProfile p = ps.Pfl,
818               bit IsDPP16 = 0> :
819    VOP_DPP<opName, p, IsDPP16> {
820  let hasSideEffects = ps.hasSideEffects;
821  let Defs = ps.Defs;
822  let SchedRW = ps.SchedRW;
823  let Uses = ps.Uses;
824
825  bits<8> vdst;
826  bits<8> src1;
827  let Inst{8-0}   = 0xfa;
828  let Inst{16-9}  = !if(p.HasSrc1, src1{7-0}, 0);
829  let Inst{24-17} = !if(p.EmitDst, vdst{7-0}, 0);
830  let Inst{30-25} = op;
831  let Inst{31}    = 0x0;
832}
833
834class Base_VOP2_DPP16<bits<6> op, VOP2_DPP_Pseudo ps,
835                 string opName = ps.OpName, VOPProfile p = ps.Pfl> :
836    VOP2_DPP<op, ps, opName, p, 1> {
837  let AssemblerPredicate = !if(p.HasExt, HasDPP16, DisableInst);
838  let SubtargetPredicate = HasDPP16;
839}
840
841class VOP2_DPP16<bits<6> op, VOP2_DPP_Pseudo ps,
842                 string opName = ps.OpName, VOPProfile p = ps.Pfl> :
843    Base_VOP2_DPP16<op, ps, opName, p>,
844    SIMCInstr <ps.PseudoInstr, SIEncodingFamily.GFX10>;
845
846class VOP2_DPP8<bits<6> op, VOP2_Pseudo ps,
847                string opName = ps.OpName, VOPProfile p = ps.Pfl> :
848    VOP_DPP8<ps.OpName, p> {
849  let hasSideEffects = ps.hasSideEffects;
850  let Defs = ps.Defs;
851  let SchedRW = ps.SchedRW;
852  let Uses = ps.Uses;
853
854  bits<8> vdst;
855  bits<8> src1;
856
857  let Inst{8-0}   = fi;
858  let Inst{16-9}  = !if(p.HasSrc1, src1{7-0}, 0);
859  let Inst{24-17} = !if(p.EmitDst, vdst{7-0}, 0);
860  let Inst{30-25} = op;
861  let Inst{31}    = 0x0;
862
863  let AssemblerPredicate = !if(p.HasExt, HasDPP8, DisableInst);
864  let SubtargetPredicate = HasDPP8;
865}
866
867//===----------------------------------------------------------------------===//
868// GFX10.
869//===----------------------------------------------------------------------===//
870
871let AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10" in {
872  //===------------------------------- VOP2 -------------------------------===//
873  multiclass VOP2Only_Real_MADK_gfx10<bits<6> op> {
874    def _gfx10 :
875      VOP2_Real<!cast<VOP2_Pseudo>(NAME), SIEncodingFamily.GFX10>,
876      VOP2_MADKe<op{5-0}, !cast<VOP2_Pseudo>(NAME).Pfl>;
877  }
878  multiclass VOP2Only_Real_MADK_gfx10_with_name<bits<6> op, string opName,
879                                                string asmName> {
880    def _gfx10 :
881        VOP2_Real<!cast<VOP2_Pseudo>(opName), SIEncodingFamily.GFX10>,
882        VOP2_MADKe<op{5-0}, !cast<VOP2_Pseudo>(opName).Pfl> {
883      VOP2_Pseudo ps = !cast<VOP2_Pseudo>(opName);
884      let AsmString = asmName # ps.AsmOperands;
885    }
886  }
887  multiclass VOP2_Real_e32_gfx10<bits<6> op> {
888    def _e32_gfx10 :
889      VOP2_Real<!cast<VOP2_Pseudo>(NAME#"_e32"), SIEncodingFamily.GFX10>,
890      VOP2e<op{5-0}, !cast<VOP2_Pseudo>(NAME#"_e32").Pfl>;
891  }
892  multiclass VOP2_Real_e64_gfx10<bits<6> op> {
893    def _e64_gfx10 :
894      VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.GFX10>,
895      VOP3e_gfx10<{0, 1, 0, 0, op{5-0}}, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>;
896  }
897  multiclass VOP2_Real_sdwa_gfx10<bits<6> op> {
898    foreach _ = BoolToList<!cast<VOP2_Pseudo>(NAME#"_e32").Pfl.HasExtSDWA9>.ret in
899    def _sdwa_gfx10 :
900      VOP_SDWA10_Real<!cast<VOP2_SDWA_Pseudo>(NAME#"_sdwa")>,
901      VOP2_SDWA9Ae<op{5-0}, !cast<VOP2_SDWA_Pseudo>(NAME#"_sdwa").Pfl> {
902      let DecoderNamespace = "SDWA10";
903    }
904  }
905  multiclass VOP2_Real_dpp_gfx10<bits<6> op> {
906    foreach _ = BoolToList<!cast<VOP2_Pseudo>(NAME#"_e32").Pfl.HasExtDPP>.ret in
907    def _dpp_gfx10 : VOP2_DPP16<op, !cast<VOP2_DPP_Pseudo>(NAME#"_dpp")> {
908      let DecoderNamespace = "SDWA10";
909    }
910  }
911  multiclass VOP2_Real_dpp8_gfx10<bits<6> op> {
912    foreach _ = BoolToList<!cast<VOP2_Pseudo>(NAME#"_e32").Pfl.HasExtDPP>.ret in
913    def _dpp8_gfx10 : VOP2_DPP8<op, !cast<VOP2_Pseudo>(NAME#"_e32")> {
914      let DecoderNamespace = "DPP8";
915    }
916  }
917
918  //===------------------------- VOP2 (with name) -------------------------===//
919  multiclass VOP2_Real_e32_gfx10_with_name<bits<6> op, string opName,
920                                           string asmName> {
921    def _e32_gfx10 :
922      VOP2_Real<!cast<VOP2_Pseudo>(opName#"_e32"), SIEncodingFamily.GFX10>,
923      VOP2e<op{5-0}, !cast<VOP2_Pseudo>(opName#"_e32").Pfl> {
924        VOP2_Pseudo ps = !cast<VOP2_Pseudo>(opName#"_e32");
925        let AsmString = asmName # ps.AsmOperands;
926      }
927  }
928  multiclass VOP2_Real_e64_gfx10_with_name<bits<6> op, string opName,
929                                           string asmName> {
930    def _e64_gfx10 :
931      VOP3_Real<!cast<VOP3_Pseudo>(opName#"_e64"), SIEncodingFamily.GFX10>,
932      VOP3e_gfx10<{0, 1, 0, 0, op{5-0}},
933                  !cast<VOP3_Pseudo>(opName#"_e64").Pfl> {
934        VOP3_Pseudo ps = !cast<VOP3_Pseudo>(opName#"_e64");
935        let AsmString = asmName # ps.AsmOperands;
936      }
937  }
938  let DecoderNamespace = "SDWA10" in {
939    multiclass VOP2_Real_sdwa_gfx10_with_name<bits<6> op, string opName,
940                                              string asmName> {
941      foreach _ = BoolToList<!cast<VOP2_Pseudo>(opName#"_e32").Pfl.HasExtSDWA9>.ret in
942      def _sdwa_gfx10 :
943        VOP_SDWA10_Real<!cast<VOP2_SDWA_Pseudo>(opName#"_sdwa")>,
944        VOP2_SDWA9Ae<op{5-0}, !cast<VOP2_SDWA_Pseudo>(opName#"_sdwa").Pfl> {
945          VOP2_SDWA_Pseudo ps = !cast<VOP2_SDWA_Pseudo>(opName#"_sdwa");
946          let AsmString = asmName # ps.AsmOperands;
947        }
948    }
949    multiclass VOP2_Real_dpp_gfx10_with_name<bits<6> op, string opName,
950                                             string asmName> {
951      foreach _ = BoolToList<!cast<VOP2_Pseudo>(opName#"_e32").Pfl.HasExtDPP>.ret in
952      def _dpp_gfx10 : VOP2_DPP16<op, !cast<VOP2_DPP_Pseudo>(opName#"_dpp")> {
953        VOP2_Pseudo ps = !cast<VOP2_Pseudo>(opName#"_e32");
954        let AsmString = asmName # ps.Pfl.AsmDPP16;
955      }
956    }
957    multiclass VOP2_Real_dpp8_gfx10_with_name<bits<6> op, string opName,
958                                              string asmName> {
959      foreach _ = BoolToList<!cast<VOP2_Pseudo>(opName#"_e32").Pfl.HasExtDPP>.ret in
960      def _dpp8_gfx10 : VOP2_DPP8<op, !cast<VOP2_Pseudo>(opName#"_e32")> {
961        VOP2_Pseudo ps = !cast<VOP2_Pseudo>(opName#"_e32");
962        let AsmString = asmName # ps.Pfl.AsmDPP8;
963        let DecoderNamespace = "DPP8";
964      }
965    }
966  } // End DecoderNamespace = "SDWA10"
967
968  //===------------------------------ VOP2be ------------------------------===//
969  multiclass VOP2be_Real_e32_gfx10<bits<6> op, string opName, string asmName> {
970    def _e32_gfx10 :
971      VOP2_Real<!cast<VOP2_Pseudo>(opName#"_e32"), SIEncodingFamily.GFX10>,
972      VOP2e<op{5-0}, !cast<VOP2_Pseudo>(opName#"_e32").Pfl> {
973        VOP2_Pseudo Ps = !cast<VOP2_Pseudo>(opName#"_e32");
974        let AsmString = asmName # !subst(", vcc", "", Ps.AsmOperands);
975      }
976  }
977  multiclass VOP2be_Real_e64_gfx10<bits<6> op, string opName, string asmName> {
978    def _e64_gfx10 :
979      VOP3_Real<!cast<VOP3_Pseudo>(opName#"_e64"), SIEncodingFamily.GFX10>,
980      VOP3be_gfx10<{0, 1, 0, 0, op{5-0}},
981                   !cast<VOP3_Pseudo>(opName#"_e64").Pfl> {
982        VOP3_Pseudo Ps = !cast<VOP3_Pseudo>(opName#"_e64");
983        let AsmString = asmName # Ps.AsmOperands;
984      }
985  }
986  multiclass VOP2be_Real_sdwa_gfx10<bits<6> op, string opName, string asmName> {
987    foreach _ = BoolToList<!cast<VOP2_Pseudo>(opName#"_e32").Pfl.HasExtSDWA9>.ret in
988    def _sdwa_gfx10 :
989      VOP_SDWA10_Real<!cast<VOP2_SDWA_Pseudo>(opName#"_sdwa")>,
990      VOP2_SDWA9Ae<op{5-0}, !cast<VOP2_SDWA_Pseudo>(opName#"_sdwa").Pfl> {
991        VOP2_SDWA_Pseudo Ps = !cast<VOP2_SDWA_Pseudo>(opName#"_sdwa");
992        let AsmString = asmName # !subst(", vcc", "", Ps.AsmOperands);
993        let DecoderNamespace = "SDWA10";
994      }
995    foreach _ = BoolToList<!cast<VOP2_Pseudo>(opName#"_e32").Pfl.HasExtSDWA9>.ret in
996    def _sdwa_w32_gfx10 :
997      Base_VOP_SDWA10_Real<!cast<VOP2_SDWA_Pseudo>(opName#"_sdwa")>,
998      VOP2_SDWA9Ae<op{5-0}, !cast<VOP2_SDWA_Pseudo>(opName#"_sdwa").Pfl> {
999        VOP2_SDWA_Pseudo Ps = !cast<VOP2_SDWA_Pseudo>(opName#"_sdwa");
1000        let AsmString = asmName # !subst("vcc", "vcc_lo", Ps.AsmOperands);
1001        let isAsmParserOnly = 1;
1002        let DecoderNamespace = "SDWA10";
1003        let WaveSizePredicate = isWave32;
1004      }
1005    foreach _ = BoolToList<!cast<VOP2_Pseudo>(opName#"_e32").Pfl.HasExtSDWA9>.ret in
1006    def _sdwa_w64_gfx10 :
1007      Base_VOP_SDWA10_Real<!cast<VOP2_SDWA_Pseudo>(opName#"_sdwa")>,
1008      VOP2_SDWA9Ae<op{5-0}, !cast<VOP2_SDWA_Pseudo>(opName#"_sdwa").Pfl> {
1009        VOP2_SDWA_Pseudo Ps = !cast<VOP2_SDWA_Pseudo>(opName#"_sdwa");
1010        let AsmString = asmName # Ps.AsmOperands;
1011        let isAsmParserOnly = 1;
1012        let DecoderNamespace = "SDWA10";
1013        let WaveSizePredicate = isWave64;
1014      }
1015  }
1016  multiclass VOP2be_Real_dpp_gfx10<bits<6> op, string opName, string asmName> {
1017    foreach _ = BoolToList<!cast<VOP2_Pseudo>(opName#"_e32").Pfl.HasExtDPP>.ret in
1018    def _dpp_gfx10 :
1019      VOP2_DPP16<op, !cast<VOP2_DPP_Pseudo>(opName#"_dpp"), asmName> {
1020        string AsmDPP = !cast<VOP2_Pseudo>(opName#"_e32").Pfl.AsmDPP16;
1021        let AsmString = asmName # !subst(", vcc", "", AsmDPP);
1022        let DecoderNamespace = "SDWA10";
1023      }
1024    foreach _ = BoolToList<!cast<VOP2_Pseudo>(opName#"_e32").Pfl.HasExtDPP>.ret in
1025    def _dpp_w32_gfx10 :
1026      Base_VOP2_DPP16<op, !cast<VOP2_DPP_Pseudo>(opName#"_dpp"), asmName> {
1027        string AsmDPP = !cast<VOP2_Pseudo>(opName#"_e32").Pfl.AsmDPP16;
1028        let AsmString = asmName # !subst("vcc", "vcc_lo", AsmDPP);
1029        let isAsmParserOnly = 1;
1030        let WaveSizePredicate = isWave32;
1031      }
1032    foreach _ = BoolToList<!cast<VOP2_Pseudo>(opName#"_e32").Pfl.HasExtDPP>.ret in
1033    def _dpp_w64_gfx10 :
1034      Base_VOP2_DPP16<op, !cast<VOP2_DPP_Pseudo>(opName#"_dpp"), asmName> {
1035        string AsmDPP = !cast<VOP2_Pseudo>(opName#"_e32").Pfl.AsmDPP16;
1036        let AsmString = asmName # AsmDPP;
1037        let isAsmParserOnly = 1;
1038        let WaveSizePredicate = isWave64;
1039      }
1040  }
1041  multiclass VOP2be_Real_dpp8_gfx10<bits<6> op, string opName, string asmName> {
1042    foreach _ = BoolToList<!cast<VOP2_Pseudo>(opName#"_e32").Pfl.HasExtDPP>.ret in
1043    def _dpp8_gfx10 :
1044      VOP2_DPP8<op, !cast<VOP2_Pseudo>(opName#"_e32"), asmName> {
1045        string AsmDPP8 = !cast<VOP2_Pseudo>(opName#"_e32").Pfl.AsmDPP8;
1046        let AsmString = asmName # !subst(", vcc", "", AsmDPP8);
1047        let DecoderNamespace = "DPP8";
1048      }
1049    foreach _ = BoolToList<!cast<VOP2_Pseudo>(opName#"_e32").Pfl.HasExtDPP>.ret in
1050    def _dpp8_w32_gfx10 :
1051      VOP2_DPP8<op, !cast<VOP2_Pseudo>(opName#"_e32"), asmName> {
1052        string AsmDPP8 = !cast<VOP2_Pseudo>(opName#"_e32").Pfl.AsmDPP8;
1053        let AsmString = asmName # !subst("vcc", "vcc_lo", AsmDPP8);
1054        let isAsmParserOnly = 1;
1055        let WaveSizePredicate = isWave32;
1056      }
1057    foreach _ = BoolToList<!cast<VOP2_Pseudo>(opName#"_e32").Pfl.HasExtDPP>.ret in
1058    def _dpp8_w64_gfx10 :
1059      VOP2_DPP8<op, !cast<VOP2_Pseudo>(opName#"_e32"), asmName> {
1060        string AsmDPP8 = !cast<VOP2_Pseudo>(opName#"_e32").Pfl.AsmDPP8;
1061        let AsmString = asmName # AsmDPP8;
1062        let isAsmParserOnly = 1;
1063        let WaveSizePredicate = isWave64;
1064      }
1065  }
1066
1067  //===----------------------------- VOP3Only -----------------------------===//
1068  multiclass VOP3Only_Real_gfx10<bits<10> op> {
1069    def _e64_gfx10 :
1070      VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.GFX10>,
1071      VOP3e_gfx10<op, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>;
1072  }
1073
1074  //===---------------------------- VOP3beOnly ----------------------------===//
1075  multiclass VOP3beOnly_Real_gfx10<bits<10> op, string opName, string asmName> {
1076    def _e64_gfx10 :
1077      VOP3_Real<!cast<VOP3_Pseudo>(opName#"_e64"), SIEncodingFamily.GFX10>,
1078      VOP3be_gfx10<op, !cast<VOP3_Pseudo>(opName#"_e64").Pfl> {
1079        VOP3_Pseudo Ps = !cast<VOP3_Pseudo>(opName#"_e64");
1080        let AsmString = asmName # Ps.AsmOperands;
1081      }
1082  }
1083} // End AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10"
1084
1085multiclass VOP2be_Real_gfx10<bits<6> op, string opName, string asmName> :
1086  VOP2be_Real_e32_gfx10<op, opName, asmName>,
1087  VOP2be_Real_e64_gfx10<op, opName, asmName>,
1088  VOP2be_Real_sdwa_gfx10<op, opName, asmName>,
1089  VOP2be_Real_dpp_gfx10<op, opName, asmName>,
1090  VOP2be_Real_dpp8_gfx10<op, opName, asmName>;
1091
1092multiclass VOP2e_Real_gfx10<bits<6> op, string opName, string asmName> :
1093  VOP2_Real_e32_gfx10<op>,
1094  VOP2_Real_e64_gfx10<op>,
1095  VOP2be_Real_sdwa_gfx10<op, opName, asmName>,
1096  VOP2be_Real_dpp_gfx10<op, opName, asmName>,
1097  VOP2be_Real_dpp8_gfx10<op, opName, asmName>;
1098
1099multiclass VOP2_Real_gfx10<bits<6> op> :
1100  VOP2_Real_e32_gfx10<op>, VOP2_Real_e64_gfx10<op>,
1101  VOP2_Real_sdwa_gfx10<op>, VOP2_Real_dpp_gfx10<op>, VOP2_Real_dpp8_gfx10<op>;
1102
1103multiclass VOP2_Real_gfx10_with_name<bits<6> op, string opName,
1104                                     string asmName> :
1105  VOP2_Real_e32_gfx10_with_name<op, opName, asmName>,
1106  VOP2_Real_e64_gfx10_with_name<op, opName, asmName>,
1107  VOP2_Real_sdwa_gfx10_with_name<op, opName, asmName>,
1108  VOP2_Real_dpp_gfx10_with_name<op, opName, asmName>,
1109  VOP2_Real_dpp8_gfx10_with_name<op, opName, asmName>;
1110
1111defm V_XNOR_B32      : VOP2_Real_gfx10<0x01e>;
1112defm V_FMAC_F32      : VOP2_Real_gfx10<0x02b>;
1113defm V_FMAMK_F32     : VOP2Only_Real_MADK_gfx10<0x02c>;
1114defm V_FMAAK_F32     : VOP2Only_Real_MADK_gfx10<0x02d>;
1115defm V_ADD_F16       : VOP2_Real_gfx10<0x032>;
1116defm V_SUB_F16       : VOP2_Real_gfx10<0x033>;
1117defm V_SUBREV_F16    : VOP2_Real_gfx10<0x034>;
1118defm V_MUL_F16       : VOP2_Real_gfx10<0x035>;
1119defm V_FMAC_F16      : VOP2_Real_gfx10<0x036>;
1120defm V_FMAMK_F16     : VOP2Only_Real_MADK_gfx10<0x037>;
1121defm V_FMAAK_F16     : VOP2Only_Real_MADK_gfx10<0x038>;
1122defm V_MAX_F16       : VOP2_Real_gfx10<0x039>;
1123defm V_MIN_F16       : VOP2_Real_gfx10<0x03a>;
1124defm V_LDEXP_F16     : VOP2_Real_gfx10<0x03b>;
1125defm V_PK_FMAC_F16   : VOP2_Real_e32_gfx10<0x03c>;
1126
1127// VOP2 no carry-in, carry-out.
1128defm V_ADD_NC_U32 :
1129  VOP2_Real_gfx10_with_name<0x025, "V_ADD_U32", "v_add_nc_u32">;
1130defm V_SUB_NC_U32 :
1131  VOP2_Real_gfx10_with_name<0x026, "V_SUB_U32", "v_sub_nc_u32">;
1132defm V_SUBREV_NC_U32 :
1133  VOP2_Real_gfx10_with_name<0x027, "V_SUBREV_U32", "v_subrev_nc_u32">;
1134
1135// VOP2 carry-in, carry-out.
1136defm V_ADD_CO_CI_U32 :
1137  VOP2be_Real_gfx10<0x028, "V_ADDC_U32", "v_add_co_ci_u32">;
1138defm V_SUB_CO_CI_U32 :
1139  VOP2be_Real_gfx10<0x029, "V_SUBB_U32", "v_sub_co_ci_u32">;
1140defm V_SUBREV_CO_CI_U32 :
1141  VOP2be_Real_gfx10<0x02a, "V_SUBBREV_U32", "v_subrev_co_ci_u32">;
1142
1143defm V_CNDMASK_B32 :
1144  VOP2e_Real_gfx10<0x001, "V_CNDMASK_B32", "v_cndmask_b32">;
1145
1146// VOP3 only.
1147defm V_BFM_B32            : VOP3Only_Real_gfx10<0x363>;
1148defm V_BCNT_U32_B32       : VOP3Only_Real_gfx10<0x364>;
1149defm V_MBCNT_LO_U32_B32   : VOP3Only_Real_gfx10<0x365>;
1150defm V_MBCNT_HI_U32_B32   : VOP3Only_Real_gfx10<0x366>;
1151defm V_LDEXP_F32          : VOP3Only_Real_gfx10<0x362>;
1152defm V_CVT_PKNORM_I16_F32 : VOP3Only_Real_gfx10<0x368>;
1153defm V_CVT_PKNORM_U16_F32 : VOP3Only_Real_gfx10<0x369>;
1154defm V_CVT_PK_U16_U32     : VOP3Only_Real_gfx10<0x36a>;
1155defm V_CVT_PK_I16_I32     : VOP3Only_Real_gfx10<0x36b>;
1156
1157// VOP3 carry-in, carry-out.
1158defm V_ADD_CO_U32 :
1159  VOP3beOnly_Real_gfx10<0x30f, "V_ADD_I32", "v_add_co_u32">;
1160defm V_SUB_CO_U32 :
1161  VOP3beOnly_Real_gfx10<0x310, "V_SUB_I32", "v_sub_co_u32">;
1162defm V_SUBREV_CO_U32 :
1163  VOP3beOnly_Real_gfx10<0x319, "V_SUBREV_I32", "v_subrev_co_u32">;
1164
1165let SubtargetPredicate = isGFX10Plus in {
1166  defm : VOP2eInstAliases<V_CNDMASK_B32_e32, V_CNDMASK_B32_e32_gfx10>;
1167
1168  defm : VOP2bInstAliases<
1169    V_ADDC_U32_e32, V_ADD_CO_CI_U32_e32_gfx10, "v_add_co_ci_u32">;
1170  defm : VOP2bInstAliases<
1171    V_SUBB_U32_e32, V_SUB_CO_CI_U32_e32_gfx10, "v_sub_co_ci_u32">;
1172  defm : VOP2bInstAliases<
1173    V_SUBBREV_U32_e32, V_SUBREV_CO_CI_U32_e32_gfx10, "v_subrev_co_ci_u32">;
1174} // End SubtargetPredicate = isGFX10Plus
1175
1176//===----------------------------------------------------------------------===//
1177// GFX6, GFX7, GFX10.
1178//===----------------------------------------------------------------------===//
1179
1180class VOP2_DPPe <bits<6> op, VOP2_DPP_Pseudo ps, VOPProfile P = ps.Pfl> :
1181  VOP_DPPe <P> {
1182  bits<8> vdst;
1183  bits<8> src1;
1184  let Inst{8-0}   = 0xfa; //dpp
1185  let Inst{16-9}  = !if(P.HasSrc1, src1{7-0}, 0);
1186  let Inst{24-17} = !if(P.EmitDst, vdst{7-0}, 0);
1187  let Inst{30-25} = op;
1188  let Inst{31}    = 0x0; //encoding
1189}
1190
1191let AssemblerPredicate = isGFX6GFX7, DecoderNamespace = "GFX6GFX7" in {
1192  multiclass VOP2Only_Real_gfx6_gfx7<bits<6> op> {
1193    def _gfx6_gfx7 :
1194      VOP2_Real<!cast<VOP2_Pseudo>(NAME), SIEncodingFamily.SI>,
1195      VOP2e<op{5-0}, !cast<VOP2_Pseudo>(NAME).Pfl>;
1196  }
1197  multiclass VOP2Only_Real_MADK_gfx6_gfx7<bits<6> op> {
1198    def _gfx6_gfx7 :
1199      VOP2_Real<!cast<VOP2_Pseudo>(NAME), SIEncodingFamily.SI>,
1200      VOP2_MADKe<op{5-0}, !cast<VOP2_Pseudo>(NAME).Pfl>;
1201  }
1202  multiclass VOP2_Real_e32_gfx6_gfx7<bits<6> op> {
1203    def _e32_gfx6_gfx7 :
1204      VOP2_Real<!cast<VOP2_Pseudo>(NAME#"_e32"), SIEncodingFamily.SI>,
1205      VOP2e<op{5-0}, !cast<VOP2_Pseudo>(NAME#"_e32").Pfl>;
1206  }
1207  multiclass VOP2_Real_e64_gfx6_gfx7<bits<6> op> {
1208    def _e64_gfx6_gfx7 :
1209      VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.SI>,
1210      VOP3e_gfx6_gfx7<{1, 0, 0, op{5-0}}, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>;
1211  }
1212  multiclass VOP2be_Real_e64_gfx6_gfx7<bits<6> op> {
1213    def _e64_gfx6_gfx7 :
1214      VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.SI>,
1215      VOP3be_gfx6_gfx7<{1, 0, 0, op{5-0}}, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>;
1216  }
1217} // End AssemblerPredicate = isGFX6GFX7, DecoderNamespace = "GFX6GFX7"
1218
1219multiclass VOP2Only_Real_MADK_gfx6_gfx7_gfx10<bits<6> op> :
1220  VOP2Only_Real_MADK_gfx6_gfx7<op>, VOP2Only_Real_MADK_gfx10<op>;
1221
1222multiclass VOP2_Real_gfx6_gfx7<bits<6> op> :
1223  VOP2_Real_e32_gfx6_gfx7<op>, VOP2_Real_e64_gfx6_gfx7<op>;
1224
1225multiclass VOP2_Real_gfx6_gfx7_gfx10<bits<6> op> :
1226  VOP2_Real_gfx6_gfx7<op>, VOP2_Real_gfx10<op>;
1227
1228multiclass VOP2be_Real_gfx6_gfx7<bits<6> op> :
1229  VOP2_Real_e32_gfx6_gfx7<op>, VOP2be_Real_e64_gfx6_gfx7<op>;
1230
1231defm V_CNDMASK_B32        : VOP2_Real_gfx6_gfx7<0x000>;
1232defm V_MIN_LEGACY_F32     : VOP2_Real_gfx6_gfx7<0x00d>;
1233defm V_MAX_LEGACY_F32     : VOP2_Real_gfx6_gfx7<0x00e>;
1234defm V_LSHR_B32           : VOP2_Real_gfx6_gfx7<0x015>;
1235defm V_ASHR_I32           : VOP2_Real_gfx6_gfx7<0x017>;
1236defm V_LSHL_B32           : VOP2_Real_gfx6_gfx7<0x019>;
1237defm V_BFM_B32            : VOP2_Real_gfx6_gfx7<0x01e>;
1238defm V_BCNT_U32_B32       : VOP2_Real_gfx6_gfx7<0x022>;
1239defm V_MBCNT_LO_U32_B32   : VOP2_Real_gfx6_gfx7<0x023>;
1240defm V_MBCNT_HI_U32_B32   : VOP2_Real_gfx6_gfx7<0x024>;
1241defm V_LDEXP_F32          : VOP2_Real_gfx6_gfx7<0x02b>;
1242defm V_CVT_PKACCUM_U8_F32 : VOP2_Real_gfx6_gfx7<0x02c>;
1243defm V_CVT_PKNORM_I16_F32 : VOP2_Real_gfx6_gfx7<0x02d>;
1244defm V_CVT_PKNORM_U16_F32 : VOP2_Real_gfx6_gfx7<0x02e>;
1245defm V_CVT_PK_U16_U32     : VOP2_Real_gfx6_gfx7<0x030>;
1246defm V_CVT_PK_I16_I32     : VOP2_Real_gfx6_gfx7<0x031>;
1247defm V_ADD_I32            : VOP2be_Real_gfx6_gfx7<0x025>;
1248defm V_SUB_I32            : VOP2be_Real_gfx6_gfx7<0x026>;
1249defm V_SUBREV_I32         : VOP2be_Real_gfx6_gfx7<0x027>;
1250defm V_ADDC_U32           : VOP2be_Real_gfx6_gfx7<0x028>;
1251defm V_SUBB_U32           : VOP2be_Real_gfx6_gfx7<0x029>;
1252defm V_SUBBREV_U32        : VOP2be_Real_gfx6_gfx7<0x02a>;
1253
1254defm V_READLANE_B32 : VOP2Only_Real_gfx6_gfx7<0x001>;
1255
1256let InOperandList = (ins SSrcOrLds_b32:$src0, SCSrc_b32:$src1, VSrc_b32:$vdst_in) in {
1257  defm V_WRITELANE_B32 : VOP2Only_Real_gfx6_gfx7<0x002>;
1258} // End InOperandList = (ins SSrcOrLds_b32:$src0, SCSrc_b32:$src1, VSrc_b32:$vdst_in)
1259
1260let SubtargetPredicate = isGFX6GFX7 in {
1261  defm : VOP2eInstAliases<V_CNDMASK_B32_e32, V_CNDMASK_B32_e32_gfx6_gfx7>;
1262} // End SubtargetPredicate = isGFX6GFX7
1263
1264defm V_ADD_F32            : VOP2_Real_gfx6_gfx7_gfx10<0x003>;
1265defm V_SUB_F32            : VOP2_Real_gfx6_gfx7_gfx10<0x004>;
1266defm V_SUBREV_F32         : VOP2_Real_gfx6_gfx7_gfx10<0x005>;
1267defm V_MAC_LEGACY_F32     : VOP2_Real_gfx6_gfx7_gfx10<0x006>;
1268defm V_MUL_LEGACY_F32     : VOP2_Real_gfx6_gfx7_gfx10<0x007>;
1269defm V_MUL_F32            : VOP2_Real_gfx6_gfx7_gfx10<0x008>;
1270defm V_MUL_I32_I24        : VOP2_Real_gfx6_gfx7_gfx10<0x009>;
1271defm V_MUL_HI_I32_I24     : VOP2_Real_gfx6_gfx7_gfx10<0x00a>;
1272defm V_MUL_U32_U24        : VOP2_Real_gfx6_gfx7_gfx10<0x00b>;
1273defm V_MUL_HI_U32_U24     : VOP2_Real_gfx6_gfx7_gfx10<0x00c>;
1274defm V_MIN_F32            : VOP2_Real_gfx6_gfx7_gfx10<0x00f>;
1275defm V_MAX_F32            : VOP2_Real_gfx6_gfx7_gfx10<0x010>;
1276defm V_MIN_I32            : VOP2_Real_gfx6_gfx7_gfx10<0x011>;
1277defm V_MAX_I32            : VOP2_Real_gfx6_gfx7_gfx10<0x012>;
1278defm V_MIN_U32            : VOP2_Real_gfx6_gfx7_gfx10<0x013>;
1279defm V_MAX_U32            : VOP2_Real_gfx6_gfx7_gfx10<0x014>;
1280defm V_LSHRREV_B32        : VOP2_Real_gfx6_gfx7_gfx10<0x016>;
1281defm V_ASHRREV_I32        : VOP2_Real_gfx6_gfx7_gfx10<0x018>;
1282defm V_LSHLREV_B32        : VOP2_Real_gfx6_gfx7_gfx10<0x01a>;
1283defm V_AND_B32            : VOP2_Real_gfx6_gfx7_gfx10<0x01b>;
1284defm V_OR_B32             : VOP2_Real_gfx6_gfx7_gfx10<0x01c>;
1285defm V_XOR_B32            : VOP2_Real_gfx6_gfx7_gfx10<0x01d>;
1286defm V_MAC_F32            : VOP2_Real_gfx6_gfx7_gfx10<0x01f>;
1287defm V_CVT_PKRTZ_F16_F32  : VOP2_Real_gfx6_gfx7_gfx10<0x02f>;
1288defm V_MADMK_F32          : VOP2Only_Real_MADK_gfx6_gfx7_gfx10<0x020>;
1289defm V_MADAK_F32          : VOP2Only_Real_MADK_gfx6_gfx7_gfx10<0x021>;
1290
1291//===----------------------------------------------------------------------===//
1292// GFX8, GFX9 (VI).
1293//===----------------------------------------------------------------------===//
1294
1295let AssemblerPredicate = isGFX8GFX9, DecoderNamespace = "GFX8" in {
1296
1297multiclass VOP2_Real_MADK_vi <bits<6> op> {
1298  def _vi : VOP2_Real<!cast<VOP2_Pseudo>(NAME), SIEncodingFamily.VI>,
1299            VOP2_MADKe<op{5-0}, !cast<VOP2_Pseudo>(NAME).Pfl>;
1300}
1301
1302multiclass VOP2_Real_e32_vi <bits<6> op> {
1303  def _e32_vi :
1304    VOP2_Real<!cast<VOP2_Pseudo>(NAME#"_e32"), SIEncodingFamily.VI>,
1305    VOP2e<op{5-0}, !cast<VOP2_Pseudo>(NAME#"_e32").Pfl>;
1306}
1307
1308multiclass VOP2_Real_e64_vi <bits<10> op> {
1309  def _e64_vi :
1310    VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>,
1311    VOP3e_vi <op, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>;
1312}
1313
1314multiclass VOP2_Real_e64only_vi <bits<10> op> {
1315  def _e64_vi :
1316    VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>,
1317    VOP3e_vi <op, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl> {
1318      // Hack to stop printing _e64
1319      VOP3_Pseudo ps = !cast<VOP3_Pseudo>(NAME#"_e64");
1320      let OutOperandList = (outs VGPR_32:$vdst);
1321      let AsmString = ps.Mnemonic # " " # ps.AsmOperands;
1322    }
1323}
1324
1325multiclass Base_VOP2_Real_e32e64_vi <bits<6> op> :
1326  VOP2_Real_e32_vi<op>,
1327  VOP2_Real_e64_vi<{0, 1, 0, 0, op{5-0}}>;
1328
1329} // End AssemblerPredicate = isGFX8GFX9, DecoderNamespace = "GFX8"
1330
1331multiclass VOP2_SDWA_Real <bits<6> op> {
1332  foreach _ = BoolToList<!cast<VOP2_Pseudo>(NAME#"_e32").Pfl.HasExtSDWA>.ret in
1333  def _sdwa_vi :
1334    VOP_SDWA_Real <!cast<VOP2_SDWA_Pseudo>(NAME#"_sdwa")>,
1335    VOP2_SDWAe <op{5-0}, !cast<VOP2_SDWA_Pseudo>(NAME#"_sdwa").Pfl>;
1336}
1337
1338multiclass VOP2_SDWA9_Real <bits<6> op> {
1339  foreach _ = BoolToList<!cast<VOP2_Pseudo>(NAME#"_e32").Pfl.HasExtSDWA9>.ret in
1340  def _sdwa_gfx9 :
1341    VOP_SDWA9_Real <!cast<VOP2_SDWA_Pseudo>(NAME#"_sdwa")>,
1342    VOP2_SDWA9Ae <op{5-0}, !cast<VOP2_SDWA_Pseudo>(NAME#"_sdwa").Pfl>;
1343}
1344
1345let AssemblerPredicate = isGFX8Only in {
1346
1347multiclass VOP2be_Real_e32e64_vi_only <bits<6> op, string OpName, string AsmName> {
1348  def _e32_vi :
1349    VOP2_Real<!cast<VOP2_Pseudo>(OpName#"_e32"), SIEncodingFamily.VI>,
1350    VOP2e<op{5-0}, !cast<VOP2_Pseudo>(OpName#"_e32").Pfl> {
1351      VOP2_Pseudo ps = !cast<VOP2_Pseudo>(OpName#"_e32");
1352      let AsmString = AsmName # ps.AsmOperands;
1353      let DecoderNamespace = "GFX8";
1354    }
1355  def _e64_vi :
1356    VOP3_Real<!cast<VOP3_Pseudo>(OpName#"_e64"), SIEncodingFamily.VI>,
1357    VOP3be_vi <{0, 1, 0, 0, op{5-0}}, !cast<VOP3_Pseudo>(OpName#"_e64").Pfl> {
1358      VOP3_Pseudo ps = !cast<VOP3_Pseudo>(OpName#"_e64");
1359      let AsmString = AsmName # ps.AsmOperands;
1360      let DecoderNamespace = "GFX8";
1361    }
1362  foreach _ = BoolToList<!cast<VOP2_Pseudo>(OpName#"_e32").Pfl.HasExtSDWA>.ret in
1363    def _sdwa_vi :
1364      VOP_SDWA_Real <!cast<VOP2_SDWA_Pseudo>(OpName#"_sdwa")>,
1365      VOP2_SDWAe <op{5-0}, !cast<VOP2_SDWA_Pseudo>(OpName#"_sdwa").Pfl> {
1366        VOP2_SDWA_Pseudo ps = !cast<VOP2_SDWA_Pseudo>(OpName#"_sdwa");
1367        let AsmString = AsmName # ps.AsmOperands;
1368      }
1369  foreach _ = BoolToList<!cast<VOP2_Pseudo>(OpName#"_e32").Pfl.HasExtDPP>.ret in
1370    def _dpp_vi :
1371      VOP_DPP_Real<!cast<VOP2_DPP_Pseudo>(OpName#"_dpp"), SIEncodingFamily.VI>,
1372      VOP2_DPPe<op, !cast<VOP2_DPP_Pseudo>(OpName#"_dpp")> {
1373        VOP2_DPP_Pseudo ps = !cast<VOP2_DPP_Pseudo>(OpName#"_dpp");
1374        let AsmString = AsmName # ps.AsmOperands;
1375      }
1376}
1377}
1378
1379let AssemblerPredicate = isGFX9Only in {
1380
1381multiclass VOP2be_Real_e32e64_gfx9 <bits<6> op, string OpName, string AsmName> {
1382  def _e32_gfx9 :
1383    VOP2_Real<!cast<VOP2_Pseudo>(OpName#"_e32"), SIEncodingFamily.GFX9>,
1384    VOP2e<op{5-0}, !cast<VOP2_Pseudo>(OpName#"_e32").Pfl> {
1385      VOP2_Pseudo ps = !cast<VOP2_Pseudo>(OpName#"_e32");
1386      let AsmString = AsmName # ps.AsmOperands;
1387      let DecoderNamespace = "GFX9";
1388    }
1389  def _e64_gfx9 :
1390    VOP3_Real<!cast<VOP3_Pseudo>(OpName#"_e64"), SIEncodingFamily.GFX9>,
1391    VOP3be_vi <{0, 1, 0, 0, op{5-0}}, !cast<VOP3_Pseudo>(OpName#"_e64").Pfl> {
1392      VOP3_Pseudo ps = !cast<VOP3_Pseudo>(OpName#"_e64");
1393      let AsmString = AsmName # ps.AsmOperands;
1394      let DecoderNamespace = "GFX9";
1395    }
1396  foreach _ = BoolToList<!cast<VOP2_Pseudo>(OpName#"_e32").Pfl.HasExtSDWA9>.ret in
1397    def _sdwa_gfx9 :
1398      VOP_SDWA9_Real <!cast<VOP2_SDWA_Pseudo>(OpName#"_sdwa")>,
1399      VOP2_SDWA9Ae <op{5-0}, !cast<VOP2_SDWA_Pseudo>(OpName#"_sdwa").Pfl> {
1400        VOP2_SDWA_Pseudo ps = !cast<VOP2_SDWA_Pseudo>(OpName#"_sdwa");
1401        let AsmString = AsmName # ps.AsmOperands;
1402      }
1403  foreach _ = BoolToList<!cast<VOP2_Pseudo>(OpName#"_e32").Pfl.HasExtDPP>.ret in
1404    def _dpp_gfx9 :
1405      VOP_DPP_Real<!cast<VOP2_DPP_Pseudo>(OpName#"_dpp"), SIEncodingFamily.GFX9>,
1406      VOP2_DPPe<op, !cast<VOP2_DPP_Pseudo>(OpName#"_dpp")> {
1407        VOP2_DPP_Pseudo ps = !cast<VOP2_DPP_Pseudo>(OpName#"_dpp");
1408        let AsmString = AsmName # ps.AsmOperands;
1409        let DecoderNamespace = "SDWA9";
1410      }
1411}
1412
1413multiclass VOP2_Real_e32e64_gfx9 <bits<6> op> {
1414  def _e32_gfx9 :
1415    VOP2_Real<!cast<VOP2_Pseudo>(NAME#"_e32"), SIEncodingFamily.GFX9>,
1416    VOP2e<op{5-0}, !cast<VOP2_Pseudo>(NAME#"_e32").Pfl>{
1417      let DecoderNamespace = "GFX9";
1418    }
1419  def _e64_gfx9 :
1420    VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.GFX9>,
1421    VOP3e_vi <{0, 1, 0, 0, op{5-0}}, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl> {
1422      let DecoderNamespace = "GFX9";
1423    }
1424  foreach _ = BoolToList<!cast<VOP2_Pseudo>(NAME#"_e32").Pfl.HasExtSDWA9>.ret in
1425    def _sdwa_gfx9 :
1426      VOP_SDWA9_Real <!cast<VOP2_SDWA_Pseudo>(NAME#"_sdwa")>,
1427      VOP2_SDWA9Ae <op{5-0}, !cast<VOP2_SDWA_Pseudo>(NAME#"_sdwa").Pfl> {
1428      }
1429  foreach _ = BoolToList<!cast<VOP2_Pseudo>(NAME#"_e32").Pfl.HasExtDPP>.ret in
1430    def _dpp_gfx9 :
1431      VOP_DPP_Real<!cast<VOP2_DPP_Pseudo>(NAME#"_dpp"), SIEncodingFamily.GFX9>,
1432      VOP2_DPPe<op, !cast<VOP2_DPP_Pseudo>(NAME#"_dpp")> {
1433        let DecoderNamespace = "SDWA9";
1434      }
1435}
1436
1437} // AssemblerPredicate = isGFX9Only
1438
1439multiclass VOP2_Real_e32e64_vi <bits<6> op> :
1440  Base_VOP2_Real_e32e64_vi<op>, VOP2_SDWA_Real<op>, VOP2_SDWA9_Real<op> {
1441
1442  foreach _ = BoolToList<!cast<VOP2_Pseudo>(NAME#"_e32").Pfl.HasExtDPP>.ret in
1443    def _dpp_vi :
1444      VOP_DPP_Real<!cast<VOP2_DPP_Pseudo>(NAME#"_dpp"), SIEncodingFamily.VI>,
1445      VOP2_DPPe<op, !cast<VOP2_DPP_Pseudo>(NAME#"_dpp")>;
1446}
1447
1448defm V_CNDMASK_B32        : VOP2_Real_e32e64_vi <0x0>;
1449defm V_ADD_F32            : VOP2_Real_e32e64_vi <0x1>;
1450defm V_SUB_F32            : VOP2_Real_e32e64_vi <0x2>;
1451defm V_SUBREV_F32         : VOP2_Real_e32e64_vi <0x3>;
1452defm V_MUL_LEGACY_F32     : VOP2_Real_e32e64_vi <0x4>;
1453defm V_MUL_F32            : VOP2_Real_e32e64_vi <0x5>;
1454defm V_MUL_I32_I24        : VOP2_Real_e32e64_vi <0x6>;
1455defm V_MUL_HI_I32_I24     : VOP2_Real_e32e64_vi <0x7>;
1456defm V_MUL_U32_U24        : VOP2_Real_e32e64_vi <0x8>;
1457defm V_MUL_HI_U32_U24     : VOP2_Real_e32e64_vi <0x9>;
1458defm V_MIN_F32            : VOP2_Real_e32e64_vi <0xa>;
1459defm V_MAX_F32            : VOP2_Real_e32e64_vi <0xb>;
1460defm V_MIN_I32            : VOP2_Real_e32e64_vi <0xc>;
1461defm V_MAX_I32            : VOP2_Real_e32e64_vi <0xd>;
1462defm V_MIN_U32            : VOP2_Real_e32e64_vi <0xe>;
1463defm V_MAX_U32            : VOP2_Real_e32e64_vi <0xf>;
1464defm V_LSHRREV_B32        : VOP2_Real_e32e64_vi <0x10>;
1465defm V_ASHRREV_I32        : VOP2_Real_e32e64_vi <0x11>;
1466defm V_LSHLREV_B32        : VOP2_Real_e32e64_vi <0x12>;
1467defm V_AND_B32            : VOP2_Real_e32e64_vi <0x13>;
1468defm V_OR_B32             : VOP2_Real_e32e64_vi <0x14>;
1469defm V_XOR_B32            : VOP2_Real_e32e64_vi <0x15>;
1470defm V_MAC_F32            : VOP2_Real_e32e64_vi <0x16>;
1471defm V_MADMK_F32          : VOP2_Real_MADK_vi <0x17>;
1472defm V_MADAK_F32          : VOP2_Real_MADK_vi <0x18>;
1473
1474defm V_ADD_U32            : VOP2be_Real_e32e64_vi_only <0x19, "V_ADD_I32",     "v_add_u32">;
1475defm V_SUB_U32            : VOP2be_Real_e32e64_vi_only <0x1a, "V_SUB_I32",     "v_sub_u32">;
1476defm V_SUBREV_U32         : VOP2be_Real_e32e64_vi_only <0x1b, "V_SUBREV_I32",  "v_subrev_u32">;
1477defm V_ADDC_U32           : VOP2be_Real_e32e64_vi_only <0x1c, "V_ADDC_U32",    "v_addc_u32">;
1478defm V_SUBB_U32           : VOP2be_Real_e32e64_vi_only <0x1d, "V_SUBB_U32",    "v_subb_u32">;
1479defm V_SUBBREV_U32        : VOP2be_Real_e32e64_vi_only <0x1e, "V_SUBBREV_U32", "v_subbrev_u32">;
1480
1481defm V_ADD_CO_U32         : VOP2be_Real_e32e64_gfx9 <0x19, "V_ADD_I32",     "v_add_co_u32">;
1482defm V_SUB_CO_U32         : VOP2be_Real_e32e64_gfx9 <0x1a, "V_SUB_I32",     "v_sub_co_u32">;
1483defm V_SUBREV_CO_U32      : VOP2be_Real_e32e64_gfx9 <0x1b, "V_SUBREV_I32",  "v_subrev_co_u32">;
1484defm V_ADDC_CO_U32        : VOP2be_Real_e32e64_gfx9 <0x1c, "V_ADDC_U32",    "v_addc_co_u32">;
1485defm V_SUBB_CO_U32        : VOP2be_Real_e32e64_gfx9 <0x1d, "V_SUBB_U32",    "v_subb_co_u32">;
1486defm V_SUBBREV_CO_U32     : VOP2be_Real_e32e64_gfx9 <0x1e, "V_SUBBREV_U32", "v_subbrev_co_u32">;
1487
1488defm V_ADD_U32            : VOP2_Real_e32e64_gfx9 <0x34>;
1489defm V_SUB_U32            : VOP2_Real_e32e64_gfx9 <0x35>;
1490defm V_SUBREV_U32         : VOP2_Real_e32e64_gfx9 <0x36>;
1491
1492defm V_BFM_B32            : VOP2_Real_e64only_vi <0x293>;
1493defm V_BCNT_U32_B32       : VOP2_Real_e64only_vi <0x28b>;
1494defm V_MBCNT_LO_U32_B32   : VOP2_Real_e64only_vi <0x28c>;
1495defm V_MBCNT_HI_U32_B32   : VOP2_Real_e64only_vi <0x28d>;
1496defm V_LDEXP_F32          : VOP2_Real_e64only_vi <0x288>;
1497defm V_CVT_PKACCUM_U8_F32 : VOP2_Real_e64only_vi <0x1f0>;
1498defm V_CVT_PKNORM_I16_F32 : VOP2_Real_e64only_vi <0x294>;
1499defm V_CVT_PKNORM_U16_F32 : VOP2_Real_e64only_vi <0x295>;
1500defm V_CVT_PKRTZ_F16_F32  : VOP2_Real_e64only_vi <0x296>;
1501defm V_CVT_PK_U16_U32     : VOP2_Real_e64only_vi <0x297>;
1502defm V_CVT_PK_I16_I32     : VOP2_Real_e64only_vi <0x298>;
1503
1504defm V_ADD_F16            : VOP2_Real_e32e64_vi <0x1f>;
1505defm V_SUB_F16            : VOP2_Real_e32e64_vi <0x20>;
1506defm V_SUBREV_F16         : VOP2_Real_e32e64_vi <0x21>;
1507defm V_MUL_F16            : VOP2_Real_e32e64_vi <0x22>;
1508defm V_MAC_F16            : VOP2_Real_e32e64_vi <0x23>;
1509defm V_MADMK_F16          : VOP2_Real_MADK_vi <0x24>;
1510defm V_MADAK_F16          : VOP2_Real_MADK_vi <0x25>;
1511defm V_ADD_U16            : VOP2_Real_e32e64_vi <0x26>;
1512defm V_SUB_U16            : VOP2_Real_e32e64_vi <0x27>;
1513defm V_SUBREV_U16         : VOP2_Real_e32e64_vi <0x28>;
1514defm V_MUL_LO_U16         : VOP2_Real_e32e64_vi <0x29>;
1515defm V_LSHLREV_B16        : VOP2_Real_e32e64_vi <0x2a>;
1516defm V_LSHRREV_B16        : VOP2_Real_e32e64_vi <0x2b>;
1517defm V_ASHRREV_I16        : VOP2_Real_e32e64_vi <0x2c>;
1518defm V_MAX_F16            : VOP2_Real_e32e64_vi <0x2d>;
1519defm V_MIN_F16            : VOP2_Real_e32e64_vi <0x2e>;
1520defm V_MAX_U16            : VOP2_Real_e32e64_vi <0x2f>;
1521defm V_MAX_I16            : VOP2_Real_e32e64_vi <0x30>;
1522defm V_MIN_U16            : VOP2_Real_e32e64_vi <0x31>;
1523defm V_MIN_I16            : VOP2_Real_e32e64_vi <0x32>;
1524defm V_LDEXP_F16          : VOP2_Real_e32e64_vi <0x33>;
1525
1526let SubtargetPredicate = isGFX8GFX9 in {
1527
1528// Aliases to simplify matching of floating-point instructions that
1529// are VOP2 on SI and VOP3 on VI.
1530class SI2_VI3Alias <string name, VOP3_Real inst> : InstAlias <
1531  name#" $dst, $src0, $src1",
1532  !if(inst.Pfl.HasOMod,
1533      (inst VGPR_32:$dst, 0, VCSrc_f32:$src0, 0, VCSrc_f32:$src1, 0, 0),
1534      (inst VGPR_32:$dst, 0, VCSrc_f32:$src0, 0, VCSrc_f32:$src1, 0))
1535>, PredicateControl {
1536  let UseInstAsmMatchConverter = 0;
1537  let AsmVariantName = AMDGPUAsmVariants.VOP3;
1538}
1539
1540def : SI2_VI3Alias <"v_ldexp_f32", V_LDEXP_F32_e64_vi>;
1541def : SI2_VI3Alias <"v_cvt_pkaccum_u8_f32", V_CVT_PKACCUM_U8_F32_e64_vi>;
1542def : SI2_VI3Alias <"v_cvt_pknorm_i16_f32", V_CVT_PKNORM_I16_F32_e64_vi>;
1543def : SI2_VI3Alias <"v_cvt_pknorm_u16_f32", V_CVT_PKNORM_U16_F32_e64_vi>;
1544def : SI2_VI3Alias <"v_cvt_pkrtz_f16_f32", V_CVT_PKRTZ_F16_F32_e64_vi>;
1545
1546defm : VOP2eInstAliases<V_CNDMASK_B32_e32, V_CNDMASK_B32_e32_vi>;
1547
1548} // End SubtargetPredicate = isGFX8GFX9
1549
1550let SubtargetPredicate = isGFX9Only in {
1551
1552defm : VOP2bInstAliases<V_ADD_I32_e32,     V_ADD_CO_U32_e32_gfx9,     "v_add_co_u32">;
1553defm : VOP2bInstAliases<V_ADDC_U32_e32,    V_ADDC_CO_U32_e32_gfx9,    "v_addc_co_u32">;
1554defm : VOP2bInstAliases<V_SUB_I32_e32,     V_SUB_CO_U32_e32_gfx9,     "v_sub_co_u32">;
1555defm : VOP2bInstAliases<V_SUBB_U32_e32,    V_SUBB_CO_U32_e32_gfx9,    "v_subb_co_u32">;
1556defm : VOP2bInstAliases<V_SUBREV_I32_e32,  V_SUBREV_CO_U32_e32_gfx9,  "v_subrev_co_u32">;
1557defm : VOP2bInstAliases<V_SUBBREV_U32_e32, V_SUBBREV_CO_U32_e32_gfx9, "v_subbrev_co_u32">;
1558
1559} // End SubtargetPredicate = isGFX9Only
1560
1561let SubtargetPredicate = HasDLInsts in {
1562
1563defm V_FMAC_F32 : VOP2_Real_e32e64_vi <0x3b>;
1564defm V_XNOR_B32 : VOP2_Real_e32e64_vi <0x3d>;
1565
1566} // End SubtargetPredicate = HasDLInsts
1567
1568multiclass VOP2_Real_DOT_ACC_gfx9<bits<6> op> : VOP2_Real_e32_vi<op> {
1569  def _dpp_vi : VOP2_DPP<op, !cast<VOP2_DPP_Pseudo>(NAME#"_dpp")>;
1570}
1571
1572multiclass VOP2_Real_DOT_ACC_gfx10<bits<6> op> :
1573  VOP2_Real_e32_gfx10<op>,
1574  VOP2_Real_dpp_gfx10<op>,
1575  VOP2_Real_dpp8_gfx10<op>;
1576
1577let SubtargetPredicate = HasDot5Insts in {
1578  defm V_DOT2C_F32_F16 : VOP2_Real_DOT_ACC_gfx9<0x37>;
1579  // NB: Opcode conflicts with V_DOT8C_I32_I4
1580  // This opcode exists in gfx 10.1* only
1581  defm V_DOT2C_F32_F16 : VOP2_Real_DOT_ACC_gfx10<0x02>;
1582}
1583
1584let SubtargetPredicate = HasDot6Insts in {
1585  defm V_DOT4C_I32_I8  : VOP2_Real_DOT_ACC_gfx9<0x39>;
1586  defm V_DOT4C_I32_I8  : VOP2_Real_DOT_ACC_gfx10<0x0d>;
1587}
1588
1589let SubtargetPredicate = HasDot4Insts in {
1590  defm V_DOT2C_I32_I16 : VOP2_Real_DOT_ACC_gfx9<0x38>;
1591}
1592let SubtargetPredicate = HasDot3Insts in {
1593  defm V_DOT8C_I32_I4  : VOP2_Real_DOT_ACC_gfx9<0x3a>;
1594}
1595
1596let SubtargetPredicate = HasPkFmacF16Inst in {
1597defm V_PK_FMAC_F16 : VOP2_Real_e32_vi<0x3c>;
1598} // End SubtargetPredicate = HasPkFmacF16Inst
1599