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 let hasSideEffects = 0 in
15 class T_Immext<Operand ImmType>
16 : EXTENDERInst<(outs), (ins ImmType:$imm),
17 "immext(#$imm)", []> {
21 let Inst{27-16} = imm{31-20};
22 let Inst{13-0} = imm{19-6};
25 def A4_ext : T_Immext<u26_6Imm>;
26 let isCodeGenOnly = 1 in {
28 def A4_ext_b : T_Immext<brtarget>;
30 def A4_ext_c : T_Immext<calltarget>;
31 def A4_ext_g : T_Immext<globaladdress>;
34 def BITPOS32 : SDNodeXForm<imm, [{
35 // Return the bit position we will set [0-31].
37 int32_t imm = N->getSExtValue();
38 return XformMskToBitPosU5Imm(imm);
41 // Fold (add (CONST32 tglobaladdr:$addr) <offset>) into a global address.
42 def FoldGlobalAddr : ComplexPattern<i32, 1, "foldGlobalAddress", [], []>;
44 // Fold (add (CONST32_GP tglobaladdr:$addr) <offset>) into a global address.
45 def FoldGlobalAddrGP : ComplexPattern<i32, 1, "foldGlobalAddressGP", [], []>;
47 def NumUsesBelowThresCONST32 : PatFrag<(ops node:$addr),
48 (HexagonCONST32 node:$addr), [{
49 return hasNumUsesBelowThresGA(N->getOperand(0).getNode());
52 // Hexagon V4 Architecture spec defines 8 instruction classes:
53 // LD ST ALU32 XTYPE J JR MEMOP NV CR SYSTEM(system is not implemented in the
57 // ========================================
58 // Loads (8/16/32/64 bit)
62 // ========================================
63 // Stores (8/16/32/64 bit)
66 // ALU32 Instructions:
67 // ========================================
68 // Arithmetic / Logical (32 bit)
71 // XTYPE Instructions (32/64 bit):
72 // ========================================
73 // Arithmetic, Logical, Bit Manipulation
74 // Multiply (Integer, Fractional, Complex)
75 // Permute / Vector Permute Operations
76 // Predicate Operations
77 // Shift / Shift with Add/Sub/Logical
79 // Vector Halfword (ALU, Shift, Multiply)
80 // Vector Word (ALU, Shift)
83 // ========================================
84 // Jump/Call PC-relative
87 // ========================================
90 // MEMOP Instructions:
91 // ========================================
92 // Operation on memory (8/16/32 bit)
95 // ========================================
100 // ========================================
101 // Control-Register Transfers
102 // Hardware Loop Setup
103 // Predicate Logicals & Reductions
105 // SYSTEM Instructions (not implemented in the compiler):
106 // ========================================
112 //===----------------------------------------------------------------------===//
114 //===----------------------------------------------------------------------===//
116 class T_ALU32_3op_not<string mnemonic, bits<3> MajOp, bits<3> MinOp,
118 : T_ALU32_3op<mnemonic, MajOp, MinOp, OpsRev, 0> {
119 let AsmString = "$Rd = "#mnemonic#"($Rs, ~$Rt)";
122 let BaseOpcode = "andn_rr", CextOpcode = "andn", isCodeGenOnly = 0 in
123 def A4_andn : T_ALU32_3op_not<"and", 0b001, 0b100, 1>;
124 let BaseOpcode = "orn_rr", CextOpcode = "orn", isCodeGenOnly = 0 in
125 def A4_orn : T_ALU32_3op_not<"or", 0b001, 0b101, 1>;
127 let CextOpcode = "rcmp.eq", isCodeGenOnly = 0 in
128 def A4_rcmpeq : T_ALU32_3op<"cmp.eq", 0b011, 0b010, 0, 1>;
129 let CextOpcode = "!rcmp.eq", isCodeGenOnly = 0 in
130 def A4_rcmpneq : T_ALU32_3op<"!cmp.eq", 0b011, 0b011, 0, 1>;
132 let isCodeGenOnly = 0 in {
133 def C4_cmpneq : T_ALU32_3op_cmp<"!cmp.eq", 0b00, 1, 1>;
134 def C4_cmplte : T_ALU32_3op_cmp<"!cmp.gt", 0b10, 1, 0>;
135 def C4_cmplteu : T_ALU32_3op_cmp<"!cmp.gtu", 0b11, 1, 0>;
138 // Pats for instruction selection.
140 // A class to embed the usual comparison patfrags within a zext to i32.
141 // The seteq/setne frags use "lhs" and "rhs" as operands, so use the same
142 // names, or else the frag's "body" won't match the operands.
143 class CmpInReg<PatFrag Op>
144 : PatFrag<(ops node:$lhs, node:$rhs),(i32 (zext (i1 Op.Fragment)))>;
146 def: T_cmp32_rr_pat<A4_rcmpeq, CmpInReg<seteq>, i32>;
147 def: T_cmp32_rr_pat<A4_rcmpneq, CmpInReg<setne>, i32>;
149 def: T_cmp32_rr_pat<C4_cmpneq, setne, i1>;
151 class T_CMP_rrbh<string mnemonic, bits<3> MinOp, bit IsComm>
152 : SInst<(outs PredRegs:$Pd), (ins IntRegs:$Rs, IntRegs:$Rt),
153 "$Pd = "#mnemonic#"($Rs, $Rt)", [], "", S_3op_tc_2early_SLOT23>,
155 let validSubTargets = HasV4SubT;
156 let InputType = "reg";
157 let CextOpcode = mnemonic;
159 let isCommutable = IsComm;
160 let hasSideEffects = 0;
167 let Inst{27-21} = 0b0111110;
168 let Inst{20-16} = Rs;
170 let Inst{7-5} = MinOp;
174 let isCodeGenOnly = 0 in {
175 def A4_cmpbeq : T_CMP_rrbh<"cmpb.eq", 0b110, 1>;
176 def A4_cmpbgt : T_CMP_rrbh<"cmpb.gt", 0b010, 0>;
177 def A4_cmpbgtu : T_CMP_rrbh<"cmpb.gtu", 0b111, 0>;
178 def A4_cmpheq : T_CMP_rrbh<"cmph.eq", 0b011, 1>;
179 def A4_cmphgt : T_CMP_rrbh<"cmph.gt", 0b100, 0>;
180 def A4_cmphgtu : T_CMP_rrbh<"cmph.gtu", 0b101, 0>;
183 class T_CMP_ribh<string mnemonic, bits<2> MajOp, bit IsHalf, bit IsComm,
184 Operand ImmType, bit IsImmExt, bit IsImmSigned, int ImmBits>
185 : ALU64Inst<(outs PredRegs:$Pd), (ins IntRegs:$Rs, ImmType:$Imm),
186 "$Pd = "#mnemonic#"($Rs, #$Imm)", [], "", ALU64_tc_2early_SLOT23>,
188 let validSubTargets = HasV4SubT;
189 let InputType = "imm";
190 let CextOpcode = mnemonic;
192 let isCommutable = IsComm;
193 let hasSideEffects = 0;
194 let isExtendable = IsImmExt;
195 let opExtendable = !if (IsImmExt, 2, 0);
196 let isExtentSigned = IsImmSigned;
197 let opExtentBits = ImmBits;
204 let Inst{27-24} = 0b1101;
205 let Inst{22-21} = MajOp;
206 let Inst{20-16} = Rs;
207 let Inst{12-5} = Imm;
209 let Inst{3} = IsHalf;
213 let isCodeGenOnly = 0 in {
214 def A4_cmpbeqi : T_CMP_ribh<"cmpb.eq", 0b00, 0, 1, u8Imm, 0, 0, 8>;
215 def A4_cmpbgti : T_CMP_ribh<"cmpb.gt", 0b01, 0, 0, s8Imm, 0, 1, 8>;
216 def A4_cmpbgtui : T_CMP_ribh<"cmpb.gtu", 0b10, 0, 0, u7Ext, 1, 0, 7>;
217 def A4_cmpheqi : T_CMP_ribh<"cmph.eq", 0b00, 1, 1, s8Ext, 1, 1, 8>;
218 def A4_cmphgti : T_CMP_ribh<"cmph.gt", 0b01, 1, 0, s8Ext, 1, 1, 8>;
219 def A4_cmphgtui : T_CMP_ribh<"cmph.gtu", 0b10, 1, 0, u7Ext, 1, 0, 7>;
221 class T_RCMP_EQ_ri<string mnemonic, bit IsNeg>
222 : ALU32_ri<(outs IntRegs:$Rd), (ins IntRegs:$Rs, s8Ext:$s8),
223 "$Rd = "#mnemonic#"($Rs, #$s8)", [], "", ALU32_2op_tc_1_SLOT0123>,
225 let validSubTargets = HasV4SubT;
226 let InputType = "imm";
227 let CextOpcode = !if (IsNeg, "!rcmp.eq", "rcmp.eq");
228 let isExtendable = 1;
229 let opExtendable = 2;
230 let isExtentSigned = 1;
231 let opExtentBits = 8;
239 let Inst{27-24} = 0b0011;
241 let Inst{21} = IsNeg;
242 let Inst{20-16} = Rs;
248 let isCodeGenOnly = 0 in {
249 def A4_rcmpeqi : T_RCMP_EQ_ri<"cmp.eq", 0>;
250 def A4_rcmpneqi : T_RCMP_EQ_ri<"!cmp.eq", 1>;
253 def: Pat<(i32 (zext (i1 (seteq (i32 IntRegs:$Rs), s8ExtPred:$s8)))),
254 (A4_rcmpeqi IntRegs:$Rs, s8ExtPred:$s8)>;
255 def: Pat<(i32 (zext (i1 (setne (i32 IntRegs:$Rs), s8ExtPred:$s8)))),
256 (A4_rcmpneqi IntRegs:$Rs, s8ExtPred:$s8)>;
258 // Preserve the S2_tstbit_r generation
259 def: Pat<(i32 (zext (i1 (setne (i32 (and (i32 (shl 1, (i32 IntRegs:$src2))),
260 (i32 IntRegs:$src1))), 0)))),
261 (C2_muxii (S2_tstbit_r IntRegs:$src1, IntRegs:$src2), 1, 0)>;
264 //===----------------------------------------------------------------------===//
266 //===----------------------------------------------------------------------===//
269 //===----------------------------------------------------------------------===//
271 //===----------------------------------------------------------------------===//
273 // Combine a word and an immediate into a register pair.
274 let hasSideEffects = 0, isExtentSigned = 1, isExtendable = 1,
276 class T_Combine1 <bits<2> MajOp, dag ins, string AsmStr>
277 : ALU32Inst <(outs DoubleRegs:$Rdd), ins, AsmStr> {
283 let Inst{27-24} = 0b0011;
284 let Inst{22-21} = MajOp;
285 let Inst{20-16} = Rs;
291 let opExtendable = 2, isCodeGenOnly = 0 in
292 def A4_combineri : T_Combine1<0b00, (ins IntRegs:$Rs, s8Ext:$s8),
293 "$Rdd = combine($Rs, #$s8)">;
295 let opExtendable = 1, isCodeGenOnly = 0 in
296 def A4_combineir : T_Combine1<0b01, (ins s8Ext:$s8, IntRegs:$Rs),
297 "$Rdd = combine(#$s8, $Rs)">;
299 def HexagonWrapperCombineRI_V4 :
300 SDNode<"HexagonISD::WrapperCombineRI_V4", SDTHexagonI64I32I32>;
301 def HexagonWrapperCombineIR_V4 :
302 SDNode<"HexagonISD::WrapperCombineIR_V4", SDTHexagonI64I32I32>;
304 def : Pat <(HexagonWrapperCombineRI_V4 IntRegs:$r, s8ExtPred:$i),
305 (A4_combineri IntRegs:$r, s8ExtPred:$i)>,
308 def : Pat <(HexagonWrapperCombineIR_V4 s8ExtPred:$i, IntRegs:$r),
309 (A4_combineir s8ExtPred:$i, IntRegs:$r)>,
312 // A4_combineii: Set two small immediates.
313 let hasSideEffects = 0, isExtendable = 1, opExtentBits = 6, opExtendable = 2 in
314 def A4_combineii: ALU32Inst<(outs DoubleRegs:$Rdd), (ins s8Imm:$s8, u6Ext:$U6),
315 "$Rdd = combine(#$s8, #$U6)"> {
321 let Inst{27-23} = 0b11001;
322 let Inst{20-16} = U6{5-1};
323 let Inst{13} = U6{0};
328 //===----------------------------------------------------------------------===//
330 //===----------------------------------------------------------------------===//
332 //===----------------------------------------------------------------------===//
334 //===----------------------------------------------------------------------===//
335 //===----------------------------------------------------------------------===//
336 // Template class for load instructions with Absolute set addressing mode.
337 //===----------------------------------------------------------------------===//
338 let isExtended = 1, opExtendable = 2, opExtentBits = 6, addrMode = AbsoluteSet,
339 hasSideEffects = 0 in
340 class T_LD_abs_set<string mnemonic, RegisterClass RC, bits<4>MajOp>:
341 LDInst<(outs RC:$dst1, IntRegs:$dst2),
343 "$dst1 = "#mnemonic#"($dst2 = #$addr)",
351 let Inst{27-25} = 0b101;
352 let Inst{24-21} = MajOp;
353 let Inst{13-12} = 0b01;
354 let Inst{4-0} = dst1;
355 let Inst{20-16} = dst2;
356 let Inst{11-8} = addr{5-2};
357 let Inst{6-5} = addr{1-0};
360 let accessSize = ByteAccess, hasNewValue = 1, isCodeGenOnly = 0 in {
361 def L4_loadrb_ap : T_LD_abs_set <"memb", IntRegs, 0b1000>;
362 def L4_loadrub_ap : T_LD_abs_set <"memub", IntRegs, 0b1001>;
365 let accessSize = HalfWordAccess, hasNewValue = 1, isCodeGenOnly = 0 in {
366 def L4_loadrh_ap : T_LD_abs_set <"memh", IntRegs, 0b1010>;
367 def L4_loadruh_ap : T_LD_abs_set <"memuh", IntRegs, 0b1011>;
370 let accessSize = WordAccess, hasNewValue = 1, isCodeGenOnly = 0 in
371 def L4_loadri_ap : T_LD_abs_set <"memw", IntRegs, 0b1100>;
373 let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
374 def L4_loadrd_ap : T_LD_abs_set <"memd", DoubleRegs, 0b1110>;
375 // Load - Indirect with long offset
376 let InputType = "imm", addrMode = BaseLongOffset, isExtended = 1,
377 opExtentBits = 6, opExtendable = 3 in
378 class T_LoadAbsReg <string mnemonic, string CextOp, RegisterClass RC,
380 : LDInst <(outs RC:$dst), (ins IntRegs:$src1, u2Imm:$src2, u6Ext:$src3),
381 "$dst = "#mnemonic#"($src1<<#$src2 + #$src3)",
387 let CextOpcode = CextOp;
388 let hasNewValue = !if (!eq(!cast<string>(RC), "DoubleRegs"), 0, 1);
391 let Inst{27-25} = 0b110;
392 let Inst{24-21} = MajOp;
393 let Inst{20-16} = src1;
394 let Inst{13} = src2{1};
396 let Inst{11-8} = src3{5-2};
397 let Inst{7} = src2{0};
398 let Inst{6-5} = src3{1-0};
402 let accessSize = ByteAccess, isCodeGenOnly = 0 in {
403 def L4_loadrb_ur : T_LoadAbsReg<"memb", "LDrib", IntRegs, 0b1000>;
404 def L4_loadrub_ur : T_LoadAbsReg<"memub", "LDriub", IntRegs, 0b1001>;
405 def L4_loadalignb_ur : T_LoadAbsReg<"memb_fifo", "LDrib_fifo",
409 let accessSize = HalfWordAccess, isCodeGenOnly = 0 in {
410 def L4_loadrh_ur : T_LoadAbsReg<"memh", "LDrih", IntRegs, 0b1010>;
411 def L4_loadruh_ur : T_LoadAbsReg<"memuh", "LDriuh", IntRegs, 0b1011>;
412 def L4_loadbsw2_ur : T_LoadAbsReg<"membh", "LDribh2", IntRegs, 0b0001>;
413 def L4_loadbzw2_ur : T_LoadAbsReg<"memubh", "LDriubh2", IntRegs, 0b0011>;
414 def L4_loadalignh_ur : T_LoadAbsReg<"memh_fifo", "LDrih_fifo",
418 let accessSize = WordAccess, isCodeGenOnly = 0 in {
419 def L4_loadri_ur : T_LoadAbsReg<"memw", "LDriw", IntRegs, 0b1100>;
420 def L4_loadbsw4_ur : T_LoadAbsReg<"membh", "LDribh4", DoubleRegs, 0b0111>;
421 def L4_loadbzw4_ur : T_LoadAbsReg<"memubh", "LDriubh4", DoubleRegs, 0b0101>;
424 let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
425 def L4_loadrd_ur : T_LoadAbsReg<"memd", "LDrid", DoubleRegs, 0b1110>;
428 multiclass T_LoadAbsReg_Pat <PatFrag ldOp, InstHexagon MI, ValueType VT = i32> {
429 def : Pat <(VT (ldOp (add (shl IntRegs:$src1, u2ImmPred:$src2),
430 (HexagonCONST32 tglobaladdr:$src3)))),
431 (MI IntRegs:$src1, u2ImmPred:$src2, tglobaladdr:$src3)>;
433 def : Pat <(VT (ldOp (add IntRegs:$src1,
434 (HexagonCONST32 tglobaladdr:$src2)))),
435 (MI IntRegs:$src1, 0, tglobaladdr:$src2)>;
438 let AddedComplexity = 60 in {
439 defm : T_LoadAbsReg_Pat <sextloadi8, L4_loadrb_ur>;
440 defm : T_LoadAbsReg_Pat <zextloadi8, L4_loadrub_ur>;
441 defm : T_LoadAbsReg_Pat <extloadi8, L4_loadrub_ur>;
443 defm : T_LoadAbsReg_Pat <sextloadi16, L4_loadrh_ur>;
444 defm : T_LoadAbsReg_Pat <zextloadi16, L4_loadruh_ur>;
445 defm : T_LoadAbsReg_Pat <extloadi16, L4_loadruh_ur>;
447 defm : T_LoadAbsReg_Pat <load, L4_loadri_ur>;
448 defm : T_LoadAbsReg_Pat <load, L4_loadrd_ur, i64>;
451 //===----------------------------------------------------------------------===//
452 // Template classes for the non-predicated load instructions with
453 // base + register offset addressing mode
454 //===----------------------------------------------------------------------===//
455 class T_load_rr <string mnemonic, RegisterClass RC, bits<3> MajOp>:
456 LDInst<(outs RC:$dst), (ins IntRegs:$src1, IntRegs:$src2, u2Imm:$u2),
457 "$dst = "#mnemonic#"($src1 + $src2<<#$u2)",
458 [], "", V4LDST_tc_ld_SLOT01>, ImmRegShl, AddrModeRel {
466 let Inst{27-24} = 0b1010;
467 let Inst{23-21} = MajOp;
468 let Inst{20-16} = src1;
469 let Inst{12-8} = src2;
470 let Inst{13} = u2{1};
475 //===----------------------------------------------------------------------===//
476 // Template classes for the predicated load instructions with
477 // base + register offset addressing mode
478 //===----------------------------------------------------------------------===//
479 let isPredicated = 1 in
480 class T_pload_rr <string mnemonic, RegisterClass RC, bits<3> MajOp,
481 bit isNot, bit isPredNew>:
482 LDInst <(outs RC:$dst),
483 (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3, u2Imm:$u2),
484 !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
485 ") ")#"$dst = "#mnemonic#"($src2+$src3<<#$u2)",
486 [], "", V4LDST_tc_ld_SLOT01>, AddrModeRel {
493 let isPredicatedFalse = isNot;
494 let isPredicatedNew = isPredNew;
498 let Inst{27-26} = 0b00;
499 let Inst{25} = isPredNew;
500 let Inst{24} = isNot;
501 let Inst{23-21} = MajOp;
502 let Inst{20-16} = src2;
503 let Inst{12-8} = src3;
504 let Inst{13} = u2{1};
506 let Inst{6-5} = src1;
510 //===----------------------------------------------------------------------===//
511 // multiclass for load instructions with base + register offset
513 //===----------------------------------------------------------------------===//
514 let hasSideEffects = 0, addrMode = BaseRegOffset in
515 multiclass ld_idxd_shl <string mnemonic, string CextOp, RegisterClass RC,
517 let CextOpcode = CextOp, BaseOpcode = CextOp#_indexed_shl,
518 InputType = "reg" in {
519 let isPredicable = 1 in
520 def L4_#NAME#_rr : T_load_rr <mnemonic, RC, MajOp>;
523 def L4_p#NAME#t_rr : T_pload_rr <mnemonic, RC, MajOp, 0, 0>;
524 def L4_p#NAME#f_rr : T_pload_rr <mnemonic, RC, MajOp, 1, 0>;
527 def L4_p#NAME#tnew_rr : T_pload_rr <mnemonic, RC, MajOp, 0, 1>;
528 def L4_p#NAME#fnew_rr : T_pload_rr <mnemonic, RC, MajOp, 1, 1>;
532 let hasNewValue = 1, accessSize = ByteAccess, isCodeGenOnly = 0 in {
533 defm loadrb : ld_idxd_shl<"memb", "LDrib", IntRegs, 0b000>;
534 defm loadrub : ld_idxd_shl<"memub", "LDriub", IntRegs, 0b001>;
537 let hasNewValue = 1, accessSize = HalfWordAccess, isCodeGenOnly = 0 in {
538 defm loadrh : ld_idxd_shl<"memh", "LDrih", IntRegs, 0b010>;
539 defm loadruh : ld_idxd_shl<"memuh", "LDriuh", IntRegs, 0b011>;
542 let hasNewValue = 1, accessSize = WordAccess, isCodeGenOnly = 0 in
543 defm loadri : ld_idxd_shl<"memw", "LDriw", IntRegs, 0b100>;
545 let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
546 defm loadrd : ld_idxd_shl<"memd", "LDrid", DoubleRegs, 0b110>;
548 // 'def pats' for load instructions with base + register offset and non-zero
549 // immediate value. Immediate value is used to left-shift the second
551 let AddedComplexity = 40 in {
552 def : Pat <(i32 (sextloadi8 (add IntRegs:$src1,
553 (shl IntRegs:$src2, u2ImmPred:$offset)))),
554 (L4_loadrb_rr IntRegs:$src1,
555 IntRegs:$src2, u2ImmPred:$offset)>,
558 def : Pat <(i32 (zextloadi8 (add IntRegs:$src1,
559 (shl IntRegs:$src2, u2ImmPred:$offset)))),
560 (L4_loadrub_rr IntRegs:$src1,
561 IntRegs:$src2, u2ImmPred:$offset)>,
564 def : Pat <(i32 (extloadi8 (add IntRegs:$src1,
565 (shl IntRegs:$src2, u2ImmPred:$offset)))),
566 (L4_loadrub_rr IntRegs:$src1,
567 IntRegs:$src2, u2ImmPred:$offset)>,
570 def : Pat <(i32 (sextloadi16 (add IntRegs:$src1,
571 (shl IntRegs:$src2, u2ImmPred:$offset)))),
572 (L4_loadrh_rr IntRegs:$src1,
573 IntRegs:$src2, u2ImmPred:$offset)>,
576 def : Pat <(i32 (zextloadi16 (add IntRegs:$src1,
577 (shl IntRegs:$src2, u2ImmPred:$offset)))),
578 (L4_loadruh_rr IntRegs:$src1,
579 IntRegs:$src2, u2ImmPred:$offset)>,
582 def : Pat <(i32 (extloadi16 (add IntRegs:$src1,
583 (shl IntRegs:$src2, u2ImmPred:$offset)))),
584 (L4_loadruh_rr IntRegs:$src1,
585 IntRegs:$src2, u2ImmPred:$offset)>,
588 def : Pat <(i32 (load (add IntRegs:$src1,
589 (shl IntRegs:$src2, u2ImmPred:$offset)))),
590 (L4_loadri_rr IntRegs:$src1,
591 IntRegs:$src2, u2ImmPred:$offset)>,
594 def : Pat <(i64 (load (add IntRegs:$src1,
595 (shl IntRegs:$src2, u2ImmPred:$offset)))),
596 (L4_loadrd_rr IntRegs:$src1,
597 IntRegs:$src2, u2ImmPred:$offset)>,
601 // 'def pats' for load instruction base + register offset and
602 // zero immediate value.
603 class Loadxs_simple_pat<PatFrag Load, ValueType VT, InstHexagon MI>
604 : Pat<(VT (Load (add (i32 IntRegs:$Rs), (i32 IntRegs:$Rt)))),
605 (VT (MI IntRegs:$Rs, IntRegs:$Rt, 0))>;
607 let AddedComplexity = 20 in {
608 def: Loadxs_simple_pat<extloadi8, i32, L4_loadrub_rr>;
609 def: Loadxs_simple_pat<zextloadi8, i32, L4_loadrub_rr>;
610 def: Loadxs_simple_pat<sextloadi8, i32, L4_loadrb_rr>;
611 def: Loadxs_simple_pat<extloadi16, i32, L4_loadruh_rr>;
612 def: Loadxs_simple_pat<zextloadi16, i32, L4_loadruh_rr>;
613 def: Loadxs_simple_pat<sextloadi16, i32, L4_loadrh_rr>;
614 def: Loadxs_simple_pat<load, i32, L4_loadri_rr>;
615 def: Loadxs_simple_pat<load, i64, L4_loadrd_rr>;
619 def : Pat <(i64 (zext (i1 PredRegs:$src1))),
620 (i64 (A4_combineir 0, (C2_muxii (i1 PredRegs:$src1), 1, 0)))>,
624 def : Pat <(i64 (zext (i32 IntRegs:$src1))),
625 (i64 (A4_combineir 0, (i32 IntRegs:$src1)))>,
628 def: Pat <(i64 (zextloadi8 ADDRriS11_0:$src1)),
629 (i64 (A4_combineir 0, (L2_loadrub_io AddrFI:$src1, 0)))>,
632 let AddedComplexity = 20 in
633 def: Pat <(i64 (zextloadi8 (add (i32 IntRegs:$src1),
634 s11_0ExtPred:$offset))),
635 (i64 (A4_combineir 0, (L2_loadrub_io IntRegs:$src1,
636 s11_0ExtPred:$offset)))>,
640 def: Pat <(i64 (zextloadi1 ADDRriS11_0:$src1)),
641 (i64 (A4_combineir 0, (L2_loadrub_io AddrFI:$src1, 0)))>,
644 let AddedComplexity = 20 in
645 def: Pat <(i64 (zextloadi1 (add (i32 IntRegs:$src1),
646 s11_0ExtPred:$offset))),
647 (i64 (A4_combineir 0, (L2_loadrub_io IntRegs:$src1,
648 s11_0ExtPred:$offset)))>,
652 def: Pat <(i64 (zextloadi16 ADDRriS11_1:$src1)),
653 (i64 (A4_combineir 0, (L2_loadruh_io AddrFI:$src1, 0)))>,
656 let AddedComplexity = 20 in
657 def: Pat <(i64 (zextloadi16 (add (i32 IntRegs:$src1),
658 s11_1ExtPred:$offset))),
659 (i64 (A4_combineir 0, (L2_loadruh_io IntRegs:$src1,
660 s11_1ExtPred:$offset)))>,
664 def: Pat <(i64 (extloadi16 ADDRriS11_2:$src1)),
665 (i64 (A4_combineir 0, (L2_loadrh_io AddrFI:$src1, 0)))>,
668 let AddedComplexity = 20 in
669 def: Pat <(i64 (extloadi16 (add (i32 IntRegs:$src1),
670 s11_1ExtPred:$offset))),
671 (i64 (A4_combineir 0, (L2_loadrh_io IntRegs:$src1,
672 s11_1ExtPred:$offset)))>,
676 def: Pat <(i64 (zextloadi32 ADDRriS11_2:$src1)),
677 (i64 (A4_combineir 0, (L2_loadri_io AddrFI:$src1, 0)))>,
680 let AddedComplexity = 100 in
681 def: Pat <(i64 (zextloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))),
682 (i64 (A4_combineir 0, (L2_loadri_io IntRegs:$src1,
683 s11_2ExtPred:$offset)))>,
687 def: Pat <(i64 (extloadi32 ADDRriS11_2:$src1)),
688 (i64 (A4_combineir 0, (L2_loadri_io AddrFI:$src1, 0)))>,
691 let AddedComplexity = 100 in
692 def: Pat <(i64 (extloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))),
693 (i64 (A4_combineir 0, (L2_loadri_io IntRegs:$src1,
694 s11_2ExtPred:$offset)))>,
699 //===----------------------------------------------------------------------===//
701 //===----------------------------------------------------------------------===//
703 //===----------------------------------------------------------------------===//
705 //===----------------------------------------------------------------------===//
707 //===----------------------------------------------------------------------===//
708 // Template class for store instructions with Absolute set addressing mode.
709 //===----------------------------------------------------------------------===//
710 let isExtended = 1, opExtendable = 1, opExtentBits = 6,
711 addrMode = AbsoluteSet, isNVStorable = 1 in
712 class T_ST_absset <string mnemonic, string BaseOp, RegisterClass RC,
713 bits<3> MajOp, MemAccessSize AccessSz, bit isHalf = 0>
714 : STInst<(outs IntRegs:$dst),
715 (ins u6Ext:$addr, RC:$src),
716 mnemonic#"($dst = #$addr) = $src"#!if(isHalf, ".h","")>, NewValueRel {
720 let accessSize = AccessSz;
721 let BaseOpcode = BaseOp#"_AbsSet";
725 let Inst{27-24} = 0b1011;
726 let Inst{23-21} = MajOp;
727 let Inst{20-16} = dst;
729 let Inst{12-8} = src;
731 let Inst{5-0} = addr;
734 def S4_storerb_ap : T_ST_absset <"memb", "STrib", IntRegs, 0b000, ByteAccess>;
735 def S4_storerh_ap : T_ST_absset <"memh", "STrih", IntRegs, 0b010,
737 def S4_storeri_ap : T_ST_absset <"memw", "STriw", IntRegs, 0b100, WordAccess>;
739 let isNVStorable = 0 in {
740 def S4_storerf_ap : T_ST_absset <"memh", "STrif", IntRegs,
741 0b011, HalfWordAccess, 1>;
742 def S4_storerd_ap : T_ST_absset <"memd", "STrid", DoubleRegs,
743 0b110, DoubleWordAccess>;
746 let opExtendable = 1, isNewValue = 1, isNVStore = 1, opNewValue = 2,
747 isExtended = 1, opExtentBits= 6 in
748 class T_ST_absset_nv <string mnemonic, string BaseOp, bits<2> MajOp,
749 MemAccessSize AccessSz >
750 : NVInst <(outs IntRegs:$dst),
751 (ins u6Ext:$addr, IntRegs:$src),
752 mnemonic#"($dst = #$addr) = $src.new">, NewValueRel {
756 let accessSize = AccessSz;
757 let BaseOpcode = BaseOp#"_AbsSet";
761 let Inst{27-21} = 0b1011101;
762 let Inst{20-16} = dst;
763 let Inst{13-11} = 0b000;
764 let Inst{12-11} = MajOp;
765 let Inst{10-8} = src;
767 let Inst{5-0} = addr;
770 let mayStore = 1, addrMode = AbsoluteSet in {
771 def S4_storerbnew_ap : T_ST_absset_nv <"memb", "STrib", 0b00, ByteAccess>;
772 def S4_storerhnew_ap : T_ST_absset_nv <"memh", "STrih", 0b01, HalfWordAccess>;
773 def S4_storerinew_ap : T_ST_absset_nv <"memw", "STriw", 0b10, WordAccess>;
776 let isExtended = 1, opExtendable = 2, opExtentBits = 6, InputType = "imm",
777 addrMode = BaseLongOffset, AddedComplexity = 40 in
778 class T_StoreAbsReg <string mnemonic, string CextOp, RegisterClass RC,
779 bits<3> MajOp, MemAccessSize AccessSz, bit isHalf = 0>
781 (ins IntRegs:$src1, u2Imm:$src2, u6Ext:$src3, RC:$src4),
782 mnemonic#"($src1<<#$src2 + #$src3) = $src4"#!if(isHalf, ".h",""),
783 []>, ImmRegShl, NewValueRel {
790 let accessSize = AccessSz;
791 let CextOpcode = CextOp;
792 let BaseOpcode = CextOp#"_shl";
795 let Inst{27-24} =0b1101;
796 let Inst{23-21} = MajOp;
797 let Inst{20-16} = src1;
798 let Inst{13} = src2{1};
799 let Inst{12-8} = src4;
801 let Inst{6} = src2{0};
802 let Inst{5-0} = src3;
805 def S4_storerb_ur : T_StoreAbsReg <"memb", "STrib", IntRegs, 0b000, ByteAccess>;
806 def S4_storerh_ur : T_StoreAbsReg <"memh", "STrih", IntRegs, 0b010,
808 def S4_storerf_ur : T_StoreAbsReg <"memh", "STrif", IntRegs, 0b011,
810 def S4_storeri_ur : T_StoreAbsReg <"memw", "STriw", IntRegs, 0b100, WordAccess>;
811 def S4_storerd_ur : T_StoreAbsReg <"memd", "STrid", DoubleRegs, 0b110,
814 let AddedComplexity = 40 in
815 multiclass T_StoreAbsReg_Pats <InstHexagon MI, RegisterClass RC, ValueType VT,
817 def : Pat<(stOp (VT RC:$src4),
818 (add (shl (i32 IntRegs:$src1), u2ImmPred:$src2),
819 u0AlwaysExtPred:$src3)),
820 (MI IntRegs:$src1, u2ImmPred:$src2, u0AlwaysExtPred:$src3, RC:$src4)>;
822 def : Pat<(stOp (VT RC:$src4),
823 (add (shl IntRegs:$src1, u2ImmPred:$src2),
824 (HexagonCONST32 tglobaladdr:$src3))),
825 (MI IntRegs:$src1, u2ImmPred:$src2, tglobaladdr:$src3, RC:$src4)>;
827 def : Pat<(stOp (VT RC:$src4),
828 (add IntRegs:$src1, (HexagonCONST32 tglobaladdr:$src3))),
829 (MI IntRegs:$src1, 0, tglobaladdr:$src3, RC:$src4)>;
832 defm : T_StoreAbsReg_Pats <S4_storerd_ur, DoubleRegs, i64, store>;
833 defm : T_StoreAbsReg_Pats <S4_storeri_ur, IntRegs, i32, store>;
834 defm : T_StoreAbsReg_Pats <S4_storerb_ur, IntRegs, i32, truncstorei8>;
835 defm : T_StoreAbsReg_Pats <S4_storerh_ur, IntRegs, i32, truncstorei16>;
837 let mayStore = 1, isNVStore = 1, isExtended = 1, addrMode = BaseLongOffset,
838 opExtentBits = 6, isNewValue = 1, opNewValue = 3, opExtendable = 2 in
839 class T_StoreAbsRegNV <string mnemonic, string CextOp, bits<2> MajOp,
840 MemAccessSize AccessSz>
842 (ins IntRegs:$src1, u2Imm:$src2, u6Ext:$src3, IntRegs:$src4),
843 mnemonic#"($src1<<#$src2 + #$src3) = $src4.new">, NewValueRel {
849 let CextOpcode = CextOp;
850 let BaseOpcode = CextOp#"_shl";
853 let Inst{27-21} = 0b1101101;
854 let Inst{12-11} = 0b00;
856 let Inst{20-16} = src1;
857 let Inst{13} = src2{1};
858 let Inst{12-11} = MajOp;
859 let Inst{10-8} = src4;
860 let Inst{6} = src2{0};
861 let Inst{5-0} = src3;
864 def S4_storerbnew_ur : T_StoreAbsRegNV <"memb", "STrib", 0b00, ByteAccess>;
865 def S4_storerhnew_ur : T_StoreAbsRegNV <"memh", "STrih", 0b01, HalfWordAccess>;
866 def S4_storerinew_ur : T_StoreAbsRegNV <"memw", "STriw", 0b10, WordAccess>;
868 //===----------------------------------------------------------------------===//
869 // Template classes for the non-predicated store instructions with
870 // base + register offset addressing mode
871 //===----------------------------------------------------------------------===//
872 let isPredicable = 1 in
873 class T_store_rr <string mnemonic, RegisterClass RC, bits<3> MajOp, bit isH>
874 : STInst < (outs ), (ins IntRegs:$Rs, IntRegs:$Ru, u2Imm:$u2, RC:$Rt),
875 mnemonic#"($Rs + $Ru<<#$u2) = $Rt"#!if(isH, ".h",""),
876 [],"",V4LDST_tc_st_SLOT01>, ImmRegShl, AddrModeRel {
885 let Inst{27-24} = 0b1011;
886 let Inst{23-21} = MajOp;
887 let Inst{20-16} = Rs;
889 let Inst{13} = u2{1};
894 //===----------------------------------------------------------------------===//
895 // Template classes for the predicated store instructions with
896 // base + register offset addressing mode
897 //===----------------------------------------------------------------------===//
898 let isPredicated = 1 in
899 class T_pstore_rr <string mnemonic, RegisterClass RC, bits<3> MajOp,
900 bit isNot, bit isPredNew, bit isH>
902 (ins PredRegs:$Pv, IntRegs:$Rs, IntRegs:$Ru, u2Imm:$u2, RC:$Rt),
904 !if(isNot, "if (!$Pv", "if ($Pv")#!if(isPredNew, ".new) ",
905 ") ")#mnemonic#"($Rs+$Ru<<#$u2) = $Rt"#!if(isH, ".h",""),
906 [], "", V4LDST_tc_st_SLOT01> , AddrModeRel{
913 let isPredicatedFalse = isNot;
914 let isPredicatedNew = isPredNew;
918 let Inst{27-26} = 0b01;
919 let Inst{25} = isPredNew;
920 let Inst{24} = isNot;
921 let Inst{23-21} = MajOp;
922 let Inst{20-16} = Rs;
924 let Inst{13} = u2{1};
930 //===----------------------------------------------------------------------===//
931 // Template classes for the new-value store instructions with
932 // base + register offset addressing mode
933 //===----------------------------------------------------------------------===//
934 let isPredicable = 1, isNewValue = 1, opNewValue = 3 in
935 class T_store_new_rr <string mnemonic, bits<2> MajOp> :
936 NVInst < (outs ), (ins IntRegs:$Rs, IntRegs:$Ru, u2Imm:$u2, IntRegs:$Nt),
937 mnemonic#"($Rs + $Ru<<#$u2) = $Nt.new",
938 [],"",V4LDST_tc_st_SLOT0>, ImmRegShl, AddrModeRel {
947 let Inst{27-21} = 0b1011101;
948 let Inst{20-16} = Rs;
950 let Inst{13} = u2{1};
952 let Inst{4-3} = MajOp;
956 //===----------------------------------------------------------------------===//
957 // Template classes for the predicated new-value store instructions with
958 // base + register offset addressing mode
959 //===----------------------------------------------------------------------===//
960 let isPredicated = 1, isNewValue = 1, opNewValue = 4 in
961 class T_pstore_new_rr <string mnemonic, bits<2> MajOp, bit isNot, bit isPredNew>
963 (ins PredRegs:$Pv, IntRegs:$Rs, IntRegs:$Ru, u2Imm:$u2, IntRegs:$Nt),
964 !if(isNot, "if (!$Pv", "if ($Pv")#!if(isPredNew, ".new) ",
965 ") ")#mnemonic#"($Rs+$Ru<<#$u2) = $Nt.new",
966 [], "", V4LDST_tc_st_SLOT0>, AddrModeRel {
973 let isPredicatedFalse = isNot;
974 let isPredicatedNew = isPredNew;
977 let Inst{27-26} = 0b01;
978 let Inst{25} = isPredNew;
979 let Inst{24} = isNot;
980 let Inst{23-21} = 0b101;
981 let Inst{20-16} = Rs;
983 let Inst{13} = u2{1};
986 let Inst{4-3} = MajOp;
990 //===----------------------------------------------------------------------===//
991 // multiclass for store instructions with base + register offset addressing
993 //===----------------------------------------------------------------------===//
994 let isNVStorable = 1 in
995 multiclass ST_Idxd_shl<string mnemonic, string CextOp, RegisterClass RC,
996 bits<3> MajOp, bit isH = 0> {
997 let CextOpcode = CextOp, BaseOpcode = CextOp#_indexed_shl in {
998 def S4_#NAME#_rr : T_store_rr <mnemonic, RC, MajOp, isH>;
1001 def S4_p#NAME#t_rr : T_pstore_rr <mnemonic, RC, MajOp, 0, 0, isH>;
1002 def S4_p#NAME#f_rr : T_pstore_rr <mnemonic, RC, MajOp, 1, 0, isH>;
1005 def S4_p#NAME#tnew_rr : T_pstore_rr <mnemonic, RC, MajOp, 0, 1, isH>;
1006 def S4_p#NAME#fnew_rr : T_pstore_rr <mnemonic, RC, MajOp, 1, 1, isH>;
1010 //===----------------------------------------------------------------------===//
1011 // multiclass for new-value store instructions with base + register offset
1013 //===----------------------------------------------------------------------===//
1014 let mayStore = 1, isNVStore = 1 in
1015 multiclass ST_Idxd_shl_nv <string mnemonic, string CextOp, RegisterClass RC,
1017 let CextOpcode = CextOp, BaseOpcode = CextOp#_indexed_shl in {
1018 def S4_#NAME#new_rr : T_store_new_rr<mnemonic, MajOp>;
1021 def S4_p#NAME#newt_rr : T_pstore_new_rr <mnemonic, MajOp, 0, 0>;
1022 def S4_p#NAME#newf_rr : T_pstore_new_rr <mnemonic, MajOp, 1, 0>;
1025 def S4_p#NAME#newtnew_rr : T_pstore_new_rr <mnemonic, MajOp, 0, 1>;
1026 def S4_p#NAME#newfnew_rr : T_pstore_new_rr <mnemonic, MajOp, 1, 1>;
1030 let addrMode = BaseRegOffset, InputType = "reg", hasSideEffects = 0 in {
1031 let accessSize = ByteAccess in
1032 defm storerb: ST_Idxd_shl<"memb", "STrib", IntRegs, 0b000>,
1033 ST_Idxd_shl_nv<"memb", "STrib", IntRegs, 0b00>;
1035 let accessSize = HalfWordAccess in
1036 defm storerh: ST_Idxd_shl<"memh", "STrih", IntRegs, 0b010>,
1037 ST_Idxd_shl_nv<"memh", "STrih", IntRegs, 0b01>;
1039 let accessSize = WordAccess in
1040 defm storeri: ST_Idxd_shl<"memw", "STriw", IntRegs, 0b100>,
1041 ST_Idxd_shl_nv<"memw", "STriw", IntRegs, 0b10>;
1043 let isNVStorable = 0, accessSize = DoubleWordAccess in
1044 defm storerd: ST_Idxd_shl<"memd", "STrid", DoubleRegs, 0b110>;
1046 let isNVStorable = 0, accessSize = HalfWordAccess in
1047 defm storerf: ST_Idxd_shl<"memh", "STrif", IntRegs, 0b011, 1>;
1050 class Storexs_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
1051 : Pat<(Store Value:$Ru, (add (i32 IntRegs:$Rs),
1052 (i32 (shl (i32 IntRegs:$Rt), u2ImmPred:$u2)))),
1053 (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2, Value:$Ru)>;
1055 let AddedComplexity = 40 in {
1056 def: Storexs_pat<truncstorei8, I32, S4_storerb_rr>;
1057 def: Storexs_pat<truncstorei16, I32, S4_storerh_rr>;
1058 def: Storexs_pat<store, I32, S4_storeri_rr>;
1059 def: Storexs_pat<store, I64, S4_storerd_rr>;
1062 // memd(Rx++#s4:3)=Rtt
1063 // memd(Rx++#s4:3:circ(Mu))=Rtt
1064 // memd(Rx++I:circ(Mu))=Rtt
1066 // memd(Rx++Mu:brev)=Rtt
1067 // memd(gp+#u16:3)=Rtt
1069 // Store doubleword conditionally.
1070 // if ([!]Pv[.new]) memd(#u6)=Rtt
1071 // TODO: needs to be implemented.
1073 //===----------------------------------------------------------------------===//
1075 //===----------------------------------------------------------------------===//
1076 let isPredicable = 1, isExtendable = 1, isExtentSigned = 1, opExtentBits = 8,
1078 class T_StoreImm <string mnemonic, Operand OffsetOp, bits<2> MajOp >
1079 : STInst <(outs ), (ins IntRegs:$Rs, OffsetOp:$offset, s8Ext:$S8),
1080 mnemonic#"($Rs+#$offset)=#$S8",
1081 [], "", V4LDST_tc_st_SLOT01>,
1082 ImmRegRel, PredNewRel {
1088 string OffsetOpStr = !cast<string>(OffsetOp);
1089 let offsetBits = !if (!eq(OffsetOpStr, "u6_2Imm"), offset{7-2},
1090 !if (!eq(OffsetOpStr, "u6_1Imm"), offset{6-1},
1091 /* u6_0Imm */ offset{5-0}));
1093 let IClass = 0b0011;
1095 let Inst{27-25} = 0b110;
1096 let Inst{22-21} = MajOp;
1097 let Inst{20-16} = Rs;
1098 let Inst{12-7} = offsetBits;
1099 let Inst{13} = S8{7};
1100 let Inst{6-0} = S8{6-0};
1103 let isPredicated = 1, isExtendable = 1, isExtentSigned = 1, opExtentBits = 6,
1105 class T_StoreImm_pred <string mnemonic, Operand OffsetOp, bits<2> MajOp,
1106 bit isPredNot, bit isPredNew >
1108 (ins PredRegs:$Pv, IntRegs:$Rs, OffsetOp:$offset, s6Ext:$S6),
1109 !if(isPredNot, "if (!$Pv", "if ($Pv")#!if(isPredNew, ".new) ",
1110 ") ")#mnemonic#"($Rs+#$offset)=#$S6",
1111 [], "", V4LDST_tc_st_SLOT01>,
1112 ImmRegRel, PredNewRel {
1119 string OffsetOpStr = !cast<string>(OffsetOp);
1120 let offsetBits = !if (!eq(OffsetOpStr, "u6_2Imm"), offset{7-2},
1121 !if (!eq(OffsetOpStr, "u6_1Imm"), offset{6-1},
1122 /* u6_0Imm */ offset{5-0}));
1123 let isPredicatedNew = isPredNew;
1124 let isPredicatedFalse = isPredNot;
1126 let IClass = 0b0011;
1128 let Inst{27-25} = 0b100;
1129 let Inst{24} = isPredNew;
1130 let Inst{23} = isPredNot;
1131 let Inst{22-21} = MajOp;
1132 let Inst{20-16} = Rs;
1133 let Inst{13} = S6{5};
1134 let Inst{12-7} = offsetBits;
1136 let Inst{4-0} = S6{4-0};
1140 //===----------------------------------------------------------------------===//
1141 // multiclass for store instructions with base + immediate offset
1142 // addressing mode and immediate stored value.
1143 // mem[bhw](Rx++#s4:3)=#s8
1144 // if ([!]Pv[.new]) mem[bhw](Rx++#s4:3)=#s6
1145 //===----------------------------------------------------------------------===//
1147 multiclass ST_Imm_Pred <string mnemonic, Operand OffsetOp, bits<2> MajOp,
1149 def _io : T_StoreImm_pred <mnemonic, OffsetOp, MajOp, PredNot, 0>;
1151 def new_io : T_StoreImm_pred <mnemonic, OffsetOp, MajOp, PredNot, 1>;
1154 multiclass ST_Imm <string mnemonic, string CextOp, Operand OffsetOp,
1156 let CextOpcode = CextOp, BaseOpcode = CextOp#_imm in {
1157 def _io : T_StoreImm <mnemonic, OffsetOp, MajOp>;
1159 defm t : ST_Imm_Pred <mnemonic, OffsetOp, MajOp, 0>;
1160 defm f : ST_Imm_Pred <mnemonic, OffsetOp, MajOp, 1>;
1164 let hasSideEffects = 0, validSubTargets = HasV4SubT, addrMode = BaseImmOffset,
1165 InputType = "imm", isCodeGenOnly = 0 in {
1166 let accessSize = ByteAccess in
1167 defm S4_storeirb : ST_Imm<"memb", "STrib", u6_0Imm, 0b00>;
1169 let accessSize = HalfWordAccess in
1170 defm S4_storeirh : ST_Imm<"memh", "STrih", u6_1Imm, 0b01>;
1172 let accessSize = WordAccess in
1173 defm S4_storeiri : ST_Imm<"memw", "STriw", u6_2Imm, 0b10>;
1176 let Predicates = [HasV4T], AddedComplexity = 10 in {
1177 def: Pat<(truncstorei8 s8ExtPred:$src3, (add IntRegs:$src1, u6_0ImmPred:$src2)),
1178 (S4_storeirb_io IntRegs:$src1, u6_0ImmPred:$src2, s8ExtPred:$src3)>;
1180 def: Pat<(truncstorei16 s8ExtPred:$src3, (add IntRegs:$src1,
1181 u6_1ImmPred:$src2)),
1182 (S4_storeirh_io IntRegs:$src1, u6_1ImmPred:$src2, s8ExtPred:$src3)>;
1184 def: Pat<(store s8ExtPred:$src3, (add IntRegs:$src1, u6_2ImmPred:$src2)),
1185 (S4_storeiri_io IntRegs:$src1, u6_2ImmPred:$src2, s8ExtPred:$src3)>;
1188 let AddedComplexity = 6 in
1189 def : Pat <(truncstorei8 s8ExtPred:$src2, (i32 IntRegs:$src1)),
1190 (S4_storeirb_io IntRegs:$src1, 0, s8ExtPred:$src2)>,
1193 // memb(Rx++#s4:0:circ(Mu))=Rt
1194 // memb(Rx++I:circ(Mu))=Rt
1196 // memb(Rx++Mu:brev)=Rt
1197 // memb(gp+#u16:0)=Rt
1201 // TODO: needs to be implemented
1202 // memh(Re=#U6)=Rt.H
1203 // memh(Rs+#s11:1)=Rt.H
1204 let AddedComplexity = 6 in
1205 def : Pat <(truncstorei16 s8ExtPred:$src2, (i32 IntRegs:$src1)),
1206 (S4_storeirh_io IntRegs:$src1, 0, s8ExtPred:$src2)>,
1209 // memh(Rs+Ru<<#u2)=Rt.H
1210 // TODO: needs to be implemented.
1212 // memh(Ru<<#u2+#U6)=Rt.H
1213 // memh(Rx++#s4:1:circ(Mu))=Rt.H
1214 // memh(Rx++#s4:1:circ(Mu))=Rt
1215 // memh(Rx++I:circ(Mu))=Rt.H
1216 // memh(Rx++I:circ(Mu))=Rt
1217 // memh(Rx++Mu)=Rt.H
1219 // memh(Rx++Mu:brev)=Rt.H
1220 // memh(Rx++Mu:brev)=Rt
1221 // memh(gp+#u16:1)=Rt
1222 // if ([!]Pv[.new]) memh(#u6)=Rt.H
1223 // if ([!]Pv[.new]) memh(#u6)=Rt
1226 // if ([!]Pv[.new]) memh(Rs+#u6:1)=Rt.H
1227 // TODO: needs to be implemented.
1229 // if ([!]Pv[.new]) memh(Rx++#s4:1)=Rt.H
1230 // TODO: Needs to be implemented.
1234 // TODO: Needs to be implemented.
1237 let hasSideEffects = 0 in
1238 def STriw_pred_V4 : STInst2<(outs),
1239 (ins MEMri:$addr, PredRegs:$src1),
1240 "Error; should not emit",
1244 let AddedComplexity = 6 in
1245 def : Pat <(store s8ExtPred:$src2, (i32 IntRegs:$src1)),
1246 (S4_storeiri_io IntRegs:$src1, 0, s8ExtPred:$src2)>,
1249 // memw(Rx++#s4:2)=Rt
1250 // memw(Rx++#s4:2:circ(Mu))=Rt
1251 // memw(Rx++I:circ(Mu))=Rt
1253 // memw(Rx++Mu:brev)=Rt
1255 //===----------------------------------------------------------------------===
1257 //===----------------------------------------------------------------------===
1260 //===----------------------------------------------------------------------===//
1262 //===----------------------------------------------------------------------===//
1264 let opNewValue = 2, opExtendable = 1, isExtentSigned = 1, isPredicable = 1 in
1265 class T_store_io_nv <string mnemonic, RegisterClass RC,
1266 Operand ImmOp, bits<2>MajOp>
1267 : NVInst_V4 <(outs),
1268 (ins IntRegs:$src1, ImmOp:$src2, RC:$src3),
1269 mnemonic#"($src1+#$src2) = $src3.new",
1270 [],"",ST_tc_st_SLOT0> {
1272 bits<13> src2; // Actual address offset
1274 bits<11> offsetBits; // Represents offset encoding
1276 let opExtentBits = !if (!eq(mnemonic, "memb"), 11,
1277 !if (!eq(mnemonic, "memh"), 12,
1278 !if (!eq(mnemonic, "memw"), 13, 0)));
1280 let opExtentAlign = !if (!eq(mnemonic, "memb"), 0,
1281 !if (!eq(mnemonic, "memh"), 1,
1282 !if (!eq(mnemonic, "memw"), 2, 0)));
1284 let offsetBits = !if (!eq(mnemonic, "memb"), src2{10-0},
1285 !if (!eq(mnemonic, "memh"), src2{11-1},
1286 !if (!eq(mnemonic, "memw"), src2{12-2}, 0)));
1288 let IClass = 0b1010;
1291 let Inst{26-25} = offsetBits{10-9};
1292 let Inst{24-21} = 0b1101;
1293 let Inst{20-16} = src1;
1294 let Inst{13} = offsetBits{8};
1295 let Inst{12-11} = MajOp;
1296 let Inst{10-8} = src3;
1297 let Inst{7-0} = offsetBits{7-0};
1300 let opExtendable = 2, opNewValue = 3, isPredicated = 1 in
1301 class T_pstore_io_nv <string mnemonic, RegisterClass RC, Operand predImmOp,
1302 bits<2>MajOp, bit PredNot, bit isPredNew>
1303 : NVInst_V4 <(outs),
1304 (ins PredRegs:$src1, IntRegs:$src2, predImmOp:$src3, RC:$src4),
1305 !if(PredNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
1306 ") ")#mnemonic#"($src2+#$src3) = $src4.new",
1307 [],"",V2LDST_tc_st_SLOT0> {
1312 bits<6> offsetBits; // Represents offset encoding
1314 let isPredicatedNew = isPredNew;
1315 let isPredicatedFalse = PredNot;
1316 let opExtentBits = !if (!eq(mnemonic, "memb"), 6,
1317 !if (!eq(mnemonic, "memh"), 7,
1318 !if (!eq(mnemonic, "memw"), 8, 0)));
1320 let opExtentAlign = !if (!eq(mnemonic, "memb"), 0,
1321 !if (!eq(mnemonic, "memh"), 1,
1322 !if (!eq(mnemonic, "memw"), 2, 0)));
1324 let offsetBits = !if (!eq(mnemonic, "memb"), src3{5-0},
1325 !if (!eq(mnemonic, "memh"), src3{6-1},
1326 !if (!eq(mnemonic, "memw"), src3{7-2}, 0)));
1328 let IClass = 0b0100;
1331 let Inst{26} = PredNot;
1332 let Inst{25} = isPredNew;
1333 let Inst{24-21} = 0b0101;
1334 let Inst{20-16} = src2;
1335 let Inst{13} = offsetBits{5};
1336 let Inst{12-11} = MajOp;
1337 let Inst{10-8} = src4;
1338 let Inst{7-3} = offsetBits{4-0};
1340 let Inst{1-0} = src1;
1343 // multiclass for new-value store instructions with base + immediate offset.
1345 let mayStore = 1, isNVStore = 1, isNewValue = 1, hasSideEffects = 0,
1347 multiclass ST_Idxd_nv<string mnemonic, string CextOp, RegisterClass RC,
1348 Operand ImmOp, Operand predImmOp, bits<2> MajOp> {
1350 let CextOpcode = CextOp, BaseOpcode = CextOp#_indexed in {
1351 def S2_#NAME#new_io : T_store_io_nv <mnemonic, RC, ImmOp, MajOp>;
1353 def S2_p#NAME#newt_io :T_pstore_io_nv <mnemonic, RC, predImmOp, MajOp, 0, 0>;
1354 def S2_p#NAME#newf_io :T_pstore_io_nv <mnemonic, RC, predImmOp, MajOp, 1, 0>;
1356 def S4_p#NAME#newtnew_io :T_pstore_io_nv <mnemonic, RC, predImmOp,
1358 def S4_p#NAME#newfnew_io :T_pstore_io_nv <mnemonic, RC, predImmOp,
1363 let addrMode = BaseImmOffset, InputType = "imm", isCodeGenOnly = 0 in {
1364 let accessSize = ByteAccess in
1365 defm storerb: ST_Idxd_nv<"memb", "STrib", IntRegs, s11_0Ext,
1366 u6_0Ext, 0b00>, AddrModeRel;
1368 let accessSize = HalfWordAccess, opExtentAlign = 1 in
1369 defm storerh: ST_Idxd_nv<"memh", "STrih", IntRegs, s11_1Ext,
1370 u6_1Ext, 0b01>, AddrModeRel;
1372 let accessSize = WordAccess, opExtentAlign = 2 in
1373 defm storeri: ST_Idxd_nv<"memw", "STriw", IntRegs, s11_2Ext,
1374 u6_2Ext, 0b10>, AddrModeRel;
1377 //===----------------------------------------------------------------------===//
1378 // Post increment loads with register offset.
1379 //===----------------------------------------------------------------------===//
1381 let hasNewValue = 1, isCodeGenOnly = 0 in
1382 def L2_loadbsw2_pr : T_load_pr <"membh", IntRegs, 0b0001, HalfWordAccess>;
1384 let isCodeGenOnly = 0 in
1385 def L2_loadbsw4_pr : T_load_pr <"membh", DoubleRegs, 0b0111, WordAccess>;
1387 //===----------------------------------------------------------------------===//
1388 // Template class for non-predicated post increment .new stores
1389 // mem[bhwd](Rx++#s4:[0123])=Nt.new
1390 //===----------------------------------------------------------------------===//
1391 let isPredicable = 1, hasSideEffects = 0, validSubTargets = HasV4SubT,
1392 addrMode = PostInc, isNVStore = 1, isNewValue = 1, opNewValue = 3 in
1393 class T_StorePI_nv <string mnemonic, Operand ImmOp, bits<2> MajOp >
1394 : NVInstPI_V4 <(outs IntRegs:$_dst_),
1395 (ins IntRegs:$src1, ImmOp:$offset, IntRegs:$src2),
1396 mnemonic#"($src1++#$offset) = $src2.new",
1397 [], "$src1 = $_dst_">,
1404 string ImmOpStr = !cast<string>(ImmOp);
1405 let offsetBits = !if (!eq(ImmOpStr, "s4_2Imm"), offset{5-2},
1406 !if (!eq(ImmOpStr, "s4_1Imm"), offset{4-1},
1407 /* s4_0Imm */ offset{3-0}));
1408 let IClass = 0b1010;
1410 let Inst{27-21} = 0b1011101;
1411 let Inst{20-16} = src1;
1413 let Inst{12-11} = MajOp;
1414 let Inst{10-8} = src2;
1416 let Inst{6-3} = offsetBits;
1420 //===----------------------------------------------------------------------===//
1421 // Template class for predicated post increment .new stores
1422 // if([!]Pv[.new]) mem[bhwd](Rx++#s4:[0123])=Nt.new
1423 //===----------------------------------------------------------------------===//
1424 let isPredicated = 1, hasSideEffects = 0, validSubTargets = HasV4SubT,
1425 addrMode = PostInc, isNVStore = 1, isNewValue = 1, opNewValue = 4 in
1426 class T_StorePI_nv_pred <string mnemonic, Operand ImmOp,
1427 bits<2> MajOp, bit isPredNot, bit isPredNew >
1428 : NVInstPI_V4 <(outs IntRegs:$_dst_),
1429 (ins PredRegs:$src1, IntRegs:$src2,
1430 ImmOp:$offset, IntRegs:$src3),
1431 !if(isPredNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
1432 ") ")#mnemonic#"($src2++#$offset) = $src3.new",
1433 [], "$src2 = $_dst_">,
1441 string ImmOpStr = !cast<string>(ImmOp);
1442 let offsetBits = !if (!eq(ImmOpStr, "s4_2Imm"), offset{5-2},
1443 !if (!eq(ImmOpStr, "s4_1Imm"), offset{4-1},
1444 /* s4_0Imm */ offset{3-0}));
1445 let isPredicatedNew = isPredNew;
1446 let isPredicatedFalse = isPredNot;
1448 let IClass = 0b1010;
1450 let Inst{27-21} = 0b1011101;
1451 let Inst{20-16} = src2;
1453 let Inst{12-11} = MajOp;
1454 let Inst{10-8} = src3;
1455 let Inst{7} = isPredNew;
1456 let Inst{6-3} = offsetBits;
1457 let Inst{2} = isPredNot;
1458 let Inst{1-0} = src1;
1461 multiclass ST_PostInc_Pred_nv<string mnemonic, Operand ImmOp,
1462 bits<2> MajOp, bit PredNot> {
1463 def _pi : T_StorePI_nv_pred <mnemonic, ImmOp, MajOp, PredNot, 0>;
1466 def new_pi : T_StorePI_nv_pred <mnemonic, ImmOp, MajOp, PredNot, 1>;
1469 multiclass ST_PostInc_nv<string mnemonic, string BaseOp, Operand ImmOp,
1471 let BaseOpcode = "POST_"#BaseOp in {
1472 def S2_#NAME#_pi : T_StorePI_nv <mnemonic, ImmOp, MajOp>;
1475 defm S2_p#NAME#t : ST_PostInc_Pred_nv <mnemonic, ImmOp, MajOp, 0>;
1476 defm S2_p#NAME#f : ST_PostInc_Pred_nv <mnemonic, ImmOp, MajOp, 1>;
1480 let accessSize = ByteAccess, isCodeGenOnly = 0 in
1481 defm storerbnew: ST_PostInc_nv <"memb", "STrib", s4_0Imm, 0b00>;
1483 let accessSize = HalfWordAccess, isCodeGenOnly = 0 in
1484 defm storerhnew: ST_PostInc_nv <"memh", "STrih", s4_1Imm, 0b01>;
1486 let accessSize = WordAccess, isCodeGenOnly = 0 in
1487 defm storerinew: ST_PostInc_nv <"memw", "STriw", s4_2Imm, 0b10>;
1489 //===----------------------------------------------------------------------===//
1490 // Template class for post increment .new stores with register offset
1491 //===----------------------------------------------------------------------===//
1492 let isNewValue = 1, mayStore = 1, isNVStore = 1, opNewValue = 3 in
1493 class T_StorePI_RegNV <string mnemonic, bits<2> MajOp, MemAccessSize AccessSz>
1494 : NVInstPI_V4 <(outs IntRegs:$_dst_),
1495 (ins IntRegs:$src1, ModRegs:$src2, IntRegs:$src3),
1496 #mnemonic#"($src1++$src2) = $src3.new",
1497 [], "$src1 = $_dst_"> {
1501 let accessSize = AccessSz;
1503 let IClass = 0b1010;
1505 let Inst{27-21} = 0b1101101;
1506 let Inst{20-16} = src1;
1507 let Inst{13} = src2;
1508 let Inst{12-11} = MajOp;
1509 let Inst{10-8} = src3;
1513 let isCodeGenOnly = 0 in {
1514 def S2_storerbnew_pr : T_StorePI_RegNV<"memb", 0b00, ByteAccess>;
1515 def S2_storerhnew_pr : T_StorePI_RegNV<"memh", 0b01, HalfWordAccess>;
1516 def S2_storerinew_pr : T_StorePI_RegNV<"memw", 0b10, WordAccess>;
1519 // memb(Rx++#s4:0:circ(Mu))=Nt.new
1520 // memb(Rx++I:circ(Mu))=Nt.new
1521 // memb(Rx++Mu)=Nt.new
1522 // memb(Rx++Mu:brev)=Nt.new
1523 // memh(Rx++#s4:1:circ(Mu))=Nt.new
1524 // memh(Rx++I:circ(Mu))=Nt.new
1525 // memh(Rx++Mu)=Nt.new
1526 // memh(Rx++Mu:brev)=Nt.new
1528 // memw(Rx++#s4:2:circ(Mu))=Nt.new
1529 // memw(Rx++I:circ(Mu))=Nt.new
1530 // memw(Rx++Mu)=Nt.new
1531 // memw(Rx++Mu:brev)=Nt.new
1533 //===----------------------------------------------------------------------===//
1535 //===----------------------------------------------------------------------===//
1537 //===----------------------------------------------------------------------===//
1539 //===----------------------------------------------------------------------===//
1541 //===----------------------------------------------------------------------===//
1542 // multiclass/template class for the new-value compare jumps with the register
1544 //===----------------------------------------------------------------------===//
1546 let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 11,
1547 opExtentAlign = 2 in
1548 class NVJrr_template<string mnemonic, bits<3> majOp, bit NvOpNum,
1549 bit isNegCond, bit isTak>
1551 (ins IntRegs:$src1, IntRegs:$src2, brtarget:$offset),
1552 "if ("#!if(isNegCond, "!","")#mnemonic#
1553 "($src1"#!if(!eq(NvOpNum, 0),".new, ",", ")#
1554 "$src2"#!if(!eq(NvOpNum, 1),".new))","))")#" jump:"
1555 #!if(isTak, "t","nt")#" $offset", []> {
1559 bits<3> Ns; // New-Value Operand
1560 bits<5> RegOp; // Non-New-Value Operand
1563 let isTaken = isTak;
1564 let isPredicatedFalse = isNegCond;
1565 let opNewValue{0} = NvOpNum;
1567 let Ns = !if(!eq(NvOpNum, 0), src1{2-0}, src2{2-0});
1568 let RegOp = !if(!eq(NvOpNum, 0), src2, src1);
1570 let IClass = 0b0010;
1572 let Inst{25-23} = majOp;
1573 let Inst{22} = isNegCond;
1574 let Inst{18-16} = Ns;
1575 let Inst{13} = isTak;
1576 let Inst{12-8} = RegOp;
1577 let Inst{21-20} = offset{10-9};
1578 let Inst{7-1} = offset{8-2};
1582 multiclass NVJrr_cond<string mnemonic, bits<3> majOp, bit NvOpNum,
1584 // Branch not taken:
1585 def _nt_V4: NVJrr_template<mnemonic, majOp, NvOpNum, isNegCond, 0>;
1587 def _t_V4: NVJrr_template<mnemonic, majOp, NvOpNum, isNegCond, 1>;
1590 // NvOpNum = 0 -> First Operand is a new-value Register
1591 // NvOpNum = 1 -> Second Operand is a new-value Register
1593 multiclass NVJrr_base<string mnemonic, string BaseOp, bits<3> majOp,
1595 let BaseOpcode = BaseOp#_NVJ in {
1596 defm _t_Jumpnv : NVJrr_cond<mnemonic, majOp, NvOpNum, 0>; // True cond
1597 defm _f_Jumpnv : NVJrr_cond<mnemonic, majOp, NvOpNum, 1>; // False cond
1601 // if ([!]cmp.eq(Ns.new,Rt)) jump:[n]t #r9:2
1602 // if ([!]cmp.gt(Ns.new,Rt)) jump:[n]t #r9:2
1603 // if ([!]cmp.gtu(Ns.new,Rt)) jump:[n]t #r9:2
1604 // if ([!]cmp.gt(Rt,Ns.new)) jump:[n]t #r9:2
1605 // if ([!]cmp.gtu(Rt,Ns.new)) jump:[n]t #r9:2
1607 let isPredicated = 1, isBranch = 1, isNewValue = 1, isTerminator = 1,
1608 Defs = [PC], hasSideEffects = 0, validSubTargets = HasV4SubT,
1609 isCodeGenOnly = 0 in {
1610 defm CMPEQrr : NVJrr_base<"cmp.eq", "CMPEQ", 0b000, 0>, PredRel;
1611 defm CMPGTrr : NVJrr_base<"cmp.gt", "CMPGT", 0b001, 0>, PredRel;
1612 defm CMPGTUrr : NVJrr_base<"cmp.gtu", "CMPGTU", 0b010, 0>, PredRel;
1613 defm CMPLTrr : NVJrr_base<"cmp.gt", "CMPLT", 0b011, 1>, PredRel;
1614 defm CMPLTUrr : NVJrr_base<"cmp.gtu", "CMPLTU", 0b100, 1>, PredRel;
1617 //===----------------------------------------------------------------------===//
1618 // multiclass/template class for the new-value compare jumps instruction
1619 // with a register and an unsigned immediate (U5) operand.
1620 //===----------------------------------------------------------------------===//
1622 let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 11,
1623 opExtentAlign = 2 in
1624 class NVJri_template<string mnemonic, bits<3> majOp, bit isNegCond,
1627 (ins IntRegs:$src1, u5Imm:$src2, brtarget:$offset),
1628 "if ("#!if(isNegCond, "!","")#mnemonic#"($src1.new, #$src2)) jump:"
1629 #!if(isTak, "t","nt")#" $offset", []> {
1631 let isTaken = isTak;
1632 let isPredicatedFalse = isNegCond;
1633 let isTaken = isTak;
1639 let IClass = 0b0010;
1641 let Inst{25-23} = majOp;
1642 let Inst{22} = isNegCond;
1643 let Inst{18-16} = src1;
1644 let Inst{13} = isTak;
1645 let Inst{12-8} = src2;
1646 let Inst{21-20} = offset{10-9};
1647 let Inst{7-1} = offset{8-2};
1650 multiclass NVJri_cond<string mnemonic, bits<3> majOp, bit isNegCond> {
1651 // Branch not taken:
1652 def _nt_V4: NVJri_template<mnemonic, majOp, isNegCond, 0>;
1654 def _t_V4: NVJri_template<mnemonic, majOp, isNegCond, 1>;
1657 multiclass NVJri_base<string mnemonic, string BaseOp, bits<3> majOp> {
1658 let BaseOpcode = BaseOp#_NVJri in {
1659 defm _t_Jumpnv : NVJri_cond<mnemonic, majOp, 0>; // True Cond
1660 defm _f_Jumpnv : NVJri_cond<mnemonic, majOp, 1>; // False cond
1664 // if ([!]cmp.eq(Ns.new,#U5)) jump:[n]t #r9:2
1665 // if ([!]cmp.gt(Ns.new,#U5)) jump:[n]t #r9:2
1666 // if ([!]cmp.gtu(Ns.new,#U5)) jump:[n]t #r9:2
1668 let isPredicated = 1, isBranch = 1, isNewValue = 1, isTerminator = 1,
1669 Defs = [PC], hasSideEffects = 0, validSubTargets = HasV4SubT,
1670 isCodeGenOnly = 0 in {
1671 defm CMPEQri : NVJri_base<"cmp.eq", "CMPEQ", 0b000>, PredRel;
1672 defm CMPGTri : NVJri_base<"cmp.gt", "CMPGT", 0b001>, PredRel;
1673 defm CMPGTUri : NVJri_base<"cmp.gtu", "CMPGTU", 0b010>, PredRel;
1676 //===----------------------------------------------------------------------===//
1677 // multiclass/template class for the new-value compare jumps instruction
1678 // with a register and an hardcoded 0/-1 immediate value.
1679 //===----------------------------------------------------------------------===//
1681 let isExtendable = 1, opExtendable = 1, isExtentSigned = 1, opExtentBits = 11,
1682 opExtentAlign = 2 in
1683 class NVJ_ConstImm_template<string mnemonic, bits<3> majOp, string ImmVal,
1684 bit isNegCond, bit isTak>
1686 (ins IntRegs:$src1, brtarget:$offset),
1687 "if ("#!if(isNegCond, "!","")#mnemonic
1688 #"($src1.new, #"#ImmVal#")) jump:"
1689 #!if(isTak, "t","nt")#" $offset", []> {
1691 let isTaken = isTak;
1692 let isPredicatedFalse = isNegCond;
1693 let isTaken = isTak;
1697 let IClass = 0b0010;
1699 let Inst{25-23} = majOp;
1700 let Inst{22} = isNegCond;
1701 let Inst{18-16} = src1;
1702 let Inst{13} = isTak;
1703 let Inst{21-20} = offset{10-9};
1704 let Inst{7-1} = offset{8-2};
1707 multiclass NVJ_ConstImm_cond<string mnemonic, bits<3> majOp, string ImmVal,
1709 // Branch not taken:
1710 def _nt_V4: NVJ_ConstImm_template<mnemonic, majOp, ImmVal, isNegCond, 0>;
1712 def _t_V4: NVJ_ConstImm_template<mnemonic, majOp, ImmVal, isNegCond, 1>;
1715 multiclass NVJ_ConstImm_base<string mnemonic, string BaseOp, bits<3> majOp,
1717 let BaseOpcode = BaseOp#_NVJ_ConstImm in {
1718 defm _t_Jumpnv : NVJ_ConstImm_cond<mnemonic, majOp, ImmVal, 0>; // True
1719 defm _f_Jumpnv : NVJ_ConstImm_cond<mnemonic, majOp, ImmVal, 1>; // False
1723 // if ([!]tstbit(Ns.new,#0)) jump:[n]t #r9:2
1724 // if ([!]cmp.eq(Ns.new,#-1)) jump:[n]t #r9:2
1725 // if ([!]cmp.gt(Ns.new,#-1)) jump:[n]t #r9:2
1727 let isPredicated = 1, isBranch = 1, isNewValue = 1, isTerminator=1,
1728 Defs = [PC], hasSideEffects = 0, isCodeGenOnly = 0 in {
1729 defm TSTBIT0 : NVJ_ConstImm_base<"tstbit", "TSTBIT", 0b011, "0">, PredRel;
1730 defm CMPEQn1 : NVJ_ConstImm_base<"cmp.eq", "CMPEQ", 0b100, "-1">, PredRel;
1731 defm CMPGTn1 : NVJ_ConstImm_base<"cmp.gt", "CMPGT", 0b101, "-1">, PredRel;
1734 // J4_hintjumpr: Hint indirect conditional jump.
1735 let isBranch = 1, isIndirectBranch = 1, hasSideEffects = 0, isCodeGenOnly = 0 in
1736 def J4_hintjumpr: JRInst <
1741 let IClass = 0b0101;
1742 let Inst{27-21} = 0b0010101;
1743 let Inst{20-16} = Rs;
1746 //===----------------------------------------------------------------------===//
1748 //===----------------------------------------------------------------------===//
1750 //===----------------------------------------------------------------------===//
1752 //===----------------------------------------------------------------------===//
1755 let hasNewValue = 1, isExtendable = 1, opExtendable = 1,
1756 isExtentSigned = 0, opExtentBits = 6, hasSideEffects = 0,
1757 Uses = [PC], validSubTargets = HasV4SubT, isCodeGenOnly = 0 in
1758 def C4_addipc : CRInst <(outs IntRegs:$Rd), (ins u6Ext:$u6),
1759 "$Rd = add(pc, #$u6)", [], "", CR_tc_2_SLOT3 > {
1763 let IClass = 0b0110;
1764 let Inst{27-16} = 0b101001001001;
1765 let Inst{12-7} = u6;
1771 let hasSideEffects = 0 in
1772 class T_LOGICAL_3OP<string MnOp1, string MnOp2, bits<2> OpBits, bit IsNeg>
1773 : CRInst<(outs PredRegs:$Pd),
1774 (ins PredRegs:$Ps, PredRegs:$Pt, PredRegs:$Pu),
1775 "$Pd = " # MnOp1 # "($Ps, " # MnOp2 # "($Pt, " #
1776 !if (IsNeg,"!","") # "$Pu))",
1777 [], "", CR_tc_2early_SLOT23> {
1783 let IClass = 0b0110;
1784 let Inst{27-24} = 0b1011;
1785 let Inst{23} = IsNeg;
1786 let Inst{22-21} = OpBits;
1788 let Inst{17-16} = Ps;
1795 let isCodeGenOnly = 0 in {
1796 def C4_and_and : T_LOGICAL_3OP<"and", "and", 0b00, 0>;
1797 def C4_and_or : T_LOGICAL_3OP<"and", "or", 0b01, 0>;
1798 def C4_or_and : T_LOGICAL_3OP<"or", "and", 0b10, 0>;
1799 def C4_or_or : T_LOGICAL_3OP<"or", "or", 0b11, 0>;
1800 def C4_and_andn : T_LOGICAL_3OP<"and", "and", 0b00, 1>;
1801 def C4_and_orn : T_LOGICAL_3OP<"and", "or", 0b01, 1>;
1802 def C4_or_andn : T_LOGICAL_3OP<"or", "and", 0b10, 1>;
1803 def C4_or_orn : T_LOGICAL_3OP<"or", "or", 0b11, 1>;
1806 //===----------------------------------------------------------------------===//
1808 //===----------------------------------------------------------------------===//
1810 //===----------------------------------------------------------------------===//
1812 //===----------------------------------------------------------------------===//
1814 // Logical with-not instructions.
1815 let validSubTargets = HasV4SubT, isCodeGenOnly = 0 in {
1816 def A4_andnp : T_ALU64_logical<"and", 0b001, 1, 0, 1>;
1817 def A4_ornp : T_ALU64_logical<"or", 0b011, 1, 0, 1>;
1820 let hasNewValue = 1, hasSideEffects = 0, isCodeGenOnly = 0 in
1821 def S4_parity: ALU64Inst<(outs IntRegs:$Rd), (ins IntRegs:$Rs, IntRegs:$Rt),
1822 "$Rd = parity($Rs, $Rt)", [], "", ALU64_tc_2_SLOT23> {
1827 let IClass = 0b1101;
1828 let Inst{27-21} = 0b0101111;
1829 let Inst{20-16} = Rs;
1830 let Inst{12-8} = Rt;
1833 // Add and accumulate.
1834 // Rd=add(Rs,add(Ru,#s6))
1835 let isExtentSigned = 1, hasNewValue = 1, isExtendable = 1, opExtentBits = 6,
1836 opExtendable = 3, isCodeGenOnly = 0 in
1837 def S4_addaddi : ALU64Inst <(outs IntRegs:$Rd),
1838 (ins IntRegs:$Rs, IntRegs:$Ru, s6Ext:$s6),
1839 "$Rd = add($Rs, add($Ru, #$s6))" ,
1840 [(set (i32 IntRegs:$Rd), (add (i32 IntRegs:$Rs),
1841 (add (i32 IntRegs:$Ru), s6_16ExtPred:$s6)))],
1842 "", ALU64_tc_2_SLOT23> {
1848 let IClass = 0b1101;
1850 let Inst{27-23} = 0b10110;
1851 let Inst{22-21} = s6{5-4};
1852 let Inst{20-16} = Rs;
1853 let Inst{13} = s6{3};
1854 let Inst{12-8} = Rd;
1855 let Inst{7-5} = s6{2-0};
1859 let isExtentSigned = 1, hasSideEffects = 0, hasNewValue = 1, isExtendable = 1,
1860 opExtentBits = 6, opExtendable = 2, isCodeGenOnly = 0 in
1861 def S4_subaddi: ALU64Inst <(outs IntRegs:$Rd),
1862 (ins IntRegs:$Rs, s6Ext:$s6, IntRegs:$Ru),
1863 "$Rd = add($Rs, sub(#$s6, $Ru))",
1864 [], "", ALU64_tc_2_SLOT23> {
1870 let IClass = 0b1101;
1872 let Inst{27-23} = 0b10111;
1873 let Inst{22-21} = s6{5-4};
1874 let Inst{20-16} = Rs;
1875 let Inst{13} = s6{3};
1876 let Inst{12-8} = Rd;
1877 let Inst{7-5} = s6{2-0};
1882 // Rdd=extract(Rss,#u6,#U6)
1883 // Rdd=extract(Rss,Rtt)
1884 // Rd=extract(Rs,Rtt)
1885 // Rd=extract(Rs,#u5,#U5)
1887 let isCodeGenOnly = 0 in {
1888 def S4_extractp_rp : T_S3op_64 < "extract", 0b11, 0b100, 0>;
1889 def S4_extractp : T_S2op_extract <"extract", 0b1010, DoubleRegs, u6Imm>;
1892 let hasNewValue = 1, isCodeGenOnly = 0 in {
1893 def S4_extract_rp : T_S3op_extract<"extract", 0b01>;
1894 def S4_extract : T_S2op_extract <"extract", 0b1101, IntRegs, u5Imm>;
1897 // Complex add/sub halfwords/words
1898 let Defs = [USR_OVF], isCodeGenOnly = 0 in {
1899 def S4_vxaddsubh : T_S3op_64 < "vxaddsubh", 0b01, 0b100, 0, 1>;
1900 def S4_vxaddsubw : T_S3op_64 < "vxaddsubw", 0b01, 0b000, 0, 1>;
1901 def S4_vxsubaddh : T_S3op_64 < "vxsubaddh", 0b01, 0b110, 0, 1>;
1902 def S4_vxsubaddw : T_S3op_64 < "vxsubaddw", 0b01, 0b010, 0, 1>;
1905 let Defs = [USR_OVF], isCodeGenOnly = 0 in {
1906 def S4_vxaddsubhr : T_S3op_64 < "vxaddsubh", 0b11, 0b000, 0, 1, 1, 1>;
1907 def S4_vxsubaddhr : T_S3op_64 < "vxsubaddh", 0b11, 0b010, 0, 1, 1, 1>;
1910 let Itinerary = M_tc_3x_SLOT23, Defs = [USR_OVF], isCodeGenOnly = 0 in {
1911 def M4_mac_up_s1_sat: T_MType_acc_rr<"+= mpy", 0b011, 0b000, 0, [], 0, 1, 1>;
1912 def M4_nac_up_s1_sat: T_MType_acc_rr<"-= mpy", 0b011, 0b001, 0, [], 0, 1, 1>;
1915 // Logical xor with xor accumulation.
1916 // Rxx^=xor(Rss,Rtt)
1917 let hasSideEffects = 0, isCodeGenOnly = 0 in
1919 : SInst <(outs DoubleRegs:$Rxx),
1920 (ins DoubleRegs:$dst2, DoubleRegs:$Rss, DoubleRegs:$Rtt),
1921 "$Rxx ^= xor($Rss, $Rtt)",
1922 [(set (i64 DoubleRegs:$Rxx),
1923 (xor (i64 DoubleRegs:$dst2), (xor (i64 DoubleRegs:$Rss),
1924 (i64 DoubleRegs:$Rtt))))],
1925 "$dst2 = $Rxx", S_3op_tc_1_SLOT23> {
1930 let IClass = 0b1100;
1932 let Inst{27-23} = 0b10101;
1933 let Inst{20-16} = Rss;
1934 let Inst{12-8} = Rtt;
1935 let Inst{4-0} = Rxx;
1938 // Rotate and reduce bytes
1939 // Rdd=vrcrotate(Rss,Rt,#u2)
1940 let hasSideEffects = 0, isCodeGenOnly = 0 in
1942 : SInst <(outs DoubleRegs:$Rdd),
1943 (ins DoubleRegs:$Rss, IntRegs:$Rt, u2Imm:$u2),
1944 "$Rdd = vrcrotate($Rss, $Rt, #$u2)",
1945 [], "", S_3op_tc_3x_SLOT23> {
1951 let IClass = 0b1100;
1953 let Inst{27-22} = 0b001111;
1954 let Inst{20-16} = Rss;
1955 let Inst{13} = u2{1};
1956 let Inst{12-8} = Rt;
1957 let Inst{7-6} = 0b11;
1958 let Inst{5} = u2{0};
1959 let Inst{4-0} = Rdd;
1962 // Rotate and reduce bytes with accumulation
1963 // Rxx+=vrcrotate(Rss,Rt,#u2)
1964 let hasSideEffects = 0, isCodeGenOnly = 0 in
1965 def S4_vrcrotate_acc
1966 : SInst <(outs DoubleRegs:$Rxx),
1967 (ins DoubleRegs:$dst2, DoubleRegs:$Rss, IntRegs:$Rt, u2Imm:$u2),
1968 "$Rxx += vrcrotate($Rss, $Rt, #$u2)", [],
1969 "$dst2 = $Rxx", S_3op_tc_3x_SLOT23> {
1975 let IClass = 0b1100;
1977 let Inst{27-21} = 0b1011101;
1978 let Inst{20-16} = Rss;
1979 let Inst{13} = u2{1};
1980 let Inst{12-8} = Rt;
1981 let Inst{5} = u2{0};
1982 let Inst{4-0} = Rxx;
1986 // Vector reduce conditional negate halfwords
1987 let hasSideEffects = 0, isCodeGenOnly = 0 in
1989 : SInst <(outs DoubleRegs:$Rxx),
1990 (ins DoubleRegs:$dst2, DoubleRegs:$Rss, IntRegs:$Rt),
1991 "$Rxx += vrcnegh($Rss, $Rt)", [],
1992 "$dst2 = $Rxx", S_3op_tc_3x_SLOT23> {
1997 let IClass = 0b1100;
1999 let Inst{27-21} = 0b1011001;
2000 let Inst{20-16} = Rss;
2002 let Inst{12-8} = Rt;
2003 let Inst{7-5} = 0b111;
2004 let Inst{4-0} = Rxx;
2008 let isCodeGenOnly = 0 in
2009 def A4_bitspliti : T_S2op_2_di <"bitsplit", 0b110, 0b100>;
2011 // Arithmetic/Convergent round
2012 let isCodeGenOnly = 0 in
2013 def A4_cround_ri : T_S2op_2_ii <"cround", 0b111, 0b000>;
2015 let isCodeGenOnly = 0 in
2016 def A4_round_ri : T_S2op_2_ii <"round", 0b111, 0b100>;
2018 let Defs = [USR_OVF], isCodeGenOnly = 0 in
2019 def A4_round_ri_sat : T_S2op_2_ii <"round", 0b111, 0b110, 1>;
2021 // Logical-logical words.
2022 // Compound or-and -- Rx=or(Ru,and(Rx,#s10))
2023 let isExtentSigned = 1, hasNewValue = 1, isExtendable = 1, opExtentBits = 10,
2024 opExtendable = 3, isCodeGenOnly = 0 in
2026 ALU64Inst<(outs IntRegs:$Rx),
2027 (ins IntRegs:$Ru, IntRegs:$_src_, s10Ext:$s10),
2028 "$Rx = or($Ru, and($_src_, #$s10))" ,
2029 [(set (i32 IntRegs:$Rx),
2030 (or (i32 IntRegs:$Ru), (and (i32 IntRegs:$_src_), s10ExtPred:$s10)))] ,
2031 "$_src_ = $Rx", ALU64_tc_2_SLOT23> {
2036 let IClass = 0b1101;
2038 let Inst{27-22} = 0b101001;
2039 let Inst{20-16} = Rx;
2040 let Inst{21} = s10{9};
2041 let Inst{13-5} = s10{8-0};
2045 // Miscellaneous ALU64 instructions.
2047 let hasNewValue = 1, hasSideEffects = 0, isCodeGenOnly = 0 in
2048 def A4_modwrapu: ALU64Inst<(outs IntRegs:$Rd), (ins IntRegs:$Rs, IntRegs:$Rt),
2049 "$Rd = modwrap($Rs, $Rt)", [], "", ALU64_tc_2_SLOT23> {
2054 let IClass = 0b1101;
2055 let Inst{27-21} = 0b0011111;
2056 let Inst{20-16} = Rs;
2057 let Inst{12-8} = Rt;
2058 let Inst{7-5} = 0b111;
2062 let hasSideEffects = 0, isCodeGenOnly = 0 in
2063 def A4_bitsplit: ALU64Inst<(outs DoubleRegs:$Rd),
2064 (ins IntRegs:$Rs, IntRegs:$Rt),
2065 "$Rd = bitsplit($Rs, $Rt)", [], "", ALU64_tc_1_SLOT23> {
2070 let IClass = 0b1101;
2071 let Inst{27-24} = 0b0100;
2073 let Inst{20-16} = Rs;
2074 let Inst{12-8} = Rt;
2078 let isCodeGenOnly = 0 in {
2079 // Rx[&|]=xor(Rs,Rt)
2080 def M4_or_xor : T_MType_acc_rr < "|= xor", 0b110, 0b001, 0>;
2081 def M4_and_xor : T_MType_acc_rr < "&= xor", 0b010, 0b010, 0>;
2083 // Rx[&|^]=or(Rs,Rt)
2084 def M4_xor_or : T_MType_acc_rr < "^= or", 0b110, 0b011, 0>;
2086 let CextOpcode = "ORr_ORr" in
2087 def M4_or_or : T_MType_acc_rr < "|= or", 0b110, 0b000, 0>;
2088 def M4_and_or : T_MType_acc_rr < "&= or", 0b010, 0b001, 0>;
2090 // Rx[&|^]=and(Rs,Rt)
2091 def M4_xor_and : T_MType_acc_rr < "^= and", 0b110, 0b010, 0>;
2093 let CextOpcode = "ORr_ANDr" in
2094 def M4_or_and : T_MType_acc_rr < "|= and", 0b010, 0b011, 0>;
2095 def M4_and_and : T_MType_acc_rr < "&= and", 0b010, 0b000, 0>;
2097 // Rx[&|^]=and(Rs,~Rt)
2098 def M4_xor_andn : T_MType_acc_rr < "^= and", 0b001, 0b010, 0, [], 1>;
2099 def M4_or_andn : T_MType_acc_rr < "|= and", 0b001, 0b000, 0, [], 1>;
2100 def M4_and_andn : T_MType_acc_rr < "&= and", 0b001, 0b001, 0, [], 1>;
2103 // Compound or-or and or-and
2104 let isExtentSigned = 1, InputType = "imm", hasNewValue = 1, isExtendable = 1,
2105 opExtentBits = 10, opExtendable = 3 in
2106 class T_CompOR <string mnemonic, bits<2> MajOp, SDNode OpNode>
2107 : MInst_acc <(outs IntRegs:$Rx),
2108 (ins IntRegs:$src1, IntRegs:$Rs, s10Ext:$s10),
2109 "$Rx |= "#mnemonic#"($Rs, #$s10)",
2110 [(set (i32 IntRegs:$Rx), (or (i32 IntRegs:$src1),
2111 (OpNode (i32 IntRegs:$Rs), s10ExtPred:$s10)))],
2112 "$src1 = $Rx", ALU64_tc_2_SLOT23>, ImmRegRel {
2117 let IClass = 0b1101;
2119 let Inst{27-24} = 0b1010;
2120 let Inst{23-22} = MajOp;
2121 let Inst{20-16} = Rs;
2122 let Inst{21} = s10{9};
2123 let Inst{13-5} = s10{8-0};
2127 let CextOpcode = "ORr_ANDr", isCodeGenOnly = 0 in
2128 def S4_or_andi : T_CompOR <"and", 0b00, and>;
2130 let CextOpcode = "ORr_ORr", isCodeGenOnly = 0 in
2131 def S4_or_ori : T_CompOR <"or", 0b10, or>;
2134 // Rd=modwrap(Rs,Rt)
2136 // Rd=cround(Rs,#u5)
2138 // Rd=round(Rs,#u5)[:sat]
2139 // Rd=round(Rs,Rt)[:sat]
2140 // Vector reduce add unsigned halfwords
2141 // Rd=vraddh(Rss,Rtt)
2143 // Rdd=vaddb(Rss,Rtt)
2144 // Vector conditional negate
2145 // Rdd=vcnegh(Rss,Rt)
2146 // Rxx+=vrcnegh(Rss,Rt)
2147 // Vector maximum bytes
2148 // Rdd=vmaxb(Rtt,Rss)
2149 // Vector reduce maximum halfwords
2150 // Rxx=vrmaxh(Rss,Ru)
2151 // Rxx=vrmaxuh(Rss,Ru)
2152 // Vector reduce maximum words
2153 // Rxx=vrmaxuw(Rss,Ru)
2154 // Rxx=vrmaxw(Rss,Ru)
2155 // Vector minimum bytes
2156 // Rdd=vminb(Rtt,Rss)
2157 // Vector reduce minimum halfwords
2158 // Rxx=vrminh(Rss,Ru)
2159 // Rxx=vrminuh(Rss,Ru)
2160 // Vector reduce minimum words
2161 // Rxx=vrminuw(Rss,Ru)
2162 // Rxx=vrminw(Rss,Ru)
2163 // Vector subtract bytes
2164 // Rdd=vsubb(Rss,Rtt)
2166 //===----------------------------------------------------------------------===//
2168 //===----------------------------------------------------------------------===//
2170 //===----------------------------------------------------------------------===//
2172 //===----------------------------------------------------------------------===//
2175 let isCodeGenOnly = 0 in
2176 def S2_brevp : T_S2op_3 <"brev", 0b11, 0b110>;
2179 let isCodeGenOnly = 0 in {
2180 def S2_ct0p : T_COUNT_LEADING_64<"ct0", 0b111, 0b010>;
2181 def S2_ct1p : T_COUNT_LEADING_64<"ct1", 0b111, 0b100>;
2182 def S4_clbpnorm : T_COUNT_LEADING_64<"normamt", 0b011, 0b000>;
2185 def: Pat<(i32 (trunc (cttz (i64 DoubleRegs:$Rss)))),
2186 (S2_ct0p (i64 DoubleRegs:$Rss))>;
2187 def: Pat<(i32 (trunc (cttz (not (i64 DoubleRegs:$Rss))))),
2188 (S2_ct1p (i64 DoubleRegs:$Rss))>;
2190 let hasSideEffects = 0, hasNewValue = 1, isCodeGenOnly = 0 in
2191 def S4_clbaddi : SInst<(outs IntRegs:$Rd), (ins IntRegs:$Rs, s6Imm:$s6),
2192 "$Rd = add(clb($Rs), #$s6)", [], "", S_2op_tc_2_SLOT23> {
2196 let IClass = 0b1000;
2197 let Inst{27-24} = 0b1100;
2198 let Inst{23-21} = 0b001;
2199 let Inst{20-16} = Rs;
2200 let Inst{13-8} = s6;
2201 let Inst{7-5} = 0b000;
2205 let hasSideEffects = 0, hasNewValue = 1, isCodeGenOnly = 0 in
2206 def S4_clbpaddi : SInst<(outs IntRegs:$Rd), (ins DoubleRegs:$Rs, s6Imm:$s6),
2207 "$Rd = add(clb($Rs), #$s6)", [], "", S_2op_tc_2_SLOT23> {
2211 let IClass = 0b1000;
2212 let Inst{27-24} = 0b1000;
2213 let Inst{23-21} = 0b011;
2214 let Inst{20-16} = Rs;
2215 let Inst{13-8} = s6;
2216 let Inst{7-5} = 0b010;
2221 // Bit test/set/clear
2222 let isCodeGenOnly = 0 in {
2223 def S4_ntstbit_i : T_TEST_BIT_IMM<"!tstbit", 0b001>;
2224 def S4_ntstbit_r : T_TEST_BIT_REG<"!tstbit", 1>;
2227 let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm.
2228 def: Pat<(i1 (seteq (and (shl 1, u5ImmPred:$u5), (i32 IntRegs:$Rs)), 0)),
2229 (S4_ntstbit_i (i32 IntRegs:$Rs), u5ImmPred:$u5)>;
2230 def: Pat<(i1 (seteq (and (shl 1, (i32 IntRegs:$Rt)), (i32 IntRegs:$Rs)), 0)),
2231 (S4_ntstbit_r (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))>;
2234 // Add extra complexity to prefer these instructions over bitsset/bitsclr.
2235 // The reason is that tstbit/ntstbit can be folded into a compound instruction:
2236 // if ([!]tstbit(...)) jump ...
2237 let AddedComplexity = 100 in
2238 def: Pat<(i1 (setne (and (i32 IntRegs:$Rs), (i32 Set5ImmPred:$u5)), (i32 0))),
2239 (S2_tstbit_i (i32 IntRegs:$Rs), (BITPOS32 Set5ImmPred:$u5))>;
2241 let AddedComplexity = 100 in
2242 def: Pat<(i1 (seteq (and (i32 IntRegs:$Rs), (i32 Set5ImmPred:$u5)), (i32 0))),
2243 (S4_ntstbit_i (i32 IntRegs:$Rs), (BITPOS32 Set5ImmPred:$u5))>;
2245 let isCodeGenOnly = 0 in {
2246 def C4_nbitsset : T_TEST_BITS_REG<"!bitsset", 0b01, 1>;
2247 def C4_nbitsclr : T_TEST_BITS_REG<"!bitsclr", 0b10, 1>;
2248 def C4_nbitsclri : T_TEST_BITS_IMM<"!bitsclr", 0b10, 1>;
2251 // Do not increase complexity of these patterns. In the DAG, "cmp i8" may be
2252 // represented as a compare against "value & 0xFF", which is an exact match
2253 // for cmpb (same for cmph). The patterns below do not contain any additional
2254 // complexity that would make them preferable, and if they were actually used
2255 // instead of cmpb/cmph, they would result in a compare against register that
2256 // is loaded with the byte/half mask (i.e. 0xFF or 0xFFFF).
2257 def: Pat<(i1 (setne (and I32:$Rs, u6ImmPred:$u6), 0)),
2258 (C4_nbitsclri I32:$Rs, u6ImmPred:$u6)>;
2259 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), 0)),
2260 (C4_nbitsclr I32:$Rs, I32:$Rt)>;
2261 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), I32:$Rt)),
2262 (C4_nbitsset I32:$Rs, I32:$Rt)>;
2264 //===----------------------------------------------------------------------===//
2266 //===----------------------------------------------------------------------===//
2268 //===----------------------------------------------------------------------===//
2270 //===----------------------------------------------------------------------===//
2272 // Rd=add(#u6,mpyi(Rs,#U6)) -- Multiply by immed and add immed.
2274 let hasNewValue = 1, isExtendable = 1, opExtentBits = 6, opExtendable = 1,
2275 isCodeGenOnly = 0 in
2276 def M4_mpyri_addi : MInst<(outs IntRegs:$Rd),
2277 (ins u6Ext:$u6, IntRegs:$Rs, u6Imm:$U6),
2278 "$Rd = add(#$u6, mpyi($Rs, #$U6))" ,
2279 [(set (i32 IntRegs:$Rd),
2280 (add (mul (i32 IntRegs:$Rs), u6ImmPred:$U6),
2281 u6ExtPred:$u6))] ,"",ALU64_tc_3x_SLOT23> {
2287 let IClass = 0b1101;
2289 let Inst{27-24} = 0b1000;
2290 let Inst{23} = U6{5};
2291 let Inst{22-21} = u6{5-4};
2292 let Inst{20-16} = Rs;
2293 let Inst{13} = u6{3};
2294 let Inst{12-8} = Rd;
2295 let Inst{7-5} = u6{2-0};
2296 let Inst{4-0} = U6{4-0};
2299 // Rd=add(#u6,mpyi(Rs,Rt))
2300 let CextOpcode = "ADD_MPY", InputType = "imm", hasNewValue = 1,
2301 isExtendable = 1, opExtentBits = 6, opExtendable = 1, isCodeGenOnly = 0 in
2302 def M4_mpyrr_addi : MInst <(outs IntRegs:$Rd),
2303 (ins u6Ext:$u6, IntRegs:$Rs, IntRegs:$Rt),
2304 "$Rd = add(#$u6, mpyi($Rs, $Rt))" ,
2305 [(set (i32 IntRegs:$Rd),
2306 (add (mul (i32 IntRegs:$Rs), (i32 IntRegs:$Rt)), u6ExtPred:$u6))],
2307 "", ALU64_tc_3x_SLOT23>, ImmRegRel {
2313 let IClass = 0b1101;
2315 let Inst{27-23} = 0b01110;
2316 let Inst{22-21} = u6{5-4};
2317 let Inst{20-16} = Rs;
2318 let Inst{13} = u6{3};
2319 let Inst{12-8} = Rt;
2320 let Inst{7-5} = u6{2-0};
2324 let hasNewValue = 1 in
2325 class T_AddMpy <bit MajOp, PatLeaf ImmPred, dag ins>
2326 : ALU64Inst <(outs IntRegs:$dst), ins,
2327 "$dst = add($src1, mpyi("#!if(MajOp,"$src3, #$src2))",
2329 [(set (i32 IntRegs:$dst),
2330 (add (i32 IntRegs:$src1), (mul (i32 IntRegs:$src3), ImmPred:$src2)))],
2331 "", ALU64_tc_3x_SLOT23> {
2337 let IClass = 0b1101;
2339 bits<6> ImmValue = !if(MajOp, src2{5-0}, src2{7-2});
2341 let Inst{27-24} = 0b1111;
2342 let Inst{23} = MajOp;
2343 let Inst{22-21} = ImmValue{5-4};
2344 let Inst{20-16} = src3;
2345 let Inst{13} = ImmValue{3};
2346 let Inst{12-8} = dst;
2347 let Inst{7-5} = ImmValue{2-0};
2348 let Inst{4-0} = src1;
2351 let isCodeGenOnly = 0 in
2352 def M4_mpyri_addr_u2 : T_AddMpy<0b0, u6_2ImmPred,
2353 (ins IntRegs:$src1, u6_2Imm:$src2, IntRegs:$src3)>;
2355 let isExtendable = 1, opExtentBits = 6, opExtendable = 3,
2356 CextOpcode = "ADD_MPY", InputType = "imm", isCodeGenOnly = 0 in
2357 def M4_mpyri_addr : T_AddMpy<0b1, u6ExtPred,
2358 (ins IntRegs:$src1, IntRegs:$src3, u6Ext:$src2)>, ImmRegRel;
2360 // Rx=add(Ru,mpyi(Rx,Rs))
2361 let validSubTargets = HasV4SubT, CextOpcode = "ADD_MPY", InputType = "reg",
2362 hasNewValue = 1, isCodeGenOnly = 0 in
2363 def M4_mpyrr_addr: MInst_acc <(outs IntRegs:$Rx),
2364 (ins IntRegs:$Ru, IntRegs:$_src_, IntRegs:$Rs),
2365 "$Rx = add($Ru, mpyi($_src_, $Rs))",
2366 [(set (i32 IntRegs:$Rx), (add (i32 IntRegs:$Ru),
2367 (mul (i32 IntRegs:$_src_), (i32 IntRegs:$Rs))))],
2368 "$_src_ = $Rx", M_tc_3x_SLOT23>, ImmRegRel {
2373 let IClass = 0b1110;
2375 let Inst{27-21} = 0b0011000;
2376 let Inst{12-8} = Rx;
2378 let Inst{20-16} = Rs;
2381 // Rd=add(##,mpyi(Rs,#U6))
2382 def : Pat <(add (mul (i32 IntRegs:$src2), u6ImmPred:$src3),
2383 (HexagonCONST32 tglobaladdr:$src1)),
2384 (i32 (M4_mpyri_addi tglobaladdr:$src1, IntRegs:$src2,
2387 // Rd=add(##,mpyi(Rs,Rt))
2388 def : Pat <(add (mul (i32 IntRegs:$src2), (i32 IntRegs:$src3)),
2389 (HexagonCONST32 tglobaladdr:$src1)),
2390 (i32 (M4_mpyrr_addi tglobaladdr:$src1, IntRegs:$src2,
2393 // Vector reduce multiply word by signed half (32x16)
2394 //Rdd=vrmpyweh(Rss,Rtt)[:<<1]
2395 let isCodeGenOnly = 0 in {
2396 def M4_vrmpyeh_s0 : T_M2_vmpy<"vrmpyweh", 0b010, 0b100, 0, 0, 0>;
2397 def M4_vrmpyeh_s1 : T_M2_vmpy<"vrmpyweh", 0b110, 0b100, 1, 0, 0>;
2400 //Rdd=vrmpywoh(Rss,Rtt)[:<<1]
2401 let isCodeGenOnly = 0 in {
2402 def M4_vrmpyoh_s0 : T_M2_vmpy<"vrmpywoh", 0b001, 0b010, 0, 0, 0>;
2403 def M4_vrmpyoh_s1 : T_M2_vmpy<"vrmpywoh", 0b101, 0b010, 1, 0, 0>;
2405 //Rdd+=vrmpyweh(Rss,Rtt)[:<<1]
2406 let isCodeGenOnly = 0 in {
2407 def M4_vrmpyeh_acc_s0: T_M2_vmpy_acc<"vrmpyweh", 0b001, 0b110, 0, 0>;
2408 def M4_vrmpyeh_acc_s1: T_M2_vmpy_acc<"vrmpyweh", 0b101, 0b110, 1, 0>;
2411 //Rdd=vrmpywoh(Rss,Rtt)[:<<1]
2412 let isCodeGenOnly = 0 in {
2413 def M4_vrmpyoh_acc_s0: T_M2_vmpy_acc<"vrmpywoh", 0b011, 0b110, 0, 0>;
2414 def M4_vrmpyoh_acc_s1: T_M2_vmpy_acc<"vrmpywoh", 0b111, 0b110, 1, 0>;
2417 // Vector multiply halfwords, signed by unsigned
2418 // Rdd=vmpyhsu(Rs,Rt)[:<<]:sat
2419 let isCodeGenOnly = 0 in {
2420 def M2_vmpy2su_s0 : T_XTYPE_mpy64 < "vmpyhsu", 0b000, 0b111, 1, 0, 0>;
2421 def M2_vmpy2su_s1 : T_XTYPE_mpy64 < "vmpyhsu", 0b100, 0b111, 1, 1, 0>;
2424 // Rxx+=vmpyhsu(Rs,Rt)[:<<1]:sat
2425 let isCodeGenOnly = 0 in {
2426 def M2_vmac2su_s0 : T_XTYPE_mpy64_acc < "vmpyhsu", "+", 0b011, 0b101, 1, 0, 0>;
2427 def M2_vmac2su_s1 : T_XTYPE_mpy64_acc < "vmpyhsu", "+", 0b111, 0b101, 1, 1, 0>;
2430 // Vector polynomial multiply halfwords
2431 // Rdd=vpmpyh(Rs,Rt)
2432 let isCodeGenOnly = 0 in
2433 def M4_vpmpyh : T_XTYPE_mpy64 < "vpmpyh", 0b110, 0b111, 0, 0, 0>;
2435 // Rxx^=vpmpyh(Rs,Rt)
2436 let isCodeGenOnly = 0 in
2437 def M4_vpmpyh_acc : T_XTYPE_mpy64_acc < "vpmpyh", "^", 0b101, 0b111, 0, 0, 0>;
2439 // Polynomial multiply words
2441 let isCodeGenOnly = 0 in
2442 def M4_pmpyw : T_XTYPE_mpy64 < "pmpyw", 0b010, 0b111, 0, 0, 0>;
2444 // Rxx^=pmpyw(Rs,Rt)
2445 let isCodeGenOnly = 0 in
2446 def M4_pmpyw_acc : T_XTYPE_mpy64_acc < "pmpyw", "^", 0b001, 0b111, 0, 0, 0>;
2448 //===----------------------------------------------------------------------===//
2450 //===----------------------------------------------------------------------===//
2453 //===----------------------------------------------------------------------===//
2454 // ALU64/Vector compare
2455 //===----------------------------------------------------------------------===//
2456 //===----------------------------------------------------------------------===//
2457 // Template class for vector compare
2458 //===----------------------------------------------------------------------===//
2460 let hasSideEffects = 0 in
2461 class T_vcmpImm <string Str, bits<2> cmpOp, bits<2> minOp, Operand ImmOprnd>
2462 : ALU64_rr <(outs PredRegs:$Pd),
2463 (ins DoubleRegs:$Rss, ImmOprnd:$Imm),
2464 "$Pd = "#Str#"($Rss, #$Imm)",
2465 [], "", ALU64_tc_2early_SLOT23> {
2470 let ImmBits{6-0} = Imm{6-0};
2471 let ImmBits{7} = !if (!eq(cmpOp,0b10), 0b0, Imm{7}); // 0 for vcmp[bhw].gtu
2473 let IClass = 0b1101;
2475 let Inst{27-24} = 0b1100;
2476 let Inst{22-21} = cmpOp;
2477 let Inst{20-16} = Rss;
2478 let Inst{12-5} = ImmBits;
2479 let Inst{4-3} = minOp;
2483 // Vector compare bytes
2484 let isCodeGenOnly = 0 in
2485 def A4_vcmpbgt : T_vcmp <"vcmpb.gt", 0b1010>;
2486 def: T_vcmp_pat<A4_vcmpbgt, setgt, v8i8>;
2488 let AsmString = "$Pd = any8(vcmpb.eq($Rss, $Rtt))" in
2489 let isCodeGenOnly = 0 in
2490 def A4_vcmpbeq_any : T_vcmp <"any8(vcmpb.gt", 0b1000>;
2492 let isCodeGenOnly = 0 in {
2493 def A4_vcmpbeqi : T_vcmpImm <"vcmpb.eq", 0b00, 0b00, u8Imm>;
2494 def A4_vcmpbgti : T_vcmpImm <"vcmpb.gt", 0b01, 0b00, s8Imm>;
2495 def A4_vcmpbgtui : T_vcmpImm <"vcmpb.gtu", 0b10, 0b00, u7Imm>;
2498 // Vector compare halfwords
2499 let isCodeGenOnly = 0 in {
2500 def A4_vcmpheqi : T_vcmpImm <"vcmph.eq", 0b00, 0b01, s8Imm>;
2501 def A4_vcmphgti : T_vcmpImm <"vcmph.gt", 0b01, 0b01, s8Imm>;
2502 def A4_vcmphgtui : T_vcmpImm <"vcmph.gtu", 0b10, 0b01, u7Imm>;
2505 // Vector compare words
2506 let isCodeGenOnly = 0 in {
2507 def A4_vcmpweqi : T_vcmpImm <"vcmpw.eq", 0b00, 0b10, s8Imm>;
2508 def A4_vcmpwgti : T_vcmpImm <"vcmpw.gt", 0b01, 0b10, s8Imm>;
2509 def A4_vcmpwgtui : T_vcmpImm <"vcmpw.gtu", 0b10, 0b10, u7Imm>;
2512 //===----------------------------------------------------------------------===//
2514 //===----------------------------------------------------------------------===//
2515 // Shift by immediate and accumulate/logical.
2516 // Rx=add(#u8,asl(Rx,#U5)) Rx=add(#u8,lsr(Rx,#U5))
2517 // Rx=sub(#u8,asl(Rx,#U5)) Rx=sub(#u8,lsr(Rx,#U5))
2518 // Rx=and(#u8,asl(Rx,#U5)) Rx=and(#u8,lsr(Rx,#U5))
2519 // Rx=or(#u8,asl(Rx,#U5)) Rx=or(#u8,lsr(Rx,#U5))
2520 let isExtendable = 1, opExtendable = 1, isExtentSigned = 0, opExtentBits = 8,
2521 hasNewValue = 1, opNewValue = 0, validSubTargets = HasV4SubT in
2522 class T_S4_ShiftOperate<string MnOp, string MnSh, SDNode Op, SDNode Sh,
2523 bit asl_lsr, bits<2> MajOp, InstrItinClass Itin>
2524 : MInst_acc<(outs IntRegs:$Rd), (ins u8Ext:$u8, IntRegs:$Rx, u5Imm:$U5),
2525 "$Rd = "#MnOp#"(#$u8, "#MnSh#"($Rx, #$U5))",
2526 [(set (i32 IntRegs:$Rd),
2527 (Op (Sh I32:$Rx, u5ImmPred:$U5), u8ExtPred:$u8))],
2528 "$Rd = $Rx", Itin> {
2535 let IClass = 0b1101;
2536 let Inst{27-24} = 0b1110;
2537 let Inst{23-21} = u8{7-5};
2538 let Inst{20-16} = Rd;
2539 let Inst{13} = u8{4};
2540 let Inst{12-8} = U5;
2541 let Inst{7-5} = u8{3-1};
2542 let Inst{4} = asl_lsr;
2543 let Inst{3} = u8{0};
2544 let Inst{2-1} = MajOp;
2547 multiclass T_ShiftOperate<string mnemonic, SDNode Op, bits<2> MajOp,
2548 InstrItinClass Itin> {
2549 def _asl_ri : T_S4_ShiftOperate<mnemonic, "asl", Op, shl, 0, MajOp, Itin>;
2550 def _lsr_ri : T_S4_ShiftOperate<mnemonic, "lsr", Op, srl, 1, MajOp, Itin>;
2553 let AddedComplexity = 200, isCodeGenOnly = 0 in {
2554 defm S4_addi : T_ShiftOperate<"add", add, 0b10, ALU64_tc_2_SLOT23>;
2555 defm S4_andi : T_ShiftOperate<"and", and, 0b00, ALU64_tc_2_SLOT23>;
2558 let AddedComplexity = 30, isCodeGenOnly = 0 in
2559 defm S4_ori : T_ShiftOperate<"or", or, 0b01, ALU64_tc_1_SLOT23>;
2561 let isCodeGenOnly = 0 in
2562 defm S4_subi : T_ShiftOperate<"sub", sub, 0b11, ALU64_tc_1_SLOT23>;
2564 // Vector conditional negate
2565 // Rdd=vcnegh(Rss,Rt)
2566 let Defs = [USR_OVF], Itinerary = S_3op_tc_2_SLOT23, isCodeGenOnly = 0 in
2567 def S2_vcnegh : T_S3op_shiftVect < "vcnegh", 0b11, 0b01>;
2569 // Rd=[cround|round](Rs,Rt)
2570 let hasNewValue = 1, Itinerary = S_3op_tc_2_SLOT23, isCodeGenOnly = 0 in {
2571 def A4_cround_rr : T_S3op_3 < "cround", IntRegs, 0b11, 0b00>;
2572 def A4_round_rr : T_S3op_3 < "round", IntRegs, 0b11, 0b10>;
2575 // Rd=round(Rs,Rt):sat
2576 let hasNewValue = 1, Defs = [USR_OVF], Itinerary = S_3op_tc_2_SLOT23,
2577 isCodeGenOnly = 0 in
2578 def A4_round_rr_sat : T_S3op_3 < "round", IntRegs, 0b11, 0b11, 1>;
2580 // Rd=[cmpyiwh|cmpyrwh](Rss,Rt):<<1:rnd:sat
2581 let Defs = [USR_OVF], Itinerary = S_3op_tc_3x_SLOT23, isCodeGenOnly = 0 in {
2582 def M4_cmpyi_wh : T_S3op_8<"cmpyiwh", 0b100, 1, 1, 1>;
2583 def M4_cmpyr_wh : T_S3op_8<"cmpyrwh", 0b110, 1, 1, 1>;
2586 // Rdd=[add|sub](Rss,Rtt,Px):carry
2587 let isPredicateLate = 1, hasSideEffects = 0 in
2588 class T_S3op_carry <string mnemonic, bits<3> MajOp>
2589 : SInst < (outs DoubleRegs:$Rdd, PredRegs:$Px),
2590 (ins DoubleRegs:$Rss, DoubleRegs:$Rtt, PredRegs:$Pu),
2591 "$Rdd = "#mnemonic#"($Rss, $Rtt, $Pu):carry",
2592 [], "$Px = $Pu", S_3op_tc_1_SLOT23 > {
2598 let IClass = 0b1100;
2600 let Inst{27-24} = 0b0010;
2601 let Inst{23-21} = MajOp;
2602 let Inst{20-16} = Rss;
2603 let Inst{12-8} = Rtt;
2605 let Inst{4-0} = Rdd;
2608 let isCodeGenOnly = 0 in {
2609 def A4_addp_c : T_S3op_carry < "add", 0b110 >;
2610 def A4_subp_c : T_S3op_carry < "sub", 0b111 >;
2613 let Itinerary = S_3op_tc_3_SLOT23, hasSideEffects = 0 in
2614 class T_S3op_6 <string mnemonic, bits<3> MinOp, bit isUnsigned>
2615 : SInst <(outs DoubleRegs:$Rxx),
2616 (ins DoubleRegs:$dst2, DoubleRegs:$Rss, IntRegs:$Ru),
2617 "$Rxx = "#mnemonic#"($Rss, $Ru)" ,
2618 [] , "$dst2 = $Rxx"> {
2623 let IClass = 0b1100;
2625 let Inst{27-21} = 0b1011001;
2626 let Inst{20-16} = Rss;
2627 let Inst{13} = isUnsigned;
2628 let Inst{12-8} = Rxx;
2629 let Inst{7-5} = MinOp;
2633 // Vector reduce maximum halfwords
2634 // Rxx=vrmax[u]h(Rss,Ru)
2635 let isCodeGenOnly = 0 in {
2636 def A4_vrmaxh : T_S3op_6 < "vrmaxh", 0b001, 0>;
2637 def A4_vrmaxuh : T_S3op_6 < "vrmaxuh", 0b001, 1>;
2639 // Vector reduce maximum words
2640 // Rxx=vrmax[u]w(Rss,Ru)
2641 let isCodeGenOnly = 0 in {
2642 def A4_vrmaxw : T_S3op_6 < "vrmaxw", 0b010, 0>;
2643 def A4_vrmaxuw : T_S3op_6 < "vrmaxuw", 0b010, 1>;
2645 // Vector reduce minimum halfwords
2646 // Rxx=vrmin[u]h(Rss,Ru)
2647 let isCodeGenOnly = 0 in {
2648 def A4_vrminh : T_S3op_6 < "vrminh", 0b101, 0>;
2649 def A4_vrminuh : T_S3op_6 < "vrminuh", 0b101, 1>;
2652 // Vector reduce minimum words
2653 // Rxx=vrmin[u]w(Rss,Ru)
2654 let isCodeGenOnly = 0 in {
2655 def A4_vrminw : T_S3op_6 < "vrminw", 0b110, 0>;
2656 def A4_vrminuw : T_S3op_6 < "vrminuw", 0b110, 1>;
2659 // Shift an immediate left by register amount.
2660 let hasNewValue = 1, hasSideEffects = 0, isCodeGenOnly = 0 in
2661 def S4_lsli: SInst <(outs IntRegs:$Rd), (ins s6Imm:$s6, IntRegs:$Rt),
2662 "$Rd = lsl(#$s6, $Rt)" ,
2663 [(set (i32 IntRegs:$Rd), (shl s6ImmPred:$s6,
2664 (i32 IntRegs:$Rt)))],
2665 "", S_3op_tc_1_SLOT23> {
2670 let IClass = 0b1100;
2672 let Inst{27-22} = 0b011010;
2673 let Inst{20-16} = s6{5-1};
2674 let Inst{12-8} = Rt;
2675 let Inst{7-6} = 0b11;
2677 let Inst{5} = s6{0};
2680 //===----------------------------------------------------------------------===//
2682 //===----------------------------------------------------------------------===//
2684 //===----------------------------------------------------------------------===//
2685 // MEMOP: Word, Half, Byte
2686 //===----------------------------------------------------------------------===//
2688 def MEMOPIMM : SDNodeXForm<imm, [{
2689 // Call the transformation function XformM5ToU5Imm to get the negative
2690 // immediate's positive counterpart.
2691 int32_t imm = N->getSExtValue();
2692 return XformM5ToU5Imm(imm);
2695 def MEMOPIMM_HALF : SDNodeXForm<imm, [{
2696 // -1 .. -31 represented as 65535..65515
2697 // assigning to a short restores our desired signed value.
2698 // Call the transformation function XformM5ToU5Imm to get the negative
2699 // immediate's positive counterpart.
2700 int16_t imm = N->getSExtValue();
2701 return XformM5ToU5Imm(imm);
2704 def MEMOPIMM_BYTE : SDNodeXForm<imm, [{
2705 // -1 .. -31 represented as 255..235
2706 // assigning to a char restores our desired signed value.
2707 // Call the transformation function XformM5ToU5Imm to get the negative
2708 // immediate's positive counterpart.
2709 int8_t imm = N->getSExtValue();
2710 return XformM5ToU5Imm(imm);
2713 def SETMEMIMM : SDNodeXForm<imm, [{
2714 // Return the bit position we will set [0-31].
2716 int32_t imm = N->getSExtValue();
2717 return XformMskToBitPosU5Imm(imm);
2720 def CLRMEMIMM : SDNodeXForm<imm, [{
2721 // Return the bit position we will clear [0-31].
2723 // we bit negate the value first
2724 int32_t imm = ~(N->getSExtValue());
2725 return XformMskToBitPosU5Imm(imm);
2728 def SETMEMIMM_SHORT : SDNodeXForm<imm, [{
2729 // Return the bit position we will set [0-15].
2731 int16_t imm = N->getSExtValue();
2732 return XformMskToBitPosU4Imm(imm);
2735 def CLRMEMIMM_SHORT : SDNodeXForm<imm, [{
2736 // Return the bit position we will clear [0-15].
2738 // we bit negate the value first
2739 int16_t imm = ~(N->getSExtValue());
2740 return XformMskToBitPosU4Imm(imm);
2743 def SETMEMIMM_BYTE : SDNodeXForm<imm, [{
2744 // Return the bit position we will set [0-7].
2746 int8_t imm = N->getSExtValue();
2747 return XformMskToBitPosU3Imm(imm);
2750 def CLRMEMIMM_BYTE : SDNodeXForm<imm, [{
2751 // Return the bit position we will clear [0-7].
2753 // we bit negate the value first
2754 int8_t imm = ~(N->getSExtValue());
2755 return XformMskToBitPosU3Imm(imm);
2758 //===----------------------------------------------------------------------===//
2759 // Template class for MemOp instructions with the register value.
2760 //===----------------------------------------------------------------------===//
2761 class MemOp_rr_base <string opc, bits<2> opcBits, Operand ImmOp,
2762 string memOp, bits<2> memOpBits> :
2764 (ins IntRegs:$base, ImmOp:$offset, IntRegs:$delta),
2765 opc#"($base+#$offset)"#memOp#"$delta",
2767 Requires<[UseMEMOP]> {
2772 bits<6> offsetBits; // memb - u6:0 , memh - u6:1, memw - u6:2
2774 let offsetBits = !if (!eq(opcBits, 0b00), offset{5-0},
2775 !if (!eq(opcBits, 0b01), offset{6-1},
2776 !if (!eq(opcBits, 0b10), offset{7-2},0)));
2778 let opExtentAlign = opcBits;
2779 let IClass = 0b0011;
2780 let Inst{27-24} = 0b1110;
2781 let Inst{22-21} = opcBits;
2782 let Inst{20-16} = base;
2784 let Inst{12-7} = offsetBits;
2785 let Inst{6-5} = memOpBits;
2786 let Inst{4-0} = delta;
2789 //===----------------------------------------------------------------------===//
2790 // Template class for MemOp instructions with the immediate value.
2791 //===----------------------------------------------------------------------===//
2792 class MemOp_ri_base <string opc, bits<2> opcBits, Operand ImmOp,
2793 string memOp, bits<2> memOpBits> :
2795 (ins IntRegs:$base, ImmOp:$offset, u5Imm:$delta),
2796 opc#"($base+#$offset)"#memOp#"#$delta"
2797 #!if(memOpBits{1},")", ""), // clrbit, setbit - include ')'
2799 Requires<[UseMEMOP]> {
2804 bits<6> offsetBits; // memb - u6:0 , memh - u6:1, memw - u6:2
2806 let offsetBits = !if (!eq(opcBits, 0b00), offset{5-0},
2807 !if (!eq(opcBits, 0b01), offset{6-1},
2808 !if (!eq(opcBits, 0b10), offset{7-2},0)));
2810 let opExtentAlign = opcBits;
2811 let IClass = 0b0011;
2812 let Inst{27-24} = 0b1111;
2813 let Inst{22-21} = opcBits;
2814 let Inst{20-16} = base;
2816 let Inst{12-7} = offsetBits;
2817 let Inst{6-5} = memOpBits;
2818 let Inst{4-0} = delta;
2821 // multiclass to define MemOp instructions with register operand.
2822 multiclass MemOp_rr<string opc, bits<2> opcBits, Operand ImmOp> {
2823 def L4_add#NAME : MemOp_rr_base <opc, opcBits, ImmOp, " += ", 0b00>; // add
2824 def L4_sub#NAME : MemOp_rr_base <opc, opcBits, ImmOp, " -= ", 0b01>; // sub
2825 def L4_and#NAME : MemOp_rr_base <opc, opcBits, ImmOp, " &= ", 0b10>; // and
2826 def L4_or#NAME : MemOp_rr_base <opc, opcBits, ImmOp, " |= ", 0b11>; // or
2829 // multiclass to define MemOp instructions with immediate Operand.
2830 multiclass MemOp_ri<string opc, bits<2> opcBits, Operand ImmOp> {
2831 def L4_iadd#NAME : MemOp_ri_base <opc, opcBits, ImmOp, " += ", 0b00 >;
2832 def L4_isub#NAME : MemOp_ri_base <opc, opcBits, ImmOp, " -= ", 0b01 >;
2833 def L4_iand#NAME : MemOp_ri_base<opc, opcBits, ImmOp, " = clrbit(", 0b10>;
2834 def L4_ior#NAME : MemOp_ri_base<opc, opcBits, ImmOp, " = setbit(", 0b11>;
2837 multiclass MemOp_base <string opc, bits<2> opcBits, Operand ImmOp> {
2838 defm _#NAME : MemOp_rr <opc, opcBits, ImmOp>;
2839 defm _#NAME : MemOp_ri <opc, opcBits, ImmOp>;
2842 // Define MemOp instructions.
2843 let isExtendable = 1, opExtendable = 1, isExtentSigned = 0,
2844 validSubTargets =HasV4SubT in {
2845 let opExtentBits = 6, accessSize = ByteAccess, isCodeGenOnly = 0 in
2846 defm memopb_io : MemOp_base <"memb", 0b00, u6_0Ext>;
2848 let opExtentBits = 7, accessSize = HalfWordAccess, isCodeGenOnly = 0 in
2849 defm memoph_io : MemOp_base <"memh", 0b01, u6_1Ext>;
2851 let opExtentBits = 8, accessSize = WordAccess, isCodeGenOnly = 0 in
2852 defm memopw_io : MemOp_base <"memw", 0b10, u6_2Ext>;
2855 //===----------------------------------------------------------------------===//
2856 // Multiclass to define 'Def Pats' for ALU operations on the memory
2857 // Here value used for the ALU operation is an immediate value.
2858 // mem[bh](Rs+#0) += #U5
2859 // mem[bh](Rs+#u6) += #U5
2860 //===----------------------------------------------------------------------===//
2862 multiclass MemOpi_u5Pats <PatFrag ldOp, PatFrag stOp, PatLeaf ExtPred,
2863 InstHexagon MI, SDNode OpNode> {
2864 let AddedComplexity = 180 in
2865 def : Pat < (stOp (OpNode (ldOp IntRegs:$addr), u5ImmPred:$addend),
2867 (MI IntRegs:$addr, #0, u5ImmPred:$addend )>;
2869 let AddedComplexity = 190 in
2870 def : Pat <(stOp (OpNode (ldOp (add IntRegs:$base, ExtPred:$offset)),
2872 (add IntRegs:$base, ExtPred:$offset)),
2873 (MI IntRegs:$base, ExtPred:$offset, u5ImmPred:$addend)>;
2876 multiclass MemOpi_u5ALUOp<PatFrag ldOp, PatFrag stOp, PatLeaf ExtPred,
2877 InstHexagon addMI, InstHexagon subMI> {
2878 defm : MemOpi_u5Pats<ldOp, stOp, ExtPred, addMI, add>;
2879 defm : MemOpi_u5Pats<ldOp, stOp, ExtPred, subMI, sub>;
2882 multiclass MemOpi_u5ExtType<PatFrag ldOpByte, PatFrag ldOpHalf > {
2884 defm : MemOpi_u5ALUOp <ldOpHalf, truncstorei16, u6_1ExtPred,
2885 L4_iadd_memoph_io, L4_isub_memoph_io>;
2887 defm : MemOpi_u5ALUOp <ldOpByte, truncstorei8, u6ExtPred,
2888 L4_iadd_memopb_io, L4_isub_memopb_io>;
2891 let Predicates = [HasV4T, UseMEMOP] in {
2892 defm : MemOpi_u5ExtType<zextloadi8, zextloadi16>; // zero extend
2893 defm : MemOpi_u5ExtType<sextloadi8, sextloadi16>; // sign extend
2894 defm : MemOpi_u5ExtType<extloadi8, extloadi16>; // any extend
2897 defm : MemOpi_u5ALUOp <load, store, u6_2ExtPred, L4_iadd_memopw_io,
2901 //===----------------------------------------------------------------------===//
2902 // multiclass to define 'Def Pats' for ALU operations on the memory.
2903 // Here value used for the ALU operation is a negative value.
2904 // mem[bh](Rs+#0) += #m5
2905 // mem[bh](Rs+#u6) += #m5
2906 //===----------------------------------------------------------------------===//
2908 multiclass MemOpi_m5Pats <PatFrag ldOp, PatFrag stOp, PatLeaf extPred,
2909 PatLeaf immPred, ComplexPattern addrPred,
2910 SDNodeXForm xformFunc, InstHexagon MI> {
2911 let AddedComplexity = 190 in
2912 def : Pat <(stOp (add (ldOp IntRegs:$addr), immPred:$subend),
2914 (MI IntRegs:$addr, #0, (xformFunc immPred:$subend) )>;
2916 let AddedComplexity = 195 in
2917 def : Pat<(stOp (add (ldOp (add IntRegs:$base, extPred:$offset)),
2919 (add IntRegs:$base, extPred:$offset)),
2920 (MI IntRegs:$base, extPred:$offset, (xformFunc immPred:$subend))>;
2923 multiclass MemOpi_m5ExtType<PatFrag ldOpByte, PatFrag ldOpHalf > {
2925 defm : MemOpi_m5Pats <ldOpHalf, truncstorei16, u6_1ExtPred, m5HImmPred,
2926 ADDRriU6_1, MEMOPIMM_HALF, L4_isub_memoph_io>;
2928 defm : MemOpi_m5Pats <ldOpByte, truncstorei8, u6ExtPred, m5BImmPred,
2929 ADDRriU6_0, MEMOPIMM_BYTE, L4_isub_memopb_io>;
2932 let Predicates = [HasV4T, UseMEMOP] in {
2933 defm : MemOpi_m5ExtType<zextloadi8, zextloadi16>; // zero extend
2934 defm : MemOpi_m5ExtType<sextloadi8, sextloadi16>; // sign extend
2935 defm : MemOpi_m5ExtType<extloadi8, extloadi16>; // any extend
2938 defm : MemOpi_m5Pats <load, store, u6_2ExtPred, m5ImmPred,
2939 ADDRriU6_2, MEMOPIMM, L4_isub_memopw_io>;
2942 //===----------------------------------------------------------------------===//
2943 // Multiclass to define 'def Pats' for bit operations on the memory.
2944 // mem[bhw](Rs+#0) = [clrbit|setbit](#U5)
2945 // mem[bhw](Rs+#u6) = [clrbit|setbit](#U5)
2946 //===----------------------------------------------------------------------===//
2948 multiclass MemOpi_bitPats <PatFrag ldOp, PatFrag stOp, PatLeaf immPred,
2949 PatLeaf extPred, ComplexPattern addrPred,
2950 SDNodeXForm xformFunc, InstHexagon MI, SDNode OpNode> {
2952 // mem[bhw](Rs+#u6:[012]) = [clrbit|setbit](#U5)
2953 let AddedComplexity = 250 in
2954 def : Pat<(stOp (OpNode (ldOp (add IntRegs:$base, extPred:$offset)),
2956 (add IntRegs:$base, extPred:$offset)),
2957 (MI IntRegs:$base, extPred:$offset, (xformFunc immPred:$bitend))>;
2959 // mem[bhw](Rs+#0) = [clrbit|setbit](#U5)
2960 let AddedComplexity = 225 in
2961 def : Pat <(stOp (OpNode (ldOp (addrPred IntRegs:$addr, extPred:$offset)),
2963 (addrPred (i32 IntRegs:$addr), extPred:$offset)),
2964 (MI IntRegs:$addr, extPred:$offset, (xformFunc immPred:$bitend))>;
2967 multiclass MemOpi_bitExtType<PatFrag ldOpByte, PatFrag ldOpHalf > {
2969 defm : MemOpi_bitPats<ldOpByte, truncstorei8, Clr3ImmPred, u6ExtPred,
2970 ADDRriU6_0, CLRMEMIMM_BYTE, L4_iand_memopb_io, and>;
2972 defm : MemOpi_bitPats<ldOpByte, truncstorei8, Set3ImmPred, u6ExtPred,
2973 ADDRriU6_0, SETMEMIMM_BYTE, L4_ior_memopb_io, or>;
2974 // Half Word - clrbit
2975 defm : MemOpi_bitPats<ldOpHalf, truncstorei16, Clr4ImmPred, u6_1ExtPred,
2976 ADDRriU6_1, CLRMEMIMM_SHORT, L4_iand_memoph_io, and>;
2977 // Half Word - setbit
2978 defm : MemOpi_bitPats<ldOpHalf, truncstorei16, Set4ImmPred, u6_1ExtPred,
2979 ADDRriU6_1, SETMEMIMM_SHORT, L4_ior_memoph_io, or>;
2982 let Predicates = [HasV4T, UseMEMOP] in {
2983 // mem[bh](Rs+#0) = [clrbit|setbit](#U5)
2984 // mem[bh](Rs+#u6:[01]) = [clrbit|setbit](#U5)
2985 defm : MemOpi_bitExtType<zextloadi8, zextloadi16>; // zero extend
2986 defm : MemOpi_bitExtType<sextloadi8, sextloadi16>; // sign extend
2987 defm : MemOpi_bitExtType<extloadi8, extloadi16>; // any extend
2989 // memw(Rs+#0) = [clrbit|setbit](#U5)
2990 // memw(Rs+#u6:2) = [clrbit|setbit](#U5)
2991 defm : MemOpi_bitPats<load, store, Clr5ImmPred, u6_2ExtPred, ADDRriU6_2,
2992 CLRMEMIMM, L4_iand_memopw_io, and>;
2993 defm : MemOpi_bitPats<load, store, Set5ImmPred, u6_2ExtPred, ADDRriU6_2,
2994 SETMEMIMM, L4_ior_memopw_io, or>;
2997 //===----------------------------------------------------------------------===//
2998 // Multiclass to define 'def Pats' for ALU operations on the memory
2999 // where addend is a register.
3000 // mem[bhw](Rs+#0) [+-&|]= Rt
3001 // mem[bhw](Rs+#U6:[012]) [+-&|]= Rt
3002 //===----------------------------------------------------------------------===//
3004 multiclass MemOpr_Pats <PatFrag ldOp, PatFrag stOp, ComplexPattern addrPred,
3005 PatLeaf extPred, InstHexagon MI, SDNode OpNode> {
3006 let AddedComplexity = 141 in
3007 // mem[bhw](Rs+#0) [+-&|]= Rt
3008 def : Pat <(stOp (OpNode (ldOp (addrPred IntRegs:$addr, extPred:$offset)),
3009 (i32 IntRegs:$addend)),
3010 (addrPred (i32 IntRegs:$addr), extPred:$offset)),
3011 (MI IntRegs:$addr, extPred:$offset, (i32 IntRegs:$addend) )>;
3013 // mem[bhw](Rs+#U6:[012]) [+-&|]= Rt
3014 let AddedComplexity = 150 in
3015 def : Pat <(stOp (OpNode (ldOp (add IntRegs:$base, extPred:$offset)),
3016 (i32 IntRegs:$orend)),
3017 (add IntRegs:$base, extPred:$offset)),
3018 (MI IntRegs:$base, extPred:$offset, (i32 IntRegs:$orend) )>;
3021 multiclass MemOPr_ALUOp<PatFrag ldOp, PatFrag stOp,
3022 ComplexPattern addrPred, PatLeaf extPred,
3023 InstHexagon addMI, InstHexagon subMI,
3024 InstHexagon andMI, InstHexagon orMI > {
3026 defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, addMI, add>;
3027 defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, subMI, sub>;
3028 defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, andMI, and>;
3029 defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, orMI, or>;
3032 multiclass MemOPr_ExtType<PatFrag ldOpByte, PatFrag ldOpHalf > {
3034 defm : MemOPr_ALUOp <ldOpHalf, truncstorei16, ADDRriU6_1, u6_1ExtPred,
3035 L4_add_memoph_io, L4_sub_memoph_io,
3036 L4_and_memoph_io, L4_or_memoph_io>;
3038 defm : MemOPr_ALUOp <ldOpByte, truncstorei8, ADDRriU6_0, u6ExtPred,
3039 L4_add_memopb_io, L4_sub_memopb_io,
3040 L4_and_memopb_io, L4_or_memopb_io>;
3043 // Define 'def Pats' for MemOps with register addend.
3044 let Predicates = [HasV4T, UseMEMOP] in {
3046 defm : MemOPr_ExtType<zextloadi8, zextloadi16>; // zero extend
3047 defm : MemOPr_ExtType<sextloadi8, sextloadi16>; // sign extend
3048 defm : MemOPr_ExtType<extloadi8, extloadi16>; // any extend
3050 defm : MemOPr_ALUOp <load, store, ADDRriU6_2, u6_2ExtPred, L4_add_memopw_io,
3051 L4_sub_memopw_io, L4_and_memopw_io, L4_or_memopw_io >;
3054 //===----------------------------------------------------------------------===//
3056 //===----------------------------------------------------------------------===//
3058 // Hexagon V4 only supports these flavors of byte/half compare instructions:
3059 // EQ/GT/GTU. Other flavors like GE/GEU/LT/LTU/LE/LEU are not supported by
3060 // hardware. However, compiler can still implement these patterns through
3061 // appropriate patterns combinations based on current implemented patterns.
3062 // The implemented patterns are: EQ/GT/GTU.
3063 // Missing patterns are: GE/GEU/LT/LTU/LE/LEU.
3065 // Following instruction is not being extended as it results into the
3066 // incorrect code for negative numbers.
3067 // Pd=cmpb.eq(Rs,#u8)
3069 // p=!cmp.eq(r1,#s10)
3070 let isCodeGenOnly = 0 in {
3071 def C4_cmpneqi : T_CMP <"cmp.eq", 0b00, 1, s10Ext>;
3072 def C4_cmpltei : T_CMP <"cmp.gt", 0b01, 1, s10Ext>;
3073 def C4_cmplteui : T_CMP <"cmp.gtu", 0b10, 1, u9Ext>;
3076 def : T_CMP_pat <C4_cmpneqi, setne, s10ExtPred>;
3077 def : T_CMP_pat <C4_cmpltei, setle, s10ExtPred>;
3078 def : T_CMP_pat <C4_cmplteui, setule, u9ImmPred>;
3080 // rs <= rt -> !(rs > rt).
3082 def: Pat<(i1 (setle (i32 IntRegs:$src1), s10ExtPred:$src2)),
3083 (C2_not (C2_cmpgti IntRegs:$src1, s10ExtPred:$src2))>;
3084 // (C4_cmpltei IntRegs:$src1, s10ExtPred:$src2)>;
3086 // Map cmplt(Rs, Imm) -> !cmpgt(Rs, Imm-1).
3087 def: Pat<(i1 (setlt (i32 IntRegs:$src1), s8ExtPred:$src2)),
3088 (C4_cmpltei IntRegs:$src1, (DEC_CONST_SIGNED s8ExtPred:$src2))>;
3090 // rs != rt -> !(rs == rt).
3091 def: Pat<(i1 (setne (i32 IntRegs:$src1), s10ExtPred:$src2)),
3092 (C4_cmpneqi IntRegs:$src1, s10ExtPred:$src2)>;
3094 // SDNode for converting immediate C to C-1.
3095 def DEC_CONST_BYTE : SDNodeXForm<imm, [{
3096 // Return the byte immediate const-1 as an SDNode.
3097 int32_t imm = N->getSExtValue();
3098 return XformU7ToU7M1Imm(imm);
3102 // zext( seteq ( and(Rs, 255), u8))
3104 // Pd=cmpb.eq(Rs, #u8)
3105 // if (Pd.new) Rd=#1
3106 // if (!Pd.new) Rd=#0
3107 def : Pat <(i32 (zext (i1 (seteq (i32 (and (i32 IntRegs:$Rs), 255)),
3109 (i32 (TFR_condset_ii (i1 (A4_cmpbeqi (i32 IntRegs:$Rs),
3115 // zext( setne ( and(Rs, 255), u8))
3117 // Pd=cmpb.eq(Rs, #u8)
3118 // if (Pd.new) Rd=#0
3119 // if (!Pd.new) Rd=#1
3120 def : Pat <(i32 (zext (i1 (setne (i32 (and (i32 IntRegs:$Rs), 255)),
3122 (i32 (TFR_condset_ii (i1 (A4_cmpbeqi (i32 IntRegs:$Rs),
3128 // zext( seteq (Rs, and(Rt, 255)))
3130 // Pd=cmpb.eq(Rs, Rt)
3131 // if (Pd.new) Rd=#1
3132 // if (!Pd.new) Rd=#0
3133 def : Pat <(i32 (zext (i1 (seteq (i32 IntRegs:$Rt),
3134 (i32 (and (i32 IntRegs:$Rs), 255)))))),
3135 (i32 (TFR_condset_ii (i1 (A4_cmpbeq (i32 IntRegs:$Rs),
3136 (i32 IntRegs:$Rt))),
3141 // zext( setne (Rs, and(Rt, 255)))
3143 // Pd=cmpb.eq(Rs, Rt)
3144 // if (Pd.new) Rd=#0
3145 // if (!Pd.new) Rd=#1
3146 def : Pat <(i32 (zext (i1 (setne (i32 IntRegs:$Rt),
3147 (i32 (and (i32 IntRegs:$Rs), 255)))))),
3148 (i32 (TFR_condset_ii (i1 (A4_cmpbeq (i32 IntRegs:$Rs),
3149 (i32 IntRegs:$Rt))),
3154 // zext( setugt ( and(Rs, 255), u8))
3156 // Pd=cmpb.gtu(Rs, #u8)
3157 // if (Pd.new) Rd=#1
3158 // if (!Pd.new) Rd=#0
3159 def : Pat <(i32 (zext (i1 (setugt (i32 (and (i32 IntRegs:$Rs), 255)),
3161 (i32 (TFR_condset_ii (i1 (A4_cmpbgtui (i32 IntRegs:$Rs),
3167 // zext( setugt ( and(Rs, 254), u8))
3169 // Pd=cmpb.gtu(Rs, #u8)
3170 // if (Pd.new) Rd=#1
3171 // if (!Pd.new) Rd=#0
3172 def : Pat <(i32 (zext (i1 (setugt (i32 (and (i32 IntRegs:$Rs), 254)),
3174 (i32 (TFR_condset_ii (i1 (A4_cmpbgtui (i32 IntRegs:$Rs),
3180 // zext( setult ( Rs, Rt))
3182 // Pd=cmp.ltu(Rs, Rt)
3183 // if (Pd.new) Rd=#1
3184 // if (!Pd.new) Rd=#0
3185 // cmp.ltu(Rs, Rt) -> cmp.gtu(Rt, Rs)
3186 def : Pat <(i32 (zext (i1 (setult (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3187 (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rt),
3188 (i32 IntRegs:$Rs))),
3193 // zext( setlt ( Rs, Rt))
3195 // Pd=cmp.lt(Rs, Rt)
3196 // if (Pd.new) Rd=#1
3197 // if (!Pd.new) Rd=#0
3198 // cmp.lt(Rs, Rt) -> cmp.gt(Rt, Rs)
3199 def : Pat <(i32 (zext (i1 (setlt (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3200 (i32 (TFR_condset_ii (i1 (C2_cmpgt (i32 IntRegs:$Rt),
3201 (i32 IntRegs:$Rs))),
3206 // zext( setugt ( Rs, Rt))
3208 // Pd=cmp.gtu(Rs, Rt)
3209 // if (Pd.new) Rd=#1
3210 // if (!Pd.new) Rd=#0
3211 def : Pat <(i32 (zext (i1 (setugt (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3212 (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rs),
3213 (i32 IntRegs:$Rt))),
3217 // This pattern interefers with coremark performance, not implementing at this
3220 // zext( setgt ( Rs, Rt))
3222 // Pd=cmp.gt(Rs, Rt)
3223 // if (Pd.new) Rd=#1
3224 // if (!Pd.new) Rd=#0
3227 // zext( setuge ( Rs, Rt))
3229 // Pd=cmp.ltu(Rs, Rt)
3230 // if (Pd.new) Rd=#0
3231 // if (!Pd.new) Rd=#1
3232 // cmp.ltu(Rs, Rt) -> cmp.gtu(Rt, Rs)
3233 def : Pat <(i32 (zext (i1 (setuge (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3234 (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rt),
3235 (i32 IntRegs:$Rs))),
3240 // zext( setge ( Rs, Rt))
3242 // Pd=cmp.lt(Rs, Rt)
3243 // if (Pd.new) Rd=#0
3244 // if (!Pd.new) Rd=#1
3245 // cmp.lt(Rs, Rt) -> cmp.gt(Rt, Rs)
3246 def : Pat <(i32 (zext (i1 (setge (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3247 (i32 (TFR_condset_ii (i1 (C2_cmpgt (i32 IntRegs:$Rt),
3248 (i32 IntRegs:$Rs))),
3253 // zext( setule ( Rs, Rt))
3255 // Pd=cmp.gtu(Rs, Rt)
3256 // if (Pd.new) Rd=#0
3257 // if (!Pd.new) Rd=#1
3258 def : Pat <(i32 (zext (i1 (setule (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3259 (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rs),
3260 (i32 IntRegs:$Rt))),
3265 // zext( setle ( Rs, Rt))
3267 // Pd=cmp.gt(Rs, Rt)
3268 // if (Pd.new) Rd=#0
3269 // if (!Pd.new) Rd=#1
3270 def : Pat <(i32 (zext (i1 (setle (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3271 (i32 (TFR_condset_ii (i1 (C2_cmpgt (i32 IntRegs:$Rs),
3272 (i32 IntRegs:$Rt))),
3277 // zext( setult ( and(Rs, 255), u8))
3278 // Use the isdigit transformation below
3280 // Generate code of the form 'mux_ii(cmpbgtu(Rdd, C-1),0,1)'
3281 // for C code of the form r = ((c>='0') & (c<='9')) ? 1 : 0;.
3282 // The isdigit transformation relies on two 'clever' aspects:
3283 // 1) The data type is unsigned which allows us to eliminate a zero test after
3284 // biasing the expression by 48. We are depending on the representation of
3285 // the unsigned types, and semantics.
3286 // 2) The front end has converted <= 9 into < 10 on entry to LLVM
3289 // retval = ((c>='0') & (c<='9')) ? 1 : 0;
3290 // The code is transformed upstream of llvm into
3291 // retval = (c-48) < 10 ? 1 : 0;
3292 let AddedComplexity = 139 in
3293 def : Pat <(i32 (zext (i1 (setult (i32 (and (i32 IntRegs:$src1), 255)),
3294 u7StrictPosImmPred:$src2)))),
3295 (i32 (C2_muxii (i1 (A4_cmpbgtui (i32 IntRegs:$src1),
3296 (DEC_CONST_BYTE u7StrictPosImmPred:$src2))),
3300 //===----------------------------------------------------------------------===//
3302 //===----------------------------------------------------------------------===//
3304 //===----------------------------------------------------------------------===//
3305 // Multiclass for DeallocReturn
3306 //===----------------------------------------------------------------------===//
3307 class L4_RETURN<string mnemonic, bit isNot, bit isPredNew, bit isTak>
3308 : LD0Inst<(outs), (ins PredRegs:$src),
3309 !if(isNot, "if (!$src", "if ($src")#
3310 !if(isPredNew, ".new) ", ") ")#mnemonic#
3311 !if(isPredNew, #!if(isTak,":t", ":nt"),""),
3312 [], "", LD_tc_3or4stall_SLOT0> {
3315 let BaseOpcode = "L4_RETURN";
3316 let isPredicatedFalse = isNot;
3317 let isPredicatedNew = isPredNew;
3318 let isTaken = isTak;
3319 let IClass = 0b1001;
3321 let Inst{27-16} = 0b011000011110;
3323 let Inst{13} = isNot;
3324 let Inst{12} = isTak;
3325 let Inst{11} = isPredNew;
3327 let Inst{9-8} = src;
3328 let Inst{4-0} = 0b11110;
3331 // Produce all predicated forms, p, !p, p.new, !p.new, :t, :nt
3332 multiclass L4_RETURN_PRED<string mnemonic, bit PredNot> {
3333 let isPredicated = 1 in {
3334 def _#NAME# : L4_RETURN <mnemonic, PredNot, 0, 1>;
3335 def _#NAME#new_pnt : L4_RETURN <mnemonic, PredNot, 1, 0>;
3336 def _#NAME#new_pt : L4_RETURN <mnemonic, PredNot, 1, 1>;
3340 multiclass LD_MISC_L4_RETURN<string mnemonic> {
3341 let isBarrier = 1, isPredicable = 1 in
3342 def NAME : LD0Inst <(outs), (ins), mnemonic, [], "",
3343 LD_tc_3or4stall_SLOT0> {
3344 let BaseOpcode = "L4_RETURN";
3345 let IClass = 0b1001;
3346 let Inst{27-16} = 0b011000011110;
3347 let Inst{13-10} = 0b0000;
3348 let Inst{4-0} = 0b11110;
3350 defm t : L4_RETURN_PRED<mnemonic, 0 >;
3351 defm f : L4_RETURN_PRED<mnemonic, 1 >;
3354 let isReturn = 1, isTerminator = 1,
3355 Defs = [R29, R30, R31, PC], Uses = [R30], hasSideEffects = 0,
3356 validSubTargets = HasV4SubT, isCodeGenOnly = 0 in
3357 defm L4_return: LD_MISC_L4_RETURN <"dealloc_return">, PredNewRel;
3359 // Restore registers and dealloc return function call.
3360 let isCall = 1, isBarrier = 1, isReturn = 1, isTerminator = 1,
3361 Defs = [R29, R30, R31, PC], isAsmParserOnly = 1 in {
3362 let validSubTargets = HasV4SubT in
3363 def RESTORE_DEALLOC_RET_JMP_V4 : JInst<(outs),
3364 (ins calltarget:$dst),
3370 // Restore registers and dealloc frame before a tail call.
3371 let isCall = 1, isBarrier = 1, isAsmParserOnly = 1,
3372 Defs = [R29, R30, R31, PC] in {
3373 let validSubTargets = HasV4SubT in
3374 def RESTORE_DEALLOC_BEFORE_TAILCALL_V4 : JInst<(outs),
3375 (ins calltarget:$dst),
3381 // Save registers function call.
3382 let isCall = 1, isBarrier = 1, isAsmParserOnly = 1,
3383 Uses = [R29, R31] in {
3384 def SAVE_REGISTERS_CALL_V4 : JInst<(outs),
3385 (ins calltarget:$dst),
3386 "call $dst // Save_calle_saved_registers",
3391 //===----------------------------------------------------------------------===//
3392 // Template class for non predicated store instructions with
3393 // GP-Relative or absolute addressing.
3394 //===----------------------------------------------------------------------===//
3395 let hasSideEffects = 0, isPredicable = 1, isNVStorable = 1 in
3396 class T_StoreAbsGP <string mnemonic, RegisterClass RC, Operand ImmOp,
3397 bits<2>MajOp, Operand AddrOp, bit isAbs, bit isHalf>
3398 : STInst<(outs), (ins AddrOp:$addr, RC:$src),
3399 mnemonic # !if(isAbs, "(##", "(#")#"$addr) = $src"#!if(isHalf, ".h",""),
3400 [], "", V2LDST_tc_st_SLOT01> {
3403 bits<16> offsetBits;
3405 string ImmOpStr = !cast<string>(ImmOp);
3406 let offsetBits = !if (!eq(ImmOpStr, "u16_3Imm"), addr{18-3},
3407 !if (!eq(ImmOpStr, "u16_2Imm"), addr{17-2},
3408 !if (!eq(ImmOpStr, "u16_1Imm"), addr{16-1},
3409 /* u16_0Imm */ addr{15-0})));
3410 let IClass = 0b0100;
3412 let Inst{26-25} = offsetBits{15-14};
3414 let Inst{23-22} = MajOp;
3415 let Inst{21} = isHalf;
3416 let Inst{20-16} = offsetBits{13-9};
3417 let Inst{13} = offsetBits{8};
3418 let Inst{12-8} = src;
3419 let Inst{7-0} = offsetBits{7-0};
3422 //===----------------------------------------------------------------------===//
3423 // Template class for predicated store instructions with
3424 // GP-Relative or absolute addressing.
3425 //===----------------------------------------------------------------------===//
3426 let hasSideEffects = 0, isPredicated = 1, isNVStorable = 1, opExtentBits = 6,
3428 class T_StoreAbs_Pred <string mnemonic, RegisterClass RC, bits<2> MajOp,
3429 bit isHalf, bit isNot, bit isNew>
3430 : STInst<(outs), (ins PredRegs:$src1, u6Ext:$absaddr, RC: $src2),
3431 !if(isNot, "if (!$src1", "if ($src1")#!if(isNew, ".new) ",
3432 ") ")#mnemonic#"(#$absaddr) = $src2"#!if(isHalf, ".h",""),
3433 [], "", ST_tc_st_SLOT01>, AddrModeRel {
3438 let isPredicatedNew = isNew;
3439 let isPredicatedFalse = isNot;
3441 let IClass = 0b1010;
3443 let Inst{27-24} = 0b1111;
3444 let Inst{23-22} = MajOp;
3445 let Inst{21} = isHalf;
3446 let Inst{17-16} = absaddr{5-4};
3447 let Inst{13} = isNew;
3448 let Inst{12-8} = src2;
3450 let Inst{6-3} = absaddr{3-0};
3451 let Inst{2} = isNot;
3452 let Inst{1-0} = src1;
3455 //===----------------------------------------------------------------------===//
3456 // Template class for predicated store instructions with absolute addressing.
3457 //===----------------------------------------------------------------------===//
3458 class T_StoreAbs <string mnemonic, RegisterClass RC, Operand ImmOp,
3459 bits<2> MajOp, bit isHalf>
3460 : T_StoreAbsGP <mnemonic, RC, ImmOp, MajOp, u0AlwaysExt, 1, isHalf>,
3462 string ImmOpStr = !cast<string>(ImmOp);
3463 let opExtentBits = !if (!eq(ImmOpStr, "u16_3Imm"), 19,
3464 !if (!eq(ImmOpStr, "u16_2Imm"), 18,
3465 !if (!eq(ImmOpStr, "u16_1Imm"), 17,
3466 /* u16_0Imm */ 16)));
3468 let opExtentAlign = !if (!eq(ImmOpStr, "u16_3Imm"), 3,
3469 !if (!eq(ImmOpStr, "u16_2Imm"), 2,
3470 !if (!eq(ImmOpStr, "u16_1Imm"), 1,
3471 /* u16_0Imm */ 0)));
3474 //===----------------------------------------------------------------------===//
3475 // Multiclass for store instructions with absolute addressing.
3476 //===----------------------------------------------------------------------===//
3477 let validSubTargets = HasV4SubT, addrMode = Absolute, isExtended = 1 in
3478 multiclass ST_Abs<string mnemonic, string CextOp, RegisterClass RC,
3479 Operand ImmOp, bits<2> MajOp, bit isHalf = 0> {
3480 let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in {
3481 let opExtendable = 0, isPredicable = 1 in
3482 def S2_#NAME#abs : T_StoreAbs <mnemonic, RC, ImmOp, MajOp, isHalf>;
3485 def S4_p#NAME#t_abs : T_StoreAbs_Pred<mnemonic, RC, MajOp, isHalf, 0, 0>;
3486 def S4_p#NAME#f_abs : T_StoreAbs_Pred<mnemonic, RC, MajOp, isHalf, 1, 0>;
3489 def S4_p#NAME#tnew_abs : T_StoreAbs_Pred<mnemonic, RC, MajOp, isHalf, 0, 1>;
3490 def S4_p#NAME#fnew_abs : T_StoreAbs_Pred<mnemonic, RC, MajOp, isHalf, 1, 1>;
3494 //===----------------------------------------------------------------------===//
3495 // Template class for non predicated new-value store instructions with
3496 // GP-Relative or absolute addressing.
3497 //===----------------------------------------------------------------------===//
3498 let hasSideEffects = 0, isPredicable = 1, mayStore = 1, isNVStore = 1,
3499 isNewValue = 1, opNewValue = 1 in
3500 class T_StoreAbsGP_NV <string mnemonic, Operand ImmOp, bits<2>MajOp, bit isAbs>
3501 : NVInst_V4<(outs), (ins u0AlwaysExt:$addr, IntRegs:$src),
3502 mnemonic # !if(isAbs, "(##", "(#")#"$addr) = $src.new",
3503 [], "", V2LDST_tc_st_SLOT0> {
3506 bits<16> offsetBits;
3508 string ImmOpStr = !cast<string>(ImmOp);
3509 let offsetBits = !if (!eq(ImmOpStr, "u16_3Imm"), addr{18-3},
3510 !if (!eq(ImmOpStr, "u16_2Imm"), addr{17-2},
3511 !if (!eq(ImmOpStr, "u16_1Imm"), addr{16-1},
3512 /* u16_0Imm */ addr{15-0})));
3513 let IClass = 0b0100;
3516 let Inst{26-25} = offsetBits{15-14};
3517 let Inst{24-21} = 0b0101;
3518 let Inst{20-16} = offsetBits{13-9};
3519 let Inst{13} = offsetBits{8};
3520 let Inst{12-11} = MajOp;
3521 let Inst{10-8} = src;
3522 let Inst{7-0} = offsetBits{7-0};
3525 //===----------------------------------------------------------------------===//
3526 // Template class for predicated new-value store instructions with
3527 // absolute addressing.
3528 //===----------------------------------------------------------------------===//
3529 let hasSideEffects = 0, isPredicated = 1, mayStore = 1, isNVStore = 1,
3530 isNewValue = 1, opNewValue = 2, opExtentBits = 6, opExtendable = 1 in
3531 class T_StoreAbs_NV_Pred <string mnemonic, bits<2> MajOp, bit isNot, bit isNew>
3532 : NVInst_V4<(outs), (ins PredRegs:$src1, u6Ext:$absaddr, IntRegs:$src2),
3533 !if(isNot, "if (!$src1", "if ($src1")#!if(isNew, ".new) ",
3534 ") ")#mnemonic#"(#$absaddr) = $src2.new",
3535 [], "", ST_tc_st_SLOT0>, AddrModeRel {
3540 let isPredicatedNew = isNew;
3541 let isPredicatedFalse = isNot;
3543 let IClass = 0b1010;
3545 let Inst{27-24} = 0b1111;
3546 let Inst{23-21} = 0b101;
3547 let Inst{17-16} = absaddr{5-4};
3548 let Inst{13} = isNew;
3549 let Inst{12-11} = MajOp;
3550 let Inst{10-8} = src2;
3552 let Inst{6-3} = absaddr{3-0};
3553 let Inst{2} = isNot;
3554 let Inst{1-0} = src1;
3557 //===----------------------------------------------------------------------===//
3558 // Template class for non-predicated new-value store instructions with
3559 // absolute addressing.
3560 //===----------------------------------------------------------------------===//
3561 class T_StoreAbs_NV <string mnemonic, Operand ImmOp, bits<2> MajOp>
3562 : T_StoreAbsGP_NV <mnemonic, ImmOp, MajOp, 1>, AddrModeRel {
3564 string ImmOpStr = !cast<string>(ImmOp);
3565 let opExtentBits = !if (!eq(ImmOpStr, "u16_3Imm"), 19,
3566 !if (!eq(ImmOpStr, "u16_2Imm"), 18,
3567 !if (!eq(ImmOpStr, "u16_1Imm"), 17,
3568 /* u16_0Imm */ 16)));
3570 let opExtentAlign = !if (!eq(ImmOpStr, "u16_3Imm"), 3,
3571 !if (!eq(ImmOpStr, "u16_2Imm"), 2,
3572 !if (!eq(ImmOpStr, "u16_1Imm"), 1,
3573 /* u16_0Imm */ 0)));
3576 //===----------------------------------------------------------------------===//
3577 // Multiclass for new-value store instructions with absolute addressing.
3578 //===----------------------------------------------------------------------===//
3579 let validSubTargets = HasV4SubT, addrMode = Absolute, isExtended = 1 in
3580 multiclass ST_Abs_NV <string mnemonic, string CextOp, Operand ImmOp,
3582 let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in {
3583 let opExtendable = 0, isPredicable = 1 in
3584 def S2_#NAME#newabs : T_StoreAbs_NV <mnemonic, ImmOp, MajOp>;
3587 def S4_p#NAME#newt_abs : T_StoreAbs_NV_Pred <mnemonic, MajOp, 0, 0>;
3588 def S4_p#NAME#newf_abs : T_StoreAbs_NV_Pred <mnemonic, MajOp, 1, 0>;
3591 def S4_p#NAME#newtnew_abs : T_StoreAbs_NV_Pred <mnemonic, MajOp, 0, 1>;
3592 def S4_p#NAME#newfnew_abs : T_StoreAbs_NV_Pred <mnemonic, MajOp, 1, 1>;
3596 //===----------------------------------------------------------------------===//
3597 // Stores with absolute addressing
3598 //===----------------------------------------------------------------------===//
3599 let accessSize = ByteAccess, isCodeGenOnly = 0 in
3600 defm storerb : ST_Abs <"memb", "STrib", IntRegs, u16_0Imm, 0b00>,
3601 ST_Abs_NV <"memb", "STrib", u16_0Imm, 0b00>;
3603 let accessSize = HalfWordAccess, isCodeGenOnly = 0 in
3604 defm storerh : ST_Abs <"memh", "STrih", IntRegs, u16_1Imm, 0b01>,
3605 ST_Abs_NV <"memh", "STrih", u16_1Imm, 0b01>;
3607 let accessSize = WordAccess, isCodeGenOnly = 0 in
3608 defm storeri : ST_Abs <"memw", "STriw", IntRegs, u16_2Imm, 0b10>,
3609 ST_Abs_NV <"memw", "STriw", u16_2Imm, 0b10>;
3611 let isNVStorable = 0, accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
3612 defm storerd : ST_Abs <"memd", "STrid", DoubleRegs, u16_3Imm, 0b11>;
3614 let isNVStorable = 0, accessSize = HalfWordAccess, isCodeGenOnly = 0 in
3615 defm storerf : ST_Abs <"memh", "STrif", IntRegs, u16_1Imm, 0b01, 1>;
3617 //===----------------------------------------------------------------------===//
3618 // GP-relative stores.
3619 // mem[bhwd](#global)=Rt
3620 // Once predicated, these instructions map to absolute addressing mode.
3621 // if ([!]Pv[.new]) mem[bhwd](##global)=Rt
3622 //===----------------------------------------------------------------------===//
3624 let validSubTargets = HasV4SubT, isAsmParserOnly = 1 in
3625 class T_StoreGP <string mnemonic, string BaseOp, RegisterClass RC,
3626 Operand ImmOp, bits<2> MajOp, bit isHalf = 0>
3627 : T_StoreAbsGP <mnemonic, RC, ImmOp, MajOp, globaladdress, 0, isHalf> {
3628 // Set BaseOpcode same as absolute addressing instructions so that
3629 // non-predicated GP-Rel instructions can have relate with predicated
3630 // Absolute instruction.
3631 let BaseOpcode = BaseOp#_abs;
3634 let validSubTargets = HasV4SubT, isAsmParserOnly = 1 in
3635 multiclass ST_GP <string mnemonic, string BaseOp, Operand ImmOp,
3636 bits<2> MajOp, bit isHalf = 0> {
3637 // Set BaseOpcode same as absolute addressing instructions so that
3638 // non-predicated GP-Rel instructions can have relate with predicated
3639 // Absolute instruction.
3640 let BaseOpcode = BaseOp#_abs in {
3641 def NAME#gp : T_StoreAbsGP <mnemonic, IntRegs, ImmOp, MajOp,
3642 globaladdress, 0, isHalf>;
3644 def NAME#newgp : T_StoreAbsGP_NV <mnemonic, ImmOp, MajOp, 0> ;
3648 let accessSize = ByteAccess in
3649 defm S2_storerb : ST_GP<"memb", "STrib", u16_0Imm, 0b00>, NewValueRel;
3651 let accessSize = HalfWordAccess in
3652 defm S2_storerh : ST_GP<"memh", "STrih", u16_1Imm, 0b01>, NewValueRel;
3654 let accessSize = WordAccess in
3655 defm S2_storeri : ST_GP<"memw", "STriw", u16_2Imm, 0b10>, NewValueRel;
3657 let isNVStorable = 0, accessSize = DoubleWordAccess in
3658 def S2_storerdgp : T_StoreGP <"memd", "STrid", DoubleRegs,
3659 u16_3Imm, 0b11>, PredNewRel;
3661 let isNVStorable = 0, accessSize = HalfWordAccess in
3662 def S2_storerfgp : T_StoreGP <"memh", "STrif", IntRegs,
3663 u16_1Imm, 0b01, 1>, PredNewRel;
3665 let Predicates = [HasV4T], AddedComplexity = 30 in {
3666 def : Pat<(truncstorei8 (i32 IntRegs:$src1),
3667 (HexagonCONST32 tglobaladdr:$absaddr)),
3668 (S2_storerbabs tglobaladdr: $absaddr, IntRegs: $src1)>;
3670 def : Pat<(truncstorei16 (i32 IntRegs:$src1),
3671 (HexagonCONST32 tglobaladdr:$absaddr)),
3672 (S2_storerhabs tglobaladdr: $absaddr, IntRegs: $src1)>;
3674 def : Pat<(store (i32 IntRegs:$src1), (HexagonCONST32 tglobaladdr:$absaddr)),
3675 (S2_storeriabs tglobaladdr: $absaddr, IntRegs: $src1)>;
3677 def : Pat<(store (i64 DoubleRegs:$src1),
3678 (HexagonCONST32 tglobaladdr:$absaddr)),
3679 (S2_storerdabs tglobaladdr: $absaddr, DoubleRegs: $src1)>;
3682 // 64 bit atomic store
3683 def : Pat <(atomic_store_64 (HexagonCONST32_GP tglobaladdr:$global),
3684 (i64 DoubleRegs:$src1)),
3685 (S2_storerdgp tglobaladdr:$global, (i64 DoubleRegs:$src1))>,
3688 // Map from store(globaladdress) -> memd(#foo)
3689 let AddedComplexity = 100 in
3690 def : Pat <(store (i64 DoubleRegs:$src1),
3691 (HexagonCONST32_GP tglobaladdr:$global)),
3692 (S2_storerdgp tglobaladdr:$global, (i64 DoubleRegs:$src1))>;
3694 // 8 bit atomic store
3695 def : Pat < (atomic_store_8 (HexagonCONST32_GP tglobaladdr:$global),
3696 (i32 IntRegs:$src1)),
3697 (S2_storerbgp tglobaladdr:$global, (i32 IntRegs:$src1))>;
3699 // Map from store(globaladdress) -> memb(#foo)
3700 let AddedComplexity = 100 in
3701 def : Pat<(truncstorei8 (i32 IntRegs:$src1),
3702 (HexagonCONST32_GP tglobaladdr:$global)),
3703 (S2_storerbgp tglobaladdr:$global, (i32 IntRegs:$src1))>;
3705 // Map from "i1 = constant<-1>; memw(CONST32(#foo)) = i1"
3706 // to "r0 = 1; memw(#foo) = r0"
3707 let AddedComplexity = 100 in
3708 def : Pat<(store (i1 -1), (HexagonCONST32_GP tglobaladdr:$global)),
3709 (S2_storerbgp tglobaladdr:$global, (A2_tfrsi 1))>;
3711 def : Pat<(atomic_store_16 (HexagonCONST32_GP tglobaladdr:$global),
3712 (i32 IntRegs:$src1)),
3713 (S2_storerhgp tglobaladdr:$global, (i32 IntRegs:$src1))>;
3715 // Map from store(globaladdress) -> memh(#foo)
3716 let AddedComplexity = 100 in
3717 def : Pat<(truncstorei16 (i32 IntRegs:$src1),
3718 (HexagonCONST32_GP tglobaladdr:$global)),
3719 (S2_storerhgp tglobaladdr:$global, (i32 IntRegs:$src1))>;
3721 // 32 bit atomic store
3722 def : Pat<(atomic_store_32 (HexagonCONST32_GP tglobaladdr:$global),
3723 (i32 IntRegs:$src1)),
3724 (S2_storerigp tglobaladdr:$global, (i32 IntRegs:$src1))>;
3726 // Map from store(globaladdress) -> memw(#foo)
3727 let AddedComplexity = 100 in
3728 def : Pat<(store (i32 IntRegs:$src1), (HexagonCONST32_GP tglobaladdr:$global)),
3729 (S2_storerigp tglobaladdr:$global, (i32 IntRegs:$src1))>;
3731 //===----------------------------------------------------------------------===//
3732 // Template class for non predicated load instructions with
3733 // absolute addressing mode.
3734 //===----------------------------------------------------------------------===//
3735 let isPredicable = 1, hasSideEffects = 0, validSubTargets = HasV4SubT in
3736 class T_LoadAbsGP <string mnemonic, RegisterClass RC, Operand ImmOp,
3737 bits<3> MajOp, Operand AddrOp, bit isAbs>
3738 : LDInst <(outs RC:$dst), (ins AddrOp:$addr),
3739 "$dst = "#mnemonic# !if(isAbs, "(##", "(#")#"$addr)",
3740 [], "", V2LDST_tc_ld_SLOT01> {
3743 bits<16> offsetBits;
3745 string ImmOpStr = !cast<string>(ImmOp);
3746 let offsetBits = !if (!eq(ImmOpStr, "u16_3Imm"), addr{18-3},
3747 !if (!eq(ImmOpStr, "u16_2Imm"), addr{17-2},
3748 !if (!eq(ImmOpStr, "u16_1Imm"), addr{16-1},
3749 /* u16_0Imm */ addr{15-0})));
3751 let IClass = 0b0100;
3754 let Inst{26-25} = offsetBits{15-14};
3756 let Inst{23-21} = MajOp;
3757 let Inst{20-16} = offsetBits{13-9};
3758 let Inst{13-5} = offsetBits{8-0};
3759 let Inst{4-0} = dst;
3762 class T_LoadAbs <string mnemonic, RegisterClass RC, Operand ImmOp,
3764 : T_LoadAbsGP <mnemonic, RC, ImmOp, MajOp, u0AlwaysExt, 1>, AddrModeRel {
3766 string ImmOpStr = !cast<string>(ImmOp);
3767 let opExtentBits = !if (!eq(ImmOpStr, "u16_3Imm"), 19,
3768 !if (!eq(ImmOpStr, "u16_2Imm"), 18,
3769 !if (!eq(ImmOpStr, "u16_1Imm"), 17,
3770 /* u16_0Imm */ 16)));
3772 let opExtentAlign = !if (!eq(ImmOpStr, "u16_3Imm"), 3,
3773 !if (!eq(ImmOpStr, "u16_2Imm"), 2,
3774 !if (!eq(ImmOpStr, "u16_1Imm"), 1,
3775 /* u16_0Imm */ 0)));
3777 //===----------------------------------------------------------------------===//
3778 // Template class for predicated load instructions with
3779 // absolute addressing mode.
3780 //===----------------------------------------------------------------------===//
3781 let isPredicated = 1, hasNewValue = 1, opExtentBits = 6, opExtendable = 2 in
3782 class T_LoadAbs_Pred <string mnemonic, RegisterClass RC, bits<3> MajOp,
3783 bit isPredNot, bit isPredNew>
3784 : LDInst <(outs RC:$dst), (ins PredRegs:$src1, u6Ext:$absaddr),
3785 !if(isPredNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
3786 ") ")#"$dst = "#mnemonic#"(#$absaddr)">, AddrModeRel {
3791 let isPredicatedNew = isPredNew;
3792 let isPredicatedFalse = isPredNot;
3794 let IClass = 0b1001;
3796 let Inst{27-24} = 0b1111;
3797 let Inst{23-21} = MajOp;
3798 let Inst{20-16} = absaddr{5-1};
3800 let Inst{12} = isPredNew;
3801 let Inst{11} = isPredNot;
3802 let Inst{10-9} = src1;
3803 let Inst{8} = absaddr{0};
3805 let Inst{4-0} = dst;
3808 //===----------------------------------------------------------------------===//
3809 // Multiclass for the load instructions with absolute addressing mode.
3810 //===----------------------------------------------------------------------===//
3811 multiclass LD_Abs_Pred<string mnemonic, RegisterClass RC, bits<3> MajOp,
3813 def _abs : T_LoadAbs_Pred <mnemonic, RC, MajOp, PredNot, 0>;
3815 def new_abs : T_LoadAbs_Pred <mnemonic, RC, MajOp, PredNot, 1>;
3818 let addrMode = Absolute, isExtended = 1 in
3819 multiclass LD_Abs<string mnemonic, string CextOp, RegisterClass RC,
3820 Operand ImmOp, bits<3> MajOp> {
3821 let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in {
3822 let opExtendable = 1, isPredicable = 1 in
3823 def L4_#NAME#_abs: T_LoadAbs <mnemonic, RC, ImmOp, MajOp>;
3826 defm L4_p#NAME#t : LD_Abs_Pred<mnemonic, RC, MajOp, 0>;
3827 defm L4_p#NAME#f : LD_Abs_Pred<mnemonic, RC, MajOp, 1>;
3831 let accessSize = ByteAccess, hasNewValue = 1, isCodeGenOnly = 0 in {
3832 defm loadrb : LD_Abs<"memb", "LDrib", IntRegs, u16_0Imm, 0b000>;
3833 defm loadrub : LD_Abs<"memub", "LDriub", IntRegs, u16_0Imm, 0b001>;
3836 let accessSize = HalfWordAccess, hasNewValue = 1, isCodeGenOnly = 0 in {
3837 defm loadrh : LD_Abs<"memh", "LDrih", IntRegs, u16_1Imm, 0b010>;
3838 defm loadruh : LD_Abs<"memuh", "LDriuh", IntRegs, u16_1Imm, 0b011>;
3841 let accessSize = WordAccess, hasNewValue = 1, isCodeGenOnly = 0 in
3842 defm loadri : LD_Abs<"memw", "LDriw", IntRegs, u16_2Imm, 0b100>;
3844 let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
3845 defm loadrd : LD_Abs<"memd", "LDrid", DoubleRegs, u16_3Imm, 0b110>;
3847 //===----------------------------------------------------------------------===//
3848 // multiclass for load instructions with GP-relative addressing mode.
3849 // Rx=mem[bhwd](##global)
3850 // Once predicated, these instructions map to absolute addressing mode.
3851 // if ([!]Pv[.new]) Rx=mem[bhwd](##global)
3852 //===----------------------------------------------------------------------===//
3854 let isAsmParserOnly = 1 in
3855 class T_LoadGP <string mnemonic, string BaseOp, RegisterClass RC, Operand ImmOp,
3857 : T_LoadAbsGP <mnemonic, RC, ImmOp, MajOp, globaladdress, 0>, PredNewRel {
3858 let BaseOpcode = BaseOp#_abs;
3861 let accessSize = ByteAccess, hasNewValue = 1 in {
3862 def L2_loadrbgp : T_LoadGP<"memb", "LDrib", IntRegs, u16_0Imm, 0b000>;
3863 def L2_loadrubgp : T_LoadGP<"memub", "LDriub", IntRegs, u16_0Imm, 0b001>;
3866 let accessSize = HalfWordAccess, hasNewValue = 1 in {
3867 def L2_loadrhgp : T_LoadGP<"memh", "LDrih", IntRegs, u16_1Imm, 0b010>;
3868 def L2_loadruhgp : T_LoadGP<"memuh", "LDriuh", IntRegs, u16_1Imm, 0b011>;
3871 let accessSize = WordAccess, hasNewValue = 1 in
3872 def L2_loadrigp : T_LoadGP<"memw", "LDriw", IntRegs, u16_2Imm, 0b100>;
3874 let accessSize = DoubleWordAccess in
3875 def L2_loadrdgp : T_LoadGP<"memd", "LDrid", DoubleRegs, u16_3Imm, 0b110>;
3877 let Predicates = [HasV4T], AddedComplexity = 30 in {
3878 def : Pat<(i32 (load (HexagonCONST32 tglobaladdr:$absaddr))),
3879 (L4_loadri_abs tglobaladdr: $absaddr)>;
3881 def : Pat<(i32 (sextloadi8 (HexagonCONST32 tglobaladdr:$absaddr))),
3882 (L4_loadrb_abs tglobaladdr:$absaddr)>;
3884 def : Pat<(i32 (zextloadi8 (HexagonCONST32 tglobaladdr:$absaddr))),
3885 (L4_loadrub_abs tglobaladdr:$absaddr)>;
3887 def : Pat<(i32 (sextloadi16 (HexagonCONST32 tglobaladdr:$absaddr))),
3888 (L4_loadrh_abs tglobaladdr:$absaddr)>;
3890 def : Pat<(i32 (zextloadi16 (HexagonCONST32 tglobaladdr:$absaddr))),
3891 (L4_loadruh_abs tglobaladdr:$absaddr)>;
3894 def : Pat <(atomic_load_64 (HexagonCONST32_GP tglobaladdr:$global)),
3895 (i64 (L2_loadrdgp tglobaladdr:$global))>;
3897 def : Pat <(atomic_load_32 (HexagonCONST32_GP tglobaladdr:$global)),
3898 (i32 (L2_loadrigp tglobaladdr:$global))>;
3900 def : Pat <(atomic_load_16 (HexagonCONST32_GP tglobaladdr:$global)),
3901 (i32 (L2_loadruhgp tglobaladdr:$global))>;
3903 def : Pat <(atomic_load_8 (HexagonCONST32_GP tglobaladdr:$global)),
3904 (i32 (L2_loadrubgp tglobaladdr:$global))>;
3906 // Map from load(globaladdress) -> memw(#foo + 0)
3907 let AddedComplexity = 100 in
3908 def : Pat <(i64 (load (HexagonCONST32_GP tglobaladdr:$global))),
3909 (i64 (L2_loadrdgp tglobaladdr:$global))>;
3911 // Map from Pd = load(globaladdress) -> Rd = memb(globaladdress), Pd = Rd
3912 let AddedComplexity = 100 in
3913 def : Pat <(i1 (load (HexagonCONST32_GP tglobaladdr:$global))),
3914 (i1 (C2_tfrrp (i32 (L2_loadrbgp tglobaladdr:$global))))>;
3916 // When the Interprocedural Global Variable optimizer realizes that a certain
3917 // global variable takes only two constant values, it shrinks the global to
3918 // a boolean. Catch those loads here in the following 3 patterns.
3919 let AddedComplexity = 100 in
3920 def : Pat <(i32 (extloadi1 (HexagonCONST32_GP tglobaladdr:$global))),
3921 (i32 (L2_loadrbgp tglobaladdr:$global))>;
3923 let AddedComplexity = 100 in
3924 def : Pat <(i32 (sextloadi1 (HexagonCONST32_GP tglobaladdr:$global))),
3925 (i32 (L2_loadrbgp tglobaladdr:$global))>;
3927 // Map from load(globaladdress) -> memb(#foo)
3928 let AddedComplexity = 100 in
3929 def : Pat <(i32 (extloadi8 (HexagonCONST32_GP tglobaladdr:$global))),
3930 (i32 (L2_loadrbgp tglobaladdr:$global))>;
3932 // Map from load(globaladdress) -> memb(#foo)
3933 let AddedComplexity = 100 in
3934 def : Pat <(i32 (sextloadi8 (HexagonCONST32_GP tglobaladdr:$global))),
3935 (i32 (L2_loadrbgp tglobaladdr:$global))>;
3937 let AddedComplexity = 100 in
3938 def : Pat <(i32 (zextloadi1 (HexagonCONST32_GP tglobaladdr:$global))),
3939 (i32 (L2_loadrubgp tglobaladdr:$global))>;
3941 // Map from load(globaladdress) -> memub(#foo)
3942 let AddedComplexity = 100 in
3943 def : Pat <(i32 (zextloadi8 (HexagonCONST32_GP tglobaladdr:$global))),
3944 (i32 (L2_loadrubgp tglobaladdr:$global))>;
3946 // Map from load(globaladdress) -> memh(#foo)
3947 let AddedComplexity = 100 in
3948 def : Pat <(i32 (extloadi16 (HexagonCONST32_GP tglobaladdr:$global))),
3949 (i32 (L2_loadrhgp tglobaladdr:$global))>;
3951 // Map from load(globaladdress) -> memh(#foo)
3952 let AddedComplexity = 100 in
3953 def : Pat <(i32 (sextloadi16 (HexagonCONST32_GP tglobaladdr:$global))),
3954 (i32 (L2_loadrhgp tglobaladdr:$global))>;
3956 // Map from load(globaladdress) -> memuh(#foo)
3957 let AddedComplexity = 100 in
3958 def : Pat <(i32 (zextloadi16 (HexagonCONST32_GP tglobaladdr:$global))),
3959 (i32 (L2_loadruhgp tglobaladdr:$global))>;
3961 // Map from load(globaladdress) -> memw(#foo)
3962 let AddedComplexity = 100 in
3963 def : Pat <(i32 (load (HexagonCONST32_GP tglobaladdr:$global))),
3964 (i32 (L2_loadrigp tglobaladdr:$global))>;
3967 // Transfer global address into a register
3968 let isExtended = 1, opExtendable = 1, AddedComplexity=50, isMoveImm = 1,
3969 isAsCheapAsAMove = 1, isReMaterializable = 1, validSubTargets = HasV4SubT in
3970 def TFRI_V4 : ALU32_ri<(outs IntRegs:$dst), (ins s16Ext:$src1),
3972 [(set IntRegs:$dst, (HexagonCONST32 tglobaladdr:$src1))]>,
3975 // Transfer a block address into a register
3976 def : Pat<(HexagonCONST32_GP tblockaddress:$src1),
3977 (TFRI_V4 tblockaddress:$src1)>,
3980 let isExtended = 1, opExtendable = 2, AddedComplexity=50,
3981 hasSideEffects = 0, isPredicated = 1, validSubTargets = HasV4SubT in
3982 def TFRI_cPt_V4 : ALU32_ri<(outs IntRegs:$dst),
3983 (ins PredRegs:$src1, s16Ext:$src2),
3984 "if($src1) $dst = #$src2",
3988 let isExtended = 1, opExtendable = 2, AddedComplexity=50, isPredicatedFalse = 1,
3989 hasSideEffects = 0, isPredicated = 1, validSubTargets = HasV4SubT in
3990 def TFRI_cNotPt_V4 : ALU32_ri<(outs IntRegs:$dst),
3991 (ins PredRegs:$src1, s16Ext:$src2),
3992 "if(!$src1) $dst = #$src2",
3996 let isExtended = 1, opExtendable = 2, AddedComplexity=50,
3997 hasSideEffects = 0, isPredicated = 1, validSubTargets = HasV4SubT in
3998 def TFRI_cdnPt_V4 : ALU32_ri<(outs IntRegs:$dst),
3999 (ins PredRegs:$src1, s16Ext:$src2),
4000 "if($src1.new) $dst = #$src2",
4004 let isExtended = 1, opExtendable = 2, AddedComplexity=50, isPredicatedFalse = 1,
4005 hasSideEffects = 0, isPredicated = 1, validSubTargets = HasV4SubT in
4006 def TFRI_cdnNotPt_V4 : ALU32_ri<(outs IntRegs:$dst),
4007 (ins PredRegs:$src1, s16Ext:$src2),
4008 "if(!$src1.new) $dst = #$src2",
4012 let AddedComplexity = 50, Predicates = [HasV4T] in
4013 def : Pat<(HexagonCONST32_GP tglobaladdr:$src1),
4014 (TFRI_V4 tglobaladdr:$src1)>,
4017 let Predicates = [HasV4T], AddedComplexity = 30 in {
4018 def : Pat<(truncstorei8 (i32 IntRegs:$src1), u0AlwaysExtPred:$src2),
4019 (S2_storerbabs u0AlwaysExtPred:$src2, IntRegs: $src1)>;
4021 def : Pat<(truncstorei16 (i32 IntRegs:$src1), u0AlwaysExtPred:$src2),
4022 (S2_storerhabs u0AlwaysExtPred:$src2, IntRegs: $src1)>;
4024 def : Pat<(store (i32 IntRegs:$src1), u0AlwaysExtPred:$src2),
4025 (S2_storeriabs u0AlwaysExtPred:$src2, IntRegs: $src1)>;
4028 let Predicates = [HasV4T], AddedComplexity = 30 in {
4029 def : Pat<(i32 (load u0AlwaysExtPred:$src)),
4030 (L4_loadri_abs u0AlwaysExtPred:$src)>;
4032 def : Pat<(i32 (sextloadi8 u0AlwaysExtPred:$src)),
4033 (L4_loadrb_abs u0AlwaysExtPred:$src)>;
4035 def : Pat<(i32 (zextloadi8 u0AlwaysExtPred:$src)),
4036 (L4_loadrub_abs u0AlwaysExtPred:$src)>;
4038 def : Pat<(i32 (sextloadi16 u0AlwaysExtPred:$src)),
4039 (L4_loadrh_abs u0AlwaysExtPred:$src)>;
4041 def : Pat<(i32 (zextloadi16 u0AlwaysExtPred:$src)),
4042 (L4_loadruh_abs u0AlwaysExtPred:$src)>;
4045 // Indexed store word - global address.
4046 // memw(Rs+#u6:2)=#S8
4047 let AddedComplexity = 10 in
4048 def STriw_offset_ext_V4 : STInst<(outs),
4049 (ins IntRegs:$src1, u6_2Imm:$src2, globaladdress:$src3),
4050 "memw($src1+#$src2) = ##$src3",
4051 [(store (HexagonCONST32 tglobaladdr:$src3),
4052 (add IntRegs:$src1, u6_2ImmPred:$src2))]>,
4055 def : Pat<(i64 (ctlz (i64 DoubleRegs:$src1))),
4056 (i64 (A4_combineir (i32 0), (i32 (S2_cl0p DoubleRegs:$src1))))>,
4059 def : Pat<(i64 (cttz (i64 DoubleRegs:$src1))),
4060 (i64 (A4_combineir (i32 0), (i32 (S2_ct0p DoubleRegs:$src1))))>,
4065 // We need a complexity of 120 here to override preceding handling of
4067 let Predicates = [HasV4T], AddedComplexity = 120 in {
4068 def: Pat <(i64 (extloadi8 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
4069 (i64 (A4_combineir 0, (L4_loadrb_abs tglobaladdr:$addr)))>;
4071 def: Pat <(i64 (zextloadi8 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
4072 (i64 (A4_combineir 0, (L4_loadrub_abs tglobaladdr:$addr)))>;
4074 def: Pat <(i64 (sextloadi8 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
4075 (i64 (A2_sxtw (L4_loadrb_abs tglobaladdr:$addr)))>;
4077 def: Pat <(i64 (extloadi8 FoldGlobalAddr:$addr)),
4078 (i64 (A4_combineir 0, (L4_loadrb_abs FoldGlobalAddr:$addr)))>;
4080 def: Pat <(i64 (zextloadi8 FoldGlobalAddr:$addr)),
4081 (i64 (A4_combineir 0, (L4_loadrub_abs FoldGlobalAddr:$addr)))>;
4083 def: Pat <(i64 (sextloadi8 FoldGlobalAddr:$addr)),
4084 (i64 (A2_sxtw (L4_loadrb_abs FoldGlobalAddr:$addr)))>;
4087 // We need a complexity of 120 here to override preceding handling of
4089 let AddedComplexity = 120 in {
4090 def: Pat <(i64 (extloadi16 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
4091 (i64 (A4_combineir 0, (L4_loadrh_abs tglobaladdr:$addr)))>,
4094 def: Pat <(i64 (zextloadi16 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
4095 (i64 (A4_combineir 0, (L4_loadruh_abs tglobaladdr:$addr)))>,
4098 def: Pat <(i64 (sextloadi16 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
4099 (i64 (A2_sxtw (L4_loadrh_abs tglobaladdr:$addr)))>,
4102 def: Pat <(i64 (extloadi16 FoldGlobalAddr:$addr)),
4103 (i64 (A4_combineir 0, (L4_loadrh_abs FoldGlobalAddr:$addr)))>,
4106 def: Pat <(i64 (zextloadi16 FoldGlobalAddr:$addr)),
4107 (i64 (A4_combineir 0, (L4_loadruh_abs FoldGlobalAddr:$addr)))>,
4110 def: Pat <(i64 (sextloadi16 FoldGlobalAddr:$addr)),
4111 (i64 (A2_sxtw (L4_loadrh_abs FoldGlobalAddr:$addr)))>,
4115 // We need a complexity of 120 here to override preceding handling of
4117 let AddedComplexity = 120 in {
4118 def: Pat <(i64 (extloadi32 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
4119 (i64 (A4_combineir 0, (L4_loadri_abs tglobaladdr:$addr)))>,
4122 def: Pat <(i64 (zextloadi32 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
4123 (i64 (A4_combineir 0, (L4_loadri_abs tglobaladdr:$addr)))>,
4126 def: Pat <(i64 (sextloadi32 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
4127 (i64 (A2_sxtw (L4_loadri_abs tglobaladdr:$addr)))>,
4130 def: Pat <(i64 (extloadi32 FoldGlobalAddr:$addr)),
4131 (i64 (A4_combineir 0, (L4_loadri_abs FoldGlobalAddr:$addr)))>,
4134 def: Pat <(i64 (zextloadi32 FoldGlobalAddr:$addr)),
4135 (i64 (A4_combineir 0, (L4_loadri_abs FoldGlobalAddr:$addr)))>,
4138 def: Pat <(i64 (sextloadi32 FoldGlobalAddr:$addr)),
4139 (i64 (A2_sxtw (L4_loadri_abs FoldGlobalAddr:$addr)))>,
4143 // Indexed store double word - global address.
4144 // memw(Rs+#u6:2)=#S8
4145 let AddedComplexity = 10 in
4146 def STrih_offset_ext_V4 : STInst<(outs),
4147 (ins IntRegs:$src1, u6_1Imm:$src2, globaladdress:$src3),
4148 "memh($src1+#$src2) = ##$src3",
4149 [(truncstorei16 (HexagonCONST32 tglobaladdr:$src3),
4150 (add IntRegs:$src1, u6_1ImmPred:$src2))]>,
4152 // Map from store(globaladdress + x) -> memd(#foo + x)
4153 let AddedComplexity = 100 in
4154 def : Pat<(store (i64 DoubleRegs:$src1),
4155 FoldGlobalAddrGP:$addr),
4156 (S2_storerdabs FoldGlobalAddrGP:$addr, (i64 DoubleRegs:$src1))>,
4159 def : Pat<(atomic_store_64 FoldGlobalAddrGP:$addr,
4160 (i64 DoubleRegs:$src1)),
4161 (S2_storerdabs FoldGlobalAddrGP:$addr, (i64 DoubleRegs:$src1))>,
4164 // Map from store(globaladdress + x) -> memb(#foo + x)
4165 let AddedComplexity = 100 in
4166 def : Pat<(truncstorei8 (i32 IntRegs:$src1), FoldGlobalAddrGP:$addr),
4167 (S2_storerbabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>,
4170 def : Pat<(atomic_store_8 FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1)),
4171 (S2_storerbabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>,
4174 // Map from store(globaladdress + x) -> memh(#foo + x)
4175 let AddedComplexity = 100 in
4176 def : Pat<(truncstorei16 (i32 IntRegs:$src1), FoldGlobalAddrGP:$addr),
4177 (S2_storerhabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>,
4180 def : Pat<(atomic_store_16 FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1)),
4181 (S2_storerhabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>,
4184 // Map from store(globaladdress + x) -> memw(#foo + x)
4185 let AddedComplexity = 100 in
4186 def : Pat<(store (i32 IntRegs:$src1), FoldGlobalAddrGP:$addr),
4187 (S2_storeriabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>,
4190 def : Pat<(atomic_store_32 FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1)),
4191 (S2_storeriabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>,
4194 // Map from load(globaladdress + x) -> memd(#foo + x)
4195 let AddedComplexity = 100 in
4196 def : Pat<(i64 (load FoldGlobalAddrGP:$addr)),
4197 (i64 (L4_loadrd_abs FoldGlobalAddrGP:$addr))>,
4200 def : Pat<(atomic_load_64 FoldGlobalAddrGP:$addr),
4201 (i64 (L4_loadrd_abs FoldGlobalAddrGP:$addr))>,
4204 // Map from load(globaladdress + x) -> memb(#foo + x)
4205 let AddedComplexity = 100 in
4206 def : Pat<(i32 (extloadi8 FoldGlobalAddrGP:$addr)),
4207 (i32 (L4_loadrb_abs FoldGlobalAddrGP:$addr))>,
4210 // Map from load(globaladdress + x) -> memb(#foo + x)
4211 let AddedComplexity = 100 in
4212 def : Pat<(i32 (sextloadi8 FoldGlobalAddrGP:$addr)),
4213 (i32 (L4_loadrb_abs FoldGlobalAddrGP:$addr))>,
4216 //let AddedComplexity = 100 in
4217 let AddedComplexity = 100 in
4218 def : Pat<(i32 (extloadi16 FoldGlobalAddrGP:$addr)),
4219 (i32 (L4_loadrh_abs FoldGlobalAddrGP:$addr))>,
4222 // Map from load(globaladdress + x) -> memh(#foo + x)
4223 let AddedComplexity = 100 in
4224 def : Pat<(i32 (sextloadi16 FoldGlobalAddrGP:$addr)),
4225 (i32 (L4_loadrh_abs FoldGlobalAddrGP:$addr))>,
4228 // Map from load(globaladdress + x) -> memuh(#foo + x)
4229 let AddedComplexity = 100 in
4230 def : Pat<(i32 (zextloadi16 FoldGlobalAddrGP:$addr)),
4231 (i32 (L4_loadruh_abs FoldGlobalAddrGP:$addr))>,
4234 def : Pat<(atomic_load_16 FoldGlobalAddrGP:$addr),
4235 (i32 (L4_loadruh_abs FoldGlobalAddrGP:$addr))>,
4238 // Map from load(globaladdress + x) -> memub(#foo + x)
4239 let AddedComplexity = 100 in
4240 def : Pat<(i32 (zextloadi8 FoldGlobalAddrGP:$addr)),
4241 (i32 (L4_loadrub_abs FoldGlobalAddrGP:$addr))>,
4244 def : Pat<(atomic_load_8 FoldGlobalAddrGP:$addr),
4245 (i32 (L4_loadrub_abs FoldGlobalAddrGP:$addr))>,
4248 // Map from load(globaladdress + x) -> memw(#foo + x)
4249 let AddedComplexity = 100 in
4250 def : Pat<(i32 (load FoldGlobalAddrGP:$addr)),
4251 (i32 (L4_loadri_abs FoldGlobalAddrGP:$addr))>,
4254 def : Pat<(atomic_load_32 FoldGlobalAddrGP:$addr),
4255 (i32 (L4_loadri_abs FoldGlobalAddrGP:$addr))>,
4258 //===----------------------------------------------------------------------===//
4259 // :raw for of boundscheck:hi:lo insns
4260 //===----------------------------------------------------------------------===//
4262 // A4_boundscheck_lo: Detect if a register is within bounds.
4263 let hasSideEffects = 0, isCodeGenOnly = 0 in
4264 def A4_boundscheck_lo: ALU64Inst <
4265 (outs PredRegs:$Pd),
4266 (ins DoubleRegs:$Rss, DoubleRegs:$Rtt),
4267 "$Pd = boundscheck($Rss, $Rtt):raw:lo"> {
4272 let IClass = 0b1101;
4274 let Inst{27-23} = 0b00100;
4276 let Inst{7-5} = 0b100;
4278 let Inst{20-16} = Rss;
4279 let Inst{12-8} = Rtt;
4282 // A4_boundscheck_hi: Detect if a register is within bounds.
4283 let hasSideEffects = 0, isCodeGenOnly = 0 in
4284 def A4_boundscheck_hi: ALU64Inst <
4285 (outs PredRegs:$Pd),
4286 (ins DoubleRegs:$Rss, DoubleRegs:$Rtt),
4287 "$Pd = boundscheck($Rss, $Rtt):raw:hi"> {
4292 let IClass = 0b1101;
4294 let Inst{27-23} = 0b00100;
4296 let Inst{7-5} = 0b101;
4298 let Inst{20-16} = Rss;
4299 let Inst{12-8} = Rtt;
4302 let hasSideEffects = 0, isAsmParserOnly = 1 in
4303 def A4_boundscheck : MInst <
4304 (outs PredRegs:$Pd), (ins IntRegs:$Rs, DoubleRegs:$Rtt),
4305 "$Pd=boundscheck($Rs,$Rtt)">;
4307 // A4_tlbmatch: Detect if a VA/ASID matches a TLB entry.
4308 let isPredicateLate = 1, hasSideEffects = 0, isCodeGenOnly = 0 in
4309 def A4_tlbmatch : ALU64Inst<(outs PredRegs:$Pd),
4310 (ins DoubleRegs:$Rs, IntRegs:$Rt),
4311 "$Pd = tlbmatch($Rs, $Rt)",
4312 [], "", ALU64_tc_2early_SLOT23> {
4317 let IClass = 0b1101;
4318 let Inst{27-23} = 0b00100;
4319 let Inst{20-16} = Rs;
4321 let Inst{12-8} = Rt;
4322 let Inst{7-5} = 0b011;
4326 // We need custom lowering of ISD::PREFETCH into HexagonISD::DCFETCH
4327 // because the SDNode ISD::PREFETCH has properties MayLoad and MayStore.
4328 // We don't really want either one here.
4329 def SDTHexagonDCFETCH : SDTypeProfile<0, 2, [SDTCisPtrTy<0>,SDTCisInt<1>]>;
4330 def HexagonDCFETCH : SDNode<"HexagonISD::DCFETCH", SDTHexagonDCFETCH,
4333 // Use LD0Inst for dcfetch, but set "mayLoad" to 0 because this doesn't
4334 // really do a load.
4335 let hasSideEffects = 1, mayLoad = 0, isCodeGenOnly = 0 in
4336 def Y2_dcfetchbo : LD0Inst<(outs), (ins IntRegs:$Rs, u11_3Imm:$u11_3),
4337 "dcfetch($Rs + #$u11_3)",
4338 [(HexagonDCFETCH IntRegs:$Rs, u11_3ImmPred:$u11_3)],
4339 "", LD_tc_ld_SLOT0> {
4343 let IClass = 0b1001;
4344 let Inst{27-21} = 0b0100000;
4345 let Inst{20-16} = Rs;
4347 let Inst{10-0} = u11_3{13-3};
4350 //===----------------------------------------------------------------------===//
4351 // Compound instructions
4352 //===----------------------------------------------------------------------===//
4354 let isBranch = 1, hasSideEffects = 0, isExtentSigned = 1,
4355 isPredicated = 1, isPredicatedNew = 1, isExtendable = 1,
4356 opExtentBits = 11, opExtentAlign = 2, opExtendable = 1,
4357 isTerminator = 1, validSubTargets = HasV4SubT in
4358 class CJInst_tstbit_R0<string px, bit np, string tnt>
4359 : InstHexagon<(outs), (ins IntRegs:$Rs, brtarget:$r9_2),
4360 ""#px#" = tstbit($Rs, #0); if ("
4361 #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2",
4362 [], "", COMPOUND, TypeCOMPOUND> {
4367 let isPredicatedFalse = np;
4368 // tnt: Taken/Not Taken
4369 let isBrTaken = !if (!eq(tnt, "t"), "true", "false");
4370 let isTaken = !if (!eq(tnt, "t"), 1, 0);
4372 let IClass = 0b0001;
4373 let Inst{27-26} = 0b00;
4374 let Inst{25} = !if (!eq(px, "!p1"), 1,
4375 !if (!eq(px, "p1"), 1, 0));
4376 let Inst{24-23} = 0b11;
4378 let Inst{21-20} = r9_2{10-9};
4379 let Inst{19-16} = Rs;
4380 let Inst{13} = !if (!eq(tnt, "t"), 1, 0);
4381 let Inst{9-8} = 0b11;
4382 let Inst{7-1} = r9_2{8-2};
4385 let Defs = [PC, P0], Uses = [P0], isCodeGenOnly = 0 in {
4386 def J4_tstbit0_tp0_jump_nt : CJInst_tstbit_R0<"p0", 0, "nt">;
4387 def J4_tstbit0_tp0_jump_t : CJInst_tstbit_R0<"p0", 0, "t">;
4388 def J4_tstbit0_fp0_jump_nt : CJInst_tstbit_R0<"p0", 1, "nt">;
4389 def J4_tstbit0_fp0_jump_t : CJInst_tstbit_R0<"p0", 1, "t">;
4392 let Defs = [PC, P1], Uses = [P1], isCodeGenOnly = 0 in {
4393 def J4_tstbit0_tp1_jump_nt : CJInst_tstbit_R0<"p1", 0, "nt">;
4394 def J4_tstbit0_tp1_jump_t : CJInst_tstbit_R0<"p1", 0, "t">;
4395 def J4_tstbit0_fp1_jump_nt : CJInst_tstbit_R0<"p1", 1, "nt">;
4396 def J4_tstbit0_fp1_jump_t : CJInst_tstbit_R0<"p1", 1, "t">;
4400 let isBranch = 1, hasSideEffects = 0,
4401 isExtentSigned = 1, isPredicated = 1, isPredicatedNew = 1,
4402 isExtendable = 1, opExtentBits = 11, opExtentAlign = 2,
4403 opExtendable = 2, isTerminator = 1, validSubTargets = HasV4SubT in
4404 class CJInst_RR<string px, string op, bit np, string tnt>
4405 : InstHexagon<(outs), (ins IntRegs:$Rs, IntRegs:$Rt, brtarget:$r9_2),
4406 ""#px#" = cmp."#op#"($Rs, $Rt); if ("
4407 #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2",
4408 [], "", COMPOUND, TypeCOMPOUND> {
4414 let isPredicatedFalse = np;
4415 // tnt: Taken/Not Taken
4416 let isBrTaken = !if (!eq(tnt, "t"), "true", "false");
4417 let isTaken = !if (!eq(tnt, "t"), 1, 0);
4419 let IClass = 0b0001;
4420 let Inst{27-23} = !if (!eq(op, "eq"), 0b01000,
4421 !if (!eq(op, "gt"), 0b01001,
4422 !if (!eq(op, "gtu"), 0b01010, 0)));
4424 let Inst{21-20} = r9_2{10-9};
4425 let Inst{19-16} = Rs;
4426 let Inst{13} = !if (!eq(tnt, "t"), 1, 0);
4427 // px: Predicate reg 0/1
4428 let Inst{12} = !if (!eq(px, "!p1"), 1,
4429 !if (!eq(px, "p1"), 1, 0));
4430 let Inst{11-8} = Rt;
4431 let Inst{7-1} = r9_2{8-2};
4434 // P[10] taken/not taken.
4435 multiclass T_tnt_CJInst_RR<string op, bit np> {
4436 let Defs = [PC, P0], Uses = [P0] in {
4437 def NAME#p0_jump_nt : CJInst_RR<"p0", op, np, "nt">;
4438 def NAME#p0_jump_t : CJInst_RR<"p0", op, np, "t">;
4440 let Defs = [PC, P1], Uses = [P1] in {
4441 def NAME#p1_jump_nt : CJInst_RR<"p1", op, np, "nt">;
4442 def NAME#p1_jump_t : CJInst_RR<"p1", op, np, "t">;
4445 // Predicate / !Predicate
4446 multiclass T_pnp_CJInst_RR<string op>{
4447 defm J4_cmp#NAME#_t : T_tnt_CJInst_RR<op, 0>;
4448 defm J4_cmp#NAME#_f : T_tnt_CJInst_RR<op, 1>;
4450 // TypeCJ Instructions compare RR and jump
4451 let isCodeGenOnly = 0 in {
4452 defm eq : T_pnp_CJInst_RR<"eq">;
4453 defm gt : T_pnp_CJInst_RR<"gt">;
4454 defm gtu : T_pnp_CJInst_RR<"gtu">;
4457 let isBranch = 1, hasSideEffects = 0, isExtentSigned = 1,
4458 isPredicated = 1, isPredicatedNew = 1, isExtendable = 1, opExtentBits = 11,
4459 opExtentAlign = 2, opExtendable = 2, isTerminator = 1,
4460 validSubTargets = HasV4SubT in
4461 class CJInst_RU5<string px, string op, bit np, string tnt>
4462 : InstHexagon<(outs), (ins IntRegs:$Rs, u5Imm:$U5, brtarget:$r9_2),
4463 ""#px#" = cmp."#op#"($Rs, #$U5); if ("
4464 #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2",
4465 [], "", COMPOUND, TypeCOMPOUND> {
4471 let isPredicatedFalse = np;
4472 // tnt: Taken/Not Taken
4473 let isBrTaken = !if (!eq(tnt, "t"), "true", "false");
4474 let isTaken = !if (!eq(tnt, "t"), 1, 0);
4476 let IClass = 0b0001;
4477 let Inst{27-26} = 0b00;
4478 // px: Predicate reg 0/1
4479 let Inst{25} = !if (!eq(px, "!p1"), 1,
4480 !if (!eq(px, "p1"), 1, 0));
4481 let Inst{24-23} = !if (!eq(op, "eq"), 0b00,
4482 !if (!eq(op, "gt"), 0b01,
4483 !if (!eq(op, "gtu"), 0b10, 0)));
4485 let Inst{21-20} = r9_2{10-9};
4486 let Inst{19-16} = Rs;
4487 let Inst{13} = !if (!eq(tnt, "t"), 1, 0);
4488 let Inst{12-8} = U5;
4489 let Inst{7-1} = r9_2{8-2};
4491 // P[10] taken/not taken.
4492 multiclass T_tnt_CJInst_RU5<string op, bit np> {
4493 let Defs = [PC, P0], Uses = [P0] in {
4494 def NAME#p0_jump_nt : CJInst_RU5<"p0", op, np, "nt">;
4495 def NAME#p0_jump_t : CJInst_RU5<"p0", op, np, "t">;
4497 let Defs = [PC, P1], Uses = [P1] in {
4498 def NAME#p1_jump_nt : CJInst_RU5<"p1", op, np, "nt">;
4499 def NAME#p1_jump_t : CJInst_RU5<"p1", op, np, "t">;
4502 // Predicate / !Predicate
4503 multiclass T_pnp_CJInst_RU5<string op>{
4504 defm J4_cmp#NAME#i_t : T_tnt_CJInst_RU5<op, 0>;
4505 defm J4_cmp#NAME#i_f : T_tnt_CJInst_RU5<op, 1>;
4507 // TypeCJ Instructions compare RI and jump
4508 let isCodeGenOnly = 0 in {
4509 defm eq : T_pnp_CJInst_RU5<"eq">;
4510 defm gt : T_pnp_CJInst_RU5<"gt">;
4511 defm gtu : T_pnp_CJInst_RU5<"gtu">;
4514 let isBranch = 1, hasSideEffects = 0, isExtentSigned = 1,
4515 isPredicated = 1, isPredicatedFalse = 1, isPredicatedNew = 1,
4516 isExtendable = 1, opExtentBits = 11, opExtentAlign = 2, opExtendable = 1,
4517 isTerminator = 1, validSubTargets = HasV4SubT in
4518 class CJInst_Rn1<string px, string op, bit np, string tnt>
4519 : InstHexagon<(outs), (ins IntRegs:$Rs, brtarget:$r9_2),
4520 ""#px#" = cmp."#op#"($Rs,#-1); if ("
4521 #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2",
4522 [], "", COMPOUND, TypeCOMPOUND> {
4527 let isPredicatedFalse = np;
4528 // tnt: Taken/Not Taken
4529 let isBrTaken = !if (!eq(tnt, "t"), "true", "false");
4530 let isTaken = !if (!eq(tnt, "t"), 1, 0);
4532 let IClass = 0b0001;
4533 let Inst{27-26} = 0b00;
4534 let Inst{25} = !if (!eq(px, "!p1"), 1,
4535 !if (!eq(px, "p1"), 1, 0));
4537 let Inst{24-23} = 0b11;
4539 let Inst{21-20} = r9_2{10-9};
4540 let Inst{19-16} = Rs;
4541 let Inst{13} = !if (!eq(tnt, "t"), 1, 0);
4542 let Inst{9-8} = !if (!eq(op, "eq"), 0b00,
4543 !if (!eq(op, "gt"), 0b01, 0));
4544 let Inst{7-1} = r9_2{8-2};
4547 // P[10] taken/not taken.
4548 multiclass T_tnt_CJInst_Rn1<string op, bit np> {
4549 let Defs = [PC, P0], Uses = [P0] in {
4550 def NAME#p0_jump_nt : CJInst_Rn1<"p0", op, np, "nt">;
4551 def NAME#p0_jump_t : CJInst_Rn1<"p0", op, np, "t">;
4553 let Defs = [PC, P1], Uses = [P1] in {
4554 def NAME#p1_jump_nt : CJInst_Rn1<"p1", op, np, "nt">;
4555 def NAME#p1_jump_t : CJInst_Rn1<"p1", op, np, "t">;
4558 // Predicate / !Predicate
4559 multiclass T_pnp_CJInst_Rn1<string op>{
4560 defm J4_cmp#NAME#n1_t : T_tnt_CJInst_Rn1<op, 0>;
4561 defm J4_cmp#NAME#n1_f : T_tnt_CJInst_Rn1<op, 1>;
4563 // TypeCJ Instructions compare -1 and jump
4564 let isCodeGenOnly = 0 in {
4565 defm eq : T_pnp_CJInst_Rn1<"eq">;
4566 defm gt : T_pnp_CJInst_Rn1<"gt">;
4569 // J4_jumpseti: Direct unconditional jump and set register to immediate.
4570 let Defs = [PC], isBranch = 1, hasSideEffects = 0, hasNewValue = 1,
4571 isExtentSigned = 1, opNewValue = 0, isExtendable = 1, opExtentBits = 11,
4572 opExtentAlign = 2, opExtendable = 2, validSubTargets = HasV4SubT,
4573 isCodeGenOnly = 0 in
4574 def J4_jumpseti: CJInst <
4576 (ins u6Imm:$U6, brtarget:$r9_2),
4577 "$Rd = #$U6 ; jump $r9_2"> {
4582 let IClass = 0b0001;
4583 let Inst{27-24} = 0b0110;
4584 let Inst{21-20} = r9_2{10-9};
4585 let Inst{19-16} = Rd;
4586 let Inst{13-8} = U6;
4587 let Inst{7-1} = r9_2{8-2};
4590 // J4_jumpsetr: Direct unconditional jump and transfer register.
4591 let Defs = [PC], isBranch = 1, hasSideEffects = 0, hasNewValue = 1,
4592 isExtentSigned = 1, opNewValue = 0, isExtendable = 1, opExtentBits = 11,
4593 opExtentAlign = 2, opExtendable = 2, validSubTargets = HasV4SubT,
4594 isCodeGenOnly = 0 in
4595 def J4_jumpsetr: CJInst <
4597 (ins IntRegs:$Rs, brtarget:$r9_2),
4598 "$Rd = $Rs ; jump $r9_2"> {
4603 let IClass = 0b0001;
4604 let Inst{27-24} = 0b0111;
4605 let Inst{21-20} = r9_2{10-9};
4606 let Inst{11-8} = Rd;
4607 let Inst{19-16} = Rs;
4608 let Inst{7-1} = r9_2{8-2};