1 // Pattern fragment that combines the value type and the register class
2 // into a single parameter.
3 // The pat frags in the definitions below need to have a named register,
4 // otherwise i32 will be assumed regardless of the register class. The
5 // name of the register does not matter.
6 def I1 : PatLeaf<(i1 PredRegs:$R)>;
7 def I32 : PatLeaf<(i32 IntRegs:$R)>;
8 def I64 : PatLeaf<(i64 DoubleRegs:$R)>;
9 def F32 : PatLeaf<(f32 IntRegs:$R)>;
10 def F64 : PatLeaf<(f64 DoubleRegs:$R)>;
12 // Pattern fragments to extract the low and high subregisters from a
14 def LoReg: OutPatFrag<(ops node:$Rs), (EXTRACT_SUBREG (i64 $Rs), isub_lo)>;
15 def HiReg: OutPatFrag<(ops node:$Rs), (EXTRACT_SUBREG (i64 $Rs), isub_hi)>;
17 def IsOrAdd: PatFrag<(ops node:$Addr, node:$off),
18 (or node:$Addr, node:$off), [{ return isOrEquivalentToAdd(N); }]>;
20 def Iss4_6 : PatLeaf<(i32 imm), [{
21 int32_t V = N->getSExtValue();
22 return isShiftedInt<4,6>(V);
25 def Iss4_7 : PatLeaf<(i32 imm), [{
26 int32_t V = N->getSExtValue();
27 return isShiftedInt<4,7>(V);
30 def IsPow2_32 : PatLeaf<(i32 imm), [{
31 uint32_t V = N->getZExtValue();
32 return isPowerOf2_32(V);
35 def IsPow2_64 : PatLeaf<(i64 imm), [{
36 uint64_t V = N->getZExtValue();
37 return isPowerOf2_64(V);
40 def IsNPow2_32 : PatLeaf<(i32 imm), [{
41 uint32_t NV = ~N->getZExtValue();
42 return isPowerOf2_32(NV);
45 def IsPow2_64L : PatLeaf<(i64 imm), [{
46 uint64_t V = N->getZExtValue();
47 return isPowerOf2_64(V) && Log2_64(V) < 32;
50 def IsPow2_64H : PatLeaf<(i64 imm), [{
51 uint64_t V = N->getZExtValue();
52 return isPowerOf2_64(V) && Log2_64(V) >= 32;
55 def IsNPow2_64L : PatLeaf<(i64 imm), [{
56 uint64_t NV = ~N->getZExtValue();
57 return isPowerOf2_64(NV) && Log2_64(NV) < 32;
60 def IsNPow2_64H : PatLeaf<(i64 imm), [{
61 uint64_t NV = ~N->getZExtValue();
62 return isPowerOf2_64(NV) && Log2_64(NV) >= 32;
65 def SDEC1 : SDNodeXForm<imm, [{
66 int32_t V = N->getSExtValue();
67 return CurDAG->getTargetConstant(V-1, SDLoc(N), MVT::i32);
70 def UDEC1 : SDNodeXForm<imm, [{
71 uint32_t V = N->getZExtValue();
73 return CurDAG->getTargetConstant(V-1, SDLoc(N), MVT::i32);
76 def UDEC32 : SDNodeXForm<imm, [{
77 uint32_t V = N->getZExtValue();
79 return CurDAG->getTargetConstant(V-32, SDLoc(N), MVT::i32);
82 def Log2_32 : SDNodeXForm<imm, [{
83 uint32_t V = N->getZExtValue();
84 return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
87 def Log2_64 : SDNodeXForm<imm, [{
88 uint64_t V = N->getZExtValue();
89 return CurDAG->getTargetConstant(Log2_64(V), SDLoc(N), MVT::i32);
92 def LogN2_32 : SDNodeXForm<imm, [{
93 uint32_t NV = ~N->getZExtValue();
94 return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
97 def LogN2_64 : SDNodeXForm<imm, [{
98 uint64_t NV = ~N->getZExtValue();
99 return CurDAG->getTargetConstant(Log2_64(NV), SDLoc(N), MVT::i32);
102 def ToZext64: OutPatFrag<(ops node:$Rs),
103 (i64 (A4_combineir 0, (i32 $Rs)))>;
104 def ToSext64: OutPatFrag<(ops node:$Rs),
105 (i64 (A2_sxtw (i32 $Rs)))>;
108 class T_CMP_pat <InstHexagon MI, PatFrag OpNode, PatLeaf ImmPred>
109 : Pat<(i1 (OpNode I32:$src1, ImmPred:$src2)),
110 (MI IntRegs:$src1, ImmPred:$src2)>;
112 def : T_CMP_pat <C2_cmpeqi, seteq, s10_0ImmPred>;
113 def : T_CMP_pat <C2_cmpgti, setgt, s10_0ImmPred>;
114 def : T_CMP_pat <C2_cmpgtui, setugt, u9_0ImmPred>;
116 def SDTHexagonI64I32I32 : SDTypeProfile<1, 2,
117 [SDTCisVT<0, i64>, SDTCisVT<1, i32>, SDTCisSameAs<1, 2>]>;
119 def HexagonCOMBINE : SDNode<"HexagonISD::COMBINE", SDTHexagonI64I32I32>;
120 def HexagonPACKHL : SDNode<"HexagonISD::PACKHL", SDTHexagonI64I32I32>;
122 // Pats for instruction selection.
123 class BinOp32_pat<SDNode Op, InstHexagon MI, ValueType ResT>
124 : Pat<(ResT (Op I32:$Rs, I32:$Rt)),
125 (ResT (MI IntRegs:$Rs, IntRegs:$Rt))>;
127 def: BinOp32_pat<add, A2_add, i32>;
128 def: BinOp32_pat<and, A2_and, i32>;
129 def: BinOp32_pat<or, A2_or, i32>;
130 def: BinOp32_pat<sub, A2_sub, i32>;
131 def: BinOp32_pat<xor, A2_xor, i32>;
133 def: BinOp32_pat<HexagonCOMBINE, A2_combinew, i64>;
134 def: BinOp32_pat<HexagonPACKHL, S2_packhl, i64>;
136 // Patfrag to convert the usual comparison patfrags (e.g. setlt) to ones
137 // that reverse the order of the operands.
138 class RevCmp<PatFrag F> : PatFrag<(ops node:$rhs, node:$lhs), F.Fragment>;
140 // Pats for compares. They use PatFrags as operands, not SDNodes,
141 // since seteq/setgt/etc. are defined as ParFrags.
142 class T_cmp32_rr_pat<InstHexagon MI, PatFrag Op, ValueType VT>
143 : Pat<(VT (Op I32:$Rs, I32:$Rt)),
144 (MI IntRegs:$Rs, IntRegs:$Rt)>;
146 def: T_cmp32_rr_pat<C2_cmpeq, seteq, i1>;
147 def: T_cmp32_rr_pat<C2_cmpgt, setgt, i1>;
148 def: T_cmp32_rr_pat<C2_cmpgtu, setugt, i1>;
150 def: T_cmp32_rr_pat<C2_cmpgt, RevCmp<setlt>, i1>;
151 def: T_cmp32_rr_pat<C2_cmpgtu, RevCmp<setult>, i1>;
153 def: Pat<(select I1:$Pu, I32:$Rs, I32:$Rt),
154 (C2_mux PredRegs:$Pu, IntRegs:$Rs, IntRegs:$Rt)>;
156 def: Pat<(add I32:$Rs, s32_0ImmPred:$s16),
157 (A2_addi I32:$Rs, imm:$s16)>;
159 def: Pat<(or I32:$Rs, s32_0ImmPred:$s10),
160 (A2_orir IntRegs:$Rs, imm:$s10)>;
161 def: Pat<(and I32:$Rs, s32_0ImmPred:$s10),
162 (A2_andir IntRegs:$Rs, imm:$s10)>;
164 def: Pat<(sub s32_0ImmPred:$s10, IntRegs:$Rs),
165 (A2_subri imm:$s10, IntRegs:$Rs)>;
167 // Rd = not(Rs) gets mapped to Rd=sub(#-1, Rs).
168 def: Pat<(not I32:$src1),
169 (A2_subri -1, IntRegs:$src1)>;
171 def TruncI64ToI32: SDNodeXForm<imm, [{
172 return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i32);
175 def: Pat<(s32_0ImmPred:$s16), (A2_tfrsi imm:$s16)>;
176 def: Pat<(s8_0Imm64Pred:$s8), (A2_tfrpi (TruncI64ToI32 $s8))>;
178 def : Pat<(select I1:$Pu, s32_0ImmPred:$s8, I32:$Rs),
179 (C2_muxri I1:$Pu, imm:$s8, I32:$Rs)>;
181 def : Pat<(select I1:$Pu, I32:$Rs, s32_0ImmPred:$s8),
182 (C2_muxir I1:$Pu, I32:$Rs, imm:$s8)>;
184 def : Pat<(select I1:$Pu, s32_0ImmPred:$s8, s8_0ImmPred:$S8),
185 (C2_muxii I1:$Pu, imm:$s8, imm:$S8)>;
187 def: Pat<(shl I32:$src1, (i32 16)), (A2_aslh I32:$src1)>;
188 def: Pat<(sra I32:$src1, (i32 16)), (A2_asrh I32:$src1)>;
189 def: Pat<(sext_inreg I32:$src1, i8), (A2_sxtb I32:$src1)>;
190 def: Pat<(sext_inreg I32:$src1, i16), (A2_sxth I32:$src1)>;
192 class T_vcmp_pat<InstHexagon MI, PatFrag Op, ValueType T>
193 : Pat<(i1 (Op (T DoubleRegs:$Rss), (T DoubleRegs:$Rtt))),
194 (i1 (MI DoubleRegs:$Rss, DoubleRegs:$Rtt))>;
196 def: T_vcmp_pat<A2_vcmpbeq, seteq, v8i8>;
197 def: T_vcmp_pat<A2_vcmpbgtu, setugt, v8i8>;
198 def: T_vcmp_pat<A2_vcmpheq, seteq, v4i16>;
199 def: T_vcmp_pat<A2_vcmphgt, setgt, v4i16>;
200 def: T_vcmp_pat<A2_vcmphgtu, setugt, v4i16>;
201 def: T_vcmp_pat<A2_vcmpweq, seteq, v2i32>;
202 def: T_vcmp_pat<A2_vcmpwgt, setgt, v2i32>;
203 def: T_vcmp_pat<A2_vcmpwgtu, setugt, v2i32>;
206 def: Pat<(sext_inreg (add I32:$src1, I32:$src2), i16),
207 (A2_addh_l16_ll I32:$src1, I32:$src2)>;
209 def: Pat<(sra (add (shl I32:$src1, (i32 16)), I32:$src2), (i32 16)),
210 (A2_addh_l16_hl I32:$src1, I32:$src2)>;
212 def: Pat<(shl (add I32:$src1, I32:$src2), (i32 16)),
213 (A2_addh_h16_ll I32:$src1, I32:$src2)>;
215 // Subtract halfword.
216 def: Pat<(sext_inreg (sub I32:$src1, I32:$src2), i16),
217 (A2_subh_l16_ll I32:$src1, I32:$src2)>;
219 def: Pat<(shl (sub I32:$src1, I32:$src2), (i32 16)),
220 (A2_subh_h16_ll I32:$src1, I32:$src2)>;
222 // Here, depending on the operand being selected, we'll either generate a
223 // min or max instruction.
225 // (a>b)?a:b --> max(a,b) => Here check performed is '>' and the value selected
226 // is the larger of two. So, the corresponding HexagonInst is passed in 'Inst'.
227 // (a>b)?b:a --> min(a,b) => Here check performed is '>' but the smaller value
228 // is selected and the corresponding HexagonInst is passed in 'SwapInst'.
230 multiclass T_MinMax_pats <PatFrag Op, PatLeaf Val,
231 InstHexagon Inst, InstHexagon SwapInst> {
232 def: Pat<(select (i1 (Op Val:$src1, Val:$src2)), Val:$src1, Val:$src2),
233 (Inst Val:$src1, Val:$src2)>;
234 def: Pat<(select (i1 (Op Val:$src1, Val:$src2)), Val:$src2, Val:$src1),
235 (SwapInst Val:$src1, Val:$src2)>;
238 def IsPosHalf : PatLeaf<(i32 IntRegs:$a), [{
239 return isPositiveHalfWord(N);
242 multiclass MinMax_pats <PatFrag Op, InstHexagon Inst, InstHexagon SwapInst> {
243 defm: T_MinMax_pats<Op, I32, Inst, SwapInst>;
245 def: Pat<(sext_inreg (select (i1 (Op IsPosHalf:$src1, IsPosHalf:$src2)),
246 IsPosHalf:$src1, IsPosHalf:$src2),
248 (Inst IntRegs:$src1, IntRegs:$src2)>;
250 def: Pat<(sext_inreg (select (i1 (Op IsPosHalf:$src1, IsPosHalf:$src2)),
251 IsPosHalf:$src2, IsPosHalf:$src1),
253 (SwapInst IntRegs:$src1, IntRegs:$src2)>;
256 let AddedComplexity = 200 in {
257 defm: MinMax_pats<setge, A2_max, A2_min>;
258 defm: MinMax_pats<setgt, A2_max, A2_min>;
259 defm: MinMax_pats<setle, A2_min, A2_max>;
260 defm: MinMax_pats<setlt, A2_min, A2_max>;
261 defm: MinMax_pats<setuge, A2_maxu, A2_minu>;
262 defm: MinMax_pats<setugt, A2_maxu, A2_minu>;
263 defm: MinMax_pats<setule, A2_minu, A2_maxu>;
264 defm: MinMax_pats<setult, A2_minu, A2_maxu>;
267 class T_cmp64_rr_pat<InstHexagon MI, PatFrag CmpOp>
268 : Pat<(i1 (CmpOp I64:$Rs, I64:$Rt)),
269 (i1 (MI DoubleRegs:$Rs, DoubleRegs:$Rt))>;
271 def: T_cmp64_rr_pat<C2_cmpeqp, seteq>;
272 def: T_cmp64_rr_pat<C2_cmpgtp, setgt>;
273 def: T_cmp64_rr_pat<C2_cmpgtup, setugt>;
274 def: T_cmp64_rr_pat<C2_cmpgtp, RevCmp<setlt>>;
275 def: T_cmp64_rr_pat<C2_cmpgtup, RevCmp<setult>>;
277 def: Pat<(i64 (add I64:$Rs, I64:$Rt)), (A2_addp I64:$Rs, I64:$Rt)>;
278 def: Pat<(i64 (sub I64:$Rs, I64:$Rt)), (A2_subp I64:$Rs, I64:$Rt)>;
280 def: Pat<(i64 (and I64:$Rs, I64:$Rt)), (A2_andp I64:$Rs, I64:$Rt)>;
281 def: Pat<(i64 (or I64:$Rs, I64:$Rt)), (A2_orp I64:$Rs, I64:$Rt)>;
282 def: Pat<(i64 (xor I64:$Rs, I64:$Rt)), (A2_xorp I64:$Rs, I64:$Rt)>;
284 def: Pat<(i1 (not I1:$Ps)), (C2_not PredRegs:$Ps)>;
286 def: Pat<(i1 (and I1:$Ps, I1:$Pt)), (C2_and I1:$Ps, I1:$Pt)>;
287 def: Pat<(i1 (or I1:$Ps, I1:$Pt)), (C2_or I1:$Ps, I1:$Pt)>;
288 def: Pat<(i1 (xor I1:$Ps, I1:$Pt)), (C2_xor I1:$Ps, I1:$Pt)>;
289 def: Pat<(i1 (and I1:$Ps, (not I1:$Pt))), (C2_andn I1:$Ps, I1:$Pt)>;
290 def: Pat<(i1 (or I1:$Ps, (not I1:$Pt))), (C2_orn I1:$Ps, I1:$Pt)>;
292 def retflag : SDNode<"HexagonISD::RET_FLAG", SDTNone,
293 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
294 def eh_return: SDNode<"HexagonISD::EH_RETURN", SDTNone, [SDNPHasChain]>;
296 def: Pat<(br bb:$dst), (J2_jump b30_2Imm:$dst)>;
297 def: Pat<(brcond I1:$src1, bb:$block), (J2_jumpt PredRegs:$src1, bb:$block)>;
298 def: Pat<(brind I32:$dst), (J2_jumpr IntRegs:$dst)>;
300 def: Pat<(retflag), (PS_jmpret (i32 R31))>;
301 def: Pat<(eh_return), (EH_RETURN_JMPR (i32 R31))>;
303 // Patterns to select load-indexed (i.e. load from base+offset).
304 multiclass Loadx_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
306 def: Pat<(VT (Load AddrFI:$fi)), (VT (MI AddrFI:$fi, 0))>;
307 def: Pat<(VT (Load (add (i32 AddrFI:$fi), ImmPred:$Off))),
308 (VT (MI AddrFI:$fi, imm:$Off))>;
309 def: Pat<(VT (Load (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off))),
310 (VT (MI AddrFI:$fi, imm:$Off))>;
311 def: Pat<(VT (Load (add I32:$Rs, ImmPred:$Off))),
312 (VT (MI IntRegs:$Rs, imm:$Off))>;
313 def: Pat<(VT (Load I32:$Rs)), (VT (MI IntRegs:$Rs, 0))>;
316 let AddedComplexity = 20 in {
317 defm: Loadx_pat<load, i32, s30_2ImmPred, L2_loadri_io>;
318 defm: Loadx_pat<load, i64, s29_3ImmPred, L2_loadrd_io>;
319 defm: Loadx_pat<atomic_load_8 , i32, s32_0ImmPred, L2_loadrub_io>;
320 defm: Loadx_pat<atomic_load_16, i32, s31_1ImmPred, L2_loadruh_io>;
321 defm: Loadx_pat<atomic_load_32, i32, s30_2ImmPred, L2_loadri_io>;
322 defm: Loadx_pat<atomic_load_64, i64, s29_3ImmPred, L2_loadrd_io>;
324 defm: Loadx_pat<extloadi1, i32, s32_0ImmPred, L2_loadrub_io>;
325 defm: Loadx_pat<extloadi8, i32, s32_0ImmPred, L2_loadrub_io>;
326 defm: Loadx_pat<extloadi16, i32, s31_1ImmPred, L2_loadruh_io>;
327 defm: Loadx_pat<sextloadi8, i32, s32_0ImmPred, L2_loadrb_io>;
328 defm: Loadx_pat<sextloadi16, i32, s31_1ImmPred, L2_loadrh_io>;
329 defm: Loadx_pat<zextloadi1, i32, s32_0ImmPred, L2_loadrub_io>;
330 defm: Loadx_pat<zextloadi8, i32, s32_0ImmPred, L2_loadrub_io>;
331 defm: Loadx_pat<zextloadi16, i32, s31_1ImmPred, L2_loadruh_io>;
335 // Sign-extending loads of i1 need to replicate the lowest bit throughout
336 // the 32-bit value. Since the loaded value can only be 0 or 1, 0-v should
338 let AddedComplexity = 20 in
339 def: Pat<(i32 (sextloadi1 I32:$Rs)),
340 (A2_subri 0, (L2_loadrub_io IntRegs:$Rs, 0))>;
342 def: Pat<(i32 (mul I32:$src1, I32:$src2)), (M2_mpyi I32:$src1, I32:$src2)>;
343 def: Pat<(i32 (mulhs I32:$src1, I32:$src2)), (M2_mpy_up I32:$src1, I32:$src2)>;
344 def: Pat<(i32 (mulhu I32:$src1, I32:$src2)), (M2_mpyu_up I32:$src1, I32:$src2)>;
346 def: Pat<(mul IntRegs:$Rs, u32_0ImmPred:$u8),
347 (M2_mpysip IntRegs:$Rs, imm:$u8)>;
348 def: Pat<(ineg (mul IntRegs:$Rs, u8_0ImmPred:$u8)),
349 (M2_mpysin IntRegs:$Rs, imm:$u8)>;
350 def: Pat<(mul IntRegs:$src1, s32_0ImmPred:$src2),
351 (M2_mpysmi IntRegs:$src1, imm:$src2)>;
352 def: Pat<(add (mul IntRegs:$src2, u32_0ImmPred:$src3), IntRegs:$src1),
353 (M2_macsip IntRegs:$src1, IntRegs:$src2, imm:$src3)>;
354 def: Pat<(add (mul I32:$src2, I32:$src3), I32:$src1),
355 (M2_maci IntRegs:$src1, IntRegs:$src2, IntRegs:$src3)>;
356 def: Pat<(add (add IntRegs:$src2, u32_0ImmPred:$src3), IntRegs:$src1),
357 (M2_accii IntRegs:$src1, IntRegs:$src2, imm:$src3)>;
358 def: Pat<(add (add I32:$src2, I32:$src3), I32:$src1),
359 (M2_acci IntRegs:$src1, IntRegs:$src2, IntRegs:$src3)>;
361 class T_MType_acc_pat1 <InstHexagon MI, SDNode firstOp, SDNode secOp,
363 : Pat <(secOp IntRegs:$src1, (firstOp IntRegs:$src2, ImmPred:$src3)),
364 (MI IntRegs:$src1, IntRegs:$src2, ImmPred:$src3)>;
366 class T_MType_acc_pat2 <InstHexagon MI, SDNode firstOp, SDNode secOp>
367 : Pat <(i32 (secOp IntRegs:$src1, (firstOp IntRegs:$src2, IntRegs:$src3))),
368 (MI IntRegs:$src1, IntRegs:$src2, IntRegs:$src3)>;
370 def : T_MType_acc_pat2 <M2_xor_xacc, xor, xor>;
371 def : T_MType_acc_pat1 <M2_macsin, mul, sub, u32_0ImmPred>;
373 def : T_MType_acc_pat1 <M2_naccii, add, sub, s32_0ImmPred>;
374 def : T_MType_acc_pat2 <M2_nacci, add, sub>;
376 def: T_MType_acc_pat2 <M4_or_xor, xor, or>;
377 def: T_MType_acc_pat2 <M4_and_xor, xor, and>;
378 def: T_MType_acc_pat2 <M4_or_and, and, or>;
379 def: T_MType_acc_pat2 <M4_and_and, and, and>;
380 def: T_MType_acc_pat2 <M4_xor_and, and, xor>;
381 def: T_MType_acc_pat2 <M4_or_or, or, or>;
382 def: T_MType_acc_pat2 <M4_and_or, or, and>;
383 def: T_MType_acc_pat2 <M4_xor_or, or, xor>;
385 class T_MType_acc_pat3 <InstHexagon MI, SDNode firstOp, SDNode secOp>
386 : Pat <(secOp I32:$src1, (firstOp I32:$src2, (not I32:$src3))),
387 (MI IntRegs:$src1, IntRegs:$src2, IntRegs:$src3)>;
389 def: T_MType_acc_pat3 <M4_or_andn, and, or>;
390 def: T_MType_acc_pat3 <M4_and_andn, and, and>;
391 def: T_MType_acc_pat3 <M4_xor_andn, and, xor>;
393 def Aext64: PatFrag<(ops node:$Rs), (i64 (anyext node:$Rs))>;
394 def Sext64: PatFrag<(ops node:$Rs), (i64 (sext node:$Rs))>;
395 def Zext64: PatFrag<(ops node:$Rs), (i64 (zext node:$Rs))>;
397 // Return true if for a 32 to 64-bit sign-extended load.
398 def Sext64Ld : PatLeaf<(i64 DoubleRegs:$src1), [{
399 LoadSDNode *LD = dyn_cast<LoadSDNode>(N);
402 return LD->getExtensionType() == ISD::SEXTLOAD &&
403 LD->getMemoryVT().getScalarType() == MVT::i32;
406 def: Pat<(mul (Aext64 I32:$src1), (Aext64 I32:$src2)),
407 (M2_dpmpyuu_s0 IntRegs:$src1, IntRegs:$src2)>;
409 def: Pat<(mul (Sext64 I32:$src1), (Sext64 I32:$src2)),
410 (M2_dpmpyss_s0 IntRegs:$src1, IntRegs:$src2)>;
412 def: Pat<(mul Sext64Ld:$src1, Sext64Ld:$src2),
413 (M2_dpmpyss_s0 (LoReg DoubleRegs:$src1), (LoReg DoubleRegs:$src2))>;
415 // Multiply and accumulate, use full result.
416 // Rxx[+-]=mpy(Rs,Rt)
418 def: Pat<(add I64:$src1, (mul (Sext64 I32:$src2), (Sext64 I32:$src3))),
419 (M2_dpmpyss_acc_s0 DoubleRegs:$src1, IntRegs:$src2, IntRegs:$src3)>;
421 def: Pat<(sub I64:$src1, (mul (Sext64 I32:$src2), (Sext64 I32:$src3))),
422 (M2_dpmpyss_nac_s0 DoubleRegs:$src1, IntRegs:$src2, IntRegs:$src3)>;
424 def: Pat<(add I64:$src1, (mul (Aext64 I32:$src2), (Aext64 I32:$src3))),
425 (M2_dpmpyuu_acc_s0 DoubleRegs:$src1, IntRegs:$src2, IntRegs:$src3)>;
427 def: Pat<(add I64:$src1, (mul (Zext64 I32:$src2), (Zext64 I32:$src3))),
428 (M2_dpmpyuu_acc_s0 DoubleRegs:$src1, IntRegs:$src2, IntRegs:$src3)>;
430 def: Pat<(sub I64:$src1, (mul (Aext64 I32:$src2), (Aext64 I32:$src3))),
431 (M2_dpmpyuu_nac_s0 DoubleRegs:$src1, IntRegs:$src2, IntRegs:$src3)>;
433 def: Pat<(sub I64:$src1, (mul (Zext64 I32:$src2), (Zext64 I32:$src3))),
434 (M2_dpmpyuu_nac_s0 DoubleRegs:$src1, IntRegs:$src2, IntRegs:$src3)>;
436 class Storepi_pat<PatFrag Store, PatFrag Value, PatFrag Offset,
438 : Pat<(Store Value:$src1, I32:$src2, Offset:$offset),
439 (MI I32:$src2, imm:$offset, Value:$src1)>;
441 def: Storepi_pat<post_truncsti8, I32, s4_0ImmPred, S2_storerb_pi>;
442 def: Storepi_pat<post_truncsti16, I32, s4_1ImmPred, S2_storerh_pi>;
443 def: Storepi_pat<post_store, I32, s4_2ImmPred, S2_storeri_pi>;
444 def: Storepi_pat<post_store, I64, s4_3ImmPred, S2_storerd_pi>;
446 // Patterns for generating stores, where the address takes different forms:
448 // - frameindex + offset,
450 // - simple (base address without offset).
451 // These would usually be used together (via Storex_pat defined below), but
452 // in some cases one may want to apply different properties (such as
453 // AddedComplexity) to the individual patterns.
454 class Storex_fi_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
455 : Pat<(Store Value:$Rs, AddrFI:$fi), (MI AddrFI:$fi, 0, Value:$Rs)>;
456 multiclass Storex_fi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
458 def: Pat<(Store Value:$Rs, (add (i32 AddrFI:$fi), ImmPred:$Off)),
459 (MI AddrFI:$fi, imm:$Off, Value:$Rs)>;
460 def: Pat<(Store Value:$Rs, (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off)),
461 (MI AddrFI:$fi, imm:$Off, Value:$Rs)>;
463 multiclass Storex_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
465 def: Pat<(Store Value:$Rt, (add I32:$Rs, ImmPred:$Off)),
466 (MI IntRegs:$Rs, imm:$Off, Value:$Rt)>;
467 def: Pat<(Store Value:$Rt, (IsOrAdd I32:$Rs, ImmPred:$Off)),
468 (MI IntRegs:$Rs, imm:$Off, Value:$Rt)>;
470 class Storex_simple_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
471 : Pat<(Store Value:$Rt, I32:$Rs),
472 (MI IntRegs:$Rs, 0, Value:$Rt)>;
474 // Patterns for generating stores, where the address takes different forms,
475 // and where the value being stored is transformed through the value modifier
476 // ValueMod. The address forms are same as above.
477 class Storexm_fi_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
479 : Pat<(Store Value:$Rs, AddrFI:$fi),
480 (MI AddrFI:$fi, 0, (ValueMod Value:$Rs))>;
481 multiclass Storexm_fi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
482 PatFrag ValueMod, InstHexagon MI> {
483 def: Pat<(Store Value:$Rs, (add (i32 AddrFI:$fi), ImmPred:$Off)),
484 (MI AddrFI:$fi, imm:$Off, (ValueMod Value:$Rs))>;
485 def: Pat<(Store Value:$Rs, (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off)),
486 (MI AddrFI:$fi, imm:$Off, (ValueMod Value:$Rs))>;
488 multiclass Storexm_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
489 PatFrag ValueMod, InstHexagon MI> {
490 def: Pat<(Store Value:$Rt, (add I32:$Rs, ImmPred:$Off)),
491 (MI IntRegs:$Rs, imm:$Off, (ValueMod Value:$Rt))>;
492 def: Pat<(Store Value:$Rt, (IsOrAdd I32:$Rs, ImmPred:$Off)),
493 (MI IntRegs:$Rs, imm:$Off, (ValueMod Value:$Rt))>;
495 class Storexm_simple_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
497 : Pat<(Store Value:$Rt, I32:$Rs),
498 (MI IntRegs:$Rs, 0, (ValueMod Value:$Rt))>;
500 multiclass Storex_pat<PatFrag Store, PatFrag Value, PatLeaf ImmPred,
502 def: Storex_fi_pat <Store, Value, MI>;
503 defm: Storex_fi_add_pat <Store, Value, ImmPred, MI>;
504 defm: Storex_add_pat <Store, Value, ImmPred, MI>;
507 multiclass Storexm_pat<PatFrag Store, PatFrag Value, PatLeaf ImmPred,
508 PatFrag ValueMod, InstHexagon MI> {
509 def: Storexm_fi_pat <Store, Value, ValueMod, MI>;
510 defm: Storexm_fi_add_pat <Store, Value, ImmPred, ValueMod, MI>;
511 defm: Storexm_add_pat <Store, Value, ImmPred, ValueMod, MI>;
514 // Regular stores in the DAG have two operands: value and address.
515 // Atomic stores also have two, but they are reversed: address, value.
516 // To use atomic stores with the patterns, they need to have their operands
517 // swapped. This relies on the knowledge that the F.Fragment uses names
519 class SwapSt<PatFrag F>
520 : PatFrag<(ops node:$val, node:$ptr), F.Fragment, F.PredicateCode,
523 let AddedComplexity = 20 in {
524 defm: Storex_pat<truncstorei8, I32, s32_0ImmPred, S2_storerb_io>;
525 defm: Storex_pat<truncstorei16, I32, s31_1ImmPred, S2_storerh_io>;
526 defm: Storex_pat<store, I32, s30_2ImmPred, S2_storeri_io>;
527 defm: Storex_pat<store, I64, s29_3ImmPred, S2_storerd_io>;
529 defm: Storex_pat<SwapSt<atomic_store_8>, I32, s32_0ImmPred, S2_storerb_io>;
530 defm: Storex_pat<SwapSt<atomic_store_16>, I32, s31_1ImmPred, S2_storerh_io>;
531 defm: Storex_pat<SwapSt<atomic_store_32>, I32, s30_2ImmPred, S2_storeri_io>;
532 defm: Storex_pat<SwapSt<atomic_store_64>, I64, s29_3ImmPred, S2_storerd_io>;
535 // Simple patterns should be tried with the least priority.
536 def: Storex_simple_pat<truncstorei8, I32, S2_storerb_io>;
537 def: Storex_simple_pat<truncstorei16, I32, S2_storerh_io>;
538 def: Storex_simple_pat<store, I32, S2_storeri_io>;
539 def: Storex_simple_pat<store, I64, S2_storerd_io>;
541 def: Storex_simple_pat<SwapSt<atomic_store_8>, I32, S2_storerb_io>;
542 def: Storex_simple_pat<SwapSt<atomic_store_16>, I32, S2_storerh_io>;
543 def: Storex_simple_pat<SwapSt<atomic_store_32>, I32, S2_storeri_io>;
544 def: Storex_simple_pat<SwapSt<atomic_store_64>, I64, S2_storerd_io>;
546 let AddedComplexity = 20 in {
547 defm: Storexm_pat<truncstorei8, I64, s32_0ImmPred, LoReg, S2_storerb_io>;
548 defm: Storexm_pat<truncstorei16, I64, s31_1ImmPred, LoReg, S2_storerh_io>;
549 defm: Storexm_pat<truncstorei32, I64, s30_2ImmPred, LoReg, S2_storeri_io>;
552 def: Storexm_simple_pat<truncstorei8, I64, LoReg, S2_storerb_io>;
553 def: Storexm_simple_pat<truncstorei16, I64, LoReg, S2_storerh_io>;
554 def: Storexm_simple_pat<truncstorei32, I64, LoReg, S2_storeri_io>;
556 def: Pat <(Sext64 I32:$src), (A2_sxtw I32:$src)>;
558 def: Pat<(select (i1 (setlt I32:$src, 0)), (sub 0, I32:$src), I32:$src),
559 (A2_abs IntRegs:$src)>;
561 let AddedComplexity = 50 in
562 def: Pat<(xor (add (sra I32:$src, (i32 31)),
564 (sra I32:$src, (i32 31))),
565 (A2_abs IntRegs:$src)>;
567 def: Pat<(sra I32:$src, u5_0ImmPred:$u5),
568 (S2_asr_i_r IntRegs:$src, imm:$u5)>;
569 def: Pat<(srl I32:$src, u5_0ImmPred:$u5),
570 (S2_lsr_i_r IntRegs:$src, imm:$u5)>;
571 def: Pat<(shl I32:$src, u5_0ImmPred:$u5),
572 (S2_asl_i_r IntRegs:$src, imm:$u5)>;
574 def: Pat<(sra (add (sra I32:$src1, u5_0ImmPred:$src2), 1), (i32 1)),
575 (S2_asr_i_r_rnd IntRegs:$src1, u5_0ImmPred:$src2)>;
577 def : Pat<(not I64:$src1),
578 (A2_notp DoubleRegs:$src1)>;
580 // Count leading zeros.
581 def: Pat<(ctlz I32:$Rs), (S2_cl0 I32:$Rs)>;
582 def: Pat<(i32 (trunc (ctlz I64:$Rss))), (S2_cl0p I64:$Rss)>;
584 // Count trailing zeros: 32-bit.
585 def: Pat<(cttz I32:$Rs), (S2_ct0 I32:$Rs)>;
587 // Count leading ones.
588 def: Pat<(ctlz (not I32:$Rs)), (S2_cl1 I32:$Rs)>;
589 def: Pat<(i32 (trunc (ctlz (not I64:$Rss)))), (S2_cl1p I64:$Rss)>;
591 // Count trailing ones: 32-bit.
592 def: Pat<(cttz (not I32:$Rs)), (S2_ct1 I32:$Rs)>;
594 let AddedComplexity = 20 in { // Complexity greater than and/or/xor
595 def: Pat<(and I32:$Rs, IsNPow2_32:$V),
596 (S2_clrbit_i IntRegs:$Rs, (LogN2_32 $V))>;
597 def: Pat<(or I32:$Rs, IsPow2_32:$V),
598 (S2_setbit_i IntRegs:$Rs, (Log2_32 $V))>;
599 def: Pat<(xor I32:$Rs, IsPow2_32:$V),
600 (S2_togglebit_i IntRegs:$Rs, (Log2_32 $V))>;
602 def: Pat<(and I32:$Rs, (not (shl 1, I32:$Rt))),
603 (S2_clrbit_r IntRegs:$Rs, IntRegs:$Rt)>;
604 def: Pat<(or I32:$Rs, (shl 1, I32:$Rt)),
605 (S2_setbit_r IntRegs:$Rs, IntRegs:$Rt)>;
606 def: Pat<(xor I32:$Rs, (shl 1, I32:$Rt)),
607 (S2_togglebit_r IntRegs:$Rs, IntRegs:$Rt)>;
610 // Clr/set/toggle bit for 64-bit values with immediate bit index.
611 let AddedComplexity = 20 in { // Complexity greater than and/or/xor
612 def: Pat<(and I64:$Rss, IsNPow2_64L:$V),
613 (REG_SEQUENCE DoubleRegs,
614 (i32 (HiReg $Rss)), isub_hi,
615 (S2_clrbit_i (LoReg $Rss), (LogN2_64 $V)), isub_lo)>;
616 def: Pat<(and I64:$Rss, IsNPow2_64H:$V),
617 (REG_SEQUENCE DoubleRegs,
618 (S2_clrbit_i (HiReg $Rss), (UDEC32 (i32 (LogN2_64 $V)))),
620 (i32 (LoReg $Rss)), isub_lo)>;
622 def: Pat<(or I64:$Rss, IsPow2_64L:$V),
623 (REG_SEQUENCE DoubleRegs,
624 (i32 (HiReg $Rss)), isub_hi,
625 (S2_setbit_i (LoReg $Rss), (Log2_64 $V)), isub_lo)>;
626 def: Pat<(or I64:$Rss, IsPow2_64H:$V),
627 (REG_SEQUENCE DoubleRegs,
628 (S2_setbit_i (HiReg $Rss), (UDEC32 (i32 (Log2_64 $V)))),
630 (i32 (LoReg $Rss)), isub_lo)>;
632 def: Pat<(xor I64:$Rss, IsPow2_64L:$V),
633 (REG_SEQUENCE DoubleRegs,
634 (i32 (HiReg $Rss)), isub_hi,
635 (S2_togglebit_i (LoReg $Rss), (Log2_64 $V)), isub_lo)>;
636 def: Pat<(xor I64:$Rss, IsPow2_64H:$V),
637 (REG_SEQUENCE DoubleRegs,
638 (S2_togglebit_i (HiReg $Rss), (UDEC32 (i32 (Log2_64 $V)))),
640 (i32 (LoReg $Rss)), isub_lo)>;
643 let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm.
644 def: Pat<(i1 (setne (and (shl 1, u5_0ImmPred:$u5), I32:$Rs), 0)),
645 (S2_tstbit_i IntRegs:$Rs, u5_0ImmPred:$u5)>;
646 def: Pat<(i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)),
647 (S2_tstbit_r IntRegs:$Rs, IntRegs:$Rt)>;
648 def: Pat<(i1 (trunc I32:$Rs)),
649 (S2_tstbit_i IntRegs:$Rs, 0)>;
650 def: Pat<(i1 (trunc I64:$Rs)),
651 (S2_tstbit_i (LoReg DoubleRegs:$Rs), 0)>;
654 let AddedComplexity = 20 in { // Complexity greater than compare reg-imm.
655 def: Pat<(i1 (seteq (and I32:$Rs, u6_0ImmPred:$u6), 0)),
656 (C2_bitsclri IntRegs:$Rs, u6_0ImmPred:$u6)>;
657 def: Pat<(i1 (seteq (and I32:$Rs, I32:$Rt), 0)),
658 (C2_bitsclr IntRegs:$Rs, IntRegs:$Rt)>;
661 let AddedComplexity = 10 in // Complexity greater than compare reg-reg.
662 def: Pat<(i1 (seteq (and I32:$Rs, I32:$Rt), IntRegs:$Rt)),
663 (C2_bitsset IntRegs:$Rs, IntRegs:$Rt)>;
665 def: Pat<(or (or (shl (or (shl (i32 (extloadi8 (add I32:$b, 3))),
667 (i32 (zextloadi8 (add I32:$b, 2)))),
669 (shl (i32 (zextloadi8 (add I32:$b, 1))), (i32 8))),
670 (zextloadi8 I32:$b)),
671 (A2_swiz (L2_loadri_io IntRegs:$b, 0))>;
673 // Patterns for loads of i1:
674 def: Pat<(i1 (load AddrFI:$fi)),
675 (C2_tfrrp (L2_loadrub_io AddrFI:$fi, 0))>;
676 def: Pat<(i1 (load (add I32:$Rs, s32_0ImmPred:$Off))),
677 (C2_tfrrp (L2_loadrub_io IntRegs:$Rs, imm:$Off))>;
678 def: Pat<(i1 (load I32:$Rs)),
679 (C2_tfrrp (L2_loadrub_io IntRegs:$Rs, 0))>;
681 def I1toI32: OutPatFrag<(ops node:$Rs),
682 (C2_muxii (i1 $Rs), 1, 0)>;
684 def I32toI1: OutPatFrag<(ops node:$Rs),
685 (i1 (C2_tfrrp (i32 $Rs)))>;
687 defm: Storexm_pat<store, I1, s32_0ImmPred, I1toI32, S2_storerb_io>;
688 def: Storexm_simple_pat<store, I1, I1toI32, S2_storerb_io>;
690 def: Pat<(sra I64:$src, u6_0ImmPred:$u6),
691 (S2_asr_i_p DoubleRegs:$src, imm:$u6)>;
692 def: Pat<(srl I64:$src, u6_0ImmPred:$u6),
693 (S2_lsr_i_p DoubleRegs:$src, imm:$u6)>;
694 def: Pat<(shl I64:$src, u6_0ImmPred:$u6),
695 (S2_asl_i_p DoubleRegs:$src, imm:$u6)>;
697 let AddedComplexity = 100 in
698 def: Pat<(add I32:$Rt, (shl I32:$Rs, u3_0ImmPred:$u3)),
699 (S2_addasl_rrri IntRegs:$Rt, IntRegs:$Rs, imm:$u3)>;
701 def HexagonBARRIER: SDNode<"HexagonISD::BARRIER", SDTNone, [SDNPHasChain]>;
702 def: Pat<(HexagonBARRIER), (Y2_barrier)>;
704 def: Pat<(IsOrAdd (i32 AddrFI:$Rs), s32_0ImmPred:$off),
705 (PS_fi (i32 AddrFI:$Rs), s32_0ImmPred:$off)>;
708 // Support for generating global address.
709 // Taken from X86InstrInfo.td.
710 def SDTHexagonCONST32 : SDTypeProfile<1, 1, [SDTCisVT<0, i32>,
713 def HexagonCONST32 : SDNode<"HexagonISD::CONST32", SDTHexagonCONST32>;
714 def HexagonCONST32_GP : SDNode<"HexagonISD::CONST32_GP", SDTHexagonCONST32>;
716 // Map TLS addressses to A2_tfrsi.
717 def: Pat<(HexagonCONST32 tglobaltlsaddr:$addr), (A2_tfrsi s32_0Imm:$addr)>;
718 def: Pat<(HexagonCONST32 bbl:$label), (A2_tfrsi s32_0Imm:$label)>;
720 def: Pat<(i64 imm:$v), (CONST64 imm:$v)>;
721 def: Pat<(i1 0), (PS_false)>;
722 def: Pat<(i1 1), (PS_true)>;
724 // Pseudo instructions.
725 def SDT_SPCallSeqStart : SDCallSeqStart<[ SDTCisVT<0, i32> ]>;
726 def SDT_SPCallSeqEnd : SDCallSeqEnd<[ SDTCisVT<0, i32>,
729 def callseq_start : SDNode<"ISD::CALLSEQ_START", SDT_SPCallSeqStart,
730 [SDNPHasChain, SDNPOutGlue]>;
731 def callseq_end : SDNode<"ISD::CALLSEQ_END", SDT_SPCallSeqEnd,
732 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
734 def SDT_SPCall : SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>;
736 // For tailcalls a HexagonTCRet SDNode has 3 SDNode Properties - a chain,
737 // Optional Flag and Variable Arguments.
738 // Its 1 Operand has pointer type.
739 def HexagonTCRet : SDNode<"HexagonISD::TC_RETURN", SDT_SPCall,
740 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
743 def: Pat<(callseq_start timm:$amt),
744 (ADJCALLSTACKDOWN imm:$amt)>;
745 def: Pat<(callseq_end timm:$amt1, timm:$amt2),
746 (ADJCALLSTACKUP imm:$amt1, imm:$amt2)>;
749 def: Pat<(HexagonTCRet tglobaladdr:$dst),
750 (PS_tailcall_i tglobaladdr:$dst)>;
751 def: Pat<(HexagonTCRet texternalsym:$dst),
752 (PS_tailcall_i texternalsym:$dst)>;
753 def: Pat<(HexagonTCRet I32:$dst),
754 (PS_tailcall_r I32:$dst)>;
756 // Map from r0 = and(r1, 65535) to r0 = zxth(r1)
757 def: Pat<(and I32:$src1, 65535),
758 (A2_zxth IntRegs:$src1)>;
760 // Map from r0 = and(r1, 255) to r0 = zxtb(r1).
761 def: Pat<(and I32:$src1, 255),
762 (A2_zxtb IntRegs:$src1)>;
764 // Map Add(p1, true) to p1 = not(p1).
765 // Add(p1, false) should never be produced,
766 // if it does, it got to be mapped to NOOP.
767 def: Pat<(add I1:$src1, -1),
768 (C2_not PredRegs:$src1)>;
770 // Map from p0 = pnot(p0); r0 = mux(p0, #i, #j) => r0 = mux(p0, #j, #i).
771 def: Pat<(select (not I1:$src1), s8_0ImmPred:$src2, s32_0ImmPred:$src3),
772 (C2_muxii PredRegs:$src1, s32_0ImmPred:$src3, s8_0ImmPred:$src2)>;
774 // Map from p0 = pnot(p0); r0 = select(p0, #i, r1)
775 // => r0 = C2_muxir(p0, r1, #i)
776 def: Pat<(select (not I1:$src1), s32_0ImmPred:$src2,
778 (C2_muxir PredRegs:$src1, IntRegs:$src3, s32_0ImmPred:$src2)>;
780 // Map from p0 = pnot(p0); r0 = mux(p0, r1, #i)
781 // => r0 = C2_muxri (p0, #i, r1)
782 def: Pat<(select (not I1:$src1), IntRegs:$src2, s32_0ImmPred:$src3),
783 (C2_muxri PredRegs:$src1, s32_0ImmPred:$src3, IntRegs:$src2)>;
785 // Map from p0 = pnot(p0); if (p0) jump => if (!p0) jump.
786 def: Pat<(brcond (not I1:$src1), bb:$offset),
787 (J2_jumpf PredRegs:$src1, bb:$offset)>;
789 // Map from Rdd = sign_extend_inreg(Rss, i32) -> Rdd = A2_sxtw(Rss.lo).
790 def: Pat<(i64 (sext_inreg I64:$src1, i32)),
791 (A2_sxtw (LoReg DoubleRegs:$src1))>;
793 // Map from Rdd = sign_extend_inreg(Rss, i16) -> Rdd = A2_sxtw(A2_sxth(Rss.lo)).
794 def: Pat<(i64 (sext_inreg I64:$src1, i16)),
795 (A2_sxtw (A2_sxth (LoReg DoubleRegs:$src1)))>;
797 // Map from Rdd = sign_extend_inreg(Rss, i8) -> Rdd = A2_sxtw(A2_sxtb(Rss.lo)).
798 def: Pat<(i64 (sext_inreg I64:$src1, i8)),
799 (A2_sxtw (A2_sxtb (LoReg DoubleRegs:$src1)))>;
801 // We want to prevent emitting pnot's as much as possible.
802 // Map brcond with an unsupported setcc to a J2_jumpf.
803 def : Pat <(brcond (i1 (setne I32:$src1, I32:$src2)),
805 (J2_jumpf (C2_cmpeq I32:$src1, I32:$src2),
808 def : Pat <(brcond (i1 (setne I32:$src1, s10_0ImmPred:$src2)),
810 (J2_jumpf (C2_cmpeqi I32:$src1, s10_0ImmPred:$src2), bb:$offset)>;
812 def: Pat<(brcond (i1 (setne I1:$src1, (i1 -1))), bb:$offset),
813 (J2_jumpf PredRegs:$src1, bb:$offset)>;
815 def: Pat<(brcond (i1 (setne I1:$src1, (i1 0))), bb:$offset),
816 (J2_jumpt PredRegs:$src1, bb:$offset)>;
818 // cmp.lt(Rs, Imm) -> !cmp.ge(Rs, Imm) -> !cmp.gt(Rs, Imm-1)
819 def: Pat<(brcond (i1 (setlt I32:$src1, s8_0ImmPred:$src2)), bb:$offset),
820 (J2_jumpf (C2_cmpgti IntRegs:$src1, (SDEC1 s8_0ImmPred:$src2)),
823 // Map from a 64-bit select to an emulated 64-bit mux.
824 // Hexagon does not support 64-bit MUXes; so emulate with combines.
825 def: Pat<(select I1:$src1, I64:$src2,
827 (A2_combinew (C2_mux PredRegs:$src1, (HiReg DoubleRegs:$src2),
828 (HiReg DoubleRegs:$src3)),
829 (C2_mux PredRegs:$src1, (LoReg DoubleRegs:$src2),
830 (LoReg DoubleRegs:$src3)))>;
832 // Map from a 1-bit select to logical ops.
833 // From LegalizeDAG.cpp: (B1 ? B2 : B3) <=> (B1 & B2)|(!B1&B3).
834 def: Pat<(select I1:$src1, I1:$src2, I1:$src3),
835 (C2_or (C2_and PredRegs:$src1, PredRegs:$src2),
836 (C2_and (C2_not PredRegs:$src1), PredRegs:$src3))>;
838 // Map for truncating from 64 immediates to 32 bit immediates.
839 def: Pat<(i32 (trunc I64:$src)),
840 (LoReg DoubleRegs:$src)>;
842 // Map for truncating from i64 immediates to i1 bit immediates.
843 def: Pat<(i1 (trunc I64:$src)),
844 (C2_tfrrp (LoReg DoubleRegs:$src))>;
846 // rs <= rt -> !(rs > rt).
847 let AddedComplexity = 30 in
848 def: Pat<(i1 (setle I32:$src1, s32_0ImmPred:$src2)),
849 (C2_not (C2_cmpgti IntRegs:$src1, s32_0ImmPred:$src2))>;
851 // rs <= rt -> !(rs > rt).
852 def : Pat<(i1 (setle I32:$src1, I32:$src2)),
853 (i1 (C2_not (C2_cmpgt I32:$src1, I32:$src2)))>;
855 // Rss <= Rtt -> !(Rss > Rtt).
856 def: Pat<(i1 (setle I64:$src1, I64:$src2)),
857 (C2_not (C2_cmpgtp DoubleRegs:$src1, DoubleRegs:$src2))>;
859 // Map cmpne -> cmpeq.
860 // Hexagon_TODO: We should improve on this.
861 // rs != rt -> !(rs == rt).
862 let AddedComplexity = 30 in
863 def: Pat<(i1 (setne I32:$src1, s32_0ImmPred:$src2)),
864 (C2_not (C2_cmpeqi IntRegs:$src1, s32_0ImmPred:$src2))>;
866 // Convert setne back to xor for hexagon since we compute w/ pred registers.
867 def: Pat<(i1 (setne I1:$src1, I1:$src2)),
868 (C2_xor PredRegs:$src1, PredRegs:$src2)>;
870 // Map cmpne(Rss) -> !cmpew(Rss).
871 // rs != rt -> !(rs == rt).
872 def: Pat<(i1 (setne I64:$src1, I64:$src2)),
873 (C2_not (C2_cmpeqp DoubleRegs:$src1, DoubleRegs:$src2))>;
875 // Map cmpge(Rs, Rt) -> !cmpgt(Rs, Rt).
876 // rs >= rt -> !(rt > rs).
877 def : Pat <(i1 (setge I32:$src1, I32:$src2)),
878 (i1 (C2_not (i1 (C2_cmpgt I32:$src2, I32:$src1))))>;
880 // cmpge(Rs, Imm) -> cmpgt(Rs, Imm-1)
881 let AddedComplexity = 30 in
882 def: Pat<(i1 (setge I32:$src1, s32_0ImmPred:$src2)),
883 (C2_cmpgti IntRegs:$src1, (SDEC1 s32_0ImmPred:$src2))>;
885 // Map cmpge(Rss, Rtt) -> !cmpgt(Rtt, Rss).
886 // rss >= rtt -> !(rtt > rss).
887 def: Pat<(i1 (setge I64:$src1, I64:$src2)),
888 (C2_not (C2_cmpgtp DoubleRegs:$src2, DoubleRegs:$src1))>;
890 // Map cmplt(Rs, Imm) -> !cmpge(Rs, Imm).
891 // !cmpge(Rs, Imm) -> !cmpgt(Rs, Imm-1).
892 // rs < rt -> !(rs >= rt).
893 let AddedComplexity = 30 in
894 def: Pat<(i1 (setlt I32:$src1, s32_0ImmPred:$src2)),
895 (C2_not (C2_cmpgti IntRegs:$src1, (SDEC1 s32_0ImmPred:$src2)))>;
897 // Generate cmpgeu(Rs, #0) -> cmpeq(Rs, Rs)
898 def: Pat<(i1 (setuge I32:$src1, 0)),
899 (C2_cmpeq IntRegs:$src1, IntRegs:$src1)>;
901 // Generate cmpgeu(Rs, #u8) -> cmpgtu(Rs, #u8 -1)
902 def: Pat<(i1 (setuge I32:$src1, u32_0ImmPred:$src2)),
903 (C2_cmpgtui IntRegs:$src1, (UDEC1 u32_0ImmPred:$src2))>;
905 // Generate cmpgtu(Rs, #u9)
906 def: Pat<(i1 (setugt I32:$src1, u32_0ImmPred:$src2)),
907 (C2_cmpgtui IntRegs:$src1, u32_0ImmPred:$src2)>;
909 // Map from Rs >= Rt -> !(Rt > Rs).
910 // rs >= rt -> !(rt > rs).
911 def: Pat<(i1 (setuge I64:$src1, I64:$src2)),
912 (C2_not (C2_cmpgtup DoubleRegs:$src2, DoubleRegs:$src1))>;
914 // Map from cmpleu(Rss, Rtt) -> !cmpgtu(Rss, Rtt-1).
915 // Map from (Rs <= Rt) -> !(Rs > Rt).
916 def: Pat<(i1 (setule I64:$src1, I64:$src2)),
917 (C2_not (C2_cmpgtup DoubleRegs:$src1, DoubleRegs:$src2))>;
921 def: Pat<(i32 (sext I1:$Pu)),
922 (C2_muxii I1:$Pu, -1, 0)>;
925 def: Pat<(i64 (sext I1:$Pu)),
926 (A2_combinew (C2_muxii PredRegs:$Pu, -1, 0),
927 (C2_muxii PredRegs:$Pu, -1, 0))>;
931 def: Pat<(i32 (zext I1:$Pu)),
932 (C2_muxii PredRegs:$Pu, 1, 0)>;
935 def: Pat<(i64 (zext I1:$Pu)),
936 (ToZext64 (C2_muxii PredRegs:$Pu, 1, 0))>;
939 def: Pat<(Zext64 I32:$Rs),
940 (ToZext64 IntRegs:$Rs)>;
942 // Map from Rs = Pd to Pd = mux(Pd, #1, #0)
943 def: Pat<(i32 (anyext I1:$Pu)),
944 (C2_muxii PredRegs:$Pu, 1, 0)>;
946 // Map from Rss = Pd to Rdd = combine(#0, (mux(Pd, #1, #0)))
947 def: Pat<(i64 (anyext I1:$Pu)),
948 (ToZext64 (C2_muxii PredRegs:$Pu, 1, 0))>;
950 // Clear the sign bit in a 64-bit register.
951 def ClearSign : OutPatFrag<(ops node:$Rss),
952 (A2_combinew (S2_clrbit_i (HiReg $Rss), 31), (LoReg $Rss))>;
954 def MulHU : OutPatFrag<(ops node:$Rss, node:$Rtt),
960 (S2_lsr_i_p (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt)), 32),
963 (A2_combinew (A2_tfrsi 0),
964 (LoReg (M2_dpmpyuu_s0 (LoReg $Rss), (HiReg $Rtt))))),
968 (S2_lsr_i_p (M2_dpmpyuu_s0 (LoReg $Rss), (HiReg $Rtt)), 32))>;
970 // Multiply 64-bit unsigned and use upper result.
971 def : Pat <(mulhu I64:$Rss, I64:$Rtt), (MulHU $Rss, $Rtt)>;
973 // Multiply 64-bit signed and use upper result.
975 // For two signed 64-bit integers A and B, let A' and B' denote A and B
976 // with the sign bit cleared. Then A = -2^63*s(A) + A', where s(A) is the
977 // sign bit of A (and identically for B). With this notation, the signed
978 // product A*B can be written as:
979 // AB = (-2^63 s(A) + A') * (-2^63 s(B) + B')
980 // = 2^126 s(A)s(B) - 2^63 [s(A)B'+s(B)A'] + A'B'
981 // = 2^126 s(A)s(B) + 2^63 [s(A)B'+s(B)A'] + A'B' - 2*2^63 [s(A)B'+s(B)A']
982 // = (unsigned product AB) - 2^64 [s(A)B'+s(B)A']
984 def : Pat <(mulhs I64:$Rss, I64:$Rtt),
988 (A2_andp (S2_asr_i_p $Rss, 63), (ClearSign $Rtt)),
989 (A2_andp (S2_asr_i_p $Rtt, 63), (ClearSign $Rss))))>;
991 // Hexagon specific ISD nodes.
992 def SDTHexagonALLOCA : SDTypeProfile<1, 2,
993 [SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
994 def HexagonALLOCA : SDNode<"HexagonISD::ALLOCA", SDTHexagonALLOCA,
998 def: Pat<(HexagonALLOCA I32:$Rs, (i32 imm:$A)),
999 (PS_alloca IntRegs:$Rs, imm:$A)>;
1001 def HexagonJT: SDNode<"HexagonISD::JT", SDTIntUnaryOp>;
1002 def HexagonCP: SDNode<"HexagonISD::CP", SDTIntUnaryOp>;
1004 def: Pat<(HexagonJT tjumptable:$dst), (A2_tfrsi imm:$dst)>;
1005 def: Pat<(HexagonCP tconstpool:$dst), (A2_tfrsi imm:$dst)>;
1007 let AddedComplexity = 100 in
1008 def: Pat<(add I32:$src1, (sra I32:$Rs, u5_0ImmPred:$u5)), (S2_asr_i_r_acc IntRegs:$src1, IntRegs:$Rs, u5_0ImmPred:$u5)>;
1009 def: Pat<(sub I32:$src1, (sra I32:$Rs, u5_0ImmPred:$u5)), (S2_asr_i_r_nac IntRegs:$src1, IntRegs:$Rs, u5_0ImmPred:$u5)>;
1010 def: Pat<(and I32:$src1, (sra I32:$Rs, u5_0ImmPred:$u5)), (S2_asr_i_r_and IntRegs:$src1, IntRegs:$Rs, u5_0ImmPred:$u5)>;
1011 def: Pat<(or I32:$src1, (sra I32:$Rs, u5_0ImmPred:$u5)), (S2_asr_i_r_or IntRegs:$src1, IntRegs:$Rs, u5_0ImmPred:$u5)>;
1013 let AddedComplexity = 100 in
1014 def: Pat<(add I64:$src1, (sra I64:$Rs, u6_0ImmPred:$u5)), (S2_asr_i_p_acc DoubleRegs:$src1, DoubleRegs:$Rs, u6_0ImmPred:$u5)>;
1015 def: Pat<(sub I64:$src1, (sra I64:$Rs, u6_0ImmPred:$u5)), (S2_asr_i_p_nac DoubleRegs:$src1, DoubleRegs:$Rs, u6_0ImmPred:$u5)>;
1016 def: Pat<(and I64:$src1, (sra I64:$Rs, u6_0ImmPred:$u5)), (S2_asr_i_p_and DoubleRegs:$src1, DoubleRegs:$Rs, u6_0ImmPred:$u5)>;
1017 def: Pat<(or I64:$src1, (sra I64:$Rs, u6_0ImmPred:$u5)), (S2_asr_i_p_or DoubleRegs:$src1, DoubleRegs:$Rs, u6_0ImmPred:$u5)>;
1019 let AddedComplexity = 100 in
1020 def: Pat<(add I32:$src1, (srl I32:$Rs, u5_0ImmPred:$u5)), (S2_lsr_i_r_acc IntRegs:$src1, IntRegs:$Rs, u5_0ImmPred:$u5)>;
1021 def: Pat<(sub I32:$src1, (srl I32:$Rs, u5_0ImmPred:$u5)), (S2_lsr_i_r_nac IntRegs:$src1, IntRegs:$Rs, u5_0ImmPred:$u5)>;
1022 def: Pat<(and I32:$src1, (srl I32:$Rs, u5_0ImmPred:$u5)), (S2_lsr_i_r_and IntRegs:$src1, IntRegs:$Rs, u5_0ImmPred:$u5)>;
1023 def: Pat<(or I32:$src1, (srl I32:$Rs, u5_0ImmPred:$u5)), (S2_lsr_i_r_or IntRegs:$src1, IntRegs:$Rs, u5_0ImmPred:$u5)>;
1024 let AddedComplexity = 100 in
1025 def: Pat<(xor I32:$src1, (srl I32:$Rs, u5_0ImmPred:$u5)), (S2_lsr_i_r_xacc IntRegs:$src1, IntRegs:$Rs, u5_0ImmPred:$u5)>;
1027 let AddedComplexity = 100 in
1028 def: Pat<(add I64:$src1, (srl I64:$Rs, u6_0ImmPred:$u5)), (S2_lsr_i_p_acc DoubleRegs:$src1, DoubleRegs:$Rs, u6_0ImmPred:$u5)>;
1029 def: Pat<(sub I64:$src1, (srl I64:$Rs, u6_0ImmPred:$u5)), (S2_lsr_i_p_nac DoubleRegs:$src1, DoubleRegs:$Rs, u6_0ImmPred:$u5)>;
1030 def: Pat<(and I64:$src1, (srl I64:$Rs, u6_0ImmPred:$u5)), (S2_lsr_i_p_and DoubleRegs:$src1, DoubleRegs:$Rs, u6_0ImmPred:$u5)>;
1031 def: Pat<(or I64:$src1, (srl I64:$Rs, u6_0ImmPred:$u5)), (S2_lsr_i_p_or DoubleRegs:$src1, DoubleRegs:$Rs, u6_0ImmPred:$u5)>;
1032 let AddedComplexity = 100 in
1033 def: Pat<(xor I64:$src1, (srl I64:$Rs, u6_0ImmPred:$u5)), (S2_lsr_i_p_xacc DoubleRegs:$src1, DoubleRegs:$Rs, u6_0ImmPred:$u5)>;
1035 let AddedComplexity = 100 in
1036 def: Pat<(add I32:$src1, (shl I32:$Rs, u5_0ImmPred:$u5)), (S2_asl_i_r_acc IntRegs:$src1, IntRegs:$Rs, u5_0ImmPred:$u5)>;
1037 def: Pat<(sub I32:$src1, (shl I32:$Rs, u5_0ImmPred:$u5)), (S2_asl_i_r_nac IntRegs:$src1, IntRegs:$Rs, u5_0ImmPred:$u5)>;
1038 def: Pat<(and I32:$src1, (shl I32:$Rs, u5_0ImmPred:$u5)), (S2_asl_i_r_and IntRegs:$src1, IntRegs:$Rs, u5_0ImmPred:$u5)>;
1039 def: Pat<(or I32:$src1, (shl I32:$Rs, u5_0ImmPred:$u5)), (S2_asl_i_r_or IntRegs:$src1, IntRegs:$Rs, u5_0ImmPred:$u5)>;
1040 let AddedComplexity = 100 in
1041 def: Pat<(xor I32:$src1, (shl I32:$Rs, u5_0ImmPred:$u5)), (S2_asl_i_r_xacc IntRegs:$src1, IntRegs:$Rs, u5_0ImmPred:$u5)>;
1043 let AddedComplexity = 100 in
1044 def: Pat<(add I64:$src1, (shl I64:$Rs, u6_0ImmPred:$u5)), (S2_asl_i_p_acc DoubleRegs:$src1, DoubleRegs:$Rs, u6_0ImmPred:$u5)>;
1045 def: Pat<(sub I64:$src1, (shl I64:$Rs, u6_0ImmPred:$u5)), (S2_asl_i_p_nac DoubleRegs:$src1, DoubleRegs:$Rs, u6_0ImmPred:$u5)>;
1046 def: Pat<(and I64:$src1, (shl I64:$Rs, u6_0ImmPred:$u5)), (S2_asl_i_p_and DoubleRegs:$src1, DoubleRegs:$Rs, u6_0ImmPred:$u5)>;
1047 def: Pat<(or I64:$src1, (shl I64:$Rs, u6_0ImmPred:$u5)), (S2_asl_i_p_or DoubleRegs:$src1, DoubleRegs:$Rs, u6_0ImmPred:$u5)>;
1048 let AddedComplexity = 100 in
1049 def: Pat<(xor I64:$src1, (shl I64:$Rs, u6_0ImmPred:$u5)), (S2_asl_i_p_xacc DoubleRegs:$src1, DoubleRegs:$Rs, u6_0ImmPred:$u5)>;
1051 let AddedComplexity = 100 in
1052 def: Pat<(add I32:$src1, (shl I32:$Rs, I32:$Rt)), (S2_asl_r_r_acc IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1053 def: Pat<(sub I32:$src1, (shl I32:$Rs, I32:$Rt)), (S2_asl_r_r_nac IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1054 def: Pat<(and I32:$src1, (shl I32:$Rs, I32:$Rt)), (S2_asl_r_r_and IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1055 def: Pat<(or I32:$src1, (shl I32:$Rs, I32:$Rt)), (S2_asl_r_r_or IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1056 let AddedComplexity = 100 in
1057 def: Pat<(add I64:$src1, (shl I64:$Rs, I32:$Rt)), (S2_asl_r_p_acc DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1058 def: Pat<(sub I64:$src1, (shl I64:$Rs, I32:$Rt)), (S2_asl_r_p_nac DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1059 def: Pat<(and I64:$src1, (shl I64:$Rs, I32:$Rt)), (S2_asl_r_p_and DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1060 def: Pat<(or I64:$src1, (shl I64:$Rs, I32:$Rt)), (S2_asl_r_p_or DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1061 def: Pat<(xor I64:$src1, (shl I64:$Rs, I32:$Rt)), (S2_asl_r_p_xor DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1063 let AddedComplexity = 100 in
1064 def: Pat<(add I32:$src1, (sra I32:$Rs, I32:$Rt)), (S2_asr_r_r_acc IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1065 def: Pat<(sub I32:$src1, (sra I32:$Rs, I32:$Rt)), (S2_asr_r_r_nac IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1066 def: Pat<(and I32:$src1, (sra I32:$Rs, I32:$Rt)), (S2_asr_r_r_and IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1067 def: Pat<(or I32:$src1, (sra I32:$Rs, I32:$Rt)), (S2_asr_r_r_or IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1068 let AddedComplexity = 100 in
1069 def: Pat<(add I64:$src1, (sra I64:$Rs, I32:$Rt)), (S2_asr_r_p_acc DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1070 def: Pat<(sub I64:$src1, (sra I64:$Rs, I32:$Rt)), (S2_asr_r_p_nac DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1071 def: Pat<(and I64:$src1, (sra I64:$Rs, I32:$Rt)), (S2_asr_r_p_and DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1072 def: Pat<(or I64:$src1, (sra I64:$Rs, I32:$Rt)), (S2_asr_r_p_or DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1073 def: Pat<(xor I64:$src1, (sra I64:$Rs, I32:$Rt)), (S2_asr_r_p_xor DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1075 let AddedComplexity = 100 in
1076 def: Pat<(add I32:$src1, (srl I32:$Rs, I32:$Rt)), (S2_lsr_r_r_acc IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1077 def: Pat<(sub I32:$src1, (srl I32:$Rs, I32:$Rt)), (S2_lsr_r_r_nac IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1078 def: Pat<(and I32:$src1, (srl I32:$Rs, I32:$Rt)), (S2_lsr_r_r_and IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1079 def: Pat<(or I32:$src1, (srl I32:$Rs, I32:$Rt)), (S2_lsr_r_r_or IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1080 let AddedComplexity = 100 in
1081 def: Pat<(add I64:$src1, (srl I64:$Rs, I32:$Rt)), (S2_lsr_r_p_acc DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1082 def: Pat<(sub I64:$src1, (srl I64:$Rs, I32:$Rt)), (S2_lsr_r_p_nac DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1083 def: Pat<(and I64:$src1, (srl I64:$Rs, I32:$Rt)), (S2_lsr_r_p_and DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1084 def: Pat<(or I64:$src1, (srl I64:$Rs, I32:$Rt)), (S2_lsr_r_p_or DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1085 def: Pat<(xor I64:$src1, (srl I64:$Rs, I32:$Rt)), (S2_lsr_r_p_xor DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1087 let AddedComplexity = 100 in
1088 def: Pat<(add I32:$src1, (shl I32:$Rs, I32:$Rt)), (S2_lsl_r_r_acc IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1089 def: Pat<(sub I32:$src1, (shl I32:$Rs, I32:$Rt)), (S2_lsl_r_r_nac IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1090 def: Pat<(and I32:$src1, (shl I32:$Rs, I32:$Rt)), (S2_lsl_r_r_and IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1091 def: Pat<(or I32:$src1, (shl I32:$Rs, I32:$Rt)), (S2_lsl_r_r_or IntRegs:$src1, IntRegs:$Rs, IntRegs:$Rt)>;
1092 let AddedComplexity = 100 in
1093 def: Pat<(add I64:$src1, (shl I64:$Rs, I32:$Rt)), (S2_lsl_r_p_acc DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1094 def: Pat<(sub I64:$src1, (shl I64:$Rs, I32:$Rt)), (S2_lsl_r_p_nac DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1095 def: Pat<(and I64:$src1, (shl I64:$Rs, I32:$Rt)), (S2_lsl_r_p_and DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1096 def: Pat<(or I64:$src1, (shl I64:$Rs, I32:$Rt)), (S2_lsl_r_p_or DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1097 def: Pat<(xor I64:$src1, (shl I64:$Rs, I32:$Rt)), (S2_lsl_r_p_xor DoubleRegs:$src1, DoubleRegs:$Rs, IntRegs:$Rt)>;
1099 def: Pat<(sra I64:$src1, I32:$src2), (S2_asr_r_p DoubleRegs:$src1, IntRegs:$src2)>;
1100 def: Pat<(srl I64:$src1, I32:$src2), (S2_lsr_r_p DoubleRegs:$src1, IntRegs:$src2)>;
1101 def: Pat<(shl I64:$src1, I32:$src2), (S2_asl_r_p DoubleRegs:$src1, IntRegs:$src2)>;
1102 def: Pat<(shl I64:$src1, I32:$src2), (S2_lsl_r_p DoubleRegs:$src1, IntRegs:$src2)>;
1104 def: Pat<(sra I32:$src1, I32:$src2), (S2_asr_r_r IntRegs:$src1, IntRegs:$src2)>;
1105 def: Pat<(srl I32:$src1, I32:$src2), (S2_lsr_r_r IntRegs:$src1, IntRegs:$src2)>;
1106 def: Pat<(shl I32:$src1, I32:$src2), (S2_asl_r_r IntRegs:$src1, IntRegs:$src2)>;
1107 def: Pat<(shl I32:$src1, I32:$src2), (S2_lsl_r_r IntRegs:$src1, IntRegs:$src2)>;
1109 def SDTHexagonINSERT:
1110 SDTypeProfile<1, 4, [SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>,
1111 SDTCisInt<0>, SDTCisVT<3, i32>, SDTCisVT<4, i32>]>;
1112 def SDTHexagonINSERTRP:
1113 SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>,
1114 SDTCisInt<0>, SDTCisVT<3, i64>]>;
1116 def HexagonINSERT : SDNode<"HexagonISD::INSERT", SDTHexagonINSERT>;
1117 def HexagonINSERTRP : SDNode<"HexagonISD::INSERTRP", SDTHexagonINSERTRP>;
1119 def: Pat<(HexagonINSERT I32:$Rs, I32:$Rt, u5_0ImmPred:$u1, u5_0ImmPred:$u2),
1120 (S2_insert I32:$Rs, I32:$Rt, u5_0ImmPred:$u1, u5_0ImmPred:$u2)>;
1121 def: Pat<(HexagonINSERT I64:$Rs, I64:$Rt, u6_0ImmPred:$u1, u6_0ImmPred:$u2),
1122 (S2_insertp I64:$Rs, I64:$Rt, u6_0ImmPred:$u1, u6_0ImmPred:$u2)>;
1123 def: Pat<(HexagonINSERTRP I32:$Rs, I32:$Rt, I64:$Ru),
1124 (S2_insert_rp I32:$Rs, I32:$Rt, I64:$Ru)>;
1125 def: Pat<(HexagonINSERTRP I64:$Rs, I64:$Rt, I64:$Ru),
1126 (S2_insertp_rp I64:$Rs, I64:$Rt, I64:$Ru)>;
1128 let AddedComplexity = 100 in
1129 def: Pat<(or (or (shl (HexagonINSERT (i32 (zextloadi8 (add I32:$b, 2))),
1130 (i32 (extloadi8 (add I32:$b, 3))),
1133 (shl (i32 (zextloadi8 (add I32:$b, 1))), (i32 8))),
1134 (zextloadi8 I32:$b)),
1135 (A2_swiz (L2_loadri_io I32:$b, 0))>;
1137 def SDTHexagonEXTRACTU:
1138 SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>, SDTCisInt<0>, SDTCisInt<1>,
1139 SDTCisVT<2, i32>, SDTCisVT<3, i32>]>;
1140 def SDTHexagonEXTRACTURP:
1141 SDTypeProfile<1, 2, [SDTCisSameAs<0, 1>, SDTCisInt<0>, SDTCisInt<1>,
1144 def HexagonEXTRACTU : SDNode<"HexagonISD::EXTRACTU", SDTHexagonEXTRACTU>;
1145 def HexagonEXTRACTURP : SDNode<"HexagonISD::EXTRACTURP", SDTHexagonEXTRACTURP>;
1147 def: Pat<(HexagonEXTRACTU I32:$src1, u5_0ImmPred:$src2, u5_0ImmPred:$src3),
1148 (S2_extractu I32:$src1, u5_0ImmPred:$src2, u5_0ImmPred:$src3)>;
1149 def: Pat<(HexagonEXTRACTU I64:$src1, u6_0ImmPred:$src2, u6_0ImmPred:$src3),
1150 (S2_extractup I64:$src1, u6_0ImmPred:$src2, u6_0ImmPred:$src3)>;
1151 def: Pat<(HexagonEXTRACTURP I32:$src1, I64:$src2),
1152 (S2_extractu_rp I32:$src1, I64:$src2)>;
1153 def: Pat<(HexagonEXTRACTURP I64:$src1, I64:$src2),
1154 (S2_extractup_rp I64:$src1, I64:$src2)>;
1156 def n8_0ImmPred: PatLeaf<(i32 imm), [{
1157 int64_t V = N->getSExtValue();
1158 return -255 <= V && V <= 0;
1161 // Change the sign of the immediate for Rd=-mpyi(Rs,#u8)
1162 def: Pat<(mul I32:$src1, (ineg n8_0ImmPred:$src2)),
1163 (M2_mpysin IntRegs:$src1, u8_0ImmPred:$src2)>;
1165 multiclass MinMax_pats_p<PatFrag Op, InstHexagon Inst, InstHexagon SwapInst> {
1166 defm: T_MinMax_pats<Op, I64, Inst, SwapInst>;
1169 def: Pat<(add (Sext64 I32:$Rs), I64:$Rt),
1170 (A2_addsp IntRegs:$Rs, DoubleRegs:$Rt)>;
1172 let AddedComplexity = 200 in {
1173 defm: MinMax_pats_p<setge, A2_maxp, A2_minp>;
1174 defm: MinMax_pats_p<setgt, A2_maxp, A2_minp>;
1175 defm: MinMax_pats_p<setle, A2_minp, A2_maxp>;
1176 defm: MinMax_pats_p<setlt, A2_minp, A2_maxp>;
1177 defm: MinMax_pats_p<setuge, A2_maxup, A2_minup>;
1178 defm: MinMax_pats_p<setugt, A2_maxup, A2_minup>;
1179 defm: MinMax_pats_p<setule, A2_minup, A2_maxup>;
1180 defm: MinMax_pats_p<setult, A2_minup, A2_maxup>;
1183 def callv3 : SDNode<"HexagonISD::CALL", SDT_SPCall,
1184 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
1186 def callv3nr : SDNode<"HexagonISD::CALLnr", SDT_SPCall,
1187 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
1190 // Map call instruction
1191 def : Pat<(callv3 I32:$dst),
1192 (J2_callr I32:$dst)>;
1193 def : Pat<(callv3 tglobaladdr:$dst),
1194 (J2_call tglobaladdr:$dst)>;
1195 def : Pat<(callv3 texternalsym:$dst),
1196 (J2_call texternalsym:$dst)>;
1197 def : Pat<(callv3 tglobaltlsaddr:$dst),
1198 (J2_call tglobaltlsaddr:$dst)>;
1200 def : Pat<(callv3nr I32:$dst),
1201 (PS_callr_nr I32:$dst)>;
1202 def : Pat<(callv3nr tglobaladdr:$dst),
1203 (PS_call_nr tglobaladdr:$dst)>;
1204 def : Pat<(callv3nr texternalsym:$dst),
1205 (PS_call_nr texternalsym:$dst)>;
1208 def addrga: PatLeaf<(i32 AddrGA:$Addr)>;
1209 def addrgp: PatLeaf<(i32 AddrGP:$Addr)>;
1212 // Pats for instruction selection.
1214 // A class to embed the usual comparison patfrags within a zext to i32.
1215 // The seteq/setne frags use "lhs" and "rhs" as operands, so use the same
1216 // names, or else the frag's "body" won't match the operands.
1217 class CmpInReg<PatFrag Op>
1218 : PatFrag<(ops node:$lhs, node:$rhs),(i32 (zext (i1 Op.Fragment)))>;
1220 def: T_cmp32_rr_pat<A4_rcmpeq, CmpInReg<seteq>, i32>;
1221 def: T_cmp32_rr_pat<A4_rcmpneq, CmpInReg<setne>, i32>;
1223 def: T_cmp32_rr_pat<C4_cmpneq, setne, i1>;
1224 def: T_cmp32_rr_pat<C4_cmplte, setle, i1>;
1225 def: T_cmp32_rr_pat<C4_cmplteu, setule, i1>;
1227 def: T_cmp32_rr_pat<C4_cmplte, RevCmp<setge>, i1>;
1228 def: T_cmp32_rr_pat<C4_cmplteu, RevCmp<setuge>, i1>;
1230 let AddedComplexity = 100 in {
1231 def: Pat<(i1 (seteq (and (xor I32:$Rs, I32:$Rt),
1233 (A4_cmpbeq IntRegs:$Rs, IntRegs:$Rt)>;
1234 def: Pat<(i1 (setne (and (xor I32:$Rs, I32:$Rt),
1236 (C2_not (A4_cmpbeq IntRegs:$Rs, IntRegs:$Rt))>;
1237 def: Pat<(i1 (seteq (and (xor I32:$Rs, I32:$Rt),
1239 (A4_cmpheq IntRegs:$Rs, IntRegs:$Rt)>;
1240 def: Pat<(i1 (setne (and (xor I32:$Rs, I32:$Rt),
1242 (C2_not (A4_cmpheq IntRegs:$Rs, IntRegs:$Rt))>;
1245 def: Pat<(i32 (zext (i1 (seteq I32:$Rs, s32_0ImmPred:$s8)))),
1246 (A4_rcmpeqi IntRegs:$Rs, s32_0ImmPred:$s8)>;
1247 def: Pat<(i32 (zext (i1 (setne I32:$Rs, s32_0ImmPred:$s8)))),
1248 (A4_rcmpneqi IntRegs:$Rs, s32_0ImmPred:$s8)>;
1250 // Preserve the S2_tstbit_r generation
1251 def: Pat<(i32 (zext (i1 (setne (i32 (and (i32 (shl 1, I32:$src2)),
1253 (C2_muxii (S2_tstbit_r IntRegs:$src1, IntRegs:$src2), 1, 0)>;
1255 // The complexity of the combines involving immediates should be greater
1256 // than the complexity of the combine with two registers.
1257 let AddedComplexity = 50 in {
1258 def: Pat<(HexagonCOMBINE IntRegs:$r, s32_0ImmPred:$i),
1259 (A4_combineri IntRegs:$r, s32_0ImmPred:$i)>;
1261 def: Pat<(HexagonCOMBINE s32_0ImmPred:$i, IntRegs:$r),
1262 (A4_combineir s32_0ImmPred:$i, IntRegs:$r)>;
1265 // The complexity of the combine with two immediates should be greater than
1266 // the complexity of a combine involving a register.
1267 let AddedComplexity = 75 in {
1268 def: Pat<(HexagonCOMBINE s8_0ImmPred:$s8, u32_0ImmPred:$u6),
1269 (A4_combineii imm:$s8, imm:$u6)>;
1270 def: Pat<(HexagonCOMBINE s32_0ImmPred:$s8, s8_0ImmPred:$S8),
1271 (A2_combineii imm:$s8, imm:$S8)>;
1275 // Patterns to generate indexed loads with different forms of the address:
1278 // - base (without offset).
1279 multiclass Loadxm_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
1280 PatLeaf ImmPred, InstHexagon MI> {
1281 def: Pat<(VT (Load AddrFI:$fi)),
1282 (VT (ValueMod (MI AddrFI:$fi, 0)))>;
1283 def: Pat<(VT (Load (add AddrFI:$fi, ImmPred:$Off))),
1284 (VT (ValueMod (MI AddrFI:$fi, imm:$Off)))>;
1285 def: Pat<(VT (Load (add IntRegs:$Rs, ImmPred:$Off))),
1286 (VT (ValueMod (MI IntRegs:$Rs, imm:$Off)))>;
1287 def: Pat<(VT (Load I32:$Rs)),
1288 (VT (ValueMod (MI IntRegs:$Rs, 0)))>;
1291 defm: Loadxm_pat<extloadi1, i64, ToZext64, s32_0ImmPred, L2_loadrub_io>;
1292 defm: Loadxm_pat<extloadi8, i64, ToZext64, s32_0ImmPred, L2_loadrub_io>;
1293 defm: Loadxm_pat<extloadi16, i64, ToZext64, s31_1ImmPred, L2_loadruh_io>;
1294 defm: Loadxm_pat<zextloadi1, i64, ToZext64, s32_0ImmPred, L2_loadrub_io>;
1295 defm: Loadxm_pat<zextloadi8, i64, ToZext64, s32_0ImmPred, L2_loadrub_io>;
1296 defm: Loadxm_pat<zextloadi16, i64, ToZext64, s31_1ImmPred, L2_loadruh_io>;
1297 defm: Loadxm_pat<sextloadi8, i64, ToSext64, s32_0ImmPred, L2_loadrb_io>;
1298 defm: Loadxm_pat<sextloadi16, i64, ToSext64, s31_1ImmPred, L2_loadrh_io>;
1300 // Map Rdd = anyext(Rs) -> Rdd = combine(#0, Rs).
1301 def: Pat<(Aext64 I32:$src1), (ToZext64 IntRegs:$src1)>;
1303 multiclass T_LoadAbsReg_Pat <PatFrag ldOp, InstHexagon MI, ValueType VT = i32> {
1304 def : Pat <(VT (ldOp (add (shl IntRegs:$src1, u2_0ImmPred:$src2),
1305 (HexagonCONST32 tglobaladdr:$src3)))),
1306 (MI IntRegs:$src1, u2_0ImmPred:$src2, tglobaladdr:$src3)>;
1307 def : Pat <(VT (ldOp (add IntRegs:$src1,
1308 (HexagonCONST32 tglobaladdr:$src2)))),
1309 (MI IntRegs:$src1, 0, tglobaladdr:$src2)>;
1311 def : Pat <(VT (ldOp (add (shl IntRegs:$src1, u2_0ImmPred:$src2),
1312 (HexagonCONST32 tconstpool:$src3)))),
1313 (MI IntRegs:$src1, u2_0ImmPred:$src2, tconstpool:$src3)>;
1314 def : Pat <(VT (ldOp (add IntRegs:$src1,
1315 (HexagonCONST32 tconstpool:$src2)))),
1316 (MI IntRegs:$src1, 0, tconstpool:$src2)>;
1318 def : Pat <(VT (ldOp (add (shl IntRegs:$src1, u2_0ImmPred:$src2),
1319 (HexagonCONST32 tjumptable:$src3)))),
1320 (MI IntRegs:$src1, u2_0ImmPred:$src2, tjumptable:$src3)>;
1321 def : Pat <(VT (ldOp (add IntRegs:$src1,
1322 (HexagonCONST32 tjumptable:$src2)))),
1323 (MI IntRegs:$src1, 0, tjumptable:$src2)>;
1326 let AddedComplexity = 60 in {
1327 defm : T_LoadAbsReg_Pat <sextloadi8, L4_loadrb_ur>;
1328 defm : T_LoadAbsReg_Pat <zextloadi8, L4_loadrub_ur>;
1329 defm : T_LoadAbsReg_Pat <extloadi8, L4_loadrub_ur>;
1331 defm : T_LoadAbsReg_Pat <sextloadi16, L4_loadrh_ur>;
1332 defm : T_LoadAbsReg_Pat <zextloadi16, L4_loadruh_ur>;
1333 defm : T_LoadAbsReg_Pat <extloadi16, L4_loadruh_ur>;
1335 defm : T_LoadAbsReg_Pat <load, L4_loadri_ur>;
1336 defm : T_LoadAbsReg_Pat <load, L4_loadrd_ur, i64>;
1339 // 'def pats' for load instructions with base + register offset and non-zero
1340 // immediate value. Immediate value is used to left-shift the second
1341 // register operand.
1342 class Loadxs_pat<PatFrag Load, ValueType VT, InstHexagon MI>
1343 : Pat<(VT (Load (add I32:$Rs,
1344 (i32 (shl I32:$Rt, u2_0ImmPred:$u2))))),
1345 (VT (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2))>;
1347 let AddedComplexity = 40 in {
1348 def: Loadxs_pat<extloadi8, i32, L4_loadrub_rr>;
1349 def: Loadxs_pat<zextloadi8, i32, L4_loadrub_rr>;
1350 def: Loadxs_pat<sextloadi8, i32, L4_loadrb_rr>;
1351 def: Loadxs_pat<extloadi16, i32, L4_loadruh_rr>;
1352 def: Loadxs_pat<zextloadi16, i32, L4_loadruh_rr>;
1353 def: Loadxs_pat<sextloadi16, i32, L4_loadrh_rr>;
1354 def: Loadxs_pat<load, i32, L4_loadri_rr>;
1355 def: Loadxs_pat<load, i64, L4_loadrd_rr>;
1358 // 'def pats' for load instruction base + register offset and
1359 // zero immediate value.
1360 class Loadxs_simple_pat<PatFrag Load, ValueType VT, InstHexagon MI>
1361 : Pat<(VT (Load (add I32:$Rs, I32:$Rt))),
1362 (VT (MI IntRegs:$Rs, IntRegs:$Rt, 0))>;
1364 let AddedComplexity = 20 in {
1365 def: Loadxs_simple_pat<extloadi8, i32, L4_loadrub_rr>;
1366 def: Loadxs_simple_pat<zextloadi8, i32, L4_loadrub_rr>;
1367 def: Loadxs_simple_pat<sextloadi8, i32, L4_loadrb_rr>;
1368 def: Loadxs_simple_pat<extloadi16, i32, L4_loadruh_rr>;
1369 def: Loadxs_simple_pat<zextloadi16, i32, L4_loadruh_rr>;
1370 def: Loadxs_simple_pat<sextloadi16, i32, L4_loadrh_rr>;
1371 def: Loadxs_simple_pat<load, i32, L4_loadri_rr>;
1372 def: Loadxs_simple_pat<load, i64, L4_loadrd_rr>;
1375 let AddedComplexity = 40 in
1376 multiclass T_StoreAbsReg_Pats <InstHexagon MI, RegisterClass RC, ValueType VT,
1378 def : Pat<(stOp (VT RC:$src4),
1379 (add (shl I32:$src1, u2_0ImmPred:$src2),
1380 u32_0ImmPred:$src3)),
1381 (MI IntRegs:$src1, u2_0ImmPred:$src2, u32_0ImmPred:$src3, RC:$src4)>;
1383 def : Pat<(stOp (VT RC:$src4),
1384 (add (shl IntRegs:$src1, u2_0ImmPred:$src2),
1385 (HexagonCONST32 tglobaladdr:$src3))),
1386 (MI IntRegs:$src1, u2_0ImmPred:$src2, tglobaladdr:$src3, RC:$src4)>;
1388 def : Pat<(stOp (VT RC:$src4),
1389 (add IntRegs:$src1, (HexagonCONST32 tglobaladdr:$src3))),
1390 (MI IntRegs:$src1, 0, tglobaladdr:$src3, RC:$src4)>;
1393 defm : T_StoreAbsReg_Pats <S4_storerd_ur, DoubleRegs, i64, store>;
1394 defm : T_StoreAbsReg_Pats <S4_storeri_ur, IntRegs, i32, store>;
1395 defm : T_StoreAbsReg_Pats <S4_storerb_ur, IntRegs, i32, truncstorei8>;
1396 defm : T_StoreAbsReg_Pats <S4_storerh_ur, IntRegs, i32, truncstorei16>;
1398 class Storexs_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
1399 : Pat<(Store Value:$Ru, (add I32:$Rs,
1400 (i32 (shl I32:$Rt, u2_0ImmPred:$u2)))),
1401 (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2, Value:$Ru)>;
1403 let AddedComplexity = 40 in {
1404 def: Storexs_pat<truncstorei8, I32, S4_storerb_rr>;
1405 def: Storexs_pat<truncstorei16, I32, S4_storerh_rr>;
1406 def: Storexs_pat<store, I32, S4_storeri_rr>;
1407 def: Storexs_pat<store, I64, S4_storerd_rr>;
1410 def s30_2ProperPred : PatLeaf<(i32 imm), [{
1411 int64_t v = (int64_t)N->getSExtValue();
1412 return isShiftedInt<30,2>(v) && !isShiftedInt<29,3>(v);
1414 def RoundTo8 : SDNodeXForm<imm, [{
1415 int32_t Imm = N->getSExtValue();
1416 return CurDAG->getTargetConstant(Imm & -8, SDLoc(N), MVT::i32);
1419 let AddedComplexity = 40 in
1420 def: Pat<(store I64:$Ru, (add I32:$Rs, s30_2ProperPred:$Off)),
1421 (S2_storerd_io (A2_addi I32:$Rs, 4), (RoundTo8 $Off), I64:$Ru)>;
1423 class Store_rr_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
1424 : Pat<(Store Value:$Ru, (add I32:$Rs, I32:$Rt)),
1425 (MI IntRegs:$Rs, IntRegs:$Rt, 0, Value:$Ru)>;
1427 let AddedComplexity = 20 in {
1428 def: Store_rr_pat<truncstorei8, I32, S4_storerb_rr>;
1429 def: Store_rr_pat<truncstorei16, I32, S4_storerh_rr>;
1430 def: Store_rr_pat<store, I32, S4_storeri_rr>;
1431 def: Store_rr_pat<store, I64, S4_storerd_rr>;
1435 def IMM_BYTE : SDNodeXForm<imm, [{
1436 // -1 etc is represented as 255 etc
1437 // assigning to a byte restores our desired signed value.
1438 int8_t imm = N->getSExtValue();
1439 return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32);
1442 def IMM_HALF : SDNodeXForm<imm, [{
1443 // -1 etc is represented as 65535 etc
1444 // assigning to a short restores our desired signed value.
1445 int16_t imm = N->getSExtValue();
1446 return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32);
1449 def IMM_WORD : SDNodeXForm<imm, [{
1450 // -1 etc can be represented as 4294967295 etc
1451 // Currently, it's not doing this. But some optimization
1452 // might convert -1 to a large +ve number.
1453 // assigning to a word restores our desired signed value.
1454 int32_t imm = N->getSExtValue();
1455 return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32);
1458 def ToImmByte : OutPatFrag<(ops node:$R), (IMM_BYTE $R)>;
1459 def ToImmHalf : OutPatFrag<(ops node:$R), (IMM_HALF $R)>;
1460 def ToImmWord : OutPatFrag<(ops node:$R), (IMM_WORD $R)>;
1462 // Emit store-immediate, but only when the stored value will not be constant-
1463 // extended. The reason for that is that there is no pass that can optimize
1464 // constant extenders in store-immediate instructions. In some cases we can
1465 // end up will a number of such stores, all of which store the same extended
1466 // value (e.g. after unrolling a loop that initializes floating point array).
1468 // Predicates to determine if the 16-bit immediate is expressible as a sign-
1469 // extended 8-bit immediate. Store-immediate-halfword will ignore any bits
1470 // beyond 0..15, so we don't care what is in there.
1472 def i16in8ImmPred: PatLeaf<(i32 imm), [{
1473 int64_t v = (int16_t)N->getSExtValue();
1474 return v == (int64_t)(int8_t)v;
1477 // Predicates to determine if the 32-bit immediate is expressible as a sign-
1478 // extended 8-bit immediate.
1479 def i32in8ImmPred: PatLeaf<(i32 imm), [{
1480 int64_t v = (int32_t)N->getSExtValue();
1481 return v == (int64_t)(int8_t)v;
1485 let AddedComplexity = 40 in {
1486 // Even though the offset is not extendable in the store-immediate, we
1487 // can still generate the fi# in the base address. If the final offset
1488 // is not valid for the instruction, we will replace it with a scratch
1490 // def: Storexm_fi_pat <truncstorei8, s32_0ImmPred, ToImmByte, S4_storeirb_io>;
1491 // def: Storexm_fi_pat <truncstorei16, i16in8ImmPred, ToImmHalf,
1493 // def: Storexm_fi_pat <store, i32in8ImmPred, ToImmWord, S4_storeiri_io>;
1495 // defm: Storexm_fi_add_pat <truncstorei8, s32_0ImmPred, u6_0ImmPred, ToImmByte,
1497 // defm: Storexm_fi_add_pat <truncstorei16, i16in8ImmPred, u6_1ImmPred,
1498 // ToImmHalf, S4_storeirh_io>;
1499 // defm: Storexm_fi_add_pat <store, i32in8ImmPred, u6_2ImmPred, ToImmWord,
1502 defm: Storexm_add_pat<truncstorei8, s32_0ImmPred, u6_0ImmPred, ToImmByte,
1504 defm: Storexm_add_pat<truncstorei16, i16in8ImmPred, u6_1ImmPred, ToImmHalf,
1506 defm: Storexm_add_pat<store, i32in8ImmPred, u6_2ImmPred, ToImmWord,
1510 def: Storexm_simple_pat<truncstorei8, s32_0ImmPred, ToImmByte, S4_storeirb_io>;
1511 def: Storexm_simple_pat<truncstorei16, s32_0ImmPred, ToImmHalf, S4_storeirh_io>;
1512 def: Storexm_simple_pat<store, s32_0ImmPred, ToImmWord, S4_storeiri_io>;
1514 // op(Ps, op(Pt, Pu))
1515 class LogLog_pat<SDNode Op1, SDNode Op2, InstHexagon MI>
1516 : Pat<(i1 (Op1 I1:$Ps, (Op2 I1:$Pt, I1:$Pu))),
1517 (MI I1:$Ps, I1:$Pt, I1:$Pu)>;
1519 // op(Ps, op(Pt, ~Pu))
1520 class LogLogNot_pat<SDNode Op1, SDNode Op2, InstHexagon MI>
1521 : Pat<(i1 (Op1 I1:$Ps, (Op2 I1:$Pt, (not I1:$Pu)))),
1522 (MI I1:$Ps, I1:$Pt, I1:$Pu)>;
1524 def: LogLog_pat<and, and, C4_and_and>;
1525 def: LogLog_pat<and, or, C4_and_or>;
1526 def: LogLog_pat<or, and, C4_or_and>;
1527 def: LogLog_pat<or, or, C4_or_or>;
1529 def: LogLogNot_pat<and, and, C4_and_andn>;
1530 def: LogLogNot_pat<and, or, C4_and_orn>;
1531 def: LogLogNot_pat<or, and, C4_or_andn>;
1532 def: LogLogNot_pat<or, or, C4_or_orn>;
1534 //===----------------------------------------------------------------------===//
1535 // PIC: Support for PIC compilations. The patterns and SD nodes defined
1536 // below are needed to support code generation for PIC
1537 //===----------------------------------------------------------------------===//
1539 def SDT_HexagonAtGot
1540 : SDTypeProfile<1, 3, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisVT<2, i32>]>;
1541 def SDT_HexagonAtPcrel
1542 : SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
1544 // AT_GOT address-of-GOT, address-of-global, offset-in-global
1545 def HexagonAtGot : SDNode<"HexagonISD::AT_GOT", SDT_HexagonAtGot>;
1546 // AT_PCREL address-of-global
1547 def HexagonAtPcrel : SDNode<"HexagonISD::AT_PCREL", SDT_HexagonAtPcrel>;
1549 def: Pat<(HexagonAtGot I32:$got, I32:$addr, (i32 0)),
1550 (L2_loadri_io I32:$got, imm:$addr)>;
1551 def: Pat<(HexagonAtGot I32:$got, I32:$addr, s30_2ImmPred:$off),
1552 (A2_addi (L2_loadri_io I32:$got, imm:$addr), imm:$off)>;
1553 def: Pat<(HexagonAtPcrel I32:$addr),
1554 (C4_addipc imm:$addr)>;
1556 def: Pat<(i64 (and I64:$Rs, (i64 (not I64:$Rt)))),
1557 (A4_andnp DoubleRegs:$Rs, DoubleRegs:$Rt)>;
1558 def: Pat<(i64 (or I64:$Rs, (i64 (not I64:$Rt)))),
1559 (A4_ornp DoubleRegs:$Rs, DoubleRegs:$Rt)>;
1561 def: Pat<(add I32:$Rs, (add I32:$Ru, s32_0ImmPred:$s6)),
1562 (S4_addaddi IntRegs:$Rs, IntRegs:$Ru, imm:$s6)>;
1564 // Rd=add(Rs,sub(#s6,Ru))
1565 def: Pat<(add I32:$src1, (sub s32_0ImmPred:$src2,
1567 (S4_subaddi IntRegs:$src1, s32_0ImmPred:$src2, IntRegs:$src3)>;
1569 // Rd=sub(add(Rs,#s6),Ru)
1570 def: Pat<(sub (add I32:$src1, s32_0ImmPred:$src2),
1572 (S4_subaddi IntRegs:$src1, s32_0ImmPred:$src2, IntRegs:$src3)>;
1574 // Rd=add(sub(Rs,Ru),#s6)
1575 def: Pat<(add (sub I32:$src1, I32:$src3),
1576 (s32_0ImmPred:$src2)),
1577 (S4_subaddi IntRegs:$src1, s32_0ImmPred:$src2, IntRegs:$src3)>;
1579 def: Pat<(xor I64:$dst2,
1580 (xor I64:$Rss, I64:$Rtt)),
1581 (M4_xor_xacc DoubleRegs:$dst2, DoubleRegs:$Rss, DoubleRegs:$Rtt)>;
1582 def: Pat<(or I32:$Ru, (and (i32 IntRegs:$_src_), s32_0ImmPred:$s10)),
1583 (S4_or_andix IntRegs:$Ru, IntRegs:$_src_, imm:$s10)>;
1585 def: Pat<(or I32:$src1, (and I32:$Rs, s32_0ImmPred:$s10)),
1586 (S4_or_andi IntRegs:$src1, IntRegs:$Rs, imm:$s10)>;
1588 def: Pat<(or I32:$src1, (or I32:$Rs, s32_0ImmPred:$s10)),
1589 (S4_or_ori IntRegs:$src1, IntRegs:$Rs, imm:$s10)>;
1593 // Count trailing zeros: 64-bit.
1594 def: Pat<(i32 (trunc (cttz I64:$Rss))), (S2_ct0p I64:$Rss)>;
1596 // Count trailing ones: 64-bit.
1597 def: Pat<(i32 (trunc (cttz (not I64:$Rss)))), (S2_ct1p I64:$Rss)>;
1599 // Define leading/trailing patterns that require zero-extensions to 64 bits.
1600 def: Pat<(i64 (ctlz I64:$Rss)), (ToZext64 (S2_cl0p I64:$Rss))>;
1601 def: Pat<(i64 (cttz I64:$Rss)), (ToZext64 (S2_ct0p I64:$Rss))>;
1602 def: Pat<(i64 (ctlz (not I64:$Rss))), (ToZext64 (S2_cl1p I64:$Rss))>;
1603 def: Pat<(i64 (cttz (not I64:$Rss))), (ToZext64 (S2_ct1p I64:$Rss))>;
1605 def: Pat<(i64 (ctpop I64:$Rss)), (ToZext64 (S5_popcountp I64:$Rss))>;
1606 def: Pat<(i32 (ctpop I32:$Rs)), (S5_popcountp (A4_combineir 0, I32:$Rs))>;
1608 def: Pat<(bitreverse I32:$Rs), (S2_brev I32:$Rs)>;
1609 def: Pat<(bitreverse I64:$Rss), (S2_brevp I64:$Rss)>;
1611 def: Pat<(bswap I32:$Rs), (A2_swiz I32:$Rs)>;
1612 def: Pat<(bswap I64:$Rss), (A2_combinew (A2_swiz (LoReg $Rss)),
1613 (A2_swiz (HiReg $Rss)))>;
1615 let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm.
1616 def: Pat<(i1 (seteq (and (shl 1, u5_0ImmPred:$u5), I32:$Rs), 0)),
1617 (S4_ntstbit_i I32:$Rs, u5_0ImmPred:$u5)>;
1618 def: Pat<(i1 (seteq (and (shl 1, I32:$Rt), I32:$Rs), 0)),
1619 (S4_ntstbit_r I32:$Rs, I32:$Rt)>;
1622 // Add extra complexity to prefer these instructions over bitsset/bitsclr.
1623 // The reason is that tstbit/ntstbit can be folded into a compound instruction:
1624 // if ([!]tstbit(...)) jump ...
1625 let AddedComplexity = 100 in
1626 def: Pat<(i1 (setne (and I32:$Rs, (i32 IsPow2_32:$u5)), (i32 0))),
1627 (S2_tstbit_i I32:$Rs, (Log2_32 imm:$u5))>;
1629 let AddedComplexity = 100 in
1630 def: Pat<(i1 (seteq (and I32:$Rs, (i32 IsPow2_32:$u5)), (i32 0))),
1631 (S4_ntstbit_i I32:$Rs, (Log2_32 imm:$u5))>;
1633 // Do not increase complexity of these patterns. In the DAG, "cmp i8" may be
1634 // represented as a compare against "value & 0xFF", which is an exact match
1635 // for cmpb (same for cmph). The patterns below do not contain any additional
1636 // complexity that would make them preferable, and if they were actually used
1637 // instead of cmpb/cmph, they would result in a compare against register that
1638 // is loaded with the byte/half mask (i.e. 0xFF or 0xFFFF).
1639 def: Pat<(i1 (setne (and I32:$Rs, u6_0ImmPred:$u6), 0)),
1640 (C4_nbitsclri I32:$Rs, u6_0ImmPred:$u6)>;
1641 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), 0)),
1642 (C4_nbitsclr I32:$Rs, I32:$Rt)>;
1643 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), I32:$Rt)),
1644 (C4_nbitsset I32:$Rs, I32:$Rt)>;
1647 def: Pat<(add (mul I32:$Rs, u6_0ImmPred:$U6), u32_0ImmPred:$u6),
1648 (M4_mpyri_addi imm:$u6, IntRegs:$Rs, imm:$U6)>;
1649 def: Pat<(add (mul I32:$Rs, I32:$Rt), u32_0ImmPred:$u6),
1650 (M4_mpyrr_addi imm:$u6, IntRegs:$Rs, IntRegs:$Rt)>;
1652 def: Pat<(add I32:$src1, (mul I32:$src3, u6_2ImmPred:$src2)),
1653 (M4_mpyri_addr_u2 IntRegs:$src1, imm:$src2, IntRegs:$src3)>;
1654 def: Pat<(add I32:$src1, (mul I32:$src3, u32_0ImmPred:$src2)),
1655 (M4_mpyri_addr IntRegs:$src1, IntRegs:$src3, imm:$src2)>;
1657 def: Pat<(add I32:$Ru, (mul (i32 IntRegs:$_src_), I32:$Rs)),
1658 (M4_mpyrr_addr IntRegs:$Ru, IntRegs:$_src_, IntRegs:$Rs)>;
1660 def: T_vcmp_pat<A4_vcmpbgt, setgt, v8i8>;
1662 class T_Shift_CommOp_pat<InstHexagon MI, SDNode Op, SDNode ShOp>
1663 : Pat<(Op (ShOp IntRegs:$Rx, u5_0ImmPred:$U5), u32_0ImmPred:$u8),
1664 (MI u32_0ImmPred:$u8, IntRegs:$Rx, u5_0ImmPred:$U5)>;
1666 let AddedComplexity = 200 in {
1667 def : T_Shift_CommOp_pat <S4_addi_asl_ri, add, shl>;
1668 def : T_Shift_CommOp_pat <S4_addi_lsr_ri, add, srl>;
1669 def : T_Shift_CommOp_pat <S4_andi_asl_ri, and, shl>;
1670 def : T_Shift_CommOp_pat <S4_andi_lsr_ri, and, srl>;
1673 let AddedComplexity = 30 in {
1674 def : T_Shift_CommOp_pat <S4_ori_asl_ri, or, shl>;
1675 def : T_Shift_CommOp_pat <S4_ori_lsr_ri, or, srl>;
1678 class T_Shift_Op_pat<InstHexagon MI, SDNode Op, SDNode ShOp>
1679 : Pat<(Op u32_0ImmPred:$u8, (ShOp IntRegs:$Rx, u5_0ImmPred:$U5)),
1680 (MI u32_0ImmPred:$u8, IntRegs:$Rx, u5_0ImmPred:$U5)>;
1682 def : T_Shift_Op_pat <S4_subi_asl_ri, sub, shl>;
1683 def : T_Shift_Op_pat <S4_subi_lsr_ri, sub, srl>;
1685 let AddedComplexity = 200 in {
1686 def: Pat<(add addrga:$addr, (shl I32:$src2, u5_0ImmPred:$src3)),
1687 (S4_addi_asl_ri addrga:$addr, IntRegs:$src2, u5_0ImmPred:$src3)>;
1688 def: Pat<(add addrga:$addr, (srl I32:$src2, u5_0ImmPred:$src3)),
1689 (S4_addi_lsr_ri addrga:$addr, IntRegs:$src2, u5_0ImmPred:$src3)>;
1690 def: Pat<(sub addrga:$addr, (shl I32:$src2, u5_0ImmPred:$src3)),
1691 (S4_subi_asl_ri addrga:$addr, IntRegs:$src2, u5_0ImmPred:$src3)>;
1692 def: Pat<(sub addrga:$addr, (srl I32:$src2, u5_0ImmPred:$src3)),
1693 (S4_subi_lsr_ri addrga:$addr, IntRegs:$src2, u5_0ImmPred:$src3)>;
1696 def: Pat<(shl s6_0ImmPred:$s6, I32:$Rt),
1697 (S4_lsli imm:$s6, IntRegs:$Rt)>;
1700 //===----------------------------------------------------------------------===//
1702 //===----------------------------------------------------------------------===//
1704 def m5_0Imm8Pred : PatLeaf<(i32 imm), [{
1705 int8_t V = N->getSExtValue();
1706 return -32 < V && V <= -1;
1709 def m5_0Imm16Pred : PatLeaf<(i32 imm), [{
1710 int16_t V = N->getSExtValue();
1711 return -32 < V && V <= -1;
1714 def m5_0ImmPred : PatLeaf<(i32 imm), [{
1715 int64_t V = N->getSExtValue();
1716 return -31 <= V && V <= -1;
1719 def IsNPow2_8 : PatLeaf<(i32 imm), [{
1720 uint8_t NV = ~N->getZExtValue();
1721 return isPowerOf2_32(NV);
1724 def IsNPow2_16 : PatLeaf<(i32 imm), [{
1725 uint16_t NV = ~N->getZExtValue();
1726 return isPowerOf2_32(NV);
1729 def Log2_8 : SDNodeXForm<imm, [{
1730 uint8_t V = N->getZExtValue();
1731 return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
1734 def Log2_16 : SDNodeXForm<imm, [{
1735 uint16_t V = N->getZExtValue();
1736 return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
1739 def LogN2_8 : SDNodeXForm<imm, [{
1740 uint8_t NV = ~N->getZExtValue();
1741 return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
1744 def LogN2_16 : SDNodeXForm<imm, [{
1745 uint16_t NV = ~N->getZExtValue();
1746 return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
1749 def NegImm8 : SDNodeXForm<imm, [{
1750 int8_t NV = -N->getSExtValue();
1751 return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
1754 def NegImm16 : SDNodeXForm<imm, [{
1755 int16_t NV = -N->getSExtValue();
1756 return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
1759 def NegImm32 : SDNodeXForm<imm, [{
1760 int32_t NV = -N->getSExtValue();
1761 return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
1764 def IdImm : SDNodeXForm<imm, [{ return SDValue(N, 0); }]>;
1766 multiclass Memopxr_simple_pat<PatFrag Load, PatFrag Store, SDNode Oper,
1769 def: Pat<(Store (Oper (Load I32:$Rs), I32:$A), I32:$Rs),
1770 (MI I32:$Rs, 0, I32:$A)>;
1772 def: Pat<(Store (Oper (Load AddrFI:$Rs), I32:$A), AddrFI:$Rs),
1773 (MI AddrFI:$Rs, 0, I32:$A)>;
1776 multiclass Memopxr_add_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
1777 SDNode Oper, InstHexagon MI> {
1779 def: Pat<(Store (Oper (Load (add I32:$Rs, ImmPred:$Off)), I32:$A),
1780 (add I32:$Rs, ImmPred:$Off)),
1781 (MI I32:$Rs, imm:$Off, I32:$A)>;
1782 def: Pat<(Store (Oper (Load (IsOrAdd I32:$Rs, ImmPred:$Off)), I32:$A),
1783 (IsOrAdd I32:$Rs, ImmPred:$Off)),
1784 (MI I32:$Rs, imm:$Off, I32:$A)>;
1786 def: Pat<(Store (Oper (Load (add AddrFI:$Rs, ImmPred:$Off)), I32:$A),
1787 (add AddrFI:$Rs, ImmPred:$Off)),
1788 (MI AddrFI:$Rs, imm:$Off, I32:$A)>;
1789 def: Pat<(Store (Oper (Load (IsOrAdd AddrFI:$Rs, ImmPred:$Off)), I32:$A),
1790 (IsOrAdd AddrFI:$Rs, ImmPred:$Off)),
1791 (MI AddrFI:$Rs, imm:$Off, I32:$A)>;
1794 multiclass Memopxr_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
1795 SDNode Oper, InstHexagon MI> {
1796 defm: Memopxr_simple_pat <Load, Store, Oper, MI>;
1797 defm: Memopxr_add_pat <Load, Store, ImmPred, Oper, MI>;
1800 let AddedComplexity = 180 in {
1802 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, add,
1803 /*anyext*/ L4_add_memopb_io>;
1804 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, add,
1805 /*sext*/ L4_add_memopb_io>;
1806 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, add,
1807 /*zext*/ L4_add_memopb_io>;
1808 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, add,
1809 /*anyext*/ L4_add_memoph_io>;
1810 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, add,
1811 /*sext*/ L4_add_memoph_io>;
1812 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, add,
1813 /*zext*/ L4_add_memoph_io>;
1814 defm: Memopxr_pat<load, store, u6_2ImmPred, add, L4_add_memopw_io>;
1817 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, sub,
1818 /*anyext*/ L4_sub_memopb_io>;
1819 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub,
1820 /*sext*/ L4_sub_memopb_io>;
1821 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub,
1822 /*zext*/ L4_sub_memopb_io>;
1823 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, sub,
1824 /*anyext*/ L4_sub_memoph_io>;
1825 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub,
1826 /*sext*/ L4_sub_memoph_io>;
1827 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub,
1828 /*zext*/ L4_sub_memoph_io>;
1829 defm: Memopxr_pat<load, store, u6_2ImmPred, sub, L4_sub_memopw_io>;
1832 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, and,
1833 /*anyext*/ L4_and_memopb_io>;
1834 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, and,
1835 /*sext*/ L4_and_memopb_io>;
1836 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, and,
1837 /*zext*/ L4_and_memopb_io>;
1838 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, and,
1839 /*anyext*/ L4_and_memoph_io>;
1840 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, and,
1841 /*sext*/ L4_and_memoph_io>;
1842 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, and,
1843 /*zext*/ L4_and_memoph_io>;
1844 defm: Memopxr_pat<load, store, u6_2ImmPred, and, L4_and_memopw_io>;
1847 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, or,
1848 /*anyext*/ L4_or_memopb_io>;
1849 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, or,
1850 /*sext*/ L4_or_memopb_io>;
1851 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, or,
1852 /*zext*/ L4_or_memopb_io>;
1853 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, or,
1854 /*anyext*/ L4_or_memoph_io>;
1855 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, or,
1856 /*sext*/ L4_or_memoph_io>;
1857 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, or,
1858 /*zext*/ L4_or_memoph_io>;
1859 defm: Memopxr_pat<load, store, u6_2ImmPred, or, L4_or_memopw_io>;
1863 multiclass Memopxi_simple_pat<PatFrag Load, PatFrag Store, SDNode Oper,
1864 PatFrag Arg, SDNodeXForm ArgMod,
1867 def: Pat<(Store (Oper (Load I32:$Rs), Arg:$A), I32:$Rs),
1868 (MI I32:$Rs, 0, (ArgMod Arg:$A))>;
1870 def: Pat<(Store (Oper (Load AddrFI:$Rs), Arg:$A), AddrFI:$Rs),
1871 (MI AddrFI:$Rs, 0, (ArgMod Arg:$A))>;
1874 multiclass Memopxi_add_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
1875 SDNode Oper, PatFrag Arg, SDNodeXForm ArgMod,
1878 def: Pat<(Store (Oper (Load (add I32:$Rs, ImmPred:$Off)), Arg:$A),
1879 (add I32:$Rs, ImmPred:$Off)),
1880 (MI I32:$Rs, imm:$Off, (ArgMod Arg:$A))>;
1881 def: Pat<(Store (Oper (Load (IsOrAdd I32:$Rs, ImmPred:$Off)), Arg:$A),
1882 (IsOrAdd I32:$Rs, ImmPred:$Off)),
1883 (MI I32:$Rs, imm:$Off, (ArgMod Arg:$A))>;
1885 def: Pat<(Store (Oper (Load (add AddrFI:$Rs, ImmPred:$Off)), Arg:$A),
1886 (add AddrFI:$Rs, ImmPred:$Off)),
1887 (MI AddrFI:$Rs, imm:$Off, (ArgMod Arg:$A))>;
1888 def: Pat<(Store (Oper (Load (IsOrAdd AddrFI:$Rs, ImmPred:$Off)), Arg:$A),
1889 (IsOrAdd AddrFI:$Rs, ImmPred:$Off)),
1890 (MI AddrFI:$Rs, imm:$Off, (ArgMod Arg:$A))>;
1893 multiclass Memopxi_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
1894 SDNode Oper, PatFrag Arg, SDNodeXForm ArgMod,
1896 defm: Memopxi_simple_pat <Load, Store, Oper, Arg, ArgMod, MI>;
1897 defm: Memopxi_add_pat <Load, Store, ImmPred, Oper, Arg, ArgMod, MI>;
1901 let AddedComplexity = 200 in {
1903 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
1904 /*anyext*/ IdImm, L4_iadd_memopb_io>;
1905 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
1906 /*sext*/ IdImm, L4_iadd_memopb_io>;
1907 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
1908 /*zext*/ IdImm, L4_iadd_memopb_io>;
1909 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
1910 /*anyext*/ IdImm, L4_iadd_memoph_io>;
1911 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
1912 /*sext*/ IdImm, L4_iadd_memoph_io>;
1913 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
1914 /*zext*/ IdImm, L4_iadd_memoph_io>;
1915 defm: Memopxi_pat<load, store, u6_2ImmPred, add, u5_0ImmPred, IdImm,
1917 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
1918 /*anyext*/ NegImm8, L4_iadd_memopb_io>;
1919 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
1920 /*sext*/ NegImm8, L4_iadd_memopb_io>;
1921 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
1922 /*zext*/ NegImm8, L4_iadd_memopb_io>;
1923 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
1924 /*anyext*/ NegImm16, L4_iadd_memoph_io>;
1925 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
1926 /*sext*/ NegImm16, L4_iadd_memoph_io>;
1927 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
1928 /*zext*/ NegImm16, L4_iadd_memoph_io>;
1929 defm: Memopxi_pat<load, store, u6_2ImmPred, sub, m5_0ImmPred, NegImm32,
1933 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
1934 /*anyext*/ IdImm, L4_isub_memopb_io>;
1935 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
1936 /*sext*/ IdImm, L4_isub_memopb_io>;
1937 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
1938 /*zext*/ IdImm, L4_isub_memopb_io>;
1939 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
1940 /*anyext*/ IdImm, L4_isub_memoph_io>;
1941 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
1942 /*sext*/ IdImm, L4_isub_memoph_io>;
1943 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
1944 /*zext*/ IdImm, L4_isub_memoph_io>;
1945 defm: Memopxi_pat<load, store, u6_2ImmPred, sub, u5_0ImmPred, IdImm,
1947 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
1948 /*anyext*/ NegImm8, L4_isub_memopb_io>;
1949 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
1950 /*sext*/ NegImm8, L4_isub_memopb_io>;
1951 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
1952 /*zext*/ NegImm8, L4_isub_memopb_io>;
1953 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
1954 /*anyext*/ NegImm16, L4_isub_memoph_io>;
1955 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
1956 /*sext*/ NegImm16, L4_isub_memoph_io>;
1957 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
1958 /*zext*/ NegImm16, L4_isub_memoph_io>;
1959 defm: Memopxi_pat<load, store, u6_2ImmPred, add, m5_0ImmPred, NegImm32,
1963 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
1964 /*anyext*/ LogN2_8, L4_iand_memopb_io>;
1965 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
1966 /*sext*/ LogN2_8, L4_iand_memopb_io>;
1967 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
1968 /*zext*/ LogN2_8, L4_iand_memopb_io>;
1969 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
1970 /*anyext*/ LogN2_16, L4_iand_memoph_io>;
1971 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
1972 /*sext*/ LogN2_16, L4_iand_memoph_io>;
1973 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
1974 /*zext*/ LogN2_16, L4_iand_memoph_io>;
1975 defm: Memopxi_pat<load, store, u6_2ImmPred, and, IsNPow2_32,
1976 LogN2_32, L4_iand_memopw_io>;
1979 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
1980 /*anyext*/ Log2_8, L4_ior_memopb_io>;
1981 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
1982 /*sext*/ Log2_8, L4_ior_memopb_io>;
1983 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
1984 /*zext*/ Log2_8, L4_ior_memopb_io>;
1985 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
1986 /*anyext*/ Log2_16, L4_ior_memoph_io>;
1987 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
1988 /*sext*/ Log2_16, L4_ior_memoph_io>;
1989 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
1990 /*zext*/ Log2_16, L4_ior_memoph_io>;
1991 defm: Memopxi_pat<load, store, u6_2ImmPred, or, IsPow2_32,
1992 Log2_32, L4_ior_memopw_io>;
1995 def : T_CMP_pat <C4_cmpneqi, setne, s32_0ImmPred>;
1996 def : T_CMP_pat <C4_cmpltei, setle, s32_0ImmPred>;
1997 def : T_CMP_pat <C4_cmplteui, setule, u9_0ImmPred>;
1999 // Map cmplt(Rs, Imm) -> !cmpgt(Rs, Imm-1).
2000 def: Pat<(i1 (setlt I32:$src1, s32_0ImmPred:$src2)),
2001 (C4_cmpltei IntRegs:$src1, (SDEC1 s32_0ImmPred:$src2))>;
2003 // rs != rt -> !(rs == rt).
2004 def: Pat<(i1 (setne I32:$src1, s32_0ImmPred:$src2)),
2005 (C4_cmpneqi IntRegs:$src1, s32_0ImmPred:$src2)>;
2008 // zext( setult ( and(Rs, 255), u8))
2009 // Use the isdigit transformation below
2012 def u7_0PosImmPred : ImmLeaf<i32, [{
2013 // True if the immediate fits in an 7-bit unsigned field and
2014 // is strictly greater than 0.
2015 return Imm > 0 && isUInt<7>(Imm);
2019 // Generate code of the form 'C2_muxii(cmpbgtui(Rdd, C-1),0,1)'
2020 // for C code of the form r = ((c>='0') & (c<='9')) ? 1 : 0;.
2021 // The isdigit transformation relies on two 'clever' aspects:
2022 // 1) The data type is unsigned which allows us to eliminate a zero test after
2023 // biasing the expression by 48. We are depending on the representation of
2024 // the unsigned types, and semantics.
2025 // 2) The front end has converted <= 9 into < 10 on entry to LLVM
2028 // retval = ((c>='0') & (c<='9')) ? 1 : 0;
2029 // The code is transformed upstream of llvm into
2030 // retval = (c-48) < 10 ? 1 : 0;
2032 let AddedComplexity = 139 in
2033 def: Pat<(i32 (zext (i1 (setult (and I32:$src1, 255), u7_0PosImmPred:$src2)))),
2034 (C2_muxii (A4_cmpbgtui IntRegs:$src1, (UDEC1 imm:$src2)), 0, 1)>;
2036 class Loada_pat<PatFrag Load, ValueType VT, PatFrag Addr, InstHexagon MI>
2037 : Pat<(VT (Load Addr:$addr)), (MI Addr:$addr)>;
2039 class Loadam_pat<PatFrag Load, ValueType VT, PatFrag Addr, PatFrag ValueMod,
2041 : Pat<(VT (Load Addr:$addr)), (ValueMod (MI Addr:$addr))>;
2043 class Storea_pat<PatFrag Store, PatFrag Value, PatFrag Addr, InstHexagon MI>
2044 : Pat<(Store Value:$val, Addr:$addr), (MI Addr:$addr, Value:$val)>;
2046 class Stoream_pat<PatFrag Store, PatFrag Value, PatFrag Addr, PatFrag ValueMod,
2048 : Pat<(Store Value:$val, Addr:$addr),
2049 (MI Addr:$addr, (ValueMod Value:$val))>;
2051 let AddedComplexity = 30 in {
2052 def: Storea_pat<truncstorei8, I32, addrga, PS_storerbabs>;
2053 def: Storea_pat<truncstorei16, I32, addrga, PS_storerhabs>;
2054 def: Storea_pat<store, I32, addrga, PS_storeriabs>;
2055 def: Storea_pat<store, I64, addrga, PS_storerdabs>;
2057 def: Stoream_pat<truncstorei8, I64, addrga, LoReg, PS_storerbabs>;
2058 def: Stoream_pat<truncstorei16, I64, addrga, LoReg, PS_storerhabs>;
2059 def: Stoream_pat<truncstorei32, I64, addrga, LoReg, PS_storeriabs>;
2062 def: Storea_pat<SwapSt<atomic_store_8>, I32, addrgp, S2_storerbgp>;
2063 def: Storea_pat<SwapSt<atomic_store_16>, I32, addrgp, S2_storerhgp>;
2064 def: Storea_pat<SwapSt<atomic_store_32>, I32, addrgp, S2_storerigp>;
2065 def: Storea_pat<SwapSt<atomic_store_64>, I64, addrgp, S2_storerdgp>;
2067 let AddedComplexity = 100 in {
2068 def: Storea_pat<truncstorei8, I32, addrgp, S2_storerbgp>;
2069 def: Storea_pat<truncstorei16, I32, addrgp, S2_storerhgp>;
2070 def: Storea_pat<store, I32, addrgp, S2_storerigp>;
2071 def: Storea_pat<store, I64, addrgp, S2_storerdgp>;
2073 // Map from "i1 = constant<-1>; memw(CONST32(#foo)) = i1"
2074 // to "r0 = 1; memw(#foo) = r0"
2075 let AddedComplexity = 100 in
2076 def: Pat<(store (i1 -1), (HexagonCONST32_GP tglobaladdr:$global)),
2077 (S2_storerbgp tglobaladdr:$global, (A2_tfrsi 1))>;
2080 class LoadAbs_pats <PatFrag ldOp, InstHexagon MI, ValueType VT = i32>
2081 : Pat <(VT (ldOp (HexagonCONST32 tglobaladdr:$absaddr))),
2082 (VT (MI tglobaladdr:$absaddr))>;
2084 let AddedComplexity = 30 in {
2085 def: LoadAbs_pats <load, PS_loadriabs>;
2086 def: LoadAbs_pats <zextloadi1, PS_loadrubabs>;
2087 def: LoadAbs_pats <sextloadi8, PS_loadrbabs>;
2088 def: LoadAbs_pats <extloadi8, PS_loadrubabs>;
2089 def: LoadAbs_pats <zextloadi8, PS_loadrubabs>;
2090 def: LoadAbs_pats <sextloadi16, PS_loadrhabs>;
2091 def: LoadAbs_pats <extloadi16, PS_loadruhabs>;
2092 def: LoadAbs_pats <zextloadi16, PS_loadruhabs>;
2093 def: LoadAbs_pats <load, PS_loadrdabs, i64>;
2096 let AddedComplexity = 30 in
2097 def: Pat<(i64 (zextloadi1 (HexagonCONST32 tglobaladdr:$absaddr))),
2098 (ToZext64 (PS_loadrubabs tglobaladdr:$absaddr))>;
2100 def: Loada_pat<atomic_load_8, i32, addrgp, L2_loadrubgp>;
2101 def: Loada_pat<atomic_load_16, i32, addrgp, L2_loadruhgp>;
2102 def: Loada_pat<atomic_load_32, i32, addrgp, L2_loadrigp>;
2103 def: Loada_pat<atomic_load_64, i64, addrgp, L2_loadrdgp>;
2105 def: Loadam_pat<load, i1, addrga, I32toI1, PS_loadrubabs>;
2106 def: Loadam_pat<load, i1, addrgp, I32toI1, L2_loadrubgp>;
2108 def: Stoream_pat<store, I1, addrga, I1toI32, PS_storerbabs>;
2109 def: Stoream_pat<store, I1, addrgp, I1toI32, S2_storerbgp>;
2111 // Map from load(globaladdress) -> mem[u][bhwd](#foo)
2112 class LoadGP_pats <PatFrag ldOp, InstHexagon MI, ValueType VT = i32>
2113 : Pat <(VT (ldOp (HexagonCONST32_GP tglobaladdr:$global))),
2114 (VT (MI tglobaladdr:$global))>;
2116 let AddedComplexity = 100 in {
2117 def: LoadGP_pats <extloadi8, L2_loadrubgp>;
2118 def: LoadGP_pats <sextloadi8, L2_loadrbgp>;
2119 def: LoadGP_pats <zextloadi8, L2_loadrubgp>;
2120 def: LoadGP_pats <extloadi16, L2_loadruhgp>;
2121 def: LoadGP_pats <sextloadi16, L2_loadrhgp>;
2122 def: LoadGP_pats <zextloadi16, L2_loadruhgp>;
2123 def: LoadGP_pats <load, L2_loadrigp>;
2124 def: LoadGP_pats <load, L2_loadrdgp, i64>;
2127 // When the Interprocedural Global Variable optimizer realizes that a certain
2128 // global variable takes only two constant values, it shrinks the global to
2129 // a boolean. Catch those loads here in the following 3 patterns.
2130 let AddedComplexity = 100 in {
2131 def: LoadGP_pats <extloadi1, L2_loadrubgp>;
2132 def: LoadGP_pats <zextloadi1, L2_loadrubgp>;
2135 // Transfer global address into a register
2136 def: Pat<(HexagonCONST32 tglobaladdr:$Rs), (A2_tfrsi imm:$Rs)>;
2137 def: Pat<(HexagonCONST32_GP tblockaddress:$Rs), (A2_tfrsi imm:$Rs)>;
2138 def: Pat<(HexagonCONST32_GP tglobaladdr:$Rs), (A2_tfrsi imm:$Rs)>;
2140 let AddedComplexity = 30 in {
2141 def: Storea_pat<truncstorei8, I32, u32_0ImmPred, PS_storerbabs>;
2142 def: Storea_pat<truncstorei16, I32, u32_0ImmPred, PS_storerhabs>;
2143 def: Storea_pat<store, I32, u32_0ImmPred, PS_storeriabs>;
2146 let AddedComplexity = 30 in {
2147 def: Loada_pat<load, i32, u32_0ImmPred, PS_loadriabs>;
2148 def: Loada_pat<sextloadi8, i32, u32_0ImmPred, PS_loadrbabs>;
2149 def: Loada_pat<zextloadi8, i32, u32_0ImmPred, PS_loadrubabs>;
2150 def: Loada_pat<sextloadi16, i32, u32_0ImmPred, PS_loadrhabs>;
2151 def: Loada_pat<zextloadi16, i32, u32_0ImmPred, PS_loadruhabs>;
2154 // Indexed store word - global address.
2155 // memw(Rs+#u6:2)=#S8
2156 let AddedComplexity = 100 in
2157 defm: Storex_add_pat<store, addrga, u6_2ImmPred, S4_storeiri_io>;
2159 // Load from a global address that has only one use in the current basic block.
2160 let AddedComplexity = 100 in {
2161 def: Loada_pat<extloadi8, i32, addrga, PS_loadrubabs>;
2162 def: Loada_pat<sextloadi8, i32, addrga, PS_loadrbabs>;
2163 def: Loada_pat<zextloadi8, i32, addrga, PS_loadrubabs>;
2165 def: Loada_pat<extloadi16, i32, addrga, PS_loadruhabs>;
2166 def: Loada_pat<sextloadi16, i32, addrga, PS_loadrhabs>;
2167 def: Loada_pat<zextloadi16, i32, addrga, PS_loadruhabs>;
2169 def: Loada_pat<load, i32, addrga, PS_loadriabs>;
2170 def: Loada_pat<load, i64, addrga, PS_loadrdabs>;
2173 // Store to a global address that has only one use in the current basic block.
2174 let AddedComplexity = 100 in {
2175 def: Storea_pat<truncstorei8, I32, addrga, PS_storerbabs>;
2176 def: Storea_pat<truncstorei16, I32, addrga, PS_storerhabs>;
2177 def: Storea_pat<store, I32, addrga, PS_storeriabs>;
2178 def: Storea_pat<store, I64, addrga, PS_storerdabs>;
2180 def: Stoream_pat<truncstorei32, I64, addrga, LoReg, PS_storeriabs>;
2183 // i8/i16/i32 -> i64 loads
2184 // We need a complexity of 120 here to override preceding handling of
2186 let AddedComplexity = 120 in {
2187 def: Loadam_pat<extloadi8, i64, addrga, ToZext64, PS_loadrubabs>;
2188 def: Loadam_pat<sextloadi8, i64, addrga, ToSext64, PS_loadrbabs>;
2189 def: Loadam_pat<zextloadi8, i64, addrga, ToZext64, PS_loadrubabs>;
2191 def: Loadam_pat<extloadi16, i64, addrga, ToZext64, PS_loadruhabs>;
2192 def: Loadam_pat<sextloadi16, i64, addrga, ToSext64, PS_loadrhabs>;
2193 def: Loadam_pat<zextloadi16, i64, addrga, ToZext64, PS_loadruhabs>;
2195 def: Loadam_pat<extloadi32, i64, addrga, ToZext64, PS_loadriabs>;
2196 def: Loadam_pat<sextloadi32, i64, addrga, ToSext64, PS_loadriabs>;
2197 def: Loadam_pat<zextloadi32, i64, addrga, ToZext64, PS_loadriabs>;
2200 let AddedComplexity = 100 in {
2201 def: Loada_pat<extloadi8, i32, addrgp, PS_loadrubabs>;
2202 def: Loada_pat<sextloadi8, i32, addrgp, PS_loadrbabs>;
2203 def: Loada_pat<zextloadi8, i32, addrgp, PS_loadrubabs>;
2205 def: Loada_pat<extloadi16, i32, addrgp, PS_loadruhabs>;
2206 def: Loada_pat<sextloadi16, i32, addrgp, PS_loadrhabs>;
2207 def: Loada_pat<zextloadi16, i32, addrgp, PS_loadruhabs>;
2209 def: Loada_pat<load, i32, addrgp, PS_loadriabs>;
2210 def: Loada_pat<load, i64, addrgp, PS_loadrdabs>;
2213 let AddedComplexity = 100 in {
2214 def: Storea_pat<truncstorei8, I32, addrgp, PS_storerbabs>;
2215 def: Storea_pat<truncstorei16, I32, addrgp, PS_storerhabs>;
2216 def: Storea_pat<store, I32, addrgp, PS_storeriabs>;
2217 def: Storea_pat<store, I64, addrgp, PS_storerdabs>;
2220 def: Loada_pat<atomic_load_8, i32, addrgp, PS_loadrubabs>;
2221 def: Loada_pat<atomic_load_16, i32, addrgp, PS_loadruhabs>;
2222 def: Loada_pat<atomic_load_32, i32, addrgp, PS_loadriabs>;
2223 def: Loada_pat<atomic_load_64, i64, addrgp, PS_loadrdabs>;
2225 def: Storea_pat<SwapSt<atomic_store_8>, I32, addrgp, PS_storerbabs>;
2226 def: Storea_pat<SwapSt<atomic_store_16>, I32, addrgp, PS_storerhabs>;
2227 def: Storea_pat<SwapSt<atomic_store_32>, I32, addrgp, PS_storeriabs>;
2228 def: Storea_pat<SwapSt<atomic_store_64>, I64, addrgp, PS_storerdabs>;
2230 def: Pat<(or (or (or (shl (i64 (zext (and I32:$b, (i32 65535)))), (i32 16)),
2231 (i64 (zext (i32 (and I32:$a, (i32 65535)))))),
2232 (shl (i64 (anyext (and I32:$c, (i32 65535)))), (i32 32))),
2233 (shl (Aext64 I32:$d), (i32 48))),
2234 (A2_combinew (A2_combine_ll I32:$d, I32:$c),
2235 (A2_combine_ll I32:$b, I32:$a))>;
2237 // We need custom lowering of ISD::PREFETCH into HexagonISD::DCFETCH
2238 // because the SDNode ISD::PREFETCH has properties MayLoad and MayStore.
2239 // We don't really want either one here.
2240 def SDTHexagonDCFETCH : SDTypeProfile<0, 2, [SDTCisPtrTy<0>,SDTCisInt<1>]>;
2241 def HexagonDCFETCH : SDNode<"HexagonISD::DCFETCH", SDTHexagonDCFETCH,
2244 def: Pat<(HexagonDCFETCH IntRegs:$Rs, u11_3ImmPred:$u11_3),
2245 (Y2_dcfetchbo IntRegs:$Rs, imm:$u11_3)>;
2246 def: Pat<(HexagonDCFETCH (i32 (add IntRegs:$Rs, u11_3ImmPred:$u11_3)), (i32 0)),
2247 (Y2_dcfetchbo IntRegs:$Rs, imm:$u11_3)>;
2249 def f32ImmPred : PatLeaf<(f32 fpimm:$F)>;
2250 def f64ImmPred : PatLeaf<(f64 fpimm:$F)>;
2252 def ftoi : SDNodeXForm<fpimm, [{
2253 APInt I = N->getValueAPF().bitcastToAPInt();
2254 return CurDAG->getTargetConstant(I.getZExtValue(), SDLoc(N),
2255 MVT::getIntegerVT(I.getBitWidth()));
2259 def: Pat<(sra (i64 (add (sra I64:$src1, u6_0ImmPred:$src2), 1)), (i32 1)),
2260 (S2_asr_i_p_rnd I64:$src1, imm:$src2)>;
2262 let AddedComplexity = 20 in {
2263 defm: Loadx_pat<load, f32, s30_2ImmPred, L2_loadri_io>;
2264 defm: Loadx_pat<load, f64, s29_3ImmPred, L2_loadrd_io>;
2267 let AddedComplexity = 60 in {
2268 defm : T_LoadAbsReg_Pat <load, L4_loadri_ur, f32>;
2269 defm : T_LoadAbsReg_Pat <load, L4_loadrd_ur, f64>;
2272 let AddedComplexity = 40 in {
2273 def: Loadxs_pat<load, f32, L4_loadri_rr>;
2274 def: Loadxs_pat<load, f64, L4_loadrd_rr>;
2277 let AddedComplexity = 20 in {
2278 def: Loadxs_simple_pat<load, f32, L4_loadri_rr>;
2279 def: Loadxs_simple_pat<load, f64, L4_loadrd_rr>;
2282 let AddedComplexity = 80 in {
2283 def: Loada_pat<load, f32, u32_0ImmPred, PS_loadriabs>;
2284 def: Loada_pat<load, f32, addrga, PS_loadriabs>;
2285 def: Loada_pat<load, f64, addrga, PS_loadrdabs>;
2288 let AddedComplexity = 100 in {
2289 def: LoadGP_pats <load, L2_loadrigp, f32>;
2290 def: LoadGP_pats <load, L2_loadrdgp, f64>;
2293 let AddedComplexity = 20 in {
2294 defm: Storex_pat<store, F32, s30_2ImmPred, S2_storeri_io>;
2295 defm: Storex_pat<store, F64, s29_3ImmPred, S2_storerd_io>;
2298 // Simple patterns should be tried with the least priority.
2299 def: Storex_simple_pat<store, F32, S2_storeri_io>;
2300 def: Storex_simple_pat<store, F64, S2_storerd_io>;
2302 let AddedComplexity = 60 in {
2303 defm : T_StoreAbsReg_Pats <S4_storeri_ur, IntRegs, f32, store>;
2304 defm : T_StoreAbsReg_Pats <S4_storerd_ur, DoubleRegs, f64, store>;
2307 let AddedComplexity = 40 in {
2308 def: Storexs_pat<store, F32, S4_storeri_rr>;
2309 def: Storexs_pat<store, F64, S4_storerd_rr>;
2312 let AddedComplexity = 20 in {
2313 def: Store_rr_pat<store, F32, S4_storeri_rr>;
2314 def: Store_rr_pat<store, F64, S4_storerd_rr>;
2317 let AddedComplexity = 80 in {
2318 def: Storea_pat<store, F32, addrga, PS_storeriabs>;
2319 def: Storea_pat<store, F64, addrga, PS_storerdabs>;
2322 let AddedComplexity = 100 in {
2323 def: Storea_pat<store, F32, addrgp, S2_storerigp>;
2324 def: Storea_pat<store, F64, addrgp, S2_storerdgp>;
2327 defm: Storex_pat<store, F32, s30_2ImmPred, S2_storeri_io>;
2328 defm: Storex_pat<store, F64, s29_3ImmPred, S2_storerd_io>;
2329 def: Storex_simple_pat<store, F32, S2_storeri_io>;
2330 def: Storex_simple_pat<store, F64, S2_storerd_io>;
2332 def: Pat<(fadd F32:$src1, F32:$src2),
2333 (F2_sfadd F32:$src1, F32:$src2)>;
2335 def: Pat<(fsub F32:$src1, F32:$src2),
2336 (F2_sfsub F32:$src1, F32:$src2)>;
2338 def: Pat<(fmul F32:$src1, F32:$src2),
2339 (F2_sfmpy F32:$src1, F32:$src2)>;
2341 let Predicates = [HasV5T] in {
2342 def: Pat<(f32 (fminnum F32:$Rs, F32:$Rt)), (F2_sfmin F32:$Rs, F32:$Rt)>;
2343 def: Pat<(f32 (fmaxnum F32:$Rs, F32:$Rt)), (F2_sfmax F32:$Rs, F32:$Rt)>;
2346 let AddedComplexity = 100, Predicates = [HasV5T] in {
2347 class SfSel12<PatFrag Cmp, InstHexagon MI>
2348 : Pat<(select (i1 (Cmp F32:$Rs, F32:$Rt)), F32:$Rs, F32:$Rt),
2349 (MI F32:$Rs, F32:$Rt)>;
2350 class SfSel21<PatFrag Cmp, InstHexagon MI>
2351 : Pat<(select (i1 (Cmp F32:$Rs, F32:$Rt)), F32:$Rt, F32:$Rs),
2352 (MI F32:$Rs, F32:$Rt)>;
2354 def: SfSel12<setolt, F2_sfmin>;
2355 def: SfSel12<setole, F2_sfmin>;
2356 def: SfSel12<setogt, F2_sfmax>;
2357 def: SfSel12<setoge, F2_sfmax>;
2358 def: SfSel21<setolt, F2_sfmax>;
2359 def: SfSel21<setole, F2_sfmax>;
2360 def: SfSel21<setogt, F2_sfmin>;
2361 def: SfSel21<setoge, F2_sfmin>;
2364 class T_fcmp32_pat<PatFrag OpNode, InstHexagon MI>
2365 : Pat<(i1 (OpNode F32:$src1, F32:$src2)),
2366 (MI F32:$src1, F32:$src2)>;
2367 class T_fcmp64_pat<PatFrag OpNode, InstHexagon MI>
2368 : Pat<(i1 (OpNode F64:$src1, F64:$src2)),
2369 (MI F64:$src1, F64:$src2)>;
2371 def: T_fcmp32_pat<setoge, F2_sfcmpge>;
2372 def: T_fcmp32_pat<setuo, F2_sfcmpuo>;
2373 def: T_fcmp32_pat<setoeq, F2_sfcmpeq>;
2374 def: T_fcmp32_pat<setogt, F2_sfcmpgt>;
2376 def: T_fcmp64_pat<setoge, F2_dfcmpge>;
2377 def: T_fcmp64_pat<setuo, F2_dfcmpuo>;
2378 def: T_fcmp64_pat<setoeq, F2_dfcmpeq>;
2379 def: T_fcmp64_pat<setogt, F2_dfcmpgt>;
2381 let Predicates = [HasV5T] in
2382 multiclass T_fcmp_pats<PatFrag cmpOp, InstHexagon IntMI, InstHexagon DoubleMI> {
2384 def: Pat<(i1 (cmpOp F32:$src1, F32:$src2)),
2385 (IntMI F32:$src1, F32:$src2)>;
2387 def: Pat<(i1 (cmpOp F64:$src1, F64:$src2)),
2388 (DoubleMI F64:$src1, F64:$src2)>;
2391 defm : T_fcmp_pats <seteq, F2_sfcmpeq, F2_dfcmpeq>;
2392 defm : T_fcmp_pats <setgt, F2_sfcmpgt, F2_dfcmpgt>;
2393 defm : T_fcmp_pats <setge, F2_sfcmpge, F2_dfcmpge>;
2395 //===----------------------------------------------------------------------===//
2396 // Multiclass to define 'Def Pats' for unordered gt, ge, eq operations.
2397 //===----------------------------------------------------------------------===//
2398 let Predicates = [HasV5T] in
2399 multiclass unord_Pats <PatFrag cmpOp, InstHexagon IntMI, InstHexagon DoubleMI> {
2401 def: Pat<(i1 (cmpOp F32:$src1, F32:$src2)),
2402 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
2403 (IntMI F32:$src1, F32:$src2))>;
2406 def: Pat<(i1 (cmpOp F64:$src1, F64:$src2)),
2407 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
2408 (DoubleMI F64:$src1, F64:$src2))>;
2411 defm : unord_Pats <setuge, F2_sfcmpge, F2_dfcmpge>;
2412 defm : unord_Pats <setugt, F2_sfcmpgt, F2_dfcmpgt>;
2413 defm : unord_Pats <setueq, F2_sfcmpeq, F2_dfcmpeq>;
2415 //===----------------------------------------------------------------------===//
2416 // Multiclass to define 'Def Pats' for the following dags:
2417 // seteq(setoeq(op1, op2), 0) -> not(setoeq(op1, op2))
2418 // seteq(setoeq(op1, op2), 1) -> setoeq(op1, op2)
2419 // setne(setoeq(op1, op2), 0) -> setoeq(op1, op2)
2420 // setne(setoeq(op1, op2), 1) -> not(setoeq(op1, op2))
2421 //===----------------------------------------------------------------------===//
2422 let Predicates = [HasV5T] in
2423 multiclass eq_ordgePats <PatFrag cmpOp, InstHexagon IntMI,
2424 InstHexagon DoubleMI> {
2426 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
2427 (C2_not (IntMI F32:$src1, F32:$src2))>;
2428 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
2429 (IntMI F32:$src1, F32:$src2)>;
2430 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
2431 (IntMI F32:$src1, F32:$src2)>;
2432 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
2433 (C2_not (IntMI F32:$src1, F32:$src2))>;
2436 def : Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
2437 (C2_not (DoubleMI F64:$src1, F64:$src2))>;
2438 def : Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 1)),
2439 (DoubleMI F64:$src1, F64:$src2)>;
2440 def : Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
2441 (DoubleMI F64:$src1, F64:$src2)>;
2442 def : Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 1)),
2443 (C2_not (DoubleMI F64:$src1, F64:$src2))>;
2446 defm : eq_ordgePats<setoeq, F2_sfcmpeq, F2_dfcmpeq>;
2447 defm : eq_ordgePats<setoge, F2_sfcmpge, F2_dfcmpge>;
2448 defm : eq_ordgePats<setogt, F2_sfcmpgt, F2_dfcmpgt>;
2450 //===----------------------------------------------------------------------===//
2451 // Multiclass to define 'Def Pats' for the following dags:
2452 // seteq(setolt(op1, op2), 0) -> not(setogt(op2, op1))
2453 // seteq(setolt(op1, op2), 1) -> setogt(op2, op1)
2454 // setne(setolt(op1, op2), 0) -> setogt(op2, op1)
2455 // setne(setolt(op1, op2), 1) -> not(setogt(op2, op1))
2456 //===----------------------------------------------------------------------===//
2457 let Predicates = [HasV5T] in
2458 multiclass eq_ordltPats <PatFrag cmpOp, InstHexagon IntMI,
2459 InstHexagon DoubleMI> {
2461 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
2462 (C2_not (IntMI F32:$src2, F32:$src1))>;
2463 def: Pat<(i1 (seteq (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
2464 (IntMI F32:$src2, F32:$src1)>;
2465 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 0)),
2466 (IntMI F32:$src2, F32:$src1)>;
2467 def: Pat<(i1 (setne (i1 (cmpOp F32:$src1, F32:$src2)), 1)),
2468 (C2_not (IntMI F32:$src2, F32:$src1))>;
2471 def: Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
2472 (C2_not (DoubleMI F64:$src2, F64:$src1))>;
2473 def: Pat<(i1 (seteq (i1 (cmpOp F64:$src1, F64:$src2)), 1)),
2474 (DoubleMI F64:$src2, F64:$src1)>;
2475 def: Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
2476 (DoubleMI F64:$src2, F64:$src1)>;
2477 def: Pat<(i1 (setne (i1 (cmpOp F64:$src1, F64:$src2)), 0)),
2478 (C2_not (DoubleMI F64:$src2, F64:$src1))>;
2481 defm : eq_ordltPats<setole, F2_sfcmpge, F2_dfcmpge>;
2482 defm : eq_ordltPats<setolt, F2_sfcmpgt, F2_dfcmpgt>;
2485 // o. seto inverse of setuo. http://llvm.org/docs/LangRef.html#i_fcmp
2486 let Predicates = [HasV5T] in {
2487 def: Pat<(i1 (seto F32:$src1, F32:$src2)),
2488 (C2_not (F2_sfcmpuo F32:$src2, F32:$src1))>;
2489 def: Pat<(i1 (seto F32:$src1, f32ImmPred:$src2)),
2490 (C2_not (F2_sfcmpuo (f32 (A2_tfrsi (ftoi $src2))), F32:$src1))>;
2491 def: Pat<(i1 (seto F64:$src1, F64:$src2)),
2492 (C2_not (F2_dfcmpuo F64:$src2, F64:$src1))>;
2493 def: Pat<(i1 (seto F64:$src1, f64ImmPred:$src2)),
2494 (C2_not (F2_dfcmpuo (CONST64 (ftoi $src2)), F64:$src1))>;
2498 let Predicates = [HasV5T] in {
2499 def: Pat<(i1 (setolt F32:$src1, F32:$src2)),
2500 (F2_sfcmpgt F32:$src2, F32:$src1)>;
2501 def: Pat<(i1 (setolt F32:$src1, f32ImmPred:$src2)),
2502 (F2_sfcmpgt (f32 (A2_tfrsi (ftoi $src2))), F32:$src1)>;
2503 def: Pat<(i1 (setolt F64:$src1, F64:$src2)),
2504 (F2_dfcmpgt F64:$src2, F64:$src1)>;
2505 def: Pat<(i1 (setolt F64:$src1, f64ImmPred:$src2)),
2506 (F2_dfcmpgt (CONST64 (ftoi $src2)), F64:$src1)>;
2510 let Predicates = [HasV5T] in {
2511 def: Pat<(i1 (setult F32:$src1, F32:$src2)),
2512 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
2513 (F2_sfcmpgt F32:$src2, F32:$src1))>;
2514 def: Pat<(i1 (setult F32:$src1, f32ImmPred:$src2)),
2515 (C2_or (F2_sfcmpuo F32:$src1, (f32 (A2_tfrsi (ftoi $src2)))),
2516 (F2_sfcmpgt (f32 (A2_tfrsi (ftoi $src2))), F32:$src1))>;
2517 def: Pat<(i1 (setult F64:$src1, F64:$src2)),
2518 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
2519 (F2_dfcmpgt F64:$src2, F64:$src1))>;
2520 def: Pat<(i1 (setult F64:$src1, f64ImmPred:$src2)),
2521 (C2_or (F2_dfcmpuo F64:$src1, (CONST64 (ftoi $src2))),
2522 (F2_dfcmpgt (CONST64 (ftoi $src2)), F64:$src1))>;
2526 let Predicates = [HasV5T] in {
2527 // rs <= rt -> rt >= rs.
2528 def: Pat<(i1 (setole F32:$src1, F32:$src2)),
2529 (F2_sfcmpge F32:$src2, F32:$src1)>;
2530 def: Pat<(i1 (setole F32:$src1, f32ImmPred:$src2)),
2531 (F2_sfcmpge (f32 (A2_tfrsi (ftoi $src2))), F32:$src1)>;
2533 // Rss <= Rtt -> Rtt >= Rss.
2534 def: Pat<(i1 (setole F64:$src1, F64:$src2)),
2535 (F2_dfcmpge F64:$src2, F64:$src1)>;
2536 def: Pat<(i1 (setole F64:$src1, f64ImmPred:$src2)),
2537 (F2_dfcmpge (CONST64 (ftoi $src2)), F64:$src1)>;
2541 let Predicates = [HasV5T] in {
2542 // rs <= rt -> rt >= rs.
2543 def: Pat<(i1 (setule F32:$src1, F32:$src2)),
2544 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
2545 (F2_sfcmpge F32:$src2, F32:$src1))>;
2546 def: Pat<(i1 (setule F32:$src1, f32ImmPred:$src2)),
2547 (C2_or (F2_sfcmpuo F32:$src1, (f32 (A2_tfrsi (ftoi $src2)))),
2548 (F2_sfcmpge (f32 (A2_tfrsi (ftoi $src2))), F32:$src1))>;
2549 def: Pat<(i1 (setule F64:$src1, F64:$src2)),
2550 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
2551 (F2_dfcmpge F64:$src2, F64:$src1))>;
2552 def: Pat<(i1 (setule F64:$src1, f64ImmPred:$src2)),
2553 (C2_or (F2_dfcmpuo F64:$src1, (CONST64 (ftoi $src2))),
2554 (F2_dfcmpge (CONST64 (ftoi $src2)), F64:$src1))>;
2558 let Predicates = [HasV5T] in {
2559 def: Pat<(i1 (setone F32:$src1, F32:$src2)),
2560 (C2_not (F2_sfcmpeq F32:$src1, F32:$src2))>;
2561 def: Pat<(i1 (setone F64:$src1, F64:$src2)),
2562 (C2_not (F2_dfcmpeq F64:$src1, F64:$src2))>;
2563 def: Pat<(i1 (setone F32:$src1, f32ImmPred:$src2)),
2564 (C2_not (F2_sfcmpeq F32:$src1, (f32 (A2_tfrsi (ftoi $src2)))))>;
2565 def: Pat<(i1 (setone F64:$src1, f64ImmPred:$src2)),
2566 (C2_not (F2_dfcmpeq F64:$src1, (CONST64 (ftoi $src2))))>;
2570 let Predicates = [HasV5T] in {
2571 def: Pat<(i1 (setune F32:$src1, F32:$src2)),
2572 (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
2573 (C2_not (F2_sfcmpeq F32:$src1, F32:$src2)))>;
2574 def: Pat<(i1 (setune F64:$src1, F64:$src2)),
2575 (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
2576 (C2_not (F2_dfcmpeq F64:$src1, F64:$src2)))>;
2577 def: Pat<(i1 (setune F32:$src1, f32ImmPred:$src2)),
2578 (C2_or (F2_sfcmpuo F32:$src1, (f32 (A2_tfrsi (ftoi $src2)))),
2579 (C2_not (F2_sfcmpeq F32:$src1,
2580 (f32 (A2_tfrsi (ftoi $src2))))))>;
2581 def: Pat<(i1 (setune F64:$src1, f64ImmPred:$src2)),
2582 (C2_or (F2_dfcmpuo F64:$src1, (CONST64 (ftoi $src2))),
2583 (C2_not (F2_dfcmpeq F64:$src1,
2584 (CONST64 (ftoi $src2)))))>;
2587 // Besides set[o|u][comparions], we also need set[comparisons].
2588 let Predicates = [HasV5T] in {
2590 def: Pat<(i1 (setlt F32:$src1, F32:$src2)),
2591 (F2_sfcmpgt F32:$src2, F32:$src1)>;
2592 def: Pat<(i1 (setlt F32:$src1, f32ImmPred:$src2)),
2593 (F2_sfcmpgt (f32 (A2_tfrsi (ftoi $src2))), F32:$src1)>;
2594 def: Pat<(i1 (setlt F64:$src1, F64:$src2)),
2595 (F2_dfcmpgt F64:$src2, F64:$src1)>;
2596 def: Pat<(i1 (setlt F64:$src1, f64ImmPred:$src2)),
2597 (F2_dfcmpgt (CONST64 (ftoi $src2)), F64:$src1)>;
2600 // rs <= rt -> rt >= rs.
2601 def: Pat<(i1 (setle F32:$src1, F32:$src2)),
2602 (F2_sfcmpge F32:$src2, F32:$src1)>;
2603 def: Pat<(i1 (setle F32:$src1, f32ImmPred:$src2)),
2604 (F2_sfcmpge (f32 (A2_tfrsi (ftoi $src2))), F32:$src1)>;
2606 // Rss <= Rtt -> Rtt >= Rss.
2607 def: Pat<(i1 (setle F64:$src1, F64:$src2)),
2608 (F2_dfcmpge F64:$src2, F64:$src1)>;
2609 def: Pat<(i1 (setle F64:$src1, f64ImmPred:$src2)),
2610 (F2_dfcmpge (CONST64 (ftoi $src2)), F64:$src1)>;
2613 def: Pat<(i1 (setne F32:$src1, F32:$src2)),
2614 (C2_not (F2_sfcmpeq F32:$src1, F32:$src2))>;
2615 def: Pat<(i1 (setne F64:$src1, F64:$src2)),
2616 (C2_not (F2_dfcmpeq F64:$src1, F64:$src2))>;
2617 def: Pat<(i1 (setne F32:$src1, f32ImmPred:$src2)),
2618 (C2_not (F2_sfcmpeq F32:$src1, (f32 (A2_tfrsi (ftoi $src2)))))>;
2619 def: Pat<(i1 (setne F64:$src1, f64ImmPred:$src2)),
2620 (C2_not (F2_dfcmpeq F64:$src1, (CONST64 (ftoi $src2))))>;
2624 def: Pat<(f64 (fpextend F32:$Rs)), (F2_conv_sf2df F32:$Rs)>;
2625 def: Pat<(f32 (fpround F64:$Rs)), (F2_conv_df2sf F64:$Rs)>;
2627 def: Pat<(f32 (sint_to_fp I32:$Rs)), (F2_conv_w2sf I32:$Rs)>;
2628 def: Pat<(f32 (sint_to_fp I64:$Rs)), (F2_conv_d2sf I64:$Rs)>;
2629 def: Pat<(f64 (sint_to_fp I32:$Rs)), (F2_conv_w2df I32:$Rs)>;
2630 def: Pat<(f64 (sint_to_fp I64:$Rs)), (F2_conv_d2df I64:$Rs)>;
2632 def: Pat<(f32 (uint_to_fp I32:$Rs)), (F2_conv_uw2sf I32:$Rs)>;
2633 def: Pat<(f32 (uint_to_fp I64:$Rs)), (F2_conv_ud2sf I64:$Rs)>;
2634 def: Pat<(f64 (uint_to_fp I32:$Rs)), (F2_conv_uw2df I32:$Rs)>;
2635 def: Pat<(f64 (uint_to_fp I64:$Rs)), (F2_conv_ud2df I64:$Rs)>;
2637 def: Pat<(i32 (fp_to_sint F32:$Rs)), (F2_conv_sf2w_chop F32:$Rs)>;
2638 def: Pat<(i32 (fp_to_sint F64:$Rs)), (F2_conv_df2w_chop F64:$Rs)>;
2639 def: Pat<(i64 (fp_to_sint F32:$Rs)), (F2_conv_sf2d_chop F32:$Rs)>;
2640 def: Pat<(i64 (fp_to_sint F64:$Rs)), (F2_conv_df2d_chop F64:$Rs)>;
2642 def: Pat<(i32 (fp_to_uint F32:$Rs)), (F2_conv_sf2uw_chop F32:$Rs)>;
2643 def: Pat<(i32 (fp_to_uint F64:$Rs)), (F2_conv_df2uw_chop F64:$Rs)>;
2644 def: Pat<(i64 (fp_to_uint F32:$Rs)), (F2_conv_sf2ud_chop F32:$Rs)>;
2645 def: Pat<(i64 (fp_to_uint F64:$Rs)), (F2_conv_df2ud_chop F64:$Rs)>;
2647 // Bitcast is different than [fp|sint|uint]_to_[sint|uint|fp].
2648 let Predicates = [HasV5T] in {
2649 def: Pat <(i32 (bitconvert F32:$src)), (I32:$src)>;
2650 def: Pat <(f32 (bitconvert I32:$src)), (F32:$src)>;
2651 def: Pat <(i64 (bitconvert F64:$src)), (I64:$src)>;
2652 def: Pat <(f64 (bitconvert I64:$src)), (F64:$src)>;
2655 def : Pat <(fma F32:$src2, F32:$src3, F32:$src1),
2656 (F2_sffma F32:$src1, F32:$src2, F32:$src3)>;
2658 def : Pat <(fma (fneg F32:$src2), F32:$src3, F32:$src1),
2659 (F2_sffms F32:$src1, F32:$src2, F32:$src3)>;
2661 def : Pat <(fma F32:$src2, (fneg F32:$src3), F32:$src1),
2662 (F2_sffms F32:$src1, F32:$src2, F32:$src3)>;
2664 def: Pat<(select I1:$Pu, F32:$Rs, f32ImmPred:$imm),
2665 (C2_muxir I1:$Pu, F32:$Rs, (ftoi $imm))>,
2668 def: Pat<(select I1:$Pu, f32ImmPred:$imm, F32:$Rt),
2669 (C2_muxri I1:$Pu, (ftoi $imm), F32:$Rt)>,
2672 def: Pat<(select I1:$src1, F32:$src2, F32:$src3),
2673 (C2_mux I1:$src1, F32:$src2, F32:$src3)>,
2676 def: Pat<(select (i1 (setult F32:$src1, F32:$src2)), F32:$src3, F32:$src4),
2677 (C2_mux (F2_sfcmpgt F32:$src2, F32:$src1), F32:$src4, F32:$src3)>,
2680 def: Pat<(select I1:$src1, F64:$src2, F64:$src3),
2681 (C2_vmux I1:$src1, F64:$src2, F64:$src3)>,
2684 def: Pat<(select (i1 (setult F64:$src1, F64:$src2)), F64:$src3, F64:$src4),
2685 (C2_vmux (F2_dfcmpgt F64:$src2, F64:$src1), F64:$src3, F64:$src4)>,
2688 // Map from p0 = pnot(p0); r0 = select(p0, #i, r1)
2689 // => r0 = mux(p0, #i, r1)
2690 def: Pat<(select (not I1:$src1), f32ImmPred:$src2, F32:$src3),
2691 (C2_muxir I1:$src1, F32:$src3, (ftoi $src2))>,
2694 // Map from p0 = pnot(p0); r0 = mux(p0, r1, #i)
2695 // => r0 = mux(p0, r1, #i)
2696 def: Pat<(select (not I1:$src1), F32:$src2, f32ImmPred:$src3),
2697 (C2_muxri I1:$src1, (ftoi $src3), F32:$src2)>,
2700 def: Pat<(i32 (fp_to_sint F64:$src1)),
2701 (LoReg (F2_conv_df2d_chop F64:$src1))>,
2704 def : Pat <(fabs F32:$src1),
2705 (S2_clrbit_i F32:$src1, 31)>,
2708 def : Pat <(fneg F32:$src1),
2709 (S2_togglebit_i F32:$src1, 31)>,
2712 def: Pat<(fabs F64:$Rs),
2713 (REG_SEQUENCE DoubleRegs,
2714 (S2_clrbit_i (HiReg $Rs), 31), isub_hi,
2715 (i32 (LoReg $Rs)), isub_lo)>;
2717 def: Pat<(fneg F64:$Rs),
2718 (REG_SEQUENCE DoubleRegs,
2719 (S2_togglebit_i (HiReg $Rs), 31), isub_hi,
2720 (i32 (LoReg $Rs)), isub_lo)>;
2722 def alignedload : PatFrag<(ops node:$addr), (load $addr), [{
2723 return isAlignedMemNode(dyn_cast<MemSDNode>(N));
2726 def unalignedload : PatFrag<(ops node:$addr), (load $addr), [{
2727 return !isAlignedMemNode(dyn_cast<MemSDNode>(N));
2730 def alignedstore : PatFrag<(ops node:$val, node:$addr), (store $val, $addr), [{
2731 return isAlignedMemNode(dyn_cast<MemSDNode>(N));
2734 def unalignedstore : PatFrag<(ops node:$val, node:$addr), (store $val, $addr), [{
2735 return !isAlignedMemNode(dyn_cast<MemSDNode>(N));
2739 multiclass vS32b_ai_pats <ValueType VTSgl, ValueType VTDbl> {
2741 def : Pat<(alignedstore (VTSgl VectorRegs:$src1), IntRegs:$addr),
2742 (V6_vS32b_ai IntRegs:$addr, 0, (VTSgl VectorRegs:$src1))>,
2743 Requires<[UseHVXSgl]>;
2744 def : Pat<(unalignedstore (VTSgl VectorRegs:$src1), IntRegs:$addr),
2745 (V6_vS32Ub_ai IntRegs:$addr, 0, (VTSgl VectorRegs:$src1))>,
2746 Requires<[UseHVXSgl]>;
2748 // 128B Aligned stores
2749 def : Pat<(alignedstore (VTDbl VectorRegs128B:$src1), IntRegs:$addr),
2750 (V6_vS32b_ai_128B IntRegs:$addr, 0, (VTDbl VectorRegs128B:$src1))>,
2751 Requires<[UseHVXDbl]>;
2752 def : Pat<(unalignedstore (VTDbl VectorRegs128B:$src1), IntRegs:$addr),
2753 (V6_vS32Ub_ai_128B IntRegs:$addr, 0, (VTDbl VectorRegs128B:$src1))>,
2754 Requires<[UseHVXDbl]>;
2756 // Fold Add R+OFF into vector store.
2757 let AddedComplexity = 10 in {
2758 def : Pat<(alignedstore (VTSgl VectorRegs:$src1),
2759 (add IntRegs:$src2, Iss4_6:$offset)),
2760 (V6_vS32b_ai IntRegs:$src2, Iss4_6:$offset,
2761 (VTSgl VectorRegs:$src1))>,
2762 Requires<[UseHVXSgl]>;
2763 def : Pat<(unalignedstore (VTSgl VectorRegs:$src1),
2764 (add IntRegs:$src2, Iss4_6:$offset)),
2765 (V6_vS32Ub_ai IntRegs:$src2, Iss4_6:$offset,
2766 (VTSgl VectorRegs:$src1))>,
2767 Requires<[UseHVXSgl]>;
2769 // Fold Add R+OFF into vector store 128B.
2770 def : Pat<(alignedstore (VTDbl VectorRegs128B:$src1),
2771 (add IntRegs:$src2, Iss4_7:$offset)),
2772 (V6_vS32b_ai_128B IntRegs:$src2, Iss4_7:$offset,
2773 (VTDbl VectorRegs128B:$src1))>,
2774 Requires<[UseHVXDbl]>;
2775 def : Pat<(unalignedstore (VTDbl VectorRegs128B:$src1),
2776 (add IntRegs:$src2, Iss4_7:$offset)),
2777 (V6_vS32Ub_ai_128B IntRegs:$src2, Iss4_7:$offset,
2778 (VTDbl VectorRegs128B:$src1))>,
2779 Requires<[UseHVXDbl]>;
2783 defm : vS32b_ai_pats <v64i8, v128i8>;
2784 defm : vS32b_ai_pats <v32i16, v64i16>;
2785 defm : vS32b_ai_pats <v16i32, v32i32>;
2786 defm : vS32b_ai_pats <v8i64, v16i64>;
2789 multiclass vL32b_ai_pats <ValueType VTSgl, ValueType VTDbl> {
2791 def : Pat < (VTSgl (alignedload IntRegs:$addr)),
2792 (V6_vL32b_ai IntRegs:$addr, 0) >,
2793 Requires<[UseHVXSgl]>;
2794 def : Pat < (VTSgl (unalignedload IntRegs:$addr)),
2795 (V6_vL32Ub_ai IntRegs:$addr, 0) >,
2796 Requires<[UseHVXSgl]>;
2799 def : Pat < (VTDbl (alignedload IntRegs:$addr)),
2800 (V6_vL32b_ai_128B IntRegs:$addr, 0) >,
2801 Requires<[UseHVXDbl]>;
2802 def : Pat < (VTDbl (unalignedload IntRegs:$addr)),
2803 (V6_vL32Ub_ai_128B IntRegs:$addr, 0) >,
2804 Requires<[UseHVXDbl]>;
2806 // Fold Add R+OFF into vector load.
2807 let AddedComplexity = 10 in {
2808 def : Pat<(VTDbl (alignedload (add IntRegs:$src2, Iss4_7:$offset))),
2809 (V6_vL32b_ai_128B IntRegs:$src2, Iss4_7:$offset)>,
2810 Requires<[UseHVXDbl]>;
2811 def : Pat<(VTDbl (unalignedload (add IntRegs:$src2, Iss4_7:$offset))),
2812 (V6_vL32Ub_ai_128B IntRegs:$src2, Iss4_7:$offset)>,
2813 Requires<[UseHVXDbl]>;
2815 def : Pat<(VTSgl (alignedload (add IntRegs:$src2, Iss4_6:$offset))),
2816 (V6_vL32b_ai IntRegs:$src2, Iss4_6:$offset)>,
2817 Requires<[UseHVXSgl]>;
2818 def : Pat<(VTSgl (unalignedload (add IntRegs:$src2, Iss4_6:$offset))),
2819 (V6_vL32Ub_ai IntRegs:$src2, Iss4_6:$offset)>,
2820 Requires<[UseHVXSgl]>;
2824 defm : vL32b_ai_pats <v64i8, v128i8>;
2825 defm : vL32b_ai_pats <v32i16, v64i16>;
2826 defm : vL32b_ai_pats <v16i32, v32i32>;
2827 defm : vL32b_ai_pats <v8i64, v16i64>;
2829 multiclass STrivv_pats <ValueType VTSgl, ValueType VTDbl> {
2830 def : Pat<(alignedstore (VTSgl VecDblRegs:$src1), IntRegs:$addr),
2831 (PS_vstorerw_ai IntRegs:$addr, 0, (VTSgl VecDblRegs:$src1))>,
2832 Requires<[UseHVXSgl]>;
2833 def : Pat<(unalignedstore (VTSgl VecDblRegs:$src1), IntRegs:$addr),
2834 (PS_vstorerwu_ai IntRegs:$addr, 0, (VTSgl VecDblRegs:$src1))>,
2835 Requires<[UseHVXSgl]>;
2837 def : Pat<(alignedstore (VTDbl VecDblRegs128B:$src1), IntRegs:$addr),
2838 (PS_vstorerw_ai_128B IntRegs:$addr, 0,
2839 (VTDbl VecDblRegs128B:$src1))>,
2840 Requires<[UseHVXDbl]>;
2841 def : Pat<(unalignedstore (VTDbl VecDblRegs128B:$src1), IntRegs:$addr),
2842 (PS_vstorerwu_ai_128B IntRegs:$addr, 0,
2843 (VTDbl VecDblRegs128B:$src1))>,
2844 Requires<[UseHVXDbl]>;
2847 defm : STrivv_pats <v128i8, v256i8>;
2848 defm : STrivv_pats <v64i16, v128i16>;
2849 defm : STrivv_pats <v32i32, v64i32>;
2850 defm : STrivv_pats <v16i64, v32i64>;
2852 multiclass LDrivv_pats <ValueType VTSgl, ValueType VTDbl> {
2853 def : Pat<(VTSgl (alignedload I32:$addr)),
2854 (PS_vloadrw_ai I32:$addr, 0)>,
2855 Requires<[UseHVXSgl]>;
2856 def : Pat<(VTSgl (unalignedload I32:$addr)),
2857 (PS_vloadrwu_ai I32:$addr, 0)>,
2858 Requires<[UseHVXSgl]>;
2860 def : Pat<(VTDbl (alignedload I32:$addr)),
2861 (PS_vloadrw_ai_128B I32:$addr, 0)>,
2862 Requires<[UseHVXDbl]>;
2863 def : Pat<(VTDbl (unalignedload I32:$addr)),
2864 (PS_vloadrwu_ai_128B I32:$addr, 0)>,
2865 Requires<[UseHVXDbl]>;
2868 defm : LDrivv_pats <v128i8, v256i8>;
2869 defm : LDrivv_pats <v64i16, v128i16>;
2870 defm : LDrivv_pats <v32i32, v64i32>;
2871 defm : LDrivv_pats <v16i64, v32i64>;
2873 let Predicates = [HasV60T,UseHVXSgl] in {
2874 def: Pat<(select I1:$Pu, (v16i32 VectorRegs:$Vs), VectorRegs:$Vt),
2875 (PS_vselect I1:$Pu, VectorRegs:$Vs, VectorRegs:$Vt)>;
2876 def: Pat<(select I1:$Pu, (v32i32 VecDblRegs:$Vs), VecDblRegs:$Vt),
2877 (PS_wselect I1:$Pu, VecDblRegs:$Vs, VecDblRegs:$Vt)>;
2879 let Predicates = [HasV60T,UseHVXDbl] in {
2880 def: Pat<(select I1:$Pu, (v32i32 VectorRegs128B:$Vs), VectorRegs128B:$Vt),
2881 (PS_vselect_128B I1:$Pu, VectorRegs128B:$Vs, VectorRegs128B:$Vt)>;
2882 def: Pat<(select I1:$Pu, (v64i32 VecDblRegs128B:$Vs), VecDblRegs128B:$Vt),
2883 (PS_wselect_128B I1:$Pu, VecDblRegs128B:$Vs, VecDblRegs128B:$Vt)>;
2887 def SDTHexagonVCOMBINE: SDTypeProfile<1, 2, [SDTCisSameAs<1, 2>,
2888 SDTCisSubVecOfVec<1, 0>]>;
2890 def HexagonVCOMBINE: SDNode<"HexagonISD::VCOMBINE", SDTHexagonVCOMBINE>;
2892 def: Pat<(v32i32 (HexagonVCOMBINE (v16i32 VectorRegs:$Vs),
2893 (v16i32 VectorRegs:$Vt))),
2894 (V6_vcombine VectorRegs:$Vs, VectorRegs:$Vt)>,
2895 Requires<[UseHVXSgl]>;
2896 def: Pat<(v64i32 (HexagonVCOMBINE (v32i32 VecDblRegs:$Vs),
2897 (v32i32 VecDblRegs:$Vt))),
2898 (V6_vcombine_128B VecDblRegs:$Vs, VecDblRegs:$Vt)>,
2899 Requires<[UseHVXDbl]>;
2901 def SDTHexagonVPACK: SDTypeProfile<1, 3, [SDTCisSameAs<1, 2>,
2904 def HexagonVPACK: SDNode<"HexagonISD::VPACK", SDTHexagonVPACK>;
2906 // 0 as the last argument denotes vpacke. 1 denotes vpacko
2907 def: Pat<(v64i8 (HexagonVPACK (v64i8 VectorRegs:$Vs),
2908 (v64i8 VectorRegs:$Vt), (i32 0))),
2909 (V6_vpackeb VectorRegs:$Vs, VectorRegs:$Vt)>,
2910 Requires<[UseHVXSgl]>;
2911 def: Pat<(v64i8 (HexagonVPACK (v64i8 VectorRegs:$Vs),
2912 (v64i8 VectorRegs:$Vt), (i32 1))),
2913 (V6_vpackob VectorRegs:$Vs, VectorRegs:$Vt)>,
2914 Requires<[UseHVXSgl]>;
2915 def: Pat<(v32i16 (HexagonVPACK (v32i16 VectorRegs:$Vs),
2916 (v32i16 VectorRegs:$Vt), (i32 0))),
2917 (V6_vpackeh VectorRegs:$Vs, VectorRegs:$Vt)>,
2918 Requires<[UseHVXSgl]>;
2919 def: Pat<(v32i16 (HexagonVPACK (v32i16 VectorRegs:$Vs),
2920 (v32i16 VectorRegs:$Vt), (i32 1))),
2921 (V6_vpackoh VectorRegs:$Vs, VectorRegs:$Vt)>,
2922 Requires<[UseHVXSgl]>;
2924 def: Pat<(v128i8 (HexagonVPACK (v128i8 VecDblRegs:$Vs),
2925 (v128i8 VecDblRegs:$Vt), (i32 0))),
2926 (V6_vpackeb_128B VecDblRegs:$Vs, VecDblRegs:$Vt)>,
2927 Requires<[UseHVXDbl]>;
2928 def: Pat<(v128i8 (HexagonVPACK (v128i8 VecDblRegs:$Vs),
2929 (v128i8 VecDblRegs:$Vt), (i32 1))),
2930 (V6_vpackob_128B VecDblRegs:$Vs, VecDblRegs:$Vt)>,
2931 Requires<[UseHVXDbl]>;
2932 def: Pat<(v64i16 (HexagonVPACK (v64i16 VecDblRegs:$Vs),
2933 (v64i16 VecDblRegs:$Vt), (i32 0))),
2934 (V6_vpackeh_128B VecDblRegs:$Vs, VecDblRegs:$Vt)>,
2935 Requires<[UseHVXDbl]>;
2936 def: Pat<(v64i16 (HexagonVPACK (v64i16 VecDblRegs:$Vs),
2937 (v64i16 VecDblRegs:$Vt), (i32 1))),
2938 (V6_vpackoh_128B VecDblRegs:$Vs, VecDblRegs:$Vt)>,
2939 Requires<[UseHVXDbl]>;
2941 def V2I1: PatLeaf<(v2i1 PredRegs:$R)>;
2942 def V4I1: PatLeaf<(v4i1 PredRegs:$R)>;
2943 def V8I1: PatLeaf<(v8i1 PredRegs:$R)>;
2944 def V4I8: PatLeaf<(v4i8 IntRegs:$R)>;
2945 def V2I16: PatLeaf<(v2i16 IntRegs:$R)>;
2946 def V8I8: PatLeaf<(v8i8 DoubleRegs:$R)>;
2947 def V4I16: PatLeaf<(v4i16 DoubleRegs:$R)>;
2948 def V2I32: PatLeaf<(v2i32 DoubleRegs:$R)>;
2951 multiclass bitconvert_32<ValueType a, ValueType b> {
2952 def : Pat <(b (bitconvert (a IntRegs:$src))),
2954 def : Pat <(a (bitconvert (b IntRegs:$src))),
2958 multiclass bitconvert_64<ValueType a, ValueType b> {
2959 def : Pat <(b (bitconvert (a DoubleRegs:$src))),
2960 (b DoubleRegs:$src)>;
2961 def : Pat <(a (bitconvert (b DoubleRegs:$src))),
2962 (a DoubleRegs:$src)>;
2965 // Bit convert vector types to integers.
2966 defm : bitconvert_32<v4i8, i32>;
2967 defm : bitconvert_32<v2i16, i32>;
2968 defm : bitconvert_64<v8i8, i64>;
2969 defm : bitconvert_64<v4i16, i64>;
2970 defm : bitconvert_64<v2i32, i64>;
2972 def: Pat<(sra (v4i16 DoubleRegs:$src1), u4_0ImmPred:$src2),
2973 (S2_asr_i_vh DoubleRegs:$src1, imm:$src2)>;
2974 def: Pat<(srl (v4i16 DoubleRegs:$src1), u4_0ImmPred:$src2),
2975 (S2_lsr_i_vh DoubleRegs:$src1, imm:$src2)>;
2976 def: Pat<(shl (v4i16 DoubleRegs:$src1), u4_0ImmPred:$src2),
2977 (S2_asl_i_vh DoubleRegs:$src1, imm:$src2)>;
2979 def: Pat<(sra (v2i32 DoubleRegs:$src1), u5_0ImmPred:$src2),
2980 (S2_asr_i_vw DoubleRegs:$src1, imm:$src2)>;
2981 def: Pat<(srl (v2i32 DoubleRegs:$src1), u5_0ImmPred:$src2),
2982 (S2_lsr_i_vw DoubleRegs:$src1, imm:$src2)>;
2983 def: Pat<(shl (v2i32 DoubleRegs:$src1), u5_0ImmPred:$src2),
2984 (S2_asl_i_vw DoubleRegs:$src1, imm:$src2)>;
2986 def : Pat<(v2i16 (add (v2i16 IntRegs:$src1), (v2i16 IntRegs:$src2))),
2987 (A2_svaddh IntRegs:$src1, IntRegs:$src2)>;
2989 def : Pat<(v2i16 (sub (v2i16 IntRegs:$src1), (v2i16 IntRegs:$src2))),
2990 (A2_svsubh IntRegs:$src1, IntRegs:$src2)>;
2992 def HexagonVSPLATB: SDNode<"HexagonISD::VSPLATB", SDTUnaryOp>;
2993 def HexagonVSPLATH: SDNode<"HexagonISD::VSPLATH", SDTUnaryOp>;
2995 // Replicate the low 8-bits from 32-bits input register into each of the
2996 // four bytes of 32-bits destination register.
2997 def: Pat<(v4i8 (HexagonVSPLATB I32:$Rs)), (S2_vsplatrb I32:$Rs)>;
2999 // Replicate the low 16-bits from 32-bits input register into each of the
3000 // four halfwords of 64-bits destination register.
3001 def: Pat<(v4i16 (HexagonVSPLATH I32:$Rs)), (S2_vsplatrh I32:$Rs)>;
3004 class VArith_pat <InstHexagon MI, SDNode Op, PatFrag Type>
3005 : Pat <(Op Type:$Rss, Type:$Rtt),
3006 (MI Type:$Rss, Type:$Rtt)>;
3008 def: VArith_pat <A2_vaddub, add, V8I8>;
3009 def: VArith_pat <A2_vaddh, add, V4I16>;
3010 def: VArith_pat <A2_vaddw, add, V2I32>;
3011 def: VArith_pat <A2_vsubub, sub, V8I8>;
3012 def: VArith_pat <A2_vsubh, sub, V4I16>;
3013 def: VArith_pat <A2_vsubw, sub, V2I32>;
3015 def: VArith_pat <A2_and, and, V2I16>;
3016 def: VArith_pat <A2_xor, xor, V2I16>;
3017 def: VArith_pat <A2_or, or, V2I16>;
3019 def: VArith_pat <A2_andp, and, V8I8>;
3020 def: VArith_pat <A2_andp, and, V4I16>;
3021 def: VArith_pat <A2_andp, and, V2I32>;
3022 def: VArith_pat <A2_orp, or, V8I8>;
3023 def: VArith_pat <A2_orp, or, V4I16>;
3024 def: VArith_pat <A2_orp, or, V2I32>;
3025 def: VArith_pat <A2_xorp, xor, V8I8>;
3026 def: VArith_pat <A2_xorp, xor, V4I16>;
3027 def: VArith_pat <A2_xorp, xor, V2I32>;
3029 def: Pat<(v2i32 (sra V2I32:$b, (i64 (HexagonCOMBINE (i32 u5_0ImmPred:$c),
3030 (i32 u5_0ImmPred:$c))))),
3031 (S2_asr_i_vw V2I32:$b, imm:$c)>;
3032 def: Pat<(v2i32 (srl V2I32:$b, (i64 (HexagonCOMBINE (i32 u5_0ImmPred:$c),
3033 (i32 u5_0ImmPred:$c))))),
3034 (S2_lsr_i_vw V2I32:$b, imm:$c)>;
3035 def: Pat<(v2i32 (shl V2I32:$b, (i64 (HexagonCOMBINE (i32 u5_0ImmPred:$c),
3036 (i32 u5_0ImmPred:$c))))),
3037 (S2_asl_i_vw V2I32:$b, imm:$c)>;
3039 def: Pat<(v4i16 (sra V4I16:$b, (v4i16 (HexagonVSPLATH (i32 (u4_0ImmPred:$c)))))),
3040 (S2_asr_i_vh V4I16:$b, imm:$c)>;
3041 def: Pat<(v4i16 (srl V4I16:$b, (v4i16 (HexagonVSPLATH (i32 (u4_0ImmPred:$c)))))),
3042 (S2_lsr_i_vh V4I16:$b, imm:$c)>;
3043 def: Pat<(v4i16 (shl V4I16:$b, (v4i16 (HexagonVSPLATH (i32 (u4_0ImmPred:$c)))))),
3044 (S2_asl_i_vh V4I16:$b, imm:$c)>;
3047 def SDTHexagon_v2i32_v2i32_i32 : SDTypeProfile<1, 2,
3048 [SDTCisSameAs<0, 1>, SDTCisVT<0, v2i32>, SDTCisInt<2>]>;
3049 def SDTHexagon_v4i16_v4i16_i32 : SDTypeProfile<1, 2,
3050 [SDTCisSameAs<0, 1>, SDTCisVT<0, v4i16>, SDTCisInt<2>]>;
3052 def HexagonVSRAW: SDNode<"HexagonISD::VSRAW", SDTHexagon_v2i32_v2i32_i32>;
3053 def HexagonVSRAH: SDNode<"HexagonISD::VSRAH", SDTHexagon_v4i16_v4i16_i32>;
3054 def HexagonVSRLW: SDNode<"HexagonISD::VSRLW", SDTHexagon_v2i32_v2i32_i32>;
3055 def HexagonVSRLH: SDNode<"HexagonISD::VSRLH", SDTHexagon_v4i16_v4i16_i32>;
3056 def HexagonVSHLW: SDNode<"HexagonISD::VSHLW", SDTHexagon_v2i32_v2i32_i32>;
3057 def HexagonVSHLH: SDNode<"HexagonISD::VSHLH", SDTHexagon_v4i16_v4i16_i32>;
3059 def: Pat<(v2i32 (HexagonVSRAW V2I32:$Rs, u5_0ImmPred:$u5)),
3060 (S2_asr_i_vw V2I32:$Rs, imm:$u5)>;
3061 def: Pat<(v4i16 (HexagonVSRAH V4I16:$Rs, u4_0ImmPred:$u4)),
3062 (S2_asr_i_vh V4I16:$Rs, imm:$u4)>;
3063 def: Pat<(v2i32 (HexagonVSRLW V2I32:$Rs, u5_0ImmPred:$u5)),
3064 (S2_lsr_i_vw V2I32:$Rs, imm:$u5)>;
3065 def: Pat<(v4i16 (HexagonVSRLH V4I16:$Rs, u4_0ImmPred:$u4)),
3066 (S2_lsr_i_vh V4I16:$Rs, imm:$u4)>;
3067 def: Pat<(v2i32 (HexagonVSHLW V2I32:$Rs, u5_0ImmPred:$u5)),
3068 (S2_asl_i_vw V2I32:$Rs, imm:$u5)>;
3069 def: Pat<(v4i16 (HexagonVSHLH V4I16:$Rs, u4_0ImmPred:$u4)),
3070 (S2_asl_i_vh V4I16:$Rs, imm:$u4)>;
3072 class vshift_rr_pat<InstHexagon MI, SDNode Op, PatFrag Value>
3073 : Pat <(Op Value:$Rs, I32:$Rt),
3074 (MI Value:$Rs, I32:$Rt)>;
3076 def: vshift_rr_pat <S2_asr_r_vw, HexagonVSRAW, V2I32>;
3077 def: vshift_rr_pat <S2_asr_r_vh, HexagonVSRAH, V4I16>;
3078 def: vshift_rr_pat <S2_lsr_r_vw, HexagonVSRLW, V2I32>;
3079 def: vshift_rr_pat <S2_lsr_r_vh, HexagonVSRLH, V4I16>;
3080 def: vshift_rr_pat <S2_asl_r_vw, HexagonVSHLW, V2I32>;
3081 def: vshift_rr_pat <S2_asl_r_vh, HexagonVSHLH, V4I16>;
3084 def SDTHexagonVecCompare_v8i8 : SDTypeProfile<1, 2,
3085 [SDTCisSameAs<1, 2>, SDTCisVT<0, i1>, SDTCisVT<1, v8i8>]>;
3086 def SDTHexagonVecCompare_v4i16 : SDTypeProfile<1, 2,
3087 [SDTCisSameAs<1, 2>, SDTCisVT<0, i1>, SDTCisVT<1, v4i16>]>;
3088 def SDTHexagonVecCompare_v2i32 : SDTypeProfile<1, 2,
3089 [SDTCisSameAs<1, 2>, SDTCisVT<0, i1>, SDTCisVT<1, v2i32>]>;
3091 def HexagonVCMPBEQ: SDNode<"HexagonISD::VCMPBEQ", SDTHexagonVecCompare_v8i8>;
3092 def HexagonVCMPBGT: SDNode<"HexagonISD::VCMPBGT", SDTHexagonVecCompare_v8i8>;
3093 def HexagonVCMPBGTU: SDNode<"HexagonISD::VCMPBGTU", SDTHexagonVecCompare_v8i8>;
3094 def HexagonVCMPHEQ: SDNode<"HexagonISD::VCMPHEQ", SDTHexagonVecCompare_v4i16>;
3095 def HexagonVCMPHGT: SDNode<"HexagonISD::VCMPHGT", SDTHexagonVecCompare_v4i16>;
3096 def HexagonVCMPHGTU: SDNode<"HexagonISD::VCMPHGTU", SDTHexagonVecCompare_v4i16>;
3097 def HexagonVCMPWEQ: SDNode<"HexagonISD::VCMPWEQ", SDTHexagonVecCompare_v2i32>;
3098 def HexagonVCMPWGT: SDNode<"HexagonISD::VCMPWGT", SDTHexagonVecCompare_v2i32>;
3099 def HexagonVCMPWGTU: SDNode<"HexagonISD::VCMPWGTU", SDTHexagonVecCompare_v2i32>;
3102 class vcmp_i1_pat<InstHexagon MI, SDNode Op, PatFrag Value>
3103 : Pat <(i1 (Op Value:$Rs, Value:$Rt)),
3104 (MI Value:$Rs, Value:$Rt)>;
3106 def: vcmp_i1_pat<A2_vcmpbeq, HexagonVCMPBEQ, V8I8>;
3107 def: vcmp_i1_pat<A4_vcmpbgt, HexagonVCMPBGT, V8I8>;
3108 def: vcmp_i1_pat<A2_vcmpbgtu, HexagonVCMPBGTU, V8I8>;
3110 def: vcmp_i1_pat<A2_vcmpheq, HexagonVCMPHEQ, V4I16>;
3111 def: vcmp_i1_pat<A2_vcmphgt, HexagonVCMPHGT, V4I16>;
3112 def: vcmp_i1_pat<A2_vcmphgtu, HexagonVCMPHGTU, V4I16>;
3114 def: vcmp_i1_pat<A2_vcmpweq, HexagonVCMPWEQ, V2I32>;
3115 def: vcmp_i1_pat<A2_vcmpwgt, HexagonVCMPWGT, V2I32>;
3116 def: vcmp_i1_pat<A2_vcmpwgtu, HexagonVCMPWGTU, V2I32>;
3119 class vcmp_vi1_pat<InstHexagon MI, PatFrag Op, PatFrag InVal, ValueType OutTy>
3120 : Pat <(OutTy (Op InVal:$Rs, InVal:$Rt)),
3121 (MI InVal:$Rs, InVal:$Rt)>;
3123 def: vcmp_vi1_pat<A2_vcmpweq, seteq, V2I32, v2i1>;
3124 def: vcmp_vi1_pat<A2_vcmpwgt, setgt, V2I32, v2i1>;
3125 def: vcmp_vi1_pat<A2_vcmpwgtu, setugt, V2I32, v2i1>;
3127 def: vcmp_vi1_pat<A2_vcmpheq, seteq, V4I16, v4i1>;
3128 def: vcmp_vi1_pat<A2_vcmphgt, setgt, V4I16, v4i1>;
3129 def: vcmp_vi1_pat<A2_vcmphgtu, setugt, V4I16, v4i1>;
3131 def: Pat<(mul V2I32:$Rs, V2I32:$Rt),
3132 (PS_vmulw DoubleRegs:$Rs, DoubleRegs:$Rt)>;
3133 def: Pat<(add V2I32:$Rx, (mul V2I32:$Rs, V2I32:$Rt)),
3134 (PS_vmulw_acc DoubleRegs:$Rx, DoubleRegs:$Rs, DoubleRegs:$Rt)>;
3137 // Adds two v4i8: Hexagon does not have an insn for this one, so we
3138 // use the double add v8i8, and use only the low part of the result.
3139 def: Pat<(v4i8 (add (v4i8 IntRegs:$Rs), (v4i8 IntRegs:$Rt))),
3140 (LoReg (A2_vaddub (ToZext64 $Rs), (ToZext64 $Rt)))>;
3142 // Subtract two v4i8: Hexagon does not have an insn for this one, so we
3143 // use the double sub v8i8, and use only the low part of the result.
3144 def: Pat<(v4i8 (sub (v4i8 IntRegs:$Rs), (v4i8 IntRegs:$Rt))),
3145 (LoReg (A2_vsubub (ToZext64 $Rs), (ToZext64 $Rt)))>;
3148 // No 32 bit vector mux.
3150 def: Pat<(v4i8 (select I1:$Pu, V4I8:$Rs, V4I8:$Rt)),
3151 (LoReg (C2_vmux I1:$Pu, (ToZext64 $Rs), (ToZext64 $Rt)))>;
3152 def: Pat<(v2i16 (select I1:$Pu, V2I16:$Rs, V2I16:$Rt)),
3153 (LoReg (C2_vmux I1:$Pu, (ToZext64 $Rs), (ToZext64 $Rt)))>;
3156 // 64-bit vector mux.
3158 def: Pat<(v8i8 (vselect V8I1:$Pu, V8I8:$Rs, V8I8:$Rt)),
3159 (C2_vmux V8I1:$Pu, V8I8:$Rs, V8I8:$Rt)>;
3160 def: Pat<(v4i16 (vselect V4I1:$Pu, V4I16:$Rs, V4I16:$Rt)),
3161 (C2_vmux V4I1:$Pu, V4I16:$Rs, V4I16:$Rt)>;
3162 def: Pat<(v2i32 (vselect V2I1:$Pu, V2I32:$Rs, V2I32:$Rt)),
3163 (C2_vmux V2I1:$Pu, V2I32:$Rs, V2I32:$Rt)>;
3166 // No 32 bit vector compare.
3168 def: Pat<(i1 (seteq V4I8:$Rs, V4I8:$Rt)),
3169 (A2_vcmpbeq (ToZext64 $Rs), (ToZext64 $Rt))>;
3170 def: Pat<(i1 (setgt V4I8:$Rs, V4I8:$Rt)),
3171 (A4_vcmpbgt (ToZext64 $Rs), (ToZext64 $Rt))>;
3172 def: Pat<(i1 (setugt V4I8:$Rs, V4I8:$Rt)),
3173 (A2_vcmpbgtu (ToZext64 $Rs), (ToZext64 $Rt))>;
3175 def: Pat<(i1 (seteq V2I16:$Rs, V2I16:$Rt)),
3176 (A2_vcmpheq (ToZext64 $Rs), (ToZext64 $Rt))>;
3177 def: Pat<(i1 (setgt V2I16:$Rs, V2I16:$Rt)),
3178 (A2_vcmphgt (ToZext64 $Rs), (ToZext64 $Rt))>;
3179 def: Pat<(i1 (setugt V2I16:$Rs, V2I16:$Rt)),
3180 (A2_vcmphgtu (ToZext64 $Rs), (ToZext64 $Rt))>;
3183 class InvertCmp_pat<InstHexagon InvMI, PatFrag CmpOp, PatFrag Value,
3185 : Pat<(CmpTy (CmpOp Value:$Rs, Value:$Rt)),
3186 (InvMI Value:$Rt, Value:$Rs)>;
3188 // Map from a compare operation to the corresponding instruction with the
3189 // order of operands reversed, e.g. x > y --> cmp.lt(y,x).
3190 def: InvertCmp_pat<A4_vcmpbgt, setlt, V8I8, i1>;
3191 def: InvertCmp_pat<A4_vcmpbgt, setlt, V8I8, v8i1>;
3192 def: InvertCmp_pat<A2_vcmphgt, setlt, V4I16, i1>;
3193 def: InvertCmp_pat<A2_vcmphgt, setlt, V4I16, v4i1>;
3194 def: InvertCmp_pat<A2_vcmpwgt, setlt, V2I32, i1>;
3195 def: InvertCmp_pat<A2_vcmpwgt, setlt, V2I32, v2i1>;
3197 def: InvertCmp_pat<A2_vcmpbgtu, setult, V8I8, i1>;
3198 def: InvertCmp_pat<A2_vcmpbgtu, setult, V8I8, v8i1>;
3199 def: InvertCmp_pat<A2_vcmphgtu, setult, V4I16, i1>;
3200 def: InvertCmp_pat<A2_vcmphgtu, setult, V4I16, v4i1>;
3201 def: InvertCmp_pat<A2_vcmpwgtu, setult, V2I32, i1>;
3202 def: InvertCmp_pat<A2_vcmpwgtu, setult, V2I32, v2i1>;
3204 // Map from vcmpne(Rss) -> !vcmpew(Rss).
3205 // rs != rt -> !(rs == rt).
3206 def: Pat<(v2i1 (setne V2I32:$Rs, V2I32:$Rt)),
3207 (C2_not (v2i1 (A2_vcmpbeq V2I32:$Rs, V2I32:$Rt)))>;
3210 // Truncate: from vector B copy all 'E'ven 'B'yte elements:
3211 // A[0] = B[0]; A[1] = B[2]; A[2] = B[4]; A[3] = B[6];
3212 def: Pat<(v4i8 (trunc V4I16:$Rs)),
3213 (S2_vtrunehb V4I16:$Rs)>;
3215 // Truncate: from vector B copy all 'O'dd 'B'yte elements:
3216 // A[0] = B[1]; A[1] = B[3]; A[2] = B[5]; A[3] = B[7];
3219 // Truncate: from vectors B and C copy all 'E'ven 'H'alf-word elements:
3220 // A[0] = B[0]; A[1] = B[2]; A[2] = C[0]; A[3] = C[2];
3223 def: Pat<(v2i16 (trunc V2I32:$Rs)),
3224 (LoReg (S2_packhl (HiReg $Rs), (LoReg $Rs)))>;
3227 def HexagonVSXTBH : SDNode<"HexagonISD::VSXTBH", SDTUnaryOp>;
3228 def HexagonVSXTBW : SDNode<"HexagonISD::VSXTBW", SDTUnaryOp>;
3230 def: Pat<(i64 (HexagonVSXTBH I32:$Rs)), (S2_vsxtbh I32:$Rs)>;
3231 def: Pat<(i64 (HexagonVSXTBW I32:$Rs)), (S2_vsxthw I32:$Rs)>;
3233 def: Pat<(v4i16 (zext V4I8:$Rs)), (S2_vzxtbh V4I8:$Rs)>;
3234 def: Pat<(v2i32 (zext V2I16:$Rs)), (S2_vzxthw V2I16:$Rs)>;
3235 def: Pat<(v4i16 (anyext V4I8:$Rs)), (S2_vzxtbh V4I8:$Rs)>;
3236 def: Pat<(v2i32 (anyext V2I16:$Rs)), (S2_vzxthw V2I16:$Rs)>;
3237 def: Pat<(v4i16 (sext V4I8:$Rs)), (S2_vsxtbh V4I8:$Rs)>;
3238 def: Pat<(v2i32 (sext V2I16:$Rs)), (S2_vsxthw V2I16:$Rs)>;
3240 // Sign extends a v2i8 into a v2i32.
3241 def: Pat<(v2i32 (sext_inreg V2I32:$Rs, v2i8)),
3242 (A2_combinew (A2_sxtb (HiReg $Rs)), (A2_sxtb (LoReg $Rs)))>;
3244 // Sign extends a v2i16 into a v2i32.
3245 def: Pat<(v2i32 (sext_inreg V2I32:$Rs, v2i16)),
3246 (A2_combinew (A2_sxth (HiReg $Rs)), (A2_sxth (LoReg $Rs)))>;
3249 // Multiplies two v2i16 and returns a v2i32. We are using here the
3250 // saturating multiply, as hexagon does not provide a non saturating
3251 // vector multiply, and saturation does not impact the result that is
3252 // in double precision of the operands.
3254 // Multiplies two v2i16 vectors: as Hexagon does not have a multiply
3255 // with the C semantics for this one, this pattern uses the half word
3256 // multiply vmpyh that takes two v2i16 and returns a v2i32. This is
3257 // then truncated to fit this back into a v2i16 and to simulate the
3258 // wrap around semantics for unsigned in C.
3259 def vmpyh: OutPatFrag<(ops node:$Rs, node:$Rt),
3260 (M2_vmpy2s_s0 (i32 $Rs), (i32 $Rt))>;
3262 def: Pat<(v2i16 (mul V2I16:$Rs, V2I16:$Rt)),
3263 (LoReg (S2_vtrunewh (A2_combineii 0, 0),
3264 (vmpyh V2I16:$Rs, V2I16:$Rt)))>;
3266 // Multiplies two v4i16 vectors.
3267 def: Pat<(v4i16 (mul V4I16:$Rs, V4I16:$Rt)),
3268 (S2_vtrunewh (vmpyh (HiReg $Rs), (HiReg $Rt)),
3269 (vmpyh (LoReg $Rs), (LoReg $Rt)))>;
3271 def VMPYB_no_V5: OutPatFrag<(ops node:$Rs, node:$Rt),
3272 (S2_vtrunewh (vmpyh (HiReg (S2_vsxtbh $Rs)), (HiReg (S2_vsxtbh $Rt))),
3273 (vmpyh (LoReg (S2_vsxtbh $Rs)), (LoReg (S2_vsxtbh $Rt))))>;
3275 // Multiplies two v4i8 vectors.
3276 def: Pat<(v4i8 (mul V4I8:$Rs, V4I8:$Rt)),
3277 (S2_vtrunehb (M5_vmpybsu V4I8:$Rs, V4I8:$Rt))>,
3280 def: Pat<(v4i8 (mul V4I8:$Rs, V4I8:$Rt)),
3281 (S2_vtrunehb (VMPYB_no_V5 V4I8:$Rs, V4I8:$Rt))>;
3283 // Multiplies two v8i8 vectors.
3284 def: Pat<(v8i8 (mul V8I8:$Rs, V8I8:$Rt)),
3285 (A2_combinew (S2_vtrunehb (M5_vmpybsu (HiReg $Rs), (HiReg $Rt))),
3286 (S2_vtrunehb (M5_vmpybsu (LoReg $Rs), (LoReg $Rt))))>,
3289 def: Pat<(v8i8 (mul V8I8:$Rs, V8I8:$Rt)),
3290 (A2_combinew (S2_vtrunehb (VMPYB_no_V5 (HiReg $Rs), (HiReg $Rt))),
3291 (S2_vtrunehb (VMPYB_no_V5 (LoReg $Rs), (LoReg $Rt))))>;
3293 def SDTHexagonBinOp64 : SDTypeProfile<1, 2,
3294 [SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>, SDTCisVT<0, i64>]>;
3296 def HexagonSHUFFEB: SDNode<"HexagonISD::SHUFFEB", SDTHexagonBinOp64>;
3297 def HexagonSHUFFEH: SDNode<"HexagonISD::SHUFFEH", SDTHexagonBinOp64>;
3298 def HexagonSHUFFOB: SDNode<"HexagonISD::SHUFFOB", SDTHexagonBinOp64>;
3299 def HexagonSHUFFOH: SDNode<"HexagonISD::SHUFFOH", SDTHexagonBinOp64>;
3301 class ShufflePat<InstHexagon MI, SDNode Op>
3302 : Pat<(i64 (Op DoubleRegs:$src1, DoubleRegs:$src2)),
3303 (i64 (MI DoubleRegs:$src1, DoubleRegs:$src2))>;
3305 // Shuffles even bytes for i=0..3: A[2*i].b = C[2*i].b; A[2*i+1].b = B[2*i].b
3306 def: ShufflePat<S2_shuffeb, HexagonSHUFFEB>;
3308 // Shuffles odd bytes for i=0..3: A[2*i].b = C[2*i+1].b; A[2*i+1].b = B[2*i+1].b
3309 def: ShufflePat<S2_shuffob, HexagonSHUFFOB>;
3311 // Shuffles even half for i=0,1: A[2*i].h = C[2*i].h; A[2*i+1].h = B[2*i].h
3312 def: ShufflePat<S2_shuffeh, HexagonSHUFFEH>;
3314 // Shuffles odd half for i=0,1: A[2*i].h = C[2*i+1].h; A[2*i+1].h = B[2*i+1].h
3315 def: ShufflePat<S2_shuffoh, HexagonSHUFFOH>;
3318 // Truncated store from v4i16 to v4i8.
3319 def truncstorev4i8: PatFrag<(ops node:$val, node:$ptr),
3320 (truncstore node:$val, node:$ptr),
3321 [{ return cast<StoreSDNode>(N)->getMemoryVT() == MVT::v4i8; }]>;
3323 // Truncated store from v2i32 to v2i16.
3324 def truncstorev2i16: PatFrag<(ops node:$val, node:$ptr),
3325 (truncstore node:$val, node:$ptr),
3326 [{ return cast<StoreSDNode>(N)->getMemoryVT() == MVT::v2i16; }]>;
3328 def: Pat<(truncstorev2i16 V2I32:$Rs, I32:$Rt),
3329 (S2_storeri_io I32:$Rt, 0, (LoReg (S2_packhl (HiReg $Rs),
3332 def: Pat<(truncstorev4i8 V4I16:$Rs, I32:$Rt),
3333 (S2_storeri_io I32:$Rt, 0, (S2_vtrunehb V4I16:$Rs))>;
3336 // Zero and sign extended load from v2i8 into v2i16.
3337 def zextloadv2i8: PatFrag<(ops node:$ptr), (zextload node:$ptr),
3338 [{ return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8; }]>;
3340 def sextloadv2i8: PatFrag<(ops node:$ptr), (sextload node:$ptr),
3341 [{ return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8; }]>;
3343 def: Pat<(v2i16 (zextloadv2i8 I32:$Rs)),
3344 (LoReg (v4i16 (S2_vzxtbh (L2_loadruh_io I32:$Rs, 0))))>;
3346 def: Pat<(v2i16 (sextloadv2i8 I32:$Rs)),
3347 (LoReg (v4i16 (S2_vsxtbh (L2_loadrh_io I32:$Rs, 0))))>;
3349 def: Pat<(v2i32 (zextloadv2i8 I32:$Rs)),
3350 (S2_vzxthw (LoReg (v4i16 (S2_vzxtbh (L2_loadruh_io I32:$Rs, 0)))))>;
3352 def: Pat<(v2i32 (sextloadv2i8 I32:$Rs)),
3353 (S2_vsxthw (LoReg (v4i16 (S2_vsxtbh (L2_loadrh_io I32:$Rs, 0)))))>;
3356 // Read cycle counter.
3358 def SDTInt64Leaf: SDTypeProfile<1, 0, [SDTCisVT<0, i64>]>;
3359 def HexagonREADCYCLE: SDNode<"HexagonISD::READCYCLE", SDTInt64Leaf,
3362 def: Pat<(HexagonREADCYCLE), (A4_tfrcpp UPCYCLE)>;