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 let isCodeGenOnly = 0 in
19 def S2_asr_i_p_rnd : S_2OpInstImm<"asr", 0b110, 0b111, u6Imm,
21 (sra (i64 (add (i64 (sra I64:$src1, u6ImmPred:$src2)), 1)),
25 let Inst{13-8} = src2;
28 def S2_asr_i_p_rnd_goodsyntax
29 : MInst<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1, u6Imm:$src2),
30 "$dst = asrrnd($src1, #$src2)">;
32 let isCodeGenOnly = 0 in
33 def C4_fastcorner9 : T_LOGICAL_2OP<"fastcorner9", 0b000, 0, 0>,
35 let Inst{13,7,4} = 0b111;
38 let isCodeGenOnly = 0 in
39 def C4_fastcorner9_not : T_LOGICAL_2OP<"!fastcorner9", 0b000, 0, 0>,
41 let Inst{20,13,7,4} = 0b1111;
44 def SDTHexagonFCONST32 : SDTypeProfile<1, 1, [
47 def HexagonFCONST32 : SDNode<"HexagonISD::FCONST32", SDTHexagonFCONST32>;
49 let isReMaterializable = 1, isMoveImm = 1 in
50 def FCONST32_nsdata : LDInst<(outs IntRegs:$dst), (ins globaladdress:$global),
51 "$dst = CONST32(#$global)",
52 [(set (f32 IntRegs:$dst),
53 (HexagonFCONST32 tglobaladdr:$global))]>,
56 let isReMaterializable = 1, isMoveImm = 1 in
57 def CONST64_Float_Real : LDInst<(outs DoubleRegs:$dst), (ins f64imm:$src1),
58 "$dst = CONST64(#$src1)",
59 [(set DoubleRegs:$dst, fpimm:$src1)]>,
62 let isReMaterializable = 1, isMoveImm = 1 in
63 def CONST32_Float_Real : LDInst<(outs IntRegs:$dst), (ins f32imm:$src1),
64 "$dst = CONST32(#$src1)",
65 [(set IntRegs:$dst, fpimm:$src1)]>,
68 // Transfer immediate float.
69 // Only works with single precision fp value.
70 // For double precision, use CONST64_float_real, as 64bit transfer
71 // can only hold 40-bit values - 32 from const ext + 8 bit immediate.
72 // Make sure that complexity is more than the CONST32 pattern in
73 // HexagonInstrInfo.td patterns.
74 let isExtended = 1, opExtendable = 1, isMoveImm = 1, isReMaterializable = 1,
75 isPredicable = 1, AddedComplexity = 30, validSubTargets = HasV5SubT,
77 def TFRI_f : ALU32_ri<(outs IntRegs:$dst), (ins f32Ext:$src1),
79 [(set IntRegs:$dst, fpimm:$src1)]>,
82 let isExtended = 1, opExtendable = 2, isPredicated = 1,
83 hasSideEffects = 0, validSubTargets = HasV5SubT in
84 def TFRI_cPt_f : ALU32_ri<(outs IntRegs:$dst),
85 (ins PredRegs:$src1, f32Ext:$src2),
86 "if ($src1) $dst = #$src2",
90 let isExtended = 1, opExtendable = 2, isPredicated = 1, isPredicatedFalse = 1,
91 hasSideEffects = 0, validSubTargets = HasV5SubT in
92 def TFRI_cNotPt_f : ALU32_ri<(outs IntRegs:$dst),
93 (ins PredRegs:$src1, f32Ext:$src2),
94 "if (!$src1) $dst =#$src2",
98 def SDTHexagonI32I64: SDTypeProfile<1, 1, [SDTCisVT<0, i32>,
101 def HexagonPOPCOUNT: SDNode<"HexagonISD::POPCOUNT", SDTHexagonI32I64>;
103 let hasNewValue = 1, validSubTargets = HasV5SubT, isCodeGenOnly = 0 in
104 def S5_popcountp : ALU64_rr<(outs IntRegs:$Rd), (ins DoubleRegs:$Rss),
105 "$Rd = popcount($Rss)",
106 [(set I32:$Rd, (HexagonPOPCOUNT I64:$Rss))], "", S_2op_tc_2_SLOT23>,
113 let Inst{27-21} = 0b1000011;
114 let Inst{7-5} = 0b011;
116 let Inst{20-16} = Rss;
119 defm: Loadx_pat<load, f32, s11_2ExtPred, L2_loadri_io>;
120 defm: Loadx_pat<load, f64, s11_3ExtPred, L2_loadrd_io>;
122 defm: Storex_pat<store, F32, s11_2ExtPred, S2_storeri_io>;
123 defm: Storex_pat<store, F64, s11_3ExtPred, S2_storerd_io>;
125 let isFP = 1, hasNewValue = 1, opNewValue = 0 in
126 class T_MInstFloat <string mnemonic, bits<3> MajOp, bits<3> MinOp>
127 : MInst<(outs IntRegs:$Rd),
128 (ins IntRegs:$Rs, IntRegs:$Rt),
129 "$Rd = "#mnemonic#"($Rs, $Rt)", [],
130 "" , M_tc_3or4x_SLOT23 > ,
138 let Inst{27-24} = 0b1011;
139 let Inst{23-21} = MajOp;
140 let Inst{20-16} = Rs;
143 let Inst{7-5} = MinOp;
147 let isCommutable = 1, isCodeGenOnly = 0 in {
148 def F2_sfadd : T_MInstFloat < "sfadd", 0b000, 0b000>;
149 def F2_sfmpy : T_MInstFloat < "sfmpy", 0b010, 0b000>;
152 let isCodeGenOnly = 0 in
153 def F2_sfsub : T_MInstFloat < "sfsub", 0b000, 0b001>;
155 def: Pat<(f32 (fadd F32:$src1, F32:$src2)),
156 (F2_sfadd F32:$src1, F32:$src2)>;
158 def: Pat<(f32 (fsub F32:$src1, F32:$src2)),
159 (F2_sfsub F32:$src1, F32:$src2)>;
161 def: Pat<(f32 (fmul F32:$src1, F32:$src2)),
162 (F2_sfmpy F32:$src1, F32:$src2)>;
164 let Itinerary = M_tc_3x_SLOT23, isCodeGenOnly = 0 in {
165 def F2_sfmax : T_MInstFloat < "sfmax", 0b100, 0b000>;
166 def F2_sfmin : T_MInstFloat < "sfmin", 0b100, 0b001>;
169 let isCodeGenOnly = 0 in {
170 def F2_sffixupn : T_MInstFloat < "sffixupn", 0b110, 0b000>;
171 def F2_sffixupd : T_MInstFloat < "sffixupd", 0b110, 0b001>;
174 // F2_sfrecipa: Reciprocal approximation for division.
175 let isPredicateLate = 1, isFP = 1,
176 hasSideEffects = 0, hasNewValue = 1, isCodeGenOnly = 0 in
177 def F2_sfrecipa: MInst <
178 (outs IntRegs:$Rd, PredRegs:$Pe),
179 (ins IntRegs:$Rs, IntRegs:$Rt),
180 "$Rd, $Pe = sfrecipa($Rs, $Rt)">,
188 let Inst{27-21} = 0b1011111;
189 let Inst{20-16} = Rs;
197 // F2_dfcmpeq: Floating point compare for equal.
198 let isCompare = 1, isFP = 1 in
199 class T_fcmp <string mnemonic, RegisterClass RC, bits<3> MinOp,
200 list<dag> pattern = [] >
201 : ALU64Inst <(outs PredRegs:$dst), (ins RC:$src1, RC:$src2),
202 "$dst = "#mnemonic#"($src1, $src2)", pattern,
203 "" , ALU64_tc_2early_SLOT23 > ,
211 let Inst{27-21} = 0b0010111;
212 let Inst{20-16} = src1;
213 let Inst{12-8} = src2;
214 let Inst{7-5} = MinOp;
218 class T_fcmp64 <string mnemonic, PatFrag OpNode, bits<3> MinOp>
219 : T_fcmp <mnemonic, DoubleRegs, MinOp,
220 [(set I1:$dst, (OpNode F64:$src1, F64:$src2))]> {
222 let Inst{27-21} = 0b0010111;
225 class T_fcmp32 <string mnemonic, PatFrag OpNode, bits<3> MinOp>
226 : T_fcmp <mnemonic, IntRegs, MinOp,
227 [(set I1:$dst, (OpNode F32:$src1, F32:$src2))]> {
229 let Inst{27-21} = 0b0111111;
232 let isCodeGenOnly = 0 in {
233 def F2_dfcmpeq : T_fcmp64<"dfcmp.eq", setoeq, 0b000>;
234 def F2_dfcmpgt : T_fcmp64<"dfcmp.gt", setogt, 0b001>;
235 def F2_dfcmpge : T_fcmp64<"dfcmp.ge", setoge, 0b010>;
236 def F2_dfcmpuo : T_fcmp64<"dfcmp.uo", setuo, 0b011>;
238 def F2_sfcmpge : T_fcmp32<"sfcmp.ge", setoge, 0b000>;
239 def F2_sfcmpuo : T_fcmp32<"sfcmp.uo", setuo, 0b001>;
240 def F2_sfcmpeq : T_fcmp32<"sfcmp.eq", setoeq, 0b011>;
241 def F2_sfcmpgt : T_fcmp32<"sfcmp.gt", setogt, 0b100>;
244 //===----------------------------------------------------------------------===//
245 // Multiclass to define 'Def Pats' for ordered gt, ge, eq operations.
246 //===----------------------------------------------------------------------===//
248 let Predicates = [HasV5T] in
249 multiclass T_fcmp_pats<PatFrag cmpOp, InstHexagon IntMI, InstHexagon DoubleMI> {
251 def: Pat<(i1 (cmpOp F32:$src1, F32:$src2)),
252 (IntMI F32:$src1, F32:$src2)>;
254 def: Pat<(i1 (cmpOp F64:$src1, F64:$src2)),
255 (DoubleMI F64:$src1, F64:$src2)>;
258 defm : T_fcmp_pats <seteq, F2_sfcmpeq, F2_dfcmpeq>;
259 defm : T_fcmp_pats <setgt, F2_sfcmpgt, F2_dfcmpgt>;
260 defm : T_fcmp_pats <setge, F2_sfcmpge, F2_dfcmpge>;
262 //===----------------------------------------------------------------------===//
263 // Multiclass to define 'Def Pats' for unordered gt, ge, eq operations.
264 //===----------------------------------------------------------------------===//
265 let Predicates = [HasV5T] in
266 multiclass unord_Pats <PatFrag cmpOp, InstHexagon IntMI, InstHexagon DoubleMI> {
268 def: Pat<(i1 (cmpOp F32:$src1, F32:$src2)),
269 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
270 (IntMI F32:$src1, F32:$src2))>;
273 def: Pat<(i1 (cmpOp F64:$src1, F64:$src2)),
274 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
275 (DoubleMI F64:$src1, F64:$src2))>;
278 defm : unord_Pats <setuge, F2_sfcmpge, F2_dfcmpge>;
279 defm : unord_Pats <setugt, F2_sfcmpgt, F2_dfcmpgt>;
280 defm : unord_Pats <setueq, F2_sfcmpeq, F2_dfcmpeq>;
282 //===----------------------------------------------------------------------===//
283 // Multiclass to define 'Def Pats' for the following dags:
284 // seteq(setoeq(op1, op2), 0) -> not(setoeq(op1, op2))
285 // seteq(setoeq(op1, op2), 1) -> setoeq(op1, op2)
286 // setne(setoeq(op1, op2), 0) -> setoeq(op1, op2)
287 // setne(setoeq(op1, op2), 1) -> not(setoeq(op1, op2))
288 //===----------------------------------------------------------------------===//
289 let Predicates = [HasV5T] in
290 multiclass eq_ordgePats <PatFrag cmpOp, InstHexagon IntMI,
291 InstHexagon DoubleMI> {
293 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
294 (C2_not (IntMI F32:$src1, F32:$src2))>;
295 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
296 (IntMI F32:$src1, F32:$src2)>;
297 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
298 (IntMI F32:$src1, F32:$src2)>;
299 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
300 (C2_not (IntMI F32:$src1, F32:$src2))>;
303 def : Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
304 (C2_not (DoubleMI F64:$src1, F64:$src2))>;
305 def : Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 1)),
306 (DoubleMI F64:$src1, F64:$src2)>;
307 def : Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
308 (DoubleMI F64:$src1, F64:$src2)>;
309 def : Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 1)),
310 (C2_not (DoubleMI F64:$src1, F64:$src2))>;
313 defm : eq_ordgePats<setoeq, F2_sfcmpeq, F2_dfcmpeq>;
314 defm : eq_ordgePats<setoge, F2_sfcmpge, F2_dfcmpge>;
315 defm : eq_ordgePats<setogt, F2_sfcmpgt, F2_dfcmpgt>;
317 //===----------------------------------------------------------------------===//
318 // Multiclass to define 'Def Pats' for the following dags:
319 // seteq(setolt(op1, op2), 0) -> not(setogt(op2, op1))
320 // seteq(setolt(op1, op2), 1) -> setogt(op2, op1)
321 // setne(setolt(op1, op2), 0) -> setogt(op2, op1)
322 // setne(setolt(op1, op2), 1) -> not(setogt(op2, op1))
323 //===----------------------------------------------------------------------===//
324 let Predicates = [HasV5T] in
325 multiclass eq_ordltPats <PatFrag cmpOp, InstHexagon IntMI,
326 InstHexagon DoubleMI> {
328 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
329 (C2_not (IntMI F32:$src2, F32:$src1))>;
330 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
331 (IntMI F32:$src2, F32:$src1)>;
332 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
333 (IntMI F32:$src2, F32:$src1)>;
334 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
335 (C2_not (IntMI F32:$src2, F32:$src1))>;
338 def: Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
339 (C2_not (DoubleMI F64:$src2, F64:$src1))>;
340 def: Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 1)),
341 (DoubleMI F64:$src2, F64:$src1)>;
342 def: Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
343 (DoubleMI F64:$src2, F64:$src1)>;
344 def: Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
345 (C2_not (DoubleMI F64:$src2, F64:$src1))>;
348 defm : eq_ordltPats<setole, F2_sfcmpge, F2_dfcmpge>;
349 defm : eq_ordltPats<setolt, F2_sfcmpgt, F2_dfcmpgt>;
352 // o. seto inverse of setuo. http://llvm.org/docs/LangRef.html#i_fcmp
353 let Predicates = [HasV5T] in {
354 def: Pat<(i1 (seto F32:$src1, F32:$src2)),
355 (C2_not (F2_sfcmpuo F32:$src2, F32:$src1))>;
356 def: Pat<(i1 (seto F32:$src1, fpimm:$src2)),
357 (C2_not (F2_sfcmpuo (TFRI_f fpimm:$src2), F32:$src1))>;
358 def: Pat<(i1 (seto F64:$src1, F64:$src2)),
359 (C2_not (F2_dfcmpuo F64:$src2, F64:$src1))>;
360 def: Pat<(i1 (seto F64:$src1, fpimm:$src2)),
361 (C2_not (F2_dfcmpuo (CONST64_Float_Real fpimm:$src2), F64:$src1))>;
365 let Predicates = [HasV5T] in {
366 def: Pat<(i1 (setolt F32:$src1, F32:$src2)),
367 (F2_sfcmpgt F32:$src2, F32:$src1)>;
368 def: Pat<(i1 (setolt F32:$src1, fpimm:$src2)),
369 (F2_sfcmpgt (f32 (TFRI_f fpimm:$src2)), F32:$src1)>;
370 def: Pat<(i1 (setolt F64:$src1, F64:$src2)),
371 (F2_dfcmpgt F64:$src2, F64:$src1)>;
372 def: Pat<(i1 (setolt F64:$src1, fpimm:$src2)),
373 (F2_dfcmpgt (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
377 let Predicates = [HasV5T] in {
378 def: Pat<(i1 (setult F32:$src1, F32:$src2)),
379 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
380 (F2_sfcmpgt F32:$src2, F32:$src1))>;
381 def: Pat<(i1 (setult F32:$src1, fpimm:$src2)),
382 (C2_or (F2_sfcmpuo F32:$src1, (TFRI_f fpimm:$src2)),
383 (F2_sfcmpgt (TFRI_f fpimm:$src2), F32:$src1))>;
384 def: Pat<(i1 (setult F64:$src1, F64:$src2)),
385 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
386 (F2_dfcmpgt F64:$src2, F64:$src1))>;
387 def: Pat<(i1 (setult F64:$src1, fpimm:$src2)),
388 (C2_or (F2_dfcmpuo F64:$src1, (CONST64_Float_Real fpimm:$src2)),
389 (F2_dfcmpgt (CONST64_Float_Real fpimm:$src2), F64:$src1))>;
393 let Predicates = [HasV5T] in {
394 // rs <= rt -> rt >= rs.
395 def: Pat<(i1 (setole F32:$src1, F32:$src2)),
396 (F2_sfcmpge F32:$src2, F32:$src1)>;
397 def: Pat<(i1 (setole F32:$src1, fpimm:$src2)),
398 (F2_sfcmpge (TFRI_f fpimm:$src2), F32:$src1)>;
400 // Rss <= Rtt -> Rtt >= Rss.
401 def: Pat<(i1 (setole F64:$src1, F64:$src2)),
402 (F2_dfcmpge F64:$src2, F64:$src1)>;
403 def: Pat<(i1 (setole F64:$src1, fpimm:$src2)),
404 (F2_dfcmpge (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
408 let Predicates = [HasV5T] in {
409 // rs <= rt -> rt >= rs.
410 def: Pat<(i1 (setule F32:$src1, F32:$src2)),
411 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
412 (F2_sfcmpge F32:$src2, F32:$src1))>;
413 def: Pat<(i1 (setule F32:$src1, fpimm:$src2)),
414 (C2_or (F2_sfcmpuo F32:$src1, (TFRI_f fpimm:$src2)),
415 (F2_sfcmpge (TFRI_f fpimm:$src2), F32:$src1))>;
416 def: Pat<(i1 (setule F64:$src1, F64:$src2)),
417 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
418 (F2_dfcmpge F64:$src2, F64:$src1))>;
419 def: Pat<(i1 (setule F64:$src1, fpimm:$src2)),
420 (C2_or (F2_dfcmpuo F64:$src1, (CONST64_Float_Real fpimm:$src2)),
421 (F2_dfcmpge (CONST64_Float_Real fpimm:$src2), F64:$src1))>;
425 let Predicates = [HasV5T] in {
426 def: Pat<(i1 (setone F32:$src1, F32:$src2)),
427 (C2_not (F2_sfcmpeq F32:$src1, F32:$src2))>;
428 def: Pat<(i1 (setone F64:$src1, F64:$src2)),
429 (C2_not (F2_dfcmpeq F64:$src1, F64:$src2))>;
430 def: Pat<(i1 (setone F32:$src1, fpimm:$src2)),
431 (C2_not (F2_sfcmpeq F32:$src1, (TFRI_f fpimm:$src2)))>;
432 def: Pat<(i1 (setone F64:$src1, fpimm:$src2)),
433 (C2_not (F2_dfcmpeq F64:$src1, (CONST64_Float_Real fpimm:$src2)))>;
437 let Predicates = [HasV5T] in {
438 def: Pat<(i1 (setune F32:$src1, F32:$src2)),
439 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
440 (C2_not (F2_sfcmpeq F32:$src1, F32:$src2)))>;
441 def: Pat<(i1 (setune F64:$src1, F64:$src2)),
442 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
443 (C2_not (F2_dfcmpeq F64:$src1, F64:$src2)))>;
444 def: Pat<(i1 (setune F32:$src1, fpimm:$src2)),
445 (C2_or (F2_sfcmpuo F32:$src1, (TFRI_f fpimm:$src2)),
446 (C2_not (F2_sfcmpeq F32:$src1, (TFRI_f fpimm:$src2))))>;
447 def: Pat<(i1 (setune F64:$src1, fpimm:$src2)),
448 (C2_or (F2_dfcmpuo F64:$src1, (CONST64_Float_Real fpimm:$src2)),
449 (C2_not (F2_dfcmpeq F64:$src1,
450 (CONST64_Float_Real fpimm:$src2))))>;
453 // Besides set[o|u][comparions], we also need set[comparisons].
454 let Predicates = [HasV5T] in {
456 def: Pat<(i1 (setlt F32:$src1, F32:$src2)),
457 (F2_sfcmpgt F32:$src2, F32:$src1)>;
458 def: Pat<(i1 (setlt F32:$src1, fpimm:$src2)),
459 (F2_sfcmpgt (TFRI_f fpimm:$src2), F32:$src1)>;
460 def: Pat<(i1 (setlt F64:$src1, F64:$src2)),
461 (F2_dfcmpgt F64:$src2, F64:$src1)>;
462 def: Pat<(i1 (setlt F64:$src1, fpimm:$src2)),
463 (F2_dfcmpgt (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
466 // rs <= rt -> rt >= rs.
467 def: Pat<(i1 (setle F32:$src1, F32:$src2)),
468 (F2_sfcmpge F32:$src2, F32:$src1)>;
469 def: Pat<(i1 (setle F32:$src1, fpimm:$src2)),
470 (F2_sfcmpge (TFRI_f fpimm:$src2), F32:$src1)>;
472 // Rss <= Rtt -> Rtt >= Rss.
473 def: Pat<(i1 (setle F64:$src1, F64:$src2)),
474 (F2_dfcmpge F64:$src2, F64:$src1)>;
475 def: Pat<(i1 (setle F64:$src1, fpimm:$src2)),
476 (F2_dfcmpge (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
479 def: Pat<(i1 (setne F32:$src1, F32:$src2)),
480 (C2_not (F2_sfcmpeq F32:$src1, F32:$src2))>;
481 def: Pat<(i1 (setne F64:$src1, F64:$src2)),
482 (C2_not (F2_dfcmpeq F64:$src1, F64:$src2))>;
483 def: Pat<(i1 (setne F32:$src1, fpimm:$src2)),
484 (C2_not (F2_sfcmpeq F32:$src1, (TFRI_f fpimm:$src2)))>;
485 def: Pat<(i1 (setne F64:$src1, fpimm:$src2)),
486 (C2_not (F2_dfcmpeq F64:$src1, (CONST64_Float_Real fpimm:$src2)))>;
489 // F2 convert template classes:
491 class F2_RDD_RSS_CONVERT<string mnemonic, bits<3> MinOp,
492 SDNode Op, PatLeaf RCOut, PatLeaf RCIn,
494 : SInst <(outs DoubleRegs:$Rdd), (ins DoubleRegs:$Rss),
495 "$Rdd = "#mnemonic#"($Rss)"#chop,
496 [(set RCOut:$Rdd, (Op RCIn:$Rss))], "",
497 S_2op_tc_3or4x_SLOT23> {
503 let Inst{27-21} = 0b0000111;
504 let Inst{20-16} = Rss;
505 let Inst{7-5} = MinOp;
510 class F2_RDD_RS_CONVERT<string mnemonic, bits<3> MinOp,
511 SDNode Op, PatLeaf RCOut, PatLeaf RCIn,
513 : SInst <(outs DoubleRegs:$Rdd), (ins IntRegs:$Rs),
514 "$Rdd = "#mnemonic#"($Rs)"#chop,
515 [(set RCOut:$Rdd, (Op RCIn:$Rs))], "",
516 S_2op_tc_3or4x_SLOT23> {
522 let Inst{27-21} = 0b0100100;
523 let Inst{20-16} = Rs;
524 let Inst{7-5} = MinOp;
528 let isFP = 1, hasNewValue = 1 in
529 class F2_RD_RSS_CONVERT<string mnemonic, bits<3> MinOp,
530 SDNode Op, PatLeaf RCOut, PatLeaf RCIn,
532 : SInst <(outs IntRegs:$Rd), (ins DoubleRegs:$Rss),
533 "$Rd = "#mnemonic#"($Rss)"#chop,
534 [(set RCOut:$Rd, (Op RCIn:$Rss))], "",
535 S_2op_tc_3or4x_SLOT23> {
541 let Inst{27-24} = 0b1000;
542 let Inst{23-21} = MinOp;
543 let Inst{20-16} = Rss;
544 let Inst{7-5} = 0b001;
548 let isFP = 1, hasNewValue = 1 in
549 class F2_RD_RS_CONVERT<string mnemonic, bits<3> MajOp, bits<3> MinOp,
550 SDNode Op, PatLeaf RCOut, PatLeaf RCIn,
552 : SInst <(outs IntRegs:$Rd), (ins IntRegs:$Rs),
553 "$Rd = "#mnemonic#"($Rs)"#chop,
554 [(set RCOut:$Rd, (Op RCIn:$Rs))], "",
555 S_2op_tc_3or4x_SLOT23> {
561 let Inst{27-24} = 0b1011;
562 let Inst{23-21} = MajOp;
563 let Inst{20-16} = Rs;
564 let Inst{7-5} = MinOp;
568 // Convert single precision to double precision and vice-versa.
569 let isCodeGenOnly = 0 in {
570 def F2_conv_sf2df : F2_RDD_RS_CONVERT <"convert_sf2df", 0b000,
573 def F2_conv_df2sf : F2_RD_RSS_CONVERT <"convert_df2sf", 0b000,
576 // Convert Integer to Floating Point.
577 def F2_conv_d2sf : F2_RD_RSS_CONVERT <"convert_d2sf", 0b010,
578 sint_to_fp, F32, I64>;
579 def F2_conv_ud2sf : F2_RD_RSS_CONVERT <"convert_ud2sf", 0b001,
580 uint_to_fp, F32, I64>;
581 def F2_conv_uw2sf : F2_RD_RS_CONVERT <"convert_uw2sf", 0b001, 0b000,
582 uint_to_fp, F32, I32>;
583 def F2_conv_w2sf : F2_RD_RS_CONVERT <"convert_w2sf", 0b010, 0b000,
584 sint_to_fp, F32, I32>;
585 def F2_conv_d2df : F2_RDD_RSS_CONVERT <"convert_d2df", 0b011,
586 sint_to_fp, F64, I64>;
587 def F2_conv_ud2df : F2_RDD_RSS_CONVERT <"convert_ud2df", 0b010,
588 uint_to_fp, F64, I64>;
589 def F2_conv_uw2df : F2_RDD_RS_CONVERT <"convert_uw2df", 0b001,
590 uint_to_fp, F64, I32>;
591 def F2_conv_w2df : F2_RDD_RS_CONVERT <"convert_w2df", 0b010,
592 sint_to_fp, F64, I32>;
594 // Convert Floating Point to Integer - default.
595 def F2_conv_df2uw_chop : F2_RD_RSS_CONVERT <"convert_df2uw", 0b101,
596 fp_to_uint, I32, F64, ":chop">;
597 def F2_conv_df2w_chop : F2_RD_RSS_CONVERT <"convert_df2w", 0b111,
598 fp_to_sint, I32, F64, ":chop">;
599 def F2_conv_sf2uw_chop : F2_RD_RS_CONVERT <"convert_sf2uw", 0b011, 0b001,
600 fp_to_uint, I32, F32, ":chop">;
601 def F2_conv_sf2w_chop : F2_RD_RS_CONVERT <"convert_sf2w", 0b100, 0b001,
602 fp_to_sint, I32, F32, ":chop">;
603 def F2_conv_df2d_chop : F2_RDD_RSS_CONVERT <"convert_df2d", 0b110,
604 fp_to_sint, I64, F64, ":chop">;
605 def F2_conv_df2ud_chop : F2_RDD_RSS_CONVERT <"convert_df2ud", 0b111,
606 fp_to_uint, I64, F64, ":chop">;
607 def F2_conv_sf2d_chop : F2_RDD_RS_CONVERT <"convert_sf2d", 0b110,
608 fp_to_sint, I64, F32, ":chop">;
609 def F2_conv_sf2ud_chop : F2_RDD_RS_CONVERT <"convert_sf2ud", 0b101,
610 fp_to_uint, I64, F32, ":chop">;
612 // Convert Floating Point to Integer: non-chopped.
613 let AddedComplexity = 20, Predicates = [HasV5T, IEEERndNearV5T] in {
614 def F2_conv_df2d : F2_RDD_RSS_CONVERT <"convert_df2d", 0b000,
615 fp_to_sint, I64, F64>;
616 def F2_conv_df2ud : F2_RDD_RSS_CONVERT <"convert_df2ud", 0b001,
617 fp_to_uint, I64, F64>;
618 def F2_conv_sf2ud : F2_RDD_RS_CONVERT <"convert_sf2ud", 0b011,
619 fp_to_uint, I64, F32>;
620 def F2_conv_sf2d : F2_RDD_RS_CONVERT <"convert_sf2d", 0b100,
621 fp_to_sint, I64, F32>;
622 def F2_conv_df2uw : F2_RD_RSS_CONVERT <"convert_df2uw", 0b011,
623 fp_to_uint, I32, F64>;
624 def F2_conv_df2w : F2_RD_RSS_CONVERT <"convert_df2w", 0b100,
625 fp_to_sint, I32, F64>;
626 def F2_conv_sf2uw : F2_RD_RS_CONVERT <"convert_sf2uw", 0b011, 0b000,
627 fp_to_uint, I32, F32>;
628 def F2_conv_sf2w : F2_RD_RS_CONVERT <"convert_sf2w", 0b100, 0b000,
629 fp_to_sint, I32, F32>;
634 let isFP = 1, hasNewValue = 1, isCodeGenOnly = 0 in
635 def F2_sffixupr: SInst<(outs IntRegs:$Rd), (ins IntRegs:$Rs),
636 "$Rd = sffixupr($Rs)",
637 [], "" , S_2op_tc_3or4x_SLOT23>, Requires<[HasV5T]> {
643 let Inst{27-21} = 0b1011101;
644 let Inst{20-16} = Rs;
645 let Inst{7-5} = 0b000;
649 // Bitcast is different than [fp|sint|uint]_to_[sint|uint|fp].
650 let Predicates = [HasV5T] in {
651 def: Pat <(i32 (bitconvert F32:$src)), (I32:$src)>;
652 def: Pat <(f32 (bitconvert I32:$src)), (F32:$src)>;
653 def: Pat <(i64 (bitconvert F64:$src)), (I64:$src)>;
654 def: Pat <(f64 (bitconvert I64:$src)), (F64:$src)>;
657 // F2_sffma: Floating-point fused multiply add.
658 let isFP = 1, hasNewValue = 1 in
659 class T_sfmpy_acc <bit isSub, bit isLib>
660 : MInst<(outs IntRegs:$Rx),
661 (ins IntRegs:$dst2, IntRegs:$Rs, IntRegs:$Rt),
662 "$Rx "#!if(isSub, "-=","+=")#" sfmpy($Rs, $Rt)"#!if(isLib, ":lib",""),
663 [], "$dst2 = $Rx" , M_tc_3_SLOT23 > ,
671 let Inst{27-21} = 0b1111000;
672 let Inst{20-16} = Rs;
681 let isCodeGenOnly = 0 in {
682 def F2_sffma: T_sfmpy_acc <0, 0>;
683 def F2_sffms: T_sfmpy_acc <1, 0>;
684 def F2_sffma_lib: T_sfmpy_acc <0, 1>;
685 def F2_sffms_lib: T_sfmpy_acc <1, 1>;
688 // Floating-point fused multiply add w/ additional scaling (2**pu).
689 let isFP = 1, hasNewValue = 1, isCodeGenOnly = 0 in
690 def F2_sffma_sc: MInst <
692 (ins IntRegs:$dst2, IntRegs:$Rs, IntRegs:$Rt, PredRegs:$Pu),
693 "$Rx += sfmpy($Rs, $Rt, $Pu):scale" ,
694 [], "$dst2 = $Rx" , M_tc_3_SLOT23 > ,
703 let Inst{27-21} = 0b1111011;
704 let Inst{20-16} = Rs;
712 let isExtended = 1, isExtentSigned = 1, opExtentBits = 8, opExtendable = 3,
713 isPseudo = 1, InputType = "imm" in
714 def MUX_ir_f : ALU32_rr<(outs IntRegs:$dst),
715 (ins PredRegs:$src1, IntRegs:$src2, f32Ext:$src3),
716 "$dst = mux($src1, $src2, #$src3)",
717 [(set F32:$dst, (f32 (select I1:$src1, F32:$src2, fpimm:$src3)))]>,
720 let isExtended = 1, isExtentSigned = 1, opExtentBits = 8, opExtendable = 2,
721 isPseudo = 1, InputType = "imm" in
722 def MUX_ri_f : ALU32_rr<(outs IntRegs:$dst),
723 (ins PredRegs:$src1, f32Ext:$src2, IntRegs:$src3),
724 "$dst = mux($src1, #$src2, $src3)",
725 [(set F32:$dst, (f32 (select I1:$src1, fpimm:$src2, F32:$src3)))]>,
728 //===----------------------------------------------------------------------===//
729 // :natural forms of vasrh and vasrhub insns
730 //===----------------------------------------------------------------------===//
731 // S5_asrhub_rnd_sat: Vector arithmetic shift right by immediate with round,
732 // saturate, and pack.
733 let Defs = [USR_OVF], hasSideEffects = 0, hasNewValue = 1, opNewValue = 0 in
734 class T_ASRHUB<bit isSat>
735 : SInst <(outs IntRegs:$Rd),
736 (ins DoubleRegs:$Rss, u4Imm:$u4),
737 "$Rd = vasrhub($Rss, #$u4):"#!if(isSat, "sat", "raw"),
738 [], "", S_2op_tc_2_SLOT23>,
746 let Inst{27-21} = 0b1000011;
747 let Inst{20-16} = Rss;
748 let Inst{13-12} = 0b00;
750 let Inst{7-6} = 0b10;
754 def S5_asrhub_sat : T_ASRHUB <1>;
756 def S5_asrhub_rnd_sat_goodsyntax
757 : SInst <(outs IntRegs:$Rd), (ins DoubleRegs:$Rss, u4Imm:$u4),
758 "$Rd = vasrhub($Rss, #$u4):rnd:sat">, Requires<[HasV5T]>;
760 // Classify floating-point value
761 let isFP = 1, isCodeGenOnly = 0 in
762 def F2_sfclass : T_TEST_BIT_IMM<"sfclass", 0b111>;
764 let isFP = 1, isCodeGenOnly = 0 in
765 def F2_dfclass: ALU64Inst<(outs PredRegs:$Pd), (ins DoubleRegs:$Rss, u5Imm:$u5),
766 "$Pd = dfclass($Rss, #$u5)",
767 [], "" , ALU64_tc_2early_SLOT23 > , Requires<[HasV5T]> {
773 let Inst{27-21} = 0b1100100;
774 let Inst{20-16} = Rss;
775 let Inst{12-10} = 0b000;
777 let Inst{4-3} = 0b10;
781 // Instructions to create floating point constant
782 let hasNewValue = 1, opNewValue = 0 in
783 class T_fimm <string mnemonic, RegisterClass RC, bits<4> RegType, bit isNeg>
784 : ALU64Inst<(outs RC:$dst), (ins u10Imm:$src),
785 "$dst = "#mnemonic#"(#$src)"#!if(isNeg, ":neg", ":pos"),
786 [], "", ALU64_tc_3x_SLOT23>, Requires<[HasV5T]> {
791 let Inst{27-24} = RegType;
793 let Inst{22} = isNeg;
794 let Inst{21} = src{9};
795 let Inst{13-5} = src{8-0};
799 let isCodeGenOnly = 0 in {
800 def F2_sfimm_p : T_fimm <"sfmake", IntRegs, 0b0110, 0>;
801 def F2_sfimm_n : T_fimm <"sfmake", IntRegs, 0b0110, 1>;
802 def F2_dfimm_p : T_fimm <"dfmake", DoubleRegs, 0b1001, 0>;
803 def F2_dfimm_n : T_fimm <"dfmake", DoubleRegs, 0b1001, 1>;
806 def : Pat <(fabs (f32 IntRegs:$src1)),
807 (S2_clrbit_i (f32 IntRegs:$src1), 31)>,
810 def : Pat <(fneg (f32 IntRegs:$src1)),
811 (S2_togglebit_i (f32 IntRegs:$src1), 31)>,
815 def : Pat <(fabs (f64 DoubleRegs:$src1)),
816 (S2_clrbit_i (f32 (EXTRACT_SUBREG DoubleRegs:$src1, subreg_hireg)), 31)>,
819 def : Pat <(fabs (f64 DoubleRegs:$src1)),
820 (S2_clrbit_i (f32 (EXTRACT_SUBREG DoubleRegs:$src1, subreg_hireg)), 31)>,