1 //==- SPUInstrInfo.td - Describe the Cell SPU Instructions -*- tablegen -*-==//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
9 // Cell SPU Instructions:
10 //===----------------------------------------------------------------------===//
12 //===----------------------------------------------------------------------===//
13 // TODO Items (not urgent today, but would be nice, low priority)
15 // ANDBI, ORBI: SPU constructs a 4-byte constant for these instructions by
16 // concatenating the byte argument b as "bbbb". Could recognize this bit pattern
17 // in 16-bit and 32-bit constants and reduce instruction count.
18 //===----------------------------------------------------------------------===//
20 //===----------------------------------------------------------------------===//
21 // Pseudo instructions:
22 //===----------------------------------------------------------------------===//
24 let hasCtrlDep = 1, Defs = [R1], Uses = [R1] in {
25 def ADJCALLSTACKDOWN : Pseudo<(outs), (ins u16imm_i32:$amt),
26 "${:comment} ADJCALLSTACKDOWN",
27 [(callseq_start timm:$amt)]>;
28 def ADJCALLSTACKUP : Pseudo<(outs), (ins u16imm_i32:$amt),
29 "${:comment} ADJCALLSTACKUP",
30 [(callseq_end timm:$amt)]>;
33 //===----------------------------------------------------------------------===//
35 // NB: The ordering is actually important, since the instruction selection
36 // will try each of the instructions in sequence, i.e., the D-form first with
37 // the 10-bit displacement, then the A-form with the 16 bit displacement, and
38 // finally the X-form with the register-register.
39 //===----------------------------------------------------------------------===//
41 let canFoldAsLoad = 1 in {
42 class LoadDFormVec<ValueType vectype>
43 : RI10Form<0b00101100, (outs VECREG:$rT), (ins dformaddr:$src),
46 [(set (vectype VECREG:$rT), (load dform_addr:$src))]>
49 class LoadDForm<RegisterClass rclass>
50 : RI10Form<0b00101100, (outs rclass:$rT), (ins dformaddr:$src),
53 [(set rclass:$rT, (load dform_addr:$src))]>
58 def v16i8: LoadDFormVec<v16i8>;
59 def v8i16: LoadDFormVec<v8i16>;
60 def v4i32: LoadDFormVec<v4i32>;
61 def v2i64: LoadDFormVec<v2i64>;
62 def v4f32: LoadDFormVec<v4f32>;
63 def v2f64: LoadDFormVec<v2f64>;
65 def r128: LoadDForm<GPRC>;
66 def r64: LoadDForm<R64C>;
67 def r32: LoadDForm<R32C>;
68 def f32: LoadDForm<R32FP>;
69 def f64: LoadDForm<R64FP>;
70 def r16: LoadDForm<R16C>;
71 def r8: LoadDForm<R8C>;
74 class LoadAFormVec<ValueType vectype>
75 : RI16Form<0b100001100, (outs VECREG:$rT), (ins addr256k:$src),
78 [(set (vectype VECREG:$rT), (load aform_addr:$src))]>
81 class LoadAForm<RegisterClass rclass>
82 : RI16Form<0b100001100, (outs rclass:$rT), (ins addr256k:$src),
85 [(set rclass:$rT, (load aform_addr:$src))]>
90 def v16i8: LoadAFormVec<v16i8>;
91 def v8i16: LoadAFormVec<v8i16>;
92 def v4i32: LoadAFormVec<v4i32>;
93 def v2i64: LoadAFormVec<v2i64>;
94 def v4f32: LoadAFormVec<v4f32>;
95 def v2f64: LoadAFormVec<v2f64>;
97 def r128: LoadAForm<GPRC>;
98 def r64: LoadAForm<R64C>;
99 def r32: LoadAForm<R32C>;
100 def f32: LoadAForm<R32FP>;
101 def f64: LoadAForm<R64FP>;
102 def r16: LoadAForm<R16C>;
103 def r8: LoadAForm<R8C>;
106 class LoadXFormVec<ValueType vectype>
107 : RRForm<0b00100011100, (outs VECREG:$rT), (ins memrr:$src),
110 [(set (vectype VECREG:$rT), (load xform_addr:$src))]>
113 class LoadXForm<RegisterClass rclass>
114 : RRForm<0b00100011100, (outs rclass:$rT), (ins memrr:$src),
117 [(set rclass:$rT, (load xform_addr:$src))]>
120 multiclass LoadXForms
122 def v16i8: LoadXFormVec<v16i8>;
123 def v8i16: LoadXFormVec<v8i16>;
124 def v4i32: LoadXFormVec<v4i32>;
125 def v2i64: LoadXFormVec<v2i64>;
126 def v4f32: LoadXFormVec<v4f32>;
127 def v2f64: LoadXFormVec<v2f64>;
129 def r128: LoadXForm<GPRC>;
130 def r64: LoadXForm<R64C>;
131 def r32: LoadXForm<R32C>;
132 def f32: LoadXForm<R32FP>;
133 def f64: LoadXForm<R64FP>;
134 def r16: LoadXForm<R16C>;
135 def r8: LoadXForm<R8C>;
138 defm LQA : LoadAForms;
139 defm LQD : LoadDForms;
140 defm LQX : LoadXForms;
142 /* Load quadword, PC relative: Not much use at this point in time.
143 Might be of use later for relocatable code. It's effectively the
144 same as LQA, but uses PC-relative addressing.
145 def LQR : RI16Form<0b111001100, (outs VECREG:$rT), (ins s16imm:$disp),
146 "lqr\t$rT, $disp", LoadStore,
147 [(set VECREG:$rT, (load iaddr:$disp))]>;
151 //===----------------------------------------------------------------------===//
153 //===----------------------------------------------------------------------===//
154 class StoreDFormVec<ValueType vectype>
155 : RI10Form<0b00100100, (outs), (ins VECREG:$rT, dformaddr:$src),
158 [(store (vectype VECREG:$rT), dform_addr:$src)]>
161 class StoreDForm<RegisterClass rclass>
162 : RI10Form<0b00100100, (outs), (ins rclass:$rT, dformaddr:$src),
165 [(store rclass:$rT, dform_addr:$src)]>
168 multiclass StoreDForms
170 def v16i8: StoreDFormVec<v16i8>;
171 def v8i16: StoreDFormVec<v8i16>;
172 def v4i32: StoreDFormVec<v4i32>;
173 def v2i64: StoreDFormVec<v2i64>;
174 def v4f32: StoreDFormVec<v4f32>;
175 def v2f64: StoreDFormVec<v2f64>;
177 def r128: StoreDForm<GPRC>;
178 def r64: StoreDForm<R64C>;
179 def r32: StoreDForm<R32C>;
180 def f32: StoreDForm<R32FP>;
181 def f64: StoreDForm<R64FP>;
182 def r16: StoreDForm<R16C>;
183 def r8: StoreDForm<R8C>;
186 class StoreAFormVec<ValueType vectype>
187 : RI16Form<0b0010010, (outs), (ins VECREG:$rT, addr256k:$src),
190 [(store (vectype VECREG:$rT), aform_addr:$src)]>;
192 class StoreAForm<RegisterClass rclass>
193 : RI16Form<0b001001, (outs), (ins rclass:$rT, addr256k:$src),
196 [(store rclass:$rT, aform_addr:$src)]>;
198 multiclass StoreAForms
200 def v16i8: StoreAFormVec<v16i8>;
201 def v8i16: StoreAFormVec<v8i16>;
202 def v4i32: StoreAFormVec<v4i32>;
203 def v2i64: StoreAFormVec<v2i64>;
204 def v4f32: StoreAFormVec<v4f32>;
205 def v2f64: StoreAFormVec<v2f64>;
207 def r128: StoreAForm<GPRC>;
208 def r64: StoreAForm<R64C>;
209 def r32: StoreAForm<R32C>;
210 def f32: StoreAForm<R32FP>;
211 def f64: StoreAForm<R64FP>;
212 def r16: StoreAForm<R16C>;
213 def r8: StoreAForm<R8C>;
216 class StoreXFormVec<ValueType vectype>
217 : RRForm<0b00100100, (outs), (ins VECREG:$rT, memrr:$src),
220 [(store (vectype VECREG:$rT), xform_addr:$src)]>
223 class StoreXForm<RegisterClass rclass>
224 : RRForm<0b00100100, (outs), (ins rclass:$rT, memrr:$src),
227 [(store rclass:$rT, xform_addr:$src)]>
230 multiclass StoreXForms
232 def v16i8: StoreXFormVec<v16i8>;
233 def v8i16: StoreXFormVec<v8i16>;
234 def v4i32: StoreXFormVec<v4i32>;
235 def v2i64: StoreXFormVec<v2i64>;
236 def v4f32: StoreXFormVec<v4f32>;
237 def v2f64: StoreXFormVec<v2f64>;
239 def r128: StoreXForm<GPRC>;
240 def r64: StoreXForm<R64C>;
241 def r32: StoreXForm<R32C>;
242 def f32: StoreXForm<R32FP>;
243 def f64: StoreXForm<R64FP>;
244 def r16: StoreXForm<R16C>;
245 def r8: StoreXForm<R8C>;
248 defm STQD : StoreDForms;
249 defm STQA : StoreAForms;
250 defm STQX : StoreXForms;
252 /* Store quadword, PC relative: Not much use at this point in time. Might
253 be useful for relocatable code.
254 def STQR : RI16Form<0b111000100, (outs), (ins VECREG:$rT, s16imm:$disp),
255 "stqr\t$rT, $disp", LoadStore,
256 [(store VECREG:$rT, iaddr:$disp)]>;
259 //===----------------------------------------------------------------------===//
260 // Generate Controls for Insertion:
261 //===----------------------------------------------------------------------===//
263 def CBD: RI7Form<0b10101111100, (outs VECREG:$rT), (ins shufaddr:$src),
264 "cbd\t$rT, $src", ShuffleOp,
265 [(set (v16i8 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
267 def CBX: RRForm<0b00101011100, (outs VECREG:$rT), (ins memrr:$src),
268 "cbx\t$rT, $src", ShuffleOp,
269 [(set (v16i8 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
271 def CHD: RI7Form<0b10101111100, (outs VECREG:$rT), (ins shufaddr:$src),
272 "chd\t$rT, $src", ShuffleOp,
273 [(set (v8i16 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
275 def CHX: RRForm<0b10101011100, (outs VECREG:$rT), (ins memrr:$src),
276 "chx\t$rT, $src", ShuffleOp,
277 [(set (v8i16 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
279 def CWD: RI7Form<0b01101111100, (outs VECREG:$rT), (ins shufaddr:$src),
280 "cwd\t$rT, $src", ShuffleOp,
281 [(set (v4i32 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
283 def CWX: RRForm<0b01101011100, (outs VECREG:$rT), (ins memrr:$src),
284 "cwx\t$rT, $src", ShuffleOp,
285 [(set (v4i32 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
287 def CWDf32: RI7Form<0b01101111100, (outs VECREG:$rT), (ins shufaddr:$src),
288 "cwd\t$rT, $src", ShuffleOp,
289 [(set (v4f32 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
291 def CWXf32: RRForm<0b01101011100, (outs VECREG:$rT), (ins memrr:$src),
292 "cwx\t$rT, $src", ShuffleOp,
293 [(set (v4f32 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
295 def CDD: RI7Form<0b11101111100, (outs VECREG:$rT), (ins shufaddr:$src),
296 "cdd\t$rT, $src", ShuffleOp,
297 [(set (v2i64 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
299 def CDX: RRForm<0b11101011100, (outs VECREG:$rT), (ins memrr:$src),
300 "cdx\t$rT, $src", ShuffleOp,
301 [(set (v2i64 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
303 def CDDf64: RI7Form<0b11101111100, (outs VECREG:$rT), (ins shufaddr:$src),
304 "cdd\t$rT, $src", ShuffleOp,
305 [(set (v2f64 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
307 def CDXf64: RRForm<0b11101011100, (outs VECREG:$rT), (ins memrr:$src),
308 "cdx\t$rT, $src", ShuffleOp,
309 [(set (v2f64 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
311 //===----------------------------------------------------------------------===//
312 // Constant formation:
313 //===----------------------------------------------------------------------===//
316 RI16Form<0b110000010, (outs VECREG:$rT), (ins s16imm:$val),
317 "ilh\t$rT, $val", ImmLoad,
318 [(set (v8i16 VECREG:$rT), (v8i16 v8i16SExt16Imm:$val))]>;
321 RI16Form<0b110000010, (outs R16C:$rT), (ins s16imm:$val),
322 "ilh\t$rT, $val", ImmLoad,
323 [(set R16C:$rT, immSExt16:$val)]>;
325 // Cell SPU doesn't have a native 8-bit immediate load, but ILH works ("with
326 // the right constant")
328 RI16Form<0b110000010, (outs R8C:$rT), (ins s16imm_i8:$val),
329 "ilh\t$rT, $val", ImmLoad,
330 [(set R8C:$rT, immSExt8:$val)]>;
332 // IL does sign extension!
334 class ILInst<dag OOL, dag IOL, list<dag> pattern>:
335 RI16Form<0b100000010, OOL, IOL, "il\t$rT, $val",
338 class ILVecInst<ValueType vectype, Operand immtype, PatLeaf xform>:
339 ILInst<(outs VECREG:$rT), (ins immtype:$val),
340 [(set (vectype VECREG:$rT), (vectype xform:$val))]>;
342 class ILRegInst<RegisterClass rclass, Operand immtype, PatLeaf xform>:
343 ILInst<(outs rclass:$rT), (ins immtype:$val),
344 [(set rclass:$rT, xform:$val)]>;
346 multiclass ImmediateLoad
348 def v2i64: ILVecInst<v2i64, s16imm_i64, v2i64SExt16Imm>;
349 def v4i32: ILVecInst<v4i32, s16imm_i32, v4i32SExt16Imm>;
351 // TODO: Need v2f64, v4f32
353 def r64: ILRegInst<R64C, s16imm_i64, immSExt16>;
354 def r32: ILRegInst<R32C, s16imm_i32, immSExt16>;
355 def f32: ILRegInst<R32FP, s16imm_f32, fpimmSExt16>;
356 def f64: ILRegInst<R64FP, s16imm_f64, fpimmSExt16>;
359 defm IL : ImmediateLoad;
361 class ILHUInst<dag OOL, dag IOL, list<dag> pattern>:
362 RI16Form<0b010000010, OOL, IOL, "ilhu\t$rT, $val",
365 class ILHUVecInst<ValueType vectype, Operand immtype, PatLeaf xform>:
366 ILHUInst<(outs VECREG:$rT), (ins immtype:$val),
367 [(set (vectype VECREG:$rT), (vectype xform:$val))]>;
369 class ILHURegInst<RegisterClass rclass, Operand immtype, PatLeaf xform>:
370 ILHUInst<(outs rclass:$rT), (ins immtype:$val),
371 [(set rclass:$rT, xform:$val)]>;
373 multiclass ImmLoadHalfwordUpper
375 def v2i64: ILHUVecInst<v2i64, u16imm_i64, immILHUvec_i64>;
376 def v4i32: ILHUVecInst<v4i32, u16imm_i32, immILHUvec>;
378 def r64: ILHURegInst<R64C, u16imm_i64, hi16>;
379 def r32: ILHURegInst<R32C, u16imm_i32, hi16>;
381 // Loads the high portion of an address
382 def hi: ILHURegInst<R32C, symbolHi, hi16>;
384 // Used in custom lowering constant SFP loads:
385 def f32: ILHURegInst<R32FP, f16imm, hi16_f32>;
388 defm ILHU : ImmLoadHalfwordUpper;
390 // Immediate load address (can also be used to load 18-bit unsigned constants,
391 // see the zext 16->32 pattern)
393 class ILAInst<dag OOL, dag IOL, list<dag> pattern>:
394 RI18Form<0b1000010, OOL, IOL, "ila\t$rT, $val",
397 class ILAVecInst<ValueType vectype, Operand immtype, PatLeaf xform>:
398 ILAInst<(outs VECREG:$rT), (ins immtype:$val),
399 [(set (vectype VECREG:$rT), (vectype xform:$val))]>;
401 class ILARegInst<RegisterClass rclass, Operand immtype, PatLeaf xform>:
402 ILAInst<(outs rclass:$rT), (ins immtype:$val),
403 [(set rclass:$rT, xform:$val)]>;
405 multiclass ImmLoadAddress
407 def v2i64: ILAVecInst<v2i64, u18imm, v2i64Uns18Imm>;
408 def v4i32: ILAVecInst<v4i32, u18imm, v4i32Uns18Imm>;
410 def r64: ILARegInst<R64C, u18imm_i64, imm18>;
411 def r32: ILARegInst<R32C, u18imm, imm18>;
412 def f32: ILARegInst<R32FP, f18imm, fpimm18>;
413 def f64: ILARegInst<R64FP, f18imm_f64, fpimm18>;
415 def hi: ILARegInst<R32C, symbolHi, imm18>;
416 def lo: ILARegInst<R32C, symbolLo, imm18>;
418 def lsa: ILAInst<(outs R32C:$rT), (ins symbolLSA:$val),
422 defm ILA : ImmLoadAddress;
424 // Immediate OR, Halfword Lower: The "other" part of loading large constants
425 // into 32-bit registers. See the anonymous pattern Pat<(i32 imm:$imm), ...>
426 // Note that these are really two operand instructions, but they're encoded
427 // as three operands with the first two arguments tied-to each other.
429 class IOHLInst<dag OOL, dag IOL, list<dag> pattern>:
430 RI16Form<0b100000110, OOL, IOL, "iohl\t$rT, $val",
432 RegConstraint<"$rS = $rT">,
435 class IOHLVecInst<ValueType vectype, Operand immtype /* , PatLeaf xform */>:
436 IOHLInst<(outs VECREG:$rT), (ins VECREG:$rS, immtype:$val),
439 class IOHLRegInst<RegisterClass rclass, Operand immtype /* , PatLeaf xform */>:
440 IOHLInst<(outs rclass:$rT), (ins rclass:$rS, immtype:$val),
443 multiclass ImmOrHalfwordLower
445 def v2i64: IOHLVecInst<v2i64, u16imm_i64>;
446 def v4i32: IOHLVecInst<v4i32, u16imm_i32>;
448 def r32: IOHLRegInst<R32C, i32imm>;
449 def f32: IOHLRegInst<R32FP, f32imm>;
451 def lo: IOHLRegInst<R32C, symbolLo>;
454 defm IOHL: ImmOrHalfwordLower;
456 // Form select mask for bytes using immediate, used in conjunction with the
459 class FSMBIVec<ValueType vectype>:
460 RI16Form<0b101001100, (outs VECREG:$rT), (ins u16imm:$val),
463 [(set (vectype VECREG:$rT), (SPUselmask (i16 immU16:$val)))]>;
465 multiclass FormSelectMaskBytesImm
467 def v16i8: FSMBIVec<v16i8>;
468 def v8i16: FSMBIVec<v8i16>;
469 def v4i32: FSMBIVec<v4i32>;
470 def v2i64: FSMBIVec<v2i64>;
473 defm FSMBI : FormSelectMaskBytesImm;
475 // fsmb: Form select mask for bytes. N.B. Input operand, $rA, is 16-bits
476 class FSMBInst<dag OOL, dag IOL, list<dag> pattern>:
477 RRForm_1<0b01101101100, OOL, IOL, "fsmb\t$rT, $rA", SelectOp,
480 class FSMBRegInst<RegisterClass rclass, ValueType vectype>:
481 FSMBInst<(outs VECREG:$rT), (ins rclass:$rA),
482 [(set (vectype VECREG:$rT), (SPUselmask rclass:$rA))]>;
484 class FSMBVecInst<ValueType vectype>:
485 FSMBInst<(outs VECREG:$rT), (ins VECREG:$rA),
486 [(set (vectype VECREG:$rT),
487 (SPUselmask (vectype VECREG:$rA)))]>;
489 multiclass FormSelectMaskBits {
490 def v16i8_r16: FSMBRegInst<R16C, v16i8>;
491 def v16i8: FSMBVecInst<v16i8>;
494 defm FSMB: FormSelectMaskBits;
496 // fsmh: Form select mask for halfwords. N.B., Input operand, $rA, is
497 // only 8-bits wide (even though it's input as 16-bits here)
499 class FSMHInst<dag OOL, dag IOL, list<dag> pattern>:
500 RRForm_1<0b10101101100, OOL, IOL, "fsmh\t$rT, $rA", SelectOp,
503 class FSMHRegInst<RegisterClass rclass, ValueType vectype>:
504 FSMHInst<(outs VECREG:$rT), (ins rclass:$rA),
505 [(set (vectype VECREG:$rT), (SPUselmask rclass:$rA))]>;
507 class FSMHVecInst<ValueType vectype>:
508 FSMHInst<(outs VECREG:$rT), (ins VECREG:$rA),
509 [(set (vectype VECREG:$rT),
510 (SPUselmask (vectype VECREG:$rA)))]>;
512 multiclass FormSelectMaskHalfword {
513 def v8i16_r16: FSMHRegInst<R16C, v8i16>;
514 def v8i16: FSMHVecInst<v8i16>;
517 defm FSMH: FormSelectMaskHalfword;
519 // fsm: Form select mask for words. Like the other fsm* instructions,
520 // only the lower 4 bits of $rA are significant.
522 class FSMInst<dag OOL, dag IOL, list<dag> pattern>:
523 RRForm_1<0b00101101100, OOL, IOL, "fsm\t$rT, $rA", SelectOp,
526 class FSMRegInst<ValueType vectype, RegisterClass rclass>:
527 FSMInst<(outs VECREG:$rT), (ins rclass:$rA),
528 [(set (vectype VECREG:$rT), (SPUselmask rclass:$rA))]>;
530 class FSMVecInst<ValueType vectype>:
531 FSMInst<(outs VECREG:$rT), (ins VECREG:$rA),
532 [(set (vectype VECREG:$rT), (SPUselmask (vectype VECREG:$rA)))]>;
534 multiclass FormSelectMaskWord {
535 def v4i32: FSMVecInst<v4i32>;
537 def r32 : FSMRegInst<v4i32, R32C>;
538 def r16 : FSMRegInst<v4i32, R16C>;
541 defm FSM : FormSelectMaskWord;
543 // Special case when used for i64 math operations
544 multiclass FormSelectMaskWord64 {
545 def r32 : FSMRegInst<v2i64, R32C>;
546 def r16 : FSMRegInst<v2i64, R16C>;
549 defm FSM64 : FormSelectMaskWord64;
551 //===----------------------------------------------------------------------===//
552 // Integer and Logical Operations:
553 //===----------------------------------------------------------------------===//
556 RRForm<0b00010011000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
557 "ah\t$rT, $rA, $rB", IntegerOp,
558 [(set (v8i16 VECREG:$rT), (int_spu_si_ah VECREG:$rA, VECREG:$rB))]>;
560 def : Pat<(add (v8i16 VECREG:$rA), (v8i16 VECREG:$rB)),
561 (AHv8i16 VECREG:$rA, VECREG:$rB)>;
564 RRForm<0b00010011000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
565 "ah\t$rT, $rA, $rB", IntegerOp,
566 [(set R16C:$rT, (add R16C:$rA, R16C:$rB))]>;
569 RI10Form<0b10111000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
570 "ahi\t$rT, $rA, $val", IntegerOp,
571 [(set (v8i16 VECREG:$rT), (add (v8i16 VECREG:$rA),
572 v8i16SExt10Imm:$val))]>;
575 RI10Form<0b10111000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
576 "ahi\t$rT, $rA, $val", IntegerOp,
577 [(set R16C:$rT, (add R16C:$rA, i16ImmSExt10:$val))]>;
579 // v4i32, i32 add instruction:
581 class AInst<dag OOL, dag IOL, list<dag> pattern>:
582 RRForm<0b00000011000, OOL, IOL,
583 "a\t$rT, $rA, $rB", IntegerOp,
586 class AVecInst<ValueType vectype>:
587 AInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
588 [(set (vectype VECREG:$rT), (add (vectype VECREG:$rA),
589 (vectype VECREG:$rB)))]>;
591 class ARegInst<RegisterClass rclass>:
592 AInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
593 [(set rclass:$rT, (add rclass:$rA, rclass:$rB))]>;
595 multiclass AddInstruction {
596 def v4i32: AVecInst<v4i32>;
597 def v16i8: AVecInst<v16i8>;
598 def r32: ARegInst<R32C>;
601 defm A : AddInstruction;
603 class AIInst<dag OOL, dag IOL, list<dag> pattern>:
604 RI10Form<0b00111000, OOL, IOL,
605 "ai\t$rT, $rA, $val", IntegerOp,
608 class AIVecInst<ValueType vectype, PatLeaf immpred>:
609 AIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
610 [(set (vectype VECREG:$rT), (add (vectype VECREG:$rA), immpred:$val))]>;
612 class AIFPVecInst<ValueType vectype, PatLeaf immpred>:
613 AIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
616 class AIRegInst<RegisterClass rclass, PatLeaf immpred>:
617 AIInst<(outs rclass:$rT), (ins rclass:$rA, s10imm_i32:$val),
618 [(set rclass:$rT, (add rclass:$rA, immpred:$val))]>;
620 // This is used to add epsilons to floating point numbers in the f32 fdiv code:
621 class AIFPInst<RegisterClass rclass, PatLeaf immpred>:
622 AIInst<(outs rclass:$rT), (ins rclass:$rA, s10imm_i32:$val),
625 multiclass AddImmediate {
626 def v4i32: AIVecInst<v4i32, v4i32SExt10Imm>;
628 def r32: AIRegInst<R32C, i32ImmSExt10>;
630 def v4f32: AIFPVecInst<v4f32, v4i32SExt10Imm>;
631 def f32: AIFPInst<R32FP, i32ImmSExt10>;
634 defm AI : AddImmediate;
637 RRForm<0b00010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
638 "sfh\t$rT, $rA, $rB", IntegerOp,
639 [(set (v8i16 VECREG:$rT), (sub (v8i16 VECREG:$rA),
640 (v8i16 VECREG:$rB)))]>;
643 RRForm<0b00010010000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
644 "sfh\t$rT, $rA, $rB", IntegerOp,
645 [(set R16C:$rT, (sub R16C:$rB, R16C:$rA))]>;
648 RI10Form<0b10110000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
649 "sfhi\t$rT, $rA, $val", IntegerOp,
650 [(set (v8i16 VECREG:$rT), (sub v8i16SExt10Imm:$val,
651 (v8i16 VECREG:$rA)))]>;
653 def SFHIr16 : RI10Form<0b10110000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
654 "sfhi\t$rT, $rA, $val", IntegerOp,
655 [(set R16C:$rT, (sub i16ImmSExt10:$val, R16C:$rA))]>;
657 def SFvec : RRForm<0b00000010000, (outs VECREG:$rT),
658 (ins VECREG:$rA, VECREG:$rB),
659 "sf\t$rT, $rA, $rB", IntegerOp,
660 [(set (v4i32 VECREG:$rT), (sub (v4i32 VECREG:$rB), (v4i32 VECREG:$rA)))]>;
663 def SFr32 : RRForm<0b00000010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
664 "sf\t$rT, $rA, $rB", IntegerOp,
665 [(set R32C:$rT, (sub R32C:$rB, R32C:$rA))]>;
668 RI10Form<0b00110000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
669 "sfi\t$rT, $rA, $val", IntegerOp,
670 [(set (v4i32 VECREG:$rT), (sub v4i32SExt10Imm:$val,
671 (v4i32 VECREG:$rA)))]>;
673 def SFIr32 : RI10Form<0b00110000, (outs R32C:$rT),
674 (ins R32C:$rA, s10imm_i32:$val),
675 "sfi\t$rT, $rA, $val", IntegerOp,
676 [(set R32C:$rT, (sub i32ImmSExt10:$val, R32C:$rA))]>;
678 // ADDX: only available in vector form, doesn't match a pattern.
679 class ADDXInst<dag OOL, dag IOL, list<dag> pattern>:
680 RRForm<0b00000010110, OOL, IOL,
681 "addx\t$rT, $rA, $rB",
684 class ADDXVecInst<ValueType vectype>:
685 ADDXInst<(outs VECREG:$rT),
686 (ins VECREG:$rA, VECREG:$rB, VECREG:$rCarry),
688 RegConstraint<"$rCarry = $rT">,
691 class ADDXRegInst<RegisterClass rclass>:
692 ADDXInst<(outs rclass:$rT),
693 (ins rclass:$rA, rclass:$rB, rclass:$rCarry),
695 RegConstraint<"$rCarry = $rT">,
698 multiclass AddExtended {
699 def v2i64 : ADDXVecInst<v2i64>;
700 def v4i32 : ADDXVecInst<v4i32>;
701 def r64 : ADDXRegInst<R64C>;
702 def r32 : ADDXRegInst<R32C>;
705 defm ADDX : AddExtended;
707 // CG: Generate carry for add
708 class CGInst<dag OOL, dag IOL, list<dag> pattern>:
709 RRForm<0b01000011000, OOL, IOL,
713 class CGVecInst<ValueType vectype>:
714 CGInst<(outs VECREG:$rT),
715 (ins VECREG:$rA, VECREG:$rB),
718 class CGRegInst<RegisterClass rclass>:
719 CGInst<(outs rclass:$rT),
720 (ins rclass:$rA, rclass:$rB),
723 multiclass CarryGenerate {
724 def v2i64 : CGVecInst<v2i64>;
725 def v4i32 : CGVecInst<v4i32>;
726 def r64 : CGRegInst<R64C>;
727 def r32 : CGRegInst<R32C>;
730 defm CG : CarryGenerate;
732 // SFX: Subract from, extended. This is used in conjunction with BG to subtract
733 // with carry (borrow, in this case)
734 class SFXInst<dag OOL, dag IOL, list<dag> pattern>:
735 RRForm<0b10000010110, OOL, IOL,
736 "sfx\t$rT, $rA, $rB",
739 class SFXVecInst<ValueType vectype>:
740 SFXInst<(outs VECREG:$rT),
741 (ins VECREG:$rA, VECREG:$rB, VECREG:$rCarry),
743 RegConstraint<"$rCarry = $rT">,
746 class SFXRegInst<RegisterClass rclass>:
747 SFXInst<(outs rclass:$rT),
748 (ins rclass:$rA, rclass:$rB, rclass:$rCarry),
750 RegConstraint<"$rCarry = $rT">,
753 multiclass SubtractExtended {
754 def v2i64 : SFXVecInst<v2i64>;
755 def v4i32 : SFXVecInst<v4i32>;
756 def r64 : SFXRegInst<R64C>;
757 def r32 : SFXRegInst<R32C>;
760 defm SFX : SubtractExtended;
762 // BG: only available in vector form, doesn't match a pattern.
763 class BGInst<dag OOL, dag IOL, list<dag> pattern>:
764 RRForm<0b01000010000, OOL, IOL,
768 class BGVecInst<ValueType vectype>:
769 BGInst<(outs VECREG:$rT),
770 (ins VECREG:$rA, VECREG:$rB),
773 class BGRegInst<RegisterClass rclass>:
774 BGInst<(outs rclass:$rT),
775 (ins rclass:$rA, rclass:$rB),
778 multiclass BorrowGenerate {
779 def v4i32 : BGVecInst<v4i32>;
780 def v2i64 : BGVecInst<v2i64>;
781 def r64 : BGRegInst<R64C>;
782 def r32 : BGRegInst<R32C>;
785 defm BG : BorrowGenerate;
787 // BGX: Borrow generate, extended.
789 RRForm<0b11000010110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB,
791 "bgx\t$rT, $rA, $rB", IntegerOp,
793 RegConstraint<"$rCarry = $rT">,
796 // Halfword multiply variants:
797 // N.B: These can be used to build up larger quantities (16x16 -> 32)
800 RRForm<0b00100011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
801 "mpy\t$rT, $rA, $rB", IntegerMulDiv,
805 RRForm<0b00100011110, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
806 "mpy\t$rT, $rA, $rB", IntegerMulDiv,
807 [(set R16C:$rT, (mul R16C:$rA, R16C:$rB))]>;
809 // Unsigned 16-bit multiply:
811 class MPYUInst<dag OOL, dag IOL, list<dag> pattern>:
812 RRForm<0b00110011110, OOL, IOL,
813 "mpyu\t$rT, $rA, $rB", IntegerMulDiv,
817 MPYUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
821 MPYUInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB),
822 [(set R32C:$rT, (mul (zext R16C:$rA), (zext R16C:$rB)))]>;
825 MPYUInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
828 // mpyi: multiply 16 x s10imm -> 32 result.
830 class MPYIInst<dag OOL, dag IOL, list<dag> pattern>:
831 RI10Form<0b00101110, OOL, IOL,
832 "mpyi\t$rT, $rA, $val", IntegerMulDiv,
836 MPYIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
837 [(set (v8i16 VECREG:$rT),
838 (mul (v8i16 VECREG:$rA), v8i16SExt10Imm:$val))]>;
841 MPYIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
842 [(set R16C:$rT, (mul R16C:$rA, i16ImmSExt10:$val))]>;
844 // mpyui: same issues as other multiplies, plus, this doesn't match a
845 // pattern... but may be used during target DAG selection or lowering
847 class MPYUIInst<dag OOL, dag IOL, list<dag> pattern>:
848 RI10Form<0b10101110, OOL, IOL,
849 "mpyui\t$rT, $rA, $val", IntegerMulDiv,
853 MPYUIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
857 MPYUIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
860 // mpya: 16 x 16 + 16 -> 32 bit result
861 class MPYAInst<dag OOL, dag IOL, list<dag> pattern>:
862 RRRForm<0b0011, OOL, IOL,
863 "mpya\t$rT, $rA, $rB, $rC", IntegerMulDiv,
867 MPYAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
868 [(set (v4i32 VECREG:$rT),
869 (add (v4i32 (bitconvert (mul (v8i16 VECREG:$rA),
870 (v8i16 VECREG:$rB)))),
871 (v4i32 VECREG:$rC)))]>;
874 MPYAInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB, R32C:$rC),
875 [(set R32C:$rT, (add (sext (mul R16C:$rA, R16C:$rB)),
879 MPYAInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB, R32C:$rC),
880 [(set R32C:$rT, (add (mul (sext R16C:$rA), (sext R16C:$rB)),
883 def MPYAr32_sextinreg:
884 MPYAInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB, R32C:$rC),
885 [(set R32C:$rT, (add (mul (sext_inreg R32C:$rA, i16),
886 (sext_inreg R32C:$rB, i16)),
889 // mpyh: multiply high, used to synthesize 32-bit multiplies
890 class MPYHInst<dag OOL, dag IOL, list<dag> pattern>:
891 RRForm<0b10100011110, OOL, IOL,
892 "mpyh\t$rT, $rA, $rB", IntegerMulDiv,
896 MPYHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
900 MPYHInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
903 // mpys: multiply high and shift right (returns the top half of
904 // a 16-bit multiply, sign extended to 32 bits.)
906 class MPYSInst<dag OOL, dag IOL>:
907 RRForm<0b11100011110, OOL, IOL,
908 "mpys\t$rT, $rA, $rB", IntegerMulDiv,
912 MPYSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
915 MPYSInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB)>;
917 // mpyhh: multiply high-high (returns the 32-bit result from multiplying
918 // the top 16 bits of the $rA, $rB)
920 class MPYHHInst<dag OOL, dag IOL>:
921 RRForm<0b01100011110, OOL, IOL,
922 "mpyhh\t$rT, $rA, $rB", IntegerMulDiv,
926 MPYHHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
929 MPYHHInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
931 // mpyhha: Multiply high-high, add to $rT:
933 class MPYHHAInst<dag OOL, dag IOL>:
934 RRForm<0b01100010110, OOL, IOL,
935 "mpyhha\t$rT, $rA, $rB", IntegerMulDiv,
939 MPYHHAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
942 MPYHHAInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
944 // mpyhhu: Multiply high-high, unsigned, e.g.:
946 // +-------+-------+ +-------+-------+ +---------+
947 // | a0 . a1 | x | b0 . b1 | = | a0 x b0 |
948 // +-------+-------+ +-------+-------+ +---------+
950 // where a0, b0 are the upper 16 bits of the 32-bit word
952 class MPYHHUInst<dag OOL, dag IOL>:
953 RRForm<0b01110011110, OOL, IOL,
954 "mpyhhu\t$rT, $rA, $rB", IntegerMulDiv,
958 MPYHHUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
961 MPYHHUInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
963 // mpyhhau: Multiply high-high, unsigned
965 class MPYHHAUInst<dag OOL, dag IOL>:
966 RRForm<0b01110010110, OOL, IOL,
967 "mpyhhau\t$rT, $rA, $rB", IntegerMulDiv,
971 MPYHHAUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
974 MPYHHAUInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
976 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
977 // clz: Count leading zeroes
978 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
979 class CLZInst<dag OOL, dag IOL, list<dag> pattern>:
980 RRForm_1<0b10100101010, OOL, IOL, "clz\t$rT, $rA",
983 class CLZRegInst<RegisterClass rclass>:
984 CLZInst<(outs rclass:$rT), (ins rclass:$rA),
985 [(set rclass:$rT, (ctlz rclass:$rA))]>;
987 class CLZVecInst<ValueType vectype>:
988 CLZInst<(outs VECREG:$rT), (ins VECREG:$rA),
989 [(set (vectype VECREG:$rT), (ctlz (vectype VECREG:$rA)))]>;
991 multiclass CountLeadingZeroes {
992 def v4i32 : CLZVecInst<v4i32>;
993 def r32 : CLZRegInst<R32C>;
996 defm CLZ : CountLeadingZeroes;
998 // cntb: Count ones in bytes (aka "population count")
1000 // NOTE: This instruction is really a vector instruction, but the custom
1001 // lowering code uses it in unorthodox ways to support CTPOP for other
1005 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
1006 "cntb\t$rT, $rA", IntegerOp,
1007 [(set (v16i8 VECREG:$rT), (SPUcntb (v16i8 VECREG:$rA)))]>;
1010 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
1011 "cntb\t$rT, $rA", IntegerOp,
1012 [(set (v8i16 VECREG:$rT), (SPUcntb (v8i16 VECREG:$rA)))]>;
1015 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
1016 "cntb\t$rT, $rA", IntegerOp,
1017 [(set (v4i32 VECREG:$rT), (SPUcntb (v4i32 VECREG:$rA)))]>;
1019 // gbb: Gather the low order bits from each byte in $rA into a single 16-bit
1020 // quantity stored into $rT's slot 0, upper 16 bits are zeroed, as are
1023 // Note: This instruction "pairs" with the fsmb instruction for all of the
1024 // various types defined here.
1026 // Note 2: The "VecInst" and "RegInst" forms refer to the result being either
1027 // a vector or register.
1029 class GBBInst<dag OOL, dag IOL, list<dag> pattern>:
1030 RRForm_1<0b01001101100, OOL, IOL, "gbb\t$rT, $rA", GatherOp, pattern>;
1032 class GBBRegInst<RegisterClass rclass, ValueType vectype>:
1033 GBBInst<(outs rclass:$rT), (ins VECREG:$rA),
1034 [/* no pattern */]>;
1036 class GBBVecInst<ValueType vectype>:
1037 GBBInst<(outs VECREG:$rT), (ins VECREG:$rA),
1038 [/* no pattern */]>;
1040 multiclass GatherBitsFromBytes {
1041 def v16i8_r32: GBBRegInst<R32C, v16i8>;
1042 def v16i8_r16: GBBRegInst<R16C, v16i8>;
1043 def v16i8: GBBVecInst<v16i8>;
1046 defm GBB: GatherBitsFromBytes;
1048 // gbh: Gather all low order bits from each halfword in $rA into a single
1049 // 8-bit quantity stored in $rT's slot 0, with the upper bits of $rT set to 0
1050 // and slots 1-3 also set to 0.
1052 // See notes for GBBInst, above.
1054 class GBHInst<dag OOL, dag IOL, list<dag> pattern>:
1055 RRForm_1<0b10001101100, OOL, IOL, "gbh\t$rT, $rA", GatherOp,
1058 class GBHRegInst<RegisterClass rclass, ValueType vectype>:
1059 GBHInst<(outs rclass:$rT), (ins VECREG:$rA),
1060 [/* no pattern */]>;
1062 class GBHVecInst<ValueType vectype>:
1063 GBHInst<(outs VECREG:$rT), (ins VECREG:$rA),
1064 [/* no pattern */]>;
1066 multiclass GatherBitsHalfword {
1067 def v8i16_r32: GBHRegInst<R32C, v8i16>;
1068 def v8i16_r16: GBHRegInst<R16C, v8i16>;
1069 def v8i16: GBHVecInst<v8i16>;
1072 defm GBH: GatherBitsHalfword;
1074 // gb: Gather all low order bits from each word in $rA into a single
1075 // 4-bit quantity stored in $rT's slot 0, upper bits in $rT set to 0,
1076 // as well as slots 1-3.
1078 // See notes for gbb, above.
1080 class GBInst<dag OOL, dag IOL, list<dag> pattern>:
1081 RRForm_1<0b00001101100, OOL, IOL, "gb\t$rT, $rA", GatherOp,
1084 class GBRegInst<RegisterClass rclass, ValueType vectype>:
1085 GBInst<(outs rclass:$rT), (ins VECREG:$rA),
1086 [/* no pattern */]>;
1088 class GBVecInst<ValueType vectype>:
1089 GBInst<(outs VECREG:$rT), (ins VECREG:$rA),
1090 [/* no pattern */]>;
1092 multiclass GatherBitsWord {
1093 def v4i32_r32: GBRegInst<R32C, v4i32>;
1094 def v4i32_r16: GBRegInst<R16C, v4i32>;
1095 def v4i32: GBVecInst<v4i32>;
1098 defm GB: GatherBitsWord;
1100 // avgb: average bytes
1102 RRForm<0b11001011000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1103 "avgb\t$rT, $rA, $rB", ByteOp,
1106 // absdb: absolute difference of bytes
1108 RRForm<0b11001010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1109 "absdb\t$rT, $rA, $rB", ByteOp,
1112 // sumb: sum bytes into halfwords
1114 RRForm<0b11001010010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1115 "sumb\t$rT, $rA, $rB", ByteOp,
1118 // Sign extension operations:
1119 class XSBHInst<dag OOL, dag IOL, list<dag> pattern>:
1120 RRForm_1<0b01101101010, OOL, IOL,
1121 "xsbh\t$rDst, $rSrc",
1122 IntegerOp, pattern>;
1124 class XSBHInRegInst<RegisterClass rclass, list<dag> pattern>:
1125 XSBHInst<(outs rclass:$rDst), (ins rclass:$rSrc),
1128 multiclass ExtendByteHalfword {
1129 def v16i8: XSBHInst<(outs VECREG:$rDst), (ins VECREG:$rSrc),
1131 /*(set (v8i16 VECREG:$rDst), (sext (v8i16 VECREG:$rSrc)))*/]>;
1132 def r8: XSBHInst<(outs R16C:$rDst), (ins R8C:$rSrc),
1133 [(set R16C:$rDst, (sext R8C:$rSrc))]>;
1134 def r16: XSBHInRegInst<R16C,
1135 [(set R16C:$rDst, (sext_inreg R16C:$rSrc, i8))]>;
1137 // 32-bit form for XSBH: used to sign extend 8-bit quantities to 16-bit
1138 // quantities to 32-bit quantities via a 32-bit register (see the sext 8->32
1139 // pattern below). Intentionally doesn't match a pattern because we want the
1140 // sext 8->32 pattern to do the work for us, namely because we need the extra
1142 def r32: XSBHInRegInst<R32C, [/* no pattern */]>;
1144 // Same as the 32-bit version, but for i64
1145 def r64: XSBHInRegInst<R64C, [/* no pattern */]>;
1148 defm XSBH : ExtendByteHalfword;
1150 // Sign extend halfwords to words:
1152 class XSHWInst<dag OOL, dag IOL, list<dag> pattern>:
1153 RRForm_1<0b01101101010, OOL, IOL, "xshw\t$rDest, $rSrc",
1154 IntegerOp, pattern>;
1156 class XSHWVecInst<ValueType in_vectype, ValueType out_vectype>:
1157 XSHWInst<(outs VECREG:$rDest), (ins VECREG:$rSrc),
1158 [(set (out_vectype VECREG:$rDest),
1159 (sext (in_vectype VECREG:$rSrc)))]>;
1161 class XSHWInRegInst<RegisterClass rclass, list<dag> pattern>:
1162 XSHWInst<(outs rclass:$rDest), (ins rclass:$rSrc),
1165 class XSHWRegInst<RegisterClass rclass>:
1166 XSHWInst<(outs rclass:$rDest), (ins R16C:$rSrc),
1167 [(set rclass:$rDest, (sext R16C:$rSrc))]>;
1169 multiclass ExtendHalfwordWord {
1170 def v4i32: XSHWVecInst<v4i32, v8i16>;
1172 def r16: XSHWRegInst<R32C>;
1174 def r32: XSHWInRegInst<R32C,
1175 [(set R32C:$rDest, (sext_inreg R32C:$rSrc, i16))]>;
1176 def r64: XSHWInRegInst<R64C, [/* no pattern */]>;
1179 defm XSHW : ExtendHalfwordWord;
1181 // Sign-extend words to doublewords (32->64 bits)
1183 class XSWDInst<dag OOL, dag IOL, list<dag> pattern>:
1184 RRForm_1<0b01100101010, OOL, IOL, "xswd\t$rDst, $rSrc",
1185 IntegerOp, pattern>;
1187 class XSWDVecInst<ValueType in_vectype, ValueType out_vectype>:
1188 XSWDInst<(outs VECREG:$rDst), (ins VECREG:$rSrc),
1189 [/*(set (out_vectype VECREG:$rDst),
1190 (sext (out_vectype VECREG:$rSrc)))*/]>;
1192 class XSWDRegInst<RegisterClass in_rclass, RegisterClass out_rclass>:
1193 XSWDInst<(outs out_rclass:$rDst), (ins in_rclass:$rSrc),
1194 [(set out_rclass:$rDst, (sext in_rclass:$rSrc))]>;
1196 multiclass ExtendWordToDoubleWord {
1197 def v2i64: XSWDVecInst<v4i32, v2i64>;
1198 def r64: XSWDRegInst<R32C, R64C>;
1200 def r64_inreg: XSWDInst<(outs R64C:$rDst), (ins R64C:$rSrc),
1201 [(set R64C:$rDst, (sext_inreg R64C:$rSrc, i32))]>;
1204 defm XSWD : ExtendWordToDoubleWord;
1208 class ANDInst<dag OOL, dag IOL, list<dag> pattern> :
1209 RRForm<0b10000011000, OOL, IOL, "and\t$rT, $rA, $rB",
1210 IntegerOp, pattern>;
1212 class ANDVecInst<ValueType vectype>:
1213 ANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1214 [(set (vectype VECREG:$rT), (and (vectype VECREG:$rA),
1215 (vectype VECREG:$rB)))]>;
1217 class ANDRegInst<RegisterClass rclass>:
1218 ANDInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1219 [(set rclass:$rT, (and rclass:$rA, rclass:$rB))]>;
1221 multiclass BitwiseAnd
1223 def v16i8: ANDVecInst<v16i8>;
1224 def v8i16: ANDVecInst<v8i16>;
1225 def v4i32: ANDVecInst<v4i32>;
1226 def v2i64: ANDVecInst<v2i64>;
1228 def r128: ANDRegInst<GPRC>;
1229 def r64: ANDRegInst<R64C>;
1230 def r32: ANDRegInst<R32C>;
1231 def r16: ANDRegInst<R16C>;
1232 def r8: ANDRegInst<R8C>;
1234 //===---------------------------------------------
1235 // Special instructions to perform the fabs instruction
1236 def fabs32: ANDInst<(outs R32FP:$rT), (ins R32FP:$rA, R32C:$rB),
1237 [/* Intentionally does not match a pattern */]>;
1239 def fabs64: ANDInst<(outs R64FP:$rT), (ins R64FP:$rA, R64C:$rB),
1240 [/* Intentionally does not match a pattern */]>;
1242 def fabsvec: ANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1243 [/* Intentionally does not match a pattern */]>;
1245 //===---------------------------------------------
1247 // Hacked form of AND to zero-extend 16-bit quantities to 32-bit
1248 // quantities -- see 16->32 zext pattern.
1250 // This pattern is somewhat artificial, since it might match some
1251 // compiler generated pattern but it is unlikely to do so.
1253 def i16i32: ANDInst<(outs R32C:$rT), (ins R16C:$rA, R32C:$rB),
1254 [(set R32C:$rT, (and (zext R16C:$rA), R32C:$rB))]>;
1257 defm AND : BitwiseAnd;
1260 def vnot_cell_conv : PatFrag<(ops node:$in),
1261 (xor node:$in, (bitconvert (v4i32 immAllOnesV)))>;
1263 // N.B.: vnot_cell_conv is one of those special target selection pattern
1265 // in which we expect there to be a bit_convert on the constant. Bear in mind
1266 // that llvm translates "not <reg>" to "xor <reg>, -1" (or in this case, a
1267 // constant -1 vector.)
1269 class ANDCInst<dag OOL, dag IOL, list<dag> pattern>:
1270 RRForm<0b10000011010, OOL, IOL, "andc\t$rT, $rA, $rB",
1271 IntegerOp, pattern>;
1273 class ANDCVecInst<ValueType vectype, PatFrag vnot_frag = vnot>:
1274 ANDCInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1275 [(set (vectype VECREG:$rT),
1276 (and (vectype VECREG:$rA),
1277 (vnot_frag (vectype VECREG:$rB))))]>;
1279 class ANDCRegInst<RegisterClass rclass>:
1280 ANDCInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1281 [(set rclass:$rT, (and rclass:$rA, (not rclass:$rB)))]>;
1283 multiclass AndComplement
1285 def v16i8: ANDCVecInst<v16i8>;
1286 def v8i16: ANDCVecInst<v8i16>;
1287 def v4i32: ANDCVecInst<v4i32>;
1288 def v2i64: ANDCVecInst<v2i64>;
1290 def r128: ANDCRegInst<GPRC>;
1291 def r64: ANDCRegInst<R64C>;
1292 def r32: ANDCRegInst<R32C>;
1293 def r16: ANDCRegInst<R16C>;
1294 def r8: ANDCRegInst<R8C>;
1296 // Sometimes, the xor pattern has a bitcast constant:
1297 def v16i8_conv: ANDCVecInst<v16i8, vnot_cell_conv>;
1300 defm ANDC : AndComplement;
1302 class ANDBIInst<dag OOL, dag IOL, list<dag> pattern>:
1303 RI10Form<0b01101000, OOL, IOL, "andbi\t$rT, $rA, $val",
1306 multiclass AndByteImm
1308 def v16i8: ANDBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1309 [(set (v16i8 VECREG:$rT),
1310 (and (v16i8 VECREG:$rA),
1311 (v16i8 v16i8U8Imm:$val)))]>;
1313 def r8: ANDBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1314 [(set R8C:$rT, (and R8C:$rA, immU8:$val))]>;
1317 defm ANDBI : AndByteImm;
1319 class ANDHIInst<dag OOL, dag IOL, list<dag> pattern> :
1320 RI10Form<0b10101000, OOL, IOL, "andhi\t$rT, $rA, $val",
1323 multiclass AndHalfwordImm
1325 def v8i16: ANDHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
1326 [(set (v8i16 VECREG:$rT),
1327 (and (v8i16 VECREG:$rA), v8i16SExt10Imm:$val))]>;
1329 def r16: ANDHIInst<(outs R16C:$rT), (ins R16C:$rA, u10imm:$val),
1330 [(set R16C:$rT, (and R16C:$rA, i16ImmUns10:$val))]>;
1332 // Zero-extend i8 to i16:
1333 def i8i16: ANDHIInst<(outs R16C:$rT), (ins R8C:$rA, u10imm:$val),
1334 [(set R16C:$rT, (and (zext R8C:$rA), i16ImmUns10:$val))]>;
1337 defm ANDHI : AndHalfwordImm;
1339 class ANDIInst<dag OOL, dag IOL, list<dag> pattern> :
1340 RI10Form<0b00101000, OOL, IOL, "andi\t$rT, $rA, $val",
1341 IntegerOp, pattern>;
1343 multiclass AndWordImm
1345 def v4i32: ANDIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
1346 [(set (v4i32 VECREG:$rT),
1347 (and (v4i32 VECREG:$rA), v4i32SExt10Imm:$val))]>;
1349 def r32: ANDIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
1350 [(set R32C:$rT, (and R32C:$rA, i32ImmSExt10:$val))]>;
1352 // Hacked form of ANDI to zero-extend i8 quantities to i32. See the zext 8->32
1354 def i8i32: ANDIInst<(outs R32C:$rT), (ins R8C:$rA, s10imm_i32:$val),
1356 (and (zext R8C:$rA), i32ImmSExt10:$val))]>;
1358 // Hacked form of ANDI to zero-extend i16 quantities to i32. See the
1359 // zext 16->32 pattern below.
1361 // Note that this pattern is somewhat artificial, since it might match
1362 // something the compiler generates but is unlikely to occur in practice.
1363 def i16i32: ANDIInst<(outs R32C:$rT), (ins R16C:$rA, s10imm_i32:$val),
1365 (and (zext R16C:$rA), i32ImmSExt10:$val))]>;
1368 defm ANDI : AndWordImm;
1370 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1371 // Bitwise OR group:
1372 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1374 // Bitwise "or" (N.B.: These are also register-register copy instructions...)
1375 class ORInst<dag OOL, dag IOL, list<dag> pattern>:
1376 RRForm<0b10000010000, OOL, IOL, "or\t$rT, $rA, $rB",
1377 IntegerOp, pattern>;
1379 class ORVecInst<ValueType vectype>:
1380 ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1381 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1382 (vectype VECREG:$rB)))]>;
1384 class ORRegInst<RegisterClass rclass>:
1385 ORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1386 [(set rclass:$rT, (or rclass:$rA, rclass:$rB))]>;
1388 // ORCvtForm: OR conversion form
1390 // This is used to "convert" the preferred slot to its vector equivalent, as
1391 // well as convert a vector back to its preferred slot.
1393 // These are effectively no-ops, but need to exist for proper type conversion
1394 // and type coercion.
1396 class ORCvtForm<dag OOL, dag IOL, list<dag> pattern = [/* no pattern */]>
1397 : SPUInstr<OOL, IOL, "or\t$rT, $rA, $rA", IntegerOp> {
1401 let Pattern = pattern;
1403 let Inst{0-10} = 0b10000010000;
1404 let Inst{11-17} = RA;
1405 let Inst{18-24} = RA;
1406 let Inst{25-31} = RT;
1409 class ORPromoteScalar<RegisterClass rclass>:
1410 ORCvtForm<(outs VECREG:$rT), (ins rclass:$rA)>;
1412 class ORExtractElt<RegisterClass rclass>:
1413 ORCvtForm<(outs rclass:$rT), (ins VECREG:$rA)>;
1415 /* class ORCvtRegGPRC<RegisterClass rclass>:
1416 ORCvtForm<(outs GPRC:$rT), (ins rclass:$rA)>; */
1418 /* class ORCvtGPRCReg<RegisterClass rclass>:
1419 ORCvtForm<(outs rclass:$rT), (ins GPRC:$rA)>; */
1421 class ORCvtFormR32Reg<RegisterClass rclass, list<dag> pattern = [ ]>:
1422 ORCvtForm<(outs rclass:$rT), (ins R32C:$rA), pattern>;
1424 class ORCvtFormRegR32<RegisterClass rclass, list<dag> pattern = [ ]>:
1425 ORCvtForm<(outs R32C:$rT), (ins rclass:$rA), pattern>;
1427 class ORCvtFormR64Reg<RegisterClass rclass, list<dag> pattern = [ ]>:
1428 ORCvtForm<(outs rclass:$rT), (ins R64C:$rA), pattern>;
1430 class ORCvtFormRegR64<RegisterClass rclass, list<dag> pattern = [ ]>:
1431 ORCvtForm<(outs R64C:$rT), (ins rclass:$rA), pattern>;
1434 ORCvtForm<(outs VECREG:$rT), (ins GPRC:$rA)>;
1437 ORCvtForm<(outs GPRC:$rT), (ins VECREG:$rA)>;
1440 ORCvtForm<(outs VECREG:$rT), (ins VECREG:$rA)>;
1442 multiclass BitwiseOr
1444 def v16i8: ORVecInst<v16i8>;
1445 def v8i16: ORVecInst<v8i16>;
1446 def v4i32: ORVecInst<v4i32>;
1447 def v2i64: ORVecInst<v2i64>;
1449 def v4f32: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1450 [(set (v4f32 VECREG:$rT),
1451 (v4f32 (bitconvert (or (v4i32 VECREG:$rA),
1452 (v4i32 VECREG:$rB)))))]>;
1454 def v2f64: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1455 [(set (v2f64 VECREG:$rT),
1456 (v2f64 (bitconvert (or (v2i64 VECREG:$rA),
1457 (v2i64 VECREG:$rB)))))]>;
1459 def r128: ORRegInst<GPRC>;
1460 def r64: ORRegInst<R64C>;
1461 def r32: ORRegInst<R32C>;
1462 def r16: ORRegInst<R16C>;
1463 def r8: ORRegInst<R8C>;
1465 // OR instructions used to copy f32 and f64 registers.
1466 def f32: ORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
1467 [/* no pattern */]>;
1469 def f64: ORInst<(outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
1470 [/* no pattern */]>;
1472 // scalar->vector promotion, prefslot2vec:
1473 def v16i8_i8: ORPromoteScalar<R8C>;
1474 def v8i16_i16: ORPromoteScalar<R16C>;
1475 def v4i32_i32: ORPromoteScalar<R32C>;
1476 def v2i64_i64: ORPromoteScalar<R64C>;
1477 def v4f32_f32: ORPromoteScalar<R32FP>;
1478 def v2f64_f64: ORPromoteScalar<R64FP>;
1480 // vector->scalar demotion, vec2prefslot:
1481 def i8_v16i8: ORExtractElt<R8C>;
1482 def i16_v8i16: ORExtractElt<R16C>;
1483 def i32_v4i32: ORExtractElt<R32C>;
1484 def i64_v2i64: ORExtractElt<R64C>;
1485 def f32_v4f32: ORExtractElt<R32FP>;
1486 def f64_v2f64: ORExtractElt<R64FP>;
1488 // Conversion from vector to GPRC
1489 def i128_vec: ORCvtVecGPRC;
1491 // Conversion from GPRC to vector
1492 def vec_i128: ORCvtGPRCVec;
1495 // Conversion from register to GPRC
1496 def i128_r64: ORCvtRegGPRC<R64C>;
1497 def i128_f64: ORCvtRegGPRC<R64FP>;
1498 def i128_r32: ORCvtRegGPRC<R32C>;
1499 def i128_f32: ORCvtRegGPRC<R32FP>;
1500 def i128_r16: ORCvtRegGPRC<R16C>;
1501 def i128_r8: ORCvtRegGPRC<R8C>;
1503 // Conversion from GPRC to register
1504 def r64_i128: ORCvtGPRCReg<R64C>;
1505 def f64_i128: ORCvtGPRCReg<R64FP>;
1506 def r32_i128: ORCvtGPRCReg<R32C>;
1507 def f32_i128: ORCvtGPRCReg<R32FP>;
1508 def r16_i128: ORCvtGPRCReg<R16C>;
1509 def r8_i128: ORCvtGPRCReg<R8C>;
1512 // Conversion from register to R32C:
1513 def r32_r16: ORCvtFormRegR32<R16C>;
1514 def r32_r8: ORCvtFormRegR32<R8C>;
1516 // Conversion from R32C to register
1517 def r32_r16: ORCvtFormR32Reg<R16C>;
1518 def r32_r8: ORCvtFormR32Reg<R8C>;
1521 // Conversion from R64C to register:
1522 def r32_r64: ORCvtFormR64Reg<R32C>;
1523 // def r16_r64: ORCvtFormR64Reg<R16C>;
1524 // def r8_r64: ORCvtFormR64Reg<R8C>;
1526 // Conversion to R64C from register:
1527 def r64_r32: ORCvtFormRegR64<R32C>;
1528 // def r64_r16: ORCvtFormRegR64<R16C>;
1529 // def r64_r8: ORCvtFormRegR64<R8C>;
1531 // bitconvert patterns:
1532 def r32_f32: ORCvtFormR32Reg<R32FP,
1533 [(set R32FP:$rT, (bitconvert R32C:$rA))]>;
1534 def f32_r32: ORCvtFormRegR32<R32FP,
1535 [(set R32C:$rT, (bitconvert R32FP:$rA))]>;
1537 def r64_f64: ORCvtFormR64Reg<R64FP,
1538 [(set R64FP:$rT, (bitconvert R64C:$rA))]>;
1539 def f64_r64: ORCvtFormRegR64<R64FP,
1540 [(set R64C:$rT, (bitconvert R64FP:$rA))]>;
1543 defm OR : BitwiseOr;
1545 // scalar->vector promotion patterns (preferred slot to vector):
1546 def : Pat<(v16i8 (SPUprefslot2vec R8C:$rA)),
1547 (ORv16i8_i8 R8C:$rA)>;
1549 def : Pat<(v8i16 (SPUprefslot2vec R16C:$rA)),
1550 (ORv8i16_i16 R16C:$rA)>;
1552 def : Pat<(v4i32 (SPUprefslot2vec R32C:$rA)),
1553 (ORv4i32_i32 R32C:$rA)>;
1555 def : Pat<(v2i64 (SPUprefslot2vec R64C:$rA)),
1556 (ORv2i64_i64 R64C:$rA)>;
1558 def : Pat<(v4f32 (SPUprefslot2vec R32FP:$rA)),
1559 (ORv4f32_f32 R32FP:$rA)>;
1561 def : Pat<(v2f64 (SPUprefslot2vec R64FP:$rA)),
1562 (ORv2f64_f64 R64FP:$rA)>;
1564 // ORi*_v*: Used to extract vector element 0 (the preferred slot), otherwise
1565 // known as converting the vector back to its preferred slot
1567 def : Pat<(SPUvec2prefslot (v16i8 VECREG:$rA)),
1568 (ORi8_v16i8 VECREG:$rA)>;
1570 def : Pat<(SPUvec2prefslot (v8i16 VECREG:$rA)),
1571 (ORi16_v8i16 VECREG:$rA)>;
1573 def : Pat<(SPUvec2prefslot (v4i32 VECREG:$rA)),
1574 (ORi32_v4i32 VECREG:$rA)>;
1576 def : Pat<(SPUvec2prefslot (v2i64 VECREG:$rA)),
1577 (ORi64_v2i64 VECREG:$rA)>;
1579 def : Pat<(SPUvec2prefslot (v4f32 VECREG:$rA)),
1580 (ORf32_v4f32 VECREG:$rA)>;
1582 def : Pat<(SPUvec2prefslot (v2f64 VECREG:$rA)),
1583 (ORf64_v2f64 VECREG:$rA)>;
1585 // Load Register: This is an assembler alias for a bitwise OR of a register
1586 // against itself. It's here because it brings some clarity to assembly
1589 let hasCtrlDep = 1 in {
1590 class LRInst<dag OOL, dag IOL>
1591 : SPUInstr<OOL, IOL, "lr\t$rT, $rA", IntegerOp> {
1595 let Pattern = [/*no pattern*/];
1597 let Inst{0-10} = 0b10000010000; /* It's an OR operation */
1598 let Inst{11-17} = RA;
1599 let Inst{18-24} = RA;
1600 let Inst{25-31} = RT;
1603 class LRVecInst<ValueType vectype>:
1604 LRInst<(outs VECREG:$rT), (ins VECREG:$rA)>;
1606 class LRRegInst<RegisterClass rclass>:
1607 LRInst<(outs rclass:$rT), (ins rclass:$rA)>;
1609 multiclass LoadRegister {
1610 def v2i64: LRVecInst<v2i64>;
1611 def v2f64: LRVecInst<v2f64>;
1612 def v4i32: LRVecInst<v4i32>;
1613 def v4f32: LRVecInst<v4f32>;
1614 def v8i16: LRVecInst<v8i16>;
1615 def v16i8: LRVecInst<v16i8>;
1617 def r128: LRRegInst<GPRC>;
1618 def r64: LRRegInst<R64C>;
1619 def f64: LRRegInst<R64FP>;
1620 def r32: LRRegInst<R32C>;
1621 def f32: LRRegInst<R32FP>;
1622 def r16: LRRegInst<R16C>;
1623 def r8: LRRegInst<R8C>;
1626 defm LR: LoadRegister;
1629 // ORC: Bitwise "or" with complement (c = a | ~b)
1631 class ORCInst<dag OOL, dag IOL, list<dag> pattern>:
1632 RRForm<0b10010010000, OOL, IOL, "orc\t$rT, $rA, $rB",
1633 IntegerOp, pattern>;
1635 class ORCVecInst<ValueType vectype>:
1636 ORCInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1637 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1638 (vnot (vectype VECREG:$rB))))]>;
1640 class ORCRegInst<RegisterClass rclass>:
1641 ORCInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1642 [(set rclass:$rT, (or rclass:$rA, (not rclass:$rB)))]>;
1644 multiclass BitwiseOrComplement
1646 def v16i8: ORCVecInst<v16i8>;
1647 def v8i16: ORCVecInst<v8i16>;
1648 def v4i32: ORCVecInst<v4i32>;
1649 def v2i64: ORCVecInst<v2i64>;
1651 def r128: ORCRegInst<GPRC>;
1652 def r64: ORCRegInst<R64C>;
1653 def r32: ORCRegInst<R32C>;
1654 def r16: ORCRegInst<R16C>;
1655 def r8: ORCRegInst<R8C>;
1658 defm ORC : BitwiseOrComplement;
1660 // OR byte immediate
1661 class ORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1662 RI10Form<0b01100000, OOL, IOL, "orbi\t$rT, $rA, $val",
1663 IntegerOp, pattern>;
1665 class ORBIVecInst<ValueType vectype, PatLeaf immpred>:
1666 ORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1667 [(set (v16i8 VECREG:$rT), (or (vectype VECREG:$rA),
1668 (vectype immpred:$val)))]>;
1670 multiclass BitwiseOrByteImm
1672 def v16i8: ORBIVecInst<v16i8, v16i8U8Imm>;
1674 def r8: ORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1675 [(set R8C:$rT, (or R8C:$rA, immU8:$val))]>;
1678 defm ORBI : BitwiseOrByteImm;
1680 // OR halfword immediate
1681 class ORHIInst<dag OOL, dag IOL, list<dag> pattern>:
1682 RI10Form<0b10100000, OOL, IOL, "orhi\t$rT, $rA, $val",
1683 IntegerOp, pattern>;
1685 class ORHIVecInst<ValueType vectype, PatLeaf immpred>:
1686 ORHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1687 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1690 multiclass BitwiseOrHalfwordImm
1692 def v8i16: ORHIVecInst<v8i16, v8i16Uns10Imm>;
1694 def r16: ORHIInst<(outs R16C:$rT), (ins R16C:$rA, u10imm:$val),
1695 [(set R16C:$rT, (or R16C:$rA, i16ImmUns10:$val))]>;
1697 // Specialized ORHI form used to promote 8-bit registers to 16-bit
1698 def i8i16: ORHIInst<(outs R16C:$rT), (ins R8C:$rA, s10imm:$val),
1699 [(set R16C:$rT, (or (anyext R8C:$rA),
1700 i16ImmSExt10:$val))]>;
1703 defm ORHI : BitwiseOrHalfwordImm;
1705 class ORIInst<dag OOL, dag IOL, list<dag> pattern>:
1706 RI10Form<0b00100000, OOL, IOL, "ori\t$rT, $rA, $val",
1707 IntegerOp, pattern>;
1709 class ORIVecInst<ValueType vectype, PatLeaf immpred>:
1710 ORIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1711 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1714 // Bitwise "or" with immediate
1715 multiclass BitwiseOrImm
1717 def v4i32: ORIVecInst<v4i32, v4i32Uns10Imm>;
1719 def r32: ORIInst<(outs R32C:$rT), (ins R32C:$rA, u10imm_i32:$val),
1720 [(set R32C:$rT, (or R32C:$rA, i32ImmUns10:$val))]>;
1722 // i16i32: hacked version of the ori instruction to extend 16-bit quantities
1723 // to 32-bit quantities. used exclusively to match "anyext" conversions (vide
1724 // infra "anyext 16->32" pattern.)
1725 def i16i32: ORIInst<(outs R32C:$rT), (ins R16C:$rA, s10imm_i32:$val),
1726 [(set R32C:$rT, (or (anyext R16C:$rA),
1727 i32ImmSExt10:$val))]>;
1729 // i8i32: Hacked version of the ORI instruction to extend 16-bit quantities
1730 // to 32-bit quantities. Used exclusively to match "anyext" conversions (vide
1731 // infra "anyext 16->32" pattern.)
1732 def i8i32: ORIInst<(outs R32C:$rT), (ins R8C:$rA, s10imm_i32:$val),
1733 [(set R32C:$rT, (or (anyext R8C:$rA),
1734 i32ImmSExt10:$val))]>;
1737 defm ORI : BitwiseOrImm;
1739 // ORX: "or" across the vector: or's $rA's word slots leaving the result in
1740 // $rT[0], slots 1-3 are zeroed.
1742 // FIXME: Needs to match an intrinsic pattern.
1744 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1745 "orx\t$rT, $rA, $rB", IntegerOp,
1750 class XORInst<dag OOL, dag IOL, list<dag> pattern> :
1751 RRForm<0b10010010000, OOL, IOL, "xor\t$rT, $rA, $rB",
1752 IntegerOp, pattern>;
1754 class XORVecInst<ValueType vectype>:
1755 XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1756 [(set (vectype VECREG:$rT), (xor (vectype VECREG:$rA),
1757 (vectype VECREG:$rB)))]>;
1759 class XORRegInst<RegisterClass rclass>:
1760 XORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1761 [(set rclass:$rT, (xor rclass:$rA, rclass:$rB))]>;
1763 multiclass BitwiseExclusiveOr
1765 def v16i8: XORVecInst<v16i8>;
1766 def v8i16: XORVecInst<v8i16>;
1767 def v4i32: XORVecInst<v4i32>;
1768 def v2i64: XORVecInst<v2i64>;
1770 def r128: XORRegInst<GPRC>;
1771 def r64: XORRegInst<R64C>;
1772 def r32: XORRegInst<R32C>;
1773 def r16: XORRegInst<R16C>;
1774 def r8: XORRegInst<R8C>;
1776 // XOR instructions used to negate f32 and f64 quantities.
1778 def fneg32: XORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32C:$rB),
1779 [/* no pattern */]>;
1781 def fneg64: XORInst<(outs R64FP:$rT), (ins R64FP:$rA, R64C:$rB),
1782 [/* no pattern */]>;
1784 def fnegvec: XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1785 [/* no pattern, see fneg{32,64} */]>;
1788 defm XOR : BitwiseExclusiveOr;
1790 //==----------------------------------------------------------
1792 class XORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1793 RI10Form<0b01100000, OOL, IOL, "xorbi\t$rT, $rA, $val",
1794 IntegerOp, pattern>;
1796 multiclass XorByteImm
1799 XORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1800 [(set (v16i8 VECREG:$rT), (xor (v16i8 VECREG:$rA), v16i8U8Imm:$val))]>;
1803 XORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1804 [(set R8C:$rT, (xor R8C:$rA, immU8:$val))]>;
1807 defm XORBI : XorByteImm;
1810 RI10Form<0b10100000, (outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1811 "xorhi\t$rT, $rA, $val", IntegerOp,
1812 [(set (v8i16 VECREG:$rT), (xor (v8i16 VECREG:$rA),
1813 v8i16SExt10Imm:$val))]>;
1816 RI10Form<0b10100000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
1817 "xorhi\t$rT, $rA, $val", IntegerOp,
1818 [(set R16C:$rT, (xor R16C:$rA, i16ImmSExt10:$val))]>;
1821 RI10Form<0b00100000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm_i32:$val),
1822 "xori\t$rT, $rA, $val", IntegerOp,
1823 [(set (v4i32 VECREG:$rT), (xor (v4i32 VECREG:$rA),
1824 v4i32SExt10Imm:$val))]>;
1827 RI10Form<0b00100000, (outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
1828 "xori\t$rT, $rA, $val", IntegerOp,
1829 [(set R32C:$rT, (xor R32C:$rA, i32ImmSExt10:$val))]>;
1833 class NANDInst<dag OOL, dag IOL, list<dag> pattern>:
1834 RRForm<0b10010011000, OOL, IOL, "nand\t$rT, $rA, $rB",
1835 IntegerOp, pattern>;
1837 class NANDVecInst<ValueType vectype>:
1838 NANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1839 [(set (vectype VECREG:$rT), (vnot (and (vectype VECREG:$rA),
1840 (vectype VECREG:$rB))))]>;
1841 class NANDRegInst<RegisterClass rclass>:
1842 NANDInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1843 [(set rclass:$rT, (not (and rclass:$rA, rclass:$rB)))]>;
1845 multiclass BitwiseNand
1847 def v16i8: NANDVecInst<v16i8>;
1848 def v8i16: NANDVecInst<v8i16>;
1849 def v4i32: NANDVecInst<v4i32>;
1850 def v2i64: NANDVecInst<v2i64>;
1852 def r128: NANDRegInst<GPRC>;
1853 def r64: NANDRegInst<R64C>;
1854 def r32: NANDRegInst<R32C>;
1855 def r16: NANDRegInst<R16C>;
1856 def r8: NANDRegInst<R8C>;
1859 defm NAND : BitwiseNand;
1863 class NORInst<dag OOL, dag IOL, list<dag> pattern>:
1864 RRForm<0b10010010000, OOL, IOL, "nor\t$rT, $rA, $rB",
1865 IntegerOp, pattern>;
1867 class NORVecInst<ValueType vectype>:
1868 NORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1869 [(set (vectype VECREG:$rT), (vnot (or (vectype VECREG:$rA),
1870 (vectype VECREG:$rB))))]>;
1871 class NORRegInst<RegisterClass rclass>:
1872 NORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1873 [(set rclass:$rT, (not (or rclass:$rA, rclass:$rB)))]>;
1875 multiclass BitwiseNor
1877 def v16i8: NORVecInst<v16i8>;
1878 def v8i16: NORVecInst<v8i16>;
1879 def v4i32: NORVecInst<v4i32>;
1880 def v2i64: NORVecInst<v2i64>;
1882 def r128: NORRegInst<GPRC>;
1883 def r64: NORRegInst<R64C>;
1884 def r32: NORRegInst<R32C>;
1885 def r16: NORRegInst<R16C>;
1886 def r8: NORRegInst<R8C>;
1889 defm NOR : BitwiseNor;
1892 class SELBInst<dag OOL, dag IOL, list<dag> pattern>:
1893 RRRForm<0b1000, OOL, IOL, "selb\t$rT, $rA, $rB, $rC",
1894 IntegerOp, pattern>;
1896 class SELBVecInst<ValueType vectype, PatFrag vnot_frag = vnot>:
1897 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1898 [(set (vectype VECREG:$rT),
1899 (or (and (vectype VECREG:$rC), (vectype VECREG:$rB)),
1900 (and (vnot_frag (vectype VECREG:$rC)),
1901 (vectype VECREG:$rA))))]>;
1903 class SELBVecVCondInst<ValueType vectype>:
1904 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1905 [(set (vectype VECREG:$rT),
1906 (select (vectype VECREG:$rC),
1907 (vectype VECREG:$rB),
1908 (vectype VECREG:$rA)))]>;
1910 class SELBVecCondInst<ValueType vectype>:
1911 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, R32C:$rC),
1912 [(set (vectype VECREG:$rT),
1914 (vectype VECREG:$rB),
1915 (vectype VECREG:$rA)))]>;
1917 class SELBRegInst<RegisterClass rclass>:
1918 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rclass:$rC),
1920 (or (and rclass:$rB, rclass:$rC),
1921 (and rclass:$rA, (not rclass:$rC))))]>;
1923 class SELBRegCondInst<RegisterClass rcond, RegisterClass rclass>:
1924 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rcond:$rC),
1926 (select rcond:$rC, rclass:$rB, rclass:$rA))]>;
1928 multiclass SelectBits
1930 def v16i8: SELBVecInst<v16i8>;
1931 def v8i16: SELBVecInst<v8i16>;
1932 def v4i32: SELBVecInst<v4i32>;
1933 def v2i64: SELBVecInst<v2i64, vnot_cell_conv>;
1935 def r128: SELBRegInst<GPRC>;
1936 def r64: SELBRegInst<R64C>;
1937 def r32: SELBRegInst<R32C>;
1938 def r16: SELBRegInst<R16C>;
1939 def r8: SELBRegInst<R8C>;
1941 def v16i8_cond: SELBVecCondInst<v16i8>;
1942 def v8i16_cond: SELBVecCondInst<v8i16>;
1943 def v4i32_cond: SELBVecCondInst<v4i32>;
1944 def v2i64_cond: SELBVecCondInst<v2i64>;
1946 def v16i8_vcond: SELBVecCondInst<v16i8>;
1947 def v8i16_vcond: SELBVecCondInst<v8i16>;
1948 def v4i32_vcond: SELBVecCondInst<v4i32>;
1949 def v2i64_vcond: SELBVecCondInst<v2i64>;
1952 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1953 [(set (v4f32 VECREG:$rT),
1954 (select (v4i32 VECREG:$rC),
1956 (v4f32 VECREG:$rA)))]>;
1958 // SELBr64_cond is defined in SPU64InstrInfo.td
1959 def r32_cond: SELBRegCondInst<R32C, R32C>;
1960 def f32_cond: SELBRegCondInst<R32C, R32FP>;
1961 def r16_cond: SELBRegCondInst<R16C, R16C>;
1962 def r8_cond: SELBRegCondInst<R8C, R8C>;
1965 defm SELB : SelectBits;
1967 class SPUselbPatVec<ValueType vectype, SPUInstr inst>:
1968 Pat<(SPUselb (vectype VECREG:$rA), (vectype VECREG:$rB), (vectype VECREG:$rC)),
1969 (inst VECREG:$rA, VECREG:$rB, VECREG:$rC)>;
1971 def : SPUselbPatVec<v16i8, SELBv16i8>;
1972 def : SPUselbPatVec<v8i16, SELBv8i16>;
1973 def : SPUselbPatVec<v4i32, SELBv4i32>;
1974 def : SPUselbPatVec<v2i64, SELBv2i64>;
1976 class SPUselbPatReg<RegisterClass rclass, SPUInstr inst>:
1977 Pat<(SPUselb rclass:$rA, rclass:$rB, rclass:$rC),
1978 (inst rclass:$rA, rclass:$rB, rclass:$rC)>;
1980 def : SPUselbPatReg<R8C, SELBr8>;
1981 def : SPUselbPatReg<R16C, SELBr16>;
1982 def : SPUselbPatReg<R32C, SELBr32>;
1983 def : SPUselbPatReg<R64C, SELBr64>;
1985 // EQV: Equivalence (1 for each same bit, otherwise 0)
1987 // Note: There are a lot of ways to match this bit operator and these patterns
1988 // attempt to be as exhaustive as possible.
1990 class EQVInst<dag OOL, dag IOL, list<dag> pattern>:
1991 RRForm<0b10010010000, OOL, IOL, "eqv\t$rT, $rA, $rB",
1992 IntegerOp, pattern>;
1994 class EQVVecInst<ValueType vectype>:
1995 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1996 [(set (vectype VECREG:$rT),
1997 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
1998 (and (vnot (vectype VECREG:$rA)),
1999 (vnot (vectype VECREG:$rB)))))]>;
2001 class EQVRegInst<RegisterClass rclass>:
2002 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2003 [(set rclass:$rT, (or (and rclass:$rA, rclass:$rB),
2004 (and (not rclass:$rA), (not rclass:$rB))))]>;
2006 class EQVVecPattern1<ValueType vectype>:
2007 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2008 [(set (vectype VECREG:$rT),
2009 (xor (vectype VECREG:$rA), (vnot (vectype VECREG:$rB))))]>;
2011 class EQVRegPattern1<RegisterClass rclass>:
2012 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2013 [(set rclass:$rT, (xor rclass:$rA, (not rclass:$rB)))]>;
2015 class EQVVecPattern2<ValueType vectype>:
2016 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2017 [(set (vectype VECREG:$rT),
2018 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
2019 (vnot (or (vectype VECREG:$rA), (vectype VECREG:$rB)))))]>;
2021 class EQVRegPattern2<RegisterClass rclass>:
2022 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2024 (or (and rclass:$rA, rclass:$rB),
2025 (not (or rclass:$rA, rclass:$rB))))]>;
2027 class EQVVecPattern3<ValueType vectype>:
2028 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2029 [(set (vectype VECREG:$rT),
2030 (not (xor (vectype VECREG:$rA), (vectype VECREG:$rB))))]>;
2032 class EQVRegPattern3<RegisterClass rclass>:
2033 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2034 [(set rclass:$rT, (not (xor rclass:$rA, rclass:$rB)))]>;
2036 multiclass BitEquivalence
2038 def v16i8: EQVVecInst<v16i8>;
2039 def v8i16: EQVVecInst<v8i16>;
2040 def v4i32: EQVVecInst<v4i32>;
2041 def v2i64: EQVVecInst<v2i64>;
2043 def v16i8_1: EQVVecPattern1<v16i8>;
2044 def v8i16_1: EQVVecPattern1<v8i16>;
2045 def v4i32_1: EQVVecPattern1<v4i32>;
2046 def v2i64_1: EQVVecPattern1<v2i64>;
2048 def v16i8_2: EQVVecPattern2<v16i8>;
2049 def v8i16_2: EQVVecPattern2<v8i16>;
2050 def v4i32_2: EQVVecPattern2<v4i32>;
2051 def v2i64_2: EQVVecPattern2<v2i64>;
2053 def v16i8_3: EQVVecPattern3<v16i8>;
2054 def v8i16_3: EQVVecPattern3<v8i16>;
2055 def v4i32_3: EQVVecPattern3<v4i32>;
2056 def v2i64_3: EQVVecPattern3<v2i64>;
2058 def r128: EQVRegInst<GPRC>;
2059 def r64: EQVRegInst<R64C>;
2060 def r32: EQVRegInst<R32C>;
2061 def r16: EQVRegInst<R16C>;
2062 def r8: EQVRegInst<R8C>;
2064 def r128_1: EQVRegPattern1<GPRC>;
2065 def r64_1: EQVRegPattern1<R64C>;
2066 def r32_1: EQVRegPattern1<R32C>;
2067 def r16_1: EQVRegPattern1<R16C>;
2068 def r8_1: EQVRegPattern1<R8C>;
2070 def r128_2: EQVRegPattern2<GPRC>;
2071 def r64_2: EQVRegPattern2<R64C>;
2072 def r32_2: EQVRegPattern2<R32C>;
2073 def r16_2: EQVRegPattern2<R16C>;
2074 def r8_2: EQVRegPattern2<R8C>;
2076 def r128_3: EQVRegPattern3<GPRC>;
2077 def r64_3: EQVRegPattern3<R64C>;
2078 def r32_3: EQVRegPattern3<R32C>;
2079 def r16_3: EQVRegPattern3<R16C>;
2080 def r8_3: EQVRegPattern3<R8C>;
2083 defm EQV: BitEquivalence;
2085 //===----------------------------------------------------------------------===//
2086 // Vector shuffle...
2087 //===----------------------------------------------------------------------===//
2088 // SPUshuffle is generated in LowerVECTOR_SHUFFLE and gets replaced with SHUFB.
2089 // See the SPUshuffle SDNode operand above, which sets up the DAG pattern
2090 // matcher to emit something when the LowerVECTOR_SHUFFLE generates a node with
2091 // the SPUISD::SHUFB opcode.
2092 //===----------------------------------------------------------------------===//
2094 class SHUFBInst<dag OOL, dag IOL, list<dag> pattern>:
2095 RRRForm<0b1000, OOL, IOL, "shufb\t$rT, $rA, $rB, $rC",
2096 IntegerOp, pattern>;
2098 class SHUFBVecInst<ValueType resultvec, ValueType maskvec>:
2099 SHUFBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
2100 [(set (resultvec VECREG:$rT),
2101 (SPUshuffle (resultvec VECREG:$rA),
2102 (resultvec VECREG:$rB),
2103 (maskvec VECREG:$rC)))]>;
2105 class SHUFBGPRCInst:
2106 SHUFBInst<(outs VECREG:$rT), (ins GPRC:$rA, GPRC:$rB, VECREG:$rC),
2107 [/* no pattern */]>;
2109 multiclass ShuffleBytes
2111 def v16i8 : SHUFBVecInst<v16i8, v16i8>;
2112 def v16i8_m32 : SHUFBVecInst<v16i8, v4i32>;
2113 def v8i16 : SHUFBVecInst<v8i16, v16i8>;
2114 def v8i16_m32 : SHUFBVecInst<v8i16, v4i32>;
2115 def v4i32 : SHUFBVecInst<v4i32, v16i8>;
2116 def v4i32_m32 : SHUFBVecInst<v4i32, v4i32>;
2117 def v2i64 : SHUFBVecInst<v2i64, v16i8>;
2118 def v2i64_m32 : SHUFBVecInst<v2i64, v4i32>;
2120 def v4f32 : SHUFBVecInst<v4f32, v16i8>;
2121 def v4f32_m32 : SHUFBVecInst<v4f32, v4i32>;
2123 def v2f64 : SHUFBVecInst<v2f64, v16i8>;
2124 def v2f64_m32 : SHUFBVecInst<v2f64, v4i32>;
2126 def gprc : SHUFBGPRCInst;
2129 defm SHUFB : ShuffleBytes;
2131 //===----------------------------------------------------------------------===//
2132 // Shift and rotate group:
2133 //===----------------------------------------------------------------------===//
2135 class SHLHInst<dag OOL, dag IOL, list<dag> pattern>:
2136 RRForm<0b11111010000, OOL, IOL, "shlh\t$rT, $rA, $rB",
2137 RotateShift, pattern>;
2139 class SHLHVecInst<ValueType vectype>:
2140 SHLHInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2141 [(set (vectype VECREG:$rT),
2142 (SPUvec_shl (vectype VECREG:$rA), R16C:$rB))]>;
2144 multiclass ShiftLeftHalfword
2146 def v8i16: SHLHVecInst<v8i16>;
2147 def r16: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
2148 [(set R16C:$rT, (shl R16C:$rA, R16C:$rB))]>;
2149 def r16_r32: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2150 [(set R16C:$rT, (shl R16C:$rA, R32C:$rB))]>;
2153 defm SHLH : ShiftLeftHalfword;
2155 //===----------------------------------------------------------------------===//
2157 class SHLHIInst<dag OOL, dag IOL, list<dag> pattern>:
2158 RI7Form<0b11111010000, OOL, IOL, "shlhi\t$rT, $rA, $val",
2159 RotateShift, pattern>;
2161 class SHLHIVecInst<ValueType vectype>:
2162 SHLHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2163 [(set (vectype VECREG:$rT),
2164 (SPUvec_shl (vectype VECREG:$rA), (i16 uimm7:$val)))]>;
2166 multiclass ShiftLeftHalfwordImm
2168 def v8i16: SHLHIVecInst<v8i16>;
2169 def r16: SHLHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2170 [(set R16C:$rT, (shl R16C:$rA, (i16 uimm7:$val)))]>;
2173 defm SHLHI : ShiftLeftHalfwordImm;
2175 def : Pat<(SPUvec_shl (v8i16 VECREG:$rA), (i32 uimm7:$val)),
2176 (SHLHIv8i16 VECREG:$rA, (TO_IMM16 uimm7:$val))>;
2178 def : Pat<(shl R16C:$rA, (i32 uimm7:$val)),
2179 (SHLHIr16 R16C:$rA, (TO_IMM16 uimm7:$val))>;
2181 //===----------------------------------------------------------------------===//
2183 class SHLInst<dag OOL, dag IOL, list<dag> pattern>:
2184 RRForm<0b11111010000, OOL, IOL, "shl\t$rT, $rA, $rB",
2185 RotateShift, pattern>;
2187 multiclass ShiftLeftWord
2190 SHLInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2191 [(set (v4i32 VECREG:$rT),
2192 (SPUvec_shl (v4i32 VECREG:$rA), R16C:$rB))]>;
2194 SHLInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2195 [(set R32C:$rT, (shl R32C:$rA, R32C:$rB))]>;
2198 defm SHL: ShiftLeftWord;
2200 //===----------------------------------------------------------------------===//
2202 class SHLIInst<dag OOL, dag IOL, list<dag> pattern>:
2203 RI7Form<0b11111010000, OOL, IOL, "shli\t$rT, $rA, $val",
2204 RotateShift, pattern>;
2206 multiclass ShiftLeftWordImm
2209 SHLIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2210 [(set (v4i32 VECREG:$rT),
2211 (SPUvec_shl (v4i32 VECREG:$rA), (i32 uimm7:$val)))]>;
2214 SHLIInst<(outs R32C:$rT), (ins R32C:$rA, u7imm_i32:$val),
2215 [(set R32C:$rT, (shl R32C:$rA, (i32 uimm7:$val)))]>;
2218 defm SHLI : ShiftLeftWordImm;
2220 //===----------------------------------------------------------------------===//
2221 // SHLQBI vec form: Note that this will shift the entire vector (the 128-bit
2222 // register) to the left. Vector form is here to ensure type correctness.
2224 // The shift count is in the lowest 3 bits (29-31) of $rB, so only a bit shift
2225 // of 7 bits is actually possible.
2227 // Note also that SHLQBI/SHLQBII are used in conjunction with SHLQBY/SHLQBYI
2228 // to shift i64 and i128. SHLQBI is the residual left over after shifting by
2229 // bytes with SHLQBY.
2231 class SHLQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2232 RRForm<0b11011011100, OOL, IOL, "shlqbi\t$rT, $rA, $rB",
2233 RotateShift, pattern>;
2235 class SHLQBIVecInst<ValueType vectype>:
2236 SHLQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2237 [(set (vectype VECREG:$rT),
2238 (SPUshlquad_l_bits (vectype VECREG:$rA), R32C:$rB))]>;
2240 class SHLQBIRegInst<RegisterClass rclass>:
2241 SHLQBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2242 [/* no pattern */]>;
2244 multiclass ShiftLeftQuadByBits
2246 def v16i8: SHLQBIVecInst<v16i8>;
2247 def v8i16: SHLQBIVecInst<v8i16>;
2248 def v4i32: SHLQBIVecInst<v4i32>;
2249 def v4f32: SHLQBIVecInst<v4f32>;
2250 def v2i64: SHLQBIVecInst<v2i64>;
2251 def v2f64: SHLQBIVecInst<v2f64>;
2253 def r128: SHLQBIRegInst<GPRC>;
2256 defm SHLQBI : ShiftLeftQuadByBits;
2258 // See note above on SHLQBI. In this case, the predicate actually does then
2259 // enforcement, whereas with SHLQBI, we have to "take it on faith."
2260 class SHLQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2261 RI7Form<0b11011111100, OOL, IOL, "shlqbii\t$rT, $rA, $val",
2262 RotateShift, pattern>;
2264 class SHLQBIIVecInst<ValueType vectype>:
2265 SHLQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2266 [(set (vectype VECREG:$rT),
2267 (SPUshlquad_l_bits (vectype VECREG:$rA), (i32 bitshift:$val)))]>;
2269 multiclass ShiftLeftQuadByBitsImm
2271 def v16i8 : SHLQBIIVecInst<v16i8>;
2272 def v8i16 : SHLQBIIVecInst<v8i16>;
2273 def v4i32 : SHLQBIIVecInst<v4i32>;
2274 def v4f32 : SHLQBIIVecInst<v4f32>;
2275 def v2i64 : SHLQBIIVecInst<v2i64>;
2276 def v2f64 : SHLQBIIVecInst<v2f64>;
2279 defm SHLQBII : ShiftLeftQuadByBitsImm;
2281 // SHLQBY, SHLQBYI vector forms: Shift the entire vector to the left by bytes,
2282 // not by bits. See notes above on SHLQBI.
2284 class SHLQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2285 RI7Form<0b11111011100, OOL, IOL, "shlqby\t$rT, $rA, $rB",
2286 RotateShift, pattern>;
2288 class SHLQBYVecInst<ValueType vectype>:
2289 SHLQBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2290 [(set (vectype VECREG:$rT),
2291 (SPUshlquad_l_bytes (vectype VECREG:$rA), R32C:$rB))]>;
2293 multiclass ShiftLeftQuadBytes
2295 def v16i8: SHLQBYVecInst<v16i8>;
2296 def v8i16: SHLQBYVecInst<v8i16>;
2297 def v4i32: SHLQBYVecInst<v4i32>;
2298 def v4f32: SHLQBYVecInst<v4f32>;
2299 def v2i64: SHLQBYVecInst<v2i64>;
2300 def v2f64: SHLQBYVecInst<v2f64>;
2301 def r128: SHLQBYInst<(outs GPRC:$rT), (ins GPRC:$rA, R32C:$rB),
2302 [(set GPRC:$rT, (SPUshlquad_l_bytes GPRC:$rA, R32C:$rB))]>;
2305 defm SHLQBY: ShiftLeftQuadBytes;
2307 class SHLQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2308 RI7Form<0b11111111100, OOL, IOL, "shlqbyi\t$rT, $rA, $val",
2309 RotateShift, pattern>;
2311 class SHLQBYIVecInst<ValueType vectype>:
2312 SHLQBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2313 [(set (vectype VECREG:$rT),
2314 (SPUshlquad_l_bytes (vectype VECREG:$rA), (i32 uimm7:$val)))]>;
2316 multiclass ShiftLeftQuadBytesImm
2318 def v16i8: SHLQBYIVecInst<v16i8>;
2319 def v8i16: SHLQBYIVecInst<v8i16>;
2320 def v4i32: SHLQBYIVecInst<v4i32>;
2321 def v4f32: SHLQBYIVecInst<v4f32>;
2322 def v2i64: SHLQBYIVecInst<v2i64>;
2323 def v2f64: SHLQBYIVecInst<v2f64>;
2324 def r128: SHLQBYIInst<(outs GPRC:$rT), (ins GPRC:$rA, u7imm_i32:$val),
2326 (SPUshlquad_l_bytes GPRC:$rA, (i32 uimm7:$val)))]>;
2329 defm SHLQBYI : ShiftLeftQuadBytesImm;
2331 class SHLQBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2332 RRForm<0b00111001111, OOL, IOL, "shlqbybi\t$rT, $rA, $rB",
2333 RotateShift, pattern>;
2335 class SHLQBYBIVecInst<ValueType vectype>:
2336 SHLQBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2337 [/* no pattern */]>;
2339 class SHLQBYBIRegInst<RegisterClass rclass>:
2340 SHLQBYBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2341 [/* no pattern */]>;
2343 multiclass ShiftLeftQuadBytesBitCount
2345 def v16i8: SHLQBYBIVecInst<v16i8>;
2346 def v8i16: SHLQBYBIVecInst<v8i16>;
2347 def v4i32: SHLQBYBIVecInst<v4i32>;
2348 def v4f32: SHLQBYBIVecInst<v4f32>;
2349 def v2i64: SHLQBYBIVecInst<v2i64>;
2350 def v2f64: SHLQBYBIVecInst<v2f64>;
2352 def r128: SHLQBYBIRegInst<GPRC>;
2355 defm SHLQBYBI : ShiftLeftQuadBytesBitCount;
2357 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2359 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2360 class ROTHInst<dag OOL, dag IOL, list<dag> pattern>:
2361 RRForm<0b00111010000, OOL, IOL, "roth\t$rT, $rA, $rB",
2362 RotateShift, pattern>;
2364 class ROTHVecInst<ValueType vectype>:
2365 ROTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2366 [(set (vectype VECREG:$rT),
2367 (SPUvec_rotl VECREG:$rA, (v8i16 VECREG:$rB)))]>;
2369 class ROTHRegInst<RegisterClass rclass>:
2370 ROTHInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2371 [(set rclass:$rT, (rotl rclass:$rA, rclass:$rB))]>;
2373 multiclass RotateLeftHalfword
2375 def v8i16: ROTHVecInst<v8i16>;
2376 def r16: ROTHRegInst<R16C>;
2379 defm ROTH: RotateLeftHalfword;
2381 def ROTHr16_r32: ROTHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2382 [(set R16C:$rT, (rotl R16C:$rA, R32C:$rB))]>;
2384 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2385 // Rotate halfword, immediate:
2386 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2387 class ROTHIInst<dag OOL, dag IOL, list<dag> pattern>:
2388 RI7Form<0b00111110000, OOL, IOL, "rothi\t$rT, $rA, $val",
2389 RotateShift, pattern>;
2391 class ROTHIVecInst<ValueType vectype>:
2392 ROTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2393 [(set (vectype VECREG:$rT),
2394 (SPUvec_rotl VECREG:$rA, (i16 uimm7:$val)))]>;
2396 multiclass RotateLeftHalfwordImm
2398 def v8i16: ROTHIVecInst<v8i16>;
2399 def r16: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2400 [(set R16C:$rT, (rotl R16C:$rA, (i16 uimm7:$val)))]>;
2401 def r16_r32: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm_i32:$val),
2402 [(set R16C:$rT, (rotl R16C:$rA, (i32 uimm7:$val)))]>;
2405 defm ROTHI: RotateLeftHalfwordImm;
2407 def : Pat<(SPUvec_rotl (v8i16 VECREG:$rA), (i32 uimm7:$val)),
2408 (ROTHIv8i16 VECREG:$rA, (TO_IMM16 imm:$val))>;
2410 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2412 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2414 class ROTInst<dag OOL, dag IOL, list<dag> pattern>:
2415 RRForm<0b00011010000, OOL, IOL, "rot\t$rT, $rA, $rB",
2416 RotateShift, pattern>;
2418 class ROTVecInst<ValueType vectype>:
2419 ROTInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2420 [(set (vectype VECREG:$rT),
2421 (SPUvec_rotl (vectype VECREG:$rA), R32C:$rB))]>;
2423 class ROTRegInst<RegisterClass rclass>:
2424 ROTInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2426 (rotl rclass:$rA, R32C:$rB))]>;
2428 multiclass RotateLeftWord
2430 def v4i32: ROTVecInst<v4i32>;
2431 def r32: ROTRegInst<R32C>;
2434 defm ROT: RotateLeftWord;
2436 // The rotate amount is in the same bits whether we've got an 8-bit, 16-bit or
2438 def ROTr32_r16_anyext:
2439 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R16C:$rB),
2440 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R16C:$rB))))]>;
2442 def : Pat<(rotl R32C:$rA, (i32 (zext R16C:$rB))),
2443 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2445 def : Pat<(rotl R32C:$rA, (i32 (sext R16C:$rB))),
2446 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2448 def ROTr32_r8_anyext:
2449 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R8C:$rB),
2450 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R8C:$rB))))]>;
2452 def : Pat<(rotl R32C:$rA, (i32 (zext R8C:$rB))),
2453 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2455 def : Pat<(rotl R32C:$rA, (i32 (sext R8C:$rB))),
2456 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2458 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2459 // Rotate word, immediate
2460 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2462 class ROTIInst<dag OOL, dag IOL, list<dag> pattern>:
2463 RI7Form<0b00011110000, OOL, IOL, "roti\t$rT, $rA, $val",
2464 RotateShift, pattern>;
2466 class ROTIVecInst<ValueType vectype, Operand optype, ValueType inttype, PatLeaf pred>:
2467 ROTIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2468 [(set (vectype VECREG:$rT),
2469 (SPUvec_rotl (vectype VECREG:$rA), (inttype pred:$val)))]>;
2471 class ROTIRegInst<RegisterClass rclass, Operand optype, ValueType inttype, PatLeaf pred>:
2472 ROTIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2473 [(set rclass:$rT, (rotl rclass:$rA, (inttype pred:$val)))]>;
2475 multiclass RotateLeftWordImm
2477 def v4i32: ROTIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2478 def v4i32_i16: ROTIVecInst<v4i32, u7imm, i16, uimm7>;
2479 def v4i32_i8: ROTIVecInst<v4i32, u7imm_i8, i8, uimm7>;
2481 def r32: ROTIRegInst<R32C, u7imm_i32, i32, uimm7>;
2482 def r32_i16: ROTIRegInst<R32C, u7imm, i16, uimm7>;
2483 def r32_i8: ROTIRegInst<R32C, u7imm_i8, i8, uimm7>;
2486 defm ROTI : RotateLeftWordImm;
2488 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2489 // Rotate quad by byte (count)
2490 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2492 class ROTQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2493 RRForm<0b00111011100, OOL, IOL, "rotqby\t$rT, $rA, $rB",
2494 RotateShift, pattern>;
2496 class ROTQBYVecInst<ValueType vectype>:
2497 ROTQBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2498 [(set (vectype VECREG:$rT),
2499 (SPUrotbytes_left (vectype VECREG:$rA), R32C:$rB))]>;
2501 multiclass RotateQuadLeftByBytes
2503 def v16i8: ROTQBYVecInst<v16i8>;
2504 def v8i16: ROTQBYVecInst<v8i16>;
2505 def v4i32: ROTQBYVecInst<v4i32>;
2506 def v4f32: ROTQBYVecInst<v4f32>;
2507 def v2i64: ROTQBYVecInst<v2i64>;
2508 def v2f64: ROTQBYVecInst<v2f64>;
2511 defm ROTQBY: RotateQuadLeftByBytes;
2513 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2514 // Rotate quad by byte (count), immediate
2515 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2517 class ROTQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2518 RI7Form<0b00111111100, OOL, IOL, "rotqbyi\t$rT, $rA, $val",
2519 RotateShift, pattern>;
2521 class ROTQBYIVecInst<ValueType vectype>:
2522 ROTQBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2523 [(set (vectype VECREG:$rT),
2524 (SPUrotbytes_left (vectype VECREG:$rA), (i16 uimm7:$val)))]>;
2526 multiclass RotateQuadByBytesImm
2528 def v16i8: ROTQBYIVecInst<v16i8>;
2529 def v8i16: ROTQBYIVecInst<v8i16>;
2530 def v4i32: ROTQBYIVecInst<v4i32>;
2531 def v4f32: ROTQBYIVecInst<v4f32>;
2532 def v2i64: ROTQBYIVecInst<v2i64>;
2533 def vfi64: ROTQBYIVecInst<v2f64>;
2536 defm ROTQBYI: RotateQuadByBytesImm;
2538 // See ROTQBY note above.
2539 class ROTQBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2540 RI7Form<0b00110011100, OOL, IOL,
2541 "rotqbybi\t$rT, $rA, $shift",
2542 RotateShift, pattern>;
2544 class ROTQBYBIVecInst<ValueType vectype, RegisterClass rclass>:
2545 ROTQBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, rclass:$shift),
2546 [(set (vectype VECREG:$rT),
2547 (SPUrotbytes_left_bits (vectype VECREG:$rA), rclass:$shift))]>;
2549 multiclass RotateQuadByBytesByBitshift {
2550 def v16i8_r32: ROTQBYBIVecInst<v16i8, R32C>;
2551 def v8i16_r32: ROTQBYBIVecInst<v8i16, R32C>;
2552 def v4i32_r32: ROTQBYBIVecInst<v4i32, R32C>;
2553 def v2i64_r32: ROTQBYBIVecInst<v2i64, R32C>;
2556 defm ROTQBYBI : RotateQuadByBytesByBitshift;
2558 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2559 // See ROTQBY note above.
2561 // Assume that the user of this instruction knows to shift the rotate count
2563 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2565 class ROTQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2566 RRForm<0b00011011100, OOL, IOL, "rotqbi\t$rT, $rA, $rB",
2567 RotateShift, pattern>;
2569 class ROTQBIVecInst<ValueType vectype>:
2570 ROTQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2571 [/* no pattern yet */]>;
2573 class ROTQBIRegInst<RegisterClass rclass>:
2574 ROTQBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2575 [/* no pattern yet */]>;
2577 multiclass RotateQuadByBitCount
2579 def v16i8: ROTQBIVecInst<v16i8>;
2580 def v8i16: ROTQBIVecInst<v8i16>;
2581 def v4i32: ROTQBIVecInst<v4i32>;
2582 def v2i64: ROTQBIVecInst<v2i64>;
2584 def r128: ROTQBIRegInst<GPRC>;
2585 def r64: ROTQBIRegInst<R64C>;
2588 defm ROTQBI: RotateQuadByBitCount;
2590 class ROTQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2591 RI7Form<0b00011111100, OOL, IOL, "rotqbii\t$rT, $rA, $val",
2592 RotateShift, pattern>;
2594 class ROTQBIIVecInst<ValueType vectype, Operand optype, ValueType inttype,
2596 ROTQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2597 [/* no pattern yet */]>;
2599 class ROTQBIIRegInst<RegisterClass rclass, Operand optype, ValueType inttype,
2601 ROTQBIIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2602 [/* no pattern yet */]>;
2604 multiclass RotateQuadByBitCountImm
2606 def v16i8: ROTQBIIVecInst<v16i8, u7imm_i32, i32, uimm7>;
2607 def v8i16: ROTQBIIVecInst<v8i16, u7imm_i32, i32, uimm7>;
2608 def v4i32: ROTQBIIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2609 def v2i64: ROTQBIIVecInst<v2i64, u7imm_i32, i32, uimm7>;
2611 def r128: ROTQBIIRegInst<GPRC, u7imm_i32, i32, uimm7>;
2612 def r64: ROTQBIIRegInst<R64C, u7imm_i32, i32, uimm7>;
2615 defm ROTQBII : RotateQuadByBitCountImm;
2617 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2618 // ROTHM v8i16 form:
2619 // NOTE(1): No vector rotate is generated by the C/C++ frontend (today),
2620 // so this only matches a synthetically generated/lowered code
2622 // NOTE(2): $rB must be negated before the right rotate!
2623 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2625 class ROTHMInst<dag OOL, dag IOL, list<dag> pattern>:
2626 RRForm<0b10111010000, OOL, IOL, "rothm\t$rT, $rA, $rB",
2627 RotateShift, pattern>;
2630 ROTHMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2631 [/* see patterns below - $rB must be negated */]>;
2633 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R32C:$rB),
2634 (ROTHMv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2636 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R16C:$rB),
2637 (ROTHMv8i16 VECREG:$rA,
2638 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2640 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R8C:$rB),
2641 (ROTHMv8i16 VECREG:$rA,
2642 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2644 // ROTHM r16 form: Rotate 16-bit quantity to right, zero fill at the left
2645 // Note: This instruction doesn't match a pattern because rB must be negated
2646 // for the instruction to work. Thus, the pattern below the instruction!
2649 ROTHMInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2650 [/* see patterns below - $rB must be negated! */]>;
2652 def : Pat<(srl R16C:$rA, R32C:$rB),
2653 (ROTHMr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2655 def : Pat<(srl R16C:$rA, R16C:$rB),
2657 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2659 def : Pat<(srl R16C:$rA, R8C:$rB),
2661 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2663 // ROTHMI v8i16 form: See the comment for ROTHM v8i16. The difference here is
2664 // that the immediate can be complemented, so that the user doesn't have to
2667 class ROTHMIInst<dag OOL, dag IOL, list<dag> pattern>:
2668 RI7Form<0b10111110000, OOL, IOL, "rothmi\t$rT, $rA, $val",
2669 RotateShift, pattern>;
2672 ROTHMIInst<(outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2673 [/* no pattern */]>;
2675 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i32 imm:$val)),
2676 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2678 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i16 imm:$val)),
2679 (ROTHMIv8i16 VECREG:$rA, (TO_IMM32 imm:$val))>;
2681 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i8 imm:$val)),
2682 (ROTHMIv8i16 VECREG:$rA, (TO_IMM32 imm:$val))>;
2685 ROTHMIInst<(outs R16C:$rT), (ins R16C:$rA, rothNeg7imm:$val),
2686 [/* no pattern */]>;
2688 def: Pat<(srl R16C:$rA, (i32 uimm7:$val)),
2689 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2691 def: Pat<(srl R16C:$rA, (i16 uimm7:$val)),
2692 (ROTHMIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
2694 def: Pat<(srl R16C:$rA, (i8 uimm7:$val)),
2695 (ROTHMIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
2697 // ROTM v4i32 form: See the ROTHM v8i16 comments.
2698 class ROTMInst<dag OOL, dag IOL, list<dag> pattern>:
2699 RRForm<0b10011010000, OOL, IOL, "rotm\t$rT, $rA, $rB",
2700 RotateShift, pattern>;
2703 ROTMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2704 [/* see patterns below - $rB must be negated */]>;
2706 def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), R32C:$rB),
2707 (ROTMv4i32 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2709 def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), R16C:$rB),
2710 (ROTMv4i32 VECREG:$rA,
2711 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2713 def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), R8C:$rB),
2714 (ROTMv4i32 VECREG:$rA,
2715 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2718 ROTMInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2719 [/* see patterns below - $rB must be negated */]>;
2721 def : Pat<(srl R32C:$rA, R32C:$rB),
2722 (ROTMr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
2724 def : Pat<(srl R32C:$rA, R16C:$rB),
2726 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2728 def : Pat<(srl R32C:$rA, R8C:$rB),
2730 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2732 // ROTMI v4i32 form: See the comment for ROTHM v8i16.
2734 RI7Form<0b10011110000, (outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2735 "rotmi\t$rT, $rA, $val", RotateShift,
2736 [(set (v4i32 VECREG:$rT),
2737 (SPUvec_srl VECREG:$rA, (i32 uimm7:$val)))]>;
2739 def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), (i16 uimm7:$val)),
2740 (ROTMIv4i32 VECREG:$rA, (TO_IMM32 uimm7:$val))>;
2742 def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), (i8 uimm7:$val)),
2743 (ROTMIv4i32 VECREG:$rA, (TO_IMM32 uimm7:$val))>;
2745 // ROTMI r32 form: know how to complement the immediate value.
2747 RI7Form<0b10011110000, (outs R32C:$rT), (ins R32C:$rA, rotNeg7imm:$val),
2748 "rotmi\t$rT, $rA, $val", RotateShift,
2749 [(set R32C:$rT, (srl R32C:$rA, (i32 uimm7:$val)))]>;
2751 def : Pat<(srl R32C:$rA, (i16 imm:$val)),
2752 (ROTMIr32 R32C:$rA, (TO_IMM32 uimm7:$val))>;
2754 def : Pat<(srl R32C:$rA, (i8 imm:$val)),
2755 (ROTMIr32 R32C:$rA, (TO_IMM32 uimm7:$val))>;
2757 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2758 // ROTQMBY: This is a vector form merely so that when used in an
2759 // instruction pattern, type checking will succeed. This instruction assumes
2760 // that the user knew to negate $rB.
2761 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2763 class ROTQMBYInst<dag OOL, dag IOL, list<dag> pattern>:
2764 RRForm<0b10111011100, OOL, IOL, "rotqmby\t$rT, $rA, $rB",
2765 RotateShift, pattern>;
2767 class ROTQMBYVecInst<ValueType vectype>:
2768 ROTQMBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2769 [/* no pattern, $rB must be negated */]>;
2771 class ROTQMBYRegInst<RegisterClass rclass>:
2772 ROTQMBYInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2773 [/* no pattern */]>;
2775 multiclass RotateQuadBytes
2777 def v16i8: ROTQMBYVecInst<v16i8>;
2778 def v8i16: ROTQMBYVecInst<v8i16>;
2779 def v4i32: ROTQMBYVecInst<v4i32>;
2780 def v2i64: ROTQMBYVecInst<v2i64>;
2782 def r128: ROTQMBYRegInst<GPRC>;
2783 def r64: ROTQMBYRegInst<R64C>;
2786 defm ROTQMBY : RotateQuadBytes;
2788 class ROTQMBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2789 RI7Form<0b10111111100, OOL, IOL, "rotqmbyi\t$rT, $rA, $val",
2790 RotateShift, pattern>;
2792 class ROTQMBYIVecInst<ValueType vectype>:
2793 ROTQMBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2794 [/* no pattern */]>;
2796 class ROTQMBYIRegInst<RegisterClass rclass, Operand optype, ValueType inttype,
2798 ROTQMBYIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2799 [/* no pattern */]>;
2801 // 128-bit zero extension form:
2802 class ROTQMBYIZExtInst<RegisterClass rclass, Operand optype, PatLeaf pred>:
2803 ROTQMBYIInst<(outs GPRC:$rT), (ins rclass:$rA, optype:$val),
2804 [/* no pattern */]>;
2806 multiclass RotateQuadBytesImm
2808 def v16i8: ROTQMBYIVecInst<v16i8>;
2809 def v8i16: ROTQMBYIVecInst<v8i16>;
2810 def v4i32: ROTQMBYIVecInst<v4i32>;
2811 def v2i64: ROTQMBYIVecInst<v2i64>;
2813 def r128: ROTQMBYIRegInst<GPRC, rotNeg7imm, i32, uimm7>;
2814 def r64: ROTQMBYIRegInst<R64C, rotNeg7imm, i32, uimm7>;
2816 def r128_zext_r8: ROTQMBYIZExtInst<R8C, rotNeg7imm, uimm7>;
2817 def r128_zext_r16: ROTQMBYIZExtInst<R16C, rotNeg7imm, uimm7>;
2818 def r128_zext_r32: ROTQMBYIZExtInst<R32C, rotNeg7imm, uimm7>;
2819 def r128_zext_r64: ROTQMBYIZExtInst<R64C, rotNeg7imm, uimm7>;
2822 defm ROTQMBYI : RotateQuadBytesImm;
2824 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2825 // Rotate right and mask by bit count
2826 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2828 class ROTQMBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2829 RRForm<0b10110011100, OOL, IOL, "rotqmbybi\t$rT, $rA, $rB",
2830 RotateShift, pattern>;
2832 class ROTQMBYBIVecInst<ValueType vectype>:
2833 ROTQMBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2834 [/* no pattern, */]>;
2836 multiclass RotateMaskQuadByBitCount
2838 def v16i8: ROTQMBYBIVecInst<v16i8>;
2839 def v8i16: ROTQMBYBIVecInst<v8i16>;
2840 def v4i32: ROTQMBYBIVecInst<v4i32>;
2841 def v2i64: ROTQMBYBIVecInst<v2i64>;
2844 defm ROTQMBYBI: RotateMaskQuadByBitCount;
2846 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2847 // Rotate quad and mask by bits
2848 // Note that the rotate amount has to be negated
2849 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2851 class ROTQMBIInst<dag OOL, dag IOL, list<dag> pattern>:
2852 RRForm<0b10011011100, OOL, IOL, "rotqmbi\t$rT, $rA, $rB",
2853 RotateShift, pattern>;
2855 class ROTQMBIVecInst<ValueType vectype>:
2856 ROTQMBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2857 [/* no pattern */]>;
2859 class ROTQMBIRegInst<RegisterClass rclass>:
2860 ROTQMBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2861 [/* no pattern */]>;
2863 multiclass RotateMaskQuadByBits
2865 def v16i8: ROTQMBIVecInst<v16i8>;
2866 def v8i16: ROTQMBIVecInst<v8i16>;
2867 def v4i32: ROTQMBIVecInst<v4i32>;
2868 def v2i64: ROTQMBIVecInst<v2i64>;
2870 def r128: ROTQMBIRegInst<GPRC>;
2871 def r64: ROTQMBIRegInst<R64C>;
2874 defm ROTQMBI: RotateMaskQuadByBits;
2876 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2877 // Rotate quad and mask by bits, immediate
2878 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2880 class ROTQMBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2881 RI7Form<0b10011111100, OOL, IOL, "rotqmbii\t$rT, $rA, $val",
2882 RotateShift, pattern>;
2884 class ROTQMBIIVecInst<ValueType vectype>:
2885 ROTQMBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2886 [/* no pattern */]>;
2888 class ROTQMBIIRegInst<RegisterClass rclass>:
2889 ROTQMBIIInst<(outs rclass:$rT), (ins rclass:$rA, rotNeg7imm:$val),
2890 [/* no pattern */]>;
2892 multiclass RotateMaskQuadByBitsImm
2894 def v16i8: ROTQMBIIVecInst<v16i8>;
2895 def v8i16: ROTQMBIIVecInst<v8i16>;
2896 def v4i32: ROTQMBIIVecInst<v4i32>;
2897 def v2i64: ROTQMBIIVecInst<v2i64>;
2899 def r128: ROTQMBIIRegInst<GPRC>;
2900 def r64: ROTQMBIIRegInst<R64C>;
2903 defm ROTQMBII: RotateMaskQuadByBitsImm;
2905 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2906 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2909 RRForm<0b01111010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2910 "rotmah\t$rT, $rA, $rB", RotateShift,
2911 [/* see patterns below - $rB must be negated */]>;
2913 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), R32C:$rB),
2914 (ROTMAHv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2916 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), R16C:$rB),
2917 (ROTMAHv8i16 VECREG:$rA,
2918 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2920 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), R8C:$rB),
2921 (ROTMAHv8i16 VECREG:$rA,
2922 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2925 RRForm<0b01111010000, (outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2926 "rotmah\t$rT, $rA, $rB", RotateShift,
2927 [/* see patterns below - $rB must be negated */]>;
2929 def : Pat<(sra R16C:$rA, R32C:$rB),
2930 (ROTMAHr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2932 def : Pat<(sra R16C:$rA, R16C:$rB),
2933 (ROTMAHr16 R16C:$rA,
2934 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2936 def : Pat<(sra R16C:$rA, R8C:$rB),
2937 (ROTMAHr16 R16C:$rA,
2938 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2941 RRForm<0b01111110000, (outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2942 "rotmahi\t$rT, $rA, $val", RotateShift,
2943 [(set (v8i16 VECREG:$rT),
2944 (SPUvec_sra (v8i16 VECREG:$rA), (i32 uimm7:$val)))]>;
2946 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i16 uimm7:$val)),
2947 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (TO_IMM32 uimm7:$val))>;
2949 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i8 uimm7:$val)),
2950 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (TO_IMM32 uimm7:$val))>;
2953 RRForm<0b01111110000, (outs R16C:$rT), (ins R16C:$rA, rothNeg7imm_i16:$val),
2954 "rotmahi\t$rT, $rA, $val", RotateShift,
2955 [(set R16C:$rT, (sra R16C:$rA, (i16 uimm7:$val)))]>;
2957 def : Pat<(sra R16C:$rA, (i32 imm:$val)),
2958 (ROTMAHIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
2960 def : Pat<(sra R16C:$rA, (i8 imm:$val)),
2961 (ROTMAHIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
2964 RRForm<0b01011010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2965 "rotma\t$rT, $rA, $rB", RotateShift,
2966 [/* see patterns below - $rB must be negated */]>;
2968 def : Pat<(SPUvec_sra (v4i32 VECREG:$rA), R32C:$rB),
2969 (ROTMAv4i32 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2971 def : Pat<(SPUvec_sra (v4i32 VECREG:$rA), R16C:$rB),
2972 (ROTMAv4i32 VECREG:$rA,
2973 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2975 def : Pat<(SPUvec_sra (v4i32 VECREG:$rA), R8C:$rB),
2976 (ROTMAv4i32 VECREG:$rA,
2977 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2980 RRForm<0b01011010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2981 "rotma\t$rT, $rA, $rB", RotateShift,
2982 [/* see patterns below - $rB must be negated */]>;
2984 def : Pat<(sra R32C:$rA, R32C:$rB),
2985 (ROTMAr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
2987 def : Pat<(sra R32C:$rA, R16C:$rB),
2989 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2991 def : Pat<(sra R32C:$rA, R8C:$rB),
2993 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2995 class ROTMAIInst<dag OOL, dag IOL, list<dag> pattern>:
2996 RRForm<0b01011110000, OOL, IOL,
2997 "rotmai\t$rT, $rA, $val",
2998 RotateShift, pattern>;
3000 class ROTMAIVecInst<ValueType vectype, Operand intop, ValueType inttype>:
3001 ROTMAIInst<(outs VECREG:$rT), (ins VECREG:$rA, intop:$val),
3002 [(set (vectype VECREG:$rT),
3003 (SPUvec_sra VECREG:$rA, (inttype uimm7:$val)))]>;
3005 class ROTMAIRegInst<RegisterClass rclass, Operand intop, ValueType inttype>:
3006 ROTMAIInst<(outs rclass:$rT), (ins rclass:$rA, intop:$val),
3007 [(set rclass:$rT, (sra rclass:$rA, (inttype uimm7:$val)))]>;
3009 multiclass RotateMaskAlgebraicImm {
3010 def v2i64_i32 : ROTMAIVecInst<v2i64, rotNeg7imm, i32>;
3011 def v4i32_i32 : ROTMAIVecInst<v4i32, rotNeg7imm, i32>;
3012 def r64_i32 : ROTMAIRegInst<R64C, rotNeg7imm, i32>;
3013 def r32_i32 : ROTMAIRegInst<R32C, rotNeg7imm, i32>;
3016 defm ROTMAI : RotateMaskAlgebraicImm;
3018 //===----------------------------------------------------------------------===//
3019 // Branch and conditionals:
3020 //===----------------------------------------------------------------------===//
3022 let isTerminator = 1, isBarrier = 1 in {
3023 // Halt If Equal (r32 preferred slot only, no vector form)
3025 RRForm_3<0b00011011110, (outs), (ins R32C:$rA, R32C:$rB),
3026 "heq\t$rA, $rB", BranchResolv,
3027 [/* no pattern to match */]>;
3030 RI10Form_2<0b11111110, (outs), (ins R32C:$rA, s10imm:$val),
3031 "heqi\t$rA, $val", BranchResolv,
3032 [/* no pattern to match */]>;
3034 // HGT/HGTI: These instructions use signed arithmetic for the comparison,
3035 // contrasting with HLGT/HLGTI, which use unsigned comparison:
3037 RRForm_3<0b00011010010, (outs), (ins R32C:$rA, R32C:$rB),
3038 "hgt\t$rA, $rB", BranchResolv,
3039 [/* no pattern to match */]>;
3042 RI10Form_2<0b11110010, (outs), (ins R32C:$rA, s10imm:$val),
3043 "hgti\t$rA, $val", BranchResolv,
3044 [/* no pattern to match */]>;
3047 RRForm_3<0b00011011010, (outs), (ins R32C:$rA, R32C:$rB),
3048 "hlgt\t$rA, $rB", BranchResolv,
3049 [/* no pattern to match */]>;
3052 RI10Form_2<0b11111010, (outs), (ins R32C:$rA, s10imm:$val),
3053 "hlgti\t$rA, $val", BranchResolv,
3054 [/* no pattern to match */]>;
3057 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3058 // Comparison operators for i8, i16 and i32:
3059 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3061 class CEQBInst<dag OOL, dag IOL, list<dag> pattern> :
3062 RRForm<0b00001011110, OOL, IOL, "ceqb\t$rT, $rA, $rB",
3065 multiclass CmpEqualByte
3068 CEQBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3069 [(set (v16i8 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
3070 (v8i16 VECREG:$rB)))]>;
3073 CEQBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3074 [(set R8C:$rT, (seteq R8C:$rA, R8C:$rB))]>;
3077 class CEQBIInst<dag OOL, dag IOL, list<dag> pattern> :
3078 RI10Form<0b01111110, OOL, IOL, "ceqbi\t$rT, $rA, $val",
3081 multiclass CmpEqualByteImm
3084 CEQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3085 [(set (v16i8 VECREG:$rT), (seteq (v16i8 VECREG:$rA),
3086 v16i8SExt8Imm:$val))]>;
3088 CEQBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3089 [(set R8C:$rT, (seteq R8C:$rA, immSExt8:$val))]>;
3092 class CEQHInst<dag OOL, dag IOL, list<dag> pattern> :
3093 RRForm<0b00010011110, OOL, IOL, "ceqh\t$rT, $rA, $rB",
3096 multiclass CmpEqualHalfword
3098 def v8i16 : CEQHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3099 [(set (v8i16 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
3100 (v8i16 VECREG:$rB)))]>;
3102 def r16 : CEQHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3103 [(set R16C:$rT, (seteq R16C:$rA, R16C:$rB))]>;
3106 class CEQHIInst<dag OOL, dag IOL, list<dag> pattern> :
3107 RI10Form<0b10111110, OOL, IOL, "ceqhi\t$rT, $rA, $val",
3110 multiclass CmpEqualHalfwordImm
3112 def v8i16 : CEQHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3113 [(set (v8i16 VECREG:$rT),
3114 (seteq (v8i16 VECREG:$rA),
3115 (v8i16 v8i16SExt10Imm:$val)))]>;
3116 def r16 : CEQHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3117 [(set R16C:$rT, (seteq R16C:$rA, i16ImmSExt10:$val))]>;
3120 class CEQInst<dag OOL, dag IOL, list<dag> pattern> :
3121 RRForm<0b00000011110, OOL, IOL, "ceq\t$rT, $rA, $rB",
3124 multiclass CmpEqualWord
3126 def v4i32 : CEQInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3127 [(set (v4i32 VECREG:$rT),
3128 (seteq (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3130 def r32 : CEQInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3131 [(set R32C:$rT, (seteq R32C:$rA, R32C:$rB))]>;
3134 class CEQIInst<dag OOL, dag IOL, list<dag> pattern> :
3135 RI10Form<0b00111110, OOL, IOL, "ceqi\t$rT, $rA, $val",
3138 multiclass CmpEqualWordImm
3140 def v4i32 : CEQIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3141 [(set (v4i32 VECREG:$rT),
3142 (seteq (v4i32 VECREG:$rA),
3143 (v4i32 v4i32SExt16Imm:$val)))]>;
3145 def r32: CEQIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3146 [(set R32C:$rT, (seteq R32C:$rA, i32ImmSExt10:$val))]>;
3149 class CGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3150 RRForm<0b00001010010, OOL, IOL, "cgtb\t$rT, $rA, $rB",
3153 multiclass CmpGtrByte
3156 CGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3157 [(set (v16i8 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3158 (v8i16 VECREG:$rB)))]>;
3161 CGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3162 [(set R8C:$rT, (setgt R8C:$rA, R8C:$rB))]>;
3165 class CGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3166 RI10Form<0b01110010, OOL, IOL, "cgtbi\t$rT, $rA, $val",
3169 multiclass CmpGtrByteImm
3172 CGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3173 [(set (v16i8 VECREG:$rT), (setgt (v16i8 VECREG:$rA),
3174 v16i8SExt8Imm:$val))]>;
3176 CGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3177 [(set R8C:$rT, (setgt R8C:$rA, immSExt8:$val))]>;
3180 class CGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3181 RRForm<0b00010010010, OOL, IOL, "cgth\t$rT, $rA, $rB",
3184 multiclass CmpGtrHalfword
3186 def v8i16 : CGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3187 [(set (v8i16 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3188 (v8i16 VECREG:$rB)))]>;
3190 def r16 : CGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3191 [(set R16C:$rT, (setgt R16C:$rA, R16C:$rB))]>;
3194 class CGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3195 RI10Form<0b10110010, OOL, IOL, "cgthi\t$rT, $rA, $val",
3198 multiclass CmpGtrHalfwordImm
3200 def v8i16 : CGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3201 [(set (v8i16 VECREG:$rT),
3202 (setgt (v8i16 VECREG:$rA),
3203 (v8i16 v8i16SExt10Imm:$val)))]>;
3204 def r16 : CGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3205 [(set R16C:$rT, (setgt R16C:$rA, i16ImmSExt10:$val))]>;
3208 class CGTInst<dag OOL, dag IOL, list<dag> pattern> :
3209 RRForm<0b00000010010, OOL, IOL, "cgt\t$rT, $rA, $rB",
3212 multiclass CmpGtrWord
3214 def v4i32 : CGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3215 [(set (v4i32 VECREG:$rT),
3216 (setgt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3218 def r32 : CGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3219 [(set R32C:$rT, (setgt R32C:$rA, R32C:$rB))]>;
3222 class CGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3223 RI10Form<0b00110010, OOL, IOL, "cgti\t$rT, $rA, $val",
3226 multiclass CmpGtrWordImm
3228 def v4i32 : CGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3229 [(set (v4i32 VECREG:$rT),
3230 (setgt (v4i32 VECREG:$rA),
3231 (v4i32 v4i32SExt16Imm:$val)))]>;
3233 def r32: CGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3234 [(set R32C:$rT, (setgt R32C:$rA, i32ImmSExt10:$val))]>;
3236 // CGTIv4f32, CGTIf32: These are used in the f32 fdiv instruction sequence:
3237 def v4f32: CGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3238 [(set (v4i32 VECREG:$rT),
3239 (setgt (v4i32 (bitconvert (v4f32 VECREG:$rA))),
3240 (v4i32 v4i32SExt16Imm:$val)))]>;
3242 def f32: CGTIInst<(outs R32C:$rT), (ins R32FP:$rA, s10imm_i32:$val),
3243 [/* no pattern */]>;
3246 class CLGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3247 RRForm<0b00001011010, OOL, IOL, "clgtb\t$rT, $rA, $rB",
3250 multiclass CmpLGtrByte
3253 CLGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3254 [(set (v16i8 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3255 (v8i16 VECREG:$rB)))]>;
3258 CLGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3259 [(set R8C:$rT, (setugt R8C:$rA, R8C:$rB))]>;
3262 class CLGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3263 RI10Form<0b01111010, OOL, IOL, "clgtbi\t$rT, $rA, $val",
3266 multiclass CmpLGtrByteImm
3269 CLGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3270 [(set (v16i8 VECREG:$rT), (setugt (v16i8 VECREG:$rA),
3271 v16i8SExt8Imm:$val))]>;
3273 CLGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3274 [(set R8C:$rT, (setugt R8C:$rA, immSExt8:$val))]>;
3277 class CLGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3278 RRForm<0b00010011010, OOL, IOL, "clgth\t$rT, $rA, $rB",
3281 multiclass CmpLGtrHalfword
3283 def v8i16 : CLGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3284 [(set (v8i16 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3285 (v8i16 VECREG:$rB)))]>;
3287 def r16 : CLGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3288 [(set R16C:$rT, (setugt R16C:$rA, R16C:$rB))]>;
3291 class CLGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3292 RI10Form<0b10111010, OOL, IOL, "clgthi\t$rT, $rA, $val",
3295 multiclass CmpLGtrHalfwordImm
3297 def v8i16 : CLGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3298 [(set (v8i16 VECREG:$rT),
3299 (setugt (v8i16 VECREG:$rA),
3300 (v8i16 v8i16SExt10Imm:$val)))]>;
3301 def r16 : CLGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3302 [(set R16C:$rT, (setugt R16C:$rA, i16ImmSExt10:$val))]>;
3305 class CLGTInst<dag OOL, dag IOL, list<dag> pattern> :
3306 RRForm<0b00000011010, OOL, IOL, "clgt\t$rT, $rA, $rB",
3309 multiclass CmpLGtrWord
3311 def v4i32 : CLGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3312 [(set (v4i32 VECREG:$rT),
3313 (setugt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3315 def r32 : CLGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3316 [(set R32C:$rT, (setugt R32C:$rA, R32C:$rB))]>;
3319 class CLGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3320 RI10Form<0b00111010, OOL, IOL, "clgti\t$rT, $rA, $val",
3323 multiclass CmpLGtrWordImm
3325 def v4i32 : CLGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3326 [(set (v4i32 VECREG:$rT),
3327 (setugt (v4i32 VECREG:$rA),
3328 (v4i32 v4i32SExt16Imm:$val)))]>;
3330 def r32: CLGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3331 [(set R32C:$rT, (setugt R32C:$rA, i32ImmSExt10:$val))]>;
3334 defm CEQB : CmpEqualByte;
3335 defm CEQBI : CmpEqualByteImm;
3336 defm CEQH : CmpEqualHalfword;
3337 defm CEQHI : CmpEqualHalfwordImm;
3338 defm CEQ : CmpEqualWord;
3339 defm CEQI : CmpEqualWordImm;
3340 defm CGTB : CmpGtrByte;
3341 defm CGTBI : CmpGtrByteImm;
3342 defm CGTH : CmpGtrHalfword;
3343 defm CGTHI : CmpGtrHalfwordImm;
3344 defm CGT : CmpGtrWord;
3345 defm CGTI : CmpGtrWordImm;
3346 defm CLGTB : CmpLGtrByte;
3347 defm CLGTBI : CmpLGtrByteImm;
3348 defm CLGTH : CmpLGtrHalfword;
3349 defm CLGTHI : CmpLGtrHalfwordImm;
3350 defm CLGT : CmpLGtrWord;
3351 defm CLGTI : CmpLGtrWordImm;
3353 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3354 // For SETCC primitives not supported above (setlt, setle, setge, etc.)
3355 // define a pattern to generate the right code, as a binary operator
3356 // (in a manner of speaking.)
3359 // 1. This only matches the setcc set of conditionals. Special pattern
3360 // matching is used for select conditionals.
3362 // 2. The "DAG" versions of these classes is almost exclusively used for
3363 // i64 comparisons. See the tblgen fundamentals documentation for what
3364 // ".ResultInstrs[0]" means; see TargetSelectionDAG.td and the Pattern
3365 // class for where ResultInstrs originates.
3366 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3368 class SETCCNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3369 SPUInstr xorinst, SPUInstr cmpare>:
3370 Pat<(cond rclass:$rA, rclass:$rB),
3371 (xorinst (cmpare rclass:$rA, rclass:$rB), (inttype -1))>;
3373 class SETCCNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3374 PatLeaf immpred, SPUInstr xorinst, SPUInstr cmpare>:
3375 Pat<(cond rclass:$rA, (inttype immpred:$imm)),
3376 (xorinst (cmpare rclass:$rA, (inttype immpred:$imm)), (inttype -1))>;
3378 def : SETCCNegCondReg<setne, R8C, i8, XORBIr8, CEQBr8>;
3379 def : SETCCNegCondImm<setne, R8C, i8, immSExt8, XORBIr8, CEQBIr8>;
3381 def : SETCCNegCondReg<setne, R16C, i16, XORHIr16, CEQHr16>;
3382 def : SETCCNegCondImm<setne, R16C, i16, i16ImmSExt10, XORHIr16, CEQHIr16>;
3384 def : SETCCNegCondReg<setne, R32C, i32, XORIr32, CEQr32>;
3385 def : SETCCNegCondImm<setne, R32C, i32, i32ImmSExt10, XORIr32, CEQIr32>;
3387 class SETCCBinOpReg<PatFrag cond, RegisterClass rclass,
3388 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3389 Pat<(cond rclass:$rA, rclass:$rB),
3390 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3391 (cmpOp2 rclass:$rA, rclass:$rB))>;
3393 class SETCCBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3395 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3396 Pat<(cond rclass:$rA, (immtype immpred:$imm)),
3397 (binop (cmpOp1 rclass:$rA, (immtype immpred:$imm)),
3398 (cmpOp2 rclass:$rA, (immtype immpred:$imm)))>;
3400 def : SETCCBinOpReg<setge, R8C, ORr8, CGTBr8, CEQBr8>;
3401 def : SETCCBinOpImm<setge, R8C, immSExt8, i8, ORr8, CGTBIr8, CEQBIr8>;
3402 def : SETCCBinOpReg<setlt, R8C, NORr8, CGTBr8, CEQBr8>;
3403 def : SETCCBinOpImm<setlt, R8C, immSExt8, i8, NORr8, CGTBIr8, CEQBIr8>;
3404 def : Pat<(setle R8C:$rA, R8C:$rB),
3405 (XORBIr8 (CGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3406 def : Pat<(setle R8C:$rA, immU8:$imm),
3407 (XORBIr8 (CGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3409 def : SETCCBinOpReg<setge, R16C, ORr16, CGTHr16, CEQHr16>;
3410 def : SETCCBinOpImm<setge, R16C, i16ImmSExt10, i16,
3411 ORr16, CGTHIr16, CEQHIr16>;
3412 def : SETCCBinOpReg<setlt, R16C, NORr16, CGTHr16, CEQHr16>;
3413 def : SETCCBinOpImm<setlt, R16C, i16ImmSExt10, i16, NORr16, CGTHIr16, CEQHIr16>;
3414 def : Pat<(setle R16C:$rA, R16C:$rB),
3415 (XORHIr16 (CGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3416 def : Pat<(setle R16C:$rA, i16ImmSExt10:$imm),
3417 (XORHIr16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3419 def : SETCCBinOpReg<setge, R32C, ORr32, CGTr32, CEQr32>;
3420 def : SETCCBinOpImm<setge, R32C, i32ImmSExt10, i32,
3421 ORr32, CGTIr32, CEQIr32>;
3422 def : SETCCBinOpReg<setlt, R32C, NORr32, CGTr32, CEQr32>;
3423 def : SETCCBinOpImm<setlt, R32C, i32ImmSExt10, i32, NORr32, CGTIr32, CEQIr32>;
3424 def : Pat<(setle R32C:$rA, R32C:$rB),
3425 (XORIr32 (CGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3426 def : Pat<(setle R32C:$rA, i32ImmSExt10:$imm),
3427 (XORIr32 (CGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3429 def : SETCCBinOpReg<setuge, R8C, ORr8, CLGTBr8, CEQBr8>;
3430 def : SETCCBinOpImm<setuge, R8C, immSExt8, i8, ORr8, CLGTBIr8, CEQBIr8>;
3431 def : SETCCBinOpReg<setult, R8C, NORr8, CLGTBr8, CEQBr8>;
3432 def : SETCCBinOpImm<setult, R8C, immSExt8, i8, NORr8, CLGTBIr8, CEQBIr8>;
3433 def : Pat<(setule R8C:$rA, R8C:$rB),
3434 (XORBIr8 (CLGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3435 def : Pat<(setule R8C:$rA, immU8:$imm),
3436 (XORBIr8 (CLGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3438 def : SETCCBinOpReg<setuge, R16C, ORr16, CLGTHr16, CEQHr16>;
3439 def : SETCCBinOpImm<setuge, R16C, i16ImmSExt10, i16,
3440 ORr16, CLGTHIr16, CEQHIr16>;
3441 def : SETCCBinOpReg<setult, R16C, NORr16, CLGTHr16, CEQHr16>;
3442 def : SETCCBinOpImm<setult, R16C, i16ImmSExt10, i16, NORr16,
3443 CLGTHIr16, CEQHIr16>;
3444 def : Pat<(setule R16C:$rA, R16C:$rB),
3445 (XORHIr16 (CLGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3446 def : Pat<(setule R16C:$rA, i16ImmSExt10:$imm),
3447 (XORHIr16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3449 def : SETCCBinOpReg<setuge, R32C, ORr32, CLGTr32, CEQr32>;
3450 def : SETCCBinOpImm<setuge, R32C, i32ImmSExt10, i32,
3451 ORr32, CLGTIr32, CEQIr32>;
3452 def : SETCCBinOpReg<setult, R32C, NORr32, CLGTr32, CEQr32>;
3453 def : SETCCBinOpImm<setult, R32C, i32ImmSExt10, i32, NORr32, CLGTIr32, CEQIr32>;
3454 def : Pat<(setule R32C:$rA, R32C:$rB),
3455 (XORIr32 (CLGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3456 def : Pat<(setule R32C:$rA, i32ImmSExt10:$imm),
3457 (XORIr32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3459 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3460 // select conditional patterns:
3461 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3463 class SELECTNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3464 SPUInstr selinstr, SPUInstr cmpare>:
3465 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3466 rclass:$rTrue, rclass:$rFalse),
3467 (selinstr rclass:$rTrue, rclass:$rFalse,
3468 (cmpare rclass:$rA, rclass:$rB))>;
3470 class SELECTNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3471 PatLeaf immpred, SPUInstr selinstr, SPUInstr cmpare>:
3472 Pat<(select (inttype (cond rclass:$rA, immpred:$imm)),
3473 rclass:$rTrue, rclass:$rFalse),
3474 (selinstr rclass:$rTrue, rclass:$rFalse,
3475 (cmpare rclass:$rA, immpred:$imm))>;
3477 def : SELECTNegCondReg<setne, R8C, i8, SELBr8, CEQBr8>;
3478 def : SELECTNegCondImm<setne, R8C, i8, immSExt8, SELBr8, CEQBIr8>;
3479 def : SELECTNegCondReg<setle, R8C, i8, SELBr8, CGTBr8>;
3480 def : SELECTNegCondImm<setle, R8C, i8, immSExt8, SELBr8, CGTBr8>;
3481 def : SELECTNegCondReg<setule, R8C, i8, SELBr8, CLGTBr8>;
3482 def : SELECTNegCondImm<setule, R8C, i8, immU8, SELBr8, CLGTBIr8>;
3484 def : SELECTNegCondReg<setne, R16C, i16, SELBr16, CEQHr16>;
3485 def : SELECTNegCondImm<setne, R16C, i16, i16ImmSExt10, SELBr16, CEQHIr16>;
3486 def : SELECTNegCondReg<setle, R16C, i16, SELBr16, CGTHr16>;
3487 def : SELECTNegCondImm<setle, R16C, i16, i16ImmSExt10, SELBr16, CGTHIr16>;
3488 def : SELECTNegCondReg<setule, R16C, i16, SELBr16, CLGTHr16>;
3489 def : SELECTNegCondImm<setule, R16C, i16, i16ImmSExt10, SELBr16, CLGTHIr16>;
3491 def : SELECTNegCondReg<setne, R32C, i32, SELBr32, CEQr32>;
3492 def : SELECTNegCondImm<setne, R32C, i32, i32ImmSExt10, SELBr32, CEQIr32>;
3493 def : SELECTNegCondReg<setle, R32C, i32, SELBr32, CGTr32>;
3494 def : SELECTNegCondImm<setle, R32C, i32, i32ImmSExt10, SELBr32, CGTIr32>;
3495 def : SELECTNegCondReg<setule, R32C, i32, SELBr32, CLGTr32>;
3496 def : SELECTNegCondImm<setule, R32C, i32, i32ImmSExt10, SELBr32, CLGTIr32>;
3498 class SELECTBinOpReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3499 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3501 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3502 rclass:$rTrue, rclass:$rFalse),
3503 (selinstr rclass:$rFalse, rclass:$rTrue,
3504 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3505 (cmpOp2 rclass:$rA, rclass:$rB)))>;
3507 class SELECTBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3509 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3511 Pat<(select (inttype (cond rclass:$rA, (inttype immpred:$imm))),
3512 rclass:$rTrue, rclass:$rFalse),
3513 (selinstr rclass:$rFalse, rclass:$rTrue,
3514 (binop (cmpOp1 rclass:$rA, (inttype immpred:$imm)),
3515 (cmpOp2 rclass:$rA, (inttype immpred:$imm))))>;
3517 def : SELECTBinOpReg<setge, R8C, i8, SELBr8, ORr8, CGTBr8, CEQBr8>;
3518 def : SELECTBinOpImm<setge, R8C, immSExt8, i8,
3519 SELBr8, ORr8, CGTBIr8, CEQBIr8>;
3521 def : SELECTBinOpReg<setge, R16C, i16, SELBr16, ORr16, CGTHr16, CEQHr16>;
3522 def : SELECTBinOpImm<setge, R16C, i16ImmSExt10, i16,
3523 SELBr16, ORr16, CGTHIr16, CEQHIr16>;
3525 def : SELECTBinOpReg<setge, R32C, i32, SELBr32, ORr32, CGTr32, CEQr32>;
3526 def : SELECTBinOpImm<setge, R32C, i32ImmSExt10, i32,
3527 SELBr32, ORr32, CGTIr32, CEQIr32>;
3529 def : SELECTBinOpReg<setuge, R8C, i8, SELBr8, ORr8, CLGTBr8, CEQBr8>;
3530 def : SELECTBinOpImm<setuge, R8C, immSExt8, i8,
3531 SELBr8, ORr8, CLGTBIr8, CEQBIr8>;
3533 def : SELECTBinOpReg<setuge, R16C, i16, SELBr16, ORr16, CLGTHr16, CEQHr16>;
3534 def : SELECTBinOpImm<setuge, R16C, i16ImmUns10, i16,
3535 SELBr16, ORr16, CLGTHIr16, CEQHIr16>;
3537 def : SELECTBinOpReg<setuge, R32C, i32, SELBr32, ORr32, CLGTr32, CEQr32>;
3538 def : SELECTBinOpImm<setuge, R32C, i32ImmUns10, i32,
3539 SELBr32, ORr32, CLGTIr32, CEQIr32>;
3541 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3544 // All calls clobber the non-callee-saved registers:
3545 Defs = [R0, R1, R2, R3, R4, R5, R6, R7, R8, R9,
3546 R10,R11,R12,R13,R14,R15,R16,R17,R18,R19,
3547 R20,R21,R22,R23,R24,R25,R26,R27,R28,R29,
3548 R30,R31,R32,R33,R34,R35,R36,R37,R38,R39,
3549 R40,R41,R42,R43,R44,R45,R46,R47,R48,R49,
3550 R50,R51,R52,R53,R54,R55,R56,R57,R58,R59,
3551 R60,R61,R62,R63,R64,R65,R66,R67,R68,R69,
3552 R70,R71,R72,R73,R74,R75,R76,R77,R78,R79],
3553 // All of these instructions use $lr (aka $0)
3555 // Branch relative and set link: Used if we actually know that the target
3556 // is within [-32768, 32767] bytes of the target
3558 BranchSetLink<0b011001100, (outs), (ins relcalltarget:$func, variable_ops),
3559 "brsl\t$$lr, $func",
3560 [(SPUcall (SPUpcrel tglobaladdr:$func, 0))]>;
3562 // Branch absolute and set link: Used if we actually know that the target
3563 // is an absolute address
3565 BranchSetLink<0b011001100, (outs), (ins calltarget:$func, variable_ops),
3566 "brasl\t$$lr, $func",
3567 [(SPUcall (SPUaform tglobaladdr:$func, 0))]>;
3569 // Branch indirect and set link if external data. These instructions are not
3570 // actually generated, matched by an intrinsic:
3571 def BISLED_00: BISLEDForm<0b11, "bisled\t$$lr, $func", [/* empty pattern */]>;
3572 def BISLED_E0: BISLEDForm<0b10, "bisled\t$$lr, $func", [/* empty pattern */]>;
3573 def BISLED_0D: BISLEDForm<0b01, "bisled\t$$lr, $func", [/* empty pattern */]>;
3574 def BISLED_ED: BISLEDForm<0b00, "bisled\t$$lr, $func", [/* empty pattern */]>;
3576 // Branch indirect and set link. This is the "X-form" address version of a
3579 BIForm<0b10010101100, "bisl\t$$lr, $func", [(SPUcall R32C:$func)]>;
3582 // Support calls to external symbols:
3583 def : Pat<(SPUcall (SPUpcrel texternalsym:$func, 0)),
3584 (BRSL texternalsym:$func)>;
3586 def : Pat<(SPUcall (SPUaform texternalsym:$func, 0)),
3587 (BRASL texternalsym:$func)>;
3589 // Unconditional branches:
3590 let isBranch = 1, isTerminator = 1, hasCtrlDep = 1 in {
3591 let isBarrier = 1 in {
3593 UncondBranch<0b001001100, (outs), (ins brtarget:$dest),
3597 // Unconditional, absolute address branch
3599 UncondBranch<0b001100000, (outs), (ins brtarget:$dest),
3601 [/* no pattern */]>;
3605 BIForm<0b00010101100, "bi\t$func", [(brind R32C:$func)]>;
3608 // Conditional branches:
3609 class BRNZInst<dag IOL, list<dag> pattern>:
3610 RI16Form<0b010000100, (outs), IOL, "brnz\t$rCond,$dest",
3611 BranchResolv, pattern>;
3613 class BRNZRegInst<RegisterClass rclass>:
3614 BRNZInst<(ins rclass:$rCond, brtarget:$dest),
3615 [(brcond rclass:$rCond, bb:$dest)]>;
3617 class BRNZVecInst<ValueType vectype>:
3618 BRNZInst<(ins VECREG:$rCond, brtarget:$dest),
3619 [(brcond (vectype VECREG:$rCond), bb:$dest)]>;
3621 multiclass BranchNotZero {
3622 def v4i32 : BRNZVecInst<v4i32>;
3623 def r32 : BRNZRegInst<R32C>;
3626 defm BRNZ : BranchNotZero;
3628 class BRZInst<dag IOL, list<dag> pattern>:
3629 RI16Form<0b000000100, (outs), IOL, "brz\t$rT,$dest",
3630 BranchResolv, pattern>;
3632 class BRZRegInst<RegisterClass rclass>:
3633 BRZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3635 class BRZVecInst<ValueType vectype>:
3636 BRZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3638 multiclass BranchZero {
3639 def v4i32: BRZVecInst<v4i32>;
3640 def r32: BRZRegInst<R32C>;
3643 defm BRZ: BranchZero;
3645 // Note: LLVM doesn't do branch conditional, indirect. Otherwise these would
3648 class BINZInst<dag IOL, list<dag> pattern>:
3649 BICondForm<0b10010100100, (outs), IOL, "binz\t$rA, $dest", pattern>;
3651 class BINZRegInst<RegisterClass rclass>:
3652 BINZInst<(ins rclass:$rA, brtarget:$dest),
3653 [(brcond rclass:$rA, R32C:$dest)]>;
3655 class BINZVecInst<ValueType vectype>:
3656 BINZInst<(ins VECREG:$rA, R32C:$dest),
3657 [(brcond (vectype VECREG:$rA), R32C:$dest)]>;
3659 multiclass BranchNotZeroIndirect {
3660 def v4i32: BINZVecInst<v4i32>;
3661 def r32: BINZRegInst<R32C>;
3664 defm BINZ: BranchNotZeroIndirect;
3666 class BIZInst<dag IOL, list<dag> pattern>:
3667 BICondForm<0b00010100100, (outs), IOL, "biz\t$rA, $func", pattern>;
3669 class BIZRegInst<RegisterClass rclass>:
3670 BIZInst<(ins rclass:$rA, R32C:$func), [/* no pattern */]>;
3672 class BIZVecInst<ValueType vectype>:
3673 BIZInst<(ins VECREG:$rA, R32C:$func), [/* no pattern */]>;
3675 multiclass BranchZeroIndirect {
3676 def v4i32: BIZVecInst<v4i32>;
3677 def r32: BIZRegInst<R32C>;
3680 defm BIZ: BranchZeroIndirect;
3683 class BRHNZInst<dag IOL, list<dag> pattern>:
3684 RI16Form<0b011000100, (outs), IOL, "brhnz\t$rCond,$dest", BranchResolv,
3687 class BRHNZRegInst<RegisterClass rclass>:
3688 BRHNZInst<(ins rclass:$rCond, brtarget:$dest),
3689 [(brcond rclass:$rCond, bb:$dest)]>;
3691 class BRHNZVecInst<ValueType vectype>:
3692 BRHNZInst<(ins VECREG:$rCond, brtarget:$dest), [/* no pattern */]>;
3694 multiclass BranchNotZeroHalfword {
3695 def v8i16: BRHNZVecInst<v8i16>;
3696 def r16: BRHNZRegInst<R16C>;
3699 defm BRHNZ: BranchNotZeroHalfword;
3701 class BRHZInst<dag IOL, list<dag> pattern>:
3702 RI16Form<0b001000100, (outs), IOL, "brhz\t$rT,$dest", BranchResolv,
3705 class BRHZRegInst<RegisterClass rclass>:
3706 BRHZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3708 class BRHZVecInst<ValueType vectype>:
3709 BRHZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3711 multiclass BranchZeroHalfword {
3712 def v8i16: BRHZVecInst<v8i16>;
3713 def r16: BRHZRegInst<R16C>;
3716 defm BRHZ: BranchZeroHalfword;
3719 //===----------------------------------------------------------------------===//
3720 // setcc and brcond patterns:
3721 //===----------------------------------------------------------------------===//
3723 def : Pat<(brcond (i16 (seteq R16C:$rA, 0)), bb:$dest),
3724 (BRHZr16 R16C:$rA, bb:$dest)>;
3725 def : Pat<(brcond (i16 (setne R16C:$rA, 0)), bb:$dest),
3726 (BRHNZr16 R16C:$rA, bb:$dest)>;
3728 def : Pat<(brcond (i32 (seteq R32C:$rA, 0)), bb:$dest),
3729 (BRZr32 R32C:$rA, bb:$dest)>;
3730 def : Pat<(brcond (i32 (setne R32C:$rA, 0)), bb:$dest),
3731 (BRNZr32 R32C:$rA, bb:$dest)>;
3733 multiclass BranchCondEQ<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3735 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3736 (brinst16 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3738 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3739 (brinst16 (CEQHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3741 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3742 (brinst32 (CEQIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3744 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3745 (brinst32 (CEQr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3748 defm BRCONDeq : BranchCondEQ<seteq, BRHNZr16, BRNZr32>;
3749 defm BRCONDne : BranchCondEQ<setne, BRHZr16, BRZr32>;
3751 multiclass BranchCondLGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3753 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3754 (brinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3756 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3757 (brinst16 (CLGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3759 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3760 (brinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3762 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3763 (brinst32 (CLGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3766 defm BRCONDugt : BranchCondLGT<setugt, BRHNZr16, BRNZr32>;
3767 defm BRCONDule : BranchCondLGT<setule, BRHZr16, BRZr32>;
3769 multiclass BranchCondLGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3770 SPUInstr orinst32, SPUInstr brinst32>
3772 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3773 (brinst16 (orinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3774 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3777 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3778 (brinst16 (orinst16 (CLGTHr16 R16C:$rA, R16:$rB),
3779 (CEQHr16 R16C:$rA, R16:$rB)),
3782 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3783 (brinst32 (orinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val),
3784 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3787 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3788 (brinst32 (orinst32 (CLGTr32 R32C:$rA, R32C:$rB),
3789 (CEQr32 R32C:$rA, R32C:$rB)),
3793 defm BRCONDuge : BranchCondLGTEQ<setuge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3794 defm BRCONDult : BranchCondLGTEQ<setult, ORr16, BRHZr16, ORr32, BRZr32>;
3796 multiclass BranchCondGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3798 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3799 (brinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3801 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3802 (brinst16 (CGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3804 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3805 (brinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3807 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3808 (brinst32 (CGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3811 defm BRCONDgt : BranchCondGT<setgt, BRHNZr16, BRNZr32>;
3812 defm BRCONDle : BranchCondGT<setle, BRHZr16, BRZr32>;
3814 multiclass BranchCondGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3815 SPUInstr orinst32, SPUInstr brinst32>
3817 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3818 (brinst16 (orinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3819 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3822 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3823 (brinst16 (orinst16 (CGTHr16 R16C:$rA, R16:$rB),
3824 (CEQHr16 R16C:$rA, R16:$rB)),
3827 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3828 (brinst32 (orinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val),
3829 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3832 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3833 (brinst32 (orinst32 (CGTr32 R32C:$rA, R32C:$rB),
3834 (CEQr32 R32C:$rA, R32C:$rB)),
3838 defm BRCONDge : BranchCondGTEQ<setge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3839 defm BRCONDlt : BranchCondGTEQ<setlt, ORr16, BRHZr16, ORr32, BRZr32>;
3841 let isTerminator = 1, isBarrier = 1 in {
3842 let isReturn = 1 in {
3844 RETForm<"bi\t$$lr", [(retflag)]>;
3848 //===----------------------------------------------------------------------===//
3849 // Single precision floating point instructions
3850 //===----------------------------------------------------------------------===//
3852 class FAInst<dag OOL, dag IOL, list<dag> pattern>:
3853 RRForm<0b01011000100, OOL, IOL, "fa\t$rT, $rA, $rB",
3856 class FAVecInst<ValueType vectype>:
3857 FAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3858 [(set (vectype VECREG:$rT),
3859 (fadd (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3863 def v4f32: FAVecInst<v4f32>;
3864 def f32: FAInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3865 [(set R32FP:$rT, (fadd R32FP:$rA, R32FP:$rB))]>;
3870 class FSInst<dag OOL, dag IOL, list<dag> pattern>:
3871 RRForm<0b01011000100, OOL, IOL, "fs\t$rT, $rA, $rB",
3874 class FSVecInst<ValueType vectype>:
3875 FSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3876 [(set (vectype VECREG:$rT),
3877 (fsub (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3881 def v4f32: FSVecInst<v4f32>;
3882 def f32: FSInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3883 [(set R32FP:$rT, (fsub R32FP:$rA, R32FP:$rB))]>;
3888 class FMInst<dag OOL, dag IOL, list<dag> pattern>:
3889 RRForm<0b01100011010, OOL, IOL,
3890 "fm\t$rT, $rA, $rB", SPrecFP,
3893 class FMVecInst<ValueType type>:
3894 FMInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3895 [(set (type VECREG:$rT),
3896 (fmul (type VECREG:$rA), (type VECREG:$rB)))]>;
3900 def v4f32: FMVecInst<v4f32>;
3901 def f32: FMInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3902 [(set R32FP:$rT, (fmul R32FP:$rA, R32FP:$rB))]>;
3907 // Floating point multiply and add
3908 // e.g. d = c + (a * b)
3910 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
3911 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
3912 [(set (v4f32 VECREG:$rT),
3913 (fadd (v4f32 VECREG:$rC),
3914 (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB))))]>;
3917 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
3918 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
3919 [(set R32FP:$rT, (fadd R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
3921 // FP multiply and subtract
3922 // Subtracts value in rC from product
3925 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
3926 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
3927 [(set (v4f32 VECREG:$rT),
3928 (fsub (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB)),
3929 (v4f32 VECREG:$rC)))]>;
3932 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
3933 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
3935 (fsub (fmul R32FP:$rA, R32FP:$rB), R32FP:$rC))]>;
3937 // Floating Negative Mulitply and Subtract
3938 // Subtracts product from value in rC
3939 // res = fneg(fms a b c)
3942 // NOTE: subtraction order
3946 RRRForm<0b1101, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
3947 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
3948 [(set R32FP:$rT, (fsub R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
3951 RRRForm<0b1101, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
3952 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
3953 [(set (v4f32 VECREG:$rT),
3954 (fsub (v4f32 VECREG:$rC),
3955 (fmul (v4f32 VECREG:$rA),
3956 (v4f32 VECREG:$rB))))]>;
3961 // Floating point reciprocal estimate
3963 class FRESTInst<dag OOL, dag IOL>:
3964 RRForm_1<0b00110111000, OOL, IOL,
3965 "frest\t$rT, $rA", SPrecFP,
3966 [/* no pattern */]>;
3969 FRESTInst<(outs VECREG:$rT), (ins VECREG:$rA)>;
3972 FRESTInst<(outs R32FP:$rT), (ins R32FP:$rA)>;
3974 // Floating point interpolate (used in conjunction with reciprocal estimate)
3976 RRForm<0b00101011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3977 "fi\t$rT, $rA, $rB", SPrecFP,
3978 [/* no pattern */]>;
3981 RRForm<0b00101011110, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3982 "fi\t$rT, $rA, $rB", SPrecFP,
3983 [/* no pattern */]>;
3985 //--------------------------------------------------------------------------
3986 // Basic single precision floating point comparisons:
3988 // Note: There is no support on SPU for single precision NaN. Consequently,
3989 // ordered and unordered comparisons are the same.
3990 //--------------------------------------------------------------------------
3993 RRForm<0b01000011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3994 "fceq\t$rT, $rA, $rB", SPrecFP,
3995 [(set R32C:$rT, (setueq R32FP:$rA, R32FP:$rB))]>;
3997 def : Pat<(setoeq R32FP:$rA, R32FP:$rB),
3998 (FCEQf32 R32FP:$rA, R32FP:$rB)>;
4001 RRForm<0b01010011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
4002 "fcmeq\t$rT, $rA, $rB", SPrecFP,
4003 [(set R32C:$rT, (setueq (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
4005 def : Pat<(setoeq (fabs R32FP:$rA), (fabs R32FP:$rB)),
4006 (FCMEQf32 R32FP:$rA, R32FP:$rB)>;
4009 RRForm<0b01000011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
4010 "fcgt\t$rT, $rA, $rB", SPrecFP,
4011 [(set R32C:$rT, (setugt R32FP:$rA, R32FP:$rB))]>;
4013 def : Pat<(setugt R32FP:$rA, R32FP:$rB),
4014 (FCGTf32 R32FP:$rA, R32FP:$rB)>;
4017 RRForm<0b01010011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
4018 "fcmgt\t$rT, $rA, $rB", SPrecFP,
4019 [(set R32C:$rT, (setugt (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
4021 def : Pat<(setugt (fabs R32FP:$rA), (fabs R32FP:$rB)),
4022 (FCMGTf32 R32FP:$rA, R32FP:$rB)>;
4024 //--------------------------------------------------------------------------
4025 // Single precision floating point comparisons and SETCC equivalents:
4026 //--------------------------------------------------------------------------
4028 def : SETCCNegCondReg<setune, R32FP, i32, XORIr32, FCEQf32>;
4029 def : SETCCNegCondReg<setone, R32FP, i32, XORIr32, FCEQf32>;
4031 def : SETCCBinOpReg<setuge, R32FP, ORr32, FCGTf32, FCEQf32>;
4032 def : SETCCBinOpReg<setoge, R32FP, ORr32, FCGTf32, FCEQf32>;
4034 def : SETCCBinOpReg<setult, R32FP, NORr32, FCGTf32, FCEQf32>;
4035 def : SETCCBinOpReg<setolt, R32FP, NORr32, FCGTf32, FCEQf32>;
4037 def : Pat<(setule R32FP:$rA, R32FP:$rB),
4038 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
4039 def : Pat<(setole R32FP:$rA, R32FP:$rB),
4040 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
4042 // FP Status and Control Register Write
4043 // Why isn't rT a don't care in the ISA?
4044 // Should we create a special RRForm_3 for this guy and zero out the rT?
4046 RRForm_1<0b01011101110, (outs R32FP:$rT), (ins R32FP:$rA),
4047 "fscrwr\t$rA", SPrecFP,
4048 [/* This instruction requires an intrinsic. Note: rT is unused. */]>;
4050 // FP Status and Control Register Read
4052 RRForm_2<0b01011101110, (outs R32FP:$rT), (ins),
4053 "fscrrd\t$rT", SPrecFP,
4054 [/* This instruction requires an intrinsic */]>;
4056 // llvm instruction space
4057 // How do these map onto cell instructions?
4059 // frest rC rB # c = 1/b (both lines)
4061 // fm rD rA rC # d = a * 1/b
4062 // fnms rB rD rB rA # b = - (d * b - a) --should == 0 in a perfect world
4063 // fma rB rB rC rD # b = b * c + d
4064 // = -(d *b -a) * c + d
4065 // = a * c - c ( a *b *c - a)
4070 // These llvm instructions will actually map to library calls.
4071 // All that's needed, then, is to check that the appropriate library is
4072 // imported and do a brsl to the proper function name.
4073 // frem # fmod(x, y): x - (x/y) * y
4074 // (Note: fmod(double, double), fmodf(float,float)
4078 // Unimplemented SPU instruction space
4079 // floating reciprocal absolute square root estimate (frsqest)
4081 // The following are probably just intrinsics
4082 // status and control register write
4083 // status and control register read
4085 //--------------------------------------
4086 // Floating Point Conversions
4087 // Signed conversions:
4089 CVTIntFPForm<0b0101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4090 "csflt\t$rT, $rA, 0", SPrecFP,
4091 [(set (v4f32 VECREG:$rT), (sint_to_fp (v4i32 VECREG:$rA)))]>;
4093 // Convert signed integer to floating point
4095 CVTIntFPForm<0b0101101110, (outs R32FP:$rT), (ins R32C:$rA),
4096 "csflt\t$rT, $rA, 0", SPrecFP,
4097 [(set R32FP:$rT, (sint_to_fp R32C:$rA))]>;
4099 // Convert unsigned into to float
4101 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4102 "cuflt\t$rT, $rA, 0", SPrecFP,
4103 [(set (v4f32 VECREG:$rT), (uint_to_fp (v4i32 VECREG:$rA)))]>;
4106 CVTIntFPForm<0b1101101110, (outs R32FP:$rT), (ins R32C:$rA),
4107 "cuflt\t$rT, $rA, 0", SPrecFP,
4108 [(set R32FP:$rT, (uint_to_fp R32C:$rA))]>;
4110 // Convert float to unsigned int
4111 // Assume that scale = 0
4114 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4115 "cfltu\t$rT, $rA, 0", SPrecFP,
4116 [(set (v4i32 VECREG:$rT), (fp_to_uint (v4f32 VECREG:$rA)))]>;
4119 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
4120 "cfltu\t$rT, $rA, 0", SPrecFP,
4121 [(set R32C:$rT, (fp_to_uint R32FP:$rA))]>;
4123 // Convert float to signed int
4124 // Assume that scale = 0
4127 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4128 "cflts\t$rT, $rA, 0", SPrecFP,
4129 [(set (v4i32 VECREG:$rT), (fp_to_sint (v4f32 VECREG:$rA)))]>;
4132 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
4133 "cflts\t$rT, $rA, 0", SPrecFP,
4134 [(set R32C:$rT, (fp_to_sint R32FP:$rA))]>;
4136 //===----------------------------------------------------------------------==//
4137 // Single<->Double precision conversions
4138 //===----------------------------------------------------------------------==//
4140 // NOTE: We use "vec" name suffix here to avoid confusion (e.g. input is a
4141 // v4f32, output is v2f64--which goes in the name?)
4143 // Floating point extend single to double
4144 // NOTE: Not sure if passing in v4f32 to FESDvec is correct since it
4145 // operates on two double-word slots (i.e. 1st and 3rd fp numbers
4148 RRForm_1<0b00011101110, (outs VECREG:$rT), (ins VECREG:$rA),
4149 "fesd\t$rT, $rA", SPrecFP,
4150 [/*(set (v2f64 VECREG:$rT), (fextend (v4f32 VECREG:$rA)))*/]>;
4153 RRForm_1<0b00011101110, (outs R64FP:$rT), (ins R32FP:$rA),
4154 "fesd\t$rT, $rA", SPrecFP,
4155 [(set R64FP:$rT, (fextend R32FP:$rA))]>;
4157 // Floating point round double to single
4159 // RRForm_1<0b10011101110, (outs VECREG:$rT), (ins VECREG:$rA),
4160 // "frds\t$rT, $rA,", SPrecFP,
4161 // [(set (v4f32 R32FP:$rT), (fround (v2f64 R64FP:$rA)))]>;
4164 RRForm_1<0b10011101110, (outs R32FP:$rT), (ins R64FP:$rA),
4165 "frds\t$rT, $rA", SPrecFP,
4166 [(set R32FP:$rT, (fround R64FP:$rA))]>;
4168 //ToDo include anyextend?
4170 //===----------------------------------------------------------------------==//
4171 // Double precision floating point instructions
4172 //===----------------------------------------------------------------------==//
4174 RRForm<0b00110011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4175 "dfa\t$rT, $rA, $rB", DPrecFP,
4176 [(set R64FP:$rT, (fadd R64FP:$rA, R64FP:$rB))]>;
4179 RRForm<0b00110011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4180 "dfa\t$rT, $rA, $rB", DPrecFP,
4181 [(set (v2f64 VECREG:$rT), (fadd (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4184 RRForm<0b10100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4185 "dfs\t$rT, $rA, $rB", DPrecFP,
4186 [(set R64FP:$rT, (fsub R64FP:$rA, R64FP:$rB))]>;
4189 RRForm<0b10100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4190 "dfs\t$rT, $rA, $rB", DPrecFP,
4191 [(set (v2f64 VECREG:$rT),
4192 (fsub (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4195 RRForm<0b01100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4196 "dfm\t$rT, $rA, $rB", DPrecFP,
4197 [(set R64FP:$rT, (fmul R64FP:$rA, R64FP:$rB))]>;
4200 RRForm<0b00100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4201 "dfm\t$rT, $rA, $rB", DPrecFP,
4202 [(set (v2f64 VECREG:$rT),
4203 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4206 RRForm<0b00111010110, (outs R64FP:$rT),
4207 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4208 "dfma\t$rT, $rA, $rB", DPrecFP,
4209 [(set R64FP:$rT, (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB)))]>,
4210 RegConstraint<"$rC = $rT">,
4214 RRForm<0b00111010110, (outs VECREG:$rT),
4215 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4216 "dfma\t$rT, $rA, $rB", DPrecFP,
4217 [(set (v2f64 VECREG:$rT),
4218 (fadd (v2f64 VECREG:$rC),
4219 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB))))]>,
4220 RegConstraint<"$rC = $rT">,
4224 RRForm<0b10111010110, (outs R64FP:$rT),
4225 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4226 "dfms\t$rT, $rA, $rB", DPrecFP,
4227 [(set R64FP:$rT, (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC))]>,
4228 RegConstraint<"$rC = $rT">,
4232 RRForm<0b10111010110, (outs VECREG:$rT),
4233 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4234 "dfms\t$rT, $rA, $rB", DPrecFP,
4235 [(set (v2f64 VECREG:$rT),
4236 (fsub (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)),
4237 (v2f64 VECREG:$rC)))]>;
4239 // DFNMS: - (a * b - c)
4240 // - (a * b) + c => c - (a * b)
4242 class DFNMSInst<dag OOL, dag IOL, list<dag> pattern>:
4243 RRForm<0b01111010110, OOL, IOL, "dfnms\t$rT, $rA, $rB",
4245 RegConstraint<"$rC = $rT">,
4248 class DFNMSVecInst<list<dag> pattern>:
4249 DFNMSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4252 class DFNMSRegInst<list<dag> pattern>:
4253 DFNMSInst<(outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4256 multiclass DFMultiplySubtract
4258 def v2f64 : DFNMSVecInst<[(set (v2f64 VECREG:$rT),
4259 (fsub (v2f64 VECREG:$rC),
4260 (fmul (v2f64 VECREG:$rA),
4261 (v2f64 VECREG:$rB))))]>;
4263 def f64 : DFNMSRegInst<[(set R64FP:$rT,
4265 (fmul R64FP:$rA, R64FP:$rB)))]>;
4268 defm DFNMS : DFMultiplySubtract;
4273 RRForm<0b11111010110, (outs R64FP:$rT),
4274 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4275 "dfnma\t$rT, $rA, $rB", DPrecFP,
4276 [(set R64FP:$rT, (fneg (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB))))]>,
4277 RegConstraint<"$rC = $rT">,
4281 RRForm<0b11111010110, (outs VECREG:$rT),
4282 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4283 "dfnma\t$rT, $rA, $rB", DPrecFP,
4284 [(set (v2f64 VECREG:$rT),
4285 (fneg (fadd (v2f64 VECREG:$rC),
4286 (fmul (v2f64 VECREG:$rA),
4287 (v2f64 VECREG:$rB)))))]>,
4288 RegConstraint<"$rC = $rT">,
4291 //===----------------------------------------------------------------------==//
4292 // Floating point negation and absolute value
4293 //===----------------------------------------------------------------------==//
4295 def : Pat<(fneg (v4f32 VECREG:$rA)),
4296 (XORfnegvec (v4f32 VECREG:$rA),
4297 (v4f32 (ILHUv4i32 0x8000)))>;
4299 def : Pat<(fneg R32FP:$rA),
4300 (XORfneg32 R32FP:$rA, (ILHUr32 0x8000))>;
4302 // Floating point absolute value
4303 // Note: f64 fabs is custom-selected.
4305 def : Pat<(fabs R32FP:$rA),
4306 (ANDfabs32 R32FP:$rA, (IOHLr32 (ILHUr32 0x7fff), 0xffff))>;
4308 def : Pat<(fabs (v4f32 VECREG:$rA)),
4309 (ANDfabsvec (v4f32 VECREG:$rA),
4310 (IOHLv4i32 (ILHUv4i32 0x7fff), 0xffff))>;
4312 //===----------------------------------------------------------------------===//
4313 // Hint for branch instructions:
4314 //===----------------------------------------------------------------------===//
4316 /* def HBR : SPUInstr<(outs), (ins), "hbr\t" */
4318 //===----------------------------------------------------------------------===//
4319 // Execution, Load NOP (execute NOPs belong in even pipeline, load NOPs belong
4320 // in the odd pipeline)
4321 //===----------------------------------------------------------------------===//
4323 def ENOP : SPUInstr<(outs), (ins), "enop", ExecNOP> {
4326 let Inst{0-10} = 0b10000000010;
4327 let Inst{11-17} = 0;
4328 let Inst{18-24} = 0;
4329 let Inst{25-31} = 0;
4332 def LNOP : SPUInstr<(outs), (ins), "lnop", LoadNOP> {
4335 let Inst{0-10} = 0b10000000000;
4336 let Inst{11-17} = 0;
4337 let Inst{18-24} = 0;
4338 let Inst{25-31} = 0;
4341 //===----------------------------------------------------------------------===//
4342 // Bit conversions (type conversions between vector/packed types)
4343 // NOTE: Promotions are handled using the XS* instructions.
4344 //===----------------------------------------------------------------------===//
4345 def : Pat<(v16i8 (bitconvert (v8i16 VECREG:$src))), (v16i8 VECREG:$src)>;
4346 def : Pat<(v16i8 (bitconvert (v4i32 VECREG:$src))), (v16i8 VECREG:$src)>;
4347 def : Pat<(v16i8 (bitconvert (v2i64 VECREG:$src))), (v16i8 VECREG:$src)>;
4348 def : Pat<(v16i8 (bitconvert (v4f32 VECREG:$src))), (v16i8 VECREG:$src)>;
4349 def : Pat<(v16i8 (bitconvert (v2f64 VECREG:$src))), (v16i8 VECREG:$src)>;
4351 def : Pat<(v8i16 (bitconvert (v16i8 VECREG:$src))), (v8i16 VECREG:$src)>;
4352 def : Pat<(v8i16 (bitconvert (v4i32 VECREG:$src))), (v8i16 VECREG:$src)>;
4353 def : Pat<(v8i16 (bitconvert (v2i64 VECREG:$src))), (v8i16 VECREG:$src)>;
4354 def : Pat<(v8i16 (bitconvert (v4f32 VECREG:$src))), (v8i16 VECREG:$src)>;
4355 def : Pat<(v8i16 (bitconvert (v2f64 VECREG:$src))), (v8i16 VECREG:$src)>;
4357 def : Pat<(v4i32 (bitconvert (v16i8 VECREG:$src))), (v4i32 VECREG:$src)>;
4358 def : Pat<(v4i32 (bitconvert (v8i16 VECREG:$src))), (v4i32 VECREG:$src)>;
4359 def : Pat<(v4i32 (bitconvert (v2i64 VECREG:$src))), (v4i32 VECREG:$src)>;
4360 def : Pat<(v4i32 (bitconvert (v4f32 VECREG:$src))), (v4i32 VECREG:$src)>;
4361 def : Pat<(v4i32 (bitconvert (v2f64 VECREG:$src))), (v4i32 VECREG:$src)>;
4363 def : Pat<(v2i64 (bitconvert (v16i8 VECREG:$src))), (v2i64 VECREG:$src)>;
4364 def : Pat<(v2i64 (bitconvert (v8i16 VECREG:$src))), (v2i64 VECREG:$src)>;
4365 def : Pat<(v2i64 (bitconvert (v4i32 VECREG:$src))), (v2i64 VECREG:$src)>;
4366 def : Pat<(v2i64 (bitconvert (v4f32 VECREG:$src))), (v2i64 VECREG:$src)>;
4367 def : Pat<(v2i64 (bitconvert (v2f64 VECREG:$src))), (v2i64 VECREG:$src)>;
4369 def : Pat<(v4f32 (bitconvert (v16i8 VECREG:$src))), (v4f32 VECREG:$src)>;
4370 def : Pat<(v4f32 (bitconvert (v8i16 VECREG:$src))), (v4f32 VECREG:$src)>;
4371 def : Pat<(v4f32 (bitconvert (v2i64 VECREG:$src))), (v4f32 VECREG:$src)>;
4372 def : Pat<(v4f32 (bitconvert (v4i32 VECREG:$src))), (v4f32 VECREG:$src)>;
4373 def : Pat<(v4f32 (bitconvert (v2f64 VECREG:$src))), (v4f32 VECREG:$src)>;
4375 def : Pat<(v2f64 (bitconvert (v16i8 VECREG:$src))), (v2f64 VECREG:$src)>;
4376 def : Pat<(v2f64 (bitconvert (v8i16 VECREG:$src))), (v2f64 VECREG:$src)>;
4377 def : Pat<(v2f64 (bitconvert (v4i32 VECREG:$src))), (v2f64 VECREG:$src)>;
4378 def : Pat<(v2f64 (bitconvert (v2i64 VECREG:$src))), (v2f64 VECREG:$src)>;
4379 def : Pat<(v2f64 (bitconvert (v4f32 VECREG:$src))), (v2f64 VECREG:$src)>;
4381 def : Pat<(i128 (bitconvert (v16i8 VECREG:$src))),
4382 (ORi128_vec VECREG:$src)>;
4383 def : Pat<(i128 (bitconvert (v8i16 VECREG:$src))),
4384 (ORi128_vec VECREG:$src)>;
4385 def : Pat<(i128 (bitconvert (v4i32 VECREG:$src))),
4386 (ORi128_vec VECREG:$src)>;
4387 def : Pat<(i128 (bitconvert (v2i64 VECREG:$src))),
4388 (ORi128_vec VECREG:$src)>;
4389 def : Pat<(i128 (bitconvert (v4f32 VECREG:$src))),
4390 (ORi128_vec VECREG:$src)>;
4391 def : Pat<(i128 (bitconvert (v2f64 VECREG:$src))),
4392 (ORi128_vec VECREG:$src)>;
4394 def : Pat<(v16i8 (bitconvert (i128 GPRC:$src))),
4395 (v16i8 (ORvec_i128 GPRC:$src))>;
4396 def : Pat<(v8i16 (bitconvert (i128 GPRC:$src))),
4397 (v8i16 (ORvec_i128 GPRC:$src))>;
4398 def : Pat<(v4i32 (bitconvert (i128 GPRC:$src))),
4399 (v4i32 (ORvec_i128 GPRC:$src))>;
4400 def : Pat<(v2i64 (bitconvert (i128 GPRC:$src))),
4401 (v2i64 (ORvec_i128 GPRC:$src))>;
4402 def : Pat<(v4f32 (bitconvert (i128 GPRC:$src))),
4403 (v4f32 (ORvec_i128 GPRC:$src))>;
4404 def : Pat<(v2f64 (bitconvert (i128 GPRC:$src))),
4405 (v2f64 (ORvec_i128 GPRC:$src))>;
4407 //===----------------------------------------------------------------------===//
4408 // Instruction patterns:
4409 //===----------------------------------------------------------------------===//
4411 // General 32-bit constants:
4412 def : Pat<(i32 imm:$imm),
4413 (IOHLr32 (ILHUr32 (HI16 imm:$imm)), (LO16 imm:$imm))>;
4415 // Single precision float constants:
4416 def : Pat<(f32 fpimm:$imm),
4417 (IOHLf32 (ILHUf32 (HI16_f32 fpimm:$imm)), (LO16_f32 fpimm:$imm))>;
4419 // General constant 32-bit vectors
4420 def : Pat<(v4i32 v4i32Imm:$imm),
4421 (IOHLv4i32 (v4i32 (ILHUv4i32 (HI16_vec v4i32Imm:$imm))),
4422 (LO16_vec v4i32Imm:$imm))>;
4425 def : Pat<(i8 imm:$imm),
4428 //===----------------------------------------------------------------------===//
4429 // Zero/Any/Sign extensions
4430 //===----------------------------------------------------------------------===//
4432 // sext 8->32: Sign extend bytes to words
4433 def : Pat<(sext_inreg R32C:$rSrc, i8),
4434 (XSHWr32 (XSBHr32 R32C:$rSrc))>;
4436 def : Pat<(i32 (sext R8C:$rSrc)),
4437 (XSHWr16 (XSBHr8 R8C:$rSrc))>;
4439 // sext 8->64: Sign extend bytes to double word
4440 def : Pat<(sext_inreg R64C:$rSrc, i8),
4441 (XSWDr64_inreg (XSHWr64 (XSBHr64 R64C:$rSrc)))>;
4443 def : Pat<(i64 (sext R8C:$rSrc)),
4444 (XSWDr64 (XSHWr16 (XSBHr8 R8C:$rSrc)))>;
4446 // zext 8->16: Zero extend bytes to halfwords
4447 def : Pat<(i16 (zext R8C:$rSrc)),
4448 (ANDHIi8i16 R8C:$rSrc, 0xff)>;
4450 // zext 8->32: Zero extend bytes to words
4451 def : Pat<(i32 (zext R8C:$rSrc)),
4452 (ANDIi8i32 R8C:$rSrc, 0xff)>;
4454 // zext 8->64: Zero extend bytes to double words
4455 def : Pat<(i64 (zext R8C:$rSrc)),
4456 (ORi64_v2i64 (SELBv4i32 (ROTQMBYv4i32
4457 (ORv4i32_i32 (ANDIi8i32 R8C:$rSrc, 0xff)),
4460 (FSMBIv4i32 0x0f0f)))>;
4462 // anyext 8->16: Extend 8->16 bits, irrespective of sign, preserves high bits
4463 def : Pat<(i16 (anyext R8C:$rSrc)),
4464 (ORHIi8i16 R8C:$rSrc, 0)>;
4466 // anyext 8->32: Extend 8->32 bits, irrespective of sign, preserves high bits
4467 def : Pat<(i32 (anyext R8C:$rSrc)),
4468 (ORIi8i32 R8C:$rSrc, 0)>;
4470 // sext 16->64: Sign extend halfword to double word
4471 def : Pat<(sext_inreg R64C:$rSrc, i16),
4472 (XSWDr64_inreg (XSHWr64 R64C:$rSrc))>;
4474 def : Pat<(sext R16C:$rSrc),
4475 (XSWDr64 (XSHWr16 R16C:$rSrc))>;
4477 // zext 16->32: Zero extend halfwords to words
4478 def : Pat<(i32 (zext R16C:$rSrc)),
4479 (ANDi16i32 R16C:$rSrc, (ILAr32 0xffff))>;
4481 def : Pat<(i32 (zext (and R16C:$rSrc, 0xf))),
4482 (ANDIi16i32 R16C:$rSrc, 0xf)>;
4484 def : Pat<(i32 (zext (and R16C:$rSrc, 0xff))),
4485 (ANDIi16i32 R16C:$rSrc, 0xff)>;
4487 def : Pat<(i32 (zext (and R16C:$rSrc, 0xfff))),
4488 (ANDIi16i32 R16C:$rSrc, 0xfff)>;
4490 // anyext 16->32: Extend 16->32 bits, irrespective of sign
4491 def : Pat<(i32 (anyext R16C:$rSrc)),
4492 (ORIi16i32 R16C:$rSrc, 0)>;
4494 //===----------------------------------------------------------------------===//
4496 // These truncates are for the SPU's supported types (i8, i16, i32). i64 and
4497 // above are custom lowered.
4498 //===----------------------------------------------------------------------===//
4500 def : Pat<(i8 (trunc GPRC:$src)),
4502 (SHUFBgprc GPRC:$src, GPRC:$src,
4503 (IOHLv4i32 (ILHUv4i32 0x0f0f), 0x0f0f)))>;
4505 def : Pat<(i8 (trunc R64C:$src)),
4508 (ORv2i64_i64 R64C:$src),
4509 (ORv2i64_i64 R64C:$src),
4510 (IOHLv4i32 (ILHUv4i32 0x0707), 0x0707)))>;
4512 def : Pat<(i8 (trunc R32C:$src)),
4515 (ORv4i32_i32 R32C:$src),
4516 (ORv4i32_i32 R32C:$src),
4517 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)))>;
4519 def : Pat<(i8 (trunc R16C:$src)),
4522 (ORv8i16_i16 R16C:$src),
4523 (ORv8i16_i16 R16C:$src),
4524 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)))>;
4526 def : Pat<(i16 (trunc GPRC:$src)),
4528 (SHUFBgprc GPRC:$src, GPRC:$src,
4529 (IOHLv4i32 (ILHUv4i32 0x0e0f), 0x0e0f)))>;
4531 def : Pat<(i16 (trunc R64C:$src)),
4534 (ORv2i64_i64 R64C:$src),
4535 (ORv2i64_i64 R64C:$src),
4536 (IOHLv4i32 (ILHUv4i32 0x0607), 0x0607)))>;
4538 def : Pat<(i16 (trunc R32C:$src)),
4541 (ORv4i32_i32 R32C:$src),
4542 (ORv4i32_i32 R32C:$src),
4543 (IOHLv4i32 (ILHUv4i32 0x0203), 0x0203)))>;
4545 def : Pat<(i32 (trunc GPRC:$src)),
4547 (SHUFBgprc GPRC:$src, GPRC:$src,
4548 (IOHLv4i32 (ILHUv4i32 0x0c0d), 0x0e0f)))>;
4550 def : Pat<(i32 (trunc R64C:$src)),
4553 (ORv2i64_i64 R64C:$src),
4554 (ORv2i64_i64 R64C:$src),
4555 (IOHLv4i32 (ILHUv4i32 0x0405), 0x0607)))>;
4557 //===----------------------------------------------------------------------===//
4558 // Address generation: SPU, like PPC, has to split addresses into high and
4559 // low parts in order to load them into a register.
4560 //===----------------------------------------------------------------------===//
4562 def : Pat<(SPUaform tglobaladdr:$in, 0), (ILAlsa tglobaladdr:$in)>;
4563 def : Pat<(SPUaform texternalsym:$in, 0), (ILAlsa texternalsym:$in)>;
4564 def : Pat<(SPUaform tjumptable:$in, 0), (ILAlsa tjumptable:$in)>;
4565 def : Pat<(SPUaform tconstpool:$in, 0), (ILAlsa tconstpool:$in)>;
4567 def : Pat<(SPUindirect (SPUhi tglobaladdr:$in, 0),
4568 (SPUlo tglobaladdr:$in, 0)),
4569 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4571 def : Pat<(SPUindirect (SPUhi texternalsym:$in, 0),
4572 (SPUlo texternalsym:$in, 0)),
4573 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4575 def : Pat<(SPUindirect (SPUhi tjumptable:$in, 0),
4576 (SPUlo tjumptable:$in, 0)),
4577 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4579 def : Pat<(SPUindirect (SPUhi tconstpool:$in, 0),
4580 (SPUlo tconstpool:$in, 0)),
4581 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4583 def : Pat<(add (SPUhi tglobaladdr:$in, 0), (SPUlo tglobaladdr:$in, 0)),
4584 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4586 def : Pat<(add (SPUhi texternalsym:$in, 0), (SPUlo texternalsym:$in, 0)),
4587 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4589 def : Pat<(add (SPUhi tjumptable:$in, 0), (SPUlo tjumptable:$in, 0)),
4590 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4592 def : Pat<(add (SPUhi tconstpool:$in, 0), (SPUlo tconstpool:$in, 0)),
4593 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4596 include "CellSDKIntrinsics.td"
4597 // Various math operator instruction sequences
4598 include "SPUMathInstr.td"
4599 // 64-bit "instructions"/support
4600 include "SPU64InstrInfo.td"
4601 // 128-bit "instructions"/support
4602 include "SPU128InstrInfo.td"