1 //=- HexagonInstrInfoV4.td - Target Desc. for Hexagon Target -*- tablegen -*-=//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the Hexagon V4 instructions in TableGen format.
12 //===----------------------------------------------------------------------===//
14 def addrga: PatLeaf<(i32 AddrGA:$Addr)>;
16 let hasSideEffects = 0 in
17 class T_Immext<Operand ImmType>
18 : EXTENDERInst<(outs), (ins ImmType:$imm),
19 "immext(#$imm)", []> {
23 let Inst{27-16} = imm{31-20};
24 let Inst{13-0} = imm{19-6};
27 def A4_ext : T_Immext<u26_6Imm>;
28 let isCodeGenOnly = 1 in {
30 def A4_ext_b : T_Immext<brtarget>;
32 def A4_ext_c : T_Immext<calltarget>;
33 def A4_ext_g : T_Immext<globaladdress>;
36 def BITPOS32 : SDNodeXForm<imm, [{
37 // Return the bit position we will set [0-31].
39 int32_t imm = N->getSExtValue();
40 return XformMskToBitPosU5Imm(imm);
43 // Fold (add (CONST32 tglobaladdr:$addr) <offset>) into a global address.
44 def FoldGlobalAddr : ComplexPattern<i32, 1, "foldGlobalAddress", [], []>;
46 // Fold (add (CONST32_GP tglobaladdr:$addr) <offset>) into a global address.
47 def FoldGlobalAddrGP : ComplexPattern<i32, 1, "foldGlobalAddressGP", [], []>;
49 def NumUsesBelowThresCONST32 : PatFrag<(ops node:$addr),
50 (HexagonCONST32 node:$addr), [{
51 return hasNumUsesBelowThresGA(N->getOperand(0).getNode());
54 // Hexagon V4 Architecture spec defines 8 instruction classes:
55 // LD ST ALU32 XTYPE J JR MEMOP NV CR SYSTEM(system is not implemented in the
59 // ========================================
60 // Loads (8/16/32/64 bit)
64 // ========================================
65 // Stores (8/16/32/64 bit)
68 // ALU32 Instructions:
69 // ========================================
70 // Arithmetic / Logical (32 bit)
73 // XTYPE Instructions (32/64 bit):
74 // ========================================
75 // Arithmetic, Logical, Bit Manipulation
76 // Multiply (Integer, Fractional, Complex)
77 // Permute / Vector Permute Operations
78 // Predicate Operations
79 // Shift / Shift with Add/Sub/Logical
81 // Vector Halfword (ALU, Shift, Multiply)
82 // Vector Word (ALU, Shift)
85 // ========================================
86 // Jump/Call PC-relative
89 // ========================================
92 // MEMOP Instructions:
93 // ========================================
94 // Operation on memory (8/16/32 bit)
97 // ========================================
102 // ========================================
103 // Control-Register Transfers
104 // Hardware Loop Setup
105 // Predicate Logicals & Reductions
107 // SYSTEM Instructions (not implemented in the compiler):
108 // ========================================
114 //===----------------------------------------------------------------------===//
116 //===----------------------------------------------------------------------===//
118 class T_ALU32_3op_not<string mnemonic, bits<3> MajOp, bits<3> MinOp,
120 : T_ALU32_3op<mnemonic, MajOp, MinOp, OpsRev, 0> {
121 let AsmString = "$Rd = "#mnemonic#"($Rs, ~$Rt)";
124 let BaseOpcode = "andn_rr", CextOpcode = "andn", isCodeGenOnly = 0 in
125 def A4_andn : T_ALU32_3op_not<"and", 0b001, 0b100, 1>;
126 let BaseOpcode = "orn_rr", CextOpcode = "orn", isCodeGenOnly = 0 in
127 def A4_orn : T_ALU32_3op_not<"or", 0b001, 0b101, 1>;
129 let CextOpcode = "rcmp.eq", isCodeGenOnly = 0 in
130 def A4_rcmpeq : T_ALU32_3op<"cmp.eq", 0b011, 0b010, 0, 1>;
131 let CextOpcode = "!rcmp.eq", isCodeGenOnly = 0 in
132 def A4_rcmpneq : T_ALU32_3op<"!cmp.eq", 0b011, 0b011, 0, 1>;
134 let isCodeGenOnly = 0 in {
135 def C4_cmpneq : T_ALU32_3op_cmp<"!cmp.eq", 0b00, 1, 1>;
136 def C4_cmplte : T_ALU32_3op_cmp<"!cmp.gt", 0b10, 1, 0>;
137 def C4_cmplteu : T_ALU32_3op_cmp<"!cmp.gtu", 0b11, 1, 0>;
140 // Pats for instruction selection.
142 // A class to embed the usual comparison patfrags within a zext to i32.
143 // The seteq/setne frags use "lhs" and "rhs" as operands, so use the same
144 // names, or else the frag's "body" won't match the operands.
145 class CmpInReg<PatFrag Op>
146 : PatFrag<(ops node:$lhs, node:$rhs),(i32 (zext (i1 Op.Fragment)))>;
148 def: T_cmp32_rr_pat<A4_rcmpeq, CmpInReg<seteq>, i32>;
149 def: T_cmp32_rr_pat<A4_rcmpneq, CmpInReg<setne>, i32>;
151 def: T_cmp32_rr_pat<C4_cmpneq, setne, i1>;
153 class T_CMP_rrbh<string mnemonic, bits<3> MinOp, bit IsComm>
154 : SInst<(outs PredRegs:$Pd), (ins IntRegs:$Rs, IntRegs:$Rt),
155 "$Pd = "#mnemonic#"($Rs, $Rt)", [], "", S_3op_tc_2early_SLOT23>,
157 let validSubTargets = HasV4SubT;
158 let InputType = "reg";
159 let CextOpcode = mnemonic;
161 let isCommutable = IsComm;
162 let hasSideEffects = 0;
169 let Inst{27-21} = 0b0111110;
170 let Inst{20-16} = Rs;
172 let Inst{7-5} = MinOp;
176 let isCodeGenOnly = 0 in {
177 def A4_cmpbeq : T_CMP_rrbh<"cmpb.eq", 0b110, 1>;
178 def A4_cmpbgt : T_CMP_rrbh<"cmpb.gt", 0b010, 0>;
179 def A4_cmpbgtu : T_CMP_rrbh<"cmpb.gtu", 0b111, 0>;
180 def A4_cmpheq : T_CMP_rrbh<"cmph.eq", 0b011, 1>;
181 def A4_cmphgt : T_CMP_rrbh<"cmph.gt", 0b100, 0>;
182 def A4_cmphgtu : T_CMP_rrbh<"cmph.gtu", 0b101, 0>;
185 class T_CMP_ribh<string mnemonic, bits<2> MajOp, bit IsHalf, bit IsComm,
186 Operand ImmType, bit IsImmExt, bit IsImmSigned, int ImmBits>
187 : ALU64Inst<(outs PredRegs:$Pd), (ins IntRegs:$Rs, ImmType:$Imm),
188 "$Pd = "#mnemonic#"($Rs, #$Imm)", [], "", ALU64_tc_2early_SLOT23>,
190 let validSubTargets = HasV4SubT;
191 let InputType = "imm";
192 let CextOpcode = mnemonic;
194 let isCommutable = IsComm;
195 let hasSideEffects = 0;
196 let isExtendable = IsImmExt;
197 let opExtendable = !if (IsImmExt, 2, 0);
198 let isExtentSigned = IsImmSigned;
199 let opExtentBits = ImmBits;
206 let Inst{27-24} = 0b1101;
207 let Inst{22-21} = MajOp;
208 let Inst{20-16} = Rs;
209 let Inst{12-5} = Imm;
211 let Inst{3} = IsHalf;
215 let isCodeGenOnly = 0 in {
216 def A4_cmpbeqi : T_CMP_ribh<"cmpb.eq", 0b00, 0, 1, u8Imm, 0, 0, 8>;
217 def A4_cmpbgti : T_CMP_ribh<"cmpb.gt", 0b01, 0, 0, s8Imm, 0, 1, 8>;
218 def A4_cmpbgtui : T_CMP_ribh<"cmpb.gtu", 0b10, 0, 0, u7Ext, 1, 0, 7>;
219 def A4_cmpheqi : T_CMP_ribh<"cmph.eq", 0b00, 1, 1, s8Ext, 1, 1, 8>;
220 def A4_cmphgti : T_CMP_ribh<"cmph.gt", 0b01, 1, 0, s8Ext, 1, 1, 8>;
221 def A4_cmphgtui : T_CMP_ribh<"cmph.gtu", 0b10, 1, 0, u7Ext, 1, 0, 7>;
223 class T_RCMP_EQ_ri<string mnemonic, bit IsNeg>
224 : ALU32_ri<(outs IntRegs:$Rd), (ins IntRegs:$Rs, s8Ext:$s8),
225 "$Rd = "#mnemonic#"($Rs, #$s8)", [], "", ALU32_2op_tc_1_SLOT0123>,
227 let validSubTargets = HasV4SubT;
228 let InputType = "imm";
229 let CextOpcode = !if (IsNeg, "!rcmp.eq", "rcmp.eq");
230 let isExtendable = 1;
231 let opExtendable = 2;
232 let isExtentSigned = 1;
233 let opExtentBits = 8;
241 let Inst{27-24} = 0b0011;
243 let Inst{21} = IsNeg;
244 let Inst{20-16} = Rs;
250 let isCodeGenOnly = 0 in {
251 def A4_rcmpeqi : T_RCMP_EQ_ri<"cmp.eq", 0>;
252 def A4_rcmpneqi : T_RCMP_EQ_ri<"!cmp.eq", 1>;
255 def: Pat<(i32 (zext (i1 (seteq (i32 IntRegs:$Rs), s8ExtPred:$s8)))),
256 (A4_rcmpeqi IntRegs:$Rs, s8ExtPred:$s8)>;
257 def: Pat<(i32 (zext (i1 (setne (i32 IntRegs:$Rs), s8ExtPred:$s8)))),
258 (A4_rcmpneqi IntRegs:$Rs, s8ExtPred:$s8)>;
260 // Preserve the S2_tstbit_r generation
261 def: Pat<(i32 (zext (i1 (setne (i32 (and (i32 (shl 1, (i32 IntRegs:$src2))),
262 (i32 IntRegs:$src1))), 0)))),
263 (C2_muxii (S2_tstbit_r IntRegs:$src1, IntRegs:$src2), 1, 0)>;
266 //===----------------------------------------------------------------------===//
268 //===----------------------------------------------------------------------===//
271 //===----------------------------------------------------------------------===//
273 //===----------------------------------------------------------------------===//
275 // Combine a word and an immediate into a register pair.
276 let hasSideEffects = 0, isExtentSigned = 1, isExtendable = 1,
278 class T_Combine1 <bits<2> MajOp, dag ins, string AsmStr>
279 : ALU32Inst <(outs DoubleRegs:$Rdd), ins, AsmStr> {
285 let Inst{27-24} = 0b0011;
286 let Inst{22-21} = MajOp;
287 let Inst{20-16} = Rs;
293 let opExtendable = 2, isCodeGenOnly = 0 in
294 def A4_combineri : T_Combine1<0b00, (ins IntRegs:$Rs, s8Ext:$s8),
295 "$Rdd = combine($Rs, #$s8)">;
297 let opExtendable = 1, isCodeGenOnly = 0 in
298 def A4_combineir : T_Combine1<0b01, (ins s8Ext:$s8, IntRegs:$Rs),
299 "$Rdd = combine(#$s8, $Rs)">;
301 def HexagonWrapperCombineRI_V4 :
302 SDNode<"HexagonISD::WrapperCombineRI_V4", SDTHexagonI64I32I32>;
303 def HexagonWrapperCombineIR_V4 :
304 SDNode<"HexagonISD::WrapperCombineIR_V4", SDTHexagonI64I32I32>;
306 def : Pat <(HexagonWrapperCombineRI_V4 IntRegs:$r, s8ExtPred:$i),
307 (A4_combineri IntRegs:$r, s8ExtPred:$i)>,
310 def : Pat <(HexagonWrapperCombineIR_V4 s8ExtPred:$i, IntRegs:$r),
311 (A4_combineir s8ExtPred:$i, IntRegs:$r)>,
314 // A4_combineii: Set two small immediates.
315 let hasSideEffects = 0, isExtendable = 1, opExtentBits = 6, opExtendable = 2 in
316 def A4_combineii: ALU32Inst<(outs DoubleRegs:$Rdd), (ins s8Imm:$s8, u6Ext:$U6),
317 "$Rdd = combine(#$s8, #$U6)"> {
323 let Inst{27-23} = 0b11001;
324 let Inst{20-16} = U6{5-1};
325 let Inst{13} = U6{0};
330 //===----------------------------------------------------------------------===//
332 //===----------------------------------------------------------------------===//
334 //===----------------------------------------------------------------------===//
336 //===----------------------------------------------------------------------===//
338 def Zext64: OutPatFrag<(ops node:$Rs),
339 (i64 (A4_combineir 0, (i32 $Rs)))>;
340 def Sext64: OutPatFrag<(ops node:$Rs),
341 (i64 (A2_sxtw (i32 $Rs)))>;
343 //===----------------------------------------------------------------------===//
344 // Template class for load instructions with Absolute set addressing mode.
345 //===----------------------------------------------------------------------===//
346 let isExtended = 1, opExtendable = 2, opExtentBits = 6, addrMode = AbsoluteSet,
347 hasSideEffects = 0 in
348 class T_LD_abs_set<string mnemonic, RegisterClass RC, bits<4>MajOp>:
349 LDInst<(outs RC:$dst1, IntRegs:$dst2),
351 "$dst1 = "#mnemonic#"($dst2 = #$addr)",
359 let Inst{27-25} = 0b101;
360 let Inst{24-21} = MajOp;
361 let Inst{13-12} = 0b01;
362 let Inst{4-0} = dst1;
363 let Inst{20-16} = dst2;
364 let Inst{11-8} = addr{5-2};
365 let Inst{6-5} = addr{1-0};
368 let accessSize = ByteAccess, hasNewValue = 1, isCodeGenOnly = 0 in {
369 def L4_loadrb_ap : T_LD_abs_set <"memb", IntRegs, 0b1000>;
370 def L4_loadrub_ap : T_LD_abs_set <"memub", IntRegs, 0b1001>;
373 let accessSize = HalfWordAccess, hasNewValue = 1, isCodeGenOnly = 0 in {
374 def L4_loadrh_ap : T_LD_abs_set <"memh", IntRegs, 0b1010>;
375 def L4_loadruh_ap : T_LD_abs_set <"memuh", IntRegs, 0b1011>;
378 let accessSize = WordAccess, hasNewValue = 1, isCodeGenOnly = 0 in
379 def L4_loadri_ap : T_LD_abs_set <"memw", IntRegs, 0b1100>;
381 let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
382 def L4_loadrd_ap : T_LD_abs_set <"memd", DoubleRegs, 0b1110>;
383 // Load - Indirect with long offset
384 let InputType = "imm", addrMode = BaseLongOffset, isExtended = 1,
385 opExtentBits = 6, opExtendable = 3 in
386 class T_LoadAbsReg <string mnemonic, string CextOp, RegisterClass RC,
388 : LDInst <(outs RC:$dst), (ins IntRegs:$src1, u2Imm:$src2, u6Ext:$src3),
389 "$dst = "#mnemonic#"($src1<<#$src2 + #$src3)",
395 let CextOpcode = CextOp;
396 let hasNewValue = !if (!eq(!cast<string>(RC), "DoubleRegs"), 0, 1);
399 let Inst{27-25} = 0b110;
400 let Inst{24-21} = MajOp;
401 let Inst{20-16} = src1;
402 let Inst{13} = src2{1};
404 let Inst{11-8} = src3{5-2};
405 let Inst{7} = src2{0};
406 let Inst{6-5} = src3{1-0};
410 let accessSize = ByteAccess, isCodeGenOnly = 0 in {
411 def L4_loadrb_ur : T_LoadAbsReg<"memb", "LDrib", IntRegs, 0b1000>;
412 def L4_loadrub_ur : T_LoadAbsReg<"memub", "LDriub", IntRegs, 0b1001>;
413 def L4_loadalignb_ur : T_LoadAbsReg<"memb_fifo", "LDrib_fifo",
417 let accessSize = HalfWordAccess, isCodeGenOnly = 0 in {
418 def L4_loadrh_ur : T_LoadAbsReg<"memh", "LDrih", IntRegs, 0b1010>;
419 def L4_loadruh_ur : T_LoadAbsReg<"memuh", "LDriuh", IntRegs, 0b1011>;
420 def L4_loadbsw2_ur : T_LoadAbsReg<"membh", "LDribh2", IntRegs, 0b0001>;
421 def L4_loadbzw2_ur : T_LoadAbsReg<"memubh", "LDriubh2", IntRegs, 0b0011>;
422 def L4_loadalignh_ur : T_LoadAbsReg<"memh_fifo", "LDrih_fifo",
426 let accessSize = WordAccess, isCodeGenOnly = 0 in {
427 def L4_loadri_ur : T_LoadAbsReg<"memw", "LDriw", IntRegs, 0b1100>;
428 def L4_loadbsw4_ur : T_LoadAbsReg<"membh", "LDribh4", DoubleRegs, 0b0111>;
429 def L4_loadbzw4_ur : T_LoadAbsReg<"memubh", "LDriubh4", DoubleRegs, 0b0101>;
432 let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
433 def L4_loadrd_ur : T_LoadAbsReg<"memd", "LDrid", DoubleRegs, 0b1110>;
436 multiclass T_LoadAbsReg_Pat <PatFrag ldOp, InstHexagon MI, ValueType VT = i32> {
437 def : Pat <(VT (ldOp (add (shl IntRegs:$src1, u2ImmPred:$src2),
438 (HexagonCONST32 tglobaladdr:$src3)))),
439 (MI IntRegs:$src1, u2ImmPred:$src2, tglobaladdr:$src3)>;
441 def : Pat <(VT (ldOp (add IntRegs:$src1,
442 (HexagonCONST32 tglobaladdr:$src2)))),
443 (MI IntRegs:$src1, 0, tglobaladdr:$src2)>;
446 let AddedComplexity = 60 in {
447 defm : T_LoadAbsReg_Pat <sextloadi8, L4_loadrb_ur>;
448 defm : T_LoadAbsReg_Pat <zextloadi8, L4_loadrub_ur>;
449 defm : T_LoadAbsReg_Pat <extloadi8, L4_loadrub_ur>;
451 defm : T_LoadAbsReg_Pat <sextloadi16, L4_loadrh_ur>;
452 defm : T_LoadAbsReg_Pat <zextloadi16, L4_loadruh_ur>;
453 defm : T_LoadAbsReg_Pat <extloadi16, L4_loadruh_ur>;
455 defm : T_LoadAbsReg_Pat <load, L4_loadri_ur>;
456 defm : T_LoadAbsReg_Pat <load, L4_loadrd_ur, i64>;
459 //===----------------------------------------------------------------------===//
460 // Template classes for the non-predicated load instructions with
461 // base + register offset addressing mode
462 //===----------------------------------------------------------------------===//
463 class T_load_rr <string mnemonic, RegisterClass RC, bits<3> MajOp>:
464 LDInst<(outs RC:$dst), (ins IntRegs:$src1, IntRegs:$src2, u2Imm:$u2),
465 "$dst = "#mnemonic#"($src1 + $src2<<#$u2)",
466 [], "", V4LDST_tc_ld_SLOT01>, ImmRegShl, AddrModeRel {
474 let Inst{27-24} = 0b1010;
475 let Inst{23-21} = MajOp;
476 let Inst{20-16} = src1;
477 let Inst{12-8} = src2;
478 let Inst{13} = u2{1};
483 //===----------------------------------------------------------------------===//
484 // Template classes for the predicated load instructions with
485 // base + register offset addressing mode
486 //===----------------------------------------------------------------------===//
487 let isPredicated = 1 in
488 class T_pload_rr <string mnemonic, RegisterClass RC, bits<3> MajOp,
489 bit isNot, bit isPredNew>:
490 LDInst <(outs RC:$dst),
491 (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3, u2Imm:$u2),
492 !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
493 ") ")#"$dst = "#mnemonic#"($src2+$src3<<#$u2)",
494 [], "", V4LDST_tc_ld_SLOT01>, AddrModeRel {
501 let isPredicatedFalse = isNot;
502 let isPredicatedNew = isPredNew;
506 let Inst{27-26} = 0b00;
507 let Inst{25} = isPredNew;
508 let Inst{24} = isNot;
509 let Inst{23-21} = MajOp;
510 let Inst{20-16} = src2;
511 let Inst{12-8} = src3;
512 let Inst{13} = u2{1};
514 let Inst{6-5} = src1;
518 //===----------------------------------------------------------------------===//
519 // multiclass for load instructions with base + register offset
521 //===----------------------------------------------------------------------===//
522 let hasSideEffects = 0, addrMode = BaseRegOffset in
523 multiclass ld_idxd_shl <string mnemonic, string CextOp, RegisterClass RC,
525 let CextOpcode = CextOp, BaseOpcode = CextOp#_indexed_shl,
526 InputType = "reg" in {
527 let isPredicable = 1 in
528 def L4_#NAME#_rr : T_load_rr <mnemonic, RC, MajOp>;
531 def L4_p#NAME#t_rr : T_pload_rr <mnemonic, RC, MajOp, 0, 0>;
532 def L4_p#NAME#f_rr : T_pload_rr <mnemonic, RC, MajOp, 1, 0>;
535 def L4_p#NAME#tnew_rr : T_pload_rr <mnemonic, RC, MajOp, 0, 1>;
536 def L4_p#NAME#fnew_rr : T_pload_rr <mnemonic, RC, MajOp, 1, 1>;
540 let hasNewValue = 1, accessSize = ByteAccess, isCodeGenOnly = 0 in {
541 defm loadrb : ld_idxd_shl<"memb", "LDrib", IntRegs, 0b000>;
542 defm loadrub : ld_idxd_shl<"memub", "LDriub", IntRegs, 0b001>;
545 let hasNewValue = 1, accessSize = HalfWordAccess, isCodeGenOnly = 0 in {
546 defm loadrh : ld_idxd_shl<"memh", "LDrih", IntRegs, 0b010>;
547 defm loadruh : ld_idxd_shl<"memuh", "LDriuh", IntRegs, 0b011>;
550 let hasNewValue = 1, accessSize = WordAccess, isCodeGenOnly = 0 in
551 defm loadri : ld_idxd_shl<"memw", "LDriw", IntRegs, 0b100>;
553 let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
554 defm loadrd : ld_idxd_shl<"memd", "LDrid", DoubleRegs, 0b110>;
556 // 'def pats' for load instructions with base + register offset and non-zero
557 // immediate value. Immediate value is used to left-shift the second
559 class Loadxs_pat<PatFrag Load, ValueType VT, InstHexagon MI>
560 : Pat<(VT (Load (add (i32 IntRegs:$Rs),
561 (i32 (shl (i32 IntRegs:$Rt), u2ImmPred:$u2))))),
562 (VT (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2))>;
564 let AddedComplexity = 40 in {
565 def: Loadxs_pat<extloadi8, i32, L4_loadrub_rr>;
566 def: Loadxs_pat<zextloadi8, i32, L4_loadrub_rr>;
567 def: Loadxs_pat<sextloadi8, i32, L4_loadrb_rr>;
568 def: Loadxs_pat<extloadi16, i32, L4_loadruh_rr>;
569 def: Loadxs_pat<zextloadi16, i32, L4_loadruh_rr>;
570 def: Loadxs_pat<sextloadi16, i32, L4_loadrh_rr>;
571 def: Loadxs_pat<load, i32, L4_loadri_rr>;
572 def: Loadxs_pat<load, i64, L4_loadrd_rr>;
575 // 'def pats' for load instruction base + register offset and
576 // zero immediate value.
577 class Loadxs_simple_pat<PatFrag Load, ValueType VT, InstHexagon MI>
578 : Pat<(VT (Load (add (i32 IntRegs:$Rs), (i32 IntRegs:$Rt)))),
579 (VT (MI IntRegs:$Rs, IntRegs:$Rt, 0))>;
581 let AddedComplexity = 20 in {
582 def: Loadxs_simple_pat<extloadi8, i32, L4_loadrub_rr>;
583 def: Loadxs_simple_pat<zextloadi8, i32, L4_loadrub_rr>;
584 def: Loadxs_simple_pat<sextloadi8, i32, L4_loadrb_rr>;
585 def: Loadxs_simple_pat<extloadi16, i32, L4_loadruh_rr>;
586 def: Loadxs_simple_pat<zextloadi16, i32, L4_loadruh_rr>;
587 def: Loadxs_simple_pat<sextloadi16, i32, L4_loadrh_rr>;
588 def: Loadxs_simple_pat<load, i32, L4_loadri_rr>;
589 def: Loadxs_simple_pat<load, i64, L4_loadrd_rr>;
593 def : Pat <(i64 (zext (i1 PredRegs:$src1))),
594 (i64 (A4_combineir 0, (C2_muxii (i1 PredRegs:$src1), 1, 0)))>,
598 def : Pat <(i64 (zext (i32 IntRegs:$src1))),
599 (i64 (A4_combineir 0, (i32 IntRegs:$src1)))>,
602 def: Pat <(i64 (zextloadi8 ADDRriS11_0:$src1)),
603 (i64 (A4_combineir 0, (L2_loadrub_io AddrFI:$src1, 0)))>,
606 let AddedComplexity = 20 in
607 def: Pat <(i64 (zextloadi8 (add (i32 IntRegs:$src1),
608 s11_0ExtPred:$offset))),
609 (i64 (A4_combineir 0, (L2_loadrub_io IntRegs:$src1,
610 s11_0ExtPred:$offset)))>,
614 def: Pat <(i64 (zextloadi1 ADDRriS11_0:$src1)),
615 (i64 (A4_combineir 0, (L2_loadrub_io AddrFI:$src1, 0)))>,
618 let AddedComplexity = 20 in
619 def: Pat <(i64 (zextloadi1 (add (i32 IntRegs:$src1),
620 s11_0ExtPred:$offset))),
621 (i64 (A4_combineir 0, (L2_loadrub_io IntRegs:$src1,
622 s11_0ExtPred:$offset)))>,
626 def: Pat <(i64 (zextloadi16 ADDRriS11_1:$src1)),
627 (i64 (A4_combineir 0, (L2_loadruh_io AddrFI:$src1, 0)))>,
630 let AddedComplexity = 20 in
631 def: Pat <(i64 (zextloadi16 (add (i32 IntRegs:$src1),
632 s11_1ExtPred:$offset))),
633 (i64 (A4_combineir 0, (L2_loadruh_io IntRegs:$src1,
634 s11_1ExtPred:$offset)))>,
638 def: Pat <(i64 (extloadi16 ADDRriS11_2:$src1)),
639 (i64 (A4_combineir 0, (L2_loadrh_io AddrFI:$src1, 0)))>,
642 let AddedComplexity = 20 in
643 def: Pat <(i64 (extloadi16 (add (i32 IntRegs:$src1),
644 s11_1ExtPred:$offset))),
645 (i64 (A4_combineir 0, (L2_loadrh_io IntRegs:$src1,
646 s11_1ExtPred:$offset)))>,
650 def: Pat <(i64 (zextloadi32 ADDRriS11_2:$src1)),
651 (i64 (A4_combineir 0, (L2_loadri_io AddrFI:$src1, 0)))>,
654 let AddedComplexity = 100 in
655 def: Pat <(i64 (zextloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))),
656 (i64 (A4_combineir 0, (L2_loadri_io IntRegs:$src1,
657 s11_2ExtPred:$offset)))>,
661 def: Pat <(i64 (extloadi32 ADDRriS11_2:$src1)),
662 (i64 (A4_combineir 0, (L2_loadri_io AddrFI:$src1, 0)))>,
665 let AddedComplexity = 100 in
666 def: Pat <(i64 (extloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))),
667 (i64 (A4_combineir 0, (L2_loadri_io IntRegs:$src1,
668 s11_2ExtPred:$offset)))>,
673 //===----------------------------------------------------------------------===//
675 //===----------------------------------------------------------------------===//
677 //===----------------------------------------------------------------------===//
679 //===----------------------------------------------------------------------===//
681 //===----------------------------------------------------------------------===//
682 // Template class for store instructions with Absolute set addressing mode.
683 //===----------------------------------------------------------------------===//
684 let isExtended = 1, opExtendable = 1, opExtentBits = 6,
685 addrMode = AbsoluteSet, isNVStorable = 1 in
686 class T_ST_absset <string mnemonic, string BaseOp, RegisterClass RC,
687 bits<3> MajOp, MemAccessSize AccessSz, bit isHalf = 0>
688 : STInst<(outs IntRegs:$dst),
689 (ins u6Ext:$addr, RC:$src),
690 mnemonic#"($dst = #$addr) = $src"#!if(isHalf, ".h","")>, NewValueRel {
694 let accessSize = AccessSz;
695 let BaseOpcode = BaseOp#"_AbsSet";
699 let Inst{27-24} = 0b1011;
700 let Inst{23-21} = MajOp;
701 let Inst{20-16} = dst;
703 let Inst{12-8} = src;
705 let Inst{5-0} = addr;
708 def S4_storerb_ap : T_ST_absset <"memb", "STrib", IntRegs, 0b000, ByteAccess>;
709 def S4_storerh_ap : T_ST_absset <"memh", "STrih", IntRegs, 0b010,
711 def S4_storeri_ap : T_ST_absset <"memw", "STriw", IntRegs, 0b100, WordAccess>;
713 let isNVStorable = 0 in {
714 def S4_storerf_ap : T_ST_absset <"memh", "STrif", IntRegs,
715 0b011, HalfWordAccess, 1>;
716 def S4_storerd_ap : T_ST_absset <"memd", "STrid", DoubleRegs,
717 0b110, DoubleWordAccess>;
720 let opExtendable = 1, isNewValue = 1, isNVStore = 1, opNewValue = 2,
721 isExtended = 1, opExtentBits= 6 in
722 class T_ST_absset_nv <string mnemonic, string BaseOp, bits<2> MajOp,
723 MemAccessSize AccessSz >
724 : NVInst <(outs IntRegs:$dst),
725 (ins u6Ext:$addr, IntRegs:$src),
726 mnemonic#"($dst = #$addr) = $src.new">, NewValueRel {
730 let accessSize = AccessSz;
731 let BaseOpcode = BaseOp#"_AbsSet";
735 let Inst{27-21} = 0b1011101;
736 let Inst{20-16} = dst;
737 let Inst{13-11} = 0b000;
738 let Inst{12-11} = MajOp;
739 let Inst{10-8} = src;
741 let Inst{5-0} = addr;
744 let mayStore = 1, addrMode = AbsoluteSet, isCodeGenOnly = 0 in {
745 def S4_storerbnew_ap : T_ST_absset_nv <"memb", "STrib", 0b00, ByteAccess>;
746 def S4_storerhnew_ap : T_ST_absset_nv <"memh", "STrih", 0b01, HalfWordAccess>;
747 def S4_storerinew_ap : T_ST_absset_nv <"memw", "STriw", 0b10, WordAccess>;
750 let isExtended = 1, opExtendable = 2, opExtentBits = 6, InputType = "imm",
751 addrMode = BaseLongOffset, AddedComplexity = 40 in
752 class T_StoreAbsReg <string mnemonic, string CextOp, RegisterClass RC,
753 bits<3> MajOp, MemAccessSize AccessSz, bit isHalf = 0>
755 (ins IntRegs:$src1, u2Imm:$src2, u6Ext:$src3, RC:$src4),
756 mnemonic#"($src1<<#$src2 + #$src3) = $src4"#!if(isHalf, ".h",""),
757 []>, ImmRegShl, NewValueRel {
764 let accessSize = AccessSz;
765 let CextOpcode = CextOp;
766 let BaseOpcode = CextOp#"_shl";
769 let Inst{27-24} =0b1101;
770 let Inst{23-21} = MajOp;
771 let Inst{20-16} = src1;
772 let Inst{13} = src2{1};
773 let Inst{12-8} = src4;
775 let Inst{6} = src2{0};
776 let Inst{5-0} = src3;
779 let isCodeGenOnly = 0 in {
780 def S4_storerb_ur : T_StoreAbsReg <"memb", "STrib", IntRegs, 0b000, ByteAccess>;
781 def S4_storerh_ur : T_StoreAbsReg <"memh", "STrih", IntRegs, 0b010,
783 def S4_storerf_ur : T_StoreAbsReg <"memh", "STrif", IntRegs, 0b011,
785 def S4_storeri_ur : T_StoreAbsReg <"memw", "STriw", IntRegs, 0b100, WordAccess>;
786 def S4_storerd_ur : T_StoreAbsReg <"memd", "STrid", DoubleRegs, 0b110,
790 let AddedComplexity = 40 in
791 multiclass T_StoreAbsReg_Pats <InstHexagon MI, RegisterClass RC, ValueType VT,
793 def : Pat<(stOp (VT RC:$src4),
794 (add (shl (i32 IntRegs:$src1), u2ImmPred:$src2),
795 u0AlwaysExtPred:$src3)),
796 (MI IntRegs:$src1, u2ImmPred:$src2, u0AlwaysExtPred:$src3, RC:$src4)>;
798 def : Pat<(stOp (VT RC:$src4),
799 (add (shl IntRegs:$src1, u2ImmPred:$src2),
800 (HexagonCONST32 tglobaladdr:$src3))),
801 (MI IntRegs:$src1, u2ImmPred:$src2, tglobaladdr:$src3, RC:$src4)>;
803 def : Pat<(stOp (VT RC:$src4),
804 (add IntRegs:$src1, (HexagonCONST32 tglobaladdr:$src3))),
805 (MI IntRegs:$src1, 0, tglobaladdr:$src3, RC:$src4)>;
808 defm : T_StoreAbsReg_Pats <S4_storerd_ur, DoubleRegs, i64, store>;
809 defm : T_StoreAbsReg_Pats <S4_storeri_ur, IntRegs, i32, store>;
810 defm : T_StoreAbsReg_Pats <S4_storerb_ur, IntRegs, i32, truncstorei8>;
811 defm : T_StoreAbsReg_Pats <S4_storerh_ur, IntRegs, i32, truncstorei16>;
813 let mayStore = 1, isNVStore = 1, isExtended = 1, addrMode = BaseLongOffset,
814 opExtentBits = 6, isNewValue = 1, opNewValue = 3, opExtendable = 2 in
815 class T_StoreAbsRegNV <string mnemonic, string CextOp, bits<2> MajOp,
816 MemAccessSize AccessSz>
818 (ins IntRegs:$src1, u2Imm:$src2, u6Ext:$src3, IntRegs:$src4),
819 mnemonic#"($src1<<#$src2 + #$src3) = $src4.new">, NewValueRel {
825 let CextOpcode = CextOp;
826 let BaseOpcode = CextOp#"_shl";
829 let Inst{27-21} = 0b1101101;
830 let Inst{12-11} = 0b00;
832 let Inst{20-16} = src1;
833 let Inst{13} = src2{1};
834 let Inst{12-11} = MajOp;
835 let Inst{10-8} = src4;
836 let Inst{6} = src2{0};
837 let Inst{5-0} = src3;
840 let isCodeGenOnly = 0 in {
841 def S4_storerbnew_ur : T_StoreAbsRegNV <"memb", "STrib", 0b00, ByteAccess>;
842 def S4_storerhnew_ur : T_StoreAbsRegNV <"memh", "STrih", 0b01, HalfWordAccess>;
843 def S4_storerinew_ur : T_StoreAbsRegNV <"memw", "STriw", 0b10, WordAccess>;
846 //===----------------------------------------------------------------------===//
847 // Template classes for the non-predicated store instructions with
848 // base + register offset addressing mode
849 //===----------------------------------------------------------------------===//
850 let isPredicable = 1 in
851 class T_store_rr <string mnemonic, RegisterClass RC, bits<3> MajOp, bit isH>
852 : STInst < (outs ), (ins IntRegs:$Rs, IntRegs:$Ru, u2Imm:$u2, RC:$Rt),
853 mnemonic#"($Rs + $Ru<<#$u2) = $Rt"#!if(isH, ".h",""),
854 [],"",V4LDST_tc_st_SLOT01>, ImmRegShl, AddrModeRel {
863 let Inst{27-24} = 0b1011;
864 let Inst{23-21} = MajOp;
865 let Inst{20-16} = Rs;
867 let Inst{13} = u2{1};
872 //===----------------------------------------------------------------------===//
873 // Template classes for the predicated store instructions with
874 // base + register offset addressing mode
875 //===----------------------------------------------------------------------===//
876 let isPredicated = 1 in
877 class T_pstore_rr <string mnemonic, RegisterClass RC, bits<3> MajOp,
878 bit isNot, bit isPredNew, bit isH>
880 (ins PredRegs:$Pv, IntRegs:$Rs, IntRegs:$Ru, u2Imm:$u2, RC:$Rt),
882 !if(isNot, "if (!$Pv", "if ($Pv")#!if(isPredNew, ".new) ",
883 ") ")#mnemonic#"($Rs+$Ru<<#$u2) = $Rt"#!if(isH, ".h",""),
884 [], "", V4LDST_tc_st_SLOT01> , AddrModeRel{
891 let isPredicatedFalse = isNot;
892 let isPredicatedNew = isPredNew;
896 let Inst{27-26} = 0b01;
897 let Inst{25} = isPredNew;
898 let Inst{24} = isNot;
899 let Inst{23-21} = MajOp;
900 let Inst{20-16} = Rs;
902 let Inst{13} = u2{1};
908 //===----------------------------------------------------------------------===//
909 // Template classes for the new-value store instructions with
910 // base + register offset addressing mode
911 //===----------------------------------------------------------------------===//
912 let isPredicable = 1, isNewValue = 1, opNewValue = 3 in
913 class T_store_new_rr <string mnemonic, bits<2> MajOp> :
914 NVInst < (outs ), (ins IntRegs:$Rs, IntRegs:$Ru, u2Imm:$u2, IntRegs:$Nt),
915 mnemonic#"($Rs + $Ru<<#$u2) = $Nt.new",
916 [],"",V4LDST_tc_st_SLOT0>, ImmRegShl, AddrModeRel {
925 let Inst{27-21} = 0b1011101;
926 let Inst{20-16} = Rs;
928 let Inst{13} = u2{1};
930 let Inst{4-3} = MajOp;
934 //===----------------------------------------------------------------------===//
935 // Template classes for the predicated new-value store instructions with
936 // base + register offset addressing mode
937 //===----------------------------------------------------------------------===//
938 let isPredicated = 1, isNewValue = 1, opNewValue = 4 in
939 class T_pstore_new_rr <string mnemonic, bits<2> MajOp, bit isNot, bit isPredNew>
941 (ins PredRegs:$Pv, IntRegs:$Rs, IntRegs:$Ru, u2Imm:$u2, IntRegs:$Nt),
942 !if(isNot, "if (!$Pv", "if ($Pv")#!if(isPredNew, ".new) ",
943 ") ")#mnemonic#"($Rs+$Ru<<#$u2) = $Nt.new",
944 [], "", V4LDST_tc_st_SLOT0>, AddrModeRel {
951 let isPredicatedFalse = isNot;
952 let isPredicatedNew = isPredNew;
955 let Inst{27-26} = 0b01;
956 let Inst{25} = isPredNew;
957 let Inst{24} = isNot;
958 let Inst{23-21} = 0b101;
959 let Inst{20-16} = Rs;
961 let Inst{13} = u2{1};
964 let Inst{4-3} = MajOp;
968 //===----------------------------------------------------------------------===//
969 // multiclass for store instructions with base + register offset addressing
971 //===----------------------------------------------------------------------===//
972 let isNVStorable = 1 in
973 multiclass ST_Idxd_shl<string mnemonic, string CextOp, RegisterClass RC,
974 bits<3> MajOp, bit isH = 0> {
975 let CextOpcode = CextOp, BaseOpcode = CextOp#_indexed_shl in {
976 def S4_#NAME#_rr : T_store_rr <mnemonic, RC, MajOp, isH>;
979 def S4_p#NAME#t_rr : T_pstore_rr <mnemonic, RC, MajOp, 0, 0, isH>;
980 def S4_p#NAME#f_rr : T_pstore_rr <mnemonic, RC, MajOp, 1, 0, isH>;
983 def S4_p#NAME#tnew_rr : T_pstore_rr <mnemonic, RC, MajOp, 0, 1, isH>;
984 def S4_p#NAME#fnew_rr : T_pstore_rr <mnemonic, RC, MajOp, 1, 1, isH>;
988 //===----------------------------------------------------------------------===//
989 // multiclass for new-value store instructions with base + register offset
991 //===----------------------------------------------------------------------===//
992 let mayStore = 1, isNVStore = 1 in
993 multiclass ST_Idxd_shl_nv <string mnemonic, string CextOp, RegisterClass RC,
995 let CextOpcode = CextOp, BaseOpcode = CextOp#_indexed_shl in {
996 def S4_#NAME#new_rr : T_store_new_rr<mnemonic, MajOp>;
999 def S4_p#NAME#newt_rr : T_pstore_new_rr <mnemonic, MajOp, 0, 0>;
1000 def S4_p#NAME#newf_rr : T_pstore_new_rr <mnemonic, MajOp, 1, 0>;
1003 def S4_p#NAME#newtnew_rr : T_pstore_new_rr <mnemonic, MajOp, 0, 1>;
1004 def S4_p#NAME#newfnew_rr : T_pstore_new_rr <mnemonic, MajOp, 1, 1>;
1008 let addrMode = BaseRegOffset, InputType = "reg", hasSideEffects = 0,
1009 isCodeGenOnly = 0 in {
1010 let accessSize = ByteAccess in
1011 defm storerb: ST_Idxd_shl<"memb", "STrib", IntRegs, 0b000>,
1012 ST_Idxd_shl_nv<"memb", "STrib", IntRegs, 0b00>;
1014 let accessSize = HalfWordAccess in
1015 defm storerh: ST_Idxd_shl<"memh", "STrih", IntRegs, 0b010>,
1016 ST_Idxd_shl_nv<"memh", "STrih", IntRegs, 0b01>;
1018 let accessSize = WordAccess in
1019 defm storeri: ST_Idxd_shl<"memw", "STriw", IntRegs, 0b100>,
1020 ST_Idxd_shl_nv<"memw", "STriw", IntRegs, 0b10>;
1022 let isNVStorable = 0, accessSize = DoubleWordAccess in
1023 defm storerd: ST_Idxd_shl<"memd", "STrid", DoubleRegs, 0b110>;
1025 let isNVStorable = 0, accessSize = HalfWordAccess in
1026 defm storerf: ST_Idxd_shl<"memh", "STrif", IntRegs, 0b011, 1>;
1029 class Storexs_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
1030 : Pat<(Store Value:$Ru, (add (i32 IntRegs:$Rs),
1031 (i32 (shl (i32 IntRegs:$Rt), u2ImmPred:$u2)))),
1032 (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2, Value:$Ru)>;
1034 let AddedComplexity = 40 in {
1035 def: Storexs_pat<truncstorei8, I32, S4_storerb_rr>;
1036 def: Storexs_pat<truncstorei16, I32, S4_storerh_rr>;
1037 def: Storexs_pat<store, I32, S4_storeri_rr>;
1038 def: Storexs_pat<store, I64, S4_storerd_rr>;
1041 // memd(Rx++#s4:3)=Rtt
1042 // memd(Rx++#s4:3:circ(Mu))=Rtt
1043 // memd(Rx++I:circ(Mu))=Rtt
1045 // memd(Rx++Mu:brev)=Rtt
1046 // memd(gp+#u16:3)=Rtt
1048 // Store doubleword conditionally.
1049 // if ([!]Pv[.new]) memd(#u6)=Rtt
1050 // TODO: needs to be implemented.
1052 //===----------------------------------------------------------------------===//
1054 //===----------------------------------------------------------------------===//
1055 let isPredicable = 1, isExtendable = 1, isExtentSigned = 1, opExtentBits = 8,
1057 class T_StoreImm <string mnemonic, Operand OffsetOp, bits<2> MajOp >
1058 : STInst <(outs ), (ins IntRegs:$Rs, OffsetOp:$offset, s8Ext:$S8),
1059 mnemonic#"($Rs+#$offset)=#$S8",
1060 [], "", V4LDST_tc_st_SLOT01>,
1061 ImmRegRel, PredNewRel {
1067 string OffsetOpStr = !cast<string>(OffsetOp);
1068 let offsetBits = !if (!eq(OffsetOpStr, "u6_2Imm"), offset{7-2},
1069 !if (!eq(OffsetOpStr, "u6_1Imm"), offset{6-1},
1070 /* u6_0Imm */ offset{5-0}));
1072 let IClass = 0b0011;
1074 let Inst{27-25} = 0b110;
1075 let Inst{22-21} = MajOp;
1076 let Inst{20-16} = Rs;
1077 let Inst{12-7} = offsetBits;
1078 let Inst{13} = S8{7};
1079 let Inst{6-0} = S8{6-0};
1082 let isPredicated = 1, isExtendable = 1, isExtentSigned = 1, opExtentBits = 6,
1084 class T_StoreImm_pred <string mnemonic, Operand OffsetOp, bits<2> MajOp,
1085 bit isPredNot, bit isPredNew >
1087 (ins PredRegs:$Pv, IntRegs:$Rs, OffsetOp:$offset, s6Ext:$S6),
1088 !if(isPredNot, "if (!$Pv", "if ($Pv")#!if(isPredNew, ".new) ",
1089 ") ")#mnemonic#"($Rs+#$offset)=#$S6",
1090 [], "", V4LDST_tc_st_SLOT01>,
1091 ImmRegRel, PredNewRel {
1098 string OffsetOpStr = !cast<string>(OffsetOp);
1099 let offsetBits = !if (!eq(OffsetOpStr, "u6_2Imm"), offset{7-2},
1100 !if (!eq(OffsetOpStr, "u6_1Imm"), offset{6-1},
1101 /* u6_0Imm */ offset{5-0}));
1102 let isPredicatedNew = isPredNew;
1103 let isPredicatedFalse = isPredNot;
1105 let IClass = 0b0011;
1107 let Inst{27-25} = 0b100;
1108 let Inst{24} = isPredNew;
1109 let Inst{23} = isPredNot;
1110 let Inst{22-21} = MajOp;
1111 let Inst{20-16} = Rs;
1112 let Inst{13} = S6{5};
1113 let Inst{12-7} = offsetBits;
1115 let Inst{4-0} = S6{4-0};
1119 //===----------------------------------------------------------------------===//
1120 // multiclass for store instructions with base + immediate offset
1121 // addressing mode and immediate stored value.
1122 // mem[bhw](Rx++#s4:3)=#s8
1123 // if ([!]Pv[.new]) mem[bhw](Rx++#s4:3)=#s6
1124 //===----------------------------------------------------------------------===//
1126 multiclass ST_Imm_Pred <string mnemonic, Operand OffsetOp, bits<2> MajOp,
1128 def _io : T_StoreImm_pred <mnemonic, OffsetOp, MajOp, PredNot, 0>;
1130 def new_io : T_StoreImm_pred <mnemonic, OffsetOp, MajOp, PredNot, 1>;
1133 multiclass ST_Imm <string mnemonic, string CextOp, Operand OffsetOp,
1135 let CextOpcode = CextOp, BaseOpcode = CextOp#_imm in {
1136 def _io : T_StoreImm <mnemonic, OffsetOp, MajOp>;
1138 defm t : ST_Imm_Pred <mnemonic, OffsetOp, MajOp, 0>;
1139 defm f : ST_Imm_Pred <mnemonic, OffsetOp, MajOp, 1>;
1143 let hasSideEffects = 0, validSubTargets = HasV4SubT, addrMode = BaseImmOffset,
1144 InputType = "imm", isCodeGenOnly = 0 in {
1145 let accessSize = ByteAccess in
1146 defm S4_storeirb : ST_Imm<"memb", "STrib", u6_0Imm, 0b00>;
1148 let accessSize = HalfWordAccess in
1149 defm S4_storeirh : ST_Imm<"memh", "STrih", u6_1Imm, 0b01>;
1151 let accessSize = WordAccess in
1152 defm S4_storeiri : ST_Imm<"memw", "STriw", u6_2Imm, 0b10>;
1155 let Predicates = [HasV4T], AddedComplexity = 10 in {
1156 def: Pat<(truncstorei8 s8ExtPred:$src3, (add IntRegs:$src1, u6_0ImmPred:$src2)),
1157 (S4_storeirb_io IntRegs:$src1, u6_0ImmPred:$src2, s8ExtPred:$src3)>;
1159 def: Pat<(truncstorei16 s8ExtPred:$src3, (add IntRegs:$src1,
1160 u6_1ImmPred:$src2)),
1161 (S4_storeirh_io IntRegs:$src1, u6_1ImmPred:$src2, s8ExtPred:$src3)>;
1163 def: Pat<(store s8ExtPred:$src3, (add IntRegs:$src1, u6_2ImmPred:$src2)),
1164 (S4_storeiri_io IntRegs:$src1, u6_2ImmPred:$src2, s8ExtPred:$src3)>;
1167 let AddedComplexity = 6 in
1168 def : Pat <(truncstorei8 s8ExtPred:$src2, (i32 IntRegs:$src1)),
1169 (S4_storeirb_io IntRegs:$src1, 0, s8ExtPred:$src2)>,
1172 // memb(Rx++#s4:0:circ(Mu))=Rt
1173 // memb(Rx++I:circ(Mu))=Rt
1175 // memb(Rx++Mu:brev)=Rt
1176 // memb(gp+#u16:0)=Rt
1180 // TODO: needs to be implemented
1181 // memh(Re=#U6)=Rt.H
1182 // memh(Rs+#s11:1)=Rt.H
1183 let AddedComplexity = 6 in
1184 def : Pat <(truncstorei16 s8ExtPred:$src2, (i32 IntRegs:$src1)),
1185 (S4_storeirh_io IntRegs:$src1, 0, s8ExtPred:$src2)>,
1188 // memh(Rs+Ru<<#u2)=Rt.H
1189 // TODO: needs to be implemented.
1191 // memh(Ru<<#u2+#U6)=Rt.H
1192 // memh(Rx++#s4:1:circ(Mu))=Rt.H
1193 // memh(Rx++#s4:1:circ(Mu))=Rt
1194 // memh(Rx++I:circ(Mu))=Rt.H
1195 // memh(Rx++I:circ(Mu))=Rt
1196 // memh(Rx++Mu)=Rt.H
1198 // memh(Rx++Mu:brev)=Rt.H
1199 // memh(Rx++Mu:brev)=Rt
1200 // memh(gp+#u16:1)=Rt
1201 // if ([!]Pv[.new]) memh(#u6)=Rt.H
1202 // if ([!]Pv[.new]) memh(#u6)=Rt
1205 // if ([!]Pv[.new]) memh(Rs+#u6:1)=Rt.H
1206 // TODO: needs to be implemented.
1208 // if ([!]Pv[.new]) memh(Rx++#s4:1)=Rt.H
1209 // TODO: Needs to be implemented.
1213 // TODO: Needs to be implemented.
1216 let hasSideEffects = 0 in
1217 def STriw_pred_V4 : STInst2<(outs),
1218 (ins MEMri:$addr, PredRegs:$src1),
1219 "Error; should not emit",
1223 let AddedComplexity = 6 in
1224 def : Pat <(store s8ExtPred:$src2, (i32 IntRegs:$src1)),
1225 (S4_storeiri_io IntRegs:$src1, 0, s8ExtPred:$src2)>,
1228 // memw(Rx++#s4:2)=Rt
1229 // memw(Rx++#s4:2:circ(Mu))=Rt
1230 // memw(Rx++I:circ(Mu))=Rt
1232 // memw(Rx++Mu:brev)=Rt
1234 //===----------------------------------------------------------------------===
1236 //===----------------------------------------------------------------------===
1239 //===----------------------------------------------------------------------===//
1241 //===----------------------------------------------------------------------===//
1243 let opNewValue = 2, opExtendable = 1, isExtentSigned = 1, isPredicable = 1 in
1244 class T_store_io_nv <string mnemonic, RegisterClass RC,
1245 Operand ImmOp, bits<2>MajOp>
1246 : NVInst_V4 <(outs),
1247 (ins IntRegs:$src1, ImmOp:$src2, RC:$src3),
1248 mnemonic#"($src1+#$src2) = $src3.new",
1249 [],"",ST_tc_st_SLOT0> {
1251 bits<13> src2; // Actual address offset
1253 bits<11> offsetBits; // Represents offset encoding
1255 let opExtentBits = !if (!eq(mnemonic, "memb"), 11,
1256 !if (!eq(mnemonic, "memh"), 12,
1257 !if (!eq(mnemonic, "memw"), 13, 0)));
1259 let opExtentAlign = !if (!eq(mnemonic, "memb"), 0,
1260 !if (!eq(mnemonic, "memh"), 1,
1261 !if (!eq(mnemonic, "memw"), 2, 0)));
1263 let offsetBits = !if (!eq(mnemonic, "memb"), src2{10-0},
1264 !if (!eq(mnemonic, "memh"), src2{11-1},
1265 !if (!eq(mnemonic, "memw"), src2{12-2}, 0)));
1267 let IClass = 0b1010;
1270 let Inst{26-25} = offsetBits{10-9};
1271 let Inst{24-21} = 0b1101;
1272 let Inst{20-16} = src1;
1273 let Inst{13} = offsetBits{8};
1274 let Inst{12-11} = MajOp;
1275 let Inst{10-8} = src3;
1276 let Inst{7-0} = offsetBits{7-0};
1279 let opExtendable = 2, opNewValue = 3, isPredicated = 1 in
1280 class T_pstore_io_nv <string mnemonic, RegisterClass RC, Operand predImmOp,
1281 bits<2>MajOp, bit PredNot, bit isPredNew>
1282 : NVInst_V4 <(outs),
1283 (ins PredRegs:$src1, IntRegs:$src2, predImmOp:$src3, RC:$src4),
1284 !if(PredNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
1285 ") ")#mnemonic#"($src2+#$src3) = $src4.new",
1286 [],"",V2LDST_tc_st_SLOT0> {
1291 bits<6> offsetBits; // Represents offset encoding
1293 let isPredicatedNew = isPredNew;
1294 let isPredicatedFalse = PredNot;
1295 let opExtentBits = !if (!eq(mnemonic, "memb"), 6,
1296 !if (!eq(mnemonic, "memh"), 7,
1297 !if (!eq(mnemonic, "memw"), 8, 0)));
1299 let opExtentAlign = !if (!eq(mnemonic, "memb"), 0,
1300 !if (!eq(mnemonic, "memh"), 1,
1301 !if (!eq(mnemonic, "memw"), 2, 0)));
1303 let offsetBits = !if (!eq(mnemonic, "memb"), src3{5-0},
1304 !if (!eq(mnemonic, "memh"), src3{6-1},
1305 !if (!eq(mnemonic, "memw"), src3{7-2}, 0)));
1307 let IClass = 0b0100;
1310 let Inst{26} = PredNot;
1311 let Inst{25} = isPredNew;
1312 let Inst{24-21} = 0b0101;
1313 let Inst{20-16} = src2;
1314 let Inst{13} = offsetBits{5};
1315 let Inst{12-11} = MajOp;
1316 let Inst{10-8} = src4;
1317 let Inst{7-3} = offsetBits{4-0};
1319 let Inst{1-0} = src1;
1322 // multiclass for new-value store instructions with base + immediate offset.
1324 let mayStore = 1, isNVStore = 1, isNewValue = 1, hasSideEffects = 0,
1326 multiclass ST_Idxd_nv<string mnemonic, string CextOp, RegisterClass RC,
1327 Operand ImmOp, Operand predImmOp, bits<2> MajOp> {
1329 let CextOpcode = CextOp, BaseOpcode = CextOp#_indexed in {
1330 def S2_#NAME#new_io : T_store_io_nv <mnemonic, RC, ImmOp, MajOp>;
1332 def S2_p#NAME#newt_io :T_pstore_io_nv <mnemonic, RC, predImmOp, MajOp, 0, 0>;
1333 def S2_p#NAME#newf_io :T_pstore_io_nv <mnemonic, RC, predImmOp, MajOp, 1, 0>;
1335 def S4_p#NAME#newtnew_io :T_pstore_io_nv <mnemonic, RC, predImmOp,
1337 def S4_p#NAME#newfnew_io :T_pstore_io_nv <mnemonic, RC, predImmOp,
1342 let addrMode = BaseImmOffset, InputType = "imm", isCodeGenOnly = 0 in {
1343 let accessSize = ByteAccess in
1344 defm storerb: ST_Idxd_nv<"memb", "STrib", IntRegs, s11_0Ext,
1345 u6_0Ext, 0b00>, AddrModeRel;
1347 let accessSize = HalfWordAccess, opExtentAlign = 1 in
1348 defm storerh: ST_Idxd_nv<"memh", "STrih", IntRegs, s11_1Ext,
1349 u6_1Ext, 0b01>, AddrModeRel;
1351 let accessSize = WordAccess, opExtentAlign = 2 in
1352 defm storeri: ST_Idxd_nv<"memw", "STriw", IntRegs, s11_2Ext,
1353 u6_2Ext, 0b10>, AddrModeRel;
1356 //===----------------------------------------------------------------------===//
1357 // Post increment loads with register offset.
1358 //===----------------------------------------------------------------------===//
1360 let hasNewValue = 1, isCodeGenOnly = 0 in
1361 def L2_loadbsw2_pr : T_load_pr <"membh", IntRegs, 0b0001, HalfWordAccess>;
1363 let isCodeGenOnly = 0 in
1364 def L2_loadbsw4_pr : T_load_pr <"membh", DoubleRegs, 0b0111, WordAccess>;
1366 //===----------------------------------------------------------------------===//
1367 // Template class for non-predicated post increment .new stores
1368 // mem[bhwd](Rx++#s4:[0123])=Nt.new
1369 //===----------------------------------------------------------------------===//
1370 let isPredicable = 1, hasSideEffects = 0, validSubTargets = HasV4SubT,
1371 addrMode = PostInc, isNVStore = 1, isNewValue = 1, opNewValue = 3 in
1372 class T_StorePI_nv <string mnemonic, Operand ImmOp, bits<2> MajOp >
1373 : NVInstPI_V4 <(outs IntRegs:$_dst_),
1374 (ins IntRegs:$src1, ImmOp:$offset, IntRegs:$src2),
1375 mnemonic#"($src1++#$offset) = $src2.new",
1376 [], "$src1 = $_dst_">,
1383 string ImmOpStr = !cast<string>(ImmOp);
1384 let offsetBits = !if (!eq(ImmOpStr, "s4_2Imm"), offset{5-2},
1385 !if (!eq(ImmOpStr, "s4_1Imm"), offset{4-1},
1386 /* s4_0Imm */ offset{3-0}));
1387 let IClass = 0b1010;
1389 let Inst{27-21} = 0b1011101;
1390 let Inst{20-16} = src1;
1392 let Inst{12-11} = MajOp;
1393 let Inst{10-8} = src2;
1395 let Inst{6-3} = offsetBits;
1399 //===----------------------------------------------------------------------===//
1400 // Template class for predicated post increment .new stores
1401 // if([!]Pv[.new]) mem[bhwd](Rx++#s4:[0123])=Nt.new
1402 //===----------------------------------------------------------------------===//
1403 let isPredicated = 1, hasSideEffects = 0, validSubTargets = HasV4SubT,
1404 addrMode = PostInc, isNVStore = 1, isNewValue = 1, opNewValue = 4 in
1405 class T_StorePI_nv_pred <string mnemonic, Operand ImmOp,
1406 bits<2> MajOp, bit isPredNot, bit isPredNew >
1407 : NVInstPI_V4 <(outs IntRegs:$_dst_),
1408 (ins PredRegs:$src1, IntRegs:$src2,
1409 ImmOp:$offset, IntRegs:$src3),
1410 !if(isPredNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
1411 ") ")#mnemonic#"($src2++#$offset) = $src3.new",
1412 [], "$src2 = $_dst_">,
1420 string ImmOpStr = !cast<string>(ImmOp);
1421 let offsetBits = !if (!eq(ImmOpStr, "s4_2Imm"), offset{5-2},
1422 !if (!eq(ImmOpStr, "s4_1Imm"), offset{4-1},
1423 /* s4_0Imm */ offset{3-0}));
1424 let isPredicatedNew = isPredNew;
1425 let isPredicatedFalse = isPredNot;
1427 let IClass = 0b1010;
1429 let Inst{27-21} = 0b1011101;
1430 let Inst{20-16} = src2;
1432 let Inst{12-11} = MajOp;
1433 let Inst{10-8} = src3;
1434 let Inst{7} = isPredNew;
1435 let Inst{6-3} = offsetBits;
1436 let Inst{2} = isPredNot;
1437 let Inst{1-0} = src1;
1440 multiclass ST_PostInc_Pred_nv<string mnemonic, Operand ImmOp,
1441 bits<2> MajOp, bit PredNot> {
1442 def _pi : T_StorePI_nv_pred <mnemonic, ImmOp, MajOp, PredNot, 0>;
1445 def new_pi : T_StorePI_nv_pred <mnemonic, ImmOp, MajOp, PredNot, 1>;
1448 multiclass ST_PostInc_nv<string mnemonic, string BaseOp, Operand ImmOp,
1450 let BaseOpcode = "POST_"#BaseOp in {
1451 def S2_#NAME#_pi : T_StorePI_nv <mnemonic, ImmOp, MajOp>;
1454 defm S2_p#NAME#t : ST_PostInc_Pred_nv <mnemonic, ImmOp, MajOp, 0>;
1455 defm S2_p#NAME#f : ST_PostInc_Pred_nv <mnemonic, ImmOp, MajOp, 1>;
1459 let accessSize = ByteAccess, isCodeGenOnly = 0 in
1460 defm storerbnew: ST_PostInc_nv <"memb", "STrib", s4_0Imm, 0b00>;
1462 let accessSize = HalfWordAccess, isCodeGenOnly = 0 in
1463 defm storerhnew: ST_PostInc_nv <"memh", "STrih", s4_1Imm, 0b01>;
1465 let accessSize = WordAccess, isCodeGenOnly = 0 in
1466 defm storerinew: ST_PostInc_nv <"memw", "STriw", s4_2Imm, 0b10>;
1468 //===----------------------------------------------------------------------===//
1469 // Template class for post increment .new stores with register offset
1470 //===----------------------------------------------------------------------===//
1471 let isNewValue = 1, mayStore = 1, isNVStore = 1, opNewValue = 3 in
1472 class T_StorePI_RegNV <string mnemonic, bits<2> MajOp, MemAccessSize AccessSz>
1473 : NVInstPI_V4 <(outs IntRegs:$_dst_),
1474 (ins IntRegs:$src1, ModRegs:$src2, IntRegs:$src3),
1475 #mnemonic#"($src1++$src2) = $src3.new",
1476 [], "$src1 = $_dst_"> {
1480 let accessSize = AccessSz;
1482 let IClass = 0b1010;
1484 let Inst{27-21} = 0b1101101;
1485 let Inst{20-16} = src1;
1486 let Inst{13} = src2;
1487 let Inst{12-11} = MajOp;
1488 let Inst{10-8} = src3;
1492 let isCodeGenOnly = 0 in {
1493 def S2_storerbnew_pr : T_StorePI_RegNV<"memb", 0b00, ByteAccess>;
1494 def S2_storerhnew_pr : T_StorePI_RegNV<"memh", 0b01, HalfWordAccess>;
1495 def S2_storerinew_pr : T_StorePI_RegNV<"memw", 0b10, WordAccess>;
1498 // memb(Rx++#s4:0:circ(Mu))=Nt.new
1499 // memb(Rx++I:circ(Mu))=Nt.new
1500 // memb(Rx++Mu)=Nt.new
1501 // memb(Rx++Mu:brev)=Nt.new
1502 // memh(Rx++#s4:1:circ(Mu))=Nt.new
1503 // memh(Rx++I:circ(Mu))=Nt.new
1504 // memh(Rx++Mu)=Nt.new
1505 // memh(Rx++Mu:brev)=Nt.new
1507 // memw(Rx++#s4:2:circ(Mu))=Nt.new
1508 // memw(Rx++I:circ(Mu))=Nt.new
1509 // memw(Rx++Mu)=Nt.new
1510 // memw(Rx++Mu:brev)=Nt.new
1512 //===----------------------------------------------------------------------===//
1514 //===----------------------------------------------------------------------===//
1516 //===----------------------------------------------------------------------===//
1518 //===----------------------------------------------------------------------===//
1520 //===----------------------------------------------------------------------===//
1521 // multiclass/template class for the new-value compare jumps with the register
1523 //===----------------------------------------------------------------------===//
1525 let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 11,
1526 opExtentAlign = 2 in
1527 class NVJrr_template<string mnemonic, bits<3> majOp, bit NvOpNum,
1528 bit isNegCond, bit isTak>
1530 (ins IntRegs:$src1, IntRegs:$src2, brtarget:$offset),
1531 "if ("#!if(isNegCond, "!","")#mnemonic#
1532 "($src1"#!if(!eq(NvOpNum, 0),".new, ",", ")#
1533 "$src2"#!if(!eq(NvOpNum, 1),".new))","))")#" jump:"
1534 #!if(isTak, "t","nt")#" $offset", []> {
1538 bits<3> Ns; // New-Value Operand
1539 bits<5> RegOp; // Non-New-Value Operand
1542 let isTaken = isTak;
1543 let isPredicatedFalse = isNegCond;
1544 let opNewValue{0} = NvOpNum;
1546 let Ns = !if(!eq(NvOpNum, 0), src1{2-0}, src2{2-0});
1547 let RegOp = !if(!eq(NvOpNum, 0), src2, src1);
1549 let IClass = 0b0010;
1551 let Inst{25-23} = majOp;
1552 let Inst{22} = isNegCond;
1553 let Inst{18-16} = Ns;
1554 let Inst{13} = isTak;
1555 let Inst{12-8} = RegOp;
1556 let Inst{21-20} = offset{10-9};
1557 let Inst{7-1} = offset{8-2};
1561 multiclass NVJrr_cond<string mnemonic, bits<3> majOp, bit NvOpNum,
1563 // Branch not taken:
1564 def _nt_V4: NVJrr_template<mnemonic, majOp, NvOpNum, isNegCond, 0>;
1566 def _t_V4: NVJrr_template<mnemonic, majOp, NvOpNum, isNegCond, 1>;
1569 // NvOpNum = 0 -> First Operand is a new-value Register
1570 // NvOpNum = 1 -> Second Operand is a new-value Register
1572 multiclass NVJrr_base<string mnemonic, string BaseOp, bits<3> majOp,
1574 let BaseOpcode = BaseOp#_NVJ in {
1575 defm _t_Jumpnv : NVJrr_cond<mnemonic, majOp, NvOpNum, 0>; // True cond
1576 defm _f_Jumpnv : NVJrr_cond<mnemonic, majOp, NvOpNum, 1>; // False cond
1580 // if ([!]cmp.eq(Ns.new,Rt)) jump:[n]t #r9:2
1581 // if ([!]cmp.gt(Ns.new,Rt)) jump:[n]t #r9:2
1582 // if ([!]cmp.gtu(Ns.new,Rt)) jump:[n]t #r9:2
1583 // if ([!]cmp.gt(Rt,Ns.new)) jump:[n]t #r9:2
1584 // if ([!]cmp.gtu(Rt,Ns.new)) jump:[n]t #r9:2
1586 let isPredicated = 1, isBranch = 1, isNewValue = 1, isTerminator = 1,
1587 Defs = [PC], hasSideEffects = 0, validSubTargets = HasV4SubT,
1588 isCodeGenOnly = 0 in {
1589 defm CMPEQrr : NVJrr_base<"cmp.eq", "CMPEQ", 0b000, 0>, PredRel;
1590 defm CMPGTrr : NVJrr_base<"cmp.gt", "CMPGT", 0b001, 0>, PredRel;
1591 defm CMPGTUrr : NVJrr_base<"cmp.gtu", "CMPGTU", 0b010, 0>, PredRel;
1592 defm CMPLTrr : NVJrr_base<"cmp.gt", "CMPLT", 0b011, 1>, PredRel;
1593 defm CMPLTUrr : NVJrr_base<"cmp.gtu", "CMPLTU", 0b100, 1>, PredRel;
1596 //===----------------------------------------------------------------------===//
1597 // multiclass/template class for the new-value compare jumps instruction
1598 // with a register and an unsigned immediate (U5) operand.
1599 //===----------------------------------------------------------------------===//
1601 let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 11,
1602 opExtentAlign = 2 in
1603 class NVJri_template<string mnemonic, bits<3> majOp, bit isNegCond,
1606 (ins IntRegs:$src1, u5Imm:$src2, brtarget:$offset),
1607 "if ("#!if(isNegCond, "!","")#mnemonic#"($src1.new, #$src2)) jump:"
1608 #!if(isTak, "t","nt")#" $offset", []> {
1610 let isTaken = isTak;
1611 let isPredicatedFalse = isNegCond;
1612 let isTaken = isTak;
1618 let IClass = 0b0010;
1620 let Inst{25-23} = majOp;
1621 let Inst{22} = isNegCond;
1622 let Inst{18-16} = src1;
1623 let Inst{13} = isTak;
1624 let Inst{12-8} = src2;
1625 let Inst{21-20} = offset{10-9};
1626 let Inst{7-1} = offset{8-2};
1629 multiclass NVJri_cond<string mnemonic, bits<3> majOp, bit isNegCond> {
1630 // Branch not taken:
1631 def _nt_V4: NVJri_template<mnemonic, majOp, isNegCond, 0>;
1633 def _t_V4: NVJri_template<mnemonic, majOp, isNegCond, 1>;
1636 multiclass NVJri_base<string mnemonic, string BaseOp, bits<3> majOp> {
1637 let BaseOpcode = BaseOp#_NVJri in {
1638 defm _t_Jumpnv : NVJri_cond<mnemonic, majOp, 0>; // True Cond
1639 defm _f_Jumpnv : NVJri_cond<mnemonic, majOp, 1>; // False cond
1643 // if ([!]cmp.eq(Ns.new,#U5)) jump:[n]t #r9:2
1644 // if ([!]cmp.gt(Ns.new,#U5)) jump:[n]t #r9:2
1645 // if ([!]cmp.gtu(Ns.new,#U5)) jump:[n]t #r9:2
1647 let isPredicated = 1, isBranch = 1, isNewValue = 1, isTerminator = 1,
1648 Defs = [PC], hasSideEffects = 0, validSubTargets = HasV4SubT,
1649 isCodeGenOnly = 0 in {
1650 defm CMPEQri : NVJri_base<"cmp.eq", "CMPEQ", 0b000>, PredRel;
1651 defm CMPGTri : NVJri_base<"cmp.gt", "CMPGT", 0b001>, PredRel;
1652 defm CMPGTUri : NVJri_base<"cmp.gtu", "CMPGTU", 0b010>, PredRel;
1655 //===----------------------------------------------------------------------===//
1656 // multiclass/template class for the new-value compare jumps instruction
1657 // with a register and an hardcoded 0/-1 immediate value.
1658 //===----------------------------------------------------------------------===//
1660 let isExtendable = 1, opExtendable = 1, isExtentSigned = 1, opExtentBits = 11,
1661 opExtentAlign = 2 in
1662 class NVJ_ConstImm_template<string mnemonic, bits<3> majOp, string ImmVal,
1663 bit isNegCond, bit isTak>
1665 (ins IntRegs:$src1, brtarget:$offset),
1666 "if ("#!if(isNegCond, "!","")#mnemonic
1667 #"($src1.new, #"#ImmVal#")) jump:"
1668 #!if(isTak, "t","nt")#" $offset", []> {
1670 let isTaken = isTak;
1671 let isPredicatedFalse = isNegCond;
1672 let isTaken = isTak;
1676 let IClass = 0b0010;
1678 let Inst{25-23} = majOp;
1679 let Inst{22} = isNegCond;
1680 let Inst{18-16} = src1;
1681 let Inst{13} = isTak;
1682 let Inst{21-20} = offset{10-9};
1683 let Inst{7-1} = offset{8-2};
1686 multiclass NVJ_ConstImm_cond<string mnemonic, bits<3> majOp, string ImmVal,
1688 // Branch not taken:
1689 def _nt_V4: NVJ_ConstImm_template<mnemonic, majOp, ImmVal, isNegCond, 0>;
1691 def _t_V4: NVJ_ConstImm_template<mnemonic, majOp, ImmVal, isNegCond, 1>;
1694 multiclass NVJ_ConstImm_base<string mnemonic, string BaseOp, bits<3> majOp,
1696 let BaseOpcode = BaseOp#_NVJ_ConstImm in {
1697 defm _t_Jumpnv : NVJ_ConstImm_cond<mnemonic, majOp, ImmVal, 0>; // True
1698 defm _f_Jumpnv : NVJ_ConstImm_cond<mnemonic, majOp, ImmVal, 1>; // False
1702 // if ([!]tstbit(Ns.new,#0)) jump:[n]t #r9:2
1703 // if ([!]cmp.eq(Ns.new,#-1)) jump:[n]t #r9:2
1704 // if ([!]cmp.gt(Ns.new,#-1)) jump:[n]t #r9:2
1706 let isPredicated = 1, isBranch = 1, isNewValue = 1, isTerminator=1,
1707 Defs = [PC], hasSideEffects = 0, isCodeGenOnly = 0 in {
1708 defm TSTBIT0 : NVJ_ConstImm_base<"tstbit", "TSTBIT", 0b011, "0">, PredRel;
1709 defm CMPEQn1 : NVJ_ConstImm_base<"cmp.eq", "CMPEQ", 0b100, "-1">, PredRel;
1710 defm CMPGTn1 : NVJ_ConstImm_base<"cmp.gt", "CMPGT", 0b101, "-1">, PredRel;
1713 // J4_hintjumpr: Hint indirect conditional jump.
1714 let isBranch = 1, isIndirectBranch = 1, hasSideEffects = 0, isCodeGenOnly = 0 in
1715 def J4_hintjumpr: JRInst <
1720 let IClass = 0b0101;
1721 let Inst{27-21} = 0b0010101;
1722 let Inst{20-16} = Rs;
1725 //===----------------------------------------------------------------------===//
1727 //===----------------------------------------------------------------------===//
1729 //===----------------------------------------------------------------------===//
1731 //===----------------------------------------------------------------------===//
1734 let hasNewValue = 1, isExtendable = 1, opExtendable = 1,
1735 isExtentSigned = 0, opExtentBits = 6, hasSideEffects = 0,
1736 Uses = [PC], validSubTargets = HasV4SubT, isCodeGenOnly = 0 in
1737 def C4_addipc : CRInst <(outs IntRegs:$Rd), (ins u6Ext:$u6),
1738 "$Rd = add(pc, #$u6)", [], "", CR_tc_2_SLOT3 > {
1742 let IClass = 0b0110;
1743 let Inst{27-16} = 0b101001001001;
1744 let Inst{12-7} = u6;
1750 let hasSideEffects = 0 in
1751 class T_LOGICAL_3OP<string MnOp1, string MnOp2, bits<2> OpBits, bit IsNeg>
1752 : CRInst<(outs PredRegs:$Pd),
1753 (ins PredRegs:$Ps, PredRegs:$Pt, PredRegs:$Pu),
1754 "$Pd = " # MnOp1 # "($Ps, " # MnOp2 # "($Pt, " #
1755 !if (IsNeg,"!","") # "$Pu))",
1756 [], "", CR_tc_2early_SLOT23> {
1762 let IClass = 0b0110;
1763 let Inst{27-24} = 0b1011;
1764 let Inst{23} = IsNeg;
1765 let Inst{22-21} = OpBits;
1767 let Inst{17-16} = Ps;
1774 let isCodeGenOnly = 0 in {
1775 def C4_and_and : T_LOGICAL_3OP<"and", "and", 0b00, 0>;
1776 def C4_and_or : T_LOGICAL_3OP<"and", "or", 0b01, 0>;
1777 def C4_or_and : T_LOGICAL_3OP<"or", "and", 0b10, 0>;
1778 def C4_or_or : T_LOGICAL_3OP<"or", "or", 0b11, 0>;
1779 def C4_and_andn : T_LOGICAL_3OP<"and", "and", 0b00, 1>;
1780 def C4_and_orn : T_LOGICAL_3OP<"and", "or", 0b01, 1>;
1781 def C4_or_andn : T_LOGICAL_3OP<"or", "and", 0b10, 1>;
1782 def C4_or_orn : T_LOGICAL_3OP<"or", "or", 0b11, 1>;
1785 //===----------------------------------------------------------------------===//
1787 //===----------------------------------------------------------------------===//
1789 //===----------------------------------------------------------------------===//
1791 //===----------------------------------------------------------------------===//
1793 // Logical with-not instructions.
1794 let validSubTargets = HasV4SubT, isCodeGenOnly = 0 in {
1795 def A4_andnp : T_ALU64_logical<"and", 0b001, 1, 0, 1>;
1796 def A4_ornp : T_ALU64_logical<"or", 0b011, 1, 0, 1>;
1799 let hasNewValue = 1, hasSideEffects = 0, isCodeGenOnly = 0 in
1800 def S4_parity: ALU64Inst<(outs IntRegs:$Rd), (ins IntRegs:$Rs, IntRegs:$Rt),
1801 "$Rd = parity($Rs, $Rt)", [], "", ALU64_tc_2_SLOT23> {
1806 let IClass = 0b1101;
1807 let Inst{27-21} = 0b0101111;
1808 let Inst{20-16} = Rs;
1809 let Inst{12-8} = Rt;
1812 // Add and accumulate.
1813 // Rd=add(Rs,add(Ru,#s6))
1814 let isExtentSigned = 1, hasNewValue = 1, isExtendable = 1, opExtentBits = 6,
1815 opExtendable = 3, isCodeGenOnly = 0 in
1816 def S4_addaddi : ALU64Inst <(outs IntRegs:$Rd),
1817 (ins IntRegs:$Rs, IntRegs:$Ru, s6Ext:$s6),
1818 "$Rd = add($Rs, add($Ru, #$s6))" ,
1819 [(set (i32 IntRegs:$Rd), (add (i32 IntRegs:$Rs),
1820 (add (i32 IntRegs:$Ru), s6_16ExtPred:$s6)))],
1821 "", ALU64_tc_2_SLOT23> {
1827 let IClass = 0b1101;
1829 let Inst{27-23} = 0b10110;
1830 let Inst{22-21} = s6{5-4};
1831 let Inst{20-16} = Rs;
1832 let Inst{13} = s6{3};
1833 let Inst{12-8} = Rd;
1834 let Inst{7-5} = s6{2-0};
1838 let isExtentSigned = 1, hasSideEffects = 0, hasNewValue = 1, isExtendable = 1,
1839 opExtentBits = 6, opExtendable = 2, isCodeGenOnly = 0 in
1840 def S4_subaddi: ALU64Inst <(outs IntRegs:$Rd),
1841 (ins IntRegs:$Rs, s6Ext:$s6, IntRegs:$Ru),
1842 "$Rd = add($Rs, sub(#$s6, $Ru))",
1843 [], "", ALU64_tc_2_SLOT23> {
1849 let IClass = 0b1101;
1851 let Inst{27-23} = 0b10111;
1852 let Inst{22-21} = s6{5-4};
1853 let Inst{20-16} = Rs;
1854 let Inst{13} = s6{3};
1855 let Inst{12-8} = Rd;
1856 let Inst{7-5} = s6{2-0};
1861 // Rdd=extract(Rss,#u6,#U6)
1862 // Rdd=extract(Rss,Rtt)
1863 // Rd=extract(Rs,Rtt)
1864 // Rd=extract(Rs,#u5,#U5)
1866 let isCodeGenOnly = 0 in {
1867 def S4_extractp_rp : T_S3op_64 < "extract", 0b11, 0b100, 0>;
1868 def S4_extractp : T_S2op_extract <"extract", 0b1010, DoubleRegs, u6Imm>;
1871 let hasNewValue = 1, isCodeGenOnly = 0 in {
1872 def S4_extract_rp : T_S3op_extract<"extract", 0b01>;
1873 def S4_extract : T_S2op_extract <"extract", 0b1101, IntRegs, u5Imm>;
1876 // Complex add/sub halfwords/words
1877 let Defs = [USR_OVF], isCodeGenOnly = 0 in {
1878 def S4_vxaddsubh : T_S3op_64 < "vxaddsubh", 0b01, 0b100, 0, 1>;
1879 def S4_vxaddsubw : T_S3op_64 < "vxaddsubw", 0b01, 0b000, 0, 1>;
1880 def S4_vxsubaddh : T_S3op_64 < "vxsubaddh", 0b01, 0b110, 0, 1>;
1881 def S4_vxsubaddw : T_S3op_64 < "vxsubaddw", 0b01, 0b010, 0, 1>;
1884 let Defs = [USR_OVF], isCodeGenOnly = 0 in {
1885 def S4_vxaddsubhr : T_S3op_64 < "vxaddsubh", 0b11, 0b000, 0, 1, 1, 1>;
1886 def S4_vxsubaddhr : T_S3op_64 < "vxsubaddh", 0b11, 0b010, 0, 1, 1, 1>;
1889 let Itinerary = M_tc_3x_SLOT23, Defs = [USR_OVF], isCodeGenOnly = 0 in {
1890 def M4_mac_up_s1_sat: T_MType_acc_rr<"+= mpy", 0b011, 0b000, 0, [], 0, 1, 1>;
1891 def M4_nac_up_s1_sat: T_MType_acc_rr<"-= mpy", 0b011, 0b001, 0, [], 0, 1, 1>;
1894 // Logical xor with xor accumulation.
1895 // Rxx^=xor(Rss,Rtt)
1896 let hasSideEffects = 0, isCodeGenOnly = 0 in
1898 : SInst <(outs DoubleRegs:$Rxx),
1899 (ins DoubleRegs:$dst2, DoubleRegs:$Rss, DoubleRegs:$Rtt),
1900 "$Rxx ^= xor($Rss, $Rtt)",
1901 [(set (i64 DoubleRegs:$Rxx),
1902 (xor (i64 DoubleRegs:$dst2), (xor (i64 DoubleRegs:$Rss),
1903 (i64 DoubleRegs:$Rtt))))],
1904 "$dst2 = $Rxx", S_3op_tc_1_SLOT23> {
1909 let IClass = 0b1100;
1911 let Inst{27-23} = 0b10101;
1912 let Inst{20-16} = Rss;
1913 let Inst{12-8} = Rtt;
1914 let Inst{4-0} = Rxx;
1917 // Rotate and reduce bytes
1918 // Rdd=vrcrotate(Rss,Rt,#u2)
1919 let hasSideEffects = 0, isCodeGenOnly = 0 in
1921 : SInst <(outs DoubleRegs:$Rdd),
1922 (ins DoubleRegs:$Rss, IntRegs:$Rt, u2Imm:$u2),
1923 "$Rdd = vrcrotate($Rss, $Rt, #$u2)",
1924 [], "", S_3op_tc_3x_SLOT23> {
1930 let IClass = 0b1100;
1932 let Inst{27-22} = 0b001111;
1933 let Inst{20-16} = Rss;
1934 let Inst{13} = u2{1};
1935 let Inst{12-8} = Rt;
1936 let Inst{7-6} = 0b11;
1937 let Inst{5} = u2{0};
1938 let Inst{4-0} = Rdd;
1941 // Rotate and reduce bytes with accumulation
1942 // Rxx+=vrcrotate(Rss,Rt,#u2)
1943 let hasSideEffects = 0, isCodeGenOnly = 0 in
1944 def S4_vrcrotate_acc
1945 : SInst <(outs DoubleRegs:$Rxx),
1946 (ins DoubleRegs:$dst2, DoubleRegs:$Rss, IntRegs:$Rt, u2Imm:$u2),
1947 "$Rxx += vrcrotate($Rss, $Rt, #$u2)", [],
1948 "$dst2 = $Rxx", S_3op_tc_3x_SLOT23> {
1954 let IClass = 0b1100;
1956 let Inst{27-21} = 0b1011101;
1957 let Inst{20-16} = Rss;
1958 let Inst{13} = u2{1};
1959 let Inst{12-8} = Rt;
1960 let Inst{5} = u2{0};
1961 let Inst{4-0} = Rxx;
1965 // Vector reduce conditional negate halfwords
1966 let hasSideEffects = 0, isCodeGenOnly = 0 in
1968 : SInst <(outs DoubleRegs:$Rxx),
1969 (ins DoubleRegs:$dst2, DoubleRegs:$Rss, IntRegs:$Rt),
1970 "$Rxx += vrcnegh($Rss, $Rt)", [],
1971 "$dst2 = $Rxx", S_3op_tc_3x_SLOT23> {
1976 let IClass = 0b1100;
1978 let Inst{27-21} = 0b1011001;
1979 let Inst{20-16} = Rss;
1981 let Inst{12-8} = Rt;
1982 let Inst{7-5} = 0b111;
1983 let Inst{4-0} = Rxx;
1987 let isCodeGenOnly = 0 in
1988 def A4_bitspliti : T_S2op_2_di <"bitsplit", 0b110, 0b100>;
1990 // Arithmetic/Convergent round
1991 let isCodeGenOnly = 0 in
1992 def A4_cround_ri : T_S2op_2_ii <"cround", 0b111, 0b000>;
1994 let isCodeGenOnly = 0 in
1995 def A4_round_ri : T_S2op_2_ii <"round", 0b111, 0b100>;
1997 let Defs = [USR_OVF], isCodeGenOnly = 0 in
1998 def A4_round_ri_sat : T_S2op_2_ii <"round", 0b111, 0b110, 1>;
2000 // Logical-logical words.
2001 // Compound or-and -- Rx=or(Ru,and(Rx,#s10))
2002 let isExtentSigned = 1, hasNewValue = 1, isExtendable = 1, opExtentBits = 10,
2003 opExtendable = 3, isCodeGenOnly = 0 in
2005 ALU64Inst<(outs IntRegs:$Rx),
2006 (ins IntRegs:$Ru, IntRegs:$_src_, s10Ext:$s10),
2007 "$Rx = or($Ru, and($_src_, #$s10))" ,
2008 [(set (i32 IntRegs:$Rx),
2009 (or (i32 IntRegs:$Ru), (and (i32 IntRegs:$_src_), s10ExtPred:$s10)))] ,
2010 "$_src_ = $Rx", ALU64_tc_2_SLOT23> {
2015 let IClass = 0b1101;
2017 let Inst{27-22} = 0b101001;
2018 let Inst{20-16} = Rx;
2019 let Inst{21} = s10{9};
2020 let Inst{13-5} = s10{8-0};
2024 // Miscellaneous ALU64 instructions.
2026 let hasNewValue = 1, hasSideEffects = 0, isCodeGenOnly = 0 in
2027 def A4_modwrapu: ALU64Inst<(outs IntRegs:$Rd), (ins IntRegs:$Rs, IntRegs:$Rt),
2028 "$Rd = modwrap($Rs, $Rt)", [], "", ALU64_tc_2_SLOT23> {
2033 let IClass = 0b1101;
2034 let Inst{27-21} = 0b0011111;
2035 let Inst{20-16} = Rs;
2036 let Inst{12-8} = Rt;
2037 let Inst{7-5} = 0b111;
2041 let hasSideEffects = 0, isCodeGenOnly = 0 in
2042 def A4_bitsplit: ALU64Inst<(outs DoubleRegs:$Rd),
2043 (ins IntRegs:$Rs, IntRegs:$Rt),
2044 "$Rd = bitsplit($Rs, $Rt)", [], "", ALU64_tc_1_SLOT23> {
2049 let IClass = 0b1101;
2050 let Inst{27-24} = 0b0100;
2052 let Inst{20-16} = Rs;
2053 let Inst{12-8} = Rt;
2057 let isCodeGenOnly = 0 in {
2058 // Rx[&|]=xor(Rs,Rt)
2059 def M4_or_xor : T_MType_acc_rr < "|= xor", 0b110, 0b001, 0>;
2060 def M4_and_xor : T_MType_acc_rr < "&= xor", 0b010, 0b010, 0>;
2062 // Rx[&|^]=or(Rs,Rt)
2063 def M4_xor_or : T_MType_acc_rr < "^= or", 0b110, 0b011, 0>;
2065 let CextOpcode = "ORr_ORr" in
2066 def M4_or_or : T_MType_acc_rr < "|= or", 0b110, 0b000, 0>;
2067 def M4_and_or : T_MType_acc_rr < "&= or", 0b010, 0b001, 0>;
2069 // Rx[&|^]=and(Rs,Rt)
2070 def M4_xor_and : T_MType_acc_rr < "^= and", 0b110, 0b010, 0>;
2072 let CextOpcode = "ORr_ANDr" in
2073 def M4_or_and : T_MType_acc_rr < "|= and", 0b010, 0b011, 0>;
2074 def M4_and_and : T_MType_acc_rr < "&= and", 0b010, 0b000, 0>;
2076 // Rx[&|^]=and(Rs,~Rt)
2077 def M4_xor_andn : T_MType_acc_rr < "^= and", 0b001, 0b010, 0, [], 1>;
2078 def M4_or_andn : T_MType_acc_rr < "|= and", 0b001, 0b000, 0, [], 1>;
2079 def M4_and_andn : T_MType_acc_rr < "&= and", 0b001, 0b001, 0, [], 1>;
2082 // Compound or-or and or-and
2083 let isExtentSigned = 1, InputType = "imm", hasNewValue = 1, isExtendable = 1,
2084 opExtentBits = 10, opExtendable = 3 in
2085 class T_CompOR <string mnemonic, bits<2> MajOp, SDNode OpNode>
2086 : MInst_acc <(outs IntRegs:$Rx),
2087 (ins IntRegs:$src1, IntRegs:$Rs, s10Ext:$s10),
2088 "$Rx |= "#mnemonic#"($Rs, #$s10)",
2089 [(set (i32 IntRegs:$Rx), (or (i32 IntRegs:$src1),
2090 (OpNode (i32 IntRegs:$Rs), s10ExtPred:$s10)))],
2091 "$src1 = $Rx", ALU64_tc_2_SLOT23>, ImmRegRel {
2096 let IClass = 0b1101;
2098 let Inst{27-24} = 0b1010;
2099 let Inst{23-22} = MajOp;
2100 let Inst{20-16} = Rs;
2101 let Inst{21} = s10{9};
2102 let Inst{13-5} = s10{8-0};
2106 let CextOpcode = "ORr_ANDr", isCodeGenOnly = 0 in
2107 def S4_or_andi : T_CompOR <"and", 0b00, and>;
2109 let CextOpcode = "ORr_ORr", isCodeGenOnly = 0 in
2110 def S4_or_ori : T_CompOR <"or", 0b10, or>;
2113 // Rd=modwrap(Rs,Rt)
2115 // Rd=cround(Rs,#u5)
2117 // Rd=round(Rs,#u5)[:sat]
2118 // Rd=round(Rs,Rt)[:sat]
2119 // Vector reduce add unsigned halfwords
2120 // Rd=vraddh(Rss,Rtt)
2122 // Rdd=vaddb(Rss,Rtt)
2123 // Vector conditional negate
2124 // Rdd=vcnegh(Rss,Rt)
2125 // Rxx+=vrcnegh(Rss,Rt)
2126 // Vector maximum bytes
2127 // Rdd=vmaxb(Rtt,Rss)
2128 // Vector reduce maximum halfwords
2129 // Rxx=vrmaxh(Rss,Ru)
2130 // Rxx=vrmaxuh(Rss,Ru)
2131 // Vector reduce maximum words
2132 // Rxx=vrmaxuw(Rss,Ru)
2133 // Rxx=vrmaxw(Rss,Ru)
2134 // Vector minimum bytes
2135 // Rdd=vminb(Rtt,Rss)
2136 // Vector reduce minimum halfwords
2137 // Rxx=vrminh(Rss,Ru)
2138 // Rxx=vrminuh(Rss,Ru)
2139 // Vector reduce minimum words
2140 // Rxx=vrminuw(Rss,Ru)
2141 // Rxx=vrminw(Rss,Ru)
2142 // Vector subtract bytes
2143 // Rdd=vsubb(Rss,Rtt)
2145 //===----------------------------------------------------------------------===//
2147 //===----------------------------------------------------------------------===//
2149 //===----------------------------------------------------------------------===//
2151 //===----------------------------------------------------------------------===//
2154 let isCodeGenOnly = 0 in
2155 def S2_brevp : T_S2op_3 <"brev", 0b11, 0b110>;
2158 let isCodeGenOnly = 0 in {
2159 def S2_ct0p : T_COUNT_LEADING_64<"ct0", 0b111, 0b010>;
2160 def S2_ct1p : T_COUNT_LEADING_64<"ct1", 0b111, 0b100>;
2161 def S4_clbpnorm : T_COUNT_LEADING_64<"normamt", 0b011, 0b000>;
2164 def: Pat<(i32 (trunc (cttz (i64 DoubleRegs:$Rss)))),
2165 (S2_ct0p (i64 DoubleRegs:$Rss))>;
2166 def: Pat<(i32 (trunc (cttz (not (i64 DoubleRegs:$Rss))))),
2167 (S2_ct1p (i64 DoubleRegs:$Rss))>;
2169 let hasSideEffects = 0, hasNewValue = 1, isCodeGenOnly = 0 in
2170 def S4_clbaddi : SInst<(outs IntRegs:$Rd), (ins IntRegs:$Rs, s6Imm:$s6),
2171 "$Rd = add(clb($Rs), #$s6)", [], "", S_2op_tc_2_SLOT23> {
2175 let IClass = 0b1000;
2176 let Inst{27-24} = 0b1100;
2177 let Inst{23-21} = 0b001;
2178 let Inst{20-16} = Rs;
2179 let Inst{13-8} = s6;
2180 let Inst{7-5} = 0b000;
2184 let hasSideEffects = 0, hasNewValue = 1, isCodeGenOnly = 0 in
2185 def S4_clbpaddi : SInst<(outs IntRegs:$Rd), (ins DoubleRegs:$Rs, s6Imm:$s6),
2186 "$Rd = add(clb($Rs), #$s6)", [], "", S_2op_tc_2_SLOT23> {
2190 let IClass = 0b1000;
2191 let Inst{27-24} = 0b1000;
2192 let Inst{23-21} = 0b011;
2193 let Inst{20-16} = Rs;
2194 let Inst{13-8} = s6;
2195 let Inst{7-5} = 0b010;
2200 // Bit test/set/clear
2201 let isCodeGenOnly = 0 in {
2202 def S4_ntstbit_i : T_TEST_BIT_IMM<"!tstbit", 0b001>;
2203 def S4_ntstbit_r : T_TEST_BIT_REG<"!tstbit", 1>;
2206 let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm.
2207 def: Pat<(i1 (seteq (and (shl 1, u5ImmPred:$u5), (i32 IntRegs:$Rs)), 0)),
2208 (S4_ntstbit_i (i32 IntRegs:$Rs), u5ImmPred:$u5)>;
2209 def: Pat<(i1 (seteq (and (shl 1, (i32 IntRegs:$Rt)), (i32 IntRegs:$Rs)), 0)),
2210 (S4_ntstbit_r (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))>;
2213 // Add extra complexity to prefer these instructions over bitsset/bitsclr.
2214 // The reason is that tstbit/ntstbit can be folded into a compound instruction:
2215 // if ([!]tstbit(...)) jump ...
2216 let AddedComplexity = 100 in
2217 def: Pat<(i1 (setne (and (i32 IntRegs:$Rs), (i32 Set5ImmPred:$u5)), (i32 0))),
2218 (S2_tstbit_i (i32 IntRegs:$Rs), (BITPOS32 Set5ImmPred:$u5))>;
2220 let AddedComplexity = 100 in
2221 def: Pat<(i1 (seteq (and (i32 IntRegs:$Rs), (i32 Set5ImmPred:$u5)), (i32 0))),
2222 (S4_ntstbit_i (i32 IntRegs:$Rs), (BITPOS32 Set5ImmPred:$u5))>;
2224 let isCodeGenOnly = 0 in {
2225 def C4_nbitsset : T_TEST_BITS_REG<"!bitsset", 0b01, 1>;
2226 def C4_nbitsclr : T_TEST_BITS_REG<"!bitsclr", 0b10, 1>;
2227 def C4_nbitsclri : T_TEST_BITS_IMM<"!bitsclr", 0b10, 1>;
2230 // Do not increase complexity of these patterns. In the DAG, "cmp i8" may be
2231 // represented as a compare against "value & 0xFF", which is an exact match
2232 // for cmpb (same for cmph). The patterns below do not contain any additional
2233 // complexity that would make them preferable, and if they were actually used
2234 // instead of cmpb/cmph, they would result in a compare against register that
2235 // is loaded with the byte/half mask (i.e. 0xFF or 0xFFFF).
2236 def: Pat<(i1 (setne (and I32:$Rs, u6ImmPred:$u6), 0)),
2237 (C4_nbitsclri I32:$Rs, u6ImmPred:$u6)>;
2238 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), 0)),
2239 (C4_nbitsclr I32:$Rs, I32:$Rt)>;
2240 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), I32:$Rt)),
2241 (C4_nbitsset I32:$Rs, I32:$Rt)>;
2243 //===----------------------------------------------------------------------===//
2245 //===----------------------------------------------------------------------===//
2247 //===----------------------------------------------------------------------===//
2249 //===----------------------------------------------------------------------===//
2251 // Rd=add(#u6,mpyi(Rs,#U6)) -- Multiply by immed and add immed.
2253 let hasNewValue = 1, isExtendable = 1, opExtentBits = 6, opExtendable = 1,
2254 isCodeGenOnly = 0 in
2255 def M4_mpyri_addi : MInst<(outs IntRegs:$Rd),
2256 (ins u6Ext:$u6, IntRegs:$Rs, u6Imm:$U6),
2257 "$Rd = add(#$u6, mpyi($Rs, #$U6))" ,
2258 [(set (i32 IntRegs:$Rd),
2259 (add (mul (i32 IntRegs:$Rs), u6ImmPred:$U6),
2260 u6ExtPred:$u6))] ,"",ALU64_tc_3x_SLOT23> {
2266 let IClass = 0b1101;
2268 let Inst{27-24} = 0b1000;
2269 let Inst{23} = U6{5};
2270 let Inst{22-21} = u6{5-4};
2271 let Inst{20-16} = Rs;
2272 let Inst{13} = u6{3};
2273 let Inst{12-8} = Rd;
2274 let Inst{7-5} = u6{2-0};
2275 let Inst{4-0} = U6{4-0};
2278 // Rd=add(#u6,mpyi(Rs,Rt))
2279 let CextOpcode = "ADD_MPY", InputType = "imm", hasNewValue = 1,
2280 isExtendable = 1, opExtentBits = 6, opExtendable = 1, isCodeGenOnly = 0 in
2281 def M4_mpyrr_addi : MInst <(outs IntRegs:$Rd),
2282 (ins u6Ext:$u6, IntRegs:$Rs, IntRegs:$Rt),
2283 "$Rd = add(#$u6, mpyi($Rs, $Rt))" ,
2284 [(set (i32 IntRegs:$Rd),
2285 (add (mul (i32 IntRegs:$Rs), (i32 IntRegs:$Rt)), u6ExtPred:$u6))],
2286 "", ALU64_tc_3x_SLOT23>, ImmRegRel {
2292 let IClass = 0b1101;
2294 let Inst{27-23} = 0b01110;
2295 let Inst{22-21} = u6{5-4};
2296 let Inst{20-16} = Rs;
2297 let Inst{13} = u6{3};
2298 let Inst{12-8} = Rt;
2299 let Inst{7-5} = u6{2-0};
2303 let hasNewValue = 1 in
2304 class T_AddMpy <bit MajOp, PatLeaf ImmPred, dag ins>
2305 : ALU64Inst <(outs IntRegs:$dst), ins,
2306 "$dst = add($src1, mpyi("#!if(MajOp,"$src3, #$src2))",
2308 [(set (i32 IntRegs:$dst),
2309 (add (i32 IntRegs:$src1), (mul (i32 IntRegs:$src3), ImmPred:$src2)))],
2310 "", ALU64_tc_3x_SLOT23> {
2316 let IClass = 0b1101;
2318 bits<6> ImmValue = !if(MajOp, src2{5-0}, src2{7-2});
2320 let Inst{27-24} = 0b1111;
2321 let Inst{23} = MajOp;
2322 let Inst{22-21} = ImmValue{5-4};
2323 let Inst{20-16} = src3;
2324 let Inst{13} = ImmValue{3};
2325 let Inst{12-8} = dst;
2326 let Inst{7-5} = ImmValue{2-0};
2327 let Inst{4-0} = src1;
2330 let isCodeGenOnly = 0 in
2331 def M4_mpyri_addr_u2 : T_AddMpy<0b0, u6_2ImmPred,
2332 (ins IntRegs:$src1, u6_2Imm:$src2, IntRegs:$src3)>;
2334 let isExtendable = 1, opExtentBits = 6, opExtendable = 3,
2335 CextOpcode = "ADD_MPY", InputType = "imm", isCodeGenOnly = 0 in
2336 def M4_mpyri_addr : T_AddMpy<0b1, u6ExtPred,
2337 (ins IntRegs:$src1, IntRegs:$src3, u6Ext:$src2)>, ImmRegRel;
2339 // Rx=add(Ru,mpyi(Rx,Rs))
2340 let validSubTargets = HasV4SubT, CextOpcode = "ADD_MPY", InputType = "reg",
2341 hasNewValue = 1, isCodeGenOnly = 0 in
2342 def M4_mpyrr_addr: MInst_acc <(outs IntRegs:$Rx),
2343 (ins IntRegs:$Ru, IntRegs:$_src_, IntRegs:$Rs),
2344 "$Rx = add($Ru, mpyi($_src_, $Rs))",
2345 [(set (i32 IntRegs:$Rx), (add (i32 IntRegs:$Ru),
2346 (mul (i32 IntRegs:$_src_), (i32 IntRegs:$Rs))))],
2347 "$_src_ = $Rx", M_tc_3x_SLOT23>, ImmRegRel {
2352 let IClass = 0b1110;
2354 let Inst{27-21} = 0b0011000;
2355 let Inst{12-8} = Rx;
2357 let Inst{20-16} = Rs;
2360 // Rd=add(##,mpyi(Rs,#U6))
2361 def : Pat <(add (mul (i32 IntRegs:$src2), u6ImmPred:$src3),
2362 (HexagonCONST32 tglobaladdr:$src1)),
2363 (i32 (M4_mpyri_addi tglobaladdr:$src1, IntRegs:$src2,
2366 // Rd=add(##,mpyi(Rs,Rt))
2367 def : Pat <(add (mul (i32 IntRegs:$src2), (i32 IntRegs:$src3)),
2368 (HexagonCONST32 tglobaladdr:$src1)),
2369 (i32 (M4_mpyrr_addi tglobaladdr:$src1, IntRegs:$src2,
2372 // Vector reduce multiply word by signed half (32x16)
2373 //Rdd=vrmpyweh(Rss,Rtt)[:<<1]
2374 let isCodeGenOnly = 0 in {
2375 def M4_vrmpyeh_s0 : T_M2_vmpy<"vrmpyweh", 0b010, 0b100, 0, 0, 0>;
2376 def M4_vrmpyeh_s1 : T_M2_vmpy<"vrmpyweh", 0b110, 0b100, 1, 0, 0>;
2379 //Rdd=vrmpywoh(Rss,Rtt)[:<<1]
2380 let isCodeGenOnly = 0 in {
2381 def M4_vrmpyoh_s0 : T_M2_vmpy<"vrmpywoh", 0b001, 0b010, 0, 0, 0>;
2382 def M4_vrmpyoh_s1 : T_M2_vmpy<"vrmpywoh", 0b101, 0b010, 1, 0, 0>;
2384 //Rdd+=vrmpyweh(Rss,Rtt)[:<<1]
2385 let isCodeGenOnly = 0 in {
2386 def M4_vrmpyeh_acc_s0: T_M2_vmpy_acc<"vrmpyweh", 0b001, 0b110, 0, 0>;
2387 def M4_vrmpyeh_acc_s1: T_M2_vmpy_acc<"vrmpyweh", 0b101, 0b110, 1, 0>;
2390 //Rdd=vrmpywoh(Rss,Rtt)[:<<1]
2391 let isCodeGenOnly = 0 in {
2392 def M4_vrmpyoh_acc_s0: T_M2_vmpy_acc<"vrmpywoh", 0b011, 0b110, 0, 0>;
2393 def M4_vrmpyoh_acc_s1: T_M2_vmpy_acc<"vrmpywoh", 0b111, 0b110, 1, 0>;
2396 // Vector multiply halfwords, signed by unsigned
2397 // Rdd=vmpyhsu(Rs,Rt)[:<<]:sat
2398 let isCodeGenOnly = 0 in {
2399 def M2_vmpy2su_s0 : T_XTYPE_mpy64 < "vmpyhsu", 0b000, 0b111, 1, 0, 0>;
2400 def M2_vmpy2su_s1 : T_XTYPE_mpy64 < "vmpyhsu", 0b100, 0b111, 1, 1, 0>;
2403 // Rxx+=vmpyhsu(Rs,Rt)[:<<1]:sat
2404 let isCodeGenOnly = 0 in {
2405 def M2_vmac2su_s0 : T_XTYPE_mpy64_acc < "vmpyhsu", "+", 0b011, 0b101, 1, 0, 0>;
2406 def M2_vmac2su_s1 : T_XTYPE_mpy64_acc < "vmpyhsu", "+", 0b111, 0b101, 1, 1, 0>;
2409 // Vector polynomial multiply halfwords
2410 // Rdd=vpmpyh(Rs,Rt)
2411 let isCodeGenOnly = 0 in
2412 def M4_vpmpyh : T_XTYPE_mpy64 < "vpmpyh", 0b110, 0b111, 0, 0, 0>;
2414 // Rxx^=vpmpyh(Rs,Rt)
2415 let isCodeGenOnly = 0 in
2416 def M4_vpmpyh_acc : T_XTYPE_mpy64_acc < "vpmpyh", "^", 0b101, 0b111, 0, 0, 0>;
2418 // Polynomial multiply words
2420 let isCodeGenOnly = 0 in
2421 def M4_pmpyw : T_XTYPE_mpy64 < "pmpyw", 0b010, 0b111, 0, 0, 0>;
2423 // Rxx^=pmpyw(Rs,Rt)
2424 let isCodeGenOnly = 0 in
2425 def M4_pmpyw_acc : T_XTYPE_mpy64_acc < "pmpyw", "^", 0b001, 0b111, 0, 0, 0>;
2427 //===----------------------------------------------------------------------===//
2429 //===----------------------------------------------------------------------===//
2432 //===----------------------------------------------------------------------===//
2433 // ALU64/Vector compare
2434 //===----------------------------------------------------------------------===//
2435 //===----------------------------------------------------------------------===//
2436 // Template class for vector compare
2437 //===----------------------------------------------------------------------===//
2439 let hasSideEffects = 0 in
2440 class T_vcmpImm <string Str, bits<2> cmpOp, bits<2> minOp, Operand ImmOprnd>
2441 : ALU64_rr <(outs PredRegs:$Pd),
2442 (ins DoubleRegs:$Rss, ImmOprnd:$Imm),
2443 "$Pd = "#Str#"($Rss, #$Imm)",
2444 [], "", ALU64_tc_2early_SLOT23> {
2449 let ImmBits{6-0} = Imm{6-0};
2450 let ImmBits{7} = !if (!eq(cmpOp,0b10), 0b0, Imm{7}); // 0 for vcmp[bhw].gtu
2452 let IClass = 0b1101;
2454 let Inst{27-24} = 0b1100;
2455 let Inst{22-21} = cmpOp;
2456 let Inst{20-16} = Rss;
2457 let Inst{12-5} = ImmBits;
2458 let Inst{4-3} = minOp;
2462 // Vector compare bytes
2463 let isCodeGenOnly = 0 in
2464 def A4_vcmpbgt : T_vcmp <"vcmpb.gt", 0b1010>;
2465 def: T_vcmp_pat<A4_vcmpbgt, setgt, v8i8>;
2467 let AsmString = "$Pd = any8(vcmpb.eq($Rss, $Rtt))" in
2468 let isCodeGenOnly = 0 in
2469 def A4_vcmpbeq_any : T_vcmp <"any8(vcmpb.gt", 0b1000>;
2471 let isCodeGenOnly = 0 in {
2472 def A4_vcmpbeqi : T_vcmpImm <"vcmpb.eq", 0b00, 0b00, u8Imm>;
2473 def A4_vcmpbgti : T_vcmpImm <"vcmpb.gt", 0b01, 0b00, s8Imm>;
2474 def A4_vcmpbgtui : T_vcmpImm <"vcmpb.gtu", 0b10, 0b00, u7Imm>;
2477 // Vector compare halfwords
2478 let isCodeGenOnly = 0 in {
2479 def A4_vcmpheqi : T_vcmpImm <"vcmph.eq", 0b00, 0b01, s8Imm>;
2480 def A4_vcmphgti : T_vcmpImm <"vcmph.gt", 0b01, 0b01, s8Imm>;
2481 def A4_vcmphgtui : T_vcmpImm <"vcmph.gtu", 0b10, 0b01, u7Imm>;
2484 // Vector compare words
2485 let isCodeGenOnly = 0 in {
2486 def A4_vcmpweqi : T_vcmpImm <"vcmpw.eq", 0b00, 0b10, s8Imm>;
2487 def A4_vcmpwgti : T_vcmpImm <"vcmpw.gt", 0b01, 0b10, s8Imm>;
2488 def A4_vcmpwgtui : T_vcmpImm <"vcmpw.gtu", 0b10, 0b10, u7Imm>;
2491 //===----------------------------------------------------------------------===//
2493 //===----------------------------------------------------------------------===//
2494 // Shift by immediate and accumulate/logical.
2495 // Rx=add(#u8,asl(Rx,#U5)) Rx=add(#u8,lsr(Rx,#U5))
2496 // Rx=sub(#u8,asl(Rx,#U5)) Rx=sub(#u8,lsr(Rx,#U5))
2497 // Rx=and(#u8,asl(Rx,#U5)) Rx=and(#u8,lsr(Rx,#U5))
2498 // Rx=or(#u8,asl(Rx,#U5)) Rx=or(#u8,lsr(Rx,#U5))
2499 let isExtendable = 1, opExtendable = 1, isExtentSigned = 0, opExtentBits = 8,
2500 hasNewValue = 1, opNewValue = 0, validSubTargets = HasV4SubT in
2501 class T_S4_ShiftOperate<string MnOp, string MnSh, SDNode Op, SDNode Sh,
2502 bit asl_lsr, bits<2> MajOp, InstrItinClass Itin>
2503 : MInst_acc<(outs IntRegs:$Rd), (ins u8Ext:$u8, IntRegs:$Rx, u5Imm:$U5),
2504 "$Rd = "#MnOp#"(#$u8, "#MnSh#"($Rx, #$U5))",
2505 [(set (i32 IntRegs:$Rd),
2506 (Op (Sh I32:$Rx, u5ImmPred:$U5), u8ExtPred:$u8))],
2507 "$Rd = $Rx", Itin> {
2514 let IClass = 0b1101;
2515 let Inst{27-24} = 0b1110;
2516 let Inst{23-21} = u8{7-5};
2517 let Inst{20-16} = Rd;
2518 let Inst{13} = u8{4};
2519 let Inst{12-8} = U5;
2520 let Inst{7-5} = u8{3-1};
2521 let Inst{4} = asl_lsr;
2522 let Inst{3} = u8{0};
2523 let Inst{2-1} = MajOp;
2526 multiclass T_ShiftOperate<string mnemonic, SDNode Op, bits<2> MajOp,
2527 InstrItinClass Itin> {
2528 def _asl_ri : T_S4_ShiftOperate<mnemonic, "asl", Op, shl, 0, MajOp, Itin>;
2529 def _lsr_ri : T_S4_ShiftOperate<mnemonic, "lsr", Op, srl, 1, MajOp, Itin>;
2532 let AddedComplexity = 200, isCodeGenOnly = 0 in {
2533 defm S4_addi : T_ShiftOperate<"add", add, 0b10, ALU64_tc_2_SLOT23>;
2534 defm S4_andi : T_ShiftOperate<"and", and, 0b00, ALU64_tc_2_SLOT23>;
2537 let AddedComplexity = 30, isCodeGenOnly = 0 in
2538 defm S4_ori : T_ShiftOperate<"or", or, 0b01, ALU64_tc_1_SLOT23>;
2540 let isCodeGenOnly = 0 in
2541 defm S4_subi : T_ShiftOperate<"sub", sub, 0b11, ALU64_tc_1_SLOT23>;
2543 // Vector conditional negate
2544 // Rdd=vcnegh(Rss,Rt)
2545 let Defs = [USR_OVF], Itinerary = S_3op_tc_2_SLOT23, isCodeGenOnly = 0 in
2546 def S2_vcnegh : T_S3op_shiftVect < "vcnegh", 0b11, 0b01>;
2548 // Rd=[cround|round](Rs,Rt)
2549 let hasNewValue = 1, Itinerary = S_3op_tc_2_SLOT23, isCodeGenOnly = 0 in {
2550 def A4_cround_rr : T_S3op_3 < "cround", IntRegs, 0b11, 0b00>;
2551 def A4_round_rr : T_S3op_3 < "round", IntRegs, 0b11, 0b10>;
2554 // Rd=round(Rs,Rt):sat
2555 let hasNewValue = 1, Defs = [USR_OVF], Itinerary = S_3op_tc_2_SLOT23,
2556 isCodeGenOnly = 0 in
2557 def A4_round_rr_sat : T_S3op_3 < "round", IntRegs, 0b11, 0b11, 1>;
2559 // Rd=[cmpyiwh|cmpyrwh](Rss,Rt):<<1:rnd:sat
2560 let Defs = [USR_OVF], Itinerary = S_3op_tc_3x_SLOT23, isCodeGenOnly = 0 in {
2561 def M4_cmpyi_wh : T_S3op_8<"cmpyiwh", 0b100, 1, 1, 1>;
2562 def M4_cmpyr_wh : T_S3op_8<"cmpyrwh", 0b110, 1, 1, 1>;
2565 // Rdd=[add|sub](Rss,Rtt,Px):carry
2566 let isPredicateLate = 1, hasSideEffects = 0 in
2567 class T_S3op_carry <string mnemonic, bits<3> MajOp>
2568 : SInst < (outs DoubleRegs:$Rdd, PredRegs:$Px),
2569 (ins DoubleRegs:$Rss, DoubleRegs:$Rtt, PredRegs:$Pu),
2570 "$Rdd = "#mnemonic#"($Rss, $Rtt, $Pu):carry",
2571 [], "$Px = $Pu", S_3op_tc_1_SLOT23 > {
2577 let IClass = 0b1100;
2579 let Inst{27-24} = 0b0010;
2580 let Inst{23-21} = MajOp;
2581 let Inst{20-16} = Rss;
2582 let Inst{12-8} = Rtt;
2584 let Inst{4-0} = Rdd;
2587 let isCodeGenOnly = 0 in {
2588 def A4_addp_c : T_S3op_carry < "add", 0b110 >;
2589 def A4_subp_c : T_S3op_carry < "sub", 0b111 >;
2592 let Itinerary = S_3op_tc_3_SLOT23, hasSideEffects = 0 in
2593 class T_S3op_6 <string mnemonic, bits<3> MinOp, bit isUnsigned>
2594 : SInst <(outs DoubleRegs:$Rxx),
2595 (ins DoubleRegs:$dst2, DoubleRegs:$Rss, IntRegs:$Ru),
2596 "$Rxx = "#mnemonic#"($Rss, $Ru)" ,
2597 [] , "$dst2 = $Rxx"> {
2602 let IClass = 0b1100;
2604 let Inst{27-21} = 0b1011001;
2605 let Inst{20-16} = Rss;
2606 let Inst{13} = isUnsigned;
2607 let Inst{12-8} = Rxx;
2608 let Inst{7-5} = MinOp;
2612 // Vector reduce maximum halfwords
2613 // Rxx=vrmax[u]h(Rss,Ru)
2614 let isCodeGenOnly = 0 in {
2615 def A4_vrmaxh : T_S3op_6 < "vrmaxh", 0b001, 0>;
2616 def A4_vrmaxuh : T_S3op_6 < "vrmaxuh", 0b001, 1>;
2618 // Vector reduce maximum words
2619 // Rxx=vrmax[u]w(Rss,Ru)
2620 let isCodeGenOnly = 0 in {
2621 def A4_vrmaxw : T_S3op_6 < "vrmaxw", 0b010, 0>;
2622 def A4_vrmaxuw : T_S3op_6 < "vrmaxuw", 0b010, 1>;
2624 // Vector reduce minimum halfwords
2625 // Rxx=vrmin[u]h(Rss,Ru)
2626 let isCodeGenOnly = 0 in {
2627 def A4_vrminh : T_S3op_6 < "vrminh", 0b101, 0>;
2628 def A4_vrminuh : T_S3op_6 < "vrminuh", 0b101, 1>;
2631 // Vector reduce minimum words
2632 // Rxx=vrmin[u]w(Rss,Ru)
2633 let isCodeGenOnly = 0 in {
2634 def A4_vrminw : T_S3op_6 < "vrminw", 0b110, 0>;
2635 def A4_vrminuw : T_S3op_6 < "vrminuw", 0b110, 1>;
2638 // Shift an immediate left by register amount.
2639 let hasNewValue = 1, hasSideEffects = 0, isCodeGenOnly = 0 in
2640 def S4_lsli: SInst <(outs IntRegs:$Rd), (ins s6Imm:$s6, IntRegs:$Rt),
2641 "$Rd = lsl(#$s6, $Rt)" ,
2642 [(set (i32 IntRegs:$Rd), (shl s6ImmPred:$s6,
2643 (i32 IntRegs:$Rt)))],
2644 "", S_3op_tc_1_SLOT23> {
2649 let IClass = 0b1100;
2651 let Inst{27-22} = 0b011010;
2652 let Inst{20-16} = s6{5-1};
2653 let Inst{12-8} = Rt;
2654 let Inst{7-6} = 0b11;
2656 let Inst{5} = s6{0};
2659 //===----------------------------------------------------------------------===//
2661 //===----------------------------------------------------------------------===//
2663 //===----------------------------------------------------------------------===//
2664 // MEMOP: Word, Half, Byte
2665 //===----------------------------------------------------------------------===//
2667 def MEMOPIMM : SDNodeXForm<imm, [{
2668 // Call the transformation function XformM5ToU5Imm to get the negative
2669 // immediate's positive counterpart.
2670 int32_t imm = N->getSExtValue();
2671 return XformM5ToU5Imm(imm);
2674 def MEMOPIMM_HALF : SDNodeXForm<imm, [{
2675 // -1 .. -31 represented as 65535..65515
2676 // assigning to a short restores our desired signed value.
2677 // Call the transformation function XformM5ToU5Imm to get the negative
2678 // immediate's positive counterpart.
2679 int16_t imm = N->getSExtValue();
2680 return XformM5ToU5Imm(imm);
2683 def MEMOPIMM_BYTE : SDNodeXForm<imm, [{
2684 // -1 .. -31 represented as 255..235
2685 // assigning to a char restores our desired signed value.
2686 // Call the transformation function XformM5ToU5Imm to get the negative
2687 // immediate's positive counterpart.
2688 int8_t imm = N->getSExtValue();
2689 return XformM5ToU5Imm(imm);
2692 def SETMEMIMM : SDNodeXForm<imm, [{
2693 // Return the bit position we will set [0-31].
2695 int32_t imm = N->getSExtValue();
2696 return XformMskToBitPosU5Imm(imm);
2699 def CLRMEMIMM : SDNodeXForm<imm, [{
2700 // Return the bit position we will clear [0-31].
2702 // we bit negate the value first
2703 int32_t imm = ~(N->getSExtValue());
2704 return XformMskToBitPosU5Imm(imm);
2707 def SETMEMIMM_SHORT : SDNodeXForm<imm, [{
2708 // Return the bit position we will set [0-15].
2710 int16_t imm = N->getSExtValue();
2711 return XformMskToBitPosU4Imm(imm);
2714 def CLRMEMIMM_SHORT : SDNodeXForm<imm, [{
2715 // Return the bit position we will clear [0-15].
2717 // we bit negate the value first
2718 int16_t imm = ~(N->getSExtValue());
2719 return XformMskToBitPosU4Imm(imm);
2722 def SETMEMIMM_BYTE : SDNodeXForm<imm, [{
2723 // Return the bit position we will set [0-7].
2725 int8_t imm = N->getSExtValue();
2726 return XformMskToBitPosU3Imm(imm);
2729 def CLRMEMIMM_BYTE : SDNodeXForm<imm, [{
2730 // Return the bit position we will clear [0-7].
2732 // we bit negate the value first
2733 int8_t imm = ~(N->getSExtValue());
2734 return XformMskToBitPosU3Imm(imm);
2737 //===----------------------------------------------------------------------===//
2738 // Template class for MemOp instructions with the register value.
2739 //===----------------------------------------------------------------------===//
2740 class MemOp_rr_base <string opc, bits<2> opcBits, Operand ImmOp,
2741 string memOp, bits<2> memOpBits> :
2743 (ins IntRegs:$base, ImmOp:$offset, IntRegs:$delta),
2744 opc#"($base+#$offset)"#memOp#"$delta",
2746 Requires<[UseMEMOP]> {
2751 bits<6> offsetBits; // memb - u6:0 , memh - u6:1, memw - u6:2
2753 let offsetBits = !if (!eq(opcBits, 0b00), offset{5-0},
2754 !if (!eq(opcBits, 0b01), offset{6-1},
2755 !if (!eq(opcBits, 0b10), offset{7-2},0)));
2757 let opExtentAlign = opcBits;
2758 let IClass = 0b0011;
2759 let Inst{27-24} = 0b1110;
2760 let Inst{22-21} = opcBits;
2761 let Inst{20-16} = base;
2763 let Inst{12-7} = offsetBits;
2764 let Inst{6-5} = memOpBits;
2765 let Inst{4-0} = delta;
2768 //===----------------------------------------------------------------------===//
2769 // Template class for MemOp instructions with the immediate value.
2770 //===----------------------------------------------------------------------===//
2771 class MemOp_ri_base <string opc, bits<2> opcBits, Operand ImmOp,
2772 string memOp, bits<2> memOpBits> :
2774 (ins IntRegs:$base, ImmOp:$offset, u5Imm:$delta),
2775 opc#"($base+#$offset)"#memOp#"#$delta"
2776 #!if(memOpBits{1},")", ""), // clrbit, setbit - include ')'
2778 Requires<[UseMEMOP]> {
2783 bits<6> offsetBits; // memb - u6:0 , memh - u6:1, memw - u6:2
2785 let offsetBits = !if (!eq(opcBits, 0b00), offset{5-0},
2786 !if (!eq(opcBits, 0b01), offset{6-1},
2787 !if (!eq(opcBits, 0b10), offset{7-2},0)));
2789 let opExtentAlign = opcBits;
2790 let IClass = 0b0011;
2791 let Inst{27-24} = 0b1111;
2792 let Inst{22-21} = opcBits;
2793 let Inst{20-16} = base;
2795 let Inst{12-7} = offsetBits;
2796 let Inst{6-5} = memOpBits;
2797 let Inst{4-0} = delta;
2800 // multiclass to define MemOp instructions with register operand.
2801 multiclass MemOp_rr<string opc, bits<2> opcBits, Operand ImmOp> {
2802 def L4_add#NAME : MemOp_rr_base <opc, opcBits, ImmOp, " += ", 0b00>; // add
2803 def L4_sub#NAME : MemOp_rr_base <opc, opcBits, ImmOp, " -= ", 0b01>; // sub
2804 def L4_and#NAME : MemOp_rr_base <opc, opcBits, ImmOp, " &= ", 0b10>; // and
2805 def L4_or#NAME : MemOp_rr_base <opc, opcBits, ImmOp, " |= ", 0b11>; // or
2808 // multiclass to define MemOp instructions with immediate Operand.
2809 multiclass MemOp_ri<string opc, bits<2> opcBits, Operand ImmOp> {
2810 def L4_iadd#NAME : MemOp_ri_base <opc, opcBits, ImmOp, " += ", 0b00 >;
2811 def L4_isub#NAME : MemOp_ri_base <opc, opcBits, ImmOp, " -= ", 0b01 >;
2812 def L4_iand#NAME : MemOp_ri_base<opc, opcBits, ImmOp, " = clrbit(", 0b10>;
2813 def L4_ior#NAME : MemOp_ri_base<opc, opcBits, ImmOp, " = setbit(", 0b11>;
2816 multiclass MemOp_base <string opc, bits<2> opcBits, Operand ImmOp> {
2817 defm _#NAME : MemOp_rr <opc, opcBits, ImmOp>;
2818 defm _#NAME : MemOp_ri <opc, opcBits, ImmOp>;
2821 // Define MemOp instructions.
2822 let isExtendable = 1, opExtendable = 1, isExtentSigned = 0,
2823 validSubTargets =HasV4SubT in {
2824 let opExtentBits = 6, accessSize = ByteAccess, isCodeGenOnly = 0 in
2825 defm memopb_io : MemOp_base <"memb", 0b00, u6_0Ext>;
2827 let opExtentBits = 7, accessSize = HalfWordAccess, isCodeGenOnly = 0 in
2828 defm memoph_io : MemOp_base <"memh", 0b01, u6_1Ext>;
2830 let opExtentBits = 8, accessSize = WordAccess, isCodeGenOnly = 0 in
2831 defm memopw_io : MemOp_base <"memw", 0b10, u6_2Ext>;
2834 //===----------------------------------------------------------------------===//
2835 // Multiclass to define 'Def Pats' for ALU operations on the memory
2836 // Here value used for the ALU operation is an immediate value.
2837 // mem[bh](Rs+#0) += #U5
2838 // mem[bh](Rs+#u6) += #U5
2839 //===----------------------------------------------------------------------===//
2841 multiclass MemOpi_u5Pats <PatFrag ldOp, PatFrag stOp, PatLeaf ExtPred,
2842 InstHexagon MI, SDNode OpNode> {
2843 let AddedComplexity = 180 in
2844 def : Pat < (stOp (OpNode (ldOp IntRegs:$addr), u5ImmPred:$addend),
2846 (MI IntRegs:$addr, #0, u5ImmPred:$addend )>;
2848 let AddedComplexity = 190 in
2849 def : Pat <(stOp (OpNode (ldOp (add IntRegs:$base, ExtPred:$offset)),
2851 (add IntRegs:$base, ExtPred:$offset)),
2852 (MI IntRegs:$base, ExtPred:$offset, u5ImmPred:$addend)>;
2855 multiclass MemOpi_u5ALUOp<PatFrag ldOp, PatFrag stOp, PatLeaf ExtPred,
2856 InstHexagon addMI, InstHexagon subMI> {
2857 defm : MemOpi_u5Pats<ldOp, stOp, ExtPred, addMI, add>;
2858 defm : MemOpi_u5Pats<ldOp, stOp, ExtPred, subMI, sub>;
2861 multiclass MemOpi_u5ExtType<PatFrag ldOpByte, PatFrag ldOpHalf > {
2863 defm : MemOpi_u5ALUOp <ldOpHalf, truncstorei16, u6_1ExtPred,
2864 L4_iadd_memoph_io, L4_isub_memoph_io>;
2866 defm : MemOpi_u5ALUOp <ldOpByte, truncstorei8, u6ExtPred,
2867 L4_iadd_memopb_io, L4_isub_memopb_io>;
2870 let Predicates = [HasV4T, UseMEMOP] in {
2871 defm : MemOpi_u5ExtType<zextloadi8, zextloadi16>; // zero extend
2872 defm : MemOpi_u5ExtType<sextloadi8, sextloadi16>; // sign extend
2873 defm : MemOpi_u5ExtType<extloadi8, extloadi16>; // any extend
2876 defm : MemOpi_u5ALUOp <load, store, u6_2ExtPred, L4_iadd_memopw_io,
2880 //===----------------------------------------------------------------------===//
2881 // multiclass to define 'Def Pats' for ALU operations on the memory.
2882 // Here value used for the ALU operation is a negative value.
2883 // mem[bh](Rs+#0) += #m5
2884 // mem[bh](Rs+#u6) += #m5
2885 //===----------------------------------------------------------------------===//
2887 multiclass MemOpi_m5Pats <PatFrag ldOp, PatFrag stOp, PatLeaf extPred,
2888 PatLeaf immPred, ComplexPattern addrPred,
2889 SDNodeXForm xformFunc, InstHexagon MI> {
2890 let AddedComplexity = 190 in
2891 def : Pat <(stOp (add (ldOp IntRegs:$addr), immPred:$subend),
2893 (MI IntRegs:$addr, #0, (xformFunc immPred:$subend) )>;
2895 let AddedComplexity = 195 in
2896 def : Pat<(stOp (add (ldOp (add IntRegs:$base, extPred:$offset)),
2898 (add IntRegs:$base, extPred:$offset)),
2899 (MI IntRegs:$base, extPred:$offset, (xformFunc immPred:$subend))>;
2902 multiclass MemOpi_m5ExtType<PatFrag ldOpByte, PatFrag ldOpHalf > {
2904 defm : MemOpi_m5Pats <ldOpHalf, truncstorei16, u6_1ExtPred, m5HImmPred,
2905 ADDRriU6_1, MEMOPIMM_HALF, L4_isub_memoph_io>;
2907 defm : MemOpi_m5Pats <ldOpByte, truncstorei8, u6ExtPred, m5BImmPred,
2908 ADDRriU6_0, MEMOPIMM_BYTE, L4_isub_memopb_io>;
2911 let Predicates = [HasV4T, UseMEMOP] in {
2912 defm : MemOpi_m5ExtType<zextloadi8, zextloadi16>; // zero extend
2913 defm : MemOpi_m5ExtType<sextloadi8, sextloadi16>; // sign extend
2914 defm : MemOpi_m5ExtType<extloadi8, extloadi16>; // any extend
2917 defm : MemOpi_m5Pats <load, store, u6_2ExtPred, m5ImmPred,
2918 ADDRriU6_2, MEMOPIMM, L4_isub_memopw_io>;
2921 //===----------------------------------------------------------------------===//
2922 // Multiclass to define 'def Pats' for bit operations on the memory.
2923 // mem[bhw](Rs+#0) = [clrbit|setbit](#U5)
2924 // mem[bhw](Rs+#u6) = [clrbit|setbit](#U5)
2925 //===----------------------------------------------------------------------===//
2927 multiclass MemOpi_bitPats <PatFrag ldOp, PatFrag stOp, PatLeaf immPred,
2928 PatLeaf extPred, ComplexPattern addrPred,
2929 SDNodeXForm xformFunc, InstHexagon MI, SDNode OpNode> {
2931 // mem[bhw](Rs+#u6:[012]) = [clrbit|setbit](#U5)
2932 let AddedComplexity = 250 in
2933 def : Pat<(stOp (OpNode (ldOp (add IntRegs:$base, extPred:$offset)),
2935 (add IntRegs:$base, extPred:$offset)),
2936 (MI IntRegs:$base, extPred:$offset, (xformFunc immPred:$bitend))>;
2938 // mem[bhw](Rs+#0) = [clrbit|setbit](#U5)
2939 let AddedComplexity = 225 in
2940 def : Pat <(stOp (OpNode (ldOp (addrPred IntRegs:$addr, extPred:$offset)),
2942 (addrPred (i32 IntRegs:$addr), extPred:$offset)),
2943 (MI IntRegs:$addr, extPred:$offset, (xformFunc immPred:$bitend))>;
2946 multiclass MemOpi_bitExtType<PatFrag ldOpByte, PatFrag ldOpHalf > {
2948 defm : MemOpi_bitPats<ldOpByte, truncstorei8, Clr3ImmPred, u6ExtPred,
2949 ADDRriU6_0, CLRMEMIMM_BYTE, L4_iand_memopb_io, and>;
2951 defm : MemOpi_bitPats<ldOpByte, truncstorei8, Set3ImmPred, u6ExtPred,
2952 ADDRriU6_0, SETMEMIMM_BYTE, L4_ior_memopb_io, or>;
2953 // Half Word - clrbit
2954 defm : MemOpi_bitPats<ldOpHalf, truncstorei16, Clr4ImmPred, u6_1ExtPred,
2955 ADDRriU6_1, CLRMEMIMM_SHORT, L4_iand_memoph_io, and>;
2956 // Half Word - setbit
2957 defm : MemOpi_bitPats<ldOpHalf, truncstorei16, Set4ImmPred, u6_1ExtPred,
2958 ADDRriU6_1, SETMEMIMM_SHORT, L4_ior_memoph_io, or>;
2961 let Predicates = [HasV4T, UseMEMOP] in {
2962 // mem[bh](Rs+#0) = [clrbit|setbit](#U5)
2963 // mem[bh](Rs+#u6:[01]) = [clrbit|setbit](#U5)
2964 defm : MemOpi_bitExtType<zextloadi8, zextloadi16>; // zero extend
2965 defm : MemOpi_bitExtType<sextloadi8, sextloadi16>; // sign extend
2966 defm : MemOpi_bitExtType<extloadi8, extloadi16>; // any extend
2968 // memw(Rs+#0) = [clrbit|setbit](#U5)
2969 // memw(Rs+#u6:2) = [clrbit|setbit](#U5)
2970 defm : MemOpi_bitPats<load, store, Clr5ImmPred, u6_2ExtPred, ADDRriU6_2,
2971 CLRMEMIMM, L4_iand_memopw_io, and>;
2972 defm : MemOpi_bitPats<load, store, Set5ImmPred, u6_2ExtPred, ADDRriU6_2,
2973 SETMEMIMM, L4_ior_memopw_io, or>;
2976 //===----------------------------------------------------------------------===//
2977 // Multiclass to define 'def Pats' for ALU operations on the memory
2978 // where addend is a register.
2979 // mem[bhw](Rs+#0) [+-&|]= Rt
2980 // mem[bhw](Rs+#U6:[012]) [+-&|]= Rt
2981 //===----------------------------------------------------------------------===//
2983 multiclass MemOpr_Pats <PatFrag ldOp, PatFrag stOp, ComplexPattern addrPred,
2984 PatLeaf extPred, InstHexagon MI, SDNode OpNode> {
2985 let AddedComplexity = 141 in
2986 // mem[bhw](Rs+#0) [+-&|]= Rt
2987 def : Pat <(stOp (OpNode (ldOp (addrPred IntRegs:$addr, extPred:$offset)),
2988 (i32 IntRegs:$addend)),
2989 (addrPred (i32 IntRegs:$addr), extPred:$offset)),
2990 (MI IntRegs:$addr, extPred:$offset, (i32 IntRegs:$addend) )>;
2992 // mem[bhw](Rs+#U6:[012]) [+-&|]= Rt
2993 let AddedComplexity = 150 in
2994 def : Pat <(stOp (OpNode (ldOp (add IntRegs:$base, extPred:$offset)),
2995 (i32 IntRegs:$orend)),
2996 (add IntRegs:$base, extPred:$offset)),
2997 (MI IntRegs:$base, extPred:$offset, (i32 IntRegs:$orend) )>;
3000 multiclass MemOPr_ALUOp<PatFrag ldOp, PatFrag stOp,
3001 ComplexPattern addrPred, PatLeaf extPred,
3002 InstHexagon addMI, InstHexagon subMI,
3003 InstHexagon andMI, InstHexagon orMI > {
3005 defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, addMI, add>;
3006 defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, subMI, sub>;
3007 defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, andMI, and>;
3008 defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, orMI, or>;
3011 multiclass MemOPr_ExtType<PatFrag ldOpByte, PatFrag ldOpHalf > {
3013 defm : MemOPr_ALUOp <ldOpHalf, truncstorei16, ADDRriU6_1, u6_1ExtPred,
3014 L4_add_memoph_io, L4_sub_memoph_io,
3015 L4_and_memoph_io, L4_or_memoph_io>;
3017 defm : MemOPr_ALUOp <ldOpByte, truncstorei8, ADDRriU6_0, u6ExtPred,
3018 L4_add_memopb_io, L4_sub_memopb_io,
3019 L4_and_memopb_io, L4_or_memopb_io>;
3022 // Define 'def Pats' for MemOps with register addend.
3023 let Predicates = [HasV4T, UseMEMOP] in {
3025 defm : MemOPr_ExtType<zextloadi8, zextloadi16>; // zero extend
3026 defm : MemOPr_ExtType<sextloadi8, sextloadi16>; // sign extend
3027 defm : MemOPr_ExtType<extloadi8, extloadi16>; // any extend
3029 defm : MemOPr_ALUOp <load, store, ADDRriU6_2, u6_2ExtPred, L4_add_memopw_io,
3030 L4_sub_memopw_io, L4_and_memopw_io, L4_or_memopw_io >;
3033 //===----------------------------------------------------------------------===//
3035 //===----------------------------------------------------------------------===//
3037 // Hexagon V4 only supports these flavors of byte/half compare instructions:
3038 // EQ/GT/GTU. Other flavors like GE/GEU/LT/LTU/LE/LEU are not supported by
3039 // hardware. However, compiler can still implement these patterns through
3040 // appropriate patterns combinations based on current implemented patterns.
3041 // The implemented patterns are: EQ/GT/GTU.
3042 // Missing patterns are: GE/GEU/LT/LTU/LE/LEU.
3044 // Following instruction is not being extended as it results into the
3045 // incorrect code for negative numbers.
3046 // Pd=cmpb.eq(Rs,#u8)
3048 // p=!cmp.eq(r1,#s10)
3049 let isCodeGenOnly = 0 in {
3050 def C4_cmpneqi : T_CMP <"cmp.eq", 0b00, 1, s10Ext>;
3051 def C4_cmpltei : T_CMP <"cmp.gt", 0b01, 1, s10Ext>;
3052 def C4_cmplteui : T_CMP <"cmp.gtu", 0b10, 1, u9Ext>;
3055 def : T_CMP_pat <C4_cmpneqi, setne, s10ExtPred>;
3056 def : T_CMP_pat <C4_cmpltei, setle, s10ExtPred>;
3057 def : T_CMP_pat <C4_cmplteui, setule, u9ImmPred>;
3059 // rs <= rt -> !(rs > rt).
3061 def: Pat<(i1 (setle (i32 IntRegs:$src1), s10ExtPred:$src2)),
3062 (C2_not (C2_cmpgti IntRegs:$src1, s10ExtPred:$src2))>;
3063 // (C4_cmpltei IntRegs:$src1, s10ExtPred:$src2)>;
3065 // Map cmplt(Rs, Imm) -> !cmpgt(Rs, Imm-1).
3066 def: Pat<(i1 (setlt (i32 IntRegs:$src1), s8ExtPred:$src2)),
3067 (C4_cmpltei IntRegs:$src1, (DEC_CONST_SIGNED s8ExtPred:$src2))>;
3069 // rs != rt -> !(rs == rt).
3070 def: Pat<(i1 (setne (i32 IntRegs:$src1), s10ExtPred:$src2)),
3071 (C4_cmpneqi IntRegs:$src1, s10ExtPred:$src2)>;
3073 // SDNode for converting immediate C to C-1.
3074 def DEC_CONST_BYTE : SDNodeXForm<imm, [{
3075 // Return the byte immediate const-1 as an SDNode.
3076 int32_t imm = N->getSExtValue();
3077 return XformU7ToU7M1Imm(imm);
3081 // zext( seteq ( and(Rs, 255), u8))
3083 // Pd=cmpb.eq(Rs, #u8)
3084 // if (Pd.new) Rd=#1
3085 // if (!Pd.new) Rd=#0
3086 def : Pat <(i32 (zext (i1 (seteq (i32 (and (i32 IntRegs:$Rs), 255)),
3088 (i32 (TFR_condset_ii (i1 (A4_cmpbeqi (i32 IntRegs:$Rs),
3094 // zext( setne ( and(Rs, 255), u8))
3096 // Pd=cmpb.eq(Rs, #u8)
3097 // if (Pd.new) Rd=#0
3098 // if (!Pd.new) Rd=#1
3099 def : Pat <(i32 (zext (i1 (setne (i32 (and (i32 IntRegs:$Rs), 255)),
3101 (i32 (TFR_condset_ii (i1 (A4_cmpbeqi (i32 IntRegs:$Rs),
3107 // zext( seteq (Rs, and(Rt, 255)))
3109 // Pd=cmpb.eq(Rs, Rt)
3110 // if (Pd.new) Rd=#1
3111 // if (!Pd.new) Rd=#0
3112 def : Pat <(i32 (zext (i1 (seteq (i32 IntRegs:$Rt),
3113 (i32 (and (i32 IntRegs:$Rs), 255)))))),
3114 (i32 (TFR_condset_ii (i1 (A4_cmpbeq (i32 IntRegs:$Rs),
3115 (i32 IntRegs:$Rt))),
3120 // zext( setne (Rs, and(Rt, 255)))
3122 // Pd=cmpb.eq(Rs, Rt)
3123 // if (Pd.new) Rd=#0
3124 // if (!Pd.new) Rd=#1
3125 def : Pat <(i32 (zext (i1 (setne (i32 IntRegs:$Rt),
3126 (i32 (and (i32 IntRegs:$Rs), 255)))))),
3127 (i32 (TFR_condset_ii (i1 (A4_cmpbeq (i32 IntRegs:$Rs),
3128 (i32 IntRegs:$Rt))),
3133 // zext( setugt ( and(Rs, 255), u8))
3135 // Pd=cmpb.gtu(Rs, #u8)
3136 // if (Pd.new) Rd=#1
3137 // if (!Pd.new) Rd=#0
3138 def : Pat <(i32 (zext (i1 (setugt (i32 (and (i32 IntRegs:$Rs), 255)),
3140 (i32 (TFR_condset_ii (i1 (A4_cmpbgtui (i32 IntRegs:$Rs),
3146 // zext( setugt ( and(Rs, 254), u8))
3148 // Pd=cmpb.gtu(Rs, #u8)
3149 // if (Pd.new) Rd=#1
3150 // if (!Pd.new) Rd=#0
3151 def : Pat <(i32 (zext (i1 (setugt (i32 (and (i32 IntRegs:$Rs), 254)),
3153 (i32 (TFR_condset_ii (i1 (A4_cmpbgtui (i32 IntRegs:$Rs),
3159 // zext( setult ( Rs, Rt))
3161 // Pd=cmp.ltu(Rs, Rt)
3162 // if (Pd.new) Rd=#1
3163 // if (!Pd.new) Rd=#0
3164 // cmp.ltu(Rs, Rt) -> cmp.gtu(Rt, Rs)
3165 def : Pat <(i32 (zext (i1 (setult (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3166 (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rt),
3167 (i32 IntRegs:$Rs))),
3172 // zext( setlt ( Rs, Rt))
3174 // Pd=cmp.lt(Rs, Rt)
3175 // if (Pd.new) Rd=#1
3176 // if (!Pd.new) Rd=#0
3177 // cmp.lt(Rs, Rt) -> cmp.gt(Rt, Rs)
3178 def : Pat <(i32 (zext (i1 (setlt (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3179 (i32 (TFR_condset_ii (i1 (C2_cmpgt (i32 IntRegs:$Rt),
3180 (i32 IntRegs:$Rs))),
3185 // zext( setugt ( Rs, Rt))
3187 // Pd=cmp.gtu(Rs, Rt)
3188 // if (Pd.new) Rd=#1
3189 // if (!Pd.new) Rd=#0
3190 def : Pat <(i32 (zext (i1 (setugt (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3191 (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rs),
3192 (i32 IntRegs:$Rt))),
3196 // This pattern interefers with coremark performance, not implementing at this
3199 // zext( setgt ( Rs, Rt))
3201 // Pd=cmp.gt(Rs, Rt)
3202 // if (Pd.new) Rd=#1
3203 // if (!Pd.new) Rd=#0
3206 // zext( setuge ( Rs, Rt))
3208 // Pd=cmp.ltu(Rs, Rt)
3209 // if (Pd.new) Rd=#0
3210 // if (!Pd.new) Rd=#1
3211 // cmp.ltu(Rs, Rt) -> cmp.gtu(Rt, Rs)
3212 def : Pat <(i32 (zext (i1 (setuge (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3213 (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rt),
3214 (i32 IntRegs:$Rs))),
3219 // zext( setge ( Rs, Rt))
3221 // Pd=cmp.lt(Rs, Rt)
3222 // if (Pd.new) Rd=#0
3223 // if (!Pd.new) Rd=#1
3224 // cmp.lt(Rs, Rt) -> cmp.gt(Rt, Rs)
3225 def : Pat <(i32 (zext (i1 (setge (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3226 (i32 (TFR_condset_ii (i1 (C2_cmpgt (i32 IntRegs:$Rt),
3227 (i32 IntRegs:$Rs))),
3232 // zext( setule ( Rs, Rt))
3234 // Pd=cmp.gtu(Rs, Rt)
3235 // if (Pd.new) Rd=#0
3236 // if (!Pd.new) Rd=#1
3237 def : Pat <(i32 (zext (i1 (setule (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3238 (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rs),
3239 (i32 IntRegs:$Rt))),
3244 // zext( setle ( Rs, Rt))
3246 // Pd=cmp.gt(Rs, Rt)
3247 // if (Pd.new) Rd=#0
3248 // if (!Pd.new) Rd=#1
3249 def : Pat <(i32 (zext (i1 (setle (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))),
3250 (i32 (TFR_condset_ii (i1 (C2_cmpgt (i32 IntRegs:$Rs),
3251 (i32 IntRegs:$Rt))),
3256 // zext( setult ( and(Rs, 255), u8))
3257 // Use the isdigit transformation below
3259 // Generate code of the form 'mux_ii(cmpbgtu(Rdd, C-1),0,1)'
3260 // for C code of the form r = ((c>='0') & (c<='9')) ? 1 : 0;.
3261 // The isdigit transformation relies on two 'clever' aspects:
3262 // 1) The data type is unsigned which allows us to eliminate a zero test after
3263 // biasing the expression by 48. We are depending on the representation of
3264 // the unsigned types, and semantics.
3265 // 2) The front end has converted <= 9 into < 10 on entry to LLVM
3268 // retval = ((c>='0') & (c<='9')) ? 1 : 0;
3269 // The code is transformed upstream of llvm into
3270 // retval = (c-48) < 10 ? 1 : 0;
3271 let AddedComplexity = 139 in
3272 def : Pat <(i32 (zext (i1 (setult (i32 (and (i32 IntRegs:$src1), 255)),
3273 u7StrictPosImmPred:$src2)))),
3274 (i32 (C2_muxii (i1 (A4_cmpbgtui (i32 IntRegs:$src1),
3275 (DEC_CONST_BYTE u7StrictPosImmPred:$src2))),
3279 //===----------------------------------------------------------------------===//
3281 //===----------------------------------------------------------------------===//
3283 //===----------------------------------------------------------------------===//
3284 // Multiclass for DeallocReturn
3285 //===----------------------------------------------------------------------===//
3286 class L4_RETURN<string mnemonic, bit isNot, bit isPredNew, bit isTak>
3287 : LD0Inst<(outs), (ins PredRegs:$src),
3288 !if(isNot, "if (!$src", "if ($src")#
3289 !if(isPredNew, ".new) ", ") ")#mnemonic#
3290 !if(isPredNew, #!if(isTak,":t", ":nt"),""),
3291 [], "", LD_tc_3or4stall_SLOT0> {
3294 let BaseOpcode = "L4_RETURN";
3295 let isPredicatedFalse = isNot;
3296 let isPredicatedNew = isPredNew;
3297 let isTaken = isTak;
3298 let IClass = 0b1001;
3300 let Inst{27-16} = 0b011000011110;
3302 let Inst{13} = isNot;
3303 let Inst{12} = isTak;
3304 let Inst{11} = isPredNew;
3306 let Inst{9-8} = src;
3307 let Inst{4-0} = 0b11110;
3310 // Produce all predicated forms, p, !p, p.new, !p.new, :t, :nt
3311 multiclass L4_RETURN_PRED<string mnemonic, bit PredNot> {
3312 let isPredicated = 1 in {
3313 def _#NAME# : L4_RETURN <mnemonic, PredNot, 0, 1>;
3314 def _#NAME#new_pnt : L4_RETURN <mnemonic, PredNot, 1, 0>;
3315 def _#NAME#new_pt : L4_RETURN <mnemonic, PredNot, 1, 1>;
3319 multiclass LD_MISC_L4_RETURN<string mnemonic> {
3320 let isBarrier = 1, isPredicable = 1 in
3321 def NAME : LD0Inst <(outs), (ins), mnemonic, [], "",
3322 LD_tc_3or4stall_SLOT0> {
3323 let BaseOpcode = "L4_RETURN";
3324 let IClass = 0b1001;
3325 let Inst{27-16} = 0b011000011110;
3326 let Inst{13-10} = 0b0000;
3327 let Inst{4-0} = 0b11110;
3329 defm t : L4_RETURN_PRED<mnemonic, 0 >;
3330 defm f : L4_RETURN_PRED<mnemonic, 1 >;
3333 let isReturn = 1, isTerminator = 1,
3334 Defs = [R29, R30, R31, PC], Uses = [R30], hasSideEffects = 0,
3335 validSubTargets = HasV4SubT, isCodeGenOnly = 0 in
3336 defm L4_return: LD_MISC_L4_RETURN <"dealloc_return">, PredNewRel;
3338 // Restore registers and dealloc return function call.
3339 let isCall = 1, isBarrier = 1, isReturn = 1, isTerminator = 1,
3340 Defs = [R29, R30, R31, PC], isAsmParserOnly = 1 in {
3341 let validSubTargets = HasV4SubT in
3342 def RESTORE_DEALLOC_RET_JMP_V4 : JInst<(outs),
3343 (ins calltarget:$dst),
3349 // Restore registers and dealloc frame before a tail call.
3350 let isCall = 1, isBarrier = 1, isAsmParserOnly = 1,
3351 Defs = [R29, R30, R31, PC] in {
3352 let validSubTargets = HasV4SubT in
3353 def RESTORE_DEALLOC_BEFORE_TAILCALL_V4 : JInst<(outs),
3354 (ins calltarget:$dst),
3360 // Save registers function call.
3361 let isCall = 1, isBarrier = 1, isAsmParserOnly = 1,
3362 Uses = [R29, R31] in {
3363 def SAVE_REGISTERS_CALL_V4 : JInst<(outs),
3364 (ins calltarget:$dst),
3365 "call $dst // Save_calle_saved_registers",
3370 //===----------------------------------------------------------------------===//
3371 // Template class for non predicated store instructions with
3372 // GP-Relative or absolute addressing.
3373 //===----------------------------------------------------------------------===//
3374 let hasSideEffects = 0, isPredicable = 1, isNVStorable = 1 in
3375 class T_StoreAbsGP <string mnemonic, RegisterClass RC, Operand ImmOp,
3376 bits<2>MajOp, Operand AddrOp, bit isAbs, bit isHalf>
3377 : STInst<(outs), (ins AddrOp:$addr, RC:$src),
3378 mnemonic # !if(isAbs, "(##", "(#")#"$addr) = $src"#!if(isHalf, ".h",""),
3379 [], "", V2LDST_tc_st_SLOT01> {
3382 bits<16> offsetBits;
3384 string ImmOpStr = !cast<string>(ImmOp);
3385 let offsetBits = !if (!eq(ImmOpStr, "u16_3Imm"), addr{18-3},
3386 !if (!eq(ImmOpStr, "u16_2Imm"), addr{17-2},
3387 !if (!eq(ImmOpStr, "u16_1Imm"), addr{16-1},
3388 /* u16_0Imm */ addr{15-0})));
3389 let IClass = 0b0100;
3391 let Inst{26-25} = offsetBits{15-14};
3393 let Inst{23-22} = MajOp;
3394 let Inst{21} = isHalf;
3395 let Inst{20-16} = offsetBits{13-9};
3396 let Inst{13} = offsetBits{8};
3397 let Inst{12-8} = src;
3398 let Inst{7-0} = offsetBits{7-0};
3401 //===----------------------------------------------------------------------===//
3402 // Template class for predicated store instructions with
3403 // GP-Relative or absolute addressing.
3404 //===----------------------------------------------------------------------===//
3405 let hasSideEffects = 0, isPredicated = 1, isNVStorable = 1, opExtentBits = 6,
3407 class T_StoreAbs_Pred <string mnemonic, RegisterClass RC, bits<2> MajOp,
3408 bit isHalf, bit isNot, bit isNew>
3409 : STInst<(outs), (ins PredRegs:$src1, u6Ext:$absaddr, RC: $src2),
3410 !if(isNot, "if (!$src1", "if ($src1")#!if(isNew, ".new) ",
3411 ") ")#mnemonic#"(#$absaddr) = $src2"#!if(isHalf, ".h",""),
3412 [], "", ST_tc_st_SLOT01>, AddrModeRel {
3417 let isPredicatedNew = isNew;
3418 let isPredicatedFalse = isNot;
3420 let IClass = 0b1010;
3422 let Inst{27-24} = 0b1111;
3423 let Inst{23-22} = MajOp;
3424 let Inst{21} = isHalf;
3425 let Inst{17-16} = absaddr{5-4};
3426 let Inst{13} = isNew;
3427 let Inst{12-8} = src2;
3429 let Inst{6-3} = absaddr{3-0};
3430 let Inst{2} = isNot;
3431 let Inst{1-0} = src1;
3434 //===----------------------------------------------------------------------===//
3435 // Template class for predicated store instructions with absolute addressing.
3436 //===----------------------------------------------------------------------===//
3437 class T_StoreAbs <string mnemonic, RegisterClass RC, Operand ImmOp,
3438 bits<2> MajOp, bit isHalf>
3439 : T_StoreAbsGP <mnemonic, RC, ImmOp, MajOp, u0AlwaysExt, 1, isHalf>,
3441 string ImmOpStr = !cast<string>(ImmOp);
3442 let opExtentBits = !if (!eq(ImmOpStr, "u16_3Imm"), 19,
3443 !if (!eq(ImmOpStr, "u16_2Imm"), 18,
3444 !if (!eq(ImmOpStr, "u16_1Imm"), 17,
3445 /* u16_0Imm */ 16)));
3447 let opExtentAlign = !if (!eq(ImmOpStr, "u16_3Imm"), 3,
3448 !if (!eq(ImmOpStr, "u16_2Imm"), 2,
3449 !if (!eq(ImmOpStr, "u16_1Imm"), 1,
3450 /* u16_0Imm */ 0)));
3453 //===----------------------------------------------------------------------===//
3454 // Multiclass for store instructions with absolute addressing.
3455 //===----------------------------------------------------------------------===//
3456 let validSubTargets = HasV4SubT, addrMode = Absolute, isExtended = 1 in
3457 multiclass ST_Abs<string mnemonic, string CextOp, RegisterClass RC,
3458 Operand ImmOp, bits<2> MajOp, bit isHalf = 0> {
3459 let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in {
3460 let opExtendable = 0, isPredicable = 1 in
3461 def S2_#NAME#abs : T_StoreAbs <mnemonic, RC, ImmOp, MajOp, isHalf>;
3464 def S4_p#NAME#t_abs : T_StoreAbs_Pred<mnemonic, RC, MajOp, isHalf, 0, 0>;
3465 def S4_p#NAME#f_abs : T_StoreAbs_Pred<mnemonic, RC, MajOp, isHalf, 1, 0>;
3468 def S4_p#NAME#tnew_abs : T_StoreAbs_Pred<mnemonic, RC, MajOp, isHalf, 0, 1>;
3469 def S4_p#NAME#fnew_abs : T_StoreAbs_Pred<mnemonic, RC, MajOp, isHalf, 1, 1>;
3473 //===----------------------------------------------------------------------===//
3474 // Template class for non predicated new-value store instructions with
3475 // GP-Relative or absolute addressing.
3476 //===----------------------------------------------------------------------===//
3477 let hasSideEffects = 0, isPredicable = 1, mayStore = 1, isNVStore = 1,
3478 isNewValue = 1, opNewValue = 1 in
3479 class T_StoreAbsGP_NV <string mnemonic, Operand ImmOp, bits<2>MajOp, bit isAbs>
3480 : NVInst_V4<(outs), (ins u0AlwaysExt:$addr, IntRegs:$src),
3481 mnemonic # !if(isAbs, "(##", "(#")#"$addr) = $src.new",
3482 [], "", V2LDST_tc_st_SLOT0> {
3485 bits<16> offsetBits;
3487 string ImmOpStr = !cast<string>(ImmOp);
3488 let offsetBits = !if (!eq(ImmOpStr, "u16_3Imm"), addr{18-3},
3489 !if (!eq(ImmOpStr, "u16_2Imm"), addr{17-2},
3490 !if (!eq(ImmOpStr, "u16_1Imm"), addr{16-1},
3491 /* u16_0Imm */ addr{15-0})));
3492 let IClass = 0b0100;
3495 let Inst{26-25} = offsetBits{15-14};
3496 let Inst{24-21} = 0b0101;
3497 let Inst{20-16} = offsetBits{13-9};
3498 let Inst{13} = offsetBits{8};
3499 let Inst{12-11} = MajOp;
3500 let Inst{10-8} = src;
3501 let Inst{7-0} = offsetBits{7-0};
3504 //===----------------------------------------------------------------------===//
3505 // Template class for predicated new-value store instructions with
3506 // absolute addressing.
3507 //===----------------------------------------------------------------------===//
3508 let hasSideEffects = 0, isPredicated = 1, mayStore = 1, isNVStore = 1,
3509 isNewValue = 1, opNewValue = 2, opExtentBits = 6, opExtendable = 1 in
3510 class T_StoreAbs_NV_Pred <string mnemonic, bits<2> MajOp, bit isNot, bit isNew>
3511 : NVInst_V4<(outs), (ins PredRegs:$src1, u6Ext:$absaddr, IntRegs:$src2),
3512 !if(isNot, "if (!$src1", "if ($src1")#!if(isNew, ".new) ",
3513 ") ")#mnemonic#"(#$absaddr) = $src2.new",
3514 [], "", ST_tc_st_SLOT0>, AddrModeRel {
3519 let isPredicatedNew = isNew;
3520 let isPredicatedFalse = isNot;
3522 let IClass = 0b1010;
3524 let Inst{27-24} = 0b1111;
3525 let Inst{23-21} = 0b101;
3526 let Inst{17-16} = absaddr{5-4};
3527 let Inst{13} = isNew;
3528 let Inst{12-11} = MajOp;
3529 let Inst{10-8} = src2;
3531 let Inst{6-3} = absaddr{3-0};
3532 let Inst{2} = isNot;
3533 let Inst{1-0} = src1;
3536 //===----------------------------------------------------------------------===//
3537 // Template class for non-predicated new-value store instructions with
3538 // absolute addressing.
3539 //===----------------------------------------------------------------------===//
3540 class T_StoreAbs_NV <string mnemonic, Operand ImmOp, bits<2> MajOp>
3541 : T_StoreAbsGP_NV <mnemonic, ImmOp, MajOp, 1>, AddrModeRel {
3543 string ImmOpStr = !cast<string>(ImmOp);
3544 let opExtentBits = !if (!eq(ImmOpStr, "u16_3Imm"), 19,
3545 !if (!eq(ImmOpStr, "u16_2Imm"), 18,
3546 !if (!eq(ImmOpStr, "u16_1Imm"), 17,
3547 /* u16_0Imm */ 16)));
3549 let opExtentAlign = !if (!eq(ImmOpStr, "u16_3Imm"), 3,
3550 !if (!eq(ImmOpStr, "u16_2Imm"), 2,
3551 !if (!eq(ImmOpStr, "u16_1Imm"), 1,
3552 /* u16_0Imm */ 0)));
3555 //===----------------------------------------------------------------------===//
3556 // Multiclass for new-value store instructions with absolute addressing.
3557 //===----------------------------------------------------------------------===//
3558 let validSubTargets = HasV4SubT, addrMode = Absolute, isExtended = 1 in
3559 multiclass ST_Abs_NV <string mnemonic, string CextOp, Operand ImmOp,
3561 let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in {
3562 let opExtendable = 0, isPredicable = 1 in
3563 def S2_#NAME#newabs : T_StoreAbs_NV <mnemonic, ImmOp, MajOp>;
3566 def S4_p#NAME#newt_abs : T_StoreAbs_NV_Pred <mnemonic, MajOp, 0, 0>;
3567 def S4_p#NAME#newf_abs : T_StoreAbs_NV_Pred <mnemonic, MajOp, 1, 0>;
3570 def S4_p#NAME#newtnew_abs : T_StoreAbs_NV_Pred <mnemonic, MajOp, 0, 1>;
3571 def S4_p#NAME#newfnew_abs : T_StoreAbs_NV_Pred <mnemonic, MajOp, 1, 1>;
3575 //===----------------------------------------------------------------------===//
3576 // Stores with absolute addressing
3577 //===----------------------------------------------------------------------===//
3578 let accessSize = ByteAccess, isCodeGenOnly = 0 in
3579 defm storerb : ST_Abs <"memb", "STrib", IntRegs, u16_0Imm, 0b00>,
3580 ST_Abs_NV <"memb", "STrib", u16_0Imm, 0b00>;
3582 let accessSize = HalfWordAccess, isCodeGenOnly = 0 in
3583 defm storerh : ST_Abs <"memh", "STrih", IntRegs, u16_1Imm, 0b01>,
3584 ST_Abs_NV <"memh", "STrih", u16_1Imm, 0b01>;
3586 let accessSize = WordAccess, isCodeGenOnly = 0 in
3587 defm storeri : ST_Abs <"memw", "STriw", IntRegs, u16_2Imm, 0b10>,
3588 ST_Abs_NV <"memw", "STriw", u16_2Imm, 0b10>;
3590 let isNVStorable = 0, accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
3591 defm storerd : ST_Abs <"memd", "STrid", DoubleRegs, u16_3Imm, 0b11>;
3593 let isNVStorable = 0, accessSize = HalfWordAccess, isCodeGenOnly = 0 in
3594 defm storerf : ST_Abs <"memh", "STrif", IntRegs, u16_1Imm, 0b01, 1>;
3596 //===----------------------------------------------------------------------===//
3597 // GP-relative stores.
3598 // mem[bhwd](#global)=Rt
3599 // Once predicated, these instructions map to absolute addressing mode.
3600 // if ([!]Pv[.new]) mem[bhwd](##global)=Rt
3601 //===----------------------------------------------------------------------===//
3603 let validSubTargets = HasV4SubT, isAsmParserOnly = 1 in
3604 class T_StoreGP <string mnemonic, string BaseOp, RegisterClass RC,
3605 Operand ImmOp, bits<2> MajOp, bit isHalf = 0>
3606 : T_StoreAbsGP <mnemonic, RC, ImmOp, MajOp, globaladdress, 0, isHalf> {
3607 // Set BaseOpcode same as absolute addressing instructions so that
3608 // non-predicated GP-Rel instructions can have relate with predicated
3609 // Absolute instruction.
3610 let BaseOpcode = BaseOp#_abs;
3613 let validSubTargets = HasV4SubT, isAsmParserOnly = 1 in
3614 multiclass ST_GP <string mnemonic, string BaseOp, Operand ImmOp,
3615 bits<2> MajOp, bit isHalf = 0> {
3616 // Set BaseOpcode same as absolute addressing instructions so that
3617 // non-predicated GP-Rel instructions can have relate with predicated
3618 // Absolute instruction.
3619 let BaseOpcode = BaseOp#_abs in {
3620 def NAME#gp : T_StoreAbsGP <mnemonic, IntRegs, ImmOp, MajOp,
3621 globaladdress, 0, isHalf>;
3623 def NAME#newgp : T_StoreAbsGP_NV <mnemonic, ImmOp, MajOp, 0> ;
3627 let accessSize = ByteAccess in
3628 defm S2_storerb : ST_GP<"memb", "STrib", u16_0Imm, 0b00>, NewValueRel;
3630 let accessSize = HalfWordAccess in
3631 defm S2_storerh : ST_GP<"memh", "STrih", u16_1Imm, 0b01>, NewValueRel;
3633 let accessSize = WordAccess in
3634 defm S2_storeri : ST_GP<"memw", "STriw", u16_2Imm, 0b10>, NewValueRel;
3636 let isNVStorable = 0, accessSize = DoubleWordAccess in
3637 def S2_storerdgp : T_StoreGP <"memd", "STrid", DoubleRegs,
3638 u16_3Imm, 0b11>, PredNewRel;
3640 let isNVStorable = 0, accessSize = HalfWordAccess in
3641 def S2_storerfgp : T_StoreGP <"memh", "STrif", IntRegs,
3642 u16_1Imm, 0b01, 1>, PredNewRel;
3644 class Loada_pat<PatFrag Load, ValueType VT, PatFrag Addr, InstHexagon MI>
3645 : Pat<(VT (Load Addr:$addr)), (MI Addr:$addr)>;
3647 class Loadam_pat<PatFrag Load, ValueType VT, PatFrag Addr, PatFrag ValueMod,
3649 : Pat<(VT (Load Addr:$addr)), (ValueMod (MI Addr:$addr))>;
3651 let Predicates = [HasV4T], AddedComplexity = 30 in {
3652 def : Pat<(truncstorei8 (i32 IntRegs:$src1),
3653 (HexagonCONST32 tglobaladdr:$absaddr)),
3654 (S2_storerbabs tglobaladdr: $absaddr, IntRegs: $src1)>;
3656 def : Pat<(truncstorei16 (i32 IntRegs:$src1),
3657 (HexagonCONST32 tglobaladdr:$absaddr)),
3658 (S2_storerhabs tglobaladdr: $absaddr, IntRegs: $src1)>;
3660 def : Pat<(store (i32 IntRegs:$src1), (HexagonCONST32 tglobaladdr:$absaddr)),
3661 (S2_storeriabs tglobaladdr: $absaddr, IntRegs: $src1)>;
3663 def : Pat<(store (i64 DoubleRegs:$src1),
3664 (HexagonCONST32 tglobaladdr:$absaddr)),
3665 (S2_storerdabs tglobaladdr: $absaddr, DoubleRegs: $src1)>;
3668 // 64 bit atomic store
3669 def : Pat <(atomic_store_64 (HexagonCONST32_GP tglobaladdr:$global),
3670 (i64 DoubleRegs:$src1)),
3671 (S2_storerdgp tglobaladdr:$global, (i64 DoubleRegs:$src1))>,
3674 // Map from store(globaladdress) -> memd(#foo)
3675 let AddedComplexity = 100 in
3676 def : Pat <(store (i64 DoubleRegs:$src1),
3677 (HexagonCONST32_GP tglobaladdr:$global)),
3678 (S2_storerdgp tglobaladdr:$global, (i64 DoubleRegs:$src1))>;
3680 // 8 bit atomic store
3681 def : Pat < (atomic_store_8 (HexagonCONST32_GP tglobaladdr:$global),
3682 (i32 IntRegs:$src1)),
3683 (S2_storerbgp tglobaladdr:$global, (i32 IntRegs:$src1))>;
3685 // Map from store(globaladdress) -> memb(#foo)
3686 let AddedComplexity = 100 in
3687 def : Pat<(truncstorei8 (i32 IntRegs:$src1),
3688 (HexagonCONST32_GP tglobaladdr:$global)),
3689 (S2_storerbgp tglobaladdr:$global, (i32 IntRegs:$src1))>;
3691 // Map from "i1 = constant<-1>; memw(CONST32(#foo)) = i1"
3692 // to "r0 = 1; memw(#foo) = r0"
3693 let AddedComplexity = 100 in
3694 def : Pat<(store (i1 -1), (HexagonCONST32_GP tglobaladdr:$global)),
3695 (S2_storerbgp tglobaladdr:$global, (A2_tfrsi 1))>;
3697 def : Pat<(atomic_store_16 (HexagonCONST32_GP tglobaladdr:$global),
3698 (i32 IntRegs:$src1)),
3699 (S2_storerhgp tglobaladdr:$global, (i32 IntRegs:$src1))>;
3701 // Map from store(globaladdress) -> memh(#foo)
3702 let AddedComplexity = 100 in
3703 def : Pat<(truncstorei16 (i32 IntRegs:$src1),
3704 (HexagonCONST32_GP tglobaladdr:$global)),
3705 (S2_storerhgp tglobaladdr:$global, (i32 IntRegs:$src1))>;
3707 // 32 bit atomic store
3708 def : Pat<(atomic_store_32 (HexagonCONST32_GP tglobaladdr:$global),
3709 (i32 IntRegs:$src1)),
3710 (S2_storerigp tglobaladdr:$global, (i32 IntRegs:$src1))>;
3712 // Map from store(globaladdress) -> memw(#foo)
3713 let AddedComplexity = 100 in
3714 def : Pat<(store (i32 IntRegs:$src1), (HexagonCONST32_GP tglobaladdr:$global)),
3715 (S2_storerigp tglobaladdr:$global, (i32 IntRegs:$src1))>;
3717 //===----------------------------------------------------------------------===//
3718 // Template class for non predicated load instructions with
3719 // absolute addressing mode.
3720 //===----------------------------------------------------------------------===//
3721 let isPredicable = 1, hasSideEffects = 0, validSubTargets = HasV4SubT in
3722 class T_LoadAbsGP <string mnemonic, RegisterClass RC, Operand ImmOp,
3723 bits<3> MajOp, Operand AddrOp, bit isAbs>
3724 : LDInst <(outs RC:$dst), (ins AddrOp:$addr),
3725 "$dst = "#mnemonic# !if(isAbs, "(##", "(#")#"$addr)",
3726 [], "", V2LDST_tc_ld_SLOT01> {
3729 bits<16> offsetBits;
3731 string ImmOpStr = !cast<string>(ImmOp);
3732 let offsetBits = !if (!eq(ImmOpStr, "u16_3Imm"), addr{18-3},
3733 !if (!eq(ImmOpStr, "u16_2Imm"), addr{17-2},
3734 !if (!eq(ImmOpStr, "u16_1Imm"), addr{16-1},
3735 /* u16_0Imm */ addr{15-0})));
3737 let IClass = 0b0100;
3740 let Inst{26-25} = offsetBits{15-14};
3742 let Inst{23-21} = MajOp;
3743 let Inst{20-16} = offsetBits{13-9};
3744 let Inst{13-5} = offsetBits{8-0};
3745 let Inst{4-0} = dst;
3748 class T_LoadAbs <string mnemonic, RegisterClass RC, Operand ImmOp,
3750 : T_LoadAbsGP <mnemonic, RC, ImmOp, MajOp, u0AlwaysExt, 1>, AddrModeRel {
3752 string ImmOpStr = !cast<string>(ImmOp);
3753 let opExtentBits = !if (!eq(ImmOpStr, "u16_3Imm"), 19,
3754 !if (!eq(ImmOpStr, "u16_2Imm"), 18,
3755 !if (!eq(ImmOpStr, "u16_1Imm"), 17,
3756 /* u16_0Imm */ 16)));
3758 let opExtentAlign = !if (!eq(ImmOpStr, "u16_3Imm"), 3,
3759 !if (!eq(ImmOpStr, "u16_2Imm"), 2,
3760 !if (!eq(ImmOpStr, "u16_1Imm"), 1,
3761 /* u16_0Imm */ 0)));
3763 //===----------------------------------------------------------------------===//
3764 // Template class for predicated load instructions with
3765 // absolute addressing mode.
3766 //===----------------------------------------------------------------------===//
3767 let isPredicated = 1, hasNewValue = 1, opExtentBits = 6, opExtendable = 2 in
3768 class T_LoadAbs_Pred <string mnemonic, RegisterClass RC, bits<3> MajOp,
3769 bit isPredNot, bit isPredNew>
3770 : LDInst <(outs RC:$dst), (ins PredRegs:$src1, u6Ext:$absaddr),
3771 !if(isPredNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
3772 ") ")#"$dst = "#mnemonic#"(#$absaddr)">, AddrModeRel {
3777 let isPredicatedNew = isPredNew;
3778 let isPredicatedFalse = isPredNot;
3780 let IClass = 0b1001;
3782 let Inst{27-24} = 0b1111;
3783 let Inst{23-21} = MajOp;
3784 let Inst{20-16} = absaddr{5-1};
3786 let Inst{12} = isPredNew;
3787 let Inst{11} = isPredNot;
3788 let Inst{10-9} = src1;
3789 let Inst{8} = absaddr{0};
3791 let Inst{4-0} = dst;
3794 //===----------------------------------------------------------------------===//
3795 // Multiclass for the load instructions with absolute addressing mode.
3796 //===----------------------------------------------------------------------===//
3797 multiclass LD_Abs_Pred<string mnemonic, RegisterClass RC, bits<3> MajOp,
3799 def _abs : T_LoadAbs_Pred <mnemonic, RC, MajOp, PredNot, 0>;
3801 def new_abs : T_LoadAbs_Pred <mnemonic, RC, MajOp, PredNot, 1>;
3804 let addrMode = Absolute, isExtended = 1 in
3805 multiclass LD_Abs<string mnemonic, string CextOp, RegisterClass RC,
3806 Operand ImmOp, bits<3> MajOp> {
3807 let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in {
3808 let opExtendable = 1, isPredicable = 1 in
3809 def L4_#NAME#_abs: T_LoadAbs <mnemonic, RC, ImmOp, MajOp>;
3812 defm L4_p#NAME#t : LD_Abs_Pred<mnemonic, RC, MajOp, 0>;
3813 defm L4_p#NAME#f : LD_Abs_Pred<mnemonic, RC, MajOp, 1>;
3817 let accessSize = ByteAccess, hasNewValue = 1, isCodeGenOnly = 0 in {
3818 defm loadrb : LD_Abs<"memb", "LDrib", IntRegs, u16_0Imm, 0b000>;
3819 defm loadrub : LD_Abs<"memub", "LDriub", IntRegs, u16_0Imm, 0b001>;
3822 let accessSize = HalfWordAccess, hasNewValue = 1, isCodeGenOnly = 0 in {
3823 defm loadrh : LD_Abs<"memh", "LDrih", IntRegs, u16_1Imm, 0b010>;
3824 defm loadruh : LD_Abs<"memuh", "LDriuh", IntRegs, u16_1Imm, 0b011>;
3827 let accessSize = WordAccess, hasNewValue = 1, isCodeGenOnly = 0 in
3828 defm loadri : LD_Abs<"memw", "LDriw", IntRegs, u16_2Imm, 0b100>;
3830 let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
3831 defm loadrd : LD_Abs<"memd", "LDrid", DoubleRegs, u16_3Imm, 0b110>;
3833 //===----------------------------------------------------------------------===//
3834 // multiclass for load instructions with GP-relative addressing mode.
3835 // Rx=mem[bhwd](##global)
3836 // Once predicated, these instructions map to absolute addressing mode.
3837 // if ([!]Pv[.new]) Rx=mem[bhwd](##global)
3838 //===----------------------------------------------------------------------===//
3840 let isAsmParserOnly = 1 in
3841 class T_LoadGP <string mnemonic, string BaseOp, RegisterClass RC, Operand ImmOp,
3843 : T_LoadAbsGP <mnemonic, RC, ImmOp, MajOp, globaladdress, 0>, PredNewRel {
3844 let BaseOpcode = BaseOp#_abs;
3847 let accessSize = ByteAccess, hasNewValue = 1 in {
3848 def L2_loadrbgp : T_LoadGP<"memb", "LDrib", IntRegs, u16_0Imm, 0b000>;
3849 def L2_loadrubgp : T_LoadGP<"memub", "LDriub", IntRegs, u16_0Imm, 0b001>;
3852 let accessSize = HalfWordAccess, hasNewValue = 1 in {
3853 def L2_loadrhgp : T_LoadGP<"memh", "LDrih", IntRegs, u16_1Imm, 0b010>;
3854 def L2_loadruhgp : T_LoadGP<"memuh", "LDriuh", IntRegs, u16_1Imm, 0b011>;
3857 let accessSize = WordAccess, hasNewValue = 1 in
3858 def L2_loadrigp : T_LoadGP<"memw", "LDriw", IntRegs, u16_2Imm, 0b100>;
3860 let accessSize = DoubleWordAccess in
3861 def L2_loadrdgp : T_LoadGP<"memd", "LDrid", DoubleRegs, u16_3Imm, 0b110>;
3863 let Predicates = [HasV4T], AddedComplexity = 30 in {
3864 def : Pat<(i32 (load (HexagonCONST32 tglobaladdr:$absaddr))),
3865 (L4_loadri_abs tglobaladdr: $absaddr)>;
3867 def : Pat<(i32 (sextloadi8 (HexagonCONST32 tglobaladdr:$absaddr))),
3868 (L4_loadrb_abs tglobaladdr:$absaddr)>;
3870 def : Pat<(i32 (zextloadi8 (HexagonCONST32 tglobaladdr:$absaddr))),
3871 (L4_loadrub_abs tglobaladdr:$absaddr)>;
3873 def : Pat<(i32 (sextloadi16 (HexagonCONST32 tglobaladdr:$absaddr))),
3874 (L4_loadrh_abs tglobaladdr:$absaddr)>;
3876 def : Pat<(i32 (zextloadi16 (HexagonCONST32 tglobaladdr:$absaddr))),
3877 (L4_loadruh_abs tglobaladdr:$absaddr)>;
3880 def : Pat <(atomic_load_64 (HexagonCONST32_GP tglobaladdr:$global)),
3881 (i64 (L2_loadrdgp tglobaladdr:$global))>;
3883 def : Pat <(atomic_load_32 (HexagonCONST32_GP tglobaladdr:$global)),
3884 (i32 (L2_loadrigp tglobaladdr:$global))>;
3886 def : Pat <(atomic_load_16 (HexagonCONST32_GP tglobaladdr:$global)),
3887 (i32 (L2_loadruhgp tglobaladdr:$global))>;
3889 def : Pat <(atomic_load_8 (HexagonCONST32_GP tglobaladdr:$global)),
3890 (i32 (L2_loadrubgp tglobaladdr:$global))>;
3892 // Map from load(globaladdress) -> memw(#foo + 0)
3893 let AddedComplexity = 100 in
3894 def : Pat <(i64 (load (HexagonCONST32_GP tglobaladdr:$global))),
3895 (i64 (L2_loadrdgp tglobaladdr:$global))>;
3897 // Map from Pd = load(globaladdress) -> Rd = memb(globaladdress), Pd = Rd
3898 let AddedComplexity = 100 in
3899 def : Pat <(i1 (load (HexagonCONST32_GP tglobaladdr:$global))),
3900 (i1 (C2_tfrrp (i32 (L2_loadrbgp tglobaladdr:$global))))>;
3902 // When the Interprocedural Global Variable optimizer realizes that a certain
3903 // global variable takes only two constant values, it shrinks the global to
3904 // a boolean. Catch those loads here in the following 3 patterns.
3905 let AddedComplexity = 100 in
3906 def : Pat <(i32 (extloadi1 (HexagonCONST32_GP tglobaladdr:$global))),
3907 (i32 (L2_loadrbgp tglobaladdr:$global))>;
3909 let AddedComplexity = 100 in
3910 def : Pat <(i32 (sextloadi1 (HexagonCONST32_GP tglobaladdr:$global))),
3911 (i32 (L2_loadrbgp tglobaladdr:$global))>;
3913 // Map from load(globaladdress) -> memb(#foo)
3914 let AddedComplexity = 100 in
3915 def : Pat <(i32 (extloadi8 (HexagonCONST32_GP tglobaladdr:$global))),
3916 (i32 (L2_loadrbgp tglobaladdr:$global))>;
3918 // Map from load(globaladdress) -> memb(#foo)
3919 let AddedComplexity = 100 in
3920 def : Pat <(i32 (sextloadi8 (HexagonCONST32_GP tglobaladdr:$global))),
3921 (i32 (L2_loadrbgp tglobaladdr:$global))>;
3923 let AddedComplexity = 100 in
3924 def : Pat <(i32 (zextloadi1 (HexagonCONST32_GP tglobaladdr:$global))),
3925 (i32 (L2_loadrubgp tglobaladdr:$global))>;
3927 // Map from load(globaladdress) -> memub(#foo)
3928 let AddedComplexity = 100 in
3929 def : Pat <(i32 (zextloadi8 (HexagonCONST32_GP tglobaladdr:$global))),
3930 (i32 (L2_loadrubgp tglobaladdr:$global))>;
3932 // Map from load(globaladdress) -> memh(#foo)
3933 let AddedComplexity = 100 in
3934 def : Pat <(i32 (extloadi16 (HexagonCONST32_GP tglobaladdr:$global))),
3935 (i32 (L2_loadrhgp tglobaladdr:$global))>;
3937 // Map from load(globaladdress) -> memh(#foo)
3938 let AddedComplexity = 100 in
3939 def : Pat <(i32 (sextloadi16 (HexagonCONST32_GP tglobaladdr:$global))),
3940 (i32 (L2_loadrhgp tglobaladdr:$global))>;
3942 // Map from load(globaladdress) -> memuh(#foo)
3943 let AddedComplexity = 100 in
3944 def : Pat <(i32 (zextloadi16 (HexagonCONST32_GP tglobaladdr:$global))),
3945 (i32 (L2_loadruhgp tglobaladdr:$global))>;
3947 // Map from load(globaladdress) -> memw(#foo)
3948 let AddedComplexity = 100 in
3949 def : Pat <(i32 (load (HexagonCONST32_GP tglobaladdr:$global))),
3950 (i32 (L2_loadrigp tglobaladdr:$global))>;
3953 // Transfer global address into a register
3954 let isExtended = 1, opExtendable = 1, AddedComplexity=50, isMoveImm = 1,
3955 isAsCheapAsAMove = 1, isReMaterializable = 1, validSubTargets = HasV4SubT in
3956 def TFRI_V4 : ALU32_ri<(outs IntRegs:$dst), (ins s16Ext:$src1),
3958 [(set IntRegs:$dst, (HexagonCONST32 tglobaladdr:$src1))]>,
3961 // Transfer a block address into a register
3962 def : Pat<(HexagonCONST32_GP tblockaddress:$src1),
3963 (TFRI_V4 tblockaddress:$src1)>,
3966 let isExtended = 1, opExtendable = 2, AddedComplexity=50,
3967 hasSideEffects = 0, isPredicated = 1, validSubTargets = HasV4SubT in
3968 def TFRI_cPt_V4 : ALU32_ri<(outs IntRegs:$dst),
3969 (ins PredRegs:$src1, s16Ext:$src2),
3970 "if($src1) $dst = #$src2",
3974 let isExtended = 1, opExtendable = 2, AddedComplexity=50, isPredicatedFalse = 1,
3975 hasSideEffects = 0, isPredicated = 1, validSubTargets = HasV4SubT in
3976 def TFRI_cNotPt_V4 : ALU32_ri<(outs IntRegs:$dst),
3977 (ins PredRegs:$src1, s16Ext:$src2),
3978 "if(!$src1) $dst = #$src2",
3982 let isExtended = 1, opExtendable = 2, AddedComplexity=50,
3983 hasSideEffects = 0, isPredicated = 1, validSubTargets = HasV4SubT in
3984 def TFRI_cdnPt_V4 : ALU32_ri<(outs IntRegs:$dst),
3985 (ins PredRegs:$src1, s16Ext:$src2),
3986 "if($src1.new) $dst = #$src2",
3990 let isExtended = 1, opExtendable = 2, AddedComplexity=50, isPredicatedFalse = 1,
3991 hasSideEffects = 0, isPredicated = 1, validSubTargets = HasV4SubT in
3992 def TFRI_cdnNotPt_V4 : ALU32_ri<(outs IntRegs:$dst),
3993 (ins PredRegs:$src1, s16Ext:$src2),
3994 "if(!$src1.new) $dst = #$src2",
3998 let AddedComplexity = 50, Predicates = [HasV4T] in
3999 def : Pat<(HexagonCONST32_GP tglobaladdr:$src1),
4000 (TFRI_V4 tglobaladdr:$src1)>,
4003 let Predicates = [HasV4T], AddedComplexity = 30 in {
4004 def : Pat<(truncstorei8 (i32 IntRegs:$src1), u0AlwaysExtPred:$src2),
4005 (S2_storerbabs u0AlwaysExtPred:$src2, IntRegs: $src1)>;
4007 def : Pat<(truncstorei16 (i32 IntRegs:$src1), u0AlwaysExtPred:$src2),
4008 (S2_storerhabs u0AlwaysExtPred:$src2, IntRegs: $src1)>;
4010 def : Pat<(store (i32 IntRegs:$src1), u0AlwaysExtPred:$src2),
4011 (S2_storeriabs u0AlwaysExtPred:$src2, IntRegs: $src1)>;
4014 let Predicates = [HasV4T], AddedComplexity = 30 in {
4015 def : Pat<(i32 (load u0AlwaysExtPred:$src)),
4016 (L4_loadri_abs u0AlwaysExtPred:$src)>;
4019 // Indexed store word - global address.
4020 // memw(Rs+#u6:2)=#S8
4021 let AddedComplexity = 10 in
4022 def STriw_offset_ext_V4 : STInst<(outs),
4023 (ins IntRegs:$src1, u6_2Imm:$src2, globaladdress:$src3),
4024 "memw($src1+#$src2) = ##$src3",
4025 [(store (HexagonCONST32 tglobaladdr:$src3),
4026 (add IntRegs:$src1, u6_2ImmPred:$src2))]>,
4029 def : Pat<(i64 (ctlz (i64 DoubleRegs:$src1))),
4030 (i64 (A4_combineir (i32 0), (i32 (S2_cl0p DoubleRegs:$src1))))>,
4033 def : Pat<(i64 (cttz (i64 DoubleRegs:$src1))),
4034 (i64 (A4_combineir (i32 0), (i32 (S2_ct0p DoubleRegs:$src1))))>,
4037 // i8/i16/i32 -> i64 loads
4038 // We need a complexity of 120 here to override preceding handling of
4040 let AddedComplexity = 120 in {
4041 def: Loadam_pat<extloadi8, i64, addrga, Zext64, L4_loadrub_abs>;
4042 def: Loadam_pat<sextloadi8, i64, addrga, Sext64, L4_loadrb_abs>;
4043 def: Loadam_pat<zextloadi8, i64, addrga, Zext64, L4_loadrub_abs>;
4045 def: Loadam_pat<extloadi16, i64, addrga, Zext64, L4_loadruh_abs>;
4046 def: Loadam_pat<sextloadi16, i64, addrga, Sext64, L4_loadrh_abs>;
4047 def: Loadam_pat<zextloadi16, i64, addrga, Zext64, L4_loadruh_abs>;
4049 def: Loadam_pat<extloadi32, i64, addrga, Zext64, L4_loadri_abs>;
4050 def: Loadam_pat<sextloadi32, i64, addrga, Sext64, L4_loadri_abs>;
4051 def: Loadam_pat<zextloadi32, i64, addrga, Zext64, L4_loadri_abs>;
4054 // Indexed store double word - global address.
4055 // memw(Rs+#u6:2)=#S8
4056 let AddedComplexity = 10 in
4057 def STrih_offset_ext_V4 : STInst<(outs),
4058 (ins IntRegs:$src1, u6_1Imm:$src2, globaladdress:$src3),
4059 "memh($src1+#$src2) = ##$src3",
4060 [(truncstorei16 (HexagonCONST32 tglobaladdr:$src3),
4061 (add IntRegs:$src1, u6_1ImmPred:$src2))]>,
4063 // Map from store(globaladdress + x) -> memd(#foo + x)
4064 let AddedComplexity = 100 in
4065 def : Pat<(store (i64 DoubleRegs:$src1),
4066 FoldGlobalAddrGP:$addr),
4067 (S2_storerdabs FoldGlobalAddrGP:$addr, (i64 DoubleRegs:$src1))>,
4070 def : Pat<(atomic_store_64 FoldGlobalAddrGP:$addr,
4071 (i64 DoubleRegs:$src1)),
4072 (S2_storerdabs FoldGlobalAddrGP:$addr, (i64 DoubleRegs:$src1))>,
4075 // Map from store(globaladdress + x) -> memb(#foo + x)
4076 let AddedComplexity = 100 in
4077 def : Pat<(truncstorei8 (i32 IntRegs:$src1), FoldGlobalAddrGP:$addr),
4078 (S2_storerbabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>,
4081 def : Pat<(atomic_store_8 FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1)),
4082 (S2_storerbabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>,
4085 // Map from store(globaladdress + x) -> memh(#foo + x)
4086 let AddedComplexity = 100 in
4087 def : Pat<(truncstorei16 (i32 IntRegs:$src1), FoldGlobalAddrGP:$addr),
4088 (S2_storerhabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>,
4091 def : Pat<(atomic_store_16 FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1)),
4092 (S2_storerhabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>,
4095 // Map from store(globaladdress + x) -> memw(#foo + x)
4096 let AddedComplexity = 100 in
4097 def : Pat<(store (i32 IntRegs:$src1), FoldGlobalAddrGP:$addr),
4098 (S2_storeriabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>,
4101 def : Pat<(atomic_store_32 FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1)),
4102 (S2_storeriabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>,
4105 // Map from load(globaladdress + x) -> memd(#foo + x)
4106 let AddedComplexity = 100 in
4107 def : Pat<(i64 (load FoldGlobalAddrGP:$addr)),
4108 (i64 (L4_loadrd_abs FoldGlobalAddrGP:$addr))>,
4111 def : Pat<(atomic_load_64 FoldGlobalAddrGP:$addr),
4112 (i64 (L4_loadrd_abs FoldGlobalAddrGP:$addr))>,
4115 // Map from load(globaladdress + x) -> memb(#foo + x)
4116 let AddedComplexity = 100 in
4117 def : Pat<(i32 (extloadi8 FoldGlobalAddrGP:$addr)),
4118 (i32 (L4_loadrb_abs FoldGlobalAddrGP:$addr))>,
4121 // Map from load(globaladdress + x) -> memb(#foo + x)
4122 let AddedComplexity = 100 in
4123 def : Pat<(i32 (sextloadi8 FoldGlobalAddrGP:$addr)),
4124 (i32 (L4_loadrb_abs FoldGlobalAddrGP:$addr))>,
4127 //let AddedComplexity = 100 in
4128 let AddedComplexity = 100 in
4129 def : Pat<(i32 (extloadi16 FoldGlobalAddrGP:$addr)),
4130 (i32 (L4_loadrh_abs FoldGlobalAddrGP:$addr))>,
4133 // Map from load(globaladdress + x) -> memh(#foo + x)
4134 let AddedComplexity = 100 in
4135 def : Pat<(i32 (sextloadi16 FoldGlobalAddrGP:$addr)),
4136 (i32 (L4_loadrh_abs FoldGlobalAddrGP:$addr))>,
4139 // Map from load(globaladdress + x) -> memuh(#foo + x)
4140 let AddedComplexity = 100 in
4141 def : Pat<(i32 (zextloadi16 FoldGlobalAddrGP:$addr)),
4142 (i32 (L4_loadruh_abs FoldGlobalAddrGP:$addr))>,
4145 def : Pat<(atomic_load_16 FoldGlobalAddrGP:$addr),
4146 (i32 (L4_loadruh_abs FoldGlobalAddrGP:$addr))>,
4149 // Map from load(globaladdress + x) -> memub(#foo + x)
4150 let AddedComplexity = 100 in
4151 def : Pat<(i32 (zextloadi8 FoldGlobalAddrGP:$addr)),
4152 (i32 (L4_loadrub_abs FoldGlobalAddrGP:$addr))>,
4155 def : Pat<(atomic_load_8 FoldGlobalAddrGP:$addr),
4156 (i32 (L4_loadrub_abs FoldGlobalAddrGP:$addr))>,
4159 // Map from load(globaladdress + x) -> memw(#foo + x)
4160 let AddedComplexity = 100 in
4161 def : Pat<(i32 (load FoldGlobalAddrGP:$addr)),
4162 (i32 (L4_loadri_abs FoldGlobalAddrGP:$addr))>,
4165 def : Pat<(atomic_load_32 FoldGlobalAddrGP:$addr),
4166 (i32 (L4_loadri_abs FoldGlobalAddrGP:$addr))>,
4169 //===----------------------------------------------------------------------===//
4170 // :raw for of boundscheck:hi:lo insns
4171 //===----------------------------------------------------------------------===//
4173 // A4_boundscheck_lo: Detect if a register is within bounds.
4174 let hasSideEffects = 0, isCodeGenOnly = 0 in
4175 def A4_boundscheck_lo: ALU64Inst <
4176 (outs PredRegs:$Pd),
4177 (ins DoubleRegs:$Rss, DoubleRegs:$Rtt),
4178 "$Pd = boundscheck($Rss, $Rtt):raw:lo"> {
4183 let IClass = 0b1101;
4185 let Inst{27-23} = 0b00100;
4187 let Inst{7-5} = 0b100;
4189 let Inst{20-16} = Rss;
4190 let Inst{12-8} = Rtt;
4193 // A4_boundscheck_hi: Detect if a register is within bounds.
4194 let hasSideEffects = 0, isCodeGenOnly = 0 in
4195 def A4_boundscheck_hi: ALU64Inst <
4196 (outs PredRegs:$Pd),
4197 (ins DoubleRegs:$Rss, DoubleRegs:$Rtt),
4198 "$Pd = boundscheck($Rss, $Rtt):raw:hi"> {
4203 let IClass = 0b1101;
4205 let Inst{27-23} = 0b00100;
4207 let Inst{7-5} = 0b101;
4209 let Inst{20-16} = Rss;
4210 let Inst{12-8} = Rtt;
4213 let hasSideEffects = 0, isAsmParserOnly = 1 in
4214 def A4_boundscheck : MInst <
4215 (outs PredRegs:$Pd), (ins IntRegs:$Rs, DoubleRegs:$Rtt),
4216 "$Pd=boundscheck($Rs,$Rtt)">;
4218 // A4_tlbmatch: Detect if a VA/ASID matches a TLB entry.
4219 let isPredicateLate = 1, hasSideEffects = 0, isCodeGenOnly = 0 in
4220 def A4_tlbmatch : ALU64Inst<(outs PredRegs:$Pd),
4221 (ins DoubleRegs:$Rs, IntRegs:$Rt),
4222 "$Pd = tlbmatch($Rs, $Rt)",
4223 [], "", ALU64_tc_2early_SLOT23> {
4228 let IClass = 0b1101;
4229 let Inst{27-23} = 0b00100;
4230 let Inst{20-16} = Rs;
4232 let Inst{12-8} = Rt;
4233 let Inst{7-5} = 0b011;
4237 // We need custom lowering of ISD::PREFETCH into HexagonISD::DCFETCH
4238 // because the SDNode ISD::PREFETCH has properties MayLoad and MayStore.
4239 // We don't really want either one here.
4240 def SDTHexagonDCFETCH : SDTypeProfile<0, 2, [SDTCisPtrTy<0>,SDTCisInt<1>]>;
4241 def HexagonDCFETCH : SDNode<"HexagonISD::DCFETCH", SDTHexagonDCFETCH,
4244 // Use LD0Inst for dcfetch, but set "mayLoad" to 0 because this doesn't
4245 // really do a load.
4246 let hasSideEffects = 1, mayLoad = 0, isCodeGenOnly = 0 in
4247 def Y2_dcfetchbo : LD0Inst<(outs), (ins IntRegs:$Rs, u11_3Imm:$u11_3),
4248 "dcfetch($Rs + #$u11_3)",
4249 [(HexagonDCFETCH IntRegs:$Rs, u11_3ImmPred:$u11_3)],
4250 "", LD_tc_ld_SLOT0> {
4254 let IClass = 0b1001;
4255 let Inst{27-21} = 0b0100000;
4256 let Inst{20-16} = Rs;
4258 let Inst{10-0} = u11_3{13-3};
4261 //===----------------------------------------------------------------------===//
4262 // Compound instructions
4263 //===----------------------------------------------------------------------===//
4265 let isBranch = 1, hasSideEffects = 0, isExtentSigned = 1,
4266 isPredicated = 1, isPredicatedNew = 1, isExtendable = 1,
4267 opExtentBits = 11, opExtentAlign = 2, opExtendable = 1,
4268 isTerminator = 1, validSubTargets = HasV4SubT in
4269 class CJInst_tstbit_R0<string px, bit np, string tnt>
4270 : InstHexagon<(outs), (ins IntRegs:$Rs, brtarget:$r9_2),
4271 ""#px#" = tstbit($Rs, #0); if ("
4272 #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2",
4273 [], "", COMPOUND, TypeCOMPOUND> {
4278 let isPredicatedFalse = np;
4279 // tnt: Taken/Not Taken
4280 let isBrTaken = !if (!eq(tnt, "t"), "true", "false");
4281 let isTaken = !if (!eq(tnt, "t"), 1, 0);
4283 let IClass = 0b0001;
4284 let Inst{27-26} = 0b00;
4285 let Inst{25} = !if (!eq(px, "!p1"), 1,
4286 !if (!eq(px, "p1"), 1, 0));
4287 let Inst{24-23} = 0b11;
4289 let Inst{21-20} = r9_2{10-9};
4290 let Inst{19-16} = Rs;
4291 let Inst{13} = !if (!eq(tnt, "t"), 1, 0);
4292 let Inst{9-8} = 0b11;
4293 let Inst{7-1} = r9_2{8-2};
4296 let Defs = [PC, P0], Uses = [P0], isCodeGenOnly = 0 in {
4297 def J4_tstbit0_tp0_jump_nt : CJInst_tstbit_R0<"p0", 0, "nt">;
4298 def J4_tstbit0_tp0_jump_t : CJInst_tstbit_R0<"p0", 0, "t">;
4299 def J4_tstbit0_fp0_jump_nt : CJInst_tstbit_R0<"p0", 1, "nt">;
4300 def J4_tstbit0_fp0_jump_t : CJInst_tstbit_R0<"p0", 1, "t">;
4303 let Defs = [PC, P1], Uses = [P1], isCodeGenOnly = 0 in {
4304 def J4_tstbit0_tp1_jump_nt : CJInst_tstbit_R0<"p1", 0, "nt">;
4305 def J4_tstbit0_tp1_jump_t : CJInst_tstbit_R0<"p1", 0, "t">;
4306 def J4_tstbit0_fp1_jump_nt : CJInst_tstbit_R0<"p1", 1, "nt">;
4307 def J4_tstbit0_fp1_jump_t : CJInst_tstbit_R0<"p1", 1, "t">;
4311 let isBranch = 1, hasSideEffects = 0,
4312 isExtentSigned = 1, isPredicated = 1, isPredicatedNew = 1,
4313 isExtendable = 1, opExtentBits = 11, opExtentAlign = 2,
4314 opExtendable = 2, isTerminator = 1, validSubTargets = HasV4SubT in
4315 class CJInst_RR<string px, string op, bit np, string tnt>
4316 : InstHexagon<(outs), (ins IntRegs:$Rs, IntRegs:$Rt, brtarget:$r9_2),
4317 ""#px#" = cmp."#op#"($Rs, $Rt); if ("
4318 #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2",
4319 [], "", COMPOUND, TypeCOMPOUND> {
4325 let isPredicatedFalse = np;
4326 // tnt: Taken/Not Taken
4327 let isBrTaken = !if (!eq(tnt, "t"), "true", "false");
4328 let isTaken = !if (!eq(tnt, "t"), 1, 0);
4330 let IClass = 0b0001;
4331 let Inst{27-23} = !if (!eq(op, "eq"), 0b01000,
4332 !if (!eq(op, "gt"), 0b01001,
4333 !if (!eq(op, "gtu"), 0b01010, 0)));
4335 let Inst{21-20} = r9_2{10-9};
4336 let Inst{19-16} = Rs;
4337 let Inst{13} = !if (!eq(tnt, "t"), 1, 0);
4338 // px: Predicate reg 0/1
4339 let Inst{12} = !if (!eq(px, "!p1"), 1,
4340 !if (!eq(px, "p1"), 1, 0));
4341 let Inst{11-8} = Rt;
4342 let Inst{7-1} = r9_2{8-2};
4345 // P[10] taken/not taken.
4346 multiclass T_tnt_CJInst_RR<string op, bit np> {
4347 let Defs = [PC, P0], Uses = [P0] in {
4348 def NAME#p0_jump_nt : CJInst_RR<"p0", op, np, "nt">;
4349 def NAME#p0_jump_t : CJInst_RR<"p0", op, np, "t">;
4351 let Defs = [PC, P1], Uses = [P1] in {
4352 def NAME#p1_jump_nt : CJInst_RR<"p1", op, np, "nt">;
4353 def NAME#p1_jump_t : CJInst_RR<"p1", op, np, "t">;
4356 // Predicate / !Predicate
4357 multiclass T_pnp_CJInst_RR<string op>{
4358 defm J4_cmp#NAME#_t : T_tnt_CJInst_RR<op, 0>;
4359 defm J4_cmp#NAME#_f : T_tnt_CJInst_RR<op, 1>;
4361 // TypeCJ Instructions compare RR and jump
4362 let isCodeGenOnly = 0 in {
4363 defm eq : T_pnp_CJInst_RR<"eq">;
4364 defm gt : T_pnp_CJInst_RR<"gt">;
4365 defm gtu : T_pnp_CJInst_RR<"gtu">;
4368 let isBranch = 1, hasSideEffects = 0, isExtentSigned = 1,
4369 isPredicated = 1, isPredicatedNew = 1, isExtendable = 1, opExtentBits = 11,
4370 opExtentAlign = 2, opExtendable = 2, isTerminator = 1,
4371 validSubTargets = HasV4SubT in
4372 class CJInst_RU5<string px, string op, bit np, string tnt>
4373 : InstHexagon<(outs), (ins IntRegs:$Rs, u5Imm:$U5, brtarget:$r9_2),
4374 ""#px#" = cmp."#op#"($Rs, #$U5); if ("
4375 #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2",
4376 [], "", COMPOUND, TypeCOMPOUND> {
4382 let isPredicatedFalse = np;
4383 // tnt: Taken/Not Taken
4384 let isBrTaken = !if (!eq(tnt, "t"), "true", "false");
4385 let isTaken = !if (!eq(tnt, "t"), 1, 0);
4387 let IClass = 0b0001;
4388 let Inst{27-26} = 0b00;
4389 // px: Predicate reg 0/1
4390 let Inst{25} = !if (!eq(px, "!p1"), 1,
4391 !if (!eq(px, "p1"), 1, 0));
4392 let Inst{24-23} = !if (!eq(op, "eq"), 0b00,
4393 !if (!eq(op, "gt"), 0b01,
4394 !if (!eq(op, "gtu"), 0b10, 0)));
4396 let Inst{21-20} = r9_2{10-9};
4397 let Inst{19-16} = Rs;
4398 let Inst{13} = !if (!eq(tnt, "t"), 1, 0);
4399 let Inst{12-8} = U5;
4400 let Inst{7-1} = r9_2{8-2};
4402 // P[10] taken/not taken.
4403 multiclass T_tnt_CJInst_RU5<string op, bit np> {
4404 let Defs = [PC, P0], Uses = [P0] in {
4405 def NAME#p0_jump_nt : CJInst_RU5<"p0", op, np, "nt">;
4406 def NAME#p0_jump_t : CJInst_RU5<"p0", op, np, "t">;
4408 let Defs = [PC, P1], Uses = [P1] in {
4409 def NAME#p1_jump_nt : CJInst_RU5<"p1", op, np, "nt">;
4410 def NAME#p1_jump_t : CJInst_RU5<"p1", op, np, "t">;
4413 // Predicate / !Predicate
4414 multiclass T_pnp_CJInst_RU5<string op>{
4415 defm J4_cmp#NAME#i_t : T_tnt_CJInst_RU5<op, 0>;
4416 defm J4_cmp#NAME#i_f : T_tnt_CJInst_RU5<op, 1>;
4418 // TypeCJ Instructions compare RI and jump
4419 let isCodeGenOnly = 0 in {
4420 defm eq : T_pnp_CJInst_RU5<"eq">;
4421 defm gt : T_pnp_CJInst_RU5<"gt">;
4422 defm gtu : T_pnp_CJInst_RU5<"gtu">;
4425 let isBranch = 1, hasSideEffects = 0, isExtentSigned = 1,
4426 isPredicated = 1, isPredicatedFalse = 1, isPredicatedNew = 1,
4427 isExtendable = 1, opExtentBits = 11, opExtentAlign = 2, opExtendable = 1,
4428 isTerminator = 1, validSubTargets = HasV4SubT in
4429 class CJInst_Rn1<string px, string op, bit np, string tnt>
4430 : InstHexagon<(outs), (ins IntRegs:$Rs, brtarget:$r9_2),
4431 ""#px#" = cmp."#op#"($Rs,#-1); if ("
4432 #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2",
4433 [], "", COMPOUND, TypeCOMPOUND> {
4438 let isPredicatedFalse = np;
4439 // tnt: Taken/Not Taken
4440 let isBrTaken = !if (!eq(tnt, "t"), "true", "false");
4441 let isTaken = !if (!eq(tnt, "t"), 1, 0);
4443 let IClass = 0b0001;
4444 let Inst{27-26} = 0b00;
4445 let Inst{25} = !if (!eq(px, "!p1"), 1,
4446 !if (!eq(px, "p1"), 1, 0));
4448 let Inst{24-23} = 0b11;
4450 let Inst{21-20} = r9_2{10-9};
4451 let Inst{19-16} = Rs;
4452 let Inst{13} = !if (!eq(tnt, "t"), 1, 0);
4453 let Inst{9-8} = !if (!eq(op, "eq"), 0b00,
4454 !if (!eq(op, "gt"), 0b01, 0));
4455 let Inst{7-1} = r9_2{8-2};
4458 // P[10] taken/not taken.
4459 multiclass T_tnt_CJInst_Rn1<string op, bit np> {
4460 let Defs = [PC, P0], Uses = [P0] in {
4461 def NAME#p0_jump_nt : CJInst_Rn1<"p0", op, np, "nt">;
4462 def NAME#p0_jump_t : CJInst_Rn1<"p0", op, np, "t">;
4464 let Defs = [PC, P1], Uses = [P1] in {
4465 def NAME#p1_jump_nt : CJInst_Rn1<"p1", op, np, "nt">;
4466 def NAME#p1_jump_t : CJInst_Rn1<"p1", op, np, "t">;
4469 // Predicate / !Predicate
4470 multiclass T_pnp_CJInst_Rn1<string op>{
4471 defm J4_cmp#NAME#n1_t : T_tnt_CJInst_Rn1<op, 0>;
4472 defm J4_cmp#NAME#n1_f : T_tnt_CJInst_Rn1<op, 1>;
4474 // TypeCJ Instructions compare -1 and jump
4475 let isCodeGenOnly = 0 in {
4476 defm eq : T_pnp_CJInst_Rn1<"eq">;
4477 defm gt : T_pnp_CJInst_Rn1<"gt">;
4480 // J4_jumpseti: Direct unconditional jump and set register to immediate.
4481 let Defs = [PC], isBranch = 1, hasSideEffects = 0, hasNewValue = 1,
4482 isExtentSigned = 1, opNewValue = 0, isExtendable = 1, opExtentBits = 11,
4483 opExtentAlign = 2, opExtendable = 2, validSubTargets = HasV4SubT,
4484 isCodeGenOnly = 0 in
4485 def J4_jumpseti: CJInst <
4487 (ins u6Imm:$U6, brtarget:$r9_2),
4488 "$Rd = #$U6 ; jump $r9_2"> {
4493 let IClass = 0b0001;
4494 let Inst{27-24} = 0b0110;
4495 let Inst{21-20} = r9_2{10-9};
4496 let Inst{19-16} = Rd;
4497 let Inst{13-8} = U6;
4498 let Inst{7-1} = r9_2{8-2};
4501 // J4_jumpsetr: Direct unconditional jump and transfer register.
4502 let Defs = [PC], isBranch = 1, hasSideEffects = 0, hasNewValue = 1,
4503 isExtentSigned = 1, opNewValue = 0, isExtendable = 1, opExtentBits = 11,
4504 opExtentAlign = 2, opExtendable = 2, validSubTargets = HasV4SubT,
4505 isCodeGenOnly = 0 in
4506 def J4_jumpsetr: CJInst <
4508 (ins IntRegs:$Rs, brtarget:$r9_2),
4509 "$Rd = $Rs ; jump $r9_2"> {
4514 let IClass = 0b0001;
4515 let Inst{27-24} = 0b0111;
4516 let Inst{21-20} = r9_2{10-9};
4517 let Inst{11-8} = Rd;
4518 let Inst{19-16} = Rs;
4519 let Inst{7-1} = r9_2{8-2};