1 //=- HexagonInstrInfoV5.td - Target Desc. for Hexagon Target -*- tablegen -*-=//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the Hexagon V5 instructions in TableGen format.
12 //===----------------------------------------------------------------------===//
14 //===----------------------------------------------------------------------===//
16 //===----------------------------------------------------------------------===//
18 // Vector multiply bytes
19 // Rdd=vmpyb[s]u(Rs,Rt)
20 let Predicates = [HasV5T], isCodeGenOnly = 0 in {
21 // Rd=vaddhub(Rss,Rtt):sat
22 let hasNewValue = 1, opNewValue = 0 in
23 def A5_vaddhubs: T_S3op_1 <"vaddhub", IntRegs, 0b01, 0b001, 0, 1>;
26 let isCodeGenOnly = 0 in
27 def S2_asr_i_p_rnd : S_2OpInstImm<"asr", 0b110, 0b111, u6Imm,
29 (sra (i64 (add (i64 (sra I64:$src1, u6ImmPred:$src2)), 1)),
33 let Inst{13-8} = src2;
36 def S2_asr_i_p_rnd_goodsyntax
37 : MInst<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1, u6Imm:$src2),
38 "$dst = asrrnd($src1, #$src2)">;
40 let isCodeGenOnly = 0 in
41 def C4_fastcorner9 : T_LOGICAL_2OP<"fastcorner9", 0b000, 0, 0>,
43 let Inst{13,7,4} = 0b111;
46 let isCodeGenOnly = 0 in
47 def C4_fastcorner9_not : T_LOGICAL_2OP<"!fastcorner9", 0b000, 0, 0>,
49 let Inst{20,13,7,4} = 0b1111;
52 def SDTHexagonFCONST32 : SDTypeProfile<1, 1, [
55 def HexagonFCONST32 : SDNode<"HexagonISD::FCONST32", SDTHexagonFCONST32>;
57 let isReMaterializable = 1, isMoveImm = 1 in
58 def FCONST32_nsdata : LDInst<(outs IntRegs:$dst), (ins globaladdress:$global),
59 "$dst = CONST32(#$global)",
60 [(set (f32 IntRegs:$dst),
61 (HexagonFCONST32 tglobaladdr:$global))]>,
64 let isReMaterializable = 1, isMoveImm = 1 in
65 def CONST64_Float_Real : LDInst<(outs DoubleRegs:$dst), (ins f64imm:$src1),
66 "$dst = CONST64(#$src1)",
67 [(set DoubleRegs:$dst, fpimm:$src1)]>,
70 let isReMaterializable = 1, isMoveImm = 1 in
71 def CONST32_Float_Real : LDInst<(outs IntRegs:$dst), (ins f32imm:$src1),
72 "$dst = CONST32(#$src1)",
73 [(set IntRegs:$dst, fpimm:$src1)]>,
76 // Transfer immediate float.
77 // Only works with single precision fp value.
78 // For double precision, use CONST64_float_real, as 64bit transfer
79 // can only hold 40-bit values - 32 from const ext + 8 bit immediate.
80 // Make sure that complexity is more than the CONST32 pattern in
81 // HexagonInstrInfo.td patterns.
82 let isExtended = 1, opExtendable = 1, isMoveImm = 1, isReMaterializable = 1,
83 isPredicable = 1, AddedComplexity = 30, validSubTargets = HasV5SubT,
85 def TFRI_f : ALU32_ri<(outs IntRegs:$dst), (ins f32Ext:$src1),
87 [(set IntRegs:$dst, fpimm:$src1)]>,
90 let isExtended = 1, opExtendable = 2, isPredicated = 1,
91 hasSideEffects = 0, validSubTargets = HasV5SubT in
92 def TFRI_cPt_f : ALU32_ri<(outs IntRegs:$dst),
93 (ins PredRegs:$src1, f32Ext:$src2),
94 "if ($src1) $dst = #$src2",
98 let isExtended = 1, opExtendable = 2, isPredicated = 1, isPredicatedFalse = 1,
99 hasSideEffects = 0, validSubTargets = HasV5SubT in
100 def TFRI_cNotPt_f : ALU32_ri<(outs IntRegs:$dst),
101 (ins PredRegs:$src1, f32Ext:$src2),
102 "if (!$src1) $dst =#$src2",
106 def SDTHexagonI32I64: SDTypeProfile<1, 1, [SDTCisVT<0, i32>,
109 def HexagonPOPCOUNT: SDNode<"HexagonISD::POPCOUNT", SDTHexagonI32I64>;
111 let hasNewValue = 1, validSubTargets = HasV5SubT, isCodeGenOnly = 0 in
112 def S5_popcountp : ALU64_rr<(outs IntRegs:$Rd), (ins DoubleRegs:$Rss),
113 "$Rd = popcount($Rss)",
114 [(set I32:$Rd, (HexagonPOPCOUNT I64:$Rss))], "", S_2op_tc_2_SLOT23>,
121 let Inst{27-21} = 0b1000011;
122 let Inst{7-5} = 0b011;
124 let Inst{20-16} = Rss;
127 defm: Loadx_pat<load, f32, s11_2ExtPred, L2_loadri_io>;
128 defm: Loadx_pat<load, f64, s11_3ExtPred, L2_loadrd_io>;
130 defm: Storex_pat<store, F32, s11_2ExtPred, S2_storeri_io>;
131 defm: Storex_pat<store, F64, s11_3ExtPred, S2_storerd_io>;
133 let isFP = 1, hasNewValue = 1, opNewValue = 0 in
134 class T_MInstFloat <string mnemonic, bits<3> MajOp, bits<3> MinOp>
135 : MInst<(outs IntRegs:$Rd),
136 (ins IntRegs:$Rs, IntRegs:$Rt),
137 "$Rd = "#mnemonic#"($Rs, $Rt)", [],
138 "" , M_tc_3or4x_SLOT23 > ,
146 let Inst{27-24} = 0b1011;
147 let Inst{23-21} = MajOp;
148 let Inst{20-16} = Rs;
151 let Inst{7-5} = MinOp;
155 let isCommutable = 1, isCodeGenOnly = 0 in {
156 def F2_sfadd : T_MInstFloat < "sfadd", 0b000, 0b000>;
157 def F2_sfmpy : T_MInstFloat < "sfmpy", 0b010, 0b000>;
160 let isCodeGenOnly = 0 in
161 def F2_sfsub : T_MInstFloat < "sfsub", 0b000, 0b001>;
163 def: Pat<(f32 (fadd F32:$src1, F32:$src2)),
164 (F2_sfadd F32:$src1, F32:$src2)>;
166 def: Pat<(f32 (fsub F32:$src1, F32:$src2)),
167 (F2_sfsub F32:$src1, F32:$src2)>;
169 def: Pat<(f32 (fmul F32:$src1, F32:$src2)),
170 (F2_sfmpy F32:$src1, F32:$src2)>;
172 let Itinerary = M_tc_3x_SLOT23, isCodeGenOnly = 0 in {
173 def F2_sfmax : T_MInstFloat < "sfmax", 0b100, 0b000>;
174 def F2_sfmin : T_MInstFloat < "sfmin", 0b100, 0b001>;
177 let isCodeGenOnly = 0 in {
178 def F2_sffixupn : T_MInstFloat < "sffixupn", 0b110, 0b000>;
179 def F2_sffixupd : T_MInstFloat < "sffixupd", 0b110, 0b001>;
182 // F2_sfrecipa: Reciprocal approximation for division.
183 let isPredicateLate = 1, isFP = 1,
184 hasSideEffects = 0, hasNewValue = 1, isCodeGenOnly = 0 in
185 def F2_sfrecipa: MInst <
186 (outs IntRegs:$Rd, PredRegs:$Pe),
187 (ins IntRegs:$Rs, IntRegs:$Rt),
188 "$Rd, $Pe = sfrecipa($Rs, $Rt)">,
196 let Inst{27-21} = 0b1011111;
197 let Inst{20-16} = Rs;
205 // F2_dfcmpeq: Floating point compare for equal.
206 let isCompare = 1, isFP = 1 in
207 class T_fcmp <string mnemonic, RegisterClass RC, bits<3> MinOp,
208 list<dag> pattern = [] >
209 : ALU64Inst <(outs PredRegs:$dst), (ins RC:$src1, RC:$src2),
210 "$dst = "#mnemonic#"($src1, $src2)", pattern,
211 "" , ALU64_tc_2early_SLOT23 > ,
219 let Inst{27-21} = 0b0010111;
220 let Inst{20-16} = src1;
221 let Inst{12-8} = src2;
222 let Inst{7-5} = MinOp;
226 class T_fcmp64 <string mnemonic, PatFrag OpNode, bits<3> MinOp>
227 : T_fcmp <mnemonic, DoubleRegs, MinOp,
228 [(set I1:$dst, (OpNode F64:$src1, F64:$src2))]> {
230 let Inst{27-21} = 0b0010111;
233 class T_fcmp32 <string mnemonic, PatFrag OpNode, bits<3> MinOp>
234 : T_fcmp <mnemonic, IntRegs, MinOp,
235 [(set I1:$dst, (OpNode F32:$src1, F32:$src2))]> {
237 let Inst{27-21} = 0b0111111;
240 let isCodeGenOnly = 0 in {
241 def F2_dfcmpeq : T_fcmp64<"dfcmp.eq", setoeq, 0b000>;
242 def F2_dfcmpgt : T_fcmp64<"dfcmp.gt", setogt, 0b001>;
243 def F2_dfcmpge : T_fcmp64<"dfcmp.ge", setoge, 0b010>;
244 def F2_dfcmpuo : T_fcmp64<"dfcmp.uo", setuo, 0b011>;
246 def F2_sfcmpge : T_fcmp32<"sfcmp.ge", setoge, 0b000>;
247 def F2_sfcmpuo : T_fcmp32<"sfcmp.uo", setuo, 0b001>;
248 def F2_sfcmpeq : T_fcmp32<"sfcmp.eq", setoeq, 0b011>;
249 def F2_sfcmpgt : T_fcmp32<"sfcmp.gt", setogt, 0b100>;
252 //===----------------------------------------------------------------------===//
253 // Multiclass to define 'Def Pats' for ordered gt, ge, eq operations.
254 //===----------------------------------------------------------------------===//
256 let Predicates = [HasV5T] in
257 multiclass T_fcmp_pats<PatFrag cmpOp, InstHexagon IntMI, InstHexagon DoubleMI> {
259 def: Pat<(i1 (cmpOp F32:$src1, F32:$src2)),
260 (IntMI F32:$src1, F32:$src2)>;
262 def: Pat<(i1 (cmpOp F64:$src1, F64:$src2)),
263 (DoubleMI F64:$src1, F64:$src2)>;
266 defm : T_fcmp_pats <seteq, F2_sfcmpeq, F2_dfcmpeq>;
267 defm : T_fcmp_pats <setgt, F2_sfcmpgt, F2_dfcmpgt>;
268 defm : T_fcmp_pats <setge, F2_sfcmpge, F2_dfcmpge>;
270 //===----------------------------------------------------------------------===//
271 // Multiclass to define 'Def Pats' for unordered gt, ge, eq operations.
272 //===----------------------------------------------------------------------===//
273 let Predicates = [HasV5T] in
274 multiclass unord_Pats <PatFrag cmpOp, InstHexagon IntMI, InstHexagon DoubleMI> {
276 def: Pat<(i1 (cmpOp F32:$src1, F32:$src2)),
277 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
278 (IntMI F32:$src1, F32:$src2))>;
281 def: Pat<(i1 (cmpOp F64:$src1, F64:$src2)),
282 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
283 (DoubleMI F64:$src1, F64:$src2))>;
286 defm : unord_Pats <setuge, F2_sfcmpge, F2_dfcmpge>;
287 defm : unord_Pats <setugt, F2_sfcmpgt, F2_dfcmpgt>;
288 defm : unord_Pats <setueq, F2_sfcmpeq, F2_dfcmpeq>;
290 //===----------------------------------------------------------------------===//
291 // Multiclass to define 'Def Pats' for the following dags:
292 // seteq(setoeq(op1, op2), 0) -> not(setoeq(op1, op2))
293 // seteq(setoeq(op1, op2), 1) -> setoeq(op1, op2)
294 // setne(setoeq(op1, op2), 0) -> setoeq(op1, op2)
295 // setne(setoeq(op1, op2), 1) -> not(setoeq(op1, op2))
296 //===----------------------------------------------------------------------===//
297 let Predicates = [HasV5T] in
298 multiclass eq_ordgePats <PatFrag cmpOp, InstHexagon IntMI,
299 InstHexagon DoubleMI> {
301 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
302 (C2_not (IntMI F32:$src1, F32:$src2))>;
303 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
304 (IntMI F32:$src1, F32:$src2)>;
305 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
306 (IntMI F32:$src1, F32:$src2)>;
307 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
308 (C2_not (IntMI F32:$src1, F32:$src2))>;
311 def : Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
312 (C2_not (DoubleMI F64:$src1, F64:$src2))>;
313 def : Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 1)),
314 (DoubleMI F64:$src1, F64:$src2)>;
315 def : Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
316 (DoubleMI F64:$src1, F64:$src2)>;
317 def : Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 1)),
318 (C2_not (DoubleMI F64:$src1, F64:$src2))>;
321 defm : eq_ordgePats<setoeq, F2_sfcmpeq, F2_dfcmpeq>;
322 defm : eq_ordgePats<setoge, F2_sfcmpge, F2_dfcmpge>;
323 defm : eq_ordgePats<setogt, F2_sfcmpgt, F2_dfcmpgt>;
325 //===----------------------------------------------------------------------===//
326 // Multiclass to define 'Def Pats' for the following dags:
327 // seteq(setolt(op1, op2), 0) -> not(setogt(op2, op1))
328 // seteq(setolt(op1, op2), 1) -> setogt(op2, op1)
329 // setne(setolt(op1, op2), 0) -> setogt(op2, op1)
330 // setne(setolt(op1, op2), 1) -> not(setogt(op2, op1))
331 //===----------------------------------------------------------------------===//
332 let Predicates = [HasV5T] in
333 multiclass eq_ordltPats <PatFrag cmpOp, InstHexagon IntMI,
334 InstHexagon DoubleMI> {
336 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
337 (C2_not (IntMI F32:$src2, F32:$src1))>;
338 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
339 (IntMI F32:$src2, F32:$src1)>;
340 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
341 (IntMI F32:$src2, F32:$src1)>;
342 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
343 (C2_not (IntMI F32:$src2, F32:$src1))>;
346 def: Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
347 (C2_not (DoubleMI F64:$src2, F64:$src1))>;
348 def: Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 1)),
349 (DoubleMI F64:$src2, F64:$src1)>;
350 def: Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
351 (DoubleMI F64:$src2, F64:$src1)>;
352 def: Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
353 (C2_not (DoubleMI F64:$src2, F64:$src1))>;
356 defm : eq_ordltPats<setole, F2_sfcmpge, F2_dfcmpge>;
357 defm : eq_ordltPats<setolt, F2_sfcmpgt, F2_dfcmpgt>;
360 // o. seto inverse of setuo. http://llvm.org/docs/LangRef.html#i_fcmp
361 let Predicates = [HasV5T] in {
362 def: Pat<(i1 (seto F32:$src1, F32:$src2)),
363 (C2_not (F2_sfcmpuo F32:$src2, F32:$src1))>;
364 def: Pat<(i1 (seto F32:$src1, fpimm:$src2)),
365 (C2_not (F2_sfcmpuo (TFRI_f fpimm:$src2), F32:$src1))>;
366 def: Pat<(i1 (seto F64:$src1, F64:$src2)),
367 (C2_not (F2_dfcmpuo F64:$src2, F64:$src1))>;
368 def: Pat<(i1 (seto F64:$src1, fpimm:$src2)),
369 (C2_not (F2_dfcmpuo (CONST64_Float_Real fpimm:$src2), F64:$src1))>;
373 let Predicates = [HasV5T] in {
374 def: Pat<(i1 (setolt F32:$src1, F32:$src2)),
375 (F2_sfcmpgt F32:$src2, F32:$src1)>;
376 def: Pat<(i1 (setolt F32:$src1, fpimm:$src2)),
377 (F2_sfcmpgt (f32 (TFRI_f fpimm:$src2)), F32:$src1)>;
378 def: Pat<(i1 (setolt F64:$src1, F64:$src2)),
379 (F2_dfcmpgt F64:$src2, F64:$src1)>;
380 def: Pat<(i1 (setolt F64:$src1, fpimm:$src2)),
381 (F2_dfcmpgt (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
385 let Predicates = [HasV5T] in {
386 def: Pat<(i1 (setult F32:$src1, F32:$src2)),
387 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
388 (F2_sfcmpgt F32:$src2, F32:$src1))>;
389 def: Pat<(i1 (setult F32:$src1, fpimm:$src2)),
390 (C2_or (F2_sfcmpuo F32:$src1, (TFRI_f fpimm:$src2)),
391 (F2_sfcmpgt (TFRI_f fpimm:$src2), F32:$src1))>;
392 def: Pat<(i1 (setult F64:$src1, F64:$src2)),
393 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
394 (F2_dfcmpgt F64:$src2, F64:$src1))>;
395 def: Pat<(i1 (setult F64:$src1, fpimm:$src2)),
396 (C2_or (F2_dfcmpuo F64:$src1, (CONST64_Float_Real fpimm:$src2)),
397 (F2_dfcmpgt (CONST64_Float_Real fpimm:$src2), F64:$src1))>;
401 let Predicates = [HasV5T] in {
402 // rs <= rt -> rt >= rs.
403 def: Pat<(i1 (setole F32:$src1, F32:$src2)),
404 (F2_sfcmpge F32:$src2, F32:$src1)>;
405 def: Pat<(i1 (setole F32:$src1, fpimm:$src2)),
406 (F2_sfcmpge (TFRI_f fpimm:$src2), F32:$src1)>;
408 // Rss <= Rtt -> Rtt >= Rss.
409 def: Pat<(i1 (setole F64:$src1, F64:$src2)),
410 (F2_dfcmpge F64:$src2, F64:$src1)>;
411 def: Pat<(i1 (setole F64:$src1, fpimm:$src2)),
412 (F2_dfcmpge (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
416 let Predicates = [HasV5T] in {
417 // rs <= rt -> rt >= rs.
418 def: Pat<(i1 (setule F32:$src1, F32:$src2)),
419 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
420 (F2_sfcmpge F32:$src2, F32:$src1))>;
421 def: Pat<(i1 (setule F32:$src1, fpimm:$src2)),
422 (C2_or (F2_sfcmpuo F32:$src1, (TFRI_f fpimm:$src2)),
423 (F2_sfcmpge (TFRI_f fpimm:$src2), F32:$src1))>;
424 def: Pat<(i1 (setule F64:$src1, F64:$src2)),
425 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
426 (F2_dfcmpge F64:$src2, F64:$src1))>;
427 def: Pat<(i1 (setule F64:$src1, fpimm:$src2)),
428 (C2_or (F2_dfcmpuo F64:$src1, (CONST64_Float_Real fpimm:$src2)),
429 (F2_dfcmpge (CONST64_Float_Real fpimm:$src2), F64:$src1))>;
433 let Predicates = [HasV5T] in {
434 def: Pat<(i1 (setone F32:$src1, F32:$src2)),
435 (C2_not (F2_sfcmpeq F32:$src1, F32:$src2))>;
436 def: Pat<(i1 (setone F64:$src1, F64:$src2)),
437 (C2_not (F2_dfcmpeq F64:$src1, F64:$src2))>;
438 def: Pat<(i1 (setone F32:$src1, fpimm:$src2)),
439 (C2_not (F2_sfcmpeq F32:$src1, (TFRI_f fpimm:$src2)))>;
440 def: Pat<(i1 (setone F64:$src1, fpimm:$src2)),
441 (C2_not (F2_dfcmpeq F64:$src1, (CONST64_Float_Real fpimm:$src2)))>;
445 let Predicates = [HasV5T] in {
446 def: Pat<(i1 (setune F32:$src1, F32:$src2)),
447 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
448 (C2_not (F2_sfcmpeq F32:$src1, F32:$src2)))>;
449 def: Pat<(i1 (setune F64:$src1, F64:$src2)),
450 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
451 (C2_not (F2_dfcmpeq F64:$src1, F64:$src2)))>;
452 def: Pat<(i1 (setune F32:$src1, fpimm:$src2)),
453 (C2_or (F2_sfcmpuo F32:$src1, (TFRI_f fpimm:$src2)),
454 (C2_not (F2_sfcmpeq F32:$src1, (TFRI_f fpimm:$src2))))>;
455 def: Pat<(i1 (setune F64:$src1, fpimm:$src2)),
456 (C2_or (F2_dfcmpuo F64:$src1, (CONST64_Float_Real fpimm:$src2)),
457 (C2_not (F2_dfcmpeq F64:$src1,
458 (CONST64_Float_Real fpimm:$src2))))>;
461 // Besides set[o|u][comparions], we also need set[comparisons].
462 let Predicates = [HasV5T] in {
464 def: Pat<(i1 (setlt F32:$src1, F32:$src2)),
465 (F2_sfcmpgt F32:$src2, F32:$src1)>;
466 def: Pat<(i1 (setlt F32:$src1, fpimm:$src2)),
467 (F2_sfcmpgt (TFRI_f fpimm:$src2), F32:$src1)>;
468 def: Pat<(i1 (setlt F64:$src1, F64:$src2)),
469 (F2_dfcmpgt F64:$src2, F64:$src1)>;
470 def: Pat<(i1 (setlt F64:$src1, fpimm:$src2)),
471 (F2_dfcmpgt (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
474 // rs <= rt -> rt >= rs.
475 def: Pat<(i1 (setle F32:$src1, F32:$src2)),
476 (F2_sfcmpge F32:$src2, F32:$src1)>;
477 def: Pat<(i1 (setle F32:$src1, fpimm:$src2)),
478 (F2_sfcmpge (TFRI_f fpimm:$src2), F32:$src1)>;
480 // Rss <= Rtt -> Rtt >= Rss.
481 def: Pat<(i1 (setle F64:$src1, F64:$src2)),
482 (F2_dfcmpge F64:$src2, F64:$src1)>;
483 def: Pat<(i1 (setle F64:$src1, fpimm:$src2)),
484 (F2_dfcmpge (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
487 def: Pat<(i1 (setne F32:$src1, F32:$src2)),
488 (C2_not (F2_sfcmpeq F32:$src1, F32:$src2))>;
489 def: Pat<(i1 (setne F64:$src1, F64:$src2)),
490 (C2_not (F2_dfcmpeq F64:$src1, F64:$src2))>;
491 def: Pat<(i1 (setne F32:$src1, fpimm:$src2)),
492 (C2_not (F2_sfcmpeq F32:$src1, (TFRI_f fpimm:$src2)))>;
493 def: Pat<(i1 (setne F64:$src1, fpimm:$src2)),
494 (C2_not (F2_dfcmpeq F64:$src1, (CONST64_Float_Real fpimm:$src2)))>;
497 // F2 convert template classes:
499 class F2_RDD_RSS_CONVERT<string mnemonic, bits<3> MinOp,
500 SDNode Op, PatLeaf RCOut, PatLeaf RCIn,
502 : SInst <(outs DoubleRegs:$Rdd), (ins DoubleRegs:$Rss),
503 "$Rdd = "#mnemonic#"($Rss)"#chop,
504 [(set RCOut:$Rdd, (Op RCIn:$Rss))], "",
505 S_2op_tc_3or4x_SLOT23> {
511 let Inst{27-21} = 0b0000111;
512 let Inst{20-16} = Rss;
513 let Inst{7-5} = MinOp;
518 class F2_RDD_RS_CONVERT<string mnemonic, bits<3> MinOp,
519 SDNode Op, PatLeaf RCOut, PatLeaf RCIn,
521 : SInst <(outs DoubleRegs:$Rdd), (ins IntRegs:$Rs),
522 "$Rdd = "#mnemonic#"($Rs)"#chop,
523 [(set RCOut:$Rdd, (Op RCIn:$Rs))], "",
524 S_2op_tc_3or4x_SLOT23> {
530 let Inst{27-21} = 0b0100100;
531 let Inst{20-16} = Rs;
532 let Inst{7-5} = MinOp;
536 let isFP = 1, hasNewValue = 1 in
537 class F2_RD_RSS_CONVERT<string mnemonic, bits<3> MinOp,
538 SDNode Op, PatLeaf RCOut, PatLeaf RCIn,
540 : SInst <(outs IntRegs:$Rd), (ins DoubleRegs:$Rss),
541 "$Rd = "#mnemonic#"($Rss)"#chop,
542 [(set RCOut:$Rd, (Op RCIn:$Rss))], "",
543 S_2op_tc_3or4x_SLOT23> {
549 let Inst{27-24} = 0b1000;
550 let Inst{23-21} = MinOp;
551 let Inst{20-16} = Rss;
552 let Inst{7-5} = 0b001;
556 let isFP = 1, hasNewValue = 1 in
557 class F2_RD_RS_CONVERT<string mnemonic, bits<3> MajOp, bits<3> MinOp,
558 SDNode Op, PatLeaf RCOut, PatLeaf RCIn,
560 : SInst <(outs IntRegs:$Rd), (ins IntRegs:$Rs),
561 "$Rd = "#mnemonic#"($Rs)"#chop,
562 [(set RCOut:$Rd, (Op RCIn:$Rs))], "",
563 S_2op_tc_3or4x_SLOT23> {
569 let Inst{27-24} = 0b1011;
570 let Inst{23-21} = MajOp;
571 let Inst{20-16} = Rs;
572 let Inst{7-5} = MinOp;
576 // Convert single precision to double precision and vice-versa.
577 let isCodeGenOnly = 0 in {
578 def F2_conv_sf2df : F2_RDD_RS_CONVERT <"convert_sf2df", 0b000,
581 def F2_conv_df2sf : F2_RD_RSS_CONVERT <"convert_df2sf", 0b000,
584 // Convert Integer to Floating Point.
585 def F2_conv_d2sf : F2_RD_RSS_CONVERT <"convert_d2sf", 0b010,
586 sint_to_fp, F32, I64>;
587 def F2_conv_ud2sf : F2_RD_RSS_CONVERT <"convert_ud2sf", 0b001,
588 uint_to_fp, F32, I64>;
589 def F2_conv_uw2sf : F2_RD_RS_CONVERT <"convert_uw2sf", 0b001, 0b000,
590 uint_to_fp, F32, I32>;
591 def F2_conv_w2sf : F2_RD_RS_CONVERT <"convert_w2sf", 0b010, 0b000,
592 sint_to_fp, F32, I32>;
593 def F2_conv_d2df : F2_RDD_RSS_CONVERT <"convert_d2df", 0b011,
594 sint_to_fp, F64, I64>;
595 def F2_conv_ud2df : F2_RDD_RSS_CONVERT <"convert_ud2df", 0b010,
596 uint_to_fp, F64, I64>;
597 def F2_conv_uw2df : F2_RDD_RS_CONVERT <"convert_uw2df", 0b001,
598 uint_to_fp, F64, I32>;
599 def F2_conv_w2df : F2_RDD_RS_CONVERT <"convert_w2df", 0b010,
600 sint_to_fp, F64, I32>;
602 // Convert Floating Point to Integer - default.
603 def F2_conv_df2uw_chop : F2_RD_RSS_CONVERT <"convert_df2uw", 0b101,
604 fp_to_uint, I32, F64, ":chop">;
605 def F2_conv_df2w_chop : F2_RD_RSS_CONVERT <"convert_df2w", 0b111,
606 fp_to_sint, I32, F64, ":chop">;
607 def F2_conv_sf2uw_chop : F2_RD_RS_CONVERT <"convert_sf2uw", 0b011, 0b001,
608 fp_to_uint, I32, F32, ":chop">;
609 def F2_conv_sf2w_chop : F2_RD_RS_CONVERT <"convert_sf2w", 0b100, 0b001,
610 fp_to_sint, I32, F32, ":chop">;
611 def F2_conv_df2d_chop : F2_RDD_RSS_CONVERT <"convert_df2d", 0b110,
612 fp_to_sint, I64, F64, ":chop">;
613 def F2_conv_df2ud_chop : F2_RDD_RSS_CONVERT <"convert_df2ud", 0b111,
614 fp_to_uint, I64, F64, ":chop">;
615 def F2_conv_sf2d_chop : F2_RDD_RS_CONVERT <"convert_sf2d", 0b110,
616 fp_to_sint, I64, F32, ":chop">;
617 def F2_conv_sf2ud_chop : F2_RDD_RS_CONVERT <"convert_sf2ud", 0b101,
618 fp_to_uint, I64, F32, ":chop">;
620 // Convert Floating Point to Integer: non-chopped.
621 let AddedComplexity = 20, Predicates = [HasV5T, IEEERndNearV5T] in {
622 def F2_conv_df2d : F2_RDD_RSS_CONVERT <"convert_df2d", 0b000,
623 fp_to_sint, I64, F64>;
624 def F2_conv_df2ud : F2_RDD_RSS_CONVERT <"convert_df2ud", 0b001,
625 fp_to_uint, I64, F64>;
626 def F2_conv_sf2ud : F2_RDD_RS_CONVERT <"convert_sf2ud", 0b011,
627 fp_to_uint, I64, F32>;
628 def F2_conv_sf2d : F2_RDD_RS_CONVERT <"convert_sf2d", 0b100,
629 fp_to_sint, I64, F32>;
630 def F2_conv_df2uw : F2_RD_RSS_CONVERT <"convert_df2uw", 0b011,
631 fp_to_uint, I32, F64>;
632 def F2_conv_df2w : F2_RD_RSS_CONVERT <"convert_df2w", 0b100,
633 fp_to_sint, I32, F64>;
634 def F2_conv_sf2uw : F2_RD_RS_CONVERT <"convert_sf2uw", 0b011, 0b000,
635 fp_to_uint, I32, F32>;
636 def F2_conv_sf2w : F2_RD_RS_CONVERT <"convert_sf2w", 0b100, 0b000,
637 fp_to_sint, I32, F32>;
642 let isFP = 1, hasNewValue = 1, isCodeGenOnly = 0 in
643 def F2_sffixupr: SInst<(outs IntRegs:$Rd), (ins IntRegs:$Rs),
644 "$Rd = sffixupr($Rs)",
645 [], "" , S_2op_tc_3or4x_SLOT23>, Requires<[HasV5T]> {
651 let Inst{27-21} = 0b1011101;
652 let Inst{20-16} = Rs;
653 let Inst{7-5} = 0b000;
657 // Bitcast is different than [fp|sint|uint]_to_[sint|uint|fp].
658 let Predicates = [HasV5T] in {
659 def: Pat <(i32 (bitconvert F32:$src)), (I32:$src)>;
660 def: Pat <(f32 (bitconvert I32:$src)), (F32:$src)>;
661 def: Pat <(i64 (bitconvert F64:$src)), (I64:$src)>;
662 def: Pat <(f64 (bitconvert I64:$src)), (F64:$src)>;
665 // F2_sffma: Floating-point fused multiply add.
666 let isFP = 1, hasNewValue = 1 in
667 class T_sfmpy_acc <bit isSub, bit isLib>
668 : MInst<(outs IntRegs:$Rx),
669 (ins IntRegs:$dst2, IntRegs:$Rs, IntRegs:$Rt),
670 "$Rx "#!if(isSub, "-=","+=")#" sfmpy($Rs, $Rt)"#!if(isLib, ":lib",""),
671 [], "$dst2 = $Rx" , M_tc_3_SLOT23 > ,
679 let Inst{27-21} = 0b1111000;
680 let Inst{20-16} = Rs;
689 let isCodeGenOnly = 0 in {
690 def F2_sffma: T_sfmpy_acc <0, 0>;
691 def F2_sffms: T_sfmpy_acc <1, 0>;
692 def F2_sffma_lib: T_sfmpy_acc <0, 1>;
693 def F2_sffms_lib: T_sfmpy_acc <1, 1>;
696 // Floating-point fused multiply add w/ additional scaling (2**pu).
697 let isFP = 1, hasNewValue = 1, isCodeGenOnly = 0 in
698 def F2_sffma_sc: MInst <
700 (ins IntRegs:$dst2, IntRegs:$Rs, IntRegs:$Rt, PredRegs:$Pu),
701 "$Rx += sfmpy($Rs, $Rt, $Pu):scale" ,
702 [], "$dst2 = $Rx" , M_tc_3_SLOT23 > ,
711 let Inst{27-21} = 0b1111011;
712 let Inst{20-16} = Rs;
720 let isExtended = 1, isExtentSigned = 1, opExtentBits = 8, opExtendable = 3,
721 isPseudo = 1, InputType = "imm" in
722 def MUX_ir_f : ALU32_rr<(outs IntRegs:$dst),
723 (ins PredRegs:$src1, IntRegs:$src2, f32Ext:$src3),
724 "$dst = mux($src1, $src2, #$src3)",
725 [(set F32:$dst, (f32 (select I1:$src1, F32:$src2, fpimm:$src3)))]>,
728 let isExtended = 1, isExtentSigned = 1, opExtentBits = 8, opExtendable = 2,
729 isPseudo = 1, InputType = "imm" in
730 def MUX_ri_f : ALU32_rr<(outs IntRegs:$dst),
731 (ins PredRegs:$src1, f32Ext:$src2, IntRegs:$src3),
732 "$dst = mux($src1, #$src2, $src3)",
733 [(set F32:$dst, (f32 (select I1:$src1, fpimm:$src2, F32:$src3)))]>,
736 //===----------------------------------------------------------------------===//
737 // :natural forms of vasrh and vasrhub insns
738 //===----------------------------------------------------------------------===//
739 // S5_asrhub_rnd_sat: Vector arithmetic shift right by immediate with round,
740 // saturate, and pack.
741 let Defs = [USR_OVF], hasSideEffects = 0, hasNewValue = 1, opNewValue = 0 in
742 class T_ASRHUB<bit isSat>
743 : SInst <(outs IntRegs:$Rd),
744 (ins DoubleRegs:$Rss, u4Imm:$u4),
745 "$Rd = vasrhub($Rss, #$u4):"#!if(isSat, "sat", "raw"),
746 [], "", S_2op_tc_2_SLOT23>,
754 let Inst{27-21} = 0b1000011;
755 let Inst{20-16} = Rss;
756 let Inst{13-12} = 0b00;
758 let Inst{7-6} = 0b10;
762 def S5_asrhub_sat : T_ASRHUB <1>;
764 def S5_asrhub_rnd_sat_goodsyntax
765 : SInst <(outs IntRegs:$Rd), (ins DoubleRegs:$Rss, u4Imm:$u4),
766 "$Rd = vasrhub($Rss, #$u4):rnd:sat">, Requires<[HasV5T]>;
768 // Complex multiply 32x16
769 let Defs = [USR_OVF], Itinerary = S_3op_tc_3x_SLOT23, isCodeGenOnly = 0 in {
770 def M4_cmpyi_whc : T_S3op_8<"cmpyiwh", 0b101, 1, 1, 1, 1>;
771 def M4_cmpyr_whc : T_S3op_8<"cmpyrwh", 0b111, 1, 1, 1, 1>;
774 // Classify floating-point value
775 let isFP = 1, isCodeGenOnly = 0 in
776 def F2_sfclass : T_TEST_BIT_IMM<"sfclass", 0b111>;
778 let isFP = 1, isCodeGenOnly = 0 in
779 def F2_dfclass: ALU64Inst<(outs PredRegs:$Pd), (ins DoubleRegs:$Rss, u5Imm:$u5),
780 "$Pd = dfclass($Rss, #$u5)",
781 [], "" , ALU64_tc_2early_SLOT23 > , Requires<[HasV5T]> {
787 let Inst{27-21} = 0b1100100;
788 let Inst{20-16} = Rss;
789 let Inst{12-10} = 0b000;
791 let Inst{4-3} = 0b10;
795 // Instructions to create floating point constant
796 let hasNewValue = 1, opNewValue = 0 in
797 class T_fimm <string mnemonic, RegisterClass RC, bits<4> RegType, bit isNeg>
798 : ALU64Inst<(outs RC:$dst), (ins u10Imm:$src),
799 "$dst = "#mnemonic#"(#$src)"#!if(isNeg, ":neg", ":pos"),
800 [], "", ALU64_tc_3x_SLOT23>, Requires<[HasV5T]> {
805 let Inst{27-24} = RegType;
807 let Inst{22} = isNeg;
808 let Inst{21} = src{9};
809 let Inst{13-5} = src{8-0};
813 let isCodeGenOnly = 0 in {
814 def F2_sfimm_p : T_fimm <"sfmake", IntRegs, 0b0110, 0>;
815 def F2_sfimm_n : T_fimm <"sfmake", IntRegs, 0b0110, 1>;
816 def F2_dfimm_p : T_fimm <"dfmake", DoubleRegs, 0b1001, 0>;
817 def F2_dfimm_n : T_fimm <"dfmake", DoubleRegs, 0b1001, 1>;
820 def : Pat <(fabs (f32 IntRegs:$src1)),
821 (S2_clrbit_i (f32 IntRegs:$src1), 31)>,
824 def : Pat <(fneg (f32 IntRegs:$src1)),
825 (S2_togglebit_i (f32 IntRegs:$src1), 31)>,
829 def : Pat <(fabs (f64 DoubleRegs:$src1)),
830 (S2_clrbit_i (f32 (EXTRACT_SUBREG DoubleRegs:$src1, subreg_hireg)), 31)>,
833 def : Pat <(fabs (f64 DoubleRegs:$src1)),
834 (S2_clrbit_i (f32 (EXTRACT_SUBREG DoubleRegs:$src1, subreg_hireg)), 31)>,