1 //=- HexagonInstrInfoV4.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 V4 instructions in TableGen format.
12 //===----------------------------------------------------------------------===//
14 def addrga: PatLeaf<(i32 AddrGA:$Addr)>;
15 def addrgp: PatLeaf<(i32 AddrGP:$Addr)>;
17 let hasSideEffects = 0 in
18 class T_Immext<Operand ImmType>
19 : EXTENDERInst<(outs), (ins ImmType:$imm),
20 "immext(#$imm)", []> {
24 let Inst{27-16} = imm{31-20};
25 let Inst{13-0} = imm{19-6};
28 def A4_ext : T_Immext<u26_6Imm>;
29 let isCodeGenOnly = 1 in {
31 def A4_ext_b : T_Immext<brtarget>;
33 def A4_ext_c : T_Immext<calltarget>;
34 def A4_ext_g : T_Immext<globaladdress>;
37 def BITPOS32 : SDNodeXForm<imm, [{
38 // Return the bit position we will set [0-31].
40 int32_t imm = N->getSExtValue();
41 return XformMskToBitPosU5Imm(imm);
44 // Fold (add (CONST32 tglobaladdr:$addr) <offset>) into a global address.
45 def FoldGlobalAddr : ComplexPattern<i32, 1, "foldGlobalAddress", [], []>;
47 // Fold (add (CONST32_GP tglobaladdr:$addr) <offset>) into a global address.
48 def FoldGlobalAddrGP : ComplexPattern<i32, 1, "foldGlobalAddressGP", [], []>;
50 def NumUsesBelowThresCONST32 : PatFrag<(ops node:$addr),
51 (HexagonCONST32 node:$addr), [{
52 return hasNumUsesBelowThresGA(N->getOperand(0).getNode());
55 // Hexagon V4 Architecture spec defines 8 instruction classes:
56 // LD ST ALU32 XTYPE J JR MEMOP NV CR SYSTEM(system is not implemented in the
60 // ========================================
61 // Loads (8/16/32/64 bit)
65 // ========================================
66 // Stores (8/16/32/64 bit)
69 // ALU32 Instructions:
70 // ========================================
71 // Arithmetic / Logical (32 bit)
74 // XTYPE Instructions (32/64 bit):
75 // ========================================
76 // Arithmetic, Logical, Bit Manipulation
77 // Multiply (Integer, Fractional, Complex)
78 // Permute / Vector Permute Operations
79 // Predicate Operations
80 // Shift / Shift with Add/Sub/Logical
82 // Vector Halfword (ALU, Shift, Multiply)
83 // Vector Word (ALU, Shift)
86 // ========================================
87 // Jump/Call PC-relative
90 // ========================================
93 // MEMOP Instructions:
94 // ========================================
95 // Operation on memory (8/16/32 bit)
98 // ========================================
103 // ========================================
104 // Control-Register Transfers
105 // Hardware Loop Setup
106 // Predicate Logicals & Reductions
108 // SYSTEM Instructions (not implemented in the compiler):
109 // ========================================
115 //===----------------------------------------------------------------------===//
117 //===----------------------------------------------------------------------===//
119 class T_ALU32_3op_not<string mnemonic, bits<3> MajOp, bits<3> MinOp,
121 : T_ALU32_3op<mnemonic, MajOp, MinOp, OpsRev, 0> {
122 let AsmString = "$Rd = "#mnemonic#"($Rs, ~$Rt)";
125 let BaseOpcode = "andn_rr", CextOpcode = "andn", isCodeGenOnly = 0 in
126 def A4_andn : T_ALU32_3op_not<"and", 0b001, 0b100, 1>;
127 let BaseOpcode = "orn_rr", CextOpcode = "orn", isCodeGenOnly = 0 in
128 def A4_orn : T_ALU32_3op_not<"or", 0b001, 0b101, 1>;
130 let CextOpcode = "rcmp.eq", isCodeGenOnly = 0 in
131 def A4_rcmpeq : T_ALU32_3op<"cmp.eq", 0b011, 0b010, 0, 1>;
132 let CextOpcode = "!rcmp.eq", isCodeGenOnly = 0 in
133 def A4_rcmpneq : T_ALU32_3op<"!cmp.eq", 0b011, 0b011, 0, 1>;
135 let isCodeGenOnly = 0 in {
136 def C4_cmpneq : T_ALU32_3op_cmp<"!cmp.eq", 0b00, 1, 1>;
137 def C4_cmplte : T_ALU32_3op_cmp<"!cmp.gt", 0b10, 1, 0>;
138 def C4_cmplteu : T_ALU32_3op_cmp<"!cmp.gtu", 0b11, 1, 0>;
141 // Pats for instruction selection.
143 // A class to embed the usual comparison patfrags within a zext to i32.
144 // The seteq/setne frags use "lhs" and "rhs" as operands, so use the same
145 // names, or else the frag's "body" won't match the operands.
146 class CmpInReg<PatFrag Op>
147 : PatFrag<(ops node:$lhs, node:$rhs),(i32 (zext (i1 Op.Fragment)))>;
149 def: T_cmp32_rr_pat<A4_rcmpeq, CmpInReg<seteq>, i32>;
150 def: T_cmp32_rr_pat<A4_rcmpneq, CmpInReg<setne>, i32>;
152 def: T_cmp32_rr_pat<C4_cmpneq, setne, i1>;
154 class T_CMP_rrbh<string mnemonic, bits<3> MinOp, bit IsComm>
155 : SInst<(outs PredRegs:$Pd), (ins IntRegs:$Rs, IntRegs:$Rt),
156 "$Pd = "#mnemonic#"($Rs, $Rt)", [], "", S_3op_tc_2early_SLOT23>,
158 let validSubTargets = HasV4SubT;
159 let InputType = "reg";
160 let CextOpcode = mnemonic;
162 let isCommutable = IsComm;
163 let hasSideEffects = 0;
170 let Inst{27-21} = 0b0111110;
171 let Inst{20-16} = Rs;
173 let Inst{7-5} = MinOp;
177 let isCodeGenOnly = 0 in {
178 def A4_cmpbeq : T_CMP_rrbh<"cmpb.eq", 0b110, 1>;
179 def A4_cmpbgt : T_CMP_rrbh<"cmpb.gt", 0b010, 0>;
180 def A4_cmpbgtu : T_CMP_rrbh<"cmpb.gtu", 0b111, 0>;
181 def A4_cmpheq : T_CMP_rrbh<"cmph.eq", 0b011, 1>;
182 def A4_cmphgt : T_CMP_rrbh<"cmph.gt", 0b100, 0>;
183 def A4_cmphgtu : T_CMP_rrbh<"cmph.gtu", 0b101, 0>;
186 class T_CMP_ribh<string mnemonic, bits<2> MajOp, bit IsHalf, bit IsComm,
187 Operand ImmType, bit IsImmExt, bit IsImmSigned, int ImmBits>
188 : ALU64Inst<(outs PredRegs:$Pd), (ins IntRegs:$Rs, ImmType:$Imm),
189 "$Pd = "#mnemonic#"($Rs, #$Imm)", [], "", ALU64_tc_2early_SLOT23>,
191 let validSubTargets = HasV4SubT;
192 let InputType = "imm";
193 let CextOpcode = mnemonic;
195 let isCommutable = IsComm;
196 let hasSideEffects = 0;
197 let isExtendable = IsImmExt;
198 let opExtendable = !if (IsImmExt, 2, 0);
199 let isExtentSigned = IsImmSigned;
200 let opExtentBits = ImmBits;
207 let Inst{27-24} = 0b1101;
208 let Inst{22-21} = MajOp;
209 let Inst{20-16} = Rs;
210 let Inst{12-5} = Imm;
212 let Inst{3} = IsHalf;
216 let isCodeGenOnly = 0 in {
217 def A4_cmpbeqi : T_CMP_ribh<"cmpb.eq", 0b00, 0, 1, u8Imm, 0, 0, 8>;
218 def A4_cmpbgti : T_CMP_ribh<"cmpb.gt", 0b01, 0, 0, s8Imm, 0, 1, 8>;
219 def A4_cmpbgtui : T_CMP_ribh<"cmpb.gtu", 0b10, 0, 0, u7Ext, 1, 0, 7>;
220 def A4_cmpheqi : T_CMP_ribh<"cmph.eq", 0b00, 1, 1, s8Ext, 1, 1, 8>;
221 def A4_cmphgti : T_CMP_ribh<"cmph.gt", 0b01, 1, 0, s8Ext, 1, 1, 8>;
222 def A4_cmphgtui : T_CMP_ribh<"cmph.gtu", 0b10, 1, 0, u7Ext, 1, 0, 7>;
224 class T_RCMP_EQ_ri<string mnemonic, bit IsNeg>
225 : ALU32_ri<(outs IntRegs:$Rd), (ins IntRegs:$Rs, s8Ext:$s8),
226 "$Rd = "#mnemonic#"($Rs, #$s8)", [], "", ALU32_2op_tc_1_SLOT0123>,
228 let validSubTargets = HasV4SubT;
229 let InputType = "imm";
230 let CextOpcode = !if (IsNeg, "!rcmp.eq", "rcmp.eq");
231 let isExtendable = 1;
232 let opExtendable = 2;
233 let isExtentSigned = 1;
234 let opExtentBits = 8;
242 let Inst{27-24} = 0b0011;
244 let Inst{21} = IsNeg;
245 let Inst{20-16} = Rs;
251 let isCodeGenOnly = 0 in {
252 def A4_rcmpeqi : T_RCMP_EQ_ri<"cmp.eq", 0>;
253 def A4_rcmpneqi : T_RCMP_EQ_ri<"!cmp.eq", 1>;
256 def: Pat<(i32 (zext (i1 (seteq (i32 IntRegs:$Rs), s8ExtPred:$s8)))),
257 (A4_rcmpeqi IntRegs:$Rs, s8ExtPred:$s8)>;
258 def: Pat<(i32 (zext (i1 (setne (i32 IntRegs:$Rs), s8ExtPred:$s8)))),
259 (A4_rcmpneqi IntRegs:$Rs, s8ExtPred:$s8)>;
261 // Preserve the S2_tstbit_r generation
262 def: Pat<(i32 (zext (i1 (setne (i32 (and (i32 (shl 1, (i32 IntRegs:$src2))),
263 (i32 IntRegs:$src1))), 0)))),
264 (C2_muxii (S2_tstbit_r IntRegs:$src1, IntRegs:$src2), 1, 0)>;
267 //===----------------------------------------------------------------------===//
269 //===----------------------------------------------------------------------===//
272 //===----------------------------------------------------------------------===//
274 //===----------------------------------------------------------------------===//
276 // Combine a word and an immediate into a register pair.
277 let hasSideEffects = 0, isExtentSigned = 1, isExtendable = 1,
279 class T_Combine1 <bits<2> MajOp, dag ins, string AsmStr>
280 : ALU32Inst <(outs DoubleRegs:$Rdd), ins, AsmStr> {
286 let Inst{27-24} = 0b0011;
287 let Inst{22-21} = MajOp;
288 let Inst{20-16} = Rs;
294 let opExtendable = 2, isCodeGenOnly = 0 in
295 def A4_combineri : T_Combine1<0b00, (ins IntRegs:$Rs, s8Ext:$s8),
296 "$Rdd = combine($Rs, #$s8)">;
298 let opExtendable = 1, isCodeGenOnly = 0 in
299 def A4_combineir : T_Combine1<0b01, (ins s8Ext:$s8, IntRegs:$Rs),
300 "$Rdd = combine(#$s8, $Rs)">;
302 def HexagonWrapperCombineRI_V4 :
303 SDNode<"HexagonISD::WrapperCombineRI_V4", SDTHexagonI64I32I32>;
304 def HexagonWrapperCombineIR_V4 :
305 SDNode<"HexagonISD::WrapperCombineIR_V4", SDTHexagonI64I32I32>;
307 def : Pat <(HexagonWrapperCombineRI_V4 IntRegs:$r, s8ExtPred:$i),
308 (A4_combineri IntRegs:$r, s8ExtPred:$i)>,
311 def : Pat <(HexagonWrapperCombineIR_V4 s8ExtPred:$i, IntRegs:$r),
312 (A4_combineir s8ExtPred:$i, IntRegs:$r)>,
315 // A4_combineii: Set two small immediates.
316 let hasSideEffects = 0, isExtendable = 1, opExtentBits = 6, opExtendable = 2 in
317 def A4_combineii: ALU32Inst<(outs DoubleRegs:$Rdd), (ins s8Imm:$s8, u6Ext:$U6),
318 "$Rdd = combine(#$s8, #$U6)"> {
324 let Inst{27-23} = 0b11001;
325 let Inst{20-16} = U6{5-1};
326 let Inst{13} = U6{0};
331 //===----------------------------------------------------------------------===//
333 //===----------------------------------------------------------------------===//
335 //===----------------------------------------------------------------------===//
337 //===----------------------------------------------------------------------===//
339 def Zext64: OutPatFrag<(ops node:$Rs),
340 (i64 (A4_combineir 0, (i32 $Rs)))>;
341 def Sext64: OutPatFrag<(ops node:$Rs),
342 (i64 (A2_sxtw (i32 $Rs)))>;
344 // Patterns to generate indexed loads with different forms of the address:
347 // - base (without offset).
348 multiclass Loadxm_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
349 PatLeaf ImmPred, InstHexagon MI> {
350 def: Pat<(VT (Load AddrFI:$fi)),
351 (VT (ValueMod (MI AddrFI:$fi, 0)))>;
352 def: Pat<(VT (Load (add IntRegs:$Rs, ImmPred:$Off))),
353 (VT (ValueMod (MI IntRegs:$Rs, imm:$Off)))>;
354 def: Pat<(VT (Load (i32 IntRegs:$Rs))),
355 (VT (ValueMod (MI IntRegs:$Rs, 0)))>;
358 defm: Loadxm_pat<extloadi1, i64, Zext64, s11_0ExtPred, L2_loadrub_io>;
359 defm: Loadxm_pat<extloadi8, i64, Zext64, s11_0ExtPred, L2_loadrub_io>;
360 defm: Loadxm_pat<extloadi16, i64, Zext64, s11_1ExtPred, L2_loadruh_io>;
361 defm: Loadxm_pat<zextloadi1, i64, Zext64, s11_0ExtPred, L2_loadrub_io>;
362 defm: Loadxm_pat<zextloadi8, i64, Zext64, s11_0ExtPred, L2_loadrub_io>;
363 defm: Loadxm_pat<zextloadi16, i64, Zext64, s11_1ExtPred, L2_loadruh_io>;
364 defm: Loadxm_pat<sextloadi8, i64, Sext64, s11_0ExtPred, L2_loadrb_io>;
365 defm: Loadxm_pat<sextloadi16, i64, Sext64, s11_1ExtPred, L2_loadrh_io>;
367 // Map Rdd = anyext(Rs) -> Rdd = combine(#0, Rs).
368 def: Pat<(i64 (anyext (i32 IntRegs:$src1))), (Zext64 IntRegs:$src1)>;
370 //===----------------------------------------------------------------------===//
371 // Template class for load instructions with Absolute set addressing mode.
372 //===----------------------------------------------------------------------===//
373 let isExtended = 1, opExtendable = 2, opExtentBits = 6, addrMode = AbsoluteSet,
374 hasSideEffects = 0 in
375 class T_LD_abs_set<string mnemonic, RegisterClass RC, bits<4>MajOp>:
376 LDInst<(outs RC:$dst1, IntRegs:$dst2),
378 "$dst1 = "#mnemonic#"($dst2 = #$addr)",
386 let Inst{27-25} = 0b101;
387 let Inst{24-21} = MajOp;
388 let Inst{13-12} = 0b01;
389 let Inst{4-0} = dst1;
390 let Inst{20-16} = dst2;
391 let Inst{11-8} = addr{5-2};
392 let Inst{6-5} = addr{1-0};
395 let accessSize = ByteAccess, hasNewValue = 1, isCodeGenOnly = 0 in {
396 def L4_loadrb_ap : T_LD_abs_set <"memb", IntRegs, 0b1000>;
397 def L4_loadrub_ap : T_LD_abs_set <"memub", IntRegs, 0b1001>;
400 let accessSize = HalfWordAccess, hasNewValue = 1, isCodeGenOnly = 0 in {
401 def L4_loadrh_ap : T_LD_abs_set <"memh", IntRegs, 0b1010>;
402 def L4_loadruh_ap : T_LD_abs_set <"memuh", IntRegs, 0b1011>;
405 let accessSize = WordAccess, hasNewValue = 1, isCodeGenOnly = 0 in
406 def L4_loadri_ap : T_LD_abs_set <"memw", IntRegs, 0b1100>;
408 let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
409 def L4_loadrd_ap : T_LD_abs_set <"memd", DoubleRegs, 0b1110>;
410 // Load - Indirect with long offset
411 let InputType = "imm", addrMode = BaseLongOffset, isExtended = 1,
412 opExtentBits = 6, opExtendable = 3 in
413 class T_LoadAbsReg <string mnemonic, string CextOp, RegisterClass RC,
415 : LDInst <(outs RC:$dst), (ins IntRegs:$src1, u2Imm:$src2, u6Ext:$src3),
416 "$dst = "#mnemonic#"($src1<<#$src2 + #$src3)",
422 let CextOpcode = CextOp;
423 let hasNewValue = !if (!eq(!cast<string>(RC), "DoubleRegs"), 0, 1);
426 let Inst{27-25} = 0b110;
427 let Inst{24-21} = MajOp;
428 let Inst{20-16} = src1;
429 let Inst{13} = src2{1};
431 let Inst{11-8} = src3{5-2};
432 let Inst{7} = src2{0};
433 let Inst{6-5} = src3{1-0};
437 let accessSize = ByteAccess, isCodeGenOnly = 0 in {
438 def L4_loadrb_ur : T_LoadAbsReg<"memb", "LDrib", IntRegs, 0b1000>;
439 def L4_loadrub_ur : T_LoadAbsReg<"memub", "LDriub", IntRegs, 0b1001>;
440 def L4_loadalignb_ur : T_LoadAbsReg<"memb_fifo", "LDrib_fifo",
444 let accessSize = HalfWordAccess, isCodeGenOnly = 0 in {
445 def L4_loadrh_ur : T_LoadAbsReg<"memh", "LDrih", IntRegs, 0b1010>;
446 def L4_loadruh_ur : T_LoadAbsReg<"memuh", "LDriuh", IntRegs, 0b1011>;
447 def L4_loadbsw2_ur : T_LoadAbsReg<"membh", "LDribh2", IntRegs, 0b0001>;
448 def L4_loadbzw2_ur : T_LoadAbsReg<"memubh", "LDriubh2", IntRegs, 0b0011>;
449 def L4_loadalignh_ur : T_LoadAbsReg<"memh_fifo", "LDrih_fifo",
453 let accessSize = WordAccess, isCodeGenOnly = 0 in {
454 def L4_loadri_ur : T_LoadAbsReg<"memw", "LDriw", IntRegs, 0b1100>;
455 def L4_loadbsw4_ur : T_LoadAbsReg<"membh", "LDribh4", DoubleRegs, 0b0111>;
456 def L4_loadbzw4_ur : T_LoadAbsReg<"memubh", "LDriubh4", DoubleRegs, 0b0101>;
459 let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
460 def L4_loadrd_ur : T_LoadAbsReg<"memd", "LDrid", DoubleRegs, 0b1110>;
463 multiclass T_LoadAbsReg_Pat <PatFrag ldOp, InstHexagon MI, ValueType VT = i32> {
464 def : Pat <(VT (ldOp (add (shl IntRegs:$src1, u2ImmPred:$src2),
465 (HexagonCONST32 tglobaladdr:$src3)))),
466 (MI IntRegs:$src1, u2ImmPred:$src2, tglobaladdr:$src3)>;
468 def : Pat <(VT (ldOp (add IntRegs:$src1,
469 (HexagonCONST32 tglobaladdr:$src2)))),
470 (MI IntRegs:$src1, 0, tglobaladdr:$src2)>;
473 let AddedComplexity = 60 in {
474 defm : T_LoadAbsReg_Pat <sextloadi8, L4_loadrb_ur>;
475 defm : T_LoadAbsReg_Pat <zextloadi8, L4_loadrub_ur>;
476 defm : T_LoadAbsReg_Pat <extloadi8, L4_loadrub_ur>;
478 defm : T_LoadAbsReg_Pat <sextloadi16, L4_loadrh_ur>;
479 defm : T_LoadAbsReg_Pat <zextloadi16, L4_loadruh_ur>;
480 defm : T_LoadAbsReg_Pat <extloadi16, L4_loadruh_ur>;
482 defm : T_LoadAbsReg_Pat <load, L4_loadri_ur>;
483 defm : T_LoadAbsReg_Pat <load, L4_loadrd_ur, i64>;
486 //===----------------------------------------------------------------------===//
487 // Template classes for the non-predicated load instructions with
488 // base + register offset addressing mode
489 //===----------------------------------------------------------------------===//
490 class T_load_rr <string mnemonic, RegisterClass RC, bits<3> MajOp>:
491 LDInst<(outs RC:$dst), (ins IntRegs:$src1, IntRegs:$src2, u2Imm:$u2),
492 "$dst = "#mnemonic#"($src1 + $src2<<#$u2)",
493 [], "", V4LDST_tc_ld_SLOT01>, ImmRegShl, AddrModeRel {
501 let Inst{27-24} = 0b1010;
502 let Inst{23-21} = MajOp;
503 let Inst{20-16} = src1;
504 let Inst{12-8} = src2;
505 let Inst{13} = u2{1};
510 //===----------------------------------------------------------------------===//
511 // Template classes for the predicated load instructions with
512 // base + register offset addressing mode
513 //===----------------------------------------------------------------------===//
514 let isPredicated = 1 in
515 class T_pload_rr <string mnemonic, RegisterClass RC, bits<3> MajOp,
516 bit isNot, bit isPredNew>:
517 LDInst <(outs RC:$dst),
518 (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3, u2Imm:$u2),
519 !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
520 ") ")#"$dst = "#mnemonic#"($src2+$src3<<#$u2)",
521 [], "", V4LDST_tc_ld_SLOT01>, AddrModeRel {
528 let isPredicatedFalse = isNot;
529 let isPredicatedNew = isPredNew;
533 let Inst{27-26} = 0b00;
534 let Inst{25} = isPredNew;
535 let Inst{24} = isNot;
536 let Inst{23-21} = MajOp;
537 let Inst{20-16} = src2;
538 let Inst{12-8} = src3;
539 let Inst{13} = u2{1};
541 let Inst{6-5} = src1;
545 //===----------------------------------------------------------------------===//
546 // multiclass for load instructions with base + register offset
548 //===----------------------------------------------------------------------===//
549 let hasSideEffects = 0, addrMode = BaseRegOffset in
550 multiclass ld_idxd_shl <string mnemonic, string CextOp, RegisterClass RC,
552 let CextOpcode = CextOp, BaseOpcode = CextOp#_indexed_shl,
553 InputType = "reg" in {
554 let isPredicable = 1 in
555 def L4_#NAME#_rr : T_load_rr <mnemonic, RC, MajOp>;
558 def L4_p#NAME#t_rr : T_pload_rr <mnemonic, RC, MajOp, 0, 0>;
559 def L4_p#NAME#f_rr : T_pload_rr <mnemonic, RC, MajOp, 1, 0>;
562 def L4_p#NAME#tnew_rr : T_pload_rr <mnemonic, RC, MajOp, 0, 1>;
563 def L4_p#NAME#fnew_rr : T_pload_rr <mnemonic, RC, MajOp, 1, 1>;
567 let hasNewValue = 1, accessSize = ByteAccess, isCodeGenOnly = 0 in {
568 defm loadrb : ld_idxd_shl<"memb", "LDrib", IntRegs, 0b000>;
569 defm loadrub : ld_idxd_shl<"memub", "LDriub", IntRegs, 0b001>;
572 let hasNewValue = 1, accessSize = HalfWordAccess, isCodeGenOnly = 0 in {
573 defm loadrh : ld_idxd_shl<"memh", "LDrih", IntRegs, 0b010>;
574 defm loadruh : ld_idxd_shl<"memuh", "LDriuh", IntRegs, 0b011>;
577 let hasNewValue = 1, accessSize = WordAccess, isCodeGenOnly = 0 in
578 defm loadri : ld_idxd_shl<"memw", "LDriw", IntRegs, 0b100>;
580 let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
581 defm loadrd : ld_idxd_shl<"memd", "LDrid", DoubleRegs, 0b110>;
583 // 'def pats' for load instructions with base + register offset and non-zero
584 // immediate value. Immediate value is used to left-shift the second
586 class Loadxs_pat<PatFrag Load, ValueType VT, InstHexagon MI>
587 : Pat<(VT (Load (add (i32 IntRegs:$Rs),
588 (i32 (shl (i32 IntRegs:$Rt), u2ImmPred:$u2))))),
589 (VT (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2))>;
591 let AddedComplexity = 40 in {
592 def: Loadxs_pat<extloadi8, i32, L4_loadrub_rr>;
593 def: Loadxs_pat<zextloadi8, i32, L4_loadrub_rr>;
594 def: Loadxs_pat<sextloadi8, i32, L4_loadrb_rr>;
595 def: Loadxs_pat<extloadi16, i32, L4_loadruh_rr>;
596 def: Loadxs_pat<zextloadi16, i32, L4_loadruh_rr>;
597 def: Loadxs_pat<sextloadi16, i32, L4_loadrh_rr>;
598 def: Loadxs_pat<load, i32, L4_loadri_rr>;
599 def: Loadxs_pat<load, i64, L4_loadrd_rr>;
602 // 'def pats' for load instruction base + register offset and
603 // zero immediate value.
604 class Loadxs_simple_pat<PatFrag Load, ValueType VT, InstHexagon MI>
605 : Pat<(VT (Load (add (i32 IntRegs:$Rs), (i32 IntRegs:$Rt)))),
606 (VT (MI IntRegs:$Rs, IntRegs:$Rt, 0))>;
608 let AddedComplexity = 20 in {
609 def: Loadxs_simple_pat<extloadi8, i32, L4_loadrub_rr>;
610 def: Loadxs_simple_pat<zextloadi8, i32, L4_loadrub_rr>;
611 def: Loadxs_simple_pat<sextloadi8, i32, L4_loadrb_rr>;
612 def: Loadxs_simple_pat<extloadi16, i32, L4_loadruh_rr>;
613 def: Loadxs_simple_pat<zextloadi16, i32, L4_loadruh_rr>;
614 def: Loadxs_simple_pat<sextloadi16, i32, L4_loadrh_rr>;
615 def: Loadxs_simple_pat<load, i32, L4_loadri_rr>;
616 def: Loadxs_simple_pat<load, i64, L4_loadrd_rr>;
620 def: Pat<(i64 (zext (i1 PredRegs:$src1))),
621 (Zext64 (C2_muxii PredRegs:$src1, 1, 0))>;
624 def: Pat<(i64 (zext (i32 IntRegs:$src1))),
625 (Zext64 IntRegs:$src1)>;
628 def: Pat <(i64 (zextloadi32 ADDRriS11_2:$src1)),
629 (i64 (A4_combineir 0, (L2_loadri_io AddrFI:$src1, 0)))>,
632 let AddedComplexity = 100 in
633 def: Pat <(i64 (zextloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))),
634 (i64 (A4_combineir 0, (L2_loadri_io IntRegs:$src1,
635 s11_2ExtPred:$offset)))>,
639 def: Pat <(i64 (extloadi32 ADDRriS11_2:$src1)),
640 (i64 (A4_combineir 0, (L2_loadri_io AddrFI:$src1, 0)))>,
643 let AddedComplexity = 100 in
644 def: Pat <(i64 (extloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))),
645 (i64 (A4_combineir 0, (L2_loadri_io IntRegs:$src1,
646 s11_2ExtPred:$offset)))>,
651 //===----------------------------------------------------------------------===//
653 //===----------------------------------------------------------------------===//
655 //===----------------------------------------------------------------------===//
657 //===----------------------------------------------------------------------===//
659 //===----------------------------------------------------------------------===//
660 // Template class for store instructions with Absolute set addressing mode.
661 //===----------------------------------------------------------------------===//
662 let isExtended = 1, opExtendable = 1, opExtentBits = 6,
663 addrMode = AbsoluteSet, isNVStorable = 1 in
664 class T_ST_absset <string mnemonic, string BaseOp, RegisterClass RC,
665 bits<3> MajOp, MemAccessSize AccessSz, bit isHalf = 0>
666 : STInst<(outs IntRegs:$dst),
667 (ins u6Ext:$addr, RC:$src),
668 mnemonic#"($dst = #$addr) = $src"#!if(isHalf, ".h","")>, NewValueRel {
672 let accessSize = AccessSz;
673 let BaseOpcode = BaseOp#"_AbsSet";
677 let Inst{27-24} = 0b1011;
678 let Inst{23-21} = MajOp;
679 let Inst{20-16} = dst;
681 let Inst{12-8} = src;
683 let Inst{5-0} = addr;
686 def S4_storerb_ap : T_ST_absset <"memb", "STrib", IntRegs, 0b000, ByteAccess>;
687 def S4_storerh_ap : T_ST_absset <"memh", "STrih", IntRegs, 0b010,
689 def S4_storeri_ap : T_ST_absset <"memw", "STriw", IntRegs, 0b100, WordAccess>;
691 let isNVStorable = 0 in {
692 def S4_storerf_ap : T_ST_absset <"memh", "STrif", IntRegs,
693 0b011, HalfWordAccess, 1>;
694 def S4_storerd_ap : T_ST_absset <"memd", "STrid", DoubleRegs,
695 0b110, DoubleWordAccess>;
698 let opExtendable = 1, isNewValue = 1, isNVStore = 1, opNewValue = 2,
699 isExtended = 1, opExtentBits= 6 in
700 class T_ST_absset_nv <string mnemonic, string BaseOp, bits<2> MajOp,
701 MemAccessSize AccessSz >
702 : NVInst <(outs IntRegs:$dst),
703 (ins u6Ext:$addr, IntRegs:$src),
704 mnemonic#"($dst = #$addr) = $src.new">, NewValueRel {
708 let accessSize = AccessSz;
709 let BaseOpcode = BaseOp#"_AbsSet";
713 let Inst{27-21} = 0b1011101;
714 let Inst{20-16} = dst;
715 let Inst{13-11} = 0b000;
716 let Inst{12-11} = MajOp;
717 let Inst{10-8} = src;
719 let Inst{5-0} = addr;
722 let mayStore = 1, addrMode = AbsoluteSet, isCodeGenOnly = 0 in {
723 def S4_storerbnew_ap : T_ST_absset_nv <"memb", "STrib", 0b00, ByteAccess>;
724 def S4_storerhnew_ap : T_ST_absset_nv <"memh", "STrih", 0b01, HalfWordAccess>;
725 def S4_storerinew_ap : T_ST_absset_nv <"memw", "STriw", 0b10, WordAccess>;
728 let isExtended = 1, opExtendable = 2, opExtentBits = 6, InputType = "imm",
729 addrMode = BaseLongOffset, AddedComplexity = 40 in
730 class T_StoreAbsReg <string mnemonic, string CextOp, RegisterClass RC,
731 bits<3> MajOp, MemAccessSize AccessSz, bit isHalf = 0>
733 (ins IntRegs:$src1, u2Imm:$src2, u6Ext:$src3, RC:$src4),
734 mnemonic#"($src1<<#$src2 + #$src3) = $src4"#!if(isHalf, ".h",""),
735 []>, ImmRegShl, NewValueRel {
742 let accessSize = AccessSz;
743 let CextOpcode = CextOp;
744 let BaseOpcode = CextOp#"_shl";
747 let Inst{27-24} =0b1101;
748 let Inst{23-21} = MajOp;
749 let Inst{20-16} = src1;
750 let Inst{13} = src2{1};
751 let Inst{12-8} = src4;
753 let Inst{6} = src2{0};
754 let Inst{5-0} = src3;
757 let isCodeGenOnly = 0 in {
758 def S4_storerb_ur : T_StoreAbsReg <"memb", "STrib", IntRegs, 0b000, ByteAccess>;
759 def S4_storerh_ur : T_StoreAbsReg <"memh", "STrih", IntRegs, 0b010,
761 def S4_storerf_ur : T_StoreAbsReg <"memh", "STrif", IntRegs, 0b011,
763 def S4_storeri_ur : T_StoreAbsReg <"memw", "STriw", IntRegs, 0b100, WordAccess>;
764 def S4_storerd_ur : T_StoreAbsReg <"memd", "STrid", DoubleRegs, 0b110,
768 let AddedComplexity = 40 in
769 multiclass T_StoreAbsReg_Pats <InstHexagon MI, RegisterClass RC, ValueType VT,
771 def : Pat<(stOp (VT RC:$src4),
772 (add (shl (i32 IntRegs:$src1), u2ImmPred:$src2),
773 u0AlwaysExtPred:$src3)),
774 (MI IntRegs:$src1, u2ImmPred:$src2, u0AlwaysExtPred:$src3, RC:$src4)>;
776 def : Pat<(stOp (VT RC:$src4),
777 (add (shl IntRegs:$src1, u2ImmPred:$src2),
778 (HexagonCONST32 tglobaladdr:$src3))),
779 (MI IntRegs:$src1, u2ImmPred:$src2, tglobaladdr:$src3, RC:$src4)>;
781 def : Pat<(stOp (VT RC:$src4),
782 (add IntRegs:$src1, (HexagonCONST32 tglobaladdr:$src3))),
783 (MI IntRegs:$src1, 0, tglobaladdr:$src3, RC:$src4)>;
786 defm : T_StoreAbsReg_Pats <S4_storerd_ur, DoubleRegs, i64, store>;
787 defm : T_StoreAbsReg_Pats <S4_storeri_ur, IntRegs, i32, store>;
788 defm : T_StoreAbsReg_Pats <S4_storerb_ur, IntRegs, i32, truncstorei8>;
789 defm : T_StoreAbsReg_Pats <S4_storerh_ur, IntRegs, i32, truncstorei16>;
791 let mayStore = 1, isNVStore = 1, isExtended = 1, addrMode = BaseLongOffset,
792 opExtentBits = 6, isNewValue = 1, opNewValue = 3, opExtendable = 2 in
793 class T_StoreAbsRegNV <string mnemonic, string CextOp, bits<2> MajOp,
794 MemAccessSize AccessSz>
796 (ins IntRegs:$src1, u2Imm:$src2, u6Ext:$src3, IntRegs:$src4),
797 mnemonic#"($src1<<#$src2 + #$src3) = $src4.new">, NewValueRel {
803 let CextOpcode = CextOp;
804 let BaseOpcode = CextOp#"_shl";
807 let Inst{27-21} = 0b1101101;
808 let Inst{12-11} = 0b00;
810 let Inst{20-16} = src1;
811 let Inst{13} = src2{1};
812 let Inst{12-11} = MajOp;
813 let Inst{10-8} = src4;
814 let Inst{6} = src2{0};
815 let Inst{5-0} = src3;
818 let isCodeGenOnly = 0 in {
819 def S4_storerbnew_ur : T_StoreAbsRegNV <"memb", "STrib", 0b00, ByteAccess>;
820 def S4_storerhnew_ur : T_StoreAbsRegNV <"memh", "STrih", 0b01, HalfWordAccess>;
821 def S4_storerinew_ur : T_StoreAbsRegNV <"memw", "STriw", 0b10, WordAccess>;
824 //===----------------------------------------------------------------------===//
825 // Template classes for the non-predicated store instructions with
826 // base + register offset addressing mode
827 //===----------------------------------------------------------------------===//
828 let isPredicable = 1 in
829 class T_store_rr <string mnemonic, RegisterClass RC, bits<3> MajOp, bit isH>
830 : STInst < (outs ), (ins IntRegs:$Rs, IntRegs:$Ru, u2Imm:$u2, RC:$Rt),
831 mnemonic#"($Rs + $Ru<<#$u2) = $Rt"#!if(isH, ".h",""),
832 [],"",V4LDST_tc_st_SLOT01>, ImmRegShl, AddrModeRel {
841 let Inst{27-24} = 0b1011;
842 let Inst{23-21} = MajOp;
843 let Inst{20-16} = Rs;
845 let Inst{13} = u2{1};
850 //===----------------------------------------------------------------------===//
851 // Template classes for the predicated store instructions with
852 // base + register offset addressing mode
853 //===----------------------------------------------------------------------===//
854 let isPredicated = 1 in
855 class T_pstore_rr <string mnemonic, RegisterClass RC, bits<3> MajOp,
856 bit isNot, bit isPredNew, bit isH>
858 (ins PredRegs:$Pv, IntRegs:$Rs, IntRegs:$Ru, u2Imm:$u2, RC:$Rt),
860 !if(isNot, "if (!$Pv", "if ($Pv")#!if(isPredNew, ".new) ",
861 ") ")#mnemonic#"($Rs+$Ru<<#$u2) = $Rt"#!if(isH, ".h",""),
862 [], "", V4LDST_tc_st_SLOT01> , AddrModeRel{
869 let isPredicatedFalse = isNot;
870 let isPredicatedNew = isPredNew;
874 let Inst{27-26} = 0b01;
875 let Inst{25} = isPredNew;
876 let Inst{24} = isNot;
877 let Inst{23-21} = MajOp;
878 let Inst{20-16} = Rs;
880 let Inst{13} = u2{1};
886 //===----------------------------------------------------------------------===//
887 // Template classes for the new-value store instructions with
888 // base + register offset addressing mode
889 //===----------------------------------------------------------------------===//
890 let isPredicable = 1, isNewValue = 1, opNewValue = 3 in
891 class T_store_new_rr <string mnemonic, bits<2> MajOp> :
892 NVInst < (outs ), (ins IntRegs:$Rs, IntRegs:$Ru, u2Imm:$u2, IntRegs:$Nt),
893 mnemonic#"($Rs + $Ru<<#$u2) = $Nt.new",
894 [],"",V4LDST_tc_st_SLOT0>, ImmRegShl, AddrModeRel {
903 let Inst{27-21} = 0b1011101;
904 let Inst{20-16} = Rs;
906 let Inst{13} = u2{1};
908 let Inst{4-3} = MajOp;
912 //===----------------------------------------------------------------------===//
913 // Template classes for the predicated new-value store instructions with
914 // base + register offset addressing mode
915 //===----------------------------------------------------------------------===//
916 let isPredicated = 1, isNewValue = 1, opNewValue = 4 in
917 class T_pstore_new_rr <string mnemonic, bits<2> MajOp, bit isNot, bit isPredNew>
919 (ins PredRegs:$Pv, IntRegs:$Rs, IntRegs:$Ru, u2Imm:$u2, IntRegs:$Nt),
920 !if(isNot, "if (!$Pv", "if ($Pv")#!if(isPredNew, ".new) ",
921 ") ")#mnemonic#"($Rs+$Ru<<#$u2) = $Nt.new",
922 [], "", V4LDST_tc_st_SLOT0>, AddrModeRel {
929 let isPredicatedFalse = isNot;
930 let isPredicatedNew = isPredNew;
933 let Inst{27-26} = 0b01;
934 let Inst{25} = isPredNew;
935 let Inst{24} = isNot;
936 let Inst{23-21} = 0b101;
937 let Inst{20-16} = Rs;
939 let Inst{13} = u2{1};
942 let Inst{4-3} = MajOp;
946 //===----------------------------------------------------------------------===//
947 // multiclass for store instructions with base + register offset addressing
949 //===----------------------------------------------------------------------===//
950 let isNVStorable = 1 in
951 multiclass ST_Idxd_shl<string mnemonic, string CextOp, RegisterClass RC,
952 bits<3> MajOp, bit isH = 0> {
953 let CextOpcode = CextOp, BaseOpcode = CextOp#_indexed_shl in {
954 def S4_#NAME#_rr : T_store_rr <mnemonic, RC, MajOp, isH>;
957 def S4_p#NAME#t_rr : T_pstore_rr <mnemonic, RC, MajOp, 0, 0, isH>;
958 def S4_p#NAME#f_rr : T_pstore_rr <mnemonic, RC, MajOp, 1, 0, isH>;
961 def S4_p#NAME#tnew_rr : T_pstore_rr <mnemonic, RC, MajOp, 0, 1, isH>;
962 def S4_p#NAME#fnew_rr : T_pstore_rr <mnemonic, RC, MajOp, 1, 1, isH>;
966 //===----------------------------------------------------------------------===//
967 // multiclass for new-value store instructions with base + register offset
969 //===----------------------------------------------------------------------===//
970 let mayStore = 1, isNVStore = 1 in
971 multiclass ST_Idxd_shl_nv <string mnemonic, string CextOp, RegisterClass RC,
973 let CextOpcode = CextOp, BaseOpcode = CextOp#_indexed_shl in {
974 def S4_#NAME#new_rr : T_store_new_rr<mnemonic, MajOp>;
977 def S4_p#NAME#newt_rr : T_pstore_new_rr <mnemonic, MajOp, 0, 0>;
978 def S4_p#NAME#newf_rr : T_pstore_new_rr <mnemonic, MajOp, 1, 0>;
981 def S4_p#NAME#newtnew_rr : T_pstore_new_rr <mnemonic, MajOp, 0, 1>;
982 def S4_p#NAME#newfnew_rr : T_pstore_new_rr <mnemonic, MajOp, 1, 1>;
986 let addrMode = BaseRegOffset, InputType = "reg", hasSideEffects = 0,
987 isCodeGenOnly = 0 in {
988 let accessSize = ByteAccess in
989 defm storerb: ST_Idxd_shl<"memb", "STrib", IntRegs, 0b000>,
990 ST_Idxd_shl_nv<"memb", "STrib", IntRegs, 0b00>;
992 let accessSize = HalfWordAccess in
993 defm storerh: ST_Idxd_shl<"memh", "STrih", IntRegs, 0b010>,
994 ST_Idxd_shl_nv<"memh", "STrih", IntRegs, 0b01>;
996 let accessSize = WordAccess in
997 defm storeri: ST_Idxd_shl<"memw", "STriw", IntRegs, 0b100>,
998 ST_Idxd_shl_nv<"memw", "STriw", IntRegs, 0b10>;
1000 let isNVStorable = 0, accessSize = DoubleWordAccess in
1001 defm storerd: ST_Idxd_shl<"memd", "STrid", DoubleRegs, 0b110>;
1003 let isNVStorable = 0, accessSize = HalfWordAccess in
1004 defm storerf: ST_Idxd_shl<"memh", "STrif", IntRegs, 0b011, 1>;
1007 class Storexs_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
1008 : Pat<(Store Value:$Ru, (add (i32 IntRegs:$Rs),
1009 (i32 (shl (i32 IntRegs:$Rt), u2ImmPred:$u2)))),
1010 (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2, Value:$Ru)>;
1012 let AddedComplexity = 40 in {
1013 def: Storexs_pat<truncstorei8, I32, S4_storerb_rr>;
1014 def: Storexs_pat<truncstorei16, I32, S4_storerh_rr>;
1015 def: Storexs_pat<store, I32, S4_storeri_rr>;
1016 def: Storexs_pat<store, I64, S4_storerd_rr>;
1019 // memd(Rx++#s4:3)=Rtt
1020 // memd(Rx++#s4:3:circ(Mu))=Rtt
1021 // memd(Rx++I:circ(Mu))=Rtt
1023 // memd(Rx++Mu:brev)=Rtt
1024 // memd(gp+#u16:3)=Rtt
1026 // Store doubleword conditionally.
1027 // if ([!]Pv[.new]) memd(#u6)=Rtt
1028 // TODO: needs to be implemented.
1030 //===----------------------------------------------------------------------===//
1032 //===----------------------------------------------------------------------===//
1033 let isPredicable = 1, isExtendable = 1, isExtentSigned = 1, opExtentBits = 8,
1035 class T_StoreImm <string mnemonic, Operand OffsetOp, bits<2> MajOp >
1036 : STInst <(outs ), (ins IntRegs:$Rs, OffsetOp:$offset, s8Ext:$S8),
1037 mnemonic#"($Rs+#$offset)=#$S8",
1038 [], "", V4LDST_tc_st_SLOT01>,
1039 ImmRegRel, PredNewRel {
1045 string OffsetOpStr = !cast<string>(OffsetOp);
1046 let offsetBits = !if (!eq(OffsetOpStr, "u6_2Imm"), offset{7-2},
1047 !if (!eq(OffsetOpStr, "u6_1Imm"), offset{6-1},
1048 /* u6_0Imm */ offset{5-0}));
1050 let IClass = 0b0011;
1052 let Inst{27-25} = 0b110;
1053 let Inst{22-21} = MajOp;
1054 let Inst{20-16} = Rs;
1055 let Inst{12-7} = offsetBits;
1056 let Inst{13} = S8{7};
1057 let Inst{6-0} = S8{6-0};
1060 let isPredicated = 1, isExtendable = 1, isExtentSigned = 1, opExtentBits = 6,
1062 class T_StoreImm_pred <string mnemonic, Operand OffsetOp, bits<2> MajOp,
1063 bit isPredNot, bit isPredNew >
1065 (ins PredRegs:$Pv, IntRegs:$Rs, OffsetOp:$offset, s6Ext:$S6),
1066 !if(isPredNot, "if (!$Pv", "if ($Pv")#!if(isPredNew, ".new) ",
1067 ") ")#mnemonic#"($Rs+#$offset)=#$S6",
1068 [], "", V4LDST_tc_st_SLOT01>,
1069 ImmRegRel, PredNewRel {
1076 string OffsetOpStr = !cast<string>(OffsetOp);
1077 let offsetBits = !if (!eq(OffsetOpStr, "u6_2Imm"), offset{7-2},
1078 !if (!eq(OffsetOpStr, "u6_1Imm"), offset{6-1},
1079 /* u6_0Imm */ offset{5-0}));
1080 let isPredicatedNew = isPredNew;
1081 let isPredicatedFalse = isPredNot;
1083 let IClass = 0b0011;
1085 let Inst{27-25} = 0b100;
1086 let Inst{24} = isPredNew;
1087 let Inst{23} = isPredNot;
1088 let Inst{22-21} = MajOp;
1089 let Inst{20-16} = Rs;
1090 let Inst{13} = S6{5};
1091 let Inst{12-7} = offsetBits;
1093 let Inst{4-0} = S6{4-0};
1097 //===----------------------------------------------------------------------===//
1098 // multiclass for store instructions with base + immediate offset
1099 // addressing mode and immediate stored value.
1100 // mem[bhw](Rx++#s4:3)=#s8
1101 // if ([!]Pv[.new]) mem[bhw](Rx++#s4:3)=#s6
1102 //===----------------------------------------------------------------------===//
1104 multiclass ST_Imm_Pred <string mnemonic, Operand OffsetOp, bits<2> MajOp,
1106 def _io : T_StoreImm_pred <mnemonic, OffsetOp, MajOp, PredNot, 0>;
1108 def new_io : T_StoreImm_pred <mnemonic, OffsetOp, MajOp, PredNot, 1>;
1111 multiclass ST_Imm <string mnemonic, string CextOp, Operand OffsetOp,
1113 let CextOpcode = CextOp, BaseOpcode = CextOp#_imm in {
1114 def _io : T_StoreImm <mnemonic, OffsetOp, MajOp>;
1116 defm t : ST_Imm_Pred <mnemonic, OffsetOp, MajOp, 0>;
1117 defm f : ST_Imm_Pred <mnemonic, OffsetOp, MajOp, 1>;
1121 let hasSideEffects = 0, validSubTargets = HasV4SubT, addrMode = BaseImmOffset,
1122 InputType = "imm", isCodeGenOnly = 0 in {
1123 let accessSize = ByteAccess in
1124 defm S4_storeirb : ST_Imm<"memb", "STrib", u6_0Imm, 0b00>;
1126 let accessSize = HalfWordAccess in
1127 defm S4_storeirh : ST_Imm<"memh", "STrih", u6_1Imm, 0b01>;
1129 let accessSize = WordAccess in
1130 defm S4_storeiri : ST_Imm<"memw", "STriw", u6_2Imm, 0b10>;
1133 let Predicates = [HasV4T], AddedComplexity = 10 in {
1134 def: Pat<(truncstorei8 s8ExtPred:$src3, (add IntRegs:$src1, u6_0ImmPred:$src2)),
1135 (S4_storeirb_io IntRegs:$src1, u6_0ImmPred:$src2, s8ExtPred:$src3)>;
1137 def: Pat<(truncstorei16 s8ExtPred:$src3, (add IntRegs:$src1,
1138 u6_1ImmPred:$src2)),
1139 (S4_storeirh_io IntRegs:$src1, u6_1ImmPred:$src2, s8ExtPred:$src3)>;
1141 def: Pat<(store s8ExtPred:$src3, (add IntRegs:$src1, u6_2ImmPred:$src2)),
1142 (S4_storeiri_io IntRegs:$src1, u6_2ImmPred:$src2, s8ExtPred:$src3)>;
1145 let AddedComplexity = 6 in
1146 def : Pat <(truncstorei8 s8ExtPred:$src2, (i32 IntRegs:$src1)),
1147 (S4_storeirb_io IntRegs:$src1, 0, s8ExtPred:$src2)>,
1150 // memb(Rx++#s4:0:circ(Mu))=Rt
1151 // memb(Rx++I:circ(Mu))=Rt
1153 // memb(Rx++Mu:brev)=Rt
1154 // memb(gp+#u16:0)=Rt
1158 // TODO: needs to be implemented
1159 // memh(Re=#U6)=Rt.H
1160 // memh(Rs+#s11:1)=Rt.H
1161 let AddedComplexity = 6 in
1162 def : Pat <(truncstorei16 s8ExtPred:$src2, (i32 IntRegs:$src1)),
1163 (S4_storeirh_io IntRegs:$src1, 0, s8ExtPred:$src2)>,
1166 // memh(Rs+Ru<<#u2)=Rt.H
1167 // TODO: needs to be implemented.
1169 // memh(Ru<<#u2+#U6)=Rt.H
1170 // memh(Rx++#s4:1:circ(Mu))=Rt.H
1171 // memh(Rx++#s4:1:circ(Mu))=Rt
1172 // memh(Rx++I:circ(Mu))=Rt.H
1173 // memh(Rx++I:circ(Mu))=Rt
1174 // memh(Rx++Mu)=Rt.H
1176 // memh(Rx++Mu:brev)=Rt.H
1177 // memh(Rx++Mu:brev)=Rt
1178 // memh(gp+#u16:1)=Rt
1179 // if ([!]Pv[.new]) memh(#u6)=Rt.H
1180 // if ([!]Pv[.new]) memh(#u6)=Rt
1183 // if ([!]Pv[.new]) memh(Rs+#u6:1)=Rt.H
1184 // TODO: needs to be implemented.
1186 // if ([!]Pv[.new]) memh(Rx++#s4:1)=Rt.H
1187 // TODO: Needs to be implemented.
1191 // TODO: Needs to be implemented.
1193 let AddedComplexity = 6 in
1194 def : Pat <(store s8ExtPred:$src2, (i32 IntRegs:$src1)),
1195 (S4_storeiri_io IntRegs:$src1, 0, s8ExtPred:$src2)>,
1198 // memw(Rx++#s4:2)=Rt
1199 // memw(Rx++#s4:2:circ(Mu))=Rt
1200 // memw(Rx++I:circ(Mu))=Rt
1202 // memw(Rx++Mu:brev)=Rt
1204 //===----------------------------------------------------------------------===
1206 //===----------------------------------------------------------------------===
1209 //===----------------------------------------------------------------------===//
1211 //===----------------------------------------------------------------------===//
1213 let opNewValue = 2, opExtendable = 1, isExtentSigned = 1, isPredicable = 1 in
1214 class T_store_io_nv <string mnemonic, RegisterClass RC,
1215 Operand ImmOp, bits<2>MajOp>
1216 : NVInst_V4 <(outs),
1217 (ins IntRegs:$src1, ImmOp:$src2, RC:$src3),
1218 mnemonic#"($src1+#$src2) = $src3.new",
1219 [],"",ST_tc_st_SLOT0> {
1221 bits<13> src2; // Actual address offset
1223 bits<11> offsetBits; // Represents offset encoding
1225 let opExtentBits = !if (!eq(mnemonic, "memb"), 11,
1226 !if (!eq(mnemonic, "memh"), 12,
1227 !if (!eq(mnemonic, "memw"), 13, 0)));
1229 let opExtentAlign = !if (!eq(mnemonic, "memb"), 0,
1230 !if (!eq(mnemonic, "memh"), 1,
1231 !if (!eq(mnemonic, "memw"), 2, 0)));
1233 let offsetBits = !if (!eq(mnemonic, "memb"), src2{10-0},
1234 !if (!eq(mnemonic, "memh"), src2{11-1},
1235 !if (!eq(mnemonic, "memw"), src2{12-2}, 0)));
1237 let IClass = 0b1010;
1240 let Inst{26-25} = offsetBits{10-9};
1241 let Inst{24-21} = 0b1101;
1242 let Inst{20-16} = src1;
1243 let Inst{13} = offsetBits{8};
1244 let Inst{12-11} = MajOp;
1245 let Inst{10-8} = src3;
1246 let Inst{7-0} = offsetBits{7-0};
1249 let opExtendable = 2, opNewValue = 3, isPredicated = 1 in
1250 class T_pstore_io_nv <string mnemonic, RegisterClass RC, Operand predImmOp,
1251 bits<2>MajOp, bit PredNot, bit isPredNew>
1252 : NVInst_V4 <(outs),
1253 (ins PredRegs:$src1, IntRegs:$src2, predImmOp:$src3, RC:$src4),
1254 !if(PredNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
1255 ") ")#mnemonic#"($src2+#$src3) = $src4.new",
1256 [],"",V2LDST_tc_st_SLOT0> {
1261 bits<6> offsetBits; // Represents offset encoding
1263 let isPredicatedNew = isPredNew;
1264 let isPredicatedFalse = PredNot;
1265 let opExtentBits = !if (!eq(mnemonic, "memb"), 6,
1266 !if (!eq(mnemonic, "memh"), 7,
1267 !if (!eq(mnemonic, "memw"), 8, 0)));
1269 let opExtentAlign = !if (!eq(mnemonic, "memb"), 0,
1270 !if (!eq(mnemonic, "memh"), 1,
1271 !if (!eq(mnemonic, "memw"), 2, 0)));
1273 let offsetBits = !if (!eq(mnemonic, "memb"), src3{5-0},
1274 !if (!eq(mnemonic, "memh"), src3{6-1},
1275 !if (!eq(mnemonic, "memw"), src3{7-2}, 0)));
1277 let IClass = 0b0100;
1280 let Inst{26} = PredNot;
1281 let Inst{25} = isPredNew;
1282 let Inst{24-21} = 0b0101;
1283 let Inst{20-16} = src2;
1284 let Inst{13} = offsetBits{5};
1285 let Inst{12-11} = MajOp;
1286 let Inst{10-8} = src4;
1287 let Inst{7-3} = offsetBits{4-0};
1289 let Inst{1-0} = src1;
1292 // multiclass for new-value store instructions with base + immediate offset.
1294 let mayStore = 1, isNVStore = 1, isNewValue = 1, hasSideEffects = 0,
1296 multiclass ST_Idxd_nv<string mnemonic, string CextOp, RegisterClass RC,
1297 Operand ImmOp, Operand predImmOp, bits<2> MajOp> {
1299 let CextOpcode = CextOp, BaseOpcode = CextOp#_indexed in {
1300 def S2_#NAME#new_io : T_store_io_nv <mnemonic, RC, ImmOp, MajOp>;
1302 def S2_p#NAME#newt_io :T_pstore_io_nv <mnemonic, RC, predImmOp, MajOp, 0, 0>;
1303 def S2_p#NAME#newf_io :T_pstore_io_nv <mnemonic, RC, predImmOp, MajOp, 1, 0>;
1305 def S4_p#NAME#newtnew_io :T_pstore_io_nv <mnemonic, RC, predImmOp,
1307 def S4_p#NAME#newfnew_io :T_pstore_io_nv <mnemonic, RC, predImmOp,
1312 let addrMode = BaseImmOffset, InputType = "imm", isCodeGenOnly = 0 in {
1313 let accessSize = ByteAccess in
1314 defm storerb: ST_Idxd_nv<"memb", "STrib", IntRegs, s11_0Ext,
1315 u6_0Ext, 0b00>, AddrModeRel;
1317 let accessSize = HalfWordAccess, opExtentAlign = 1 in
1318 defm storerh: ST_Idxd_nv<"memh", "STrih", IntRegs, s11_1Ext,
1319 u6_1Ext, 0b01>, AddrModeRel;
1321 let accessSize = WordAccess, opExtentAlign = 2 in
1322 defm storeri: ST_Idxd_nv<"memw", "STriw", IntRegs, s11_2Ext,
1323 u6_2Ext, 0b10>, AddrModeRel;
1326 //===----------------------------------------------------------------------===//
1327 // Post increment loads with register offset.
1328 //===----------------------------------------------------------------------===//
1330 let hasNewValue = 1, isCodeGenOnly = 0 in
1331 def L2_loadbsw2_pr : T_load_pr <"membh", IntRegs, 0b0001, HalfWordAccess>;
1333 let isCodeGenOnly = 0 in
1334 def L2_loadbsw4_pr : T_load_pr <"membh", DoubleRegs, 0b0111, WordAccess>;
1336 //===----------------------------------------------------------------------===//
1337 // Template class for non-predicated post increment .new stores
1338 // mem[bhwd](Rx++#s4:[0123])=Nt.new
1339 //===----------------------------------------------------------------------===//
1340 let isPredicable = 1, hasSideEffects = 0, validSubTargets = HasV4SubT,
1341 addrMode = PostInc, isNVStore = 1, isNewValue = 1, opNewValue = 3 in
1342 class T_StorePI_nv <string mnemonic, Operand ImmOp, bits<2> MajOp >
1343 : NVInstPI_V4 <(outs IntRegs:$_dst_),
1344 (ins IntRegs:$src1, ImmOp:$offset, IntRegs:$src2),
1345 mnemonic#"($src1++#$offset) = $src2.new",
1346 [], "$src1 = $_dst_">,
1353 string ImmOpStr = !cast<string>(ImmOp);
1354 let offsetBits = !if (!eq(ImmOpStr, "s4_2Imm"), offset{5-2},
1355 !if (!eq(ImmOpStr, "s4_1Imm"), offset{4-1},
1356 /* s4_0Imm */ offset{3-0}));
1357 let IClass = 0b1010;
1359 let Inst{27-21} = 0b1011101;
1360 let Inst{20-16} = src1;
1362 let Inst{12-11} = MajOp;
1363 let Inst{10-8} = src2;
1365 let Inst{6-3} = offsetBits;
1369 //===----------------------------------------------------------------------===//
1370 // Template class for predicated post increment .new stores
1371 // if([!]Pv[.new]) mem[bhwd](Rx++#s4:[0123])=Nt.new
1372 //===----------------------------------------------------------------------===//
1373 let isPredicated = 1, hasSideEffects = 0, validSubTargets = HasV4SubT,
1374 addrMode = PostInc, isNVStore = 1, isNewValue = 1, opNewValue = 4 in
1375 class T_StorePI_nv_pred <string mnemonic, Operand ImmOp,
1376 bits<2> MajOp, bit isPredNot, bit isPredNew >
1377 : NVInstPI_V4 <(outs IntRegs:$_dst_),
1378 (ins PredRegs:$src1, IntRegs:$src2,
1379 ImmOp:$offset, IntRegs:$src3),
1380 !if(isPredNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
1381 ") ")#mnemonic#"($src2++#$offset) = $src3.new",
1382 [], "$src2 = $_dst_">,
1390 string ImmOpStr = !cast<string>(ImmOp);
1391 let offsetBits = !if (!eq(ImmOpStr, "s4_2Imm"), offset{5-2},
1392 !if (!eq(ImmOpStr, "s4_1Imm"), offset{4-1},
1393 /* s4_0Imm */ offset{3-0}));
1394 let isPredicatedNew = isPredNew;
1395 let isPredicatedFalse = isPredNot;
1397 let IClass = 0b1010;
1399 let Inst{27-21} = 0b1011101;
1400 let Inst{20-16} = src2;
1402 let Inst{12-11} = MajOp;
1403 let Inst{10-8} = src3;
1404 let Inst{7} = isPredNew;
1405 let Inst{6-3} = offsetBits;
1406 let Inst{2} = isPredNot;
1407 let Inst{1-0} = src1;
1410 multiclass ST_PostInc_Pred_nv<string mnemonic, Operand ImmOp,
1411 bits<2> MajOp, bit PredNot> {
1412 def _pi : T_StorePI_nv_pred <mnemonic, ImmOp, MajOp, PredNot, 0>;
1415 def new_pi : T_StorePI_nv_pred <mnemonic, ImmOp, MajOp, PredNot, 1>;
1418 multiclass ST_PostInc_nv<string mnemonic, string BaseOp, Operand ImmOp,
1420 let BaseOpcode = "POST_"#BaseOp in {
1421 def S2_#NAME#_pi : T_StorePI_nv <mnemonic, ImmOp, MajOp>;
1424 defm S2_p#NAME#t : ST_PostInc_Pred_nv <mnemonic, ImmOp, MajOp, 0>;
1425 defm S2_p#NAME#f : ST_PostInc_Pred_nv <mnemonic, ImmOp, MajOp, 1>;
1429 let accessSize = ByteAccess, isCodeGenOnly = 0 in
1430 defm storerbnew: ST_PostInc_nv <"memb", "STrib", s4_0Imm, 0b00>;
1432 let accessSize = HalfWordAccess, isCodeGenOnly = 0 in
1433 defm storerhnew: ST_PostInc_nv <"memh", "STrih", s4_1Imm, 0b01>;
1435 let accessSize = WordAccess, isCodeGenOnly = 0 in
1436 defm storerinew: ST_PostInc_nv <"memw", "STriw", s4_2Imm, 0b10>;
1438 //===----------------------------------------------------------------------===//
1439 // Template class for post increment .new stores with register offset
1440 //===----------------------------------------------------------------------===//
1441 let isNewValue = 1, mayStore = 1, isNVStore = 1, opNewValue = 3 in
1442 class T_StorePI_RegNV <string mnemonic, bits<2> MajOp, MemAccessSize AccessSz>
1443 : NVInstPI_V4 <(outs IntRegs:$_dst_),
1444 (ins IntRegs:$src1, ModRegs:$src2, IntRegs:$src3),
1445 #mnemonic#"($src1++$src2) = $src3.new",
1446 [], "$src1 = $_dst_"> {
1450 let accessSize = AccessSz;
1452 let IClass = 0b1010;
1454 let Inst{27-21} = 0b1101101;
1455 let Inst{20-16} = src1;
1456 let Inst{13} = src2;
1457 let Inst{12-11} = MajOp;
1458 let Inst{10-8} = src3;
1462 let isCodeGenOnly = 0 in {
1463 def S2_storerbnew_pr : T_StorePI_RegNV<"memb", 0b00, ByteAccess>;
1464 def S2_storerhnew_pr : T_StorePI_RegNV<"memh", 0b01, HalfWordAccess>;
1465 def S2_storerinew_pr : T_StorePI_RegNV<"memw", 0b10, WordAccess>;
1468 // memb(Rx++#s4:0:circ(Mu))=Nt.new
1469 // memb(Rx++I:circ(Mu))=Nt.new
1470 // memb(Rx++Mu)=Nt.new
1471 // memb(Rx++Mu:brev)=Nt.new
1472 // memh(Rx++#s4:1:circ(Mu))=Nt.new
1473 // memh(Rx++I:circ(Mu))=Nt.new
1474 // memh(Rx++Mu)=Nt.new
1475 // memh(Rx++Mu:brev)=Nt.new
1477 // memw(Rx++#s4:2:circ(Mu))=Nt.new
1478 // memw(Rx++I:circ(Mu))=Nt.new
1479 // memw(Rx++Mu)=Nt.new
1480 // memw(Rx++Mu:brev)=Nt.new
1482 //===----------------------------------------------------------------------===//
1484 //===----------------------------------------------------------------------===//
1486 //===----------------------------------------------------------------------===//
1488 //===----------------------------------------------------------------------===//
1490 //===----------------------------------------------------------------------===//
1491 // multiclass/template class for the new-value compare jumps with the register
1493 //===----------------------------------------------------------------------===//
1495 let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 11,
1496 opExtentAlign = 2 in
1497 class NVJrr_template<string mnemonic, bits<3> majOp, bit NvOpNum,
1498 bit isNegCond, bit isTak>
1500 (ins IntRegs:$src1, IntRegs:$src2, brtarget:$offset),
1501 "if ("#!if(isNegCond, "!","")#mnemonic#
1502 "($src1"#!if(!eq(NvOpNum, 0),".new, ",", ")#
1503 "$src2"#!if(!eq(NvOpNum, 1),".new))","))")#" jump:"
1504 #!if(isTak, "t","nt")#" $offset", []> {
1508 bits<3> Ns; // New-Value Operand
1509 bits<5> RegOp; // Non-New-Value Operand
1512 let isTaken = isTak;
1513 let isPredicatedFalse = isNegCond;
1514 let opNewValue{0} = NvOpNum;
1516 let Ns = !if(!eq(NvOpNum, 0), src1{2-0}, src2{2-0});
1517 let RegOp = !if(!eq(NvOpNum, 0), src2, src1);
1519 let IClass = 0b0010;
1521 let Inst{25-23} = majOp;
1522 let Inst{22} = isNegCond;
1523 let Inst{18-16} = Ns;
1524 let Inst{13} = isTak;
1525 let Inst{12-8} = RegOp;
1526 let Inst{21-20} = offset{10-9};
1527 let Inst{7-1} = offset{8-2};
1531 multiclass NVJrr_cond<string mnemonic, bits<3> majOp, bit NvOpNum,
1533 // Branch not taken:
1534 def _nt_V4: NVJrr_template<mnemonic, majOp, NvOpNum, isNegCond, 0>;
1536 def _t_V4: NVJrr_template<mnemonic, majOp, NvOpNum, isNegCond, 1>;
1539 // NvOpNum = 0 -> First Operand is a new-value Register
1540 // NvOpNum = 1 -> Second Operand is a new-value Register
1542 multiclass NVJrr_base<string mnemonic, string BaseOp, bits<3> majOp,
1544 let BaseOpcode = BaseOp#_NVJ in {
1545 defm _t_Jumpnv : NVJrr_cond<mnemonic, majOp, NvOpNum, 0>; // True cond
1546 defm _f_Jumpnv : NVJrr_cond<mnemonic, majOp, NvOpNum, 1>; // False cond
1550 // if ([!]cmp.eq(Ns.new,Rt)) jump:[n]t #r9:2
1551 // if ([!]cmp.gt(Ns.new,Rt)) jump:[n]t #r9:2
1552 // if ([!]cmp.gtu(Ns.new,Rt)) jump:[n]t #r9:2
1553 // if ([!]cmp.gt(Rt,Ns.new)) jump:[n]t #r9:2
1554 // if ([!]cmp.gtu(Rt,Ns.new)) jump:[n]t #r9:2
1556 let isPredicated = 1, isBranch = 1, isNewValue = 1, isTerminator = 1,
1557 Defs = [PC], hasSideEffects = 0, validSubTargets = HasV4SubT,
1558 isCodeGenOnly = 0 in {
1559 defm CMPEQrr : NVJrr_base<"cmp.eq", "CMPEQ", 0b000, 0>, PredRel;
1560 defm CMPGTrr : NVJrr_base<"cmp.gt", "CMPGT", 0b001, 0>, PredRel;
1561 defm CMPGTUrr : NVJrr_base<"cmp.gtu", "CMPGTU", 0b010, 0>, PredRel;
1562 defm CMPLTrr : NVJrr_base<"cmp.gt", "CMPLT", 0b011, 1>, PredRel;
1563 defm CMPLTUrr : NVJrr_base<"cmp.gtu", "CMPLTU", 0b100, 1>, PredRel;
1566 //===----------------------------------------------------------------------===//
1567 // multiclass/template class for the new-value compare jumps instruction
1568 // with a register and an unsigned immediate (U5) operand.
1569 //===----------------------------------------------------------------------===//
1571 let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 11,
1572 opExtentAlign = 2 in
1573 class NVJri_template<string mnemonic, bits<3> majOp, bit isNegCond,
1576 (ins IntRegs:$src1, u5Imm:$src2, brtarget:$offset),
1577 "if ("#!if(isNegCond, "!","")#mnemonic#"($src1.new, #$src2)) jump:"
1578 #!if(isTak, "t","nt")#" $offset", []> {
1580 let isTaken = isTak;
1581 let isPredicatedFalse = isNegCond;
1582 let isTaken = isTak;
1588 let IClass = 0b0010;
1590 let Inst{25-23} = majOp;
1591 let Inst{22} = isNegCond;
1592 let Inst{18-16} = src1;
1593 let Inst{13} = isTak;
1594 let Inst{12-8} = src2;
1595 let Inst{21-20} = offset{10-9};
1596 let Inst{7-1} = offset{8-2};
1599 multiclass NVJri_cond<string mnemonic, bits<3> majOp, bit isNegCond> {
1600 // Branch not taken:
1601 def _nt_V4: NVJri_template<mnemonic, majOp, isNegCond, 0>;
1603 def _t_V4: NVJri_template<mnemonic, majOp, isNegCond, 1>;
1606 multiclass NVJri_base<string mnemonic, string BaseOp, bits<3> majOp> {
1607 let BaseOpcode = BaseOp#_NVJri in {
1608 defm _t_Jumpnv : NVJri_cond<mnemonic, majOp, 0>; // True Cond
1609 defm _f_Jumpnv : NVJri_cond<mnemonic, majOp, 1>; // False cond
1613 // if ([!]cmp.eq(Ns.new,#U5)) jump:[n]t #r9:2
1614 // if ([!]cmp.gt(Ns.new,#U5)) jump:[n]t #r9:2
1615 // if ([!]cmp.gtu(Ns.new,#U5)) jump:[n]t #r9:2
1617 let isPredicated = 1, isBranch = 1, isNewValue = 1, isTerminator = 1,
1618 Defs = [PC], hasSideEffects = 0, validSubTargets = HasV4SubT,
1619 isCodeGenOnly = 0 in {
1620 defm CMPEQri : NVJri_base<"cmp.eq", "CMPEQ", 0b000>, PredRel;
1621 defm CMPGTri : NVJri_base<"cmp.gt", "CMPGT", 0b001>, PredRel;
1622 defm CMPGTUri : NVJri_base<"cmp.gtu", "CMPGTU", 0b010>, PredRel;
1625 //===----------------------------------------------------------------------===//
1626 // multiclass/template class for the new-value compare jumps instruction
1627 // with a register and an hardcoded 0/-1 immediate value.
1628 //===----------------------------------------------------------------------===//
1630 let isExtendable = 1, opExtendable = 1, isExtentSigned = 1, opExtentBits = 11,
1631 opExtentAlign = 2 in
1632 class NVJ_ConstImm_template<string mnemonic, bits<3> majOp, string ImmVal,
1633 bit isNegCond, bit isTak>
1635 (ins IntRegs:$src1, brtarget:$offset),
1636 "if ("#!if(isNegCond, "!","")#mnemonic
1637 #"($src1.new, #"#ImmVal#")) jump:"
1638 #!if(isTak, "t","nt")#" $offset", []> {
1640 let isTaken = isTak;
1641 let isPredicatedFalse = isNegCond;
1642 let isTaken = isTak;
1646 let IClass = 0b0010;
1648 let Inst{25-23} = majOp;
1649 let Inst{22} = isNegCond;
1650 let Inst{18-16} = src1;
1651 let Inst{13} = isTak;
1652 let Inst{21-20} = offset{10-9};
1653 let Inst{7-1} = offset{8-2};
1656 multiclass NVJ_ConstImm_cond<string mnemonic, bits<3> majOp, string ImmVal,
1658 // Branch not taken:
1659 def _nt_V4: NVJ_ConstImm_template<mnemonic, majOp, ImmVal, isNegCond, 0>;
1661 def _t_V4: NVJ_ConstImm_template<mnemonic, majOp, ImmVal, isNegCond, 1>;
1664 multiclass NVJ_ConstImm_base<string mnemonic, string BaseOp, bits<3> majOp,
1666 let BaseOpcode = BaseOp#_NVJ_ConstImm in {
1667 defm _t_Jumpnv : NVJ_ConstImm_cond<mnemonic, majOp, ImmVal, 0>; // True
1668 defm _f_Jumpnv : NVJ_ConstImm_cond<mnemonic, majOp, ImmVal, 1>; // False
1672 // if ([!]tstbit(Ns.new,#0)) jump:[n]t #r9:2
1673 // if ([!]cmp.eq(Ns.new,#-1)) jump:[n]t #r9:2
1674 // if ([!]cmp.gt(Ns.new,#-1)) jump:[n]t #r9:2
1676 let isPredicated = 1, isBranch = 1, isNewValue = 1, isTerminator=1,
1677 Defs = [PC], hasSideEffects = 0, isCodeGenOnly = 0 in {
1678 defm TSTBIT0 : NVJ_ConstImm_base<"tstbit", "TSTBIT", 0b011, "0">, PredRel;
1679 defm CMPEQn1 : NVJ_ConstImm_base<"cmp.eq", "CMPEQ", 0b100, "-1">, PredRel;
1680 defm CMPGTn1 : NVJ_ConstImm_base<"cmp.gt", "CMPGT", 0b101, "-1">, PredRel;
1683 // J4_hintjumpr: Hint indirect conditional jump.
1684 let isBranch = 1, isIndirectBranch = 1, hasSideEffects = 0, isCodeGenOnly = 0 in
1685 def J4_hintjumpr: JRInst <
1690 let IClass = 0b0101;
1691 let Inst{27-21} = 0b0010101;
1692 let Inst{20-16} = Rs;
1695 //===----------------------------------------------------------------------===//
1697 //===----------------------------------------------------------------------===//
1699 //===----------------------------------------------------------------------===//
1701 //===----------------------------------------------------------------------===//
1704 let hasNewValue = 1, isExtendable = 1, opExtendable = 1,
1705 isExtentSigned = 0, opExtentBits = 6, hasSideEffects = 0,
1706 Uses = [PC], validSubTargets = HasV4SubT, isCodeGenOnly = 0 in
1707 def C4_addipc : CRInst <(outs IntRegs:$Rd), (ins u6Ext:$u6),
1708 "$Rd = add(pc, #$u6)", [], "", CR_tc_2_SLOT3 > {
1712 let IClass = 0b0110;
1713 let Inst{27-16} = 0b101001001001;
1714 let Inst{12-7} = u6;
1720 let hasSideEffects = 0 in
1721 class T_LOGICAL_3OP<string MnOp1, string MnOp2, bits<2> OpBits, bit IsNeg>
1722 : CRInst<(outs PredRegs:$Pd),
1723 (ins PredRegs:$Ps, PredRegs:$Pt, PredRegs:$Pu),
1724 "$Pd = " # MnOp1 # "($Ps, " # MnOp2 # "($Pt, " #
1725 !if (IsNeg,"!","") # "$Pu))",
1726 [], "", CR_tc_2early_SLOT23> {
1732 let IClass = 0b0110;
1733 let Inst{27-24} = 0b1011;
1734 let Inst{23} = IsNeg;
1735 let Inst{22-21} = OpBits;
1737 let Inst{17-16} = Ps;
1744 let isCodeGenOnly = 0 in {
1745 def C4_and_and : T_LOGICAL_3OP<"and", "and", 0b00, 0>;
1746 def C4_and_or : T_LOGICAL_3OP<"and", "or", 0b01, 0>;
1747 def C4_or_and : T_LOGICAL_3OP<"or", "and", 0b10, 0>;
1748 def C4_or_or : T_LOGICAL_3OP<"or", "or", 0b11, 0>;
1749 def C4_and_andn : T_LOGICAL_3OP<"and", "and", 0b00, 1>;
1750 def C4_and_orn : T_LOGICAL_3OP<"and", "or", 0b01, 1>;
1751 def C4_or_andn : T_LOGICAL_3OP<"or", "and", 0b10, 1>;
1752 def C4_or_orn : T_LOGICAL_3OP<"or", "or", 0b11, 1>;
1755 //===----------------------------------------------------------------------===//
1757 //===----------------------------------------------------------------------===//
1759 //===----------------------------------------------------------------------===//
1761 //===----------------------------------------------------------------------===//
1763 // Logical with-not instructions.
1764 let validSubTargets = HasV4SubT, isCodeGenOnly = 0 in {
1765 def A4_andnp : T_ALU64_logical<"and", 0b001, 1, 0, 1>;
1766 def A4_ornp : T_ALU64_logical<"or", 0b011, 1, 0, 1>;
1769 let hasNewValue = 1, hasSideEffects = 0, isCodeGenOnly = 0 in
1770 def S4_parity: ALU64Inst<(outs IntRegs:$Rd), (ins IntRegs:$Rs, IntRegs:$Rt),
1771 "$Rd = parity($Rs, $Rt)", [], "", ALU64_tc_2_SLOT23> {
1776 let IClass = 0b1101;
1777 let Inst{27-21} = 0b0101111;
1778 let Inst{20-16} = Rs;
1779 let Inst{12-8} = Rt;
1782 // Add and accumulate.
1783 // Rd=add(Rs,add(Ru,#s6))
1784 let isExtentSigned = 1, hasNewValue = 1, isExtendable = 1, opExtentBits = 6,
1785 opExtendable = 3, isCodeGenOnly = 0 in
1786 def S4_addaddi : ALU64Inst <(outs IntRegs:$Rd),
1787 (ins IntRegs:$Rs, IntRegs:$Ru, s6Ext:$s6),
1788 "$Rd = add($Rs, add($Ru, #$s6))" ,
1789 [(set (i32 IntRegs:$Rd), (add (i32 IntRegs:$Rs),
1790 (add (i32 IntRegs:$Ru), s6_16ExtPred:$s6)))],
1791 "", ALU64_tc_2_SLOT23> {
1797 let IClass = 0b1101;
1799 let Inst{27-23} = 0b10110;
1800 let Inst{22-21} = s6{5-4};
1801 let Inst{20-16} = Rs;
1802 let Inst{13} = s6{3};
1803 let Inst{12-8} = Rd;
1804 let Inst{7-5} = s6{2-0};
1808 let isExtentSigned = 1, hasSideEffects = 0, hasNewValue = 1, isExtendable = 1,
1809 opExtentBits = 6, opExtendable = 2, isCodeGenOnly = 0 in
1810 def S4_subaddi: ALU64Inst <(outs IntRegs:$Rd),
1811 (ins IntRegs:$Rs, s6Ext:$s6, IntRegs:$Ru),
1812 "$Rd = add($Rs, sub(#$s6, $Ru))",
1813 [], "", ALU64_tc_2_SLOT23> {
1819 let IClass = 0b1101;
1821 let Inst{27-23} = 0b10111;
1822 let Inst{22-21} = s6{5-4};
1823 let Inst{20-16} = Rs;
1824 let Inst{13} = s6{3};
1825 let Inst{12-8} = Rd;
1826 let Inst{7-5} = s6{2-0};
1831 // Rdd=extract(Rss,#u6,#U6)
1832 // Rdd=extract(Rss,Rtt)
1833 // Rd=extract(Rs,Rtt)
1834 // Rd=extract(Rs,#u5,#U5)
1836 let isCodeGenOnly = 0 in {
1837 def S4_extractp_rp : T_S3op_64 < "extract", 0b11, 0b100, 0>;
1838 def S4_extractp : T_S2op_extract <"extract", 0b1010, DoubleRegs, u6Imm>;
1841 let hasNewValue = 1, isCodeGenOnly = 0 in {
1842 def S4_extract_rp : T_S3op_extract<"extract", 0b01>;
1843 def S4_extract : T_S2op_extract <"extract", 0b1101, IntRegs, u5Imm>;
1846 // Complex add/sub halfwords/words
1847 let Defs = [USR_OVF], isCodeGenOnly = 0 in {
1848 def S4_vxaddsubh : T_S3op_64 < "vxaddsubh", 0b01, 0b100, 0, 1>;
1849 def S4_vxaddsubw : T_S3op_64 < "vxaddsubw", 0b01, 0b000, 0, 1>;
1850 def S4_vxsubaddh : T_S3op_64 < "vxsubaddh", 0b01, 0b110, 0, 1>;
1851 def S4_vxsubaddw : T_S3op_64 < "vxsubaddw", 0b01, 0b010, 0, 1>;
1854 let Defs = [USR_OVF], isCodeGenOnly = 0 in {
1855 def S4_vxaddsubhr : T_S3op_64 < "vxaddsubh", 0b11, 0b000, 0, 1, 1, 1>;
1856 def S4_vxsubaddhr : T_S3op_64 < "vxsubaddh", 0b11, 0b010, 0, 1, 1, 1>;
1859 let Itinerary = M_tc_3x_SLOT23, Defs = [USR_OVF], isCodeGenOnly = 0 in {
1860 def M4_mac_up_s1_sat: T_MType_acc_rr<"+= mpy", 0b011, 0b000, 0, [], 0, 1, 1>;
1861 def M4_nac_up_s1_sat: T_MType_acc_rr<"-= mpy", 0b011, 0b001, 0, [], 0, 1, 1>;
1864 // Logical xor with xor accumulation.
1865 // Rxx^=xor(Rss,Rtt)
1866 let hasSideEffects = 0, isCodeGenOnly = 0 in
1868 : SInst <(outs DoubleRegs:$Rxx),
1869 (ins DoubleRegs:$dst2, DoubleRegs:$Rss, DoubleRegs:$Rtt),
1870 "$Rxx ^= xor($Rss, $Rtt)",
1871 [(set (i64 DoubleRegs:$Rxx),
1872 (xor (i64 DoubleRegs:$dst2), (xor (i64 DoubleRegs:$Rss),
1873 (i64 DoubleRegs:$Rtt))))],
1874 "$dst2 = $Rxx", S_3op_tc_1_SLOT23> {
1879 let IClass = 0b1100;
1881 let Inst{27-23} = 0b10101;
1882 let Inst{20-16} = Rss;
1883 let Inst{12-8} = Rtt;
1884 let Inst{4-0} = Rxx;
1887 // Rotate and reduce bytes
1888 // Rdd=vrcrotate(Rss,Rt,#u2)
1889 let hasSideEffects = 0, isCodeGenOnly = 0 in
1891 : SInst <(outs DoubleRegs:$Rdd),
1892 (ins DoubleRegs:$Rss, IntRegs:$Rt, u2Imm:$u2),
1893 "$Rdd = vrcrotate($Rss, $Rt, #$u2)",
1894 [], "", S_3op_tc_3x_SLOT23> {
1900 let IClass = 0b1100;
1902 let Inst{27-22} = 0b001111;
1903 let Inst{20-16} = Rss;
1904 let Inst{13} = u2{1};
1905 let Inst{12-8} = Rt;
1906 let Inst{7-6} = 0b11;
1907 let Inst{5} = u2{0};
1908 let Inst{4-0} = Rdd;
1911 // Rotate and reduce bytes with accumulation
1912 // Rxx+=vrcrotate(Rss,Rt,#u2)
1913 let hasSideEffects = 0, isCodeGenOnly = 0 in
1914 def S4_vrcrotate_acc
1915 : SInst <(outs DoubleRegs:$Rxx),
1916 (ins DoubleRegs:$dst2, DoubleRegs:$Rss, IntRegs:$Rt, u2Imm:$u2),
1917 "$Rxx += vrcrotate($Rss, $Rt, #$u2)", [],
1918 "$dst2 = $Rxx", S_3op_tc_3x_SLOT23> {
1924 let IClass = 0b1100;
1926 let Inst{27-21} = 0b1011101;
1927 let Inst{20-16} = Rss;
1928 let Inst{13} = u2{1};
1929 let Inst{12-8} = Rt;
1930 let Inst{5} = u2{0};
1931 let Inst{4-0} = Rxx;
1935 // Vector reduce conditional negate halfwords
1936 let hasSideEffects = 0, isCodeGenOnly = 0 in
1938 : SInst <(outs DoubleRegs:$Rxx),
1939 (ins DoubleRegs:$dst2, DoubleRegs:$Rss, IntRegs:$Rt),
1940 "$Rxx += vrcnegh($Rss, $Rt)", [],
1941 "$dst2 = $Rxx", S_3op_tc_3x_SLOT23> {
1946 let IClass = 0b1100;
1948 let Inst{27-21} = 0b1011001;
1949 let Inst{20-16} = Rss;
1951 let Inst{12-8} = Rt;
1952 let Inst{7-5} = 0b111;
1953 let Inst{4-0} = Rxx;
1957 let isCodeGenOnly = 0 in
1958 def A4_bitspliti : T_S2op_2_di <"bitsplit", 0b110, 0b100>;
1960 // Arithmetic/Convergent round
1961 let isCodeGenOnly = 0 in
1962 def A4_cround_ri : T_S2op_2_ii <"cround", 0b111, 0b000>;
1964 let isCodeGenOnly = 0 in
1965 def A4_round_ri : T_S2op_2_ii <"round", 0b111, 0b100>;
1967 let Defs = [USR_OVF], isCodeGenOnly = 0 in
1968 def A4_round_ri_sat : T_S2op_2_ii <"round", 0b111, 0b110, 1>;
1970 // Logical-logical words.
1971 // Compound or-and -- Rx=or(Ru,and(Rx,#s10))
1972 let isExtentSigned = 1, hasNewValue = 1, isExtendable = 1, opExtentBits = 10,
1973 opExtendable = 3, isCodeGenOnly = 0 in
1975 ALU64Inst<(outs IntRegs:$Rx),
1976 (ins IntRegs:$Ru, IntRegs:$_src_, s10Ext:$s10),
1977 "$Rx = or($Ru, and($_src_, #$s10))" ,
1978 [(set (i32 IntRegs:$Rx),
1979 (or (i32 IntRegs:$Ru), (and (i32 IntRegs:$_src_), s10ExtPred:$s10)))] ,
1980 "$_src_ = $Rx", ALU64_tc_2_SLOT23> {
1985 let IClass = 0b1101;
1987 let Inst{27-22} = 0b101001;
1988 let Inst{20-16} = Rx;
1989 let Inst{21} = s10{9};
1990 let Inst{13-5} = s10{8-0};
1994 // Miscellaneous ALU64 instructions.
1996 let hasNewValue = 1, hasSideEffects = 0, isCodeGenOnly = 0 in
1997 def A4_modwrapu: ALU64Inst<(outs IntRegs:$Rd), (ins IntRegs:$Rs, IntRegs:$Rt),
1998 "$Rd = modwrap($Rs, $Rt)", [], "", ALU64_tc_2_SLOT23> {
2003 let IClass = 0b1101;
2004 let Inst{27-21} = 0b0011111;
2005 let Inst{20-16} = Rs;
2006 let Inst{12-8} = Rt;
2007 let Inst{7-5} = 0b111;
2011 let hasSideEffects = 0, isCodeGenOnly = 0 in
2012 def A4_bitsplit: ALU64Inst<(outs DoubleRegs:$Rd),
2013 (ins IntRegs:$Rs, IntRegs:$Rt),
2014 "$Rd = bitsplit($Rs, $Rt)", [], "", ALU64_tc_1_SLOT23> {
2019 let IClass = 0b1101;
2020 let Inst{27-24} = 0b0100;
2022 let Inst{20-16} = Rs;
2023 let Inst{12-8} = Rt;
2027 let isCodeGenOnly = 0 in {
2028 // Rx[&|]=xor(Rs,Rt)
2029 def M4_or_xor : T_MType_acc_rr < "|= xor", 0b110, 0b001, 0>;
2030 def M4_and_xor : T_MType_acc_rr < "&= xor", 0b010, 0b010, 0>;
2032 // Rx[&|^]=or(Rs,Rt)
2033 def M4_xor_or : T_MType_acc_rr < "^= or", 0b110, 0b011, 0>;
2035 let CextOpcode = "ORr_ORr" in
2036 def M4_or_or : T_MType_acc_rr < "|= or", 0b110, 0b000, 0>;
2037 def M4_and_or : T_MType_acc_rr < "&= or", 0b010, 0b001, 0>;
2039 // Rx[&|^]=and(Rs,Rt)
2040 def M4_xor_and : T_MType_acc_rr < "^= and", 0b110, 0b010, 0>;
2042 let CextOpcode = "ORr_ANDr" in
2043 def M4_or_and : T_MType_acc_rr < "|= and", 0b010, 0b011, 0>;
2044 def M4_and_and : T_MType_acc_rr < "&= and", 0b010, 0b000, 0>;
2046 // Rx[&|^]=and(Rs,~Rt)
2047 def M4_xor_andn : T_MType_acc_rr < "^= and", 0b001, 0b010, 0, [], 1>;
2048 def M4_or_andn : T_MType_acc_rr < "|= and", 0b001, 0b000, 0, [], 1>;
2049 def M4_and_andn : T_MType_acc_rr < "&= and", 0b001, 0b001, 0, [], 1>;
2052 // Compound or-or and or-and
2053 let isExtentSigned = 1, InputType = "imm", hasNewValue = 1, isExtendable = 1,
2054 opExtentBits = 10, opExtendable = 3 in
2055 class T_CompOR <string mnemonic, bits<2> MajOp, SDNode OpNode>
2056 : MInst_acc <(outs IntRegs:$Rx),
2057 (ins IntRegs:$src1, IntRegs:$Rs, s10Ext:$s10),
2058 "$Rx |= "#mnemonic#"($Rs, #$s10)",
2059 [(set (i32 IntRegs:$Rx), (or (i32 IntRegs:$src1),
2060 (OpNode (i32 IntRegs:$Rs), s10ExtPred:$s10)))],
2061 "$src1 = $Rx", ALU64_tc_2_SLOT23>, ImmRegRel {
2066 let IClass = 0b1101;
2068 let Inst{27-24} = 0b1010;
2069 let Inst{23-22} = MajOp;
2070 let Inst{20-16} = Rs;
2071 let Inst{21} = s10{9};
2072 let Inst{13-5} = s10{8-0};
2076 let CextOpcode = "ORr_ANDr", isCodeGenOnly = 0 in
2077 def S4_or_andi : T_CompOR <"and", 0b00, and>;
2079 let CextOpcode = "ORr_ORr", isCodeGenOnly = 0 in
2080 def S4_or_ori : T_CompOR <"or", 0b10, or>;
2083 // Rd=modwrap(Rs,Rt)
2085 // Rd=cround(Rs,#u5)
2087 // Rd=round(Rs,#u5)[:sat]
2088 // Rd=round(Rs,Rt)[:sat]
2089 // Vector reduce add unsigned halfwords
2090 // Rd=vraddh(Rss,Rtt)
2092 // Rdd=vaddb(Rss,Rtt)
2093 // Vector conditional negate
2094 // Rdd=vcnegh(Rss,Rt)
2095 // Rxx+=vrcnegh(Rss,Rt)
2096 // Vector maximum bytes
2097 // Rdd=vmaxb(Rtt,Rss)
2098 // Vector reduce maximum halfwords
2099 // Rxx=vrmaxh(Rss,Ru)
2100 // Rxx=vrmaxuh(Rss,Ru)
2101 // Vector reduce maximum words
2102 // Rxx=vrmaxuw(Rss,Ru)
2103 // Rxx=vrmaxw(Rss,Ru)
2104 // Vector minimum bytes
2105 // Rdd=vminb(Rtt,Rss)
2106 // Vector reduce minimum halfwords
2107 // Rxx=vrminh(Rss,Ru)
2108 // Rxx=vrminuh(Rss,Ru)
2109 // Vector reduce minimum words
2110 // Rxx=vrminuw(Rss,Ru)
2111 // Rxx=vrminw(Rss,Ru)
2112 // Vector subtract bytes
2113 // Rdd=vsubb(Rss,Rtt)
2115 //===----------------------------------------------------------------------===//
2117 //===----------------------------------------------------------------------===//
2119 //===----------------------------------------------------------------------===//
2121 //===----------------------------------------------------------------------===//
2124 let isCodeGenOnly = 0 in
2125 def S2_brevp : T_S2op_3 <"brev", 0b11, 0b110>;
2128 let isCodeGenOnly = 0 in {
2129 def S2_ct0p : T_COUNT_LEADING_64<"ct0", 0b111, 0b010>;
2130 def S2_ct1p : T_COUNT_LEADING_64<"ct1", 0b111, 0b100>;
2131 def S4_clbpnorm : T_COUNT_LEADING_64<"normamt", 0b011, 0b000>;
2134 def: Pat<(i32 (trunc (cttz (i64 DoubleRegs:$Rss)))),
2135 (S2_ct0p (i64 DoubleRegs:$Rss))>;
2136 def: Pat<(i32 (trunc (cttz (not (i64 DoubleRegs:$Rss))))),
2137 (S2_ct1p (i64 DoubleRegs:$Rss))>;
2139 let hasSideEffects = 0, hasNewValue = 1, isCodeGenOnly = 0 in
2140 def S4_clbaddi : SInst<(outs IntRegs:$Rd), (ins IntRegs:$Rs, s6Imm:$s6),
2141 "$Rd = add(clb($Rs), #$s6)", [], "", S_2op_tc_2_SLOT23> {
2145 let IClass = 0b1000;
2146 let Inst{27-24} = 0b1100;
2147 let Inst{23-21} = 0b001;
2148 let Inst{20-16} = Rs;
2149 let Inst{13-8} = s6;
2150 let Inst{7-5} = 0b000;
2154 let hasSideEffects = 0, hasNewValue = 1, isCodeGenOnly = 0 in
2155 def S4_clbpaddi : SInst<(outs IntRegs:$Rd), (ins DoubleRegs:$Rs, s6Imm:$s6),
2156 "$Rd = add(clb($Rs), #$s6)", [], "", S_2op_tc_2_SLOT23> {
2160 let IClass = 0b1000;
2161 let Inst{27-24} = 0b1000;
2162 let Inst{23-21} = 0b011;
2163 let Inst{20-16} = Rs;
2164 let Inst{13-8} = s6;
2165 let Inst{7-5} = 0b010;
2170 // Bit test/set/clear
2171 let isCodeGenOnly = 0 in {
2172 def S4_ntstbit_i : T_TEST_BIT_IMM<"!tstbit", 0b001>;
2173 def S4_ntstbit_r : T_TEST_BIT_REG<"!tstbit", 1>;
2176 let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm.
2177 def: Pat<(i1 (seteq (and (shl 1, u5ImmPred:$u5), (i32 IntRegs:$Rs)), 0)),
2178 (S4_ntstbit_i (i32 IntRegs:$Rs), u5ImmPred:$u5)>;
2179 def: Pat<(i1 (seteq (and (shl 1, (i32 IntRegs:$Rt)), (i32 IntRegs:$Rs)), 0)),
2180 (S4_ntstbit_r (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))>;
2183 // Add extra complexity to prefer these instructions over bitsset/bitsclr.
2184 // The reason is that tstbit/ntstbit can be folded into a compound instruction:
2185 // if ([!]tstbit(...)) jump ...
2186 let AddedComplexity = 100 in
2187 def: Pat<(i1 (setne (and (i32 IntRegs:$Rs), (i32 Set5ImmPred:$u5)), (i32 0))),
2188 (S2_tstbit_i (i32 IntRegs:$Rs), (BITPOS32 Set5ImmPred:$u5))>;
2190 let AddedComplexity = 100 in
2191 def: Pat<(i1 (seteq (and (i32 IntRegs:$Rs), (i32 Set5ImmPred:$u5)), (i32 0))),
2192 (S4_ntstbit_i (i32 IntRegs:$Rs), (BITPOS32 Set5ImmPred:$u5))>;
2194 let isCodeGenOnly = 0 in {
2195 def C4_nbitsset : T_TEST_BITS_REG<"!bitsset", 0b01, 1>;
2196 def C4_nbitsclr : T_TEST_BITS_REG<"!bitsclr", 0b10, 1>;
2197 def C4_nbitsclri : T_TEST_BITS_IMM<"!bitsclr", 0b10, 1>;
2200 // Do not increase complexity of these patterns. In the DAG, "cmp i8" may be
2201 // represented as a compare against "value & 0xFF", which is an exact match
2202 // for cmpb (same for cmph). The patterns below do not contain any additional
2203 // complexity that would make them preferable, and if they were actually used
2204 // instead of cmpb/cmph, they would result in a compare against register that
2205 // is loaded with the byte/half mask (i.e. 0xFF or 0xFFFF).
2206 def: Pat<(i1 (setne (and I32:$Rs, u6ImmPred:$u6), 0)),
2207 (C4_nbitsclri I32:$Rs, u6ImmPred:$u6)>;
2208 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), 0)),
2209 (C4_nbitsclr I32:$Rs, I32:$Rt)>;
2210 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), I32:$Rt)),
2211 (C4_nbitsset I32:$Rs, I32:$Rt)>;
2213 //===----------------------------------------------------------------------===//
2215 //===----------------------------------------------------------------------===//
2217 //===----------------------------------------------------------------------===//
2219 //===----------------------------------------------------------------------===//
2221 // Rd=add(#u6,mpyi(Rs,#U6)) -- Multiply by immed and add immed.
2223 let hasNewValue = 1, isExtendable = 1, opExtentBits = 6, opExtendable = 1,
2224 isCodeGenOnly = 0 in
2225 def M4_mpyri_addi : MInst<(outs IntRegs:$Rd),
2226 (ins u6Ext:$u6, IntRegs:$Rs, u6Imm:$U6),
2227 "$Rd = add(#$u6, mpyi($Rs, #$U6))" ,
2228 [(set (i32 IntRegs:$Rd),
2229 (add (mul (i32 IntRegs:$Rs), u6ImmPred:$U6),
2230 u6ExtPred:$u6))] ,"",ALU64_tc_3x_SLOT23> {
2236 let IClass = 0b1101;
2238 let Inst{27-24} = 0b1000;
2239 let Inst{23} = U6{5};
2240 let Inst{22-21} = u6{5-4};
2241 let Inst{20-16} = Rs;
2242 let Inst{13} = u6{3};
2243 let Inst{12-8} = Rd;
2244 let Inst{7-5} = u6{2-0};
2245 let Inst{4-0} = U6{4-0};
2248 // Rd=add(#u6,mpyi(Rs,Rt))
2249 let CextOpcode = "ADD_MPY", InputType = "imm", hasNewValue = 1,
2250 isExtendable = 1, opExtentBits = 6, opExtendable = 1, isCodeGenOnly = 0 in
2251 def M4_mpyrr_addi : MInst <(outs IntRegs:$Rd),
2252 (ins u6Ext:$u6, IntRegs:$Rs, IntRegs:$Rt),
2253 "$Rd = add(#$u6, mpyi($Rs, $Rt))" ,
2254 [(set (i32 IntRegs:$Rd),
2255 (add (mul (i32 IntRegs:$Rs), (i32 IntRegs:$Rt)), u6ExtPred:$u6))],
2256 "", ALU64_tc_3x_SLOT23>, ImmRegRel {
2262 let IClass = 0b1101;
2264 let Inst{27-23} = 0b01110;
2265 let Inst{22-21} = u6{5-4};
2266 let Inst{20-16} = Rs;
2267 let Inst{13} = u6{3};
2268 let Inst{12-8} = Rt;
2269 let Inst{7-5} = u6{2-0};
2273 let hasNewValue = 1 in
2274 class T_AddMpy <bit MajOp, PatLeaf ImmPred, dag ins>
2275 : ALU64Inst <(outs IntRegs:$dst), ins,
2276 "$dst = add($src1, mpyi("#!if(MajOp,"$src3, #$src2))",
2278 [(set (i32 IntRegs:$dst),
2279 (add (i32 IntRegs:$src1), (mul (i32 IntRegs:$src3), ImmPred:$src2)))],
2280 "", ALU64_tc_3x_SLOT23> {
2286 let IClass = 0b1101;
2288 bits<6> ImmValue = !if(MajOp, src2{5-0}, src2{7-2});
2290 let Inst{27-24} = 0b1111;
2291 let Inst{23} = MajOp;
2292 let Inst{22-21} = ImmValue{5-4};
2293 let Inst{20-16} = src3;
2294 let Inst{13} = ImmValue{3};
2295 let Inst{12-8} = dst;
2296 let Inst{7-5} = ImmValue{2-0};
2297 let Inst{4-0} = src1;
2300 let isCodeGenOnly = 0 in
2301 def M4_mpyri_addr_u2 : T_AddMpy<0b0, u6_2ImmPred,
2302 (ins IntRegs:$src1, u6_2Imm:$src2, IntRegs:$src3)>;
2304 let isExtendable = 1, opExtentBits = 6, opExtendable = 3,
2305 CextOpcode = "ADD_MPY", InputType = "imm", isCodeGenOnly = 0 in
2306 def M4_mpyri_addr : T_AddMpy<0b1, u6ExtPred,
2307 (ins IntRegs:$src1, IntRegs:$src3, u6Ext:$src2)>, ImmRegRel;
2309 // Rx=add(Ru,mpyi(Rx,Rs))
2310 let validSubTargets = HasV4SubT, CextOpcode = "ADD_MPY", InputType = "reg",
2311 hasNewValue = 1, isCodeGenOnly = 0 in
2312 def M4_mpyrr_addr: MInst_acc <(outs IntRegs:$Rx),
2313 (ins IntRegs:$Ru, IntRegs:$_src_, IntRegs:$Rs),
2314 "$Rx = add($Ru, mpyi($_src_, $Rs))",
2315 [(set (i32 IntRegs:$Rx), (add (i32 IntRegs:$Ru),
2316 (mul (i32 IntRegs:$_src_), (i32 IntRegs:$Rs))))],
2317 "$_src_ = $Rx", M_tc_3x_SLOT23>, ImmRegRel {
2322 let IClass = 0b1110;
2324 let Inst{27-21} = 0b0011000;
2325 let Inst{12-8} = Rx;
2327 let Inst{20-16} = Rs;
2330 // Rd=add(##,mpyi(Rs,#U6))
2331 def : Pat <(add (mul (i32 IntRegs:$src2), u6ImmPred:$src3),
2332 (HexagonCONST32 tglobaladdr:$src1)),
2333 (i32 (M4_mpyri_addi tglobaladdr:$src1, IntRegs:$src2,
2336 // Rd=add(##,mpyi(Rs,Rt))
2337 def : Pat <(add (mul (i32 IntRegs:$src2), (i32 IntRegs:$src3)),
2338 (HexagonCONST32 tglobaladdr:$src1)),
2339 (i32 (M4_mpyrr_addi tglobaladdr:$src1, IntRegs:$src2,
2342 // Vector reduce multiply word by signed half (32x16)
2343 //Rdd=vrmpyweh(Rss,Rtt)[:<<1]
2344 let isCodeGenOnly = 0 in {
2345 def M4_vrmpyeh_s0 : T_M2_vmpy<"vrmpyweh", 0b010, 0b100, 0, 0, 0>;
2346 def M4_vrmpyeh_s1 : T_M2_vmpy<"vrmpyweh", 0b110, 0b100, 1, 0, 0>;
2349 //Rdd=vrmpywoh(Rss,Rtt)[:<<1]
2350 let isCodeGenOnly = 0 in {
2351 def M4_vrmpyoh_s0 : T_M2_vmpy<"vrmpywoh", 0b001, 0b010, 0, 0, 0>;
2352 def M4_vrmpyoh_s1 : T_M2_vmpy<"vrmpywoh", 0b101, 0b010, 1, 0, 0>;
2354 //Rdd+=vrmpyweh(Rss,Rtt)[:<<1]
2355 let isCodeGenOnly = 0 in {
2356 def M4_vrmpyeh_acc_s0: T_M2_vmpy_acc<"vrmpyweh", 0b001, 0b110, 0, 0>;
2357 def M4_vrmpyeh_acc_s1: T_M2_vmpy_acc<"vrmpyweh", 0b101, 0b110, 1, 0>;
2360 //Rdd=vrmpywoh(Rss,Rtt)[:<<1]
2361 let isCodeGenOnly = 0 in {
2362 def M4_vrmpyoh_acc_s0: T_M2_vmpy_acc<"vrmpywoh", 0b011, 0b110, 0, 0>;
2363 def M4_vrmpyoh_acc_s1: T_M2_vmpy_acc<"vrmpywoh", 0b111, 0b110, 1, 0>;
2366 // Vector multiply halfwords, signed by unsigned
2367 // Rdd=vmpyhsu(Rs,Rt)[:<<]:sat
2368 let isCodeGenOnly = 0 in {
2369 def M2_vmpy2su_s0 : T_XTYPE_mpy64 < "vmpyhsu", 0b000, 0b111, 1, 0, 0>;
2370 def M2_vmpy2su_s1 : T_XTYPE_mpy64 < "vmpyhsu", 0b100, 0b111, 1, 1, 0>;
2373 // Rxx+=vmpyhsu(Rs,Rt)[:<<1]:sat
2374 let isCodeGenOnly = 0 in {
2375 def M2_vmac2su_s0 : T_XTYPE_mpy64_acc < "vmpyhsu", "+", 0b011, 0b101, 1, 0, 0>;
2376 def M2_vmac2su_s1 : T_XTYPE_mpy64_acc < "vmpyhsu", "+", 0b111, 0b101, 1, 1, 0>;
2379 // Vector polynomial multiply halfwords
2380 // Rdd=vpmpyh(Rs,Rt)
2381 let isCodeGenOnly = 0 in
2382 def M4_vpmpyh : T_XTYPE_mpy64 < "vpmpyh", 0b110, 0b111, 0, 0, 0>;
2384 // Rxx^=vpmpyh(Rs,Rt)
2385 let isCodeGenOnly = 0 in
2386 def M4_vpmpyh_acc : T_XTYPE_mpy64_acc < "vpmpyh", "^", 0b101, 0b111, 0, 0, 0>;
2388 // Polynomial multiply words
2390 let isCodeGenOnly = 0 in
2391 def M4_pmpyw : T_XTYPE_mpy64 < "pmpyw", 0b010, 0b111, 0, 0, 0>;
2393 // Rxx^=pmpyw(Rs,Rt)
2394 let isCodeGenOnly = 0 in
2395 def M4_pmpyw_acc : T_XTYPE_mpy64_acc < "pmpyw", "^", 0b001, 0b111, 0, 0, 0>;
2397 //===----------------------------------------------------------------------===//
2399 //===----------------------------------------------------------------------===//
2402 //===----------------------------------------------------------------------===//
2403 // ALU64/Vector compare
2404 //===----------------------------------------------------------------------===//
2405 //===----------------------------------------------------------------------===//
2406 // Template class for vector compare
2407 //===----------------------------------------------------------------------===//
2409 let hasSideEffects = 0 in
2410 class T_vcmpImm <string Str, bits<2> cmpOp, bits<2> minOp, Operand ImmOprnd>
2411 : ALU64_rr <(outs PredRegs:$Pd),
2412 (ins DoubleRegs:$Rss, ImmOprnd:$Imm),
2413 "$Pd = "#Str#"($Rss, #$Imm)",
2414 [], "", ALU64_tc_2early_SLOT23> {
2419 let ImmBits{6-0} = Imm{6-0};
2420 let ImmBits{7} = !if (!eq(cmpOp,0b10), 0b0, Imm{7}); // 0 for vcmp[bhw].gtu
2422 let IClass = 0b1101;
2424 let Inst{27-24} = 0b1100;
2425 let Inst{22-21} = cmpOp;
2426 let Inst{20-16} = Rss;
2427 let Inst{12-5} = ImmBits;
2428 let Inst{4-3} = minOp;
2432 // Vector compare bytes
2433 let isCodeGenOnly = 0 in
2434 def A4_vcmpbgt : T_vcmp <"vcmpb.gt", 0b1010>;
2435 def: T_vcmp_pat<A4_vcmpbgt, setgt, v8i8>;
2437 let AsmString = "$Pd = any8(vcmpb.eq($Rss, $Rtt))" in
2438 let isCodeGenOnly = 0 in
2439 def A4_vcmpbeq_any : T_vcmp <"any8(vcmpb.gt", 0b1000>;
2441 let isCodeGenOnly = 0 in {
2442 def A4_vcmpbeqi : T_vcmpImm <"vcmpb.eq", 0b00, 0b00, u8Imm>;
2443 def A4_vcmpbgti : T_vcmpImm <"vcmpb.gt", 0b01, 0b00, s8Imm>;
2444 def A4_vcmpbgtui : T_vcmpImm <"vcmpb.gtu", 0b10, 0b00, u7Imm>;
2447 // Vector compare halfwords
2448 let isCodeGenOnly = 0 in {
2449 def A4_vcmpheqi : T_vcmpImm <"vcmph.eq", 0b00, 0b01, s8Imm>;
2450 def A4_vcmphgti : T_vcmpImm <"vcmph.gt", 0b01, 0b01, s8Imm>;
2451 def A4_vcmphgtui : T_vcmpImm <"vcmph.gtu", 0b10, 0b01, u7Imm>;
2454 // Vector compare words
2455 let isCodeGenOnly = 0 in {
2456 def A4_vcmpweqi : T_vcmpImm <"vcmpw.eq", 0b00, 0b10, s8Imm>;
2457 def A4_vcmpwgti : T_vcmpImm <"vcmpw.gt", 0b01, 0b10, s8Imm>;
2458 def A4_vcmpwgtui : T_vcmpImm <"vcmpw.gtu", 0b10, 0b10, u7Imm>;
2461 //===----------------------------------------------------------------------===//
2463 //===----------------------------------------------------------------------===//
2464 // Shift by immediate and accumulate/logical.
2465 // Rx=add(#u8,asl(Rx,#U5)) Rx=add(#u8,lsr(Rx,#U5))
2466 // Rx=sub(#u8,asl(Rx,#U5)) Rx=sub(#u8,lsr(Rx,#U5))
2467 // Rx=and(#u8,asl(Rx,#U5)) Rx=and(#u8,lsr(Rx,#U5))
2468 // Rx=or(#u8,asl(Rx,#U5)) Rx=or(#u8,lsr(Rx,#U5))
2469 let isExtendable = 1, opExtendable = 1, isExtentSigned = 0, opExtentBits = 8,
2470 hasNewValue = 1, opNewValue = 0, validSubTargets = HasV4SubT in
2471 class T_S4_ShiftOperate<string MnOp, string MnSh, SDNode Op, SDNode Sh,
2472 bit asl_lsr, bits<2> MajOp, InstrItinClass Itin>
2473 : MInst_acc<(outs IntRegs:$Rd), (ins u8Ext:$u8, IntRegs:$Rx, u5Imm:$U5),
2474 "$Rd = "#MnOp#"(#$u8, "#MnSh#"($Rx, #$U5))",
2475 [(set (i32 IntRegs:$Rd),
2476 (Op (Sh I32:$Rx, u5ImmPred:$U5), u8ExtPred:$u8))],
2477 "$Rd = $Rx", Itin> {
2484 let IClass = 0b1101;
2485 let Inst{27-24} = 0b1110;
2486 let Inst{23-21} = u8{7-5};
2487 let Inst{20-16} = Rd;
2488 let Inst{13} = u8{4};
2489 let Inst{12-8} = U5;
2490 let Inst{7-5} = u8{3-1};
2491 let Inst{4} = asl_lsr;
2492 let Inst{3} = u8{0};
2493 let Inst{2-1} = MajOp;
2496 multiclass T_ShiftOperate<string mnemonic, SDNode Op, bits<2> MajOp,
2497 InstrItinClass Itin> {
2498 def _asl_ri : T_S4_ShiftOperate<mnemonic, "asl", Op, shl, 0, MajOp, Itin>;
2499 def _lsr_ri : T_S4_ShiftOperate<mnemonic, "lsr", Op, srl, 1, MajOp, Itin>;
2502 let AddedComplexity = 200, isCodeGenOnly = 0 in {
2503 defm S4_addi : T_ShiftOperate<"add", add, 0b10, ALU64_tc_2_SLOT23>;
2504 defm S4_andi : T_ShiftOperate<"and", and, 0b00, ALU64_tc_2_SLOT23>;
2507 let AddedComplexity = 30, isCodeGenOnly = 0 in
2508 defm S4_ori : T_ShiftOperate<"or", or, 0b01, ALU64_tc_1_SLOT23>;
2510 let isCodeGenOnly = 0 in
2511 defm S4_subi : T_ShiftOperate<"sub", sub, 0b11, ALU64_tc_1_SLOT23>;
2513 // Vector conditional negate
2514 // Rdd=vcnegh(Rss,Rt)
2515 let Defs = [USR_OVF], Itinerary = S_3op_tc_2_SLOT23, isCodeGenOnly = 0 in
2516 def S2_vcnegh : T_S3op_shiftVect < "vcnegh", 0b11, 0b01>;
2518 // Rd=[cround|round](Rs,Rt)
2519 let hasNewValue = 1, Itinerary = S_3op_tc_2_SLOT23, isCodeGenOnly = 0 in {
2520 def A4_cround_rr : T_S3op_3 < "cround", IntRegs, 0b11, 0b00>;
2521 def A4_round_rr : T_S3op_3 < "round", IntRegs, 0b11, 0b10>;
2524 // Rd=round(Rs,Rt):sat
2525 let hasNewValue = 1, Defs = [USR_OVF], Itinerary = S_3op_tc_2_SLOT23,
2526 isCodeGenOnly = 0 in
2527 def A4_round_rr_sat : T_S3op_3 < "round", IntRegs, 0b11, 0b11, 1>;
2529 // Rd=[cmpyiwh|cmpyrwh](Rss,Rt):<<1:rnd:sat
2530 let Defs = [USR_OVF], Itinerary = S_3op_tc_3x_SLOT23, isCodeGenOnly = 0 in {
2531 def M4_cmpyi_wh : T_S3op_8<"cmpyiwh", 0b100, 1, 1, 1>;
2532 def M4_cmpyr_wh : T_S3op_8<"cmpyrwh", 0b110, 1, 1, 1>;
2535 // Rdd=[add|sub](Rss,Rtt,Px):carry
2536 let isPredicateLate = 1, hasSideEffects = 0 in
2537 class T_S3op_carry <string mnemonic, bits<3> MajOp>
2538 : SInst < (outs DoubleRegs:$Rdd, PredRegs:$Px),
2539 (ins DoubleRegs:$Rss, DoubleRegs:$Rtt, PredRegs:$Pu),
2540 "$Rdd = "#mnemonic#"($Rss, $Rtt, $Pu):carry",
2541 [], "$Px = $Pu", S_3op_tc_1_SLOT23 > {
2547 let IClass = 0b1100;
2549 let Inst{27-24} = 0b0010;
2550 let Inst{23-21} = MajOp;
2551 let Inst{20-16} = Rss;
2552 let Inst{12-8} = Rtt;
2554 let Inst{4-0} = Rdd;
2557 let isCodeGenOnly = 0 in {
2558 def A4_addp_c : T_S3op_carry < "add", 0b110 >;
2559 def A4_subp_c : T_S3op_carry < "sub", 0b111 >;
2562 let Itinerary = S_3op_tc_3_SLOT23, hasSideEffects = 0 in
2563 class T_S3op_6 <string mnemonic, bits<3> MinOp, bit isUnsigned>
2564 : SInst <(outs DoubleRegs:$Rxx),
2565 (ins DoubleRegs:$dst2, DoubleRegs:$Rss, IntRegs:$Ru),
2566 "$Rxx = "#mnemonic#"($Rss, $Ru)" ,
2567 [] , "$dst2 = $Rxx"> {
2572 let IClass = 0b1100;
2574 let Inst{27-21} = 0b1011001;
2575 let Inst{20-16} = Rss;
2576 let Inst{13} = isUnsigned;
2577 let Inst{12-8} = Rxx;
2578 let Inst{7-5} = MinOp;
2582 // Vector reduce maximum halfwords
2583 // Rxx=vrmax[u]h(Rss,Ru)
2584 let isCodeGenOnly = 0 in {
2585 def A4_vrmaxh : T_S3op_6 < "vrmaxh", 0b001, 0>;
2586 def A4_vrmaxuh : T_S3op_6 < "vrmaxuh", 0b001, 1>;
2588 // Vector reduce maximum words
2589 // Rxx=vrmax[u]w(Rss,Ru)
2590 let isCodeGenOnly = 0 in {
2591 def A4_vrmaxw : T_S3op_6 < "vrmaxw", 0b010, 0>;
2592 def A4_vrmaxuw : T_S3op_6 < "vrmaxuw", 0b010, 1>;
2594 // Vector reduce minimum halfwords
2595 // Rxx=vrmin[u]h(Rss,Ru)
2596 let isCodeGenOnly = 0 in {
2597 def A4_vrminh : T_S3op_6 < "vrminh", 0b101, 0>;
2598 def A4_vrminuh : T_S3op_6 < "vrminuh", 0b101, 1>;
2601 // Vector reduce minimum words
2602 // Rxx=vrmin[u]w(Rss,Ru)
2603 let isCodeGenOnly = 0 in {
2604 def A4_vrminw : T_S3op_6 < "vrminw", 0b110, 0>;
2605 def A4_vrminuw : T_S3op_6 < "vrminuw", 0b110, 1>;
2608 // Shift an immediate left by register amount.
2609 let hasNewValue = 1, hasSideEffects = 0, isCodeGenOnly = 0 in
2610 def S4_lsli: SInst <(outs IntRegs:$Rd), (ins s6Imm:$s6, IntRegs:$Rt),
2611 "$Rd = lsl(#$s6, $Rt)" ,
2612 [(set (i32 IntRegs:$Rd), (shl s6ImmPred:$s6,
2613 (i32 IntRegs:$Rt)))],
2614 "", S_3op_tc_1_SLOT23> {
2619 let IClass = 0b1100;
2621 let Inst{27-22} = 0b011010;
2622 let Inst{20-16} = s6{5-1};
2623 let Inst{12-8} = Rt;
2624 let Inst{7-6} = 0b11;
2626 let Inst{5} = s6{0};
2629 //===----------------------------------------------------------------------===//
2631 //===----------------------------------------------------------------------===//
2633 //===----------------------------------------------------------------------===//
2634 // MEMOP: Word, Half, Byte
2635 //===----------------------------------------------------------------------===//
2637 def MEMOPIMM : SDNodeXForm<imm, [{
2638 // Call the transformation function XformM5ToU5Imm to get the negative
2639 // immediate's positive counterpart.
2640 int32_t imm = N->getSExtValue();
2641 return XformM5ToU5Imm(imm);
2644 def MEMOPIMM_HALF : SDNodeXForm<imm, [{
2645 // -1 .. -31 represented as 65535..65515
2646 // assigning to a short restores our desired signed value.
2647 // Call the transformation function XformM5ToU5Imm to get the negative
2648 // immediate's positive counterpart.
2649 int16_t imm = N->getSExtValue();
2650 return XformM5ToU5Imm(imm);
2653 def MEMOPIMM_BYTE : SDNodeXForm<imm, [{
2654 // -1 .. -31 represented as 255..235
2655 // assigning to a char restores our desired signed value.
2656 // Call the transformation function XformM5ToU5Imm to get the negative
2657 // immediate's positive counterpart.
2658 int8_t imm = N->getSExtValue();
2659 return XformM5ToU5Imm(imm);
2662 def SETMEMIMM : SDNodeXForm<imm, [{
2663 // Return the bit position we will set [0-31].
2665 int32_t imm = N->getSExtValue();
2666 return XformMskToBitPosU5Imm(imm);
2669 def CLRMEMIMM : SDNodeXForm<imm, [{
2670 // Return the bit position we will clear [0-31].
2672 // we bit negate the value first
2673 int32_t imm = ~(N->getSExtValue());
2674 return XformMskToBitPosU5Imm(imm);
2677 def SETMEMIMM_SHORT : SDNodeXForm<imm, [{
2678 // Return the bit position we will set [0-15].
2680 int16_t imm = N->getSExtValue();
2681 return XformMskToBitPosU4Imm(imm);
2684 def CLRMEMIMM_SHORT : SDNodeXForm<imm, [{
2685 // Return the bit position we will clear [0-15].
2687 // we bit negate the value first
2688 int16_t imm = ~(N->getSExtValue());
2689 return XformMskToBitPosU4Imm(imm);
2692 def SETMEMIMM_BYTE : SDNodeXForm<imm, [{
2693 // Return the bit position we will set [0-7].
2695 int8_t imm = N->getSExtValue();
2696 return XformMskToBitPosU3Imm(imm);
2699 def CLRMEMIMM_BYTE : SDNodeXForm<imm, [{
2700 // Return the bit position we will clear [0-7].
2702 // we bit negate the value first
2703 int8_t imm = ~(N->getSExtValue());
2704 return XformMskToBitPosU3Imm(imm);
2707 //===----------------------------------------------------------------------===//
2708 // Template class for MemOp instructions with the register value.
2709 //===----------------------------------------------------------------------===//
2710 class MemOp_rr_base <string opc, bits<2> opcBits, Operand ImmOp,
2711 string memOp, bits<2> memOpBits> :
2713 (ins IntRegs:$base, ImmOp:$offset, IntRegs:$delta),
2714 opc#"($base+#$offset)"#memOp#"$delta",
2716 Requires<[UseMEMOP]> {
2721 bits<6> offsetBits; // memb - u6:0 , memh - u6:1, memw - u6:2
2723 let offsetBits = !if (!eq(opcBits, 0b00), offset{5-0},
2724 !if (!eq(opcBits, 0b01), offset{6-1},
2725 !if (!eq(opcBits, 0b10), offset{7-2},0)));
2727 let opExtentAlign = opcBits;
2728 let IClass = 0b0011;
2729 let Inst{27-24} = 0b1110;
2730 let Inst{22-21} = opcBits;
2731 let Inst{20-16} = base;
2733 let Inst{12-7} = offsetBits;
2734 let Inst{6-5} = memOpBits;
2735 let Inst{4-0} = delta;
2738 //===----------------------------------------------------------------------===//
2739 // Template class for MemOp instructions with the immediate value.
2740 //===----------------------------------------------------------------------===//
2741 class MemOp_ri_base <string opc, bits<2> opcBits, Operand ImmOp,
2742 string memOp, bits<2> memOpBits> :
2744 (ins IntRegs:$base, ImmOp:$offset, u5Imm:$delta),
2745 opc#"($base+#$offset)"#memOp#"#$delta"
2746 #!if(memOpBits{1},")", ""), // clrbit, setbit - include ')'
2748 Requires<[UseMEMOP]> {
2753 bits<6> offsetBits; // memb - u6:0 , memh - u6:1, memw - u6:2
2755 let offsetBits = !if (!eq(opcBits, 0b00), offset{5-0},
2756 !if (!eq(opcBits, 0b01), offset{6-1},
2757 !if (!eq(opcBits, 0b10), offset{7-2},0)));
2759 let opExtentAlign = opcBits;
2760 let IClass = 0b0011;
2761 let Inst{27-24} = 0b1111;
2762 let Inst{22-21} = opcBits;
2763 let Inst{20-16} = base;
2765 let Inst{12-7} = offsetBits;
2766 let Inst{6-5} = memOpBits;
2767 let Inst{4-0} = delta;
2770 // multiclass to define MemOp instructions with register operand.
2771 multiclass MemOp_rr<string opc, bits<2> opcBits, Operand ImmOp> {
2772 def L4_add#NAME : MemOp_rr_base <opc, opcBits, ImmOp, " += ", 0b00>; // add
2773 def L4_sub#NAME : MemOp_rr_base <opc, opcBits, ImmOp, " -= ", 0b01>; // sub
2774 def L4_and#NAME : MemOp_rr_base <opc, opcBits, ImmOp, " &= ", 0b10>; // and
2775 def L4_or#NAME : MemOp_rr_base <opc, opcBits, ImmOp, " |= ", 0b11>; // or
2778 // multiclass to define MemOp instructions with immediate Operand.
2779 multiclass MemOp_ri<string opc, bits<2> opcBits, Operand ImmOp> {
2780 def L4_iadd#NAME : MemOp_ri_base <opc, opcBits, ImmOp, " += ", 0b00 >;
2781 def L4_isub#NAME : MemOp_ri_base <opc, opcBits, ImmOp, " -= ", 0b01 >;
2782 def L4_iand#NAME : MemOp_ri_base<opc, opcBits, ImmOp, " = clrbit(", 0b10>;
2783 def L4_ior#NAME : MemOp_ri_base<opc, opcBits, ImmOp, " = setbit(", 0b11>;
2786 multiclass MemOp_base <string opc, bits<2> opcBits, Operand ImmOp> {
2787 defm _#NAME : MemOp_rr <opc, opcBits, ImmOp>;
2788 defm _#NAME : MemOp_ri <opc, opcBits, ImmOp>;
2791 // Define MemOp instructions.
2792 let isExtendable = 1, opExtendable = 1, isExtentSigned = 0,
2793 validSubTargets =HasV4SubT in {
2794 let opExtentBits = 6, accessSize = ByteAccess, isCodeGenOnly = 0 in
2795 defm memopb_io : MemOp_base <"memb", 0b00, u6_0Ext>;
2797 let opExtentBits = 7, accessSize = HalfWordAccess, isCodeGenOnly = 0 in
2798 defm memoph_io : MemOp_base <"memh", 0b01, u6_1Ext>;
2800 let opExtentBits = 8, accessSize = WordAccess, isCodeGenOnly = 0 in
2801 defm memopw_io : MemOp_base <"memw", 0b10, u6_2Ext>;
2804 //===----------------------------------------------------------------------===//
2805 // Multiclass to define 'Def Pats' for ALU operations on the memory
2806 // Here value used for the ALU operation is an immediate value.
2807 // mem[bh](Rs+#0) += #U5
2808 // mem[bh](Rs+#u6) += #U5
2809 //===----------------------------------------------------------------------===//
2811 multiclass MemOpi_u5Pats <PatFrag ldOp, PatFrag stOp, PatLeaf ExtPred,
2812 InstHexagon MI, SDNode OpNode> {
2813 let AddedComplexity = 180 in
2814 def : Pat < (stOp (OpNode (ldOp IntRegs:$addr), u5ImmPred:$addend),
2816 (MI IntRegs:$addr, #0, u5ImmPred:$addend )>;
2818 let AddedComplexity = 190 in
2819 def : Pat <(stOp (OpNode (ldOp (add IntRegs:$base, ExtPred:$offset)),
2821 (add IntRegs:$base, ExtPred:$offset)),
2822 (MI IntRegs:$base, ExtPred:$offset, u5ImmPred:$addend)>;
2825 multiclass MemOpi_u5ALUOp<PatFrag ldOp, PatFrag stOp, PatLeaf ExtPred,
2826 InstHexagon addMI, InstHexagon subMI> {
2827 defm : MemOpi_u5Pats<ldOp, stOp, ExtPred, addMI, add>;
2828 defm : MemOpi_u5Pats<ldOp, stOp, ExtPred, subMI, sub>;
2831 multiclass MemOpi_u5ExtType<PatFrag ldOpByte, PatFrag ldOpHalf > {
2833 defm : MemOpi_u5ALUOp <ldOpHalf, truncstorei16, u6_1ExtPred,
2834 L4_iadd_memoph_io, L4_isub_memoph_io>;
2836 defm : MemOpi_u5ALUOp <ldOpByte, truncstorei8, u6ExtPred,
2837 L4_iadd_memopb_io, L4_isub_memopb_io>;
2840 let Predicates = [HasV4T, UseMEMOP] in {
2841 defm : MemOpi_u5ExtType<zextloadi8, zextloadi16>; // zero extend
2842 defm : MemOpi_u5ExtType<sextloadi8, sextloadi16>; // sign extend
2843 defm : MemOpi_u5ExtType<extloadi8, extloadi16>; // any extend
2846 defm : MemOpi_u5ALUOp <load, store, u6_2ExtPred, L4_iadd_memopw_io,
2850 //===----------------------------------------------------------------------===//
2851 // multiclass to define 'Def Pats' for ALU operations on the memory.
2852 // Here value used for the ALU operation is a negative value.
2853 // mem[bh](Rs+#0) += #m5
2854 // mem[bh](Rs+#u6) += #m5
2855 //===----------------------------------------------------------------------===//
2857 multiclass MemOpi_m5Pats <PatFrag ldOp, PatFrag stOp, PatLeaf extPred,
2858 PatLeaf immPred, ComplexPattern addrPred,
2859 SDNodeXForm xformFunc, InstHexagon MI> {
2860 let AddedComplexity = 190 in
2861 def : Pat <(stOp (add (ldOp IntRegs:$addr), immPred:$subend),
2863 (MI IntRegs:$addr, #0, (xformFunc immPred:$subend) )>;
2865 let AddedComplexity = 195 in
2866 def : Pat<(stOp (add (ldOp (add IntRegs:$base, extPred:$offset)),
2868 (add IntRegs:$base, extPred:$offset)),
2869 (MI IntRegs:$base, extPred:$offset, (xformFunc immPred:$subend))>;
2872 multiclass MemOpi_m5ExtType<PatFrag ldOpByte, PatFrag ldOpHalf > {
2874 defm : MemOpi_m5Pats <ldOpHalf, truncstorei16, u6_1ExtPred, m5HImmPred,
2875 ADDRriU6_1, MEMOPIMM_HALF, L4_isub_memoph_io>;
2877 defm : MemOpi_m5Pats <ldOpByte, truncstorei8, u6ExtPred, m5BImmPred,
2878 ADDRriU6_0, MEMOPIMM_BYTE, L4_isub_memopb_io>;
2881 let Predicates = [HasV4T, UseMEMOP] in {
2882 defm : MemOpi_m5ExtType<zextloadi8, zextloadi16>; // zero extend
2883 defm : MemOpi_m5ExtType<sextloadi8, sextloadi16>; // sign extend
2884 defm : MemOpi_m5ExtType<extloadi8, extloadi16>; // any extend
2887 defm : MemOpi_m5Pats <load, store, u6_2ExtPred, m5ImmPred,
2888 ADDRriU6_2, MEMOPIMM, L4_isub_memopw_io>;
2891 //===----------------------------------------------------------------------===//
2892 // Multiclass to define 'def Pats' for bit operations on the memory.
2893 // mem[bhw](Rs+#0) = [clrbit|setbit](#U5)
2894 // mem[bhw](Rs+#u6) = [clrbit|setbit](#U5)
2895 //===----------------------------------------------------------------------===//
2897 multiclass MemOpi_bitPats <PatFrag ldOp, PatFrag stOp, PatLeaf immPred,
2898 PatLeaf extPred, ComplexPattern addrPred,
2899 SDNodeXForm xformFunc, InstHexagon MI, SDNode OpNode> {
2901 // mem[bhw](Rs+#u6:[012]) = [clrbit|setbit](#U5)
2902 let AddedComplexity = 250 in
2903 def : Pat<(stOp (OpNode (ldOp (add IntRegs:$base, extPred:$offset)),
2905 (add IntRegs:$base, extPred:$offset)),
2906 (MI IntRegs:$base, extPred:$offset, (xformFunc immPred:$bitend))>;
2908 // mem[bhw](Rs+#0) = [clrbit|setbit](#U5)
2909 let AddedComplexity = 225 in
2910 def : Pat <(stOp (OpNode (ldOp (addrPred IntRegs:$addr, extPred:$offset)),
2912 (addrPred (i32 IntRegs:$addr), extPred:$offset)),
2913 (MI IntRegs:$addr, extPred:$offset, (xformFunc immPred:$bitend))>;
2916 multiclass MemOpi_bitExtType<PatFrag ldOpByte, PatFrag ldOpHalf > {
2918 defm : MemOpi_bitPats<ldOpByte, truncstorei8, Clr3ImmPred, u6ExtPred,
2919 ADDRriU6_0, CLRMEMIMM_BYTE, L4_iand_memopb_io, and>;
2921 defm : MemOpi_bitPats<ldOpByte, truncstorei8, Set3ImmPred, u6ExtPred,
2922 ADDRriU6_0, SETMEMIMM_BYTE, L4_ior_memopb_io, or>;
2923 // Half Word - clrbit
2924 defm : MemOpi_bitPats<ldOpHalf, truncstorei16, Clr4ImmPred, u6_1ExtPred,
2925 ADDRriU6_1, CLRMEMIMM_SHORT, L4_iand_memoph_io, and>;
2926 // Half Word - setbit
2927 defm : MemOpi_bitPats<ldOpHalf, truncstorei16, Set4ImmPred, u6_1ExtPred,
2928 ADDRriU6_1, SETMEMIMM_SHORT, L4_ior_memoph_io, or>;
2931 let Predicates = [HasV4T, UseMEMOP] in {
2932 // mem[bh](Rs+#0) = [clrbit|setbit](#U5)
2933 // mem[bh](Rs+#u6:[01]) = [clrbit|setbit](#U5)
2934 defm : MemOpi_bitExtType<zextloadi8, zextloadi16>; // zero extend
2935 defm : MemOpi_bitExtType<sextloadi8, sextloadi16>; // sign extend
2936 defm : MemOpi_bitExtType<extloadi8, extloadi16>; // any extend
2938 // memw(Rs+#0) = [clrbit|setbit](#U5)
2939 // memw(Rs+#u6:2) = [clrbit|setbit](#U5)
2940 defm : MemOpi_bitPats<load, store, Clr5ImmPred, u6_2ExtPred, ADDRriU6_2,
2941 CLRMEMIMM, L4_iand_memopw_io, and>;
2942 defm : MemOpi_bitPats<load, store, Set5ImmPred, u6_2ExtPred, ADDRriU6_2,
2943 SETMEMIMM, L4_ior_memopw_io, or>;
2946 //===----------------------------------------------------------------------===//
2947 // Multiclass to define 'def Pats' for ALU operations on the memory
2948 // where addend is a register.
2949 // mem[bhw](Rs+#0) [+-&|]= Rt
2950 // mem[bhw](Rs+#U6:[012]) [+-&|]= Rt
2951 //===----------------------------------------------------------------------===//
2953 multiclass MemOpr_Pats <PatFrag ldOp, PatFrag stOp, ComplexPattern addrPred,
2954 PatLeaf extPred, InstHexagon MI, SDNode OpNode> {
2955 let AddedComplexity = 141 in
2956 // mem[bhw](Rs+#0) [+-&|]= Rt
2957 def : Pat <(stOp (OpNode (ldOp (addrPred IntRegs:$addr, extPred:$offset)),
2958 (i32 IntRegs:$addend)),
2959 (addrPred (i32 IntRegs:$addr), extPred:$offset)),
2960 (MI IntRegs:$addr, extPred:$offset, (i32 IntRegs:$addend) )>;
2962 // mem[bhw](Rs+#U6:[012]) [+-&|]= Rt
2963 let AddedComplexity = 150 in
2964 def : Pat <(stOp (OpNode (ldOp (add IntRegs:$base, extPred:$offset)),
2965 (i32 IntRegs:$orend)),
2966 (add IntRegs:$base, extPred:$offset)),
2967 (MI IntRegs:$base, extPred:$offset, (i32 IntRegs:$orend) )>;
2970 multiclass MemOPr_ALUOp<PatFrag ldOp, PatFrag stOp,
2971 ComplexPattern addrPred, PatLeaf extPred,
2972 InstHexagon addMI, InstHexagon subMI,
2973 InstHexagon andMI, InstHexagon orMI > {
2975 defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, addMI, add>;
2976 defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, subMI, sub>;
2977 defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, andMI, and>;
2978 defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, orMI, or>;
2981 multiclass MemOPr_ExtType<PatFrag ldOpByte, PatFrag ldOpHalf > {
2983 defm : MemOPr_ALUOp <ldOpHalf, truncstorei16, ADDRriU6_1, u6_1ExtPred,
2984 L4_add_memoph_io, L4_sub_memoph_io,
2985 L4_and_memoph_io, L4_or_memoph_io>;
2987 defm : MemOPr_ALUOp <ldOpByte, truncstorei8, ADDRriU6_0, u6ExtPred,
2988 L4_add_memopb_io, L4_sub_memopb_io,
2989 L4_and_memopb_io, L4_or_memopb_io>;
2992 // Define 'def Pats' for MemOps with register addend.
2993 let Predicates = [HasV4T, UseMEMOP] in {
2995 defm : MemOPr_ExtType<zextloadi8, zextloadi16>; // zero extend
2996 defm : MemOPr_ExtType<sextloadi8, sextloadi16>; // sign extend
2997 defm : MemOPr_ExtType<extloadi8, extloadi16>; // any extend
2999 defm : MemOPr_ALUOp <load, store, ADDRriU6_2, u6_2ExtPred, L4_add_memopw_io,
3000 L4_sub_memopw_io, L4_and_memopw_io, L4_or_memopw_io >;
3003 //===----------------------------------------------------------------------===//
3005 //===----------------------------------------------------------------------===//
3007 // Hexagon V4 only supports these flavors of byte/half compare instructions:
3008 // EQ/GT/GTU. Other flavors like GE/GEU/LT/LTU/LE/LEU are not supported by
3009 // hardware. However, compiler can still implement these patterns through
3010 // appropriate patterns combinations based on current implemented patterns.
3011 // The implemented patterns are: EQ/GT/GTU.
3012 // Missing patterns are: GE/GEU/LT/LTU/LE/LEU.
3014 // Following instruction is not being extended as it results into the
3015 // incorrect code for negative numbers.
3016 // Pd=cmpb.eq(Rs,#u8)
3018 // p=!cmp.eq(r1,#s10)
3019 let isCodeGenOnly = 0 in {
3020 def C4_cmpneqi : T_CMP <"cmp.eq", 0b00, 1, s10Ext>;
3021 def C4_cmpltei : T_CMP <"cmp.gt", 0b01, 1, s10Ext>;
3022 def C4_cmplteui : T_CMP <"cmp.gtu", 0b10, 1, u9Ext>;
3025 def : T_CMP_pat <C4_cmpneqi, setne, s10ExtPred>;
3026 def : T_CMP_pat <C4_cmpltei, setle, s10ExtPred>;
3027 def : T_CMP_pat <C4_cmplteui, setule, u9ImmPred>;
3029 // rs <= rt -> !(rs > rt).
3031 def: Pat<(i1 (setle (i32 IntRegs:$src1), s10ExtPred:$src2)),
3032 (C2_not (C2_cmpgti IntRegs:$src1, s10ExtPred:$src2))>;
3033 // (C4_cmpltei IntRegs:$src1, s10ExtPred:$src2)>;
3035 // Map cmplt(Rs, Imm) -> !cmpgt(Rs, Imm-1).
3036 def: Pat<(i1 (setlt (i32 IntRegs:$src1), s8ExtPred:$src2)),
3037 (C4_cmpltei IntRegs:$src1, (DEC_CONST_SIGNED s8ExtPred:$src2))>;
3039 // rs != rt -> !(rs == rt).
3040 def: Pat<(i1 (setne (i32 IntRegs:$src1), s10ExtPred:$src2)),
3041 (C4_cmpneqi IntRegs:$src1, s10ExtPred:$src2)>;
3043 // SDNode for converting immediate C to C-1.
3044 def DEC_CONST_BYTE : SDNodeXForm<imm, [{
3045 // Return the byte immediate const-1 as an SDNode.
3046 int32_t imm = N->getSExtValue();
3047 return XformU7ToU7M1Imm(imm);
3051 // zext( seteq ( and(Rs, 255), u8))
3053 // Pd=cmpb.eq(Rs, #u8)
3054 // if (Pd.new) Rd=#1
3055 // if (!Pd.new) Rd=#0
3056 def : Pat <(i32 (zext (i1 (seteq (i32 (and (i32 IntRegs:$Rs), 255)),
3058 (i32 (TFR_condset_ii (i1 (A4_cmpbeqi (i32 IntRegs:$Rs),
3064 // zext( setne ( and(Rs, 255), u8))
3066 // Pd=cmpb.eq(Rs, #u8)
3067 // if (Pd.new) Rd=#0
3068 // if (!Pd.new) Rd=#1
3069 def : Pat <(i32 (zext (i1 (setne (i32 (and (i32 IntRegs:$Rs), 255)),
3071 (i32 (TFR_condset_ii (i1 (A4_cmpbeqi (i32 IntRegs:$Rs),
3077 // zext( seteq (Rs, and(Rt, 255)))
3079 // Pd=cmpb.eq(Rs, Rt)
3080 // if (Pd.new) Rd=#1
3081 // if (!Pd.new) Rd=#0
3082 def : Pat <(i32 (zext (i1 (seteq (i32 IntRegs:$Rt),
3083 (i32 (and (i32 IntRegs:$Rs), 255)))))),
3084 (i32 (TFR_condset_ii (i1 (A4_cmpbeq (i32 IntRegs:$Rs),
3085 (i32 IntRegs:$Rt))),
3090 // zext( setne (Rs, and(Rt, 255)))
3092 // Pd=cmpb.eq(Rs, Rt)
3093 // if (Pd.new) Rd=#0
3094 // if (!Pd.new) Rd=#1
3095 def : Pat <(i32 (zext (i1 (setne (i32 IntRegs:$Rt),
3096 (i32 (and (i32 IntRegs:$Rs), 255)))))),
3097 (i32 (TFR_condset_ii (i1 (A4_cmpbeq (i32 IntRegs:$Rs),
3098 (i32 IntRegs:$Rt))),
3103 // zext( setugt ( and(Rs, 255), u8))
3105 // Pd=cmpb.gtu(Rs, #u8)
3106 // if (Pd.new) Rd=#1
3107 // if (!Pd.new) Rd=#0
3108 def : Pat <(i32 (zext (i1 (setugt (i32 (and (i32 IntRegs:$Rs), 255)),
3110 (i32 (TFR_condset_ii (i1 (A4_cmpbgtui (i32 IntRegs:$Rs),
3116 // zext( setugt ( and(Rs, 254), u8))
3118 // Pd=cmpb.gtu(Rs, #u8)
3119 // if (Pd.new) Rd=#1
3120 // if (!Pd.new) Rd=#0
3121 def : Pat <(i32 (zext (i1 (setugt (i32 (and (i32 IntRegs:$Rs), 254)),
3123 (i32 (TFR_condset_ii (i1 (A4_cmpbgtui (i32 IntRegs:$Rs),
3129 // zext( setult ( Rs, Rt))
3131 // Pd=cmp.ltu(Rs, Rt)
3132 // if (Pd.new) Rd=#1
3133 // if (!Pd.new) Rd=#0
3134 // cmp.ltu(Rs, Rt) -> cmp.gtu(Rt, Rs)
3135 def : Pat <(i32 (zext (i1 (setult (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3136 (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rt),
3137 (i32 IntRegs:$Rs))),
3142 // zext( setlt ( Rs, Rt))
3144 // Pd=cmp.lt(Rs, Rt)
3145 // if (Pd.new) Rd=#1
3146 // if (!Pd.new) Rd=#0
3147 // cmp.lt(Rs, Rt) -> cmp.gt(Rt, Rs)
3148 def : Pat <(i32 (zext (i1 (setlt (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3149 (i32 (TFR_condset_ii (i1 (C2_cmpgt (i32 IntRegs:$Rt),
3150 (i32 IntRegs:$Rs))),
3155 // zext( setugt ( Rs, Rt))
3157 // Pd=cmp.gtu(Rs, Rt)
3158 // if (Pd.new) Rd=#1
3159 // if (!Pd.new) Rd=#0
3160 def : Pat <(i32 (zext (i1 (setugt (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3161 (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rs),
3162 (i32 IntRegs:$Rt))),
3166 // This pattern interefers with coremark performance, not implementing at this
3169 // zext( setgt ( Rs, Rt))
3171 // Pd=cmp.gt(Rs, Rt)
3172 // if (Pd.new) Rd=#1
3173 // if (!Pd.new) Rd=#0
3176 // zext( setuge ( Rs, Rt))
3178 // Pd=cmp.ltu(Rs, Rt)
3179 // if (Pd.new) Rd=#0
3180 // if (!Pd.new) Rd=#1
3181 // cmp.ltu(Rs, Rt) -> cmp.gtu(Rt, Rs)
3182 def : Pat <(i32 (zext (i1 (setuge (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3183 (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rt),
3184 (i32 IntRegs:$Rs))),
3189 // zext( setge ( Rs, Rt))
3191 // Pd=cmp.lt(Rs, Rt)
3192 // if (Pd.new) Rd=#0
3193 // if (!Pd.new) Rd=#1
3194 // cmp.lt(Rs, Rt) -> cmp.gt(Rt, Rs)
3195 def : Pat <(i32 (zext (i1 (setge (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3196 (i32 (TFR_condset_ii (i1 (C2_cmpgt (i32 IntRegs:$Rt),
3197 (i32 IntRegs:$Rs))),
3202 // zext( setule ( Rs, Rt))
3204 // Pd=cmp.gtu(Rs, Rt)
3205 // if (Pd.new) Rd=#0
3206 // if (!Pd.new) Rd=#1
3207 def : Pat <(i32 (zext (i1 (setule (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3208 (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rs),
3209 (i32 IntRegs:$Rt))),
3214 // zext( setle ( Rs, Rt))
3216 // Pd=cmp.gt(Rs, Rt)
3217 // if (Pd.new) Rd=#0
3218 // if (!Pd.new) Rd=#1
3219 def : Pat <(i32 (zext (i1 (setle (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3220 (i32 (TFR_condset_ii (i1 (C2_cmpgt (i32 IntRegs:$Rs),
3221 (i32 IntRegs:$Rt))),
3226 // zext( setult ( and(Rs, 255), u8))
3227 // Use the isdigit transformation below
3229 // Generate code of the form 'mux_ii(cmpbgtu(Rdd, C-1),0,1)'
3230 // for C code of the form r = ((c>='0') & (c<='9')) ? 1 : 0;.
3231 // The isdigit transformation relies on two 'clever' aspects:
3232 // 1) The data type is unsigned which allows us to eliminate a zero test after
3233 // biasing the expression by 48. We are depending on the representation of
3234 // the unsigned types, and semantics.
3235 // 2) The front end has converted <= 9 into < 10 on entry to LLVM
3238 // retval = ((c>='0') & (c<='9')) ? 1 : 0;
3239 // The code is transformed upstream of llvm into
3240 // retval = (c-48) < 10 ? 1 : 0;
3241 let AddedComplexity = 139 in
3242 def : Pat <(i32 (zext (i1 (setult (i32 (and (i32 IntRegs:$src1), 255)),
3243 u7StrictPosImmPred:$src2)))),
3244 (i32 (C2_muxii (i1 (A4_cmpbgtui (i32 IntRegs:$src1),
3245 (DEC_CONST_BYTE u7StrictPosImmPred:$src2))),
3249 //===----------------------------------------------------------------------===//
3251 //===----------------------------------------------------------------------===//
3253 //===----------------------------------------------------------------------===//
3254 // Multiclass for DeallocReturn
3255 //===----------------------------------------------------------------------===//
3256 class L4_RETURN<string mnemonic, bit isNot, bit isPredNew, bit isTak>
3257 : LD0Inst<(outs), (ins PredRegs:$src),
3258 !if(isNot, "if (!$src", "if ($src")#
3259 !if(isPredNew, ".new) ", ") ")#mnemonic#
3260 !if(isPredNew, #!if(isTak,":t", ":nt"),""),
3261 [], "", LD_tc_3or4stall_SLOT0> {
3264 let BaseOpcode = "L4_RETURN";
3265 let isPredicatedFalse = isNot;
3266 let isPredicatedNew = isPredNew;
3267 let isTaken = isTak;
3268 let IClass = 0b1001;
3270 let Inst{27-16} = 0b011000011110;
3272 let Inst{13} = isNot;
3273 let Inst{12} = isTak;
3274 let Inst{11} = isPredNew;
3276 let Inst{9-8} = src;
3277 let Inst{4-0} = 0b11110;
3280 // Produce all predicated forms, p, !p, p.new, !p.new, :t, :nt
3281 multiclass L4_RETURN_PRED<string mnemonic, bit PredNot> {
3282 let isPredicated = 1 in {
3283 def _#NAME# : L4_RETURN <mnemonic, PredNot, 0, 1>;
3284 def _#NAME#new_pnt : L4_RETURN <mnemonic, PredNot, 1, 0>;
3285 def _#NAME#new_pt : L4_RETURN <mnemonic, PredNot, 1, 1>;
3289 multiclass LD_MISC_L4_RETURN<string mnemonic> {
3290 let isBarrier = 1, isPredicable = 1 in
3291 def NAME : LD0Inst <(outs), (ins), mnemonic, [], "",
3292 LD_tc_3or4stall_SLOT0> {
3293 let BaseOpcode = "L4_RETURN";
3294 let IClass = 0b1001;
3295 let Inst{27-16} = 0b011000011110;
3296 let Inst{13-10} = 0b0000;
3297 let Inst{4-0} = 0b11110;
3299 defm t : L4_RETURN_PRED<mnemonic, 0 >;
3300 defm f : L4_RETURN_PRED<mnemonic, 1 >;
3303 let isReturn = 1, isTerminator = 1,
3304 Defs = [R29, R30, R31, PC], Uses = [R30], hasSideEffects = 0,
3305 validSubTargets = HasV4SubT, isCodeGenOnly = 0 in
3306 defm L4_return: LD_MISC_L4_RETURN <"dealloc_return">, PredNewRel;
3308 // Restore registers and dealloc return function call.
3309 let isCall = 1, isBarrier = 1, isReturn = 1, isTerminator = 1,
3310 Defs = [R29, R30, R31, PC], isAsmParserOnly = 1 in {
3311 let validSubTargets = HasV4SubT in
3312 def RESTORE_DEALLOC_RET_JMP_V4 : JInst<(outs),
3313 (ins calltarget:$dst),
3319 // Restore registers and dealloc frame before a tail call.
3320 let isCall = 1, isBarrier = 1, isAsmParserOnly = 1,
3321 Defs = [R29, R30, R31, PC] in {
3322 let validSubTargets = HasV4SubT in
3323 def RESTORE_DEALLOC_BEFORE_TAILCALL_V4 : JInst<(outs),
3324 (ins calltarget:$dst),
3330 // Save registers function call.
3331 let isCall = 1, isBarrier = 1, isAsmParserOnly = 1,
3332 Uses = [R29, R31] in {
3333 def SAVE_REGISTERS_CALL_V4 : JInst<(outs),
3334 (ins calltarget:$dst),
3335 "call $dst // Save_calle_saved_registers",
3340 //===----------------------------------------------------------------------===//
3341 // Template class for non predicated store instructions with
3342 // GP-Relative or absolute addressing.
3343 //===----------------------------------------------------------------------===//
3344 let hasSideEffects = 0, isPredicable = 1, isNVStorable = 1 in
3345 class T_StoreAbsGP <string mnemonic, RegisterClass RC, Operand ImmOp,
3346 bits<2>MajOp, Operand AddrOp, bit isAbs, bit isHalf>
3347 : STInst<(outs), (ins AddrOp:$addr, RC:$src),
3348 mnemonic # !if(isAbs, "(##", "(#")#"$addr) = $src"#!if(isHalf, ".h",""),
3349 [], "", V2LDST_tc_st_SLOT01> {
3352 bits<16> offsetBits;
3354 string ImmOpStr = !cast<string>(ImmOp);
3355 let offsetBits = !if (!eq(ImmOpStr, "u16_3Imm"), addr{18-3},
3356 !if (!eq(ImmOpStr, "u16_2Imm"), addr{17-2},
3357 !if (!eq(ImmOpStr, "u16_1Imm"), addr{16-1},
3358 /* u16_0Imm */ addr{15-0})));
3359 let IClass = 0b0100;
3361 let Inst{26-25} = offsetBits{15-14};
3363 let Inst{23-22} = MajOp;
3364 let Inst{21} = isHalf;
3365 let Inst{20-16} = offsetBits{13-9};
3366 let Inst{13} = offsetBits{8};
3367 let Inst{12-8} = src;
3368 let Inst{7-0} = offsetBits{7-0};
3371 //===----------------------------------------------------------------------===//
3372 // Template class for predicated store instructions with
3373 // GP-Relative or absolute addressing.
3374 //===----------------------------------------------------------------------===//
3375 let hasSideEffects = 0, isPredicated = 1, isNVStorable = 1, opExtentBits = 6,
3377 class T_StoreAbs_Pred <string mnemonic, RegisterClass RC, bits<2> MajOp,
3378 bit isHalf, bit isNot, bit isNew>
3379 : STInst<(outs), (ins PredRegs:$src1, u6Ext:$absaddr, RC: $src2),
3380 !if(isNot, "if (!$src1", "if ($src1")#!if(isNew, ".new) ",
3381 ") ")#mnemonic#"(#$absaddr) = $src2"#!if(isHalf, ".h",""),
3382 [], "", ST_tc_st_SLOT01>, AddrModeRel {
3387 let isPredicatedNew = isNew;
3388 let isPredicatedFalse = isNot;
3390 let IClass = 0b1010;
3392 let Inst{27-24} = 0b1111;
3393 let Inst{23-22} = MajOp;
3394 let Inst{21} = isHalf;
3395 let Inst{17-16} = absaddr{5-4};
3396 let Inst{13} = isNew;
3397 let Inst{12-8} = src2;
3399 let Inst{6-3} = absaddr{3-0};
3400 let Inst{2} = isNot;
3401 let Inst{1-0} = src1;
3404 //===----------------------------------------------------------------------===//
3405 // Template class for predicated store instructions with absolute addressing.
3406 //===----------------------------------------------------------------------===//
3407 class T_StoreAbs <string mnemonic, RegisterClass RC, Operand ImmOp,
3408 bits<2> MajOp, bit isHalf>
3409 : T_StoreAbsGP <mnemonic, RC, ImmOp, MajOp, u0AlwaysExt, 1, isHalf>,
3411 string ImmOpStr = !cast<string>(ImmOp);
3412 let opExtentBits = !if (!eq(ImmOpStr, "u16_3Imm"), 19,
3413 !if (!eq(ImmOpStr, "u16_2Imm"), 18,
3414 !if (!eq(ImmOpStr, "u16_1Imm"), 17,
3415 /* u16_0Imm */ 16)));
3417 let opExtentAlign = !if (!eq(ImmOpStr, "u16_3Imm"), 3,
3418 !if (!eq(ImmOpStr, "u16_2Imm"), 2,
3419 !if (!eq(ImmOpStr, "u16_1Imm"), 1,
3420 /* u16_0Imm */ 0)));
3423 //===----------------------------------------------------------------------===//
3424 // Multiclass for store instructions with absolute addressing.
3425 //===----------------------------------------------------------------------===//
3426 let validSubTargets = HasV4SubT, addrMode = Absolute, isExtended = 1 in
3427 multiclass ST_Abs<string mnemonic, string CextOp, RegisterClass RC,
3428 Operand ImmOp, bits<2> MajOp, bit isHalf = 0> {
3429 let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in {
3430 let opExtendable = 0, isPredicable = 1 in
3431 def S2_#NAME#abs : T_StoreAbs <mnemonic, RC, ImmOp, MajOp, isHalf>;
3434 def S4_p#NAME#t_abs : T_StoreAbs_Pred<mnemonic, RC, MajOp, isHalf, 0, 0>;
3435 def S4_p#NAME#f_abs : T_StoreAbs_Pred<mnemonic, RC, MajOp, isHalf, 1, 0>;
3438 def S4_p#NAME#tnew_abs : T_StoreAbs_Pred<mnemonic, RC, MajOp, isHalf, 0, 1>;
3439 def S4_p#NAME#fnew_abs : T_StoreAbs_Pred<mnemonic, RC, MajOp, isHalf, 1, 1>;
3443 //===----------------------------------------------------------------------===//
3444 // Template class for non predicated new-value store instructions with
3445 // GP-Relative or absolute addressing.
3446 //===----------------------------------------------------------------------===//
3447 let hasSideEffects = 0, isPredicable = 1, mayStore = 1, isNVStore = 1,
3448 isNewValue = 1, opNewValue = 1 in
3449 class T_StoreAbsGP_NV <string mnemonic, Operand ImmOp, bits<2>MajOp, bit isAbs>
3450 : NVInst_V4<(outs), (ins u0AlwaysExt:$addr, IntRegs:$src),
3451 mnemonic # !if(isAbs, "(##", "(#")#"$addr) = $src.new",
3452 [], "", V2LDST_tc_st_SLOT0> {
3455 bits<16> offsetBits;
3457 string ImmOpStr = !cast<string>(ImmOp);
3458 let offsetBits = !if (!eq(ImmOpStr, "u16_3Imm"), addr{18-3},
3459 !if (!eq(ImmOpStr, "u16_2Imm"), addr{17-2},
3460 !if (!eq(ImmOpStr, "u16_1Imm"), addr{16-1},
3461 /* u16_0Imm */ addr{15-0})));
3462 let IClass = 0b0100;
3465 let Inst{26-25} = offsetBits{15-14};
3466 let Inst{24-21} = 0b0101;
3467 let Inst{20-16} = offsetBits{13-9};
3468 let Inst{13} = offsetBits{8};
3469 let Inst{12-11} = MajOp;
3470 let Inst{10-8} = src;
3471 let Inst{7-0} = offsetBits{7-0};
3474 //===----------------------------------------------------------------------===//
3475 // Template class for predicated new-value store instructions with
3476 // absolute addressing.
3477 //===----------------------------------------------------------------------===//
3478 let hasSideEffects = 0, isPredicated = 1, mayStore = 1, isNVStore = 1,
3479 isNewValue = 1, opNewValue = 2, opExtentBits = 6, opExtendable = 1 in
3480 class T_StoreAbs_NV_Pred <string mnemonic, bits<2> MajOp, bit isNot, bit isNew>
3481 : NVInst_V4<(outs), (ins PredRegs:$src1, u6Ext:$absaddr, IntRegs:$src2),
3482 !if(isNot, "if (!$src1", "if ($src1")#!if(isNew, ".new) ",
3483 ") ")#mnemonic#"(#$absaddr) = $src2.new",
3484 [], "", ST_tc_st_SLOT0>, AddrModeRel {
3489 let isPredicatedNew = isNew;
3490 let isPredicatedFalse = isNot;
3492 let IClass = 0b1010;
3494 let Inst{27-24} = 0b1111;
3495 let Inst{23-21} = 0b101;
3496 let Inst{17-16} = absaddr{5-4};
3497 let Inst{13} = isNew;
3498 let Inst{12-11} = MajOp;
3499 let Inst{10-8} = src2;
3501 let Inst{6-3} = absaddr{3-0};
3502 let Inst{2} = isNot;
3503 let Inst{1-0} = src1;
3506 //===----------------------------------------------------------------------===//
3507 // Template class for non-predicated new-value store instructions with
3508 // absolute addressing.
3509 //===----------------------------------------------------------------------===//
3510 class T_StoreAbs_NV <string mnemonic, Operand ImmOp, bits<2> MajOp>
3511 : T_StoreAbsGP_NV <mnemonic, ImmOp, MajOp, 1>, AddrModeRel {
3513 string ImmOpStr = !cast<string>(ImmOp);
3514 let opExtentBits = !if (!eq(ImmOpStr, "u16_3Imm"), 19,
3515 !if (!eq(ImmOpStr, "u16_2Imm"), 18,
3516 !if (!eq(ImmOpStr, "u16_1Imm"), 17,
3517 /* u16_0Imm */ 16)));
3519 let opExtentAlign = !if (!eq(ImmOpStr, "u16_3Imm"), 3,
3520 !if (!eq(ImmOpStr, "u16_2Imm"), 2,
3521 !if (!eq(ImmOpStr, "u16_1Imm"), 1,
3522 /* u16_0Imm */ 0)));
3525 //===----------------------------------------------------------------------===//
3526 // Multiclass for new-value store instructions with absolute addressing.
3527 //===----------------------------------------------------------------------===//
3528 let validSubTargets = HasV4SubT, addrMode = Absolute, isExtended = 1 in
3529 multiclass ST_Abs_NV <string mnemonic, string CextOp, Operand ImmOp,
3531 let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in {
3532 let opExtendable = 0, isPredicable = 1 in
3533 def S2_#NAME#newabs : T_StoreAbs_NV <mnemonic, ImmOp, MajOp>;
3536 def S4_p#NAME#newt_abs : T_StoreAbs_NV_Pred <mnemonic, MajOp, 0, 0>;
3537 def S4_p#NAME#newf_abs : T_StoreAbs_NV_Pred <mnemonic, MajOp, 1, 0>;
3540 def S4_p#NAME#newtnew_abs : T_StoreAbs_NV_Pred <mnemonic, MajOp, 0, 1>;
3541 def S4_p#NAME#newfnew_abs : T_StoreAbs_NV_Pred <mnemonic, MajOp, 1, 1>;
3545 //===----------------------------------------------------------------------===//
3546 // Stores with absolute addressing
3547 //===----------------------------------------------------------------------===//
3548 let accessSize = ByteAccess, isCodeGenOnly = 0 in
3549 defm storerb : ST_Abs <"memb", "STrib", IntRegs, u16_0Imm, 0b00>,
3550 ST_Abs_NV <"memb", "STrib", u16_0Imm, 0b00>;
3552 let accessSize = HalfWordAccess, isCodeGenOnly = 0 in
3553 defm storerh : ST_Abs <"memh", "STrih", IntRegs, u16_1Imm, 0b01>,
3554 ST_Abs_NV <"memh", "STrih", u16_1Imm, 0b01>;
3556 let accessSize = WordAccess, isCodeGenOnly = 0 in
3557 defm storeri : ST_Abs <"memw", "STriw", IntRegs, u16_2Imm, 0b10>,
3558 ST_Abs_NV <"memw", "STriw", u16_2Imm, 0b10>;
3560 let isNVStorable = 0, accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
3561 defm storerd : ST_Abs <"memd", "STrid", DoubleRegs, u16_3Imm, 0b11>;
3563 let isNVStorable = 0, accessSize = HalfWordAccess, isCodeGenOnly = 0 in
3564 defm storerf : ST_Abs <"memh", "STrif", IntRegs, u16_1Imm, 0b01, 1>;
3566 //===----------------------------------------------------------------------===//
3567 // GP-relative stores.
3568 // mem[bhwd](#global)=Rt
3569 // Once predicated, these instructions map to absolute addressing mode.
3570 // if ([!]Pv[.new]) mem[bhwd](##global)=Rt
3571 //===----------------------------------------------------------------------===//
3573 let validSubTargets = HasV4SubT, isAsmParserOnly = 1 in
3574 class T_StoreGP <string mnemonic, string BaseOp, RegisterClass RC,
3575 Operand ImmOp, bits<2> MajOp, bit isHalf = 0>
3576 : T_StoreAbsGP <mnemonic, RC, ImmOp, MajOp, globaladdress, 0, isHalf> {
3577 // Set BaseOpcode same as absolute addressing instructions so that
3578 // non-predicated GP-Rel instructions can have relate with predicated
3579 // Absolute instruction.
3580 let BaseOpcode = BaseOp#_abs;
3583 let validSubTargets = HasV4SubT, isAsmParserOnly = 1 in
3584 multiclass ST_GP <string mnemonic, string BaseOp, Operand ImmOp,
3585 bits<2> MajOp, bit isHalf = 0> {
3586 // Set BaseOpcode same as absolute addressing instructions so that
3587 // non-predicated GP-Rel instructions can have relate with predicated
3588 // Absolute instruction.
3589 let BaseOpcode = BaseOp#_abs in {
3590 def NAME#gp : T_StoreAbsGP <mnemonic, IntRegs, ImmOp, MajOp,
3591 globaladdress, 0, isHalf>;
3593 def NAME#newgp : T_StoreAbsGP_NV <mnemonic, ImmOp, MajOp, 0> ;
3597 let accessSize = ByteAccess in
3598 defm S2_storerb : ST_GP<"memb", "STrib", u16_0Imm, 0b00>, NewValueRel;
3600 let accessSize = HalfWordAccess in
3601 defm S2_storerh : ST_GP<"memh", "STrih", u16_1Imm, 0b01>, NewValueRel;
3603 let accessSize = WordAccess in
3604 defm S2_storeri : ST_GP<"memw", "STriw", u16_2Imm, 0b10>, NewValueRel;
3606 let isNVStorable = 0, accessSize = DoubleWordAccess in
3607 def S2_storerdgp : T_StoreGP <"memd", "STrid", DoubleRegs,
3608 u16_3Imm, 0b11>, PredNewRel;
3610 let isNVStorable = 0, accessSize = HalfWordAccess in
3611 def S2_storerfgp : T_StoreGP <"memh", "STrif", IntRegs,
3612 u16_1Imm, 0b01, 1>, PredNewRel;
3614 class Loada_pat<PatFrag Load, ValueType VT, PatFrag Addr, InstHexagon MI>
3615 : Pat<(VT (Load Addr:$addr)), (MI Addr:$addr)>;
3617 class Loadam_pat<PatFrag Load, ValueType VT, PatFrag Addr, PatFrag ValueMod,
3619 : Pat<(VT (Load Addr:$addr)), (ValueMod (MI Addr:$addr))>;
3621 class Storea_pat<PatFrag Store, PatFrag Value, PatFrag Addr, InstHexagon MI>
3622 : Pat<(Store Value:$val, Addr:$addr), (MI Addr:$addr, Value:$val)>;
3624 class Stoream_pat<PatFrag Store, PatFrag Value, PatFrag Addr, PatFrag ValueMod,
3626 : Pat<(Store Value:$val, Addr:$addr),
3627 (MI Addr:$addr, (ValueMod Value:$val))>;
3629 def: Storea_pat<SwapSt<atomic_store_8>, I32, addrgp, S2_storerbgp>;
3630 def: Storea_pat<SwapSt<atomic_store_16>, I32, addrgp, S2_storerhgp>;
3631 def: Storea_pat<SwapSt<atomic_store_32>, I32, addrgp, S2_storerigp>;
3632 def: Storea_pat<SwapSt<atomic_store_64>, I64, addrgp, S2_storerdgp>;
3634 let AddedComplexity = 100 in {
3635 def: Storea_pat<truncstorei8, I32, addrgp, S2_storerbgp>;
3636 def: Storea_pat<truncstorei16, I32, addrgp, S2_storerhgp>;
3637 def: Storea_pat<store, I32, addrgp, S2_storerigp>;
3638 def: Storea_pat<store, I64, addrgp, S2_storerdgp>;
3640 // Map from "i1 = constant<-1>; memw(CONST32(#foo)) = i1"
3641 // to "r0 = 1; memw(#foo) = r0"
3642 let AddedComplexity = 100 in
3643 def: Pat<(store (i1 -1), (HexagonCONST32_GP tglobaladdr:$global)),
3644 (S2_storerbgp tglobaladdr:$global, (A2_tfrsi 1))>;
3647 //===----------------------------------------------------------------------===//
3648 // Template class for non predicated load instructions with
3649 // absolute addressing mode.
3650 //===----------------------------------------------------------------------===//
3651 let isPredicable = 1, hasSideEffects = 0, validSubTargets = HasV4SubT in
3652 class T_LoadAbsGP <string mnemonic, RegisterClass RC, Operand ImmOp,
3653 bits<3> MajOp, Operand AddrOp, bit isAbs>
3654 : LDInst <(outs RC:$dst), (ins AddrOp:$addr),
3655 "$dst = "#mnemonic# !if(isAbs, "(##", "(#")#"$addr)",
3656 [], "", V2LDST_tc_ld_SLOT01> {
3659 bits<16> offsetBits;
3661 string ImmOpStr = !cast<string>(ImmOp);
3662 let offsetBits = !if (!eq(ImmOpStr, "u16_3Imm"), addr{18-3},
3663 !if (!eq(ImmOpStr, "u16_2Imm"), addr{17-2},
3664 !if (!eq(ImmOpStr, "u16_1Imm"), addr{16-1},
3665 /* u16_0Imm */ addr{15-0})));
3667 let IClass = 0b0100;
3670 let Inst{26-25} = offsetBits{15-14};
3672 let Inst{23-21} = MajOp;
3673 let Inst{20-16} = offsetBits{13-9};
3674 let Inst{13-5} = offsetBits{8-0};
3675 let Inst{4-0} = dst;
3678 class T_LoadAbs <string mnemonic, RegisterClass RC, Operand ImmOp,
3680 : T_LoadAbsGP <mnemonic, RC, ImmOp, MajOp, u0AlwaysExt, 1>, AddrModeRel {
3682 string ImmOpStr = !cast<string>(ImmOp);
3683 let opExtentBits = !if (!eq(ImmOpStr, "u16_3Imm"), 19,
3684 !if (!eq(ImmOpStr, "u16_2Imm"), 18,
3685 !if (!eq(ImmOpStr, "u16_1Imm"), 17,
3686 /* u16_0Imm */ 16)));
3688 let opExtentAlign = !if (!eq(ImmOpStr, "u16_3Imm"), 3,
3689 !if (!eq(ImmOpStr, "u16_2Imm"), 2,
3690 !if (!eq(ImmOpStr, "u16_1Imm"), 1,
3691 /* u16_0Imm */ 0)));
3693 //===----------------------------------------------------------------------===//
3694 // Template class for predicated load instructions with
3695 // absolute addressing mode.
3696 //===----------------------------------------------------------------------===//
3697 let isPredicated = 1, hasNewValue = 1, opExtentBits = 6, opExtendable = 2 in
3698 class T_LoadAbs_Pred <string mnemonic, RegisterClass RC, bits<3> MajOp,
3699 bit isPredNot, bit isPredNew>
3700 : LDInst <(outs RC:$dst), (ins PredRegs:$src1, u6Ext:$absaddr),
3701 !if(isPredNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
3702 ") ")#"$dst = "#mnemonic#"(#$absaddr)">, AddrModeRel {
3707 let isPredicatedNew = isPredNew;
3708 let isPredicatedFalse = isPredNot;
3710 let IClass = 0b1001;
3712 let Inst{27-24} = 0b1111;
3713 let Inst{23-21} = MajOp;
3714 let Inst{20-16} = absaddr{5-1};
3716 let Inst{12} = isPredNew;
3717 let Inst{11} = isPredNot;
3718 let Inst{10-9} = src1;
3719 let Inst{8} = absaddr{0};
3721 let Inst{4-0} = dst;
3724 //===----------------------------------------------------------------------===//
3725 // Multiclass for the load instructions with absolute addressing mode.
3726 //===----------------------------------------------------------------------===//
3727 multiclass LD_Abs_Pred<string mnemonic, RegisterClass RC, bits<3> MajOp,
3729 def _abs : T_LoadAbs_Pred <mnemonic, RC, MajOp, PredNot, 0>;
3731 def new_abs : T_LoadAbs_Pred <mnemonic, RC, MajOp, PredNot, 1>;
3734 let addrMode = Absolute, isExtended = 1 in
3735 multiclass LD_Abs<string mnemonic, string CextOp, RegisterClass RC,
3736 Operand ImmOp, bits<3> MajOp> {
3737 let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in {
3738 let opExtendable = 1, isPredicable = 1 in
3739 def L4_#NAME#_abs: T_LoadAbs <mnemonic, RC, ImmOp, MajOp>;
3742 defm L4_p#NAME#t : LD_Abs_Pred<mnemonic, RC, MajOp, 0>;
3743 defm L4_p#NAME#f : LD_Abs_Pred<mnemonic, RC, MajOp, 1>;
3747 let accessSize = ByteAccess, hasNewValue = 1, isCodeGenOnly = 0 in {
3748 defm loadrb : LD_Abs<"memb", "LDrib", IntRegs, u16_0Imm, 0b000>;
3749 defm loadrub : LD_Abs<"memub", "LDriub", IntRegs, u16_0Imm, 0b001>;
3752 let accessSize = HalfWordAccess, hasNewValue = 1, isCodeGenOnly = 0 in {
3753 defm loadrh : LD_Abs<"memh", "LDrih", IntRegs, u16_1Imm, 0b010>;
3754 defm loadruh : LD_Abs<"memuh", "LDriuh", IntRegs, u16_1Imm, 0b011>;
3757 let accessSize = WordAccess, hasNewValue = 1, isCodeGenOnly = 0 in
3758 defm loadri : LD_Abs<"memw", "LDriw", IntRegs, u16_2Imm, 0b100>;
3760 let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
3761 defm loadrd : LD_Abs<"memd", "LDrid", DoubleRegs, u16_3Imm, 0b110>;
3763 //===----------------------------------------------------------------------===//
3764 // multiclass for load instructions with GP-relative addressing mode.
3765 // Rx=mem[bhwd](##global)
3766 // Once predicated, these instructions map to absolute addressing mode.
3767 // if ([!]Pv[.new]) Rx=mem[bhwd](##global)
3768 //===----------------------------------------------------------------------===//
3770 let isAsmParserOnly = 1 in
3771 class T_LoadGP <string mnemonic, string BaseOp, RegisterClass RC, Operand ImmOp,
3773 : T_LoadAbsGP <mnemonic, RC, ImmOp, MajOp, globaladdress, 0>, PredNewRel {
3774 let BaseOpcode = BaseOp#_abs;
3777 let accessSize = ByteAccess, hasNewValue = 1 in {
3778 def L2_loadrbgp : T_LoadGP<"memb", "LDrib", IntRegs, u16_0Imm, 0b000>;
3779 def L2_loadrubgp : T_LoadGP<"memub", "LDriub", IntRegs, u16_0Imm, 0b001>;
3782 let accessSize = HalfWordAccess, hasNewValue = 1 in {
3783 def L2_loadrhgp : T_LoadGP<"memh", "LDrih", IntRegs, u16_1Imm, 0b010>;
3784 def L2_loadruhgp : T_LoadGP<"memuh", "LDriuh", IntRegs, u16_1Imm, 0b011>;
3787 let accessSize = WordAccess, hasNewValue = 1 in
3788 def L2_loadrigp : T_LoadGP<"memw", "LDriw", IntRegs, u16_2Imm, 0b100>;
3790 let accessSize = DoubleWordAccess in
3791 def L2_loadrdgp : T_LoadGP<"memd", "LDrid", DoubleRegs, u16_3Imm, 0b110>;
3793 def: Loada_pat<atomic_load_8, i32, addrgp, L2_loadrubgp>;
3794 def: Loada_pat<atomic_load_16, i32, addrgp, L2_loadruhgp>;
3795 def: Loada_pat<atomic_load_32, i32, addrgp, L2_loadrigp>;
3796 def: Loada_pat<atomic_load_64, i64, addrgp, L2_loadrdgp>;
3798 // Map from load(globaladdress) -> mem[u][bhwd](#foo)
3799 class LoadGP_pats <PatFrag ldOp, InstHexagon MI, ValueType VT = i32>
3800 : Pat <(VT (ldOp (HexagonCONST32_GP tglobaladdr:$global))),
3801 (VT (MI tglobaladdr:$global))>;
3803 let AddedComplexity = 100 in {
3804 def: LoadGP_pats <extloadi8, L2_loadrbgp>;
3805 def: LoadGP_pats <sextloadi8, L2_loadrbgp>;
3806 def: LoadGP_pats <zextloadi8, L2_loadrubgp>;
3807 def: LoadGP_pats <extloadi16, L2_loadrhgp>;
3808 def: LoadGP_pats <sextloadi16, L2_loadrhgp>;
3809 def: LoadGP_pats <zextloadi16, L2_loadruhgp>;
3810 def: LoadGP_pats <load, L2_loadrigp>;
3811 def: LoadGP_pats <load, L2_loadrdgp, i64>;
3814 // When the Interprocedural Global Variable optimizer realizes that a certain
3815 // global variable takes only two constant values, it shrinks the global to
3816 // a boolean. Catch those loads here in the following 3 patterns.
3817 let AddedComplexity = 100 in {
3818 def: LoadGP_pats <extloadi1, L2_loadrubgp>;
3819 def: LoadGP_pats <zextloadi1, L2_loadrubgp>;
3822 def: Pat<(i64 (ctlz I64:$src1)), (Zext64 (S2_cl0p I64:$src1))>;
3823 def: Pat<(i64 (cttz I64:$src1)), (Zext64 (S2_ct0p I64:$src1))>;
3825 let AddedComplexity = 30 in {
3826 def: Storea_pat<truncstorei8, I32, u0AlwaysExtPred, S2_storerbabs>;
3827 def: Storea_pat<truncstorei16, I32, u0AlwaysExtPred, S2_storerhabs>;
3828 def: Storea_pat<store, I32, u0AlwaysExtPred, S2_storeriabs>;
3831 let AddedComplexity = 30 in {
3832 def: Loada_pat<load, i32, u0AlwaysExtPred, L4_loadri_abs>;
3833 def: Loada_pat<sextloadi8, i32, u0AlwaysExtPred, L4_loadrb_abs>;
3834 def: Loada_pat<zextloadi8, i32, u0AlwaysExtPred, L4_loadrub_abs>;
3835 def: Loada_pat<sextloadi16, i32, u0AlwaysExtPred, L4_loadrh_abs>;
3836 def: Loada_pat<zextloadi16, i32, u0AlwaysExtPred, L4_loadruh_abs>;
3839 // Indexed store word - global address.
3840 // memw(Rs+#u6:2)=#S8
3841 let AddedComplexity = 100 in
3842 def: Storex_add_pat<store, addrga, u6_2ImmPred, S4_storeiri_io>;
3844 // Load from a global address that has only one use in the current basic block.
3845 let AddedComplexity = 100 in {
3846 def: Loada_pat<extloadi8, i32, addrga, L4_loadrub_abs>;
3847 def: Loada_pat<sextloadi8, i32, addrga, L4_loadrb_abs>;
3848 def: Loada_pat<zextloadi8, i32, addrga, L4_loadrub_abs>;
3850 def: Loada_pat<extloadi16, i32, addrga, L4_loadruh_abs>;
3851 def: Loada_pat<sextloadi16, i32, addrga, L4_loadrh_abs>;
3852 def: Loada_pat<zextloadi16, i32, addrga, L4_loadruh_abs>;
3854 def: Loada_pat<load, i32, addrga, L4_loadri_abs>;
3855 def: Loada_pat<load, i64, addrga, L4_loadrd_abs>;
3858 // Store to a global address that has only one use in the current basic block.
3859 let AddedComplexity = 100 in {
3860 def: Storea_pat<truncstorei8, I32, addrga, S2_storerbabs>;
3861 def: Storea_pat<truncstorei16, I32, addrga, S2_storerhabs>;
3862 def: Storea_pat<store, I32, addrga, S2_storeriabs>;
3863 def: Storea_pat<store, I64, addrga, S2_storerdabs>;
3865 def: Stoream_pat<truncstorei32, I64, addrga, LoReg, S2_storeriabs>;
3868 // Map from Pd = load(globaladdress) -> Rd = memb(globaladdress), Pd = Rd
3869 let AddedComplexity = 100 in
3870 def : Pat <(i1 (load (HexagonCONST32_GP tglobaladdr:$global))),
3871 (i1 (C2_tfrrp (i32 (L2_loadrbgp tglobaladdr:$global))))>;
3873 // Transfer global address into a register
3874 let isExtended = 1, opExtendable = 1, AddedComplexity=50, isMoveImm = 1,
3875 isAsCheapAsAMove = 1, isReMaterializable = 1, validSubTargets = HasV4SubT,
3876 isCodeGenOnly = 1 in
3877 def TFRI_V4 : ALU32_ri<(outs IntRegs:$dst), (ins s16Ext:$src1),
3879 [(set IntRegs:$dst, (HexagonCONST32 tglobaladdr:$src1))]>,
3882 // Transfer a block address into a register
3883 def : Pat<(HexagonCONST32_GP tblockaddress:$src1),
3884 (TFRI_V4 tblockaddress:$src1)>,
3887 let AddedComplexity = 50, Predicates = [HasV4T] in
3888 def : Pat<(HexagonCONST32_GP tglobaladdr:$src1),
3889 (TFRI_V4 tglobaladdr:$src1)>,
3892 // i8/i16/i32 -> i64 loads
3893 // We need a complexity of 120 here to override preceding handling of
3895 let AddedComplexity = 120 in {
3896 def: Loadam_pat<extloadi8, i64, addrga, Zext64, L4_loadrub_abs>;
3897 def: Loadam_pat<sextloadi8, i64, addrga, Sext64, L4_loadrb_abs>;
3898 def: Loadam_pat<zextloadi8, i64, addrga, Zext64, L4_loadrub_abs>;
3900 def: Loadam_pat<extloadi16, i64, addrga, Zext64, L4_loadruh_abs>;
3901 def: Loadam_pat<sextloadi16, i64, addrga, Sext64, L4_loadrh_abs>;
3902 def: Loadam_pat<zextloadi16, i64, addrga, Zext64, L4_loadruh_abs>;
3904 def: Loadam_pat<extloadi32, i64, addrga, Zext64, L4_loadri_abs>;
3905 def: Loadam_pat<sextloadi32, i64, addrga, Sext64, L4_loadri_abs>;
3906 def: Loadam_pat<zextloadi32, i64, addrga, Zext64, L4_loadri_abs>;
3908 let AddedComplexity = 100 in {
3909 def: Loada_pat<extloadi8, i32, addrgp, L4_loadrub_abs>;
3910 def: Loada_pat<sextloadi8, i32, addrgp, L4_loadrb_abs>;
3911 def: Loada_pat<zextloadi8, i32, addrgp, L4_loadrub_abs>;
3913 def: Loada_pat<extloadi16, i32, addrgp, L4_loadruh_abs>;
3914 def: Loada_pat<sextloadi16, i32, addrgp, L4_loadrh_abs>;
3915 def: Loada_pat<zextloadi16, i32, addrgp, L4_loadruh_abs>;
3917 def: Loada_pat<load, i32, addrgp, L4_loadri_abs>;
3918 def: Loada_pat<load, i64, addrgp, L4_loadrd_abs>;
3921 let AddedComplexity = 100 in {
3922 def: Storea_pat<truncstorei8, I32, addrgp, S2_storerbabs>;
3923 def: Storea_pat<truncstorei16, I32, addrgp, S2_storerhabs>;
3924 def: Storea_pat<store, I32, addrgp, S2_storeriabs>;
3925 def: Storea_pat<store, I64, addrgp, S2_storerdabs>;
3928 def: Loada_pat<atomic_load_8, i32, addrgp, L4_loadrub_abs>;
3929 def: Loada_pat<atomic_load_16, i32, addrgp, L4_loadruh_abs>;
3930 def: Loada_pat<atomic_load_32, i32, addrgp, L4_loadri_abs>;
3931 def: Loada_pat<atomic_load_64, i64, addrgp, L4_loadrd_abs>;
3933 def: Storea_pat<SwapSt<atomic_store_8>, I32, addrgp, S2_storerbabs>;
3934 def: Storea_pat<SwapSt<atomic_store_16>, I32, addrgp, S2_storerhabs>;
3935 def: Storea_pat<SwapSt<atomic_store_32>, I32, addrgp, S2_storeriabs>;
3936 def: Storea_pat<SwapSt<atomic_store_64>, I64, addrgp, S2_storerdabs>;
3938 //===----------------------------------------------------------------------===//
3939 // :raw for of boundscheck:hi:lo insns
3940 //===----------------------------------------------------------------------===//
3942 // A4_boundscheck_lo: Detect if a register is within bounds.
3943 let hasSideEffects = 0, isCodeGenOnly = 0 in
3944 def A4_boundscheck_lo: ALU64Inst <
3945 (outs PredRegs:$Pd),
3946 (ins DoubleRegs:$Rss, DoubleRegs:$Rtt),
3947 "$Pd = boundscheck($Rss, $Rtt):raw:lo"> {
3952 let IClass = 0b1101;
3954 let Inst{27-23} = 0b00100;
3956 let Inst{7-5} = 0b100;
3958 let Inst{20-16} = Rss;
3959 let Inst{12-8} = Rtt;
3962 // A4_boundscheck_hi: Detect if a register is within bounds.
3963 let hasSideEffects = 0, isCodeGenOnly = 0 in
3964 def A4_boundscheck_hi: ALU64Inst <
3965 (outs PredRegs:$Pd),
3966 (ins DoubleRegs:$Rss, DoubleRegs:$Rtt),
3967 "$Pd = boundscheck($Rss, $Rtt):raw:hi"> {
3972 let IClass = 0b1101;
3974 let Inst{27-23} = 0b00100;
3976 let Inst{7-5} = 0b101;
3978 let Inst{20-16} = Rss;
3979 let Inst{12-8} = Rtt;
3982 let hasSideEffects = 0, isAsmParserOnly = 1 in
3983 def A4_boundscheck : MInst <
3984 (outs PredRegs:$Pd), (ins IntRegs:$Rs, DoubleRegs:$Rtt),
3985 "$Pd=boundscheck($Rs,$Rtt)">;
3987 // A4_tlbmatch: Detect if a VA/ASID matches a TLB entry.
3988 let isPredicateLate = 1, hasSideEffects = 0, isCodeGenOnly = 0 in
3989 def A4_tlbmatch : ALU64Inst<(outs PredRegs:$Pd),
3990 (ins DoubleRegs:$Rs, IntRegs:$Rt),
3991 "$Pd = tlbmatch($Rs, $Rt)",
3992 [], "", ALU64_tc_2early_SLOT23> {
3997 let IClass = 0b1101;
3998 let Inst{27-23} = 0b00100;
3999 let Inst{20-16} = Rs;
4001 let Inst{12-8} = Rt;
4002 let Inst{7-5} = 0b011;
4006 // We need custom lowering of ISD::PREFETCH into HexagonISD::DCFETCH
4007 // because the SDNode ISD::PREFETCH has properties MayLoad and MayStore.
4008 // We don't really want either one here.
4009 def SDTHexagonDCFETCH : SDTypeProfile<0, 2, [SDTCisPtrTy<0>,SDTCisInt<1>]>;
4010 def HexagonDCFETCH : SDNode<"HexagonISD::DCFETCH", SDTHexagonDCFETCH,
4013 // Use LD0Inst for dcfetch, but set "mayLoad" to 0 because this doesn't
4014 // really do a load.
4015 let hasSideEffects = 1, mayLoad = 0, isCodeGenOnly = 0 in
4016 def Y2_dcfetchbo : LD0Inst<(outs), (ins IntRegs:$Rs, u11_3Imm:$u11_3),
4017 "dcfetch($Rs + #$u11_3)",
4018 [(HexagonDCFETCH IntRegs:$Rs, u11_3ImmPred:$u11_3)],
4019 "", LD_tc_ld_SLOT0> {
4023 let IClass = 0b1001;
4024 let Inst{27-21} = 0b0100000;
4025 let Inst{20-16} = Rs;
4027 let Inst{10-0} = u11_3{13-3};
4030 //===----------------------------------------------------------------------===//
4031 // Compound instructions
4032 //===----------------------------------------------------------------------===//
4034 let isBranch = 1, hasSideEffects = 0, isExtentSigned = 1,
4035 isPredicated = 1, isPredicatedNew = 1, isExtendable = 1,
4036 opExtentBits = 11, opExtentAlign = 2, opExtendable = 1,
4037 isTerminator = 1, validSubTargets = HasV4SubT in
4038 class CJInst_tstbit_R0<string px, bit np, string tnt>
4039 : InstHexagon<(outs), (ins IntRegs:$Rs, brtarget:$r9_2),
4040 ""#px#" = tstbit($Rs, #0); if ("
4041 #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2",
4042 [], "", COMPOUND, TypeCOMPOUND> {
4047 let isPredicatedFalse = np;
4048 // tnt: Taken/Not Taken
4049 let isBrTaken = !if (!eq(tnt, "t"), "true", "false");
4050 let isTaken = !if (!eq(tnt, "t"), 1, 0);
4052 let IClass = 0b0001;
4053 let Inst{27-26} = 0b00;
4054 let Inst{25} = !if (!eq(px, "!p1"), 1,
4055 !if (!eq(px, "p1"), 1, 0));
4056 let Inst{24-23} = 0b11;
4058 let Inst{21-20} = r9_2{10-9};
4059 let Inst{19-16} = Rs;
4060 let Inst{13} = !if (!eq(tnt, "t"), 1, 0);
4061 let Inst{9-8} = 0b11;
4062 let Inst{7-1} = r9_2{8-2};
4065 let Defs = [PC, P0], Uses = [P0], isCodeGenOnly = 0 in {
4066 def J4_tstbit0_tp0_jump_nt : CJInst_tstbit_R0<"p0", 0, "nt">;
4067 def J4_tstbit0_tp0_jump_t : CJInst_tstbit_R0<"p0", 0, "t">;
4068 def J4_tstbit0_fp0_jump_nt : CJInst_tstbit_R0<"p0", 1, "nt">;
4069 def J4_tstbit0_fp0_jump_t : CJInst_tstbit_R0<"p0", 1, "t">;
4072 let Defs = [PC, P1], Uses = [P1], isCodeGenOnly = 0 in {
4073 def J4_tstbit0_tp1_jump_nt : CJInst_tstbit_R0<"p1", 0, "nt">;
4074 def J4_tstbit0_tp1_jump_t : CJInst_tstbit_R0<"p1", 0, "t">;
4075 def J4_tstbit0_fp1_jump_nt : CJInst_tstbit_R0<"p1", 1, "nt">;
4076 def J4_tstbit0_fp1_jump_t : CJInst_tstbit_R0<"p1", 1, "t">;
4080 let isBranch = 1, hasSideEffects = 0,
4081 isExtentSigned = 1, isPredicated = 1, isPredicatedNew = 1,
4082 isExtendable = 1, opExtentBits = 11, opExtentAlign = 2,
4083 opExtendable = 2, isTerminator = 1, validSubTargets = HasV4SubT in
4084 class CJInst_RR<string px, string op, bit np, string tnt>
4085 : InstHexagon<(outs), (ins IntRegs:$Rs, IntRegs:$Rt, brtarget:$r9_2),
4086 ""#px#" = cmp."#op#"($Rs, $Rt); if ("
4087 #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2",
4088 [], "", COMPOUND, TypeCOMPOUND> {
4094 let isPredicatedFalse = np;
4095 // tnt: Taken/Not Taken
4096 let isBrTaken = !if (!eq(tnt, "t"), "true", "false");
4097 let isTaken = !if (!eq(tnt, "t"), 1, 0);
4099 let IClass = 0b0001;
4100 let Inst{27-23} = !if (!eq(op, "eq"), 0b01000,
4101 !if (!eq(op, "gt"), 0b01001,
4102 !if (!eq(op, "gtu"), 0b01010, 0)));
4104 let Inst{21-20} = r9_2{10-9};
4105 let Inst{19-16} = Rs;
4106 let Inst{13} = !if (!eq(tnt, "t"), 1, 0);
4107 // px: Predicate reg 0/1
4108 let Inst{12} = !if (!eq(px, "!p1"), 1,
4109 !if (!eq(px, "p1"), 1, 0));
4110 let Inst{11-8} = Rt;
4111 let Inst{7-1} = r9_2{8-2};
4114 // P[10] taken/not taken.
4115 multiclass T_tnt_CJInst_RR<string op, bit np> {
4116 let Defs = [PC, P0], Uses = [P0] in {
4117 def NAME#p0_jump_nt : CJInst_RR<"p0", op, np, "nt">;
4118 def NAME#p0_jump_t : CJInst_RR<"p0", op, np, "t">;
4120 let Defs = [PC, P1], Uses = [P1] in {
4121 def NAME#p1_jump_nt : CJInst_RR<"p1", op, np, "nt">;
4122 def NAME#p1_jump_t : CJInst_RR<"p1", op, np, "t">;
4125 // Predicate / !Predicate
4126 multiclass T_pnp_CJInst_RR<string op>{
4127 defm J4_cmp#NAME#_t : T_tnt_CJInst_RR<op, 0>;
4128 defm J4_cmp#NAME#_f : T_tnt_CJInst_RR<op, 1>;
4130 // TypeCJ Instructions compare RR and jump
4131 let isCodeGenOnly = 0 in {
4132 defm eq : T_pnp_CJInst_RR<"eq">;
4133 defm gt : T_pnp_CJInst_RR<"gt">;
4134 defm gtu : T_pnp_CJInst_RR<"gtu">;
4137 let isBranch = 1, hasSideEffects = 0, isExtentSigned = 1,
4138 isPredicated = 1, isPredicatedNew = 1, isExtendable = 1, opExtentBits = 11,
4139 opExtentAlign = 2, opExtendable = 2, isTerminator = 1,
4140 validSubTargets = HasV4SubT in
4141 class CJInst_RU5<string px, string op, bit np, string tnt>
4142 : InstHexagon<(outs), (ins IntRegs:$Rs, u5Imm:$U5, brtarget:$r9_2),
4143 ""#px#" = cmp."#op#"($Rs, #$U5); if ("
4144 #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2",
4145 [], "", COMPOUND, TypeCOMPOUND> {
4151 let isPredicatedFalse = np;
4152 // tnt: Taken/Not Taken
4153 let isBrTaken = !if (!eq(tnt, "t"), "true", "false");
4154 let isTaken = !if (!eq(tnt, "t"), 1, 0);
4156 let IClass = 0b0001;
4157 let Inst{27-26} = 0b00;
4158 // px: Predicate reg 0/1
4159 let Inst{25} = !if (!eq(px, "!p1"), 1,
4160 !if (!eq(px, "p1"), 1, 0));
4161 let Inst{24-23} = !if (!eq(op, "eq"), 0b00,
4162 !if (!eq(op, "gt"), 0b01,
4163 !if (!eq(op, "gtu"), 0b10, 0)));
4165 let Inst{21-20} = r9_2{10-9};
4166 let Inst{19-16} = Rs;
4167 let Inst{13} = !if (!eq(tnt, "t"), 1, 0);
4168 let Inst{12-8} = U5;
4169 let Inst{7-1} = r9_2{8-2};
4171 // P[10] taken/not taken.
4172 multiclass T_tnt_CJInst_RU5<string op, bit np> {
4173 let Defs = [PC, P0], Uses = [P0] in {
4174 def NAME#p0_jump_nt : CJInst_RU5<"p0", op, np, "nt">;
4175 def NAME#p0_jump_t : CJInst_RU5<"p0", op, np, "t">;
4177 let Defs = [PC, P1], Uses = [P1] in {
4178 def NAME#p1_jump_nt : CJInst_RU5<"p1", op, np, "nt">;
4179 def NAME#p1_jump_t : CJInst_RU5<"p1", op, np, "t">;
4182 // Predicate / !Predicate
4183 multiclass T_pnp_CJInst_RU5<string op>{
4184 defm J4_cmp#NAME#i_t : T_tnt_CJInst_RU5<op, 0>;
4185 defm J4_cmp#NAME#i_f : T_tnt_CJInst_RU5<op, 1>;
4187 // TypeCJ Instructions compare RI and jump
4188 let isCodeGenOnly = 0 in {
4189 defm eq : T_pnp_CJInst_RU5<"eq">;
4190 defm gt : T_pnp_CJInst_RU5<"gt">;
4191 defm gtu : T_pnp_CJInst_RU5<"gtu">;
4194 let isBranch = 1, hasSideEffects = 0, isExtentSigned = 1,
4195 isPredicated = 1, isPredicatedFalse = 1, isPredicatedNew = 1,
4196 isExtendable = 1, opExtentBits = 11, opExtentAlign = 2, opExtendable = 1,
4197 isTerminator = 1, validSubTargets = HasV4SubT in
4198 class CJInst_Rn1<string px, string op, bit np, string tnt>
4199 : InstHexagon<(outs), (ins IntRegs:$Rs, brtarget:$r9_2),
4200 ""#px#" = cmp."#op#"($Rs,#-1); if ("
4201 #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2",
4202 [], "", COMPOUND, TypeCOMPOUND> {
4207 let isPredicatedFalse = np;
4208 // tnt: Taken/Not Taken
4209 let isBrTaken = !if (!eq(tnt, "t"), "true", "false");
4210 let isTaken = !if (!eq(tnt, "t"), 1, 0);
4212 let IClass = 0b0001;
4213 let Inst{27-26} = 0b00;
4214 let Inst{25} = !if (!eq(px, "!p1"), 1,
4215 !if (!eq(px, "p1"), 1, 0));
4217 let Inst{24-23} = 0b11;
4219 let Inst{21-20} = r9_2{10-9};
4220 let Inst{19-16} = Rs;
4221 let Inst{13} = !if (!eq(tnt, "t"), 1, 0);
4222 let Inst{9-8} = !if (!eq(op, "eq"), 0b00,
4223 !if (!eq(op, "gt"), 0b01, 0));
4224 let Inst{7-1} = r9_2{8-2};
4227 // P[10] taken/not taken.
4228 multiclass T_tnt_CJInst_Rn1<string op, bit np> {
4229 let Defs = [PC, P0], Uses = [P0] in {
4230 def NAME#p0_jump_nt : CJInst_Rn1<"p0", op, np, "nt">;
4231 def NAME#p0_jump_t : CJInst_Rn1<"p0", op, np, "t">;
4233 let Defs = [PC, P1], Uses = [P1] in {
4234 def NAME#p1_jump_nt : CJInst_Rn1<"p1", op, np, "nt">;
4235 def NAME#p1_jump_t : CJInst_Rn1<"p1", op, np, "t">;
4238 // Predicate / !Predicate
4239 multiclass T_pnp_CJInst_Rn1<string op>{
4240 defm J4_cmp#NAME#n1_t : T_tnt_CJInst_Rn1<op, 0>;
4241 defm J4_cmp#NAME#n1_f : T_tnt_CJInst_Rn1<op, 1>;
4243 // TypeCJ Instructions compare -1 and jump
4244 let isCodeGenOnly = 0 in {
4245 defm eq : T_pnp_CJInst_Rn1<"eq">;
4246 defm gt : T_pnp_CJInst_Rn1<"gt">;
4249 // J4_jumpseti: Direct unconditional jump and set register to immediate.
4250 let Defs = [PC], isBranch = 1, hasSideEffects = 0, hasNewValue = 1,
4251 isExtentSigned = 1, opNewValue = 0, isExtendable = 1, opExtentBits = 11,
4252 opExtentAlign = 2, opExtendable = 2, validSubTargets = HasV4SubT,
4253 isCodeGenOnly = 0 in
4254 def J4_jumpseti: CJInst <
4256 (ins u6Imm:$U6, brtarget:$r9_2),
4257 "$Rd = #$U6 ; jump $r9_2"> {
4262 let IClass = 0b0001;
4263 let Inst{27-24} = 0b0110;
4264 let Inst{21-20} = r9_2{10-9};
4265 let Inst{19-16} = Rd;
4266 let Inst{13-8} = U6;
4267 let Inst{7-1} = r9_2{8-2};
4270 // J4_jumpsetr: Direct unconditional jump and transfer register.
4271 let Defs = [PC], isBranch = 1, hasSideEffects = 0, hasNewValue = 1,
4272 isExtentSigned = 1, opNewValue = 0, isExtendable = 1, opExtentBits = 11,
4273 opExtentAlign = 2, opExtendable = 2, validSubTargets = HasV4SubT,
4274 isCodeGenOnly = 0 in
4275 def J4_jumpsetr: CJInst <
4277 (ins IntRegs:$Rs, brtarget:$r9_2),
4278 "$Rd = $Rs ; jump $r9_2"> {
4283 let IClass = 0b0001;
4284 let Inst{27-24} = 0b0111;
4285 let Inst{21-20} = r9_2{10-9};
4286 let Inst{11-8} = Rd;
4287 let Inst{19-16} = Rs;
4288 let Inst{7-1} = r9_2{8-2};