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 //Rdd[+]=vrmpybsu(Rss,Rtt)
19 let Predicates = [HasV5T] in {
20 def M5_vrmpybsu: T_XTYPE_Vect<"vrmpybsu", 0b110, 0b001, 0>;
21 def M5_vrmacbsu: T_XTYPE_Vect_acc<"vrmpybsu", 0b110, 0b001, 0>;
23 //Rdd[+]=vrmpybu(Rss,Rtt)
24 def M5_vrmpybuu: T_XTYPE_Vect<"vrmpybu", 0b100, 0b001, 0>;
25 def M5_vrmacbuu: T_XTYPE_Vect_acc<"vrmpybu", 0b100, 0b001, 0>;
27 def M5_vdmpybsu: T_M2_vmpy<"vdmpybsu", 0b101, 0b001, 0, 0, 1>;
28 def M5_vdmacbsu: T_M2_vmpy_acc_sat <"vdmpybsu", 0b001, 0b001, 0, 0>;
31 // Vector multiply bytes
32 // Rdd=vmpyb[s]u(Rs,Rt)
33 let Predicates = [HasV5T] in {
34 def M5_vmpybsu: T_XTYPE_mpy64 <"vmpybsu", 0b010, 0b001, 0, 0, 0>;
35 def M5_vmpybuu: T_XTYPE_mpy64 <"vmpybu", 0b100, 0b001, 0, 0, 0>;
37 // Rxx+=vmpyb[s]u(Rs,Rt)
38 def M5_vmacbsu: T_XTYPE_mpy64_acc <"vmpybsu", "+", 0b110, 0b001, 0, 0, 0>;
39 def M5_vmacbuu: T_XTYPE_mpy64_acc <"vmpybu", "+", 0b100, 0b001, 0, 0, 0>;
41 // Rd=vaddhub(Rss,Rtt):sat
42 let hasNewValue = 1, opNewValue = 0 in
43 def A5_vaddhubs: T_S3op_1 <"vaddhub", IntRegs, 0b01, 0b001, 0, 1>;
46 def S2_asr_i_p_rnd : S_2OpInstImm<"asr", 0b110, 0b111, u6Imm,
48 (sra (i64 (add (i64 (sra I64:$src1, u6ImmPred:$src2)), 1)),
52 let Inst{13-8} = src2;
55 let isAsmParserOnly = 1 in
56 def S2_asr_i_p_rnd_goodsyntax
57 : MInst<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1, u6Imm:$src2),
58 "$dst = asrrnd($src1, #$src2)">;
60 def C4_fastcorner9 : T_LOGICAL_2OP<"fastcorner9", 0b000, 0, 0>,
62 let Inst{13,7,4} = 0b111;
65 def C4_fastcorner9_not : T_LOGICAL_2OP<"!fastcorner9", 0b000, 0, 0>,
67 let Inst{20,13,7,4} = 0b1111;
70 def SDTHexagonFCONST32 : SDTypeProfile<1, 1, [SDTCisVT<0, f32>,
72 def HexagonFCONST32 : SDNode<"HexagonISD::FCONST32", SDTHexagonFCONST32>;
74 let isReMaterializable = 1, isMoveImm = 1, isAsmParserOnly = 1 in
75 def FCONST32_nsdata : LDInst<(outs IntRegs:$dst), (ins globaladdress:$global),
76 "$dst = CONST32(#$global)",
78 (HexagonFCONST32 tglobaladdr:$global))]>,
81 let isReMaterializable = 1, isMoveImm = 1, isAsmParserOnly = 1 in
82 def CONST64_Float_Real : LDInst<(outs DoubleRegs:$dst), (ins f64imm:$src1),
83 "$dst = CONST64(#$src1)",
84 [(set F64:$dst, fpimm:$src1)]>,
87 let isReMaterializable = 1, isMoveImm = 1, isAsmParserOnly = 1 in
88 def CONST32_Float_Real : LDInst<(outs IntRegs:$dst), (ins f32imm:$src1),
89 "$dst = CONST32(#$src1)",
90 [(set F32:$dst, fpimm:$src1)]>,
93 // Transfer immediate float.
94 // Only works with single precision fp value.
95 // For double precision, use CONST64_float_real, as 64bit transfer
96 // can only hold 40-bit values - 32 from const ext + 8 bit immediate.
97 // Make sure that complexity is more than the CONST32 pattern in
98 // HexagonInstrInfo.td patterns.
99 let isExtended = 1, opExtendable = 1, isMoveImm = 1, isReMaterializable = 1,
100 isPredicable = 1, AddedComplexity = 30, validSubTargets = HasV5SubT,
102 def TFRI_f : ALU32_ri<(outs IntRegs:$dst), (ins f32Ext:$src1),
104 [(set F32:$dst, fpimm:$src1)]>,
107 let isExtended = 1, opExtendable = 2, isPredicated = 1,
108 hasSideEffects = 0, validSubTargets = HasV5SubT, isCodeGenOnly = 1 in
109 def TFRI_cPt_f : ALU32_ri<(outs IntRegs:$dst),
110 (ins PredRegs:$src1, f32Ext:$src2),
111 "if ($src1) $dst = #$src2", []>,
114 let isPseudo = 1, isExtended = 1, opExtendable = 2, isPredicated = 1,
115 isPredicatedFalse = 1, hasSideEffects = 0, validSubTargets = HasV5SubT in
116 def TFRI_cNotPt_f : ALU32_ri<(outs IntRegs:$dst),
117 (ins PredRegs:$src1, f32Ext:$src2),
118 "if (!$src1) $dst = #$src2", []>,
121 def SDTHexagonI32I64: SDTypeProfile<1, 1, [SDTCisVT<0, i32>,
124 def HexagonPOPCOUNT: SDNode<"HexagonISD::POPCOUNT", SDTHexagonI32I64>;
126 let hasNewValue = 1, validSubTargets = HasV5SubT in
127 def S5_popcountp : ALU64_rr<(outs IntRegs:$Rd), (ins DoubleRegs:$Rss),
128 "$Rd = popcount($Rss)",
129 [(set I32:$Rd, (HexagonPOPCOUNT I64:$Rss))], "", S_2op_tc_2_SLOT23>,
136 let Inst{27-21} = 0b1000011;
137 let Inst{7-5} = 0b011;
139 let Inst{20-16} = Rss;
142 defm: Loadx_pat<load, f32, s11_2ExtPred, L2_loadri_io>;
143 defm: Loadx_pat<load, f64, s11_3ExtPred, L2_loadrd_io>;
145 defm: Storex_pat<store, F32, s11_2ExtPred, S2_storeri_io>;
146 defm: Storex_pat<store, F64, s11_3ExtPred, S2_storerd_io>;
148 let isFP = 1, hasNewValue = 1, opNewValue = 0 in
149 class T_MInstFloat <string mnemonic, bits<3> MajOp, bits<3> MinOp>
150 : MInst<(outs IntRegs:$Rd),
151 (ins IntRegs:$Rs, IntRegs:$Rt),
152 "$Rd = "#mnemonic#"($Rs, $Rt)", [],
153 "" , M_tc_3or4x_SLOT23 > ,
161 let Inst{27-24} = 0b1011;
162 let Inst{23-21} = MajOp;
163 let Inst{20-16} = Rs;
166 let Inst{7-5} = MinOp;
170 let isCommutable = 1 in {
171 def F2_sfadd : T_MInstFloat < "sfadd", 0b000, 0b000>;
172 def F2_sfmpy : T_MInstFloat < "sfmpy", 0b010, 0b000>;
175 def F2_sfsub : T_MInstFloat < "sfsub", 0b000, 0b001>;
177 def: Pat<(f32 (fadd F32:$src1, F32:$src2)),
178 (F2_sfadd F32:$src1, F32:$src2)>;
180 def: Pat<(f32 (fsub F32:$src1, F32:$src2)),
181 (F2_sfsub F32:$src1, F32:$src2)>;
183 def: Pat<(f32 (fmul F32:$src1, F32:$src2)),
184 (F2_sfmpy F32:$src1, F32:$src2)>;
186 let Itinerary = M_tc_3x_SLOT23 in {
187 def F2_sfmax : T_MInstFloat < "sfmax", 0b100, 0b000>;
188 def F2_sfmin : T_MInstFloat < "sfmin", 0b100, 0b001>;
191 def F2_sffixupn : T_MInstFloat < "sffixupn", 0b110, 0b000>;
192 def F2_sffixupd : T_MInstFloat < "sffixupd", 0b110, 0b001>;
194 // F2_sfrecipa: Reciprocal approximation for division.
195 let isPredicateLate = 1, isFP = 1,
196 hasSideEffects = 0, hasNewValue = 1 in
197 def F2_sfrecipa: MInst <
198 (outs IntRegs:$Rd, PredRegs:$Pe),
199 (ins IntRegs:$Rs, IntRegs:$Rt),
200 "$Rd, $Pe = sfrecipa($Rs, $Rt)">,
208 let Inst{27-21} = 0b1011111;
209 let Inst{20-16} = Rs;
217 // F2_dfcmpeq: Floating point compare for equal.
218 let isCompare = 1, isFP = 1 in
219 class T_fcmp <string mnemonic, RegisterClass RC, bits<3> MinOp,
220 list<dag> pattern = [] >
221 : ALU64Inst <(outs PredRegs:$dst), (ins RC:$src1, RC:$src2),
222 "$dst = "#mnemonic#"($src1, $src2)", pattern,
223 "" , ALU64_tc_2early_SLOT23 > ,
231 let Inst{27-21} = 0b0010111;
232 let Inst{20-16} = src1;
233 let Inst{12-8} = src2;
234 let Inst{7-5} = MinOp;
238 class T_fcmp64 <string mnemonic, PatFrag OpNode, bits<3> MinOp>
239 : T_fcmp <mnemonic, DoubleRegs, MinOp,
240 [(set I1:$dst, (OpNode F64:$src1, F64:$src2))]> {
242 let Inst{27-21} = 0b0010111;
245 class T_fcmp32 <string mnemonic, PatFrag OpNode, bits<3> MinOp>
246 : T_fcmp <mnemonic, IntRegs, MinOp,
247 [(set I1:$dst, (OpNode F32:$src1, F32:$src2))]> {
249 let Inst{27-21} = 0b0111111;
252 def F2_dfcmpeq : T_fcmp64<"dfcmp.eq", setoeq, 0b000>;
253 def F2_dfcmpgt : T_fcmp64<"dfcmp.gt", setogt, 0b001>;
254 def F2_dfcmpge : T_fcmp64<"dfcmp.ge", setoge, 0b010>;
255 def F2_dfcmpuo : T_fcmp64<"dfcmp.uo", setuo, 0b011>;
257 def F2_sfcmpge : T_fcmp32<"sfcmp.ge", setoge, 0b000>;
258 def F2_sfcmpuo : T_fcmp32<"sfcmp.uo", setuo, 0b001>;
259 def F2_sfcmpeq : T_fcmp32<"sfcmp.eq", setoeq, 0b011>;
260 def F2_sfcmpgt : T_fcmp32<"sfcmp.gt", setogt, 0b100>;
262 //===----------------------------------------------------------------------===//
263 // Multiclass to define 'Def Pats' for ordered gt, ge, eq operations.
264 //===----------------------------------------------------------------------===//
266 let Predicates = [HasV5T] in
267 multiclass T_fcmp_pats<PatFrag cmpOp, InstHexagon IntMI, InstHexagon DoubleMI> {
269 def: Pat<(i1 (cmpOp F32:$src1, F32:$src2)),
270 (IntMI F32:$src1, F32:$src2)>;
272 def: Pat<(i1 (cmpOp F64:$src1, F64:$src2)),
273 (DoubleMI F64:$src1, F64:$src2)>;
276 defm : T_fcmp_pats <seteq, F2_sfcmpeq, F2_dfcmpeq>;
277 defm : T_fcmp_pats <setgt, F2_sfcmpgt, F2_dfcmpgt>;
278 defm : T_fcmp_pats <setge, F2_sfcmpge, F2_dfcmpge>;
280 //===----------------------------------------------------------------------===//
281 // Multiclass to define 'Def Pats' for unordered gt, ge, eq operations.
282 //===----------------------------------------------------------------------===//
283 let Predicates = [HasV5T] in
284 multiclass unord_Pats <PatFrag cmpOp, InstHexagon IntMI, InstHexagon DoubleMI> {
286 def: Pat<(i1 (cmpOp F32:$src1, F32:$src2)),
287 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
288 (IntMI F32:$src1, F32:$src2))>;
291 def: Pat<(i1 (cmpOp F64:$src1, F64:$src2)),
292 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
293 (DoubleMI F64:$src1, F64:$src2))>;
296 defm : unord_Pats <setuge, F2_sfcmpge, F2_dfcmpge>;
297 defm : unord_Pats <setugt, F2_sfcmpgt, F2_dfcmpgt>;
298 defm : unord_Pats <setueq, F2_sfcmpeq, F2_dfcmpeq>;
300 //===----------------------------------------------------------------------===//
301 // Multiclass to define 'Def Pats' for the following dags:
302 // seteq(setoeq(op1, op2), 0) -> not(setoeq(op1, op2))
303 // seteq(setoeq(op1, op2), 1) -> setoeq(op1, op2)
304 // setne(setoeq(op1, op2), 0) -> setoeq(op1, op2)
305 // setne(setoeq(op1, op2), 1) -> not(setoeq(op1, op2))
306 //===----------------------------------------------------------------------===//
307 let Predicates = [HasV5T] in
308 multiclass eq_ordgePats <PatFrag cmpOp, InstHexagon IntMI,
309 InstHexagon DoubleMI> {
311 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
312 (C2_not (IntMI F32:$src1, F32:$src2))>;
313 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
314 (IntMI F32:$src1, F32:$src2)>;
315 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
316 (IntMI F32:$src1, F32:$src2)>;
317 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
318 (C2_not (IntMI F32:$src1, F32:$src2))>;
321 def : Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
322 (C2_not (DoubleMI F64:$src1, F64:$src2))>;
323 def : Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 1)),
324 (DoubleMI F64:$src1, F64:$src2)>;
325 def : Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
326 (DoubleMI F64:$src1, F64:$src2)>;
327 def : Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 1)),
328 (C2_not (DoubleMI F64:$src1, F64:$src2))>;
331 defm : eq_ordgePats<setoeq, F2_sfcmpeq, F2_dfcmpeq>;
332 defm : eq_ordgePats<setoge, F2_sfcmpge, F2_dfcmpge>;
333 defm : eq_ordgePats<setogt, F2_sfcmpgt, F2_dfcmpgt>;
335 //===----------------------------------------------------------------------===//
336 // Multiclass to define 'Def Pats' for the following dags:
337 // seteq(setolt(op1, op2), 0) -> not(setogt(op2, op1))
338 // seteq(setolt(op1, op2), 1) -> setogt(op2, op1)
339 // setne(setolt(op1, op2), 0) -> setogt(op2, op1)
340 // setne(setolt(op1, op2), 1) -> not(setogt(op2, op1))
341 //===----------------------------------------------------------------------===//
342 let Predicates = [HasV5T] in
343 multiclass eq_ordltPats <PatFrag cmpOp, InstHexagon IntMI,
344 InstHexagon DoubleMI> {
346 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
347 (C2_not (IntMI F32:$src2, F32:$src1))>;
348 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
349 (IntMI F32:$src2, F32:$src1)>;
350 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
351 (IntMI F32:$src2, F32:$src1)>;
352 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
353 (C2_not (IntMI F32:$src2, F32:$src1))>;
356 def: Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
357 (C2_not (DoubleMI F64:$src2, F64:$src1))>;
358 def: Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 1)),
359 (DoubleMI F64:$src2, F64:$src1)>;
360 def: Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
361 (DoubleMI F64:$src2, F64:$src1)>;
362 def: Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
363 (C2_not (DoubleMI F64:$src2, F64:$src1))>;
366 defm : eq_ordltPats<setole, F2_sfcmpge, F2_dfcmpge>;
367 defm : eq_ordltPats<setolt, F2_sfcmpgt, F2_dfcmpgt>;
370 // o. seto inverse of setuo. http://llvm.org/docs/LangRef.html#i_fcmp
371 let Predicates = [HasV5T] in {
372 def: Pat<(i1 (seto F32:$src1, F32:$src2)),
373 (C2_not (F2_sfcmpuo F32:$src2, F32:$src1))>;
374 def: Pat<(i1 (seto F32:$src1, fpimm:$src2)),
375 (C2_not (F2_sfcmpuo (TFRI_f fpimm:$src2), F32:$src1))>;
376 def: Pat<(i1 (seto F64:$src1, F64:$src2)),
377 (C2_not (F2_dfcmpuo F64:$src2, F64:$src1))>;
378 def: Pat<(i1 (seto F64:$src1, fpimm:$src2)),
379 (C2_not (F2_dfcmpuo (CONST64_Float_Real fpimm:$src2), F64:$src1))>;
383 let Predicates = [HasV5T] in {
384 def: Pat<(i1 (setolt F32:$src1, F32:$src2)),
385 (F2_sfcmpgt F32:$src2, F32:$src1)>;
386 def: Pat<(i1 (setolt F32:$src1, fpimm:$src2)),
387 (F2_sfcmpgt (f32 (TFRI_f fpimm:$src2)), F32:$src1)>;
388 def: Pat<(i1 (setolt F64:$src1, F64:$src2)),
389 (F2_dfcmpgt F64:$src2, F64:$src1)>;
390 def: Pat<(i1 (setolt F64:$src1, fpimm:$src2)),
391 (F2_dfcmpgt (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
395 let Predicates = [HasV5T] in {
396 def: Pat<(i1 (setult F32:$src1, F32:$src2)),
397 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
398 (F2_sfcmpgt F32:$src2, F32:$src1))>;
399 def: Pat<(i1 (setult F32:$src1, fpimm:$src2)),
400 (C2_or (F2_sfcmpuo F32:$src1, (TFRI_f fpimm:$src2)),
401 (F2_sfcmpgt (TFRI_f fpimm:$src2), F32:$src1))>;
402 def: Pat<(i1 (setult F64:$src1, F64:$src2)),
403 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
404 (F2_dfcmpgt F64:$src2, F64:$src1))>;
405 def: Pat<(i1 (setult F64:$src1, fpimm:$src2)),
406 (C2_or (F2_dfcmpuo F64:$src1, (CONST64_Float_Real fpimm:$src2)),
407 (F2_dfcmpgt (CONST64_Float_Real fpimm:$src2), F64:$src1))>;
411 let Predicates = [HasV5T] in {
412 // rs <= rt -> rt >= rs.
413 def: Pat<(i1 (setole F32:$src1, F32:$src2)),
414 (F2_sfcmpge F32:$src2, F32:$src1)>;
415 def: Pat<(i1 (setole F32:$src1, fpimm:$src2)),
416 (F2_sfcmpge (TFRI_f fpimm:$src2), F32:$src1)>;
418 // Rss <= Rtt -> Rtt >= Rss.
419 def: Pat<(i1 (setole F64:$src1, F64:$src2)),
420 (F2_dfcmpge F64:$src2, F64:$src1)>;
421 def: Pat<(i1 (setole F64:$src1, fpimm:$src2)),
422 (F2_dfcmpge (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
426 let Predicates = [HasV5T] in {
427 // rs <= rt -> rt >= rs.
428 def: Pat<(i1 (setule F32:$src1, F32:$src2)),
429 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
430 (F2_sfcmpge F32:$src2, F32:$src1))>;
431 def: Pat<(i1 (setule F32:$src1, fpimm:$src2)),
432 (C2_or (F2_sfcmpuo F32:$src1, (TFRI_f fpimm:$src2)),
433 (F2_sfcmpge (TFRI_f fpimm:$src2), F32:$src1))>;
434 def: Pat<(i1 (setule F64:$src1, F64:$src2)),
435 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
436 (F2_dfcmpge F64:$src2, F64:$src1))>;
437 def: Pat<(i1 (setule F64:$src1, fpimm:$src2)),
438 (C2_or (F2_dfcmpuo F64:$src1, (CONST64_Float_Real fpimm:$src2)),
439 (F2_dfcmpge (CONST64_Float_Real fpimm:$src2), F64:$src1))>;
443 let Predicates = [HasV5T] in {
444 def: Pat<(i1 (setone F32:$src1, F32:$src2)),
445 (C2_not (F2_sfcmpeq F32:$src1, F32:$src2))>;
446 def: Pat<(i1 (setone F64:$src1, F64:$src2)),
447 (C2_not (F2_dfcmpeq F64:$src1, F64:$src2))>;
448 def: Pat<(i1 (setone F32:$src1, fpimm:$src2)),
449 (C2_not (F2_sfcmpeq F32:$src1, (TFRI_f fpimm:$src2)))>;
450 def: Pat<(i1 (setone F64:$src1, fpimm:$src2)),
451 (C2_not (F2_dfcmpeq F64:$src1, (CONST64_Float_Real fpimm:$src2)))>;
455 let Predicates = [HasV5T] in {
456 def: Pat<(i1 (setune F32:$src1, F32:$src2)),
457 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
458 (C2_not (F2_sfcmpeq F32:$src1, F32:$src2)))>;
459 def: Pat<(i1 (setune F64:$src1, F64:$src2)),
460 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
461 (C2_not (F2_dfcmpeq F64:$src1, F64:$src2)))>;
462 def: Pat<(i1 (setune F32:$src1, fpimm:$src2)),
463 (C2_or (F2_sfcmpuo F32:$src1, (TFRI_f fpimm:$src2)),
464 (C2_not (F2_sfcmpeq F32:$src1, (TFRI_f fpimm:$src2))))>;
465 def: Pat<(i1 (setune F64:$src1, fpimm:$src2)),
466 (C2_or (F2_dfcmpuo F64:$src1, (CONST64_Float_Real fpimm:$src2)),
467 (C2_not (F2_dfcmpeq F64:$src1,
468 (CONST64_Float_Real fpimm:$src2))))>;
471 // Besides set[o|u][comparions], we also need set[comparisons].
472 let Predicates = [HasV5T] in {
474 def: Pat<(i1 (setlt F32:$src1, F32:$src2)),
475 (F2_sfcmpgt F32:$src2, F32:$src1)>;
476 def: Pat<(i1 (setlt F32:$src1, fpimm:$src2)),
477 (F2_sfcmpgt (TFRI_f fpimm:$src2), F32:$src1)>;
478 def: Pat<(i1 (setlt F64:$src1, F64:$src2)),
479 (F2_dfcmpgt F64:$src2, F64:$src1)>;
480 def: Pat<(i1 (setlt F64:$src1, fpimm:$src2)),
481 (F2_dfcmpgt (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
484 // rs <= rt -> rt >= rs.
485 def: Pat<(i1 (setle F32:$src1, F32:$src2)),
486 (F2_sfcmpge F32:$src2, F32:$src1)>;
487 def: Pat<(i1 (setle F32:$src1, fpimm:$src2)),
488 (F2_sfcmpge (TFRI_f fpimm:$src2), F32:$src1)>;
490 // Rss <= Rtt -> Rtt >= Rss.
491 def: Pat<(i1 (setle F64:$src1, F64:$src2)),
492 (F2_dfcmpge F64:$src2, F64:$src1)>;
493 def: Pat<(i1 (setle F64:$src1, fpimm:$src2)),
494 (F2_dfcmpge (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
497 def: Pat<(i1 (setne F32:$src1, F32:$src2)),
498 (C2_not (F2_sfcmpeq F32:$src1, F32:$src2))>;
499 def: Pat<(i1 (setne F64:$src1, F64:$src2)),
500 (C2_not (F2_dfcmpeq F64:$src1, F64:$src2))>;
501 def: Pat<(i1 (setne F32:$src1, fpimm:$src2)),
502 (C2_not (F2_sfcmpeq F32:$src1, (TFRI_f fpimm:$src2)))>;
503 def: Pat<(i1 (setne F64:$src1, fpimm:$src2)),
504 (C2_not (F2_dfcmpeq F64:$src1, (CONST64_Float_Real fpimm:$src2)))>;
507 // F2 convert template classes:
509 class F2_RDD_RSS_CONVERT<string mnemonic, bits<3> MinOp,
510 SDNode Op, PatLeaf RCOut, PatLeaf RCIn,
512 : SInst <(outs DoubleRegs:$Rdd), (ins DoubleRegs:$Rss),
513 "$Rdd = "#mnemonic#"($Rss)"#chop,
514 [(set RCOut:$Rdd, (Op RCIn:$Rss))], "",
515 S_2op_tc_3or4x_SLOT23> {
521 let Inst{27-21} = 0b0000111;
522 let Inst{20-16} = Rss;
523 let Inst{7-5} = MinOp;
528 class F2_RDD_RS_CONVERT<string mnemonic, bits<3> MinOp,
529 SDNode Op, PatLeaf RCOut, PatLeaf RCIn,
531 : SInst <(outs DoubleRegs:$Rdd), (ins IntRegs:$Rs),
532 "$Rdd = "#mnemonic#"($Rs)"#chop,
533 [(set RCOut:$Rdd, (Op RCIn:$Rs))], "",
534 S_2op_tc_3or4x_SLOT23> {
540 let Inst{27-21} = 0b0100100;
541 let Inst{20-16} = Rs;
542 let Inst{7-5} = MinOp;
546 let isFP = 1, hasNewValue = 1 in
547 class F2_RD_RSS_CONVERT<string mnemonic, bits<3> MinOp,
548 SDNode Op, PatLeaf RCOut, PatLeaf RCIn,
550 : SInst <(outs IntRegs:$Rd), (ins DoubleRegs:$Rss),
551 "$Rd = "#mnemonic#"($Rss)"#chop,
552 [(set RCOut:$Rd, (Op RCIn:$Rss))], "",
553 S_2op_tc_3or4x_SLOT23> {
559 let Inst{27-24} = 0b1000;
560 let Inst{23-21} = MinOp;
561 let Inst{20-16} = Rss;
562 let Inst{7-5} = 0b001;
566 let isFP = 1, hasNewValue = 1 in
567 class F2_RD_RS_CONVERT<string mnemonic, bits<3> MajOp, bits<3> MinOp,
568 SDNode Op, PatLeaf RCOut, PatLeaf RCIn,
570 : SInst <(outs IntRegs:$Rd), (ins IntRegs:$Rs),
571 "$Rd = "#mnemonic#"($Rs)"#chop,
572 [(set RCOut:$Rd, (Op RCIn:$Rs))], "",
573 S_2op_tc_3or4x_SLOT23> {
579 let Inst{27-24} = 0b1011;
580 let Inst{23-21} = MajOp;
581 let Inst{20-16} = Rs;
582 let Inst{7-5} = MinOp;
586 // Convert single precision to double precision and vice-versa.
587 def F2_conv_sf2df : F2_RDD_RS_CONVERT <"convert_sf2df", 0b000,
590 def F2_conv_df2sf : F2_RD_RSS_CONVERT <"convert_df2sf", 0b000,
593 // Convert Integer to Floating Point.
594 def F2_conv_d2sf : F2_RD_RSS_CONVERT <"convert_d2sf", 0b010,
595 sint_to_fp, F32, I64>;
596 def F2_conv_ud2sf : F2_RD_RSS_CONVERT <"convert_ud2sf", 0b001,
597 uint_to_fp, F32, I64>;
598 def F2_conv_uw2sf : F2_RD_RS_CONVERT <"convert_uw2sf", 0b001, 0b000,
599 uint_to_fp, F32, I32>;
600 def F2_conv_w2sf : F2_RD_RS_CONVERT <"convert_w2sf", 0b010, 0b000,
601 sint_to_fp, F32, I32>;
602 def F2_conv_d2df : F2_RDD_RSS_CONVERT <"convert_d2df", 0b011,
603 sint_to_fp, F64, I64>;
604 def F2_conv_ud2df : F2_RDD_RSS_CONVERT <"convert_ud2df", 0b010,
605 uint_to_fp, F64, I64>;
606 def F2_conv_uw2df : F2_RDD_RS_CONVERT <"convert_uw2df", 0b001,
607 uint_to_fp, F64, I32>;
608 def F2_conv_w2df : F2_RDD_RS_CONVERT <"convert_w2df", 0b010,
609 sint_to_fp, F64, I32>;
611 // Convert Floating Point to Integer - default.
612 def F2_conv_df2uw_chop : F2_RD_RSS_CONVERT <"convert_df2uw", 0b101,
613 fp_to_uint, I32, F64, ":chop">;
614 def F2_conv_df2w_chop : F2_RD_RSS_CONVERT <"convert_df2w", 0b111,
615 fp_to_sint, I32, F64, ":chop">;
616 def F2_conv_sf2uw_chop : F2_RD_RS_CONVERT <"convert_sf2uw", 0b011, 0b001,
617 fp_to_uint, I32, F32, ":chop">;
618 def F2_conv_sf2w_chop : F2_RD_RS_CONVERT <"convert_sf2w", 0b100, 0b001,
619 fp_to_sint, I32, F32, ":chop">;
620 def F2_conv_df2d_chop : F2_RDD_RSS_CONVERT <"convert_df2d", 0b110,
621 fp_to_sint, I64, F64, ":chop">;
622 def F2_conv_df2ud_chop : F2_RDD_RSS_CONVERT <"convert_df2ud", 0b111,
623 fp_to_uint, I64, F64, ":chop">;
624 def F2_conv_sf2d_chop : F2_RDD_RS_CONVERT <"convert_sf2d", 0b110,
625 fp_to_sint, I64, F32, ":chop">;
626 def F2_conv_sf2ud_chop : F2_RDD_RS_CONVERT <"convert_sf2ud", 0b101,
627 fp_to_uint, I64, F32, ":chop">;
629 // Convert Floating Point to Integer: non-chopped.
630 let AddedComplexity = 20, Predicates = [HasV5T, IEEERndNearV5T] in {
631 def F2_conv_df2d : F2_RDD_RSS_CONVERT <"convert_df2d", 0b000,
632 fp_to_sint, I64, F64>;
633 def F2_conv_df2ud : F2_RDD_RSS_CONVERT <"convert_df2ud", 0b001,
634 fp_to_uint, I64, F64>;
635 def F2_conv_sf2ud : F2_RDD_RS_CONVERT <"convert_sf2ud", 0b011,
636 fp_to_uint, I64, F32>;
637 def F2_conv_sf2d : F2_RDD_RS_CONVERT <"convert_sf2d", 0b100,
638 fp_to_sint, I64, F32>;
639 def F2_conv_df2uw : F2_RD_RSS_CONVERT <"convert_df2uw", 0b011,
640 fp_to_uint, I32, F64>;
641 def F2_conv_df2w : F2_RD_RSS_CONVERT <"convert_df2w", 0b100,
642 fp_to_sint, I32, F64>;
643 def F2_conv_sf2uw : F2_RD_RS_CONVERT <"convert_sf2uw", 0b011, 0b000,
644 fp_to_uint, I32, F32>;
645 def F2_conv_sf2w : F2_RD_RS_CONVERT <"convert_sf2w", 0b100, 0b000,
646 fp_to_sint, I32, F32>;
650 let isFP = 1, hasNewValue = 1 in
651 def F2_sffixupr: SInst<(outs IntRegs:$Rd), (ins IntRegs:$Rs),
652 "$Rd = sffixupr($Rs)",
653 [], "" , S_2op_tc_3or4x_SLOT23>, Requires<[HasV5T]> {
659 let Inst{27-21} = 0b1011101;
660 let Inst{20-16} = Rs;
661 let Inst{7-5} = 0b000;
665 // Bitcast is different than [fp|sint|uint]_to_[sint|uint|fp].
666 let Predicates = [HasV5T] in {
667 def: Pat <(i32 (bitconvert F32:$src)), (I32:$src)>;
668 def: Pat <(f32 (bitconvert I32:$src)), (F32:$src)>;
669 def: Pat <(i64 (bitconvert F64:$src)), (I64:$src)>;
670 def: Pat <(f64 (bitconvert I64:$src)), (F64:$src)>;
673 // F2_sffma: Floating-point fused multiply add.
674 let isFP = 1, hasNewValue = 1 in
675 class T_sfmpy_acc <bit isSub, bit isLib>
676 : MInst<(outs IntRegs:$Rx),
677 (ins IntRegs:$dst2, IntRegs:$Rs, IntRegs:$Rt),
678 "$Rx "#!if(isSub, "-=","+=")#" sfmpy($Rs, $Rt)"#!if(isLib, ":lib",""),
679 [], "$dst2 = $Rx" , M_tc_3_SLOT23 > ,
687 let Inst{27-21} = 0b1111000;
688 let Inst{20-16} = Rs;
697 def F2_sffma: T_sfmpy_acc <0, 0>;
698 def F2_sffms: T_sfmpy_acc <1, 0>;
699 def F2_sffma_lib: T_sfmpy_acc <0, 1>;
700 def F2_sffms_lib: T_sfmpy_acc <1, 1>;
702 // Floating-point fused multiply add w/ additional scaling (2**pu).
703 let isFP = 1, hasNewValue = 1 in
704 def F2_sffma_sc: MInst <
706 (ins IntRegs:$dst2, IntRegs:$Rs, IntRegs:$Rt, PredRegs:$Pu),
707 "$Rx += sfmpy($Rs, $Rt, $Pu):scale" ,
708 [], "$dst2 = $Rx" , M_tc_3_SLOT23 > ,
717 let Inst{27-21} = 0b1111011;
718 let Inst{20-16} = Rs;
726 let isExtended = 1, isExtentSigned = 1, opExtentBits = 8, opExtendable = 3,
727 isPseudo = 1, InputType = "imm" in
728 def MUX_ir_f : ALU32_rr<(outs IntRegs:$dst),
729 (ins PredRegs:$src1, IntRegs:$src2, f32Ext:$src3),
730 "$dst = mux($src1, $src2, #$src3)",
731 [(set F32:$dst, (f32 (select I1:$src1, F32:$src2, fpimm:$src3)))]>,
734 let isExtended = 1, isExtentSigned = 1, opExtentBits = 8, opExtendable = 2,
735 isPseudo = 1, InputType = "imm" in
736 def MUX_ri_f : ALU32_rr<(outs IntRegs:$dst),
737 (ins PredRegs:$src1, f32Ext:$src2, IntRegs:$src3),
738 "$dst = mux($src1, #$src2, $src3)",
739 [(set F32:$dst, (f32 (select I1:$src1, fpimm:$src2, F32:$src3)))]>,
742 //===----------------------------------------------------------------------===//
743 // :natural forms of vasrh and vasrhub insns
744 //===----------------------------------------------------------------------===//
745 // S5_asrhub_rnd_sat: Vector arithmetic shift right by immediate with round,
746 // saturate, and pack.
747 let Defs = [USR_OVF], hasSideEffects = 0, hasNewValue = 1, opNewValue = 0 in
748 class T_ASRHUB<bit isSat>
749 : SInst <(outs IntRegs:$Rd),
750 (ins DoubleRegs:$Rss, u4Imm:$u4),
751 "$Rd = vasrhub($Rss, #$u4):"#!if(isSat, "sat", "raw"),
752 [], "", S_2op_tc_2_SLOT23>,
760 let Inst{27-21} = 0b1000011;
761 let Inst{20-16} = Rss;
762 let Inst{13-12} = 0b00;
764 let Inst{7-6} = 0b10;
769 def S5_asrhub_rnd_sat : T_ASRHUB <0>;
770 def S5_asrhub_sat : T_ASRHUB <1>;
772 let isAsmParserOnly = 1 in
773 def S5_asrhub_rnd_sat_goodsyntax
774 : SInst <(outs IntRegs:$Rd), (ins DoubleRegs:$Rss, u4Imm:$u4),
775 "$Rd = vasrhub($Rss, #$u4):rnd:sat">, Requires<[HasV5T]>;
777 // S5_vasrhrnd: Vector arithmetic shift right by immediate with round.
778 let hasSideEffects = 0 in
779 def S5_vasrhrnd : SInst <(outs DoubleRegs:$Rdd),
780 (ins DoubleRegs:$Rss, u4Imm:$u4),
781 "$Rdd = vasrh($Rss, #$u4):raw">,
789 let Inst{27-21} = 0b0000001;
790 let Inst{20-16} = Rss;
791 let Inst{13-12} = 0b00;
793 let Inst{7-5} = 0b000;
797 let isAsmParserOnly = 1 in
798 def S5_vasrhrnd_goodsyntax
799 : SInst <(outs DoubleRegs:$Rdd), (ins DoubleRegs:$Rss, u4Imm:$u4),
800 "$Rdd = vasrh($Rss,#$u4):rnd">, Requires<[HasV5T]>;
802 // Floating point reciprocal square root approximation
803 let Uses = [USR], isPredicateLate = 1, isFP = 1,
804 hasSideEffects = 0, hasNewValue = 1, opNewValue = 0,
805 validSubTargets = HasV5SubT in
806 def F2_sfinvsqrta: SInst <
807 (outs IntRegs:$Rd, PredRegs:$Pe),
809 "$Rd, $Pe = sfinvsqrta($Rs)" > ,
817 let Inst{27-21} = 0b1011111;
818 let Inst{20-16} = Rs;
824 // Complex multiply 32x16
825 let Defs = [USR_OVF], Itinerary = S_3op_tc_3x_SLOT23 in {
826 def M4_cmpyi_whc : T_S3op_8<"cmpyiwh", 0b101, 1, 1, 1, 1>;
827 def M4_cmpyr_whc : T_S3op_8<"cmpyrwh", 0b111, 1, 1, 1, 1>;
830 // Classify floating-point value
832 def F2_sfclass : T_TEST_BIT_IMM<"sfclass", 0b111>;
835 def F2_dfclass: ALU64Inst<(outs PredRegs:$Pd), (ins DoubleRegs:$Rss, u5Imm:$u5),
836 "$Pd = dfclass($Rss, #$u5)",
837 [], "" , ALU64_tc_2early_SLOT23 > , Requires<[HasV5T]> {
843 let Inst{27-21} = 0b1100100;
844 let Inst{20-16} = Rss;
845 let Inst{12-10} = 0b000;
847 let Inst{4-3} = 0b10;
851 // Instructions to create floating point constant
852 let hasNewValue = 1, opNewValue = 0 in
853 class T_fimm <string mnemonic, RegisterClass RC, bits<4> RegType, bit isNeg>
854 : ALU64Inst<(outs RC:$dst), (ins u10Imm:$src),
855 "$dst = "#mnemonic#"(#$src)"#!if(isNeg, ":neg", ":pos"),
856 [], "", ALU64_tc_3x_SLOT23>, Requires<[HasV5T]> {
861 let Inst{27-24} = RegType;
863 let Inst{22} = isNeg;
864 let Inst{21} = src{9};
865 let Inst{13-5} = src{8-0};
869 let hasNewValue = 1, opNewValue = 0 in {
870 def F2_sfimm_p : T_fimm <"sfmake", IntRegs, 0b0110, 0>;
871 def F2_sfimm_n : T_fimm <"sfmake", IntRegs, 0b0110, 1>;
874 def F2_dfimm_p : T_fimm <"dfmake", DoubleRegs, 0b1001, 0>;
875 def F2_dfimm_n : T_fimm <"dfmake", DoubleRegs, 0b1001, 1>;
877 def : Pat <(fabs (f32 IntRegs:$src1)),
878 (S2_clrbit_i (f32 IntRegs:$src1), 31)>,
881 def : Pat <(fneg (f32 IntRegs:$src1)),
882 (S2_togglebit_i (f32 IntRegs:$src1), 31)>,