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:$rB))]>;
1389 multiclass BitwiseOr
1391 def v16i8: ORVecInst<v16i8>;
1392 def v8i16: ORVecInst<v8i16>;
1393 def v4i32: ORVecInst<v4i32>;
1394 def v2i64: ORVecInst<v2i64>;
1396 def v4f32: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1397 [(set (v4f32 VECREG:$rT),
1398 (v4f32 (bitconvert (or (v4i32 VECREG:$rA),
1399 (v4i32 VECREG:$rB)))))]>;
1401 def v2f64: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1402 [(set (v2f64 VECREG:$rT),
1403 (v2f64 (bitconvert (or (v2i64 VECREG:$rA),
1404 (v2i64 VECREG:$rB)))))]>;
1406 def r128: ORRegInst<GPRC>;
1407 def r64: ORRegInst<R64C>;
1408 def r32: ORRegInst<R32C>;
1409 def r16: ORRegInst<R16C>;
1410 def r8: ORRegInst<R8C>;
1412 // OR instructions used to copy f32 and f64 registers.
1413 def f32: ORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
1414 [/* no pattern */]>;
1416 def f64: ORInst<(outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
1417 [/* no pattern */]>;
1420 defm OR : BitwiseOr;
1422 //===----------------------------------------------------------------------===//
1423 // SPU::PREFSLOT2VEC and VEC2PREFSLOT re-interpretations of registers
1424 //===----------------------------------------------------------------------===//
1425 def : Pat<(v16i8 (SPUprefslot2vec R8C:$rA)),
1426 (COPY_TO_REGCLASS R8C:$rA, VECREG)>;
1428 def : Pat<(v8i16 (SPUprefslot2vec R16C:$rA)),
1429 (COPY_TO_REGCLASS R16C:$rA, VECREG)>;
1431 def : Pat<(v4i32 (SPUprefslot2vec R32C:$rA)),
1432 (COPY_TO_REGCLASS R32C:$rA, VECREG)>;
1434 def : Pat<(v2i64 (SPUprefslot2vec R64C:$rA)),
1435 (COPY_TO_REGCLASS R64C:$rA, VECREG)>;
1437 def : Pat<(v4f32 (SPUprefslot2vec R32FP:$rA)),
1438 (COPY_TO_REGCLASS R32FP:$rA, VECREG)>;
1440 def : Pat<(v2f64 (SPUprefslot2vec R64FP:$rA)),
1441 (COPY_TO_REGCLASS R64FP:$rA, VECREG)>;
1443 def : Pat<(i8 (SPUvec2prefslot (v16i8 VECREG:$rA))),
1444 (COPY_TO_REGCLASS (v16i8 VECREG:$rA), R8C)>;
1446 def : Pat<(i16 (SPUvec2prefslot (v8i16 VECREG:$rA))),
1447 (COPY_TO_REGCLASS (v8i16 VECREG:$rA), R16C)>;
1449 def : Pat<(i32 (SPUvec2prefslot (v4i32 VECREG:$rA))),
1450 (COPY_TO_REGCLASS (v4i32 VECREG:$rA), R32C)>;
1452 def : Pat<(i64 (SPUvec2prefslot (v2i64 VECREG:$rA))),
1453 (COPY_TO_REGCLASS (v2i64 VECREG:$rA), R64C)>;
1455 def : Pat<(f32 (SPUvec2prefslot (v4f32 VECREG:$rA))),
1456 (COPY_TO_REGCLASS (v4f32 VECREG:$rA), R32FP)>;
1458 def : Pat<(f64 (SPUvec2prefslot (v2f64 VECREG:$rA))),
1459 (COPY_TO_REGCLASS (v2f64 VECREG:$rA), R64FP)>;
1461 // Load Register: This is an assembler alias for a bitwise OR of a register
1462 // against itself. It's here because it brings some clarity to assembly
1465 let hasCtrlDep = 1 in {
1466 class LRInst<dag OOL, dag IOL>
1467 : SPUInstr<OOL, IOL, "lr\t$rT, $rA", IntegerOp> {
1471 let Pattern = [/*no pattern*/];
1473 let Inst{0-10} = 0b10000010000; /* It's an OR operation */
1474 let Inst{11-17} = RA;
1475 let Inst{18-24} = RA;
1476 let Inst{25-31} = RT;
1479 class LRVecInst<ValueType vectype>:
1480 LRInst<(outs VECREG:$rT), (ins VECREG:$rA)>;
1482 class LRRegInst<RegisterClass rclass>:
1483 LRInst<(outs rclass:$rT), (ins rclass:$rA)>;
1485 multiclass LoadRegister {
1486 def v2i64: LRVecInst<v2i64>;
1487 def v2f64: LRVecInst<v2f64>;
1488 def v4i32: LRVecInst<v4i32>;
1489 def v4f32: LRVecInst<v4f32>;
1490 def v8i16: LRVecInst<v8i16>;
1491 def v16i8: LRVecInst<v16i8>;
1493 def r128: LRRegInst<GPRC>;
1494 def r64: LRRegInst<R64C>;
1495 def f64: LRRegInst<R64FP>;
1496 def r32: LRRegInst<R32C>;
1497 def f32: LRRegInst<R32FP>;
1498 def r16: LRRegInst<R16C>;
1499 def r8: LRRegInst<R8C>;
1502 defm LR: LoadRegister;
1505 // ORC: Bitwise "or" with complement (c = a | ~b)
1507 class ORCInst<dag OOL, dag IOL, list<dag> pattern>:
1508 RRForm<0b10010010000, OOL, IOL, "orc\t$rT, $rA, $rB",
1509 IntegerOp, pattern>;
1511 class ORCVecInst<ValueType vectype>:
1512 ORCInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1513 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1514 (vnot (vectype VECREG:$rB))))]>;
1516 class ORCRegInst<RegisterClass rclass>:
1517 ORCInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1518 [(set rclass:$rT, (or rclass:$rA, (not rclass:$rB)))]>;
1520 multiclass BitwiseOrComplement
1522 def v16i8: ORCVecInst<v16i8>;
1523 def v8i16: ORCVecInst<v8i16>;
1524 def v4i32: ORCVecInst<v4i32>;
1525 def v2i64: ORCVecInst<v2i64>;
1527 def r128: ORCRegInst<GPRC>;
1528 def r64: ORCRegInst<R64C>;
1529 def r32: ORCRegInst<R32C>;
1530 def r16: ORCRegInst<R16C>;
1531 def r8: ORCRegInst<R8C>;
1534 defm ORC : BitwiseOrComplement;
1536 // OR byte immediate
1537 class ORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1538 RI10Form<0b01100000, OOL, IOL, "orbi\t$rT, $rA, $val",
1539 IntegerOp, pattern>;
1541 class ORBIVecInst<ValueType vectype, PatLeaf immpred>:
1542 ORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1543 [(set (v16i8 VECREG:$rT), (or (vectype VECREG:$rA),
1544 (vectype immpred:$val)))]>;
1546 multiclass BitwiseOrByteImm
1548 def v16i8: ORBIVecInst<v16i8, v16i8U8Imm>;
1550 def r8: ORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1551 [(set R8C:$rT, (or R8C:$rA, immU8:$val))]>;
1554 defm ORBI : BitwiseOrByteImm;
1556 // OR halfword immediate
1557 class ORHIInst<dag OOL, dag IOL, list<dag> pattern>:
1558 RI10Form<0b10100000, OOL, IOL, "orhi\t$rT, $rA, $val",
1559 IntegerOp, pattern>;
1561 class ORHIVecInst<ValueType vectype, PatLeaf immpred>:
1562 ORHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1563 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1566 multiclass BitwiseOrHalfwordImm
1568 def v8i16: ORHIVecInst<v8i16, v8i16Uns10Imm>;
1570 def r16: ORHIInst<(outs R16C:$rT), (ins R16C:$rA, u10imm:$val),
1571 [(set R16C:$rT, (or R16C:$rA, i16ImmUns10:$val))]>;
1573 // Specialized ORHI form used to promote 8-bit registers to 16-bit
1574 def i8i16: ORHIInst<(outs R16C:$rT), (ins R8C:$rA, s10imm:$val),
1575 [(set R16C:$rT, (or (anyext R8C:$rA),
1576 i16ImmSExt10:$val))]>;
1579 defm ORHI : BitwiseOrHalfwordImm;
1581 class ORIInst<dag OOL, dag IOL, list<dag> pattern>:
1582 RI10Form<0b00100000, OOL, IOL, "ori\t$rT, $rA, $val",
1583 IntegerOp, pattern>;
1585 class ORIVecInst<ValueType vectype, PatLeaf immpred>:
1586 ORIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1587 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1590 // Bitwise "or" with immediate
1591 multiclass BitwiseOrImm
1593 def v4i32: ORIVecInst<v4i32, v4i32Uns10Imm>;
1595 def r32: ORIInst<(outs R32C:$rT), (ins R32C:$rA, u10imm_i32:$val),
1596 [(set R32C:$rT, (or R32C:$rA, i32ImmUns10:$val))]>;
1598 // i16i32: hacked version of the ori instruction to extend 16-bit quantities
1599 // to 32-bit quantities. used exclusively to match "anyext" conversions (vide
1600 // infra "anyext 16->32" pattern.)
1601 def i16i32: ORIInst<(outs R32C:$rT), (ins R16C:$rA, s10imm_i32:$val),
1602 [(set R32C:$rT, (or (anyext R16C:$rA),
1603 i32ImmSExt10:$val))]>;
1605 // i8i32: Hacked version of the ORI instruction to extend 16-bit quantities
1606 // to 32-bit quantities. Used exclusively to match "anyext" conversions (vide
1607 // infra "anyext 16->32" pattern.)
1608 def i8i32: ORIInst<(outs R32C:$rT), (ins R8C:$rA, s10imm_i32:$val),
1609 [(set R32C:$rT, (or (anyext R8C:$rA),
1610 i32ImmSExt10:$val))]>;
1613 defm ORI : BitwiseOrImm;
1615 // ORX: "or" across the vector: or's $rA's word slots leaving the result in
1616 // $rT[0], slots 1-3 are zeroed.
1618 // FIXME: Needs to match an intrinsic pattern.
1620 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1621 "orx\t$rT, $rA, $rB", IntegerOp,
1626 class XORInst<dag OOL, dag IOL, list<dag> pattern> :
1627 RRForm<0b10010010000, OOL, IOL, "xor\t$rT, $rA, $rB",
1628 IntegerOp, pattern>;
1630 class XORVecInst<ValueType vectype>:
1631 XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1632 [(set (vectype VECREG:$rT), (xor (vectype VECREG:$rA),
1633 (vectype VECREG:$rB)))]>;
1635 class XORRegInst<RegisterClass rclass>:
1636 XORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1637 [(set rclass:$rT, (xor rclass:$rA, rclass:$rB))]>;
1639 multiclass BitwiseExclusiveOr
1641 def v16i8: XORVecInst<v16i8>;
1642 def v8i16: XORVecInst<v8i16>;
1643 def v4i32: XORVecInst<v4i32>;
1644 def v2i64: XORVecInst<v2i64>;
1646 def r128: XORRegInst<GPRC>;
1647 def r64: XORRegInst<R64C>;
1648 def r32: XORRegInst<R32C>;
1649 def r16: XORRegInst<R16C>;
1650 def r8: XORRegInst<R8C>;
1652 // XOR instructions used to negate f32 and f64 quantities.
1654 def fneg32: XORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32C:$rB),
1655 [/* no pattern */]>;
1657 def fneg64: XORInst<(outs R64FP:$rT), (ins R64FP:$rA, R64C:$rB),
1658 [/* no pattern */]>;
1660 def fnegvec: XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1661 [/* no pattern, see fneg{32,64} */]>;
1664 defm XOR : BitwiseExclusiveOr;
1666 //==----------------------------------------------------------
1668 class XORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1669 RI10Form<0b01100000, OOL, IOL, "xorbi\t$rT, $rA, $val",
1670 IntegerOp, pattern>;
1672 multiclass XorByteImm
1675 XORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1676 [(set (v16i8 VECREG:$rT), (xor (v16i8 VECREG:$rA), v16i8U8Imm:$val))]>;
1679 XORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1680 [(set R8C:$rT, (xor R8C:$rA, immU8:$val))]>;
1683 defm XORBI : XorByteImm;
1686 RI10Form<0b10100000, (outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1687 "xorhi\t$rT, $rA, $val", IntegerOp,
1688 [(set (v8i16 VECREG:$rT), (xor (v8i16 VECREG:$rA),
1689 v8i16SExt10Imm:$val))]>;
1692 RI10Form<0b10100000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
1693 "xorhi\t$rT, $rA, $val", IntegerOp,
1694 [(set R16C:$rT, (xor R16C:$rA, i16ImmSExt10:$val))]>;
1697 RI10Form<0b00100000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm_i32:$val),
1698 "xori\t$rT, $rA, $val", IntegerOp,
1699 [(set (v4i32 VECREG:$rT), (xor (v4i32 VECREG:$rA),
1700 v4i32SExt10Imm:$val))]>;
1703 RI10Form<0b00100000, (outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
1704 "xori\t$rT, $rA, $val", IntegerOp,
1705 [(set R32C:$rT, (xor R32C:$rA, i32ImmSExt10:$val))]>;
1709 class NANDInst<dag OOL, dag IOL, list<dag> pattern>:
1710 RRForm<0b10010011000, OOL, IOL, "nand\t$rT, $rA, $rB",
1711 IntegerOp, pattern>;
1713 class NANDVecInst<ValueType vectype>:
1714 NANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1715 [(set (vectype VECREG:$rT), (vnot (and (vectype VECREG:$rA),
1716 (vectype VECREG:$rB))))]>;
1717 class NANDRegInst<RegisterClass rclass>:
1718 NANDInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1719 [(set rclass:$rT, (not (and rclass:$rA, rclass:$rB)))]>;
1721 multiclass BitwiseNand
1723 def v16i8: NANDVecInst<v16i8>;
1724 def v8i16: NANDVecInst<v8i16>;
1725 def v4i32: NANDVecInst<v4i32>;
1726 def v2i64: NANDVecInst<v2i64>;
1728 def r128: NANDRegInst<GPRC>;
1729 def r64: NANDRegInst<R64C>;
1730 def r32: NANDRegInst<R32C>;
1731 def r16: NANDRegInst<R16C>;
1732 def r8: NANDRegInst<R8C>;
1735 defm NAND : BitwiseNand;
1739 class NORInst<dag OOL, dag IOL, list<dag> pattern>:
1740 RRForm<0b10010010000, OOL, IOL, "nor\t$rT, $rA, $rB",
1741 IntegerOp, pattern>;
1743 class NORVecInst<ValueType vectype>:
1744 NORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1745 [(set (vectype VECREG:$rT), (vnot (or (vectype VECREG:$rA),
1746 (vectype VECREG:$rB))))]>;
1747 class NORRegInst<RegisterClass rclass>:
1748 NORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1749 [(set rclass:$rT, (not (or rclass:$rA, rclass:$rB)))]>;
1751 multiclass BitwiseNor
1753 def v16i8: NORVecInst<v16i8>;
1754 def v8i16: NORVecInst<v8i16>;
1755 def v4i32: NORVecInst<v4i32>;
1756 def v2i64: NORVecInst<v2i64>;
1758 def r128: NORRegInst<GPRC>;
1759 def r64: NORRegInst<R64C>;
1760 def r32: NORRegInst<R32C>;
1761 def r16: NORRegInst<R16C>;
1762 def r8: NORRegInst<R8C>;
1765 defm NOR : BitwiseNor;
1768 class SELBInst<dag OOL, dag IOL, list<dag> pattern>:
1769 RRRForm<0b1000, OOL, IOL, "selb\t$rT, $rA, $rB, $rC",
1770 IntegerOp, pattern>;
1772 class SELBVecInst<ValueType vectype, PatFrag vnot_frag = vnot>:
1773 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1774 [(set (vectype VECREG:$rT),
1775 (or (and (vectype VECREG:$rC), (vectype VECREG:$rB)),
1776 (and (vnot_frag (vectype VECREG:$rC)),
1777 (vectype VECREG:$rA))))]>;
1779 class SELBVecVCondInst<ValueType vectype>:
1780 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1781 [(set (vectype VECREG:$rT),
1782 (select (vectype VECREG:$rC),
1783 (vectype VECREG:$rB),
1784 (vectype VECREG:$rA)))]>;
1786 class SELBVecCondInst<ValueType vectype>:
1787 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, R32C:$rC),
1788 [(set (vectype VECREG:$rT),
1790 (vectype VECREG:$rB),
1791 (vectype VECREG:$rA)))]>;
1793 class SELBRegInst<RegisterClass rclass>:
1794 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rclass:$rC),
1796 (or (and rclass:$rB, rclass:$rC),
1797 (and rclass:$rA, (not rclass:$rC))))]>;
1799 class SELBRegCondInst<RegisterClass rcond, RegisterClass rclass>:
1800 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rcond:$rC),
1802 (select rcond:$rC, rclass:$rB, rclass:$rA))]>;
1804 multiclass SelectBits
1806 def v16i8: SELBVecInst<v16i8>;
1807 def v8i16: SELBVecInst<v8i16>;
1808 def v4i32: SELBVecInst<v4i32>;
1809 def v2i64: SELBVecInst<v2i64, vnot_cell_conv>;
1811 def r128: SELBRegInst<GPRC>;
1812 def r64: SELBRegInst<R64C>;
1813 def r32: SELBRegInst<R32C>;
1814 def r16: SELBRegInst<R16C>;
1815 def r8: SELBRegInst<R8C>;
1817 def v16i8_cond: SELBVecCondInst<v16i8>;
1818 def v8i16_cond: SELBVecCondInst<v8i16>;
1819 def v4i32_cond: SELBVecCondInst<v4i32>;
1820 def v2i64_cond: SELBVecCondInst<v2i64>;
1822 def v16i8_vcond: SELBVecCondInst<v16i8>;
1823 def v8i16_vcond: SELBVecCondInst<v8i16>;
1824 def v4i32_vcond: SELBVecCondInst<v4i32>;
1825 def v2i64_vcond: SELBVecCondInst<v2i64>;
1828 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1829 [(set (v4f32 VECREG:$rT),
1830 (select (v4i32 VECREG:$rC),
1832 (v4f32 VECREG:$rA)))]>;
1834 // SELBr64_cond is defined in SPU64InstrInfo.td
1835 def r32_cond: SELBRegCondInst<R32C, R32C>;
1836 def f32_cond: SELBRegCondInst<R32C, R32FP>;
1837 def r16_cond: SELBRegCondInst<R16C, R16C>;
1838 def r8_cond: SELBRegCondInst<R8C, R8C>;
1841 defm SELB : SelectBits;
1843 class SPUselbPatVec<ValueType vectype, SPUInstr inst>:
1844 Pat<(SPUselb (vectype VECREG:$rA), (vectype VECREG:$rB), (vectype VECREG:$rC)),
1845 (inst VECREG:$rA, VECREG:$rB, VECREG:$rC)>;
1847 def : SPUselbPatVec<v16i8, SELBv16i8>;
1848 def : SPUselbPatVec<v8i16, SELBv8i16>;
1849 def : SPUselbPatVec<v4i32, SELBv4i32>;
1850 def : SPUselbPatVec<v2i64, SELBv2i64>;
1852 class SPUselbPatReg<RegisterClass rclass, SPUInstr inst>:
1853 Pat<(SPUselb rclass:$rA, rclass:$rB, rclass:$rC),
1854 (inst rclass:$rA, rclass:$rB, rclass:$rC)>;
1856 def : SPUselbPatReg<R8C, SELBr8>;
1857 def : SPUselbPatReg<R16C, SELBr16>;
1858 def : SPUselbPatReg<R32C, SELBr32>;
1859 def : SPUselbPatReg<R64C, SELBr64>;
1861 // EQV: Equivalence (1 for each same bit, otherwise 0)
1863 // Note: There are a lot of ways to match this bit operator and these patterns
1864 // attempt to be as exhaustive as possible.
1866 class EQVInst<dag OOL, dag IOL, list<dag> pattern>:
1867 RRForm<0b10010010000, OOL, IOL, "eqv\t$rT, $rA, $rB",
1868 IntegerOp, pattern>;
1870 class EQVVecInst<ValueType vectype>:
1871 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1872 [(set (vectype VECREG:$rT),
1873 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
1874 (and (vnot (vectype VECREG:$rA)),
1875 (vnot (vectype VECREG:$rB)))))]>;
1877 class EQVRegInst<RegisterClass rclass>:
1878 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1879 [(set rclass:$rT, (or (and rclass:$rA, rclass:$rB),
1880 (and (not rclass:$rA), (not rclass:$rB))))]>;
1882 class EQVVecPattern1<ValueType vectype>:
1883 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1884 [(set (vectype VECREG:$rT),
1885 (xor (vectype VECREG:$rA), (vnot (vectype VECREG:$rB))))]>;
1887 class EQVRegPattern1<RegisterClass rclass>:
1888 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1889 [(set rclass:$rT, (xor rclass:$rA, (not rclass:$rB)))]>;
1891 class EQVVecPattern2<ValueType vectype>:
1892 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1893 [(set (vectype VECREG:$rT),
1894 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
1895 (vnot (or (vectype VECREG:$rA), (vectype VECREG:$rB)))))]>;
1897 class EQVRegPattern2<RegisterClass rclass>:
1898 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1900 (or (and rclass:$rA, rclass:$rB),
1901 (not (or rclass:$rA, rclass:$rB))))]>;
1903 class EQVVecPattern3<ValueType vectype>:
1904 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1905 [(set (vectype VECREG:$rT),
1906 (not (xor (vectype VECREG:$rA), (vectype VECREG:$rB))))]>;
1908 class EQVRegPattern3<RegisterClass rclass>:
1909 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1910 [(set rclass:$rT, (not (xor rclass:$rA, rclass:$rB)))]>;
1912 multiclass BitEquivalence
1914 def v16i8: EQVVecInst<v16i8>;
1915 def v8i16: EQVVecInst<v8i16>;
1916 def v4i32: EQVVecInst<v4i32>;
1917 def v2i64: EQVVecInst<v2i64>;
1919 def v16i8_1: EQVVecPattern1<v16i8>;
1920 def v8i16_1: EQVVecPattern1<v8i16>;
1921 def v4i32_1: EQVVecPattern1<v4i32>;
1922 def v2i64_1: EQVVecPattern1<v2i64>;
1924 def v16i8_2: EQVVecPattern2<v16i8>;
1925 def v8i16_2: EQVVecPattern2<v8i16>;
1926 def v4i32_2: EQVVecPattern2<v4i32>;
1927 def v2i64_2: EQVVecPattern2<v2i64>;
1929 def v16i8_3: EQVVecPattern3<v16i8>;
1930 def v8i16_3: EQVVecPattern3<v8i16>;
1931 def v4i32_3: EQVVecPattern3<v4i32>;
1932 def v2i64_3: EQVVecPattern3<v2i64>;
1934 def r128: EQVRegInst<GPRC>;
1935 def r64: EQVRegInst<R64C>;
1936 def r32: EQVRegInst<R32C>;
1937 def r16: EQVRegInst<R16C>;
1938 def r8: EQVRegInst<R8C>;
1940 def r128_1: EQVRegPattern1<GPRC>;
1941 def r64_1: EQVRegPattern1<R64C>;
1942 def r32_1: EQVRegPattern1<R32C>;
1943 def r16_1: EQVRegPattern1<R16C>;
1944 def r8_1: EQVRegPattern1<R8C>;
1946 def r128_2: EQVRegPattern2<GPRC>;
1947 def r64_2: EQVRegPattern2<R64C>;
1948 def r32_2: EQVRegPattern2<R32C>;
1949 def r16_2: EQVRegPattern2<R16C>;
1950 def r8_2: EQVRegPattern2<R8C>;
1952 def r128_3: EQVRegPattern3<GPRC>;
1953 def r64_3: EQVRegPattern3<R64C>;
1954 def r32_3: EQVRegPattern3<R32C>;
1955 def r16_3: EQVRegPattern3<R16C>;
1956 def r8_3: EQVRegPattern3<R8C>;
1959 defm EQV: BitEquivalence;
1961 //===----------------------------------------------------------------------===//
1962 // Vector shuffle...
1963 //===----------------------------------------------------------------------===//
1964 // SPUshuffle is generated in LowerVECTOR_SHUFFLE and gets replaced with SHUFB.
1965 // See the SPUshuffle SDNode operand above, which sets up the DAG pattern
1966 // matcher to emit something when the LowerVECTOR_SHUFFLE generates a node with
1967 // the SPUISD::SHUFB opcode.
1968 //===----------------------------------------------------------------------===//
1970 class SHUFBInst<dag OOL, dag IOL, list<dag> pattern>:
1971 RRRForm<0b1000, OOL, IOL, "shufb\t$rT, $rA, $rB, $rC",
1972 IntegerOp, pattern>;
1974 class SHUFBVecInst<ValueType resultvec, ValueType maskvec>:
1975 SHUFBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1976 [(set (resultvec VECREG:$rT),
1977 (SPUshuffle (resultvec VECREG:$rA),
1978 (resultvec VECREG:$rB),
1979 (maskvec VECREG:$rC)))]>;
1981 class SHUFBGPRCInst:
1982 SHUFBInst<(outs VECREG:$rT), (ins GPRC:$rA, GPRC:$rB, VECREG:$rC),
1983 [/* no pattern */]>;
1985 multiclass ShuffleBytes
1987 def v16i8 : SHUFBVecInst<v16i8, v16i8>;
1988 def v16i8_m32 : SHUFBVecInst<v16i8, v4i32>;
1989 def v8i16 : SHUFBVecInst<v8i16, v16i8>;
1990 def v8i16_m32 : SHUFBVecInst<v8i16, v4i32>;
1991 def v4i32 : SHUFBVecInst<v4i32, v16i8>;
1992 def v4i32_m32 : SHUFBVecInst<v4i32, v4i32>;
1993 def v2i64 : SHUFBVecInst<v2i64, v16i8>;
1994 def v2i64_m32 : SHUFBVecInst<v2i64, v4i32>;
1996 def v4f32 : SHUFBVecInst<v4f32, v16i8>;
1997 def v4f32_m32 : SHUFBVecInst<v4f32, v4i32>;
1999 def v2f64 : SHUFBVecInst<v2f64, v16i8>;
2000 def v2f64_m32 : SHUFBVecInst<v2f64, v4i32>;
2002 def gprc : SHUFBGPRCInst;
2005 defm SHUFB : ShuffleBytes;
2007 //===----------------------------------------------------------------------===//
2008 // Shift and rotate group:
2009 //===----------------------------------------------------------------------===//
2011 class SHLHInst<dag OOL, dag IOL, list<dag> pattern>:
2012 RRForm<0b11111010000, OOL, IOL, "shlh\t$rT, $rA, $rB",
2013 RotateShift, pattern>;
2015 class SHLHVecInst<ValueType vectype>:
2016 SHLHInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2017 [(set (vectype VECREG:$rT),
2018 (SPUvec_shl (vectype VECREG:$rA), R16C:$rB))]>;
2020 multiclass ShiftLeftHalfword
2022 def v8i16: SHLHVecInst<v8i16>;
2023 def r16: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
2024 [(set R16C:$rT, (shl R16C:$rA, R16C:$rB))]>;
2025 def r16_r32: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2026 [(set R16C:$rT, (shl R16C:$rA, R32C:$rB))]>;
2029 defm SHLH : ShiftLeftHalfword;
2031 //===----------------------------------------------------------------------===//
2033 class SHLHIInst<dag OOL, dag IOL, list<dag> pattern>:
2034 RI7Form<0b11111010000, OOL, IOL, "shlhi\t$rT, $rA, $val",
2035 RotateShift, pattern>;
2037 class SHLHIVecInst<ValueType vectype>:
2038 SHLHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2039 [(set (vectype VECREG:$rT),
2040 (SPUvec_shl (vectype VECREG:$rA), (i16 uimm7:$val)))]>;
2042 multiclass ShiftLeftHalfwordImm
2044 def v8i16: SHLHIVecInst<v8i16>;
2045 def r16: SHLHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2046 [(set R16C:$rT, (shl R16C:$rA, (i16 uimm7:$val)))]>;
2049 defm SHLHI : ShiftLeftHalfwordImm;
2051 def : Pat<(SPUvec_shl (v8i16 VECREG:$rA), (i32 uimm7:$val)),
2052 (SHLHIv8i16 VECREG:$rA, (TO_IMM16 uimm7:$val))>;
2054 def : Pat<(shl R16C:$rA, (i32 uimm7:$val)),
2055 (SHLHIr16 R16C:$rA, (TO_IMM16 uimm7:$val))>;
2057 //===----------------------------------------------------------------------===//
2059 class SHLInst<dag OOL, dag IOL, list<dag> pattern>:
2060 RRForm<0b11111010000, OOL, IOL, "shl\t$rT, $rA, $rB",
2061 RotateShift, pattern>;
2063 multiclass ShiftLeftWord
2066 SHLInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2067 [(set (v4i32 VECREG:$rT),
2068 (SPUvec_shl (v4i32 VECREG:$rA), R16C:$rB))]>;
2070 SHLInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2071 [(set R32C:$rT, (shl R32C:$rA, R32C:$rB))]>;
2074 defm SHL: ShiftLeftWord;
2076 //===----------------------------------------------------------------------===//
2078 class SHLIInst<dag OOL, dag IOL, list<dag> pattern>:
2079 RI7Form<0b11111010000, OOL, IOL, "shli\t$rT, $rA, $val",
2080 RotateShift, pattern>;
2082 multiclass ShiftLeftWordImm
2085 SHLIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2086 [(set (v4i32 VECREG:$rT),
2087 (SPUvec_shl (v4i32 VECREG:$rA), (i32 uimm7:$val)))]>;
2090 SHLIInst<(outs R32C:$rT), (ins R32C:$rA, u7imm_i32:$val),
2091 [(set R32C:$rT, (shl R32C:$rA, (i32 uimm7:$val)))]>;
2094 defm SHLI : ShiftLeftWordImm;
2096 //===----------------------------------------------------------------------===//
2097 // SHLQBI vec form: Note that this will shift the entire vector (the 128-bit
2098 // register) to the left. Vector form is here to ensure type correctness.
2100 // The shift count is in the lowest 3 bits (29-31) of $rB, so only a bit shift
2101 // of 7 bits is actually possible.
2103 // Note also that SHLQBI/SHLQBII are used in conjunction with SHLQBY/SHLQBYI
2104 // to shift i64 and i128. SHLQBI is the residual left over after shifting by
2105 // bytes with SHLQBY.
2107 class SHLQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2108 RRForm<0b11011011100, OOL, IOL, "shlqbi\t$rT, $rA, $rB",
2109 RotateShift, pattern>;
2111 class SHLQBIVecInst<ValueType vectype>:
2112 SHLQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2113 [(set (vectype VECREG:$rT),
2114 (SPUshlquad_l_bits (vectype VECREG:$rA), R32C:$rB))]>;
2116 class SHLQBIRegInst<RegisterClass rclass>:
2117 SHLQBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2118 [/* no pattern */]>;
2120 multiclass ShiftLeftQuadByBits
2122 def v16i8: SHLQBIVecInst<v16i8>;
2123 def v8i16: SHLQBIVecInst<v8i16>;
2124 def v4i32: SHLQBIVecInst<v4i32>;
2125 def v4f32: SHLQBIVecInst<v4f32>;
2126 def v2i64: SHLQBIVecInst<v2i64>;
2127 def v2f64: SHLQBIVecInst<v2f64>;
2129 def r128: SHLQBIRegInst<GPRC>;
2132 defm SHLQBI : ShiftLeftQuadByBits;
2134 // See note above on SHLQBI. In this case, the predicate actually does then
2135 // enforcement, whereas with SHLQBI, we have to "take it on faith."
2136 class SHLQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2137 RI7Form<0b11011111100, OOL, IOL, "shlqbii\t$rT, $rA, $val",
2138 RotateShift, pattern>;
2140 class SHLQBIIVecInst<ValueType vectype>:
2141 SHLQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2142 [(set (vectype VECREG:$rT),
2143 (SPUshlquad_l_bits (vectype VECREG:$rA), (i32 bitshift:$val)))]>;
2145 multiclass ShiftLeftQuadByBitsImm
2147 def v16i8 : SHLQBIIVecInst<v16i8>;
2148 def v8i16 : SHLQBIIVecInst<v8i16>;
2149 def v4i32 : SHLQBIIVecInst<v4i32>;
2150 def v4f32 : SHLQBIIVecInst<v4f32>;
2151 def v2i64 : SHLQBIIVecInst<v2i64>;
2152 def v2f64 : SHLQBIIVecInst<v2f64>;
2155 defm SHLQBII : ShiftLeftQuadByBitsImm;
2157 // SHLQBY, SHLQBYI vector forms: Shift the entire vector to the left by bytes,
2158 // not by bits. See notes above on SHLQBI.
2160 class SHLQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2161 RI7Form<0b11111011100, OOL, IOL, "shlqby\t$rT, $rA, $rB",
2162 RotateShift, pattern>;
2164 class SHLQBYVecInst<ValueType vectype>:
2165 SHLQBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2166 [(set (vectype VECREG:$rT),
2167 (SPUshlquad_l_bytes (vectype VECREG:$rA), R32C:$rB))]>;
2169 multiclass ShiftLeftQuadBytes
2171 def v16i8: SHLQBYVecInst<v16i8>;
2172 def v8i16: SHLQBYVecInst<v8i16>;
2173 def v4i32: SHLQBYVecInst<v4i32>;
2174 def v4f32: SHLQBYVecInst<v4f32>;
2175 def v2i64: SHLQBYVecInst<v2i64>;
2176 def v2f64: SHLQBYVecInst<v2f64>;
2177 def r128: SHLQBYInst<(outs GPRC:$rT), (ins GPRC:$rA, R32C:$rB),
2178 [(set GPRC:$rT, (SPUshlquad_l_bytes GPRC:$rA, R32C:$rB))]>;
2181 defm SHLQBY: ShiftLeftQuadBytes;
2183 class SHLQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2184 RI7Form<0b11111111100, OOL, IOL, "shlqbyi\t$rT, $rA, $val",
2185 RotateShift, pattern>;
2187 class SHLQBYIVecInst<ValueType vectype>:
2188 SHLQBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2189 [(set (vectype VECREG:$rT),
2190 (SPUshlquad_l_bytes (vectype VECREG:$rA), (i32 uimm7:$val)))]>;
2192 multiclass ShiftLeftQuadBytesImm
2194 def v16i8: SHLQBYIVecInst<v16i8>;
2195 def v8i16: SHLQBYIVecInst<v8i16>;
2196 def v4i32: SHLQBYIVecInst<v4i32>;
2197 def v4f32: SHLQBYIVecInst<v4f32>;
2198 def v2i64: SHLQBYIVecInst<v2i64>;
2199 def v2f64: SHLQBYIVecInst<v2f64>;
2200 def r128: SHLQBYIInst<(outs GPRC:$rT), (ins GPRC:$rA, u7imm_i32:$val),
2202 (SPUshlquad_l_bytes GPRC:$rA, (i32 uimm7:$val)))]>;
2205 defm SHLQBYI : ShiftLeftQuadBytesImm;
2207 class SHLQBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2208 RRForm<0b00111001111, OOL, IOL, "shlqbybi\t$rT, $rA, $rB",
2209 RotateShift, pattern>;
2211 class SHLQBYBIVecInst<ValueType vectype>:
2212 SHLQBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2213 [/* no pattern */]>;
2215 class SHLQBYBIRegInst<RegisterClass rclass>:
2216 SHLQBYBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2217 [/* no pattern */]>;
2219 multiclass ShiftLeftQuadBytesBitCount
2221 def v16i8: SHLQBYBIVecInst<v16i8>;
2222 def v8i16: SHLQBYBIVecInst<v8i16>;
2223 def v4i32: SHLQBYBIVecInst<v4i32>;
2224 def v4f32: SHLQBYBIVecInst<v4f32>;
2225 def v2i64: SHLQBYBIVecInst<v2i64>;
2226 def v2f64: SHLQBYBIVecInst<v2f64>;
2228 def r128: SHLQBYBIRegInst<GPRC>;
2231 defm SHLQBYBI : ShiftLeftQuadBytesBitCount;
2233 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2235 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2236 class ROTHInst<dag OOL, dag IOL, list<dag> pattern>:
2237 RRForm<0b00111010000, OOL, IOL, "roth\t$rT, $rA, $rB",
2238 RotateShift, pattern>;
2240 class ROTHVecInst<ValueType vectype>:
2241 ROTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2242 [(set (vectype VECREG:$rT),
2243 (SPUvec_rotl VECREG:$rA, (v8i16 VECREG:$rB)))]>;
2245 class ROTHRegInst<RegisterClass rclass>:
2246 ROTHInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2247 [(set rclass:$rT, (rotl rclass:$rA, rclass:$rB))]>;
2249 multiclass RotateLeftHalfword
2251 def v8i16: ROTHVecInst<v8i16>;
2252 def r16: ROTHRegInst<R16C>;
2255 defm ROTH: RotateLeftHalfword;
2257 def ROTHr16_r32: ROTHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2258 [(set R16C:$rT, (rotl R16C:$rA, R32C:$rB))]>;
2260 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2261 // Rotate halfword, immediate:
2262 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2263 class ROTHIInst<dag OOL, dag IOL, list<dag> pattern>:
2264 RI7Form<0b00111110000, OOL, IOL, "rothi\t$rT, $rA, $val",
2265 RotateShift, pattern>;
2267 class ROTHIVecInst<ValueType vectype>:
2268 ROTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2269 [(set (vectype VECREG:$rT),
2270 (SPUvec_rotl VECREG:$rA, (i16 uimm7:$val)))]>;
2272 multiclass RotateLeftHalfwordImm
2274 def v8i16: ROTHIVecInst<v8i16>;
2275 def r16: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2276 [(set R16C:$rT, (rotl R16C:$rA, (i16 uimm7:$val)))]>;
2277 def r16_r32: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm_i32:$val),
2278 [(set R16C:$rT, (rotl R16C:$rA, (i32 uimm7:$val)))]>;
2281 defm ROTHI: RotateLeftHalfwordImm;
2283 def : Pat<(SPUvec_rotl (v8i16 VECREG:$rA), (i32 uimm7:$val)),
2284 (ROTHIv8i16 VECREG:$rA, (TO_IMM16 imm:$val))>;
2286 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2288 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2290 class ROTInst<dag OOL, dag IOL, list<dag> pattern>:
2291 RRForm<0b00011010000, OOL, IOL, "rot\t$rT, $rA, $rB",
2292 RotateShift, pattern>;
2294 class ROTVecInst<ValueType vectype>:
2295 ROTInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2296 [(set (vectype VECREG:$rT),
2297 (SPUvec_rotl (vectype VECREG:$rA), R32C:$rB))]>;
2299 class ROTRegInst<RegisterClass rclass>:
2300 ROTInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2302 (rotl rclass:$rA, R32C:$rB))]>;
2304 multiclass RotateLeftWord
2306 def v4i32: ROTVecInst<v4i32>;
2307 def r32: ROTRegInst<R32C>;
2310 defm ROT: RotateLeftWord;
2312 // The rotate amount is in the same bits whether we've got an 8-bit, 16-bit or
2314 def ROTr32_r16_anyext:
2315 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R16C:$rB),
2316 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R16C:$rB))))]>;
2318 def : Pat<(rotl R32C:$rA, (i32 (zext R16C:$rB))),
2319 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2321 def : Pat<(rotl R32C:$rA, (i32 (sext R16C:$rB))),
2322 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2324 def ROTr32_r8_anyext:
2325 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R8C:$rB),
2326 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R8C:$rB))))]>;
2328 def : Pat<(rotl R32C:$rA, (i32 (zext R8C:$rB))),
2329 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2331 def : Pat<(rotl R32C:$rA, (i32 (sext R8C:$rB))),
2332 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2334 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2335 // Rotate word, immediate
2336 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2338 class ROTIInst<dag OOL, dag IOL, list<dag> pattern>:
2339 RI7Form<0b00011110000, OOL, IOL, "roti\t$rT, $rA, $val",
2340 RotateShift, pattern>;
2342 class ROTIVecInst<ValueType vectype, Operand optype, ValueType inttype, PatLeaf pred>:
2343 ROTIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2344 [(set (vectype VECREG:$rT),
2345 (SPUvec_rotl (vectype VECREG:$rA), (inttype pred:$val)))]>;
2347 class ROTIRegInst<RegisterClass rclass, Operand optype, ValueType inttype, PatLeaf pred>:
2348 ROTIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2349 [(set rclass:$rT, (rotl rclass:$rA, (inttype pred:$val)))]>;
2351 multiclass RotateLeftWordImm
2353 def v4i32: ROTIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2354 def v4i32_i16: ROTIVecInst<v4i32, u7imm, i16, uimm7>;
2355 def v4i32_i8: ROTIVecInst<v4i32, u7imm_i8, i8, uimm7>;
2357 def r32: ROTIRegInst<R32C, u7imm_i32, i32, uimm7>;
2358 def r32_i16: ROTIRegInst<R32C, u7imm, i16, uimm7>;
2359 def r32_i8: ROTIRegInst<R32C, u7imm_i8, i8, uimm7>;
2362 defm ROTI : RotateLeftWordImm;
2364 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2365 // Rotate quad by byte (count)
2366 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2368 class ROTQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2369 RRForm<0b00111011100, OOL, IOL, "rotqby\t$rT, $rA, $rB",
2370 RotateShift, pattern>;
2372 class ROTQBYGenInst<ValueType type, RegisterClass rc>:
2373 ROTQBYInst<(outs rc:$rT), (ins rc:$rA, R32C:$rB),
2374 [(set (type rc:$rT),
2375 (SPUrotbytes_left (type rc:$rA), R32C:$rB))]>;
2377 class ROTQBYVecInst<ValueType type>:
2378 ROTQBYGenInst<type, VECREG>;
2380 multiclass RotateQuadLeftByBytes
2382 def v16i8: ROTQBYVecInst<v16i8>;
2383 def v8i16: ROTQBYVecInst<v8i16>;
2384 def v4i32: ROTQBYVecInst<v4i32>;
2385 def v4f32: ROTQBYVecInst<v4f32>;
2386 def v2i64: ROTQBYVecInst<v2i64>;
2387 def v2f64: ROTQBYVecInst<v2f64>;
2388 def i128: ROTQBYGenInst<i128, GPRC>;
2391 defm ROTQBY: RotateQuadLeftByBytes;
2393 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2394 // Rotate quad by byte (count), immediate
2395 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2397 class ROTQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2398 RI7Form<0b00111111100, OOL, IOL, "rotqbyi\t$rT, $rA, $val",
2399 RotateShift, pattern>;
2401 class ROTQBYIGenInst<ValueType type, RegisterClass rclass>:
2402 ROTQBYIInst<(outs rclass:$rT), (ins rclass:$rA, u7imm:$val),
2403 [(set (type rclass:$rT),
2404 (SPUrotbytes_left (type rclass:$rA), (i16 uimm7:$val)))]>;
2406 class ROTQBYIVecInst<ValueType vectype>:
2407 ROTQBYIGenInst<vectype, VECREG>;
2409 multiclass RotateQuadByBytesImm
2411 def v16i8: ROTQBYIVecInst<v16i8>;
2412 def v8i16: ROTQBYIVecInst<v8i16>;
2413 def v4i32: ROTQBYIVecInst<v4i32>;
2414 def v4f32: ROTQBYIVecInst<v4f32>;
2415 def v2i64: ROTQBYIVecInst<v2i64>;
2416 def vfi64: ROTQBYIVecInst<v2f64>;
2417 def i128: ROTQBYIGenInst<i128, GPRC>;
2420 defm ROTQBYI: RotateQuadByBytesImm;
2422 // See ROTQBY note above.
2423 class ROTQBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2424 RI7Form<0b00110011100, OOL, IOL,
2425 "rotqbybi\t$rT, $rA, $shift",
2426 RotateShift, pattern>;
2428 class ROTQBYBIVecInst<ValueType vectype, RegisterClass rclass>:
2429 ROTQBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, rclass:$shift),
2430 [(set (vectype VECREG:$rT),
2431 (SPUrotbytes_left_bits (vectype VECREG:$rA), rclass:$shift))]>;
2433 multiclass RotateQuadByBytesByBitshift {
2434 def v16i8_r32: ROTQBYBIVecInst<v16i8, R32C>;
2435 def v8i16_r32: ROTQBYBIVecInst<v8i16, R32C>;
2436 def v4i32_r32: ROTQBYBIVecInst<v4i32, R32C>;
2437 def v2i64_r32: ROTQBYBIVecInst<v2i64, R32C>;
2440 defm ROTQBYBI : RotateQuadByBytesByBitshift;
2442 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2443 // See ROTQBY note above.
2445 // Assume that the user of this instruction knows to shift the rotate count
2447 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2449 class ROTQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2450 RRForm<0b00011011100, OOL, IOL, "rotqbi\t$rT, $rA, $rB",
2451 RotateShift, pattern>;
2453 class ROTQBIVecInst<ValueType vectype>:
2454 ROTQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2455 [/* no pattern yet */]>;
2457 class ROTQBIRegInst<RegisterClass rclass>:
2458 ROTQBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2459 [/* no pattern yet */]>;
2461 multiclass RotateQuadByBitCount
2463 def v16i8: ROTQBIVecInst<v16i8>;
2464 def v8i16: ROTQBIVecInst<v8i16>;
2465 def v4i32: ROTQBIVecInst<v4i32>;
2466 def v2i64: ROTQBIVecInst<v2i64>;
2468 def r128: ROTQBIRegInst<GPRC>;
2469 def r64: ROTQBIRegInst<R64C>;
2472 defm ROTQBI: RotateQuadByBitCount;
2474 class ROTQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2475 RI7Form<0b00011111100, OOL, IOL, "rotqbii\t$rT, $rA, $val",
2476 RotateShift, pattern>;
2478 class ROTQBIIVecInst<ValueType vectype, Operand optype, ValueType inttype,
2480 ROTQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2481 [/* no pattern yet */]>;
2483 class ROTQBIIRegInst<RegisterClass rclass, Operand optype, ValueType inttype,
2485 ROTQBIIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2486 [/* no pattern yet */]>;
2488 multiclass RotateQuadByBitCountImm
2490 def v16i8: ROTQBIIVecInst<v16i8, u7imm_i32, i32, uimm7>;
2491 def v8i16: ROTQBIIVecInst<v8i16, u7imm_i32, i32, uimm7>;
2492 def v4i32: ROTQBIIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2493 def v2i64: ROTQBIIVecInst<v2i64, u7imm_i32, i32, uimm7>;
2495 def r128: ROTQBIIRegInst<GPRC, u7imm_i32, i32, uimm7>;
2496 def r64: ROTQBIIRegInst<R64C, u7imm_i32, i32, uimm7>;
2499 defm ROTQBII : RotateQuadByBitCountImm;
2501 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2502 // ROTHM v8i16 form:
2503 // NOTE(1): No vector rotate is generated by the C/C++ frontend (today),
2504 // so this only matches a synthetically generated/lowered code
2506 // NOTE(2): $rB must be negated before the right rotate!
2507 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2509 class ROTHMInst<dag OOL, dag IOL, list<dag> pattern>:
2510 RRForm<0b10111010000, OOL, IOL, "rothm\t$rT, $rA, $rB",
2511 RotateShift, pattern>;
2514 ROTHMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2515 [/* see patterns below - $rB must be negated */]>;
2517 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R32C:$rB),
2518 (ROTHMv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2520 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R16C:$rB),
2521 (ROTHMv8i16 VECREG:$rA,
2522 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2524 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R8C:$rB),
2525 (ROTHMv8i16 VECREG:$rA,
2526 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2528 // ROTHM r16 form: Rotate 16-bit quantity to right, zero fill at the left
2529 // Note: This instruction doesn't match a pattern because rB must be negated
2530 // for the instruction to work. Thus, the pattern below the instruction!
2533 ROTHMInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2534 [/* see patterns below - $rB must be negated! */]>;
2536 def : Pat<(srl R16C:$rA, R32C:$rB),
2537 (ROTHMr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2539 def : Pat<(srl R16C:$rA, R16C:$rB),
2541 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2543 def : Pat<(srl R16C:$rA, R8C:$rB),
2545 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2547 // ROTHMI v8i16 form: See the comment for ROTHM v8i16. The difference here is
2548 // that the immediate can be complemented, so that the user doesn't have to
2551 class ROTHMIInst<dag OOL, dag IOL, list<dag> pattern>:
2552 RI7Form<0b10111110000, OOL, IOL, "rothmi\t$rT, $rA, $val",
2553 RotateShift, pattern>;
2556 ROTHMIInst<(outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2557 [/* no pattern */]>;
2559 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i32 imm:$val)),
2560 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2562 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i16 imm:$val)),
2563 (ROTHMIv8i16 VECREG:$rA, (TO_IMM32 imm:$val))>;
2565 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i8 imm:$val)),
2566 (ROTHMIv8i16 VECREG:$rA, (TO_IMM32 imm:$val))>;
2569 ROTHMIInst<(outs R16C:$rT), (ins R16C:$rA, rothNeg7imm:$val),
2570 [/* no pattern */]>;
2572 def: Pat<(srl R16C:$rA, (i32 uimm7:$val)),
2573 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2575 def: Pat<(srl R16C:$rA, (i16 uimm7:$val)),
2576 (ROTHMIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
2578 def: Pat<(srl R16C:$rA, (i8 uimm7:$val)),
2579 (ROTHMIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
2581 // ROTM v4i32 form: See the ROTHM v8i16 comments.
2582 class ROTMInst<dag OOL, dag IOL, list<dag> pattern>:
2583 RRForm<0b10011010000, OOL, IOL, "rotm\t$rT, $rA, $rB",
2584 RotateShift, pattern>;
2587 ROTMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2588 [/* see patterns below - $rB must be negated */]>;
2590 def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), R32C:$rB),
2591 (ROTMv4i32 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2593 def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), R16C:$rB),
2594 (ROTMv4i32 VECREG:$rA,
2595 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2597 def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), R8C:$rB),
2598 (ROTMv4i32 VECREG:$rA,
2599 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2602 ROTMInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2603 [/* see patterns below - $rB must be negated */]>;
2605 def : Pat<(srl R32C:$rA, R32C:$rB),
2606 (ROTMr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
2608 def : Pat<(srl R32C:$rA, R16C:$rB),
2610 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2612 def : Pat<(srl R32C:$rA, R8C:$rB),
2614 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2616 // ROTMI v4i32 form: See the comment for ROTHM v8i16.
2618 RI7Form<0b10011110000, (outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2619 "rotmi\t$rT, $rA, $val", RotateShift,
2620 [(set (v4i32 VECREG:$rT),
2621 (SPUvec_srl VECREG:$rA, (i32 uimm7:$val)))]>;
2623 def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), (i16 uimm7:$val)),
2624 (ROTMIv4i32 VECREG:$rA, (TO_IMM32 uimm7:$val))>;
2626 def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), (i8 uimm7:$val)),
2627 (ROTMIv4i32 VECREG:$rA, (TO_IMM32 uimm7:$val))>;
2629 // ROTMI r32 form: know how to complement the immediate value.
2631 RI7Form<0b10011110000, (outs R32C:$rT), (ins R32C:$rA, rotNeg7imm:$val),
2632 "rotmi\t$rT, $rA, $val", RotateShift,
2633 [(set R32C:$rT, (srl R32C:$rA, (i32 uimm7:$val)))]>;
2635 def : Pat<(srl R32C:$rA, (i16 imm:$val)),
2636 (ROTMIr32 R32C:$rA, (TO_IMM32 uimm7:$val))>;
2638 def : Pat<(srl R32C:$rA, (i8 imm:$val)),
2639 (ROTMIr32 R32C:$rA, (TO_IMM32 uimm7:$val))>;
2641 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2642 // ROTQMBY: This is a vector form merely so that when used in an
2643 // instruction pattern, type checking will succeed. This instruction assumes
2644 // that the user knew to negate $rB.
2645 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2647 class ROTQMBYInst<dag OOL, dag IOL, list<dag> pattern>:
2648 RRForm<0b10111011100, OOL, IOL, "rotqmby\t$rT, $rA, $rB",
2649 RotateShift, pattern>;
2651 class ROTQMBYVecInst<ValueType vectype>:
2652 ROTQMBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2653 [/* no pattern, $rB must be negated */]>;
2655 class ROTQMBYRegInst<RegisterClass rclass>:
2656 ROTQMBYInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2657 [/* no pattern */]>;
2659 multiclass RotateQuadBytes
2661 def v16i8: ROTQMBYVecInst<v16i8>;
2662 def v8i16: ROTQMBYVecInst<v8i16>;
2663 def v4i32: ROTQMBYVecInst<v4i32>;
2664 def v2i64: ROTQMBYVecInst<v2i64>;
2666 def r128: ROTQMBYRegInst<GPRC>;
2667 def r64: ROTQMBYRegInst<R64C>;
2670 defm ROTQMBY : RotateQuadBytes;
2672 def : Pat<(SPUsrl_bytes GPRC:$rA, R32C:$rB),
2673 (ROTQMBYr128 GPRC:$rA,
2674 (SFIr32 R32C:$rB, 0))>;
2676 class ROTQMBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2677 RI7Form<0b10111111100, OOL, IOL, "rotqmbyi\t$rT, $rA, $val",
2678 RotateShift, pattern>;
2680 class ROTQMBYIVecInst<ValueType vectype>:
2681 ROTQMBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2682 [/* no pattern */]>;
2684 class ROTQMBYIRegInst<RegisterClass rclass, Operand optype, ValueType inttype,
2686 ROTQMBYIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2687 [/* no pattern */]>;
2689 // 128-bit zero extension form:
2690 class ROTQMBYIZExtInst<RegisterClass rclass, Operand optype, PatLeaf pred>:
2691 ROTQMBYIInst<(outs GPRC:$rT), (ins rclass:$rA, optype:$val),
2692 [/* no pattern */]>;
2694 multiclass RotateQuadBytesImm
2696 def v16i8: ROTQMBYIVecInst<v16i8>;
2697 def v8i16: ROTQMBYIVecInst<v8i16>;
2698 def v4i32: ROTQMBYIVecInst<v4i32>;
2699 def v2i64: ROTQMBYIVecInst<v2i64>;
2701 def r128: ROTQMBYIRegInst<GPRC, rotNeg7imm, i32, uimm7>;
2702 def r64: ROTQMBYIRegInst<R64C, rotNeg7imm, i32, uimm7>;
2704 def r128_zext_r8: ROTQMBYIZExtInst<R8C, rotNeg7imm, uimm7>;
2705 def r128_zext_r16: ROTQMBYIZExtInst<R16C, rotNeg7imm, uimm7>;
2706 def r128_zext_r32: ROTQMBYIZExtInst<R32C, rotNeg7imm, uimm7>;
2707 def r128_zext_r64: ROTQMBYIZExtInst<R64C, rotNeg7imm, uimm7>;
2710 defm ROTQMBYI : RotateQuadBytesImm;
2712 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2713 // Rotate right and mask by bit count
2714 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2716 class ROTQMBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2717 RRForm<0b10110011100, OOL, IOL, "rotqmbybi\t$rT, $rA, $rB",
2718 RotateShift, pattern>;
2720 class ROTQMBYBIVecInst<ValueType vectype>:
2721 ROTQMBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2722 [/* no pattern, */]>;
2724 multiclass RotateMaskQuadByBitCount
2726 def v16i8: ROTQMBYBIVecInst<v16i8>;
2727 def v8i16: ROTQMBYBIVecInst<v8i16>;
2728 def v4i32: ROTQMBYBIVecInst<v4i32>;
2729 def v2i64: ROTQMBYBIVecInst<v2i64>;
2732 defm ROTQMBYBI: RotateMaskQuadByBitCount;
2734 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2735 // Rotate quad and mask by bits
2736 // Note that the rotate amount has to be negated
2737 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2739 class ROTQMBIInst<dag OOL, dag IOL, list<dag> pattern>:
2740 RRForm<0b10011011100, OOL, IOL, "rotqmbi\t$rT, $rA, $rB",
2741 RotateShift, pattern>;
2743 class ROTQMBIVecInst<ValueType vectype>:
2744 ROTQMBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2745 [/* no pattern */]>;
2747 class ROTQMBIRegInst<RegisterClass rclass>:
2748 ROTQMBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2749 [/* no pattern */]>;
2751 multiclass RotateMaskQuadByBits
2753 def v16i8: ROTQMBIVecInst<v16i8>;
2754 def v8i16: ROTQMBIVecInst<v8i16>;
2755 def v4i32: ROTQMBIVecInst<v4i32>;
2756 def v2i64: ROTQMBIVecInst<v2i64>;
2758 def r128: ROTQMBIRegInst<GPRC>;
2759 def r64: ROTQMBIRegInst<R64C>;
2762 defm ROTQMBI: RotateMaskQuadByBits;
2764 def : Pat<(srl GPRC:$rA, R32C:$rB),
2765 (ROTQMBIr128 GPRC:$rA,
2766 (SFIr32 R32C:$rB, 0))>;
2769 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2770 // Rotate quad and mask by bits, immediate
2771 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2773 class ROTQMBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2774 RI7Form<0b10011111100, OOL, IOL, "rotqmbii\t$rT, $rA, $val",
2775 RotateShift, pattern>;
2777 class ROTQMBIIVecInst<ValueType vectype>:
2778 ROTQMBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2779 [/* no pattern */]>;
2781 class ROTQMBIIRegInst<RegisterClass rclass>:
2782 ROTQMBIIInst<(outs rclass:$rT), (ins rclass:$rA, rotNeg7imm:$val),
2783 [/* no pattern */]>;
2785 multiclass RotateMaskQuadByBitsImm
2787 def v16i8: ROTQMBIIVecInst<v16i8>;
2788 def v8i16: ROTQMBIIVecInst<v8i16>;
2789 def v4i32: ROTQMBIIVecInst<v4i32>;
2790 def v2i64: ROTQMBIIVecInst<v2i64>;
2792 def r128: ROTQMBIIRegInst<GPRC>;
2793 def r64: ROTQMBIIRegInst<R64C>;
2796 defm ROTQMBII: RotateMaskQuadByBitsImm;
2798 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2799 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2802 RRForm<0b01111010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2803 "rotmah\t$rT, $rA, $rB", RotateShift,
2804 [/* see patterns below - $rB must be negated */]>;
2806 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), R32C:$rB),
2807 (ROTMAHv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2809 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), R16C:$rB),
2810 (ROTMAHv8i16 VECREG:$rA,
2811 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2813 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), R8C:$rB),
2814 (ROTMAHv8i16 VECREG:$rA,
2815 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2818 RRForm<0b01111010000, (outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2819 "rotmah\t$rT, $rA, $rB", RotateShift,
2820 [/* see patterns below - $rB must be negated */]>;
2822 def : Pat<(sra R16C:$rA, R32C:$rB),
2823 (ROTMAHr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2825 def : Pat<(sra R16C:$rA, R16C:$rB),
2826 (ROTMAHr16 R16C:$rA,
2827 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2829 def : Pat<(sra R16C:$rA, R8C:$rB),
2830 (ROTMAHr16 R16C:$rA,
2831 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2834 RRForm<0b01111110000, (outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2835 "rotmahi\t$rT, $rA, $val", RotateShift,
2836 [(set (v8i16 VECREG:$rT),
2837 (SPUvec_sra (v8i16 VECREG:$rA), (i32 uimm7:$val)))]>;
2839 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i16 uimm7:$val)),
2840 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (TO_IMM32 uimm7:$val))>;
2842 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i8 uimm7:$val)),
2843 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (TO_IMM32 uimm7:$val))>;
2846 RRForm<0b01111110000, (outs R16C:$rT), (ins R16C:$rA, rothNeg7imm_i16:$val),
2847 "rotmahi\t$rT, $rA, $val", RotateShift,
2848 [(set R16C:$rT, (sra R16C:$rA, (i16 uimm7:$val)))]>;
2850 def : Pat<(sra R16C:$rA, (i32 imm:$val)),
2851 (ROTMAHIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
2853 def : Pat<(sra R16C:$rA, (i8 imm:$val)),
2854 (ROTMAHIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
2857 RRForm<0b01011010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2858 "rotma\t$rT, $rA, $rB", RotateShift,
2859 [/* see patterns below - $rB must be negated */]>;
2861 def : Pat<(SPUvec_sra (v4i32 VECREG:$rA), R32C:$rB),
2862 (ROTMAv4i32 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2864 def : Pat<(SPUvec_sra (v4i32 VECREG:$rA), R16C:$rB),
2865 (ROTMAv4i32 VECREG:$rA,
2866 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2868 def : Pat<(SPUvec_sra (v4i32 VECREG:$rA), R8C:$rB),
2869 (ROTMAv4i32 VECREG:$rA,
2870 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2873 RRForm<0b01011010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2874 "rotma\t$rT, $rA, $rB", RotateShift,
2875 [/* see patterns below - $rB must be negated */]>;
2877 def : Pat<(sra R32C:$rA, R32C:$rB),
2878 (ROTMAr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
2880 def : Pat<(sra R32C:$rA, R16C:$rB),
2882 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2884 def : Pat<(sra R32C:$rA, R8C:$rB),
2886 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2888 class ROTMAIInst<dag OOL, dag IOL, list<dag> pattern>:
2889 RRForm<0b01011110000, OOL, IOL,
2890 "rotmai\t$rT, $rA, $val",
2891 RotateShift, pattern>;
2893 class ROTMAIVecInst<ValueType vectype, Operand intop, ValueType inttype>:
2894 ROTMAIInst<(outs VECREG:$rT), (ins VECREG:$rA, intop:$val),
2895 [(set (vectype VECREG:$rT),
2896 (SPUvec_sra VECREG:$rA, (inttype uimm7:$val)))]>;
2898 class ROTMAIRegInst<RegisterClass rclass, Operand intop, ValueType inttype>:
2899 ROTMAIInst<(outs rclass:$rT), (ins rclass:$rA, intop:$val),
2900 [(set rclass:$rT, (sra rclass:$rA, (inttype uimm7:$val)))]>;
2902 multiclass RotateMaskAlgebraicImm {
2903 def v2i64_i32 : ROTMAIVecInst<v2i64, rotNeg7imm, i32>;
2904 def v4i32_i32 : ROTMAIVecInst<v4i32, rotNeg7imm, i32>;
2905 def r64_i32 : ROTMAIRegInst<R64C, rotNeg7imm, i32>;
2906 def r32_i32 : ROTMAIRegInst<R32C, rotNeg7imm, i32>;
2909 defm ROTMAI : RotateMaskAlgebraicImm;
2911 //===----------------------------------------------------------------------===//
2912 // Branch and conditionals:
2913 //===----------------------------------------------------------------------===//
2915 let isTerminator = 1, isBarrier = 1 in {
2916 // Halt If Equal (r32 preferred slot only, no vector form)
2918 RRForm_3<0b00011011110, (outs), (ins R32C:$rA, R32C:$rB),
2919 "heq\t$rA, $rB", BranchResolv,
2920 [/* no pattern to match */]>;
2923 RI10Form_2<0b11111110, (outs), (ins R32C:$rA, s10imm:$val),
2924 "heqi\t$rA, $val", BranchResolv,
2925 [/* no pattern to match */]>;
2927 // HGT/HGTI: These instructions use signed arithmetic for the comparison,
2928 // contrasting with HLGT/HLGTI, which use unsigned comparison:
2930 RRForm_3<0b00011010010, (outs), (ins R32C:$rA, R32C:$rB),
2931 "hgt\t$rA, $rB", BranchResolv,
2932 [/* no pattern to match */]>;
2935 RI10Form_2<0b11110010, (outs), (ins R32C:$rA, s10imm:$val),
2936 "hgti\t$rA, $val", BranchResolv,
2937 [/* no pattern to match */]>;
2940 RRForm_3<0b00011011010, (outs), (ins R32C:$rA, R32C:$rB),
2941 "hlgt\t$rA, $rB", BranchResolv,
2942 [/* no pattern to match */]>;
2945 RI10Form_2<0b11111010, (outs), (ins R32C:$rA, s10imm:$val),
2946 "hlgti\t$rA, $val", BranchResolv,
2947 [/* no pattern to match */]>;
2950 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2951 // Comparison operators for i8, i16 and i32:
2952 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2954 class CEQBInst<dag OOL, dag IOL, list<dag> pattern> :
2955 RRForm<0b00001011110, OOL, IOL, "ceqb\t$rT, $rA, $rB",
2958 multiclass CmpEqualByte
2961 CEQBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2962 [(set (v16i8 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
2963 (v8i16 VECREG:$rB)))]>;
2966 CEQBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
2967 [(set R8C:$rT, (seteq R8C:$rA, R8C:$rB))]>;
2970 class CEQBIInst<dag OOL, dag IOL, list<dag> pattern> :
2971 RI10Form<0b01111110, OOL, IOL, "ceqbi\t$rT, $rA, $val",
2974 multiclass CmpEqualByteImm
2977 CEQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
2978 [(set (v16i8 VECREG:$rT), (seteq (v16i8 VECREG:$rA),
2979 v16i8SExt8Imm:$val))]>;
2981 CEQBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
2982 [(set R8C:$rT, (seteq R8C:$rA, immSExt8:$val))]>;
2985 class CEQHInst<dag OOL, dag IOL, list<dag> pattern> :
2986 RRForm<0b00010011110, OOL, IOL, "ceqh\t$rT, $rA, $rB",
2989 multiclass CmpEqualHalfword
2991 def v8i16 : CEQHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2992 [(set (v8i16 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
2993 (v8i16 VECREG:$rB)))]>;
2995 def r16 : CEQHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
2996 [(set R16C:$rT, (seteq R16C:$rA, R16C:$rB))]>;
2999 class CEQHIInst<dag OOL, dag IOL, list<dag> pattern> :
3000 RI10Form<0b10111110, OOL, IOL, "ceqhi\t$rT, $rA, $val",
3003 multiclass CmpEqualHalfwordImm
3005 def v8i16 : CEQHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3006 [(set (v8i16 VECREG:$rT),
3007 (seteq (v8i16 VECREG:$rA),
3008 (v8i16 v8i16SExt10Imm:$val)))]>;
3009 def r16 : CEQHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3010 [(set R16C:$rT, (seteq R16C:$rA, i16ImmSExt10:$val))]>;
3013 class CEQInst<dag OOL, dag IOL, list<dag> pattern> :
3014 RRForm<0b00000011110, OOL, IOL, "ceq\t$rT, $rA, $rB",
3017 multiclass CmpEqualWord
3019 def v4i32 : CEQInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3020 [(set (v4i32 VECREG:$rT),
3021 (seteq (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3023 def r32 : CEQInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3024 [(set R32C:$rT, (seteq R32C:$rA, R32C:$rB))]>;
3027 class CEQIInst<dag OOL, dag IOL, list<dag> pattern> :
3028 RI10Form<0b00111110, OOL, IOL, "ceqi\t$rT, $rA, $val",
3031 multiclass CmpEqualWordImm
3033 def v4i32 : CEQIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3034 [(set (v4i32 VECREG:$rT),
3035 (seteq (v4i32 VECREG:$rA),
3036 (v4i32 v4i32SExt16Imm:$val)))]>;
3038 def r32: CEQIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3039 [(set R32C:$rT, (seteq R32C:$rA, i32ImmSExt10:$val))]>;
3042 class CGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3043 RRForm<0b00001010010, OOL, IOL, "cgtb\t$rT, $rA, $rB",
3046 multiclass CmpGtrByte
3049 CGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3050 [(set (v16i8 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3051 (v8i16 VECREG:$rB)))]>;
3054 CGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3055 [(set R8C:$rT, (setgt R8C:$rA, R8C:$rB))]>;
3058 class CGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3059 RI10Form<0b01110010, OOL, IOL, "cgtbi\t$rT, $rA, $val",
3062 multiclass CmpGtrByteImm
3065 CGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3066 [(set (v16i8 VECREG:$rT), (setgt (v16i8 VECREG:$rA),
3067 v16i8SExt8Imm:$val))]>;
3069 CGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3070 [(set R8C:$rT, (setgt R8C:$rA, immSExt8:$val))]>;
3073 class CGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3074 RRForm<0b00010010010, OOL, IOL, "cgth\t$rT, $rA, $rB",
3077 multiclass CmpGtrHalfword
3079 def v8i16 : CGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3080 [(set (v8i16 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3081 (v8i16 VECREG:$rB)))]>;
3083 def r16 : CGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3084 [(set R16C:$rT, (setgt R16C:$rA, R16C:$rB))]>;
3087 class CGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3088 RI10Form<0b10110010, OOL, IOL, "cgthi\t$rT, $rA, $val",
3091 multiclass CmpGtrHalfwordImm
3093 def v8i16 : CGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3094 [(set (v8i16 VECREG:$rT),
3095 (setgt (v8i16 VECREG:$rA),
3096 (v8i16 v8i16SExt10Imm:$val)))]>;
3097 def r16 : CGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3098 [(set R16C:$rT, (setgt R16C:$rA, i16ImmSExt10:$val))]>;
3101 class CGTInst<dag OOL, dag IOL, list<dag> pattern> :
3102 RRForm<0b00000010010, OOL, IOL, "cgt\t$rT, $rA, $rB",
3105 multiclass CmpGtrWord
3107 def v4i32 : CGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3108 [(set (v4i32 VECREG:$rT),
3109 (setgt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3111 def r32 : CGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3112 [(set R32C:$rT, (setgt R32C:$rA, R32C:$rB))]>;
3115 class CGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3116 RI10Form<0b00110010, OOL, IOL, "cgti\t$rT, $rA, $val",
3119 multiclass CmpGtrWordImm
3121 def v4i32 : CGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3122 [(set (v4i32 VECREG:$rT),
3123 (setgt (v4i32 VECREG:$rA),
3124 (v4i32 v4i32SExt16Imm:$val)))]>;
3126 def r32: CGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3127 [(set R32C:$rT, (setgt R32C:$rA, i32ImmSExt10:$val))]>;
3129 // CGTIv4f32, CGTIf32: These are used in the f32 fdiv instruction sequence:
3130 def v4f32: CGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3131 [(set (v4i32 VECREG:$rT),
3132 (setgt (v4i32 (bitconvert (v4f32 VECREG:$rA))),
3133 (v4i32 v4i32SExt16Imm:$val)))]>;
3135 def f32: CGTIInst<(outs R32C:$rT), (ins R32FP:$rA, s10imm_i32:$val),
3136 [/* no pattern */]>;
3139 class CLGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3140 RRForm<0b00001011010, OOL, IOL, "clgtb\t$rT, $rA, $rB",
3143 multiclass CmpLGtrByte
3146 CLGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3147 [(set (v16i8 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3148 (v8i16 VECREG:$rB)))]>;
3151 CLGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3152 [(set R8C:$rT, (setugt R8C:$rA, R8C:$rB))]>;
3155 class CLGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3156 RI10Form<0b01111010, OOL, IOL, "clgtbi\t$rT, $rA, $val",
3159 multiclass CmpLGtrByteImm
3162 CLGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3163 [(set (v16i8 VECREG:$rT), (setugt (v16i8 VECREG:$rA),
3164 v16i8SExt8Imm:$val))]>;
3166 CLGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3167 [(set R8C:$rT, (setugt R8C:$rA, immSExt8:$val))]>;
3170 class CLGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3171 RRForm<0b00010011010, OOL, IOL, "clgth\t$rT, $rA, $rB",
3174 multiclass CmpLGtrHalfword
3176 def v8i16 : CLGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3177 [(set (v8i16 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3178 (v8i16 VECREG:$rB)))]>;
3180 def r16 : CLGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3181 [(set R16C:$rT, (setugt R16C:$rA, R16C:$rB))]>;
3184 class CLGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3185 RI10Form<0b10111010, OOL, IOL, "clgthi\t$rT, $rA, $val",
3188 multiclass CmpLGtrHalfwordImm
3190 def v8i16 : CLGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3191 [(set (v8i16 VECREG:$rT),
3192 (setugt (v8i16 VECREG:$rA),
3193 (v8i16 v8i16SExt10Imm:$val)))]>;
3194 def r16 : CLGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3195 [(set R16C:$rT, (setugt R16C:$rA, i16ImmSExt10:$val))]>;
3198 class CLGTInst<dag OOL, dag IOL, list<dag> pattern> :
3199 RRForm<0b00000011010, OOL, IOL, "clgt\t$rT, $rA, $rB",
3202 multiclass CmpLGtrWord
3204 def v4i32 : CLGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3205 [(set (v4i32 VECREG:$rT),
3206 (setugt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3208 def r32 : CLGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3209 [(set R32C:$rT, (setugt R32C:$rA, R32C:$rB))]>;
3212 class CLGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3213 RI10Form<0b00111010, OOL, IOL, "clgti\t$rT, $rA, $val",
3216 multiclass CmpLGtrWordImm
3218 def v4i32 : CLGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3219 [(set (v4i32 VECREG:$rT),
3220 (setugt (v4i32 VECREG:$rA),
3221 (v4i32 v4i32SExt16Imm:$val)))]>;
3223 def r32: CLGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3224 [(set R32C:$rT, (setugt R32C:$rA, i32ImmSExt10:$val))]>;
3227 defm CEQB : CmpEqualByte;
3228 defm CEQBI : CmpEqualByteImm;
3229 defm CEQH : CmpEqualHalfword;
3230 defm CEQHI : CmpEqualHalfwordImm;
3231 defm CEQ : CmpEqualWord;
3232 defm CEQI : CmpEqualWordImm;
3233 defm CGTB : CmpGtrByte;
3234 defm CGTBI : CmpGtrByteImm;
3235 defm CGTH : CmpGtrHalfword;
3236 defm CGTHI : CmpGtrHalfwordImm;
3237 defm CGT : CmpGtrWord;
3238 defm CGTI : CmpGtrWordImm;
3239 defm CLGTB : CmpLGtrByte;
3240 defm CLGTBI : CmpLGtrByteImm;
3241 defm CLGTH : CmpLGtrHalfword;
3242 defm CLGTHI : CmpLGtrHalfwordImm;
3243 defm CLGT : CmpLGtrWord;
3244 defm CLGTI : CmpLGtrWordImm;
3246 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3247 // For SETCC primitives not supported above (setlt, setle, setge, etc.)
3248 // define a pattern to generate the right code, as a binary operator
3249 // (in a manner of speaking.)
3252 // 1. This only matches the setcc set of conditionals. Special pattern
3253 // matching is used for select conditionals.
3255 // 2. The "DAG" versions of these classes is almost exclusively used for
3256 // i64 comparisons. See the tblgen fundamentals documentation for what
3257 // ".ResultInstrs[0]" means; see TargetSelectionDAG.td and the Pattern
3258 // class for where ResultInstrs originates.
3259 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3261 class SETCCNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3262 SPUInstr xorinst, SPUInstr cmpare>:
3263 Pat<(cond rclass:$rA, rclass:$rB),
3264 (xorinst (cmpare rclass:$rA, rclass:$rB), (inttype -1))>;
3266 class SETCCNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3267 PatLeaf immpred, SPUInstr xorinst, SPUInstr cmpare>:
3268 Pat<(cond rclass:$rA, (inttype immpred:$imm)),
3269 (xorinst (cmpare rclass:$rA, (inttype immpred:$imm)), (inttype -1))>;
3271 def : SETCCNegCondReg<setne, R8C, i8, XORBIr8, CEQBr8>;
3272 def : SETCCNegCondImm<setne, R8C, i8, immSExt8, XORBIr8, CEQBIr8>;
3274 def : SETCCNegCondReg<setne, R16C, i16, XORHIr16, CEQHr16>;
3275 def : SETCCNegCondImm<setne, R16C, i16, i16ImmSExt10, XORHIr16, CEQHIr16>;
3277 def : SETCCNegCondReg<setne, R32C, i32, XORIr32, CEQr32>;
3278 def : SETCCNegCondImm<setne, R32C, i32, i32ImmSExt10, XORIr32, CEQIr32>;
3280 class SETCCBinOpReg<PatFrag cond, RegisterClass rclass,
3281 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3282 Pat<(cond rclass:$rA, rclass:$rB),
3283 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3284 (cmpOp2 rclass:$rA, rclass:$rB))>;
3286 class SETCCBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3288 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3289 Pat<(cond rclass:$rA, (immtype immpred:$imm)),
3290 (binop (cmpOp1 rclass:$rA, (immtype immpred:$imm)),
3291 (cmpOp2 rclass:$rA, (immtype immpred:$imm)))>;
3293 def : SETCCBinOpReg<setge, R8C, ORr8, CGTBr8, CEQBr8>;
3294 def : SETCCBinOpImm<setge, R8C, immSExt8, i8, ORr8, CGTBIr8, CEQBIr8>;
3295 def : SETCCBinOpReg<setlt, R8C, NORr8, CGTBr8, CEQBr8>;
3296 def : SETCCBinOpImm<setlt, R8C, immSExt8, i8, NORr8, CGTBIr8, CEQBIr8>;
3297 def : Pat<(setle R8C:$rA, R8C:$rB),
3298 (XORBIr8 (CGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3299 def : Pat<(setle R8C:$rA, immU8:$imm),
3300 (XORBIr8 (CGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3302 def : SETCCBinOpReg<setge, R16C, ORr16, CGTHr16, CEQHr16>;
3303 def : SETCCBinOpImm<setge, R16C, i16ImmSExt10, i16,
3304 ORr16, CGTHIr16, CEQHIr16>;
3305 def : SETCCBinOpReg<setlt, R16C, NORr16, CGTHr16, CEQHr16>;
3306 def : SETCCBinOpImm<setlt, R16C, i16ImmSExt10, i16, NORr16, CGTHIr16, CEQHIr16>;
3307 def : Pat<(setle R16C:$rA, R16C:$rB),
3308 (XORHIr16 (CGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3309 def : Pat<(setle R16C:$rA, i16ImmSExt10:$imm),
3310 (XORHIr16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3312 def : SETCCBinOpReg<setge, R32C, ORr32, CGTr32, CEQr32>;
3313 def : SETCCBinOpImm<setge, R32C, i32ImmSExt10, i32,
3314 ORr32, CGTIr32, CEQIr32>;
3315 def : SETCCBinOpReg<setlt, R32C, NORr32, CGTr32, CEQr32>;
3316 def : SETCCBinOpImm<setlt, R32C, i32ImmSExt10, i32, NORr32, CGTIr32, CEQIr32>;
3317 def : Pat<(setle R32C:$rA, R32C:$rB),
3318 (XORIr32 (CGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3319 def : Pat<(setle R32C:$rA, i32ImmSExt10:$imm),
3320 (XORIr32 (CGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3322 def : SETCCBinOpReg<setuge, R8C, ORr8, CLGTBr8, CEQBr8>;
3323 def : SETCCBinOpImm<setuge, R8C, immSExt8, i8, ORr8, CLGTBIr8, CEQBIr8>;
3324 def : SETCCBinOpReg<setult, R8C, NORr8, CLGTBr8, CEQBr8>;
3325 def : SETCCBinOpImm<setult, R8C, immSExt8, i8, NORr8, CLGTBIr8, CEQBIr8>;
3326 def : Pat<(setule R8C:$rA, R8C:$rB),
3327 (XORBIr8 (CLGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3328 def : Pat<(setule R8C:$rA, immU8:$imm),
3329 (XORBIr8 (CLGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3331 def : SETCCBinOpReg<setuge, R16C, ORr16, CLGTHr16, CEQHr16>;
3332 def : SETCCBinOpImm<setuge, R16C, i16ImmSExt10, i16,
3333 ORr16, CLGTHIr16, CEQHIr16>;
3334 def : SETCCBinOpReg<setult, R16C, NORr16, CLGTHr16, CEQHr16>;
3335 def : SETCCBinOpImm<setult, R16C, i16ImmSExt10, i16, NORr16,
3336 CLGTHIr16, CEQHIr16>;
3337 def : Pat<(setule R16C:$rA, R16C:$rB),
3338 (XORHIr16 (CLGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3339 def : Pat<(setule R16C:$rA, i16ImmSExt10:$imm),
3340 (XORHIr16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3342 def : SETCCBinOpReg<setuge, R32C, ORr32, CLGTr32, CEQr32>;
3343 def : SETCCBinOpImm<setuge, R32C, i32ImmSExt10, i32,
3344 ORr32, CLGTIr32, CEQIr32>;
3345 def : SETCCBinOpReg<setult, R32C, NORr32, CLGTr32, CEQr32>;
3346 def : SETCCBinOpImm<setult, R32C, i32ImmSExt10, i32, NORr32, CLGTIr32, CEQIr32>;
3347 def : Pat<(setule R32C:$rA, R32C:$rB),
3348 (XORIr32 (CLGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3349 def : Pat<(setule R32C:$rA, i32ImmSExt10:$imm),
3350 (XORIr32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3352 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3353 // select conditional patterns:
3354 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3356 class SELECTNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3357 SPUInstr selinstr, SPUInstr cmpare>:
3358 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3359 rclass:$rTrue, rclass:$rFalse),
3360 (selinstr rclass:$rTrue, rclass:$rFalse,
3361 (cmpare rclass:$rA, rclass:$rB))>;
3363 class SELECTNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3364 PatLeaf immpred, SPUInstr selinstr, SPUInstr cmpare>:
3365 Pat<(select (inttype (cond rclass:$rA, immpred:$imm)),
3366 rclass:$rTrue, rclass:$rFalse),
3367 (selinstr rclass:$rTrue, rclass:$rFalse,
3368 (cmpare rclass:$rA, immpred:$imm))>;
3370 def : SELECTNegCondReg<setne, R8C, i8, SELBr8, CEQBr8>;
3371 def : SELECTNegCondImm<setne, R8C, i8, immSExt8, SELBr8, CEQBIr8>;
3372 def : SELECTNegCondReg<setle, R8C, i8, SELBr8, CGTBr8>;
3373 def : SELECTNegCondImm<setle, R8C, i8, immSExt8, SELBr8, CGTBr8>;
3374 def : SELECTNegCondReg<setule, R8C, i8, SELBr8, CLGTBr8>;
3375 def : SELECTNegCondImm<setule, R8C, i8, immU8, SELBr8, CLGTBIr8>;
3377 def : SELECTNegCondReg<setne, R16C, i16, SELBr16, CEQHr16>;
3378 def : SELECTNegCondImm<setne, R16C, i16, i16ImmSExt10, SELBr16, CEQHIr16>;
3379 def : SELECTNegCondReg<setle, R16C, i16, SELBr16, CGTHr16>;
3380 def : SELECTNegCondImm<setle, R16C, i16, i16ImmSExt10, SELBr16, CGTHIr16>;
3381 def : SELECTNegCondReg<setule, R16C, i16, SELBr16, CLGTHr16>;
3382 def : SELECTNegCondImm<setule, R16C, i16, i16ImmSExt10, SELBr16, CLGTHIr16>;
3384 def : SELECTNegCondReg<setne, R32C, i32, SELBr32, CEQr32>;
3385 def : SELECTNegCondImm<setne, R32C, i32, i32ImmSExt10, SELBr32, CEQIr32>;
3386 def : SELECTNegCondReg<setle, R32C, i32, SELBr32, CGTr32>;
3387 def : SELECTNegCondImm<setle, R32C, i32, i32ImmSExt10, SELBr32, CGTIr32>;
3388 def : SELECTNegCondReg<setule, R32C, i32, SELBr32, CLGTr32>;
3389 def : SELECTNegCondImm<setule, R32C, i32, i32ImmSExt10, SELBr32, CLGTIr32>;
3391 class SELECTBinOpReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3392 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3394 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3395 rclass:$rTrue, rclass:$rFalse),
3396 (selinstr rclass:$rFalse, rclass:$rTrue,
3397 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3398 (cmpOp2 rclass:$rA, rclass:$rB)))>;
3400 class SELECTBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3402 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3404 Pat<(select (inttype (cond rclass:$rA, (inttype immpred:$imm))),
3405 rclass:$rTrue, rclass:$rFalse),
3406 (selinstr rclass:$rFalse, rclass:$rTrue,
3407 (binop (cmpOp1 rclass:$rA, (inttype immpred:$imm)),
3408 (cmpOp2 rclass:$rA, (inttype immpred:$imm))))>;
3410 def : SELECTBinOpReg<setge, R8C, i8, SELBr8, ORr8, CGTBr8, CEQBr8>;
3411 def : SELECTBinOpImm<setge, R8C, immSExt8, i8,
3412 SELBr8, ORr8, CGTBIr8, CEQBIr8>;
3414 def : SELECTBinOpReg<setge, R16C, i16, SELBr16, ORr16, CGTHr16, CEQHr16>;
3415 def : SELECTBinOpImm<setge, R16C, i16ImmSExt10, i16,
3416 SELBr16, ORr16, CGTHIr16, CEQHIr16>;
3418 def : SELECTBinOpReg<setge, R32C, i32, SELBr32, ORr32, CGTr32, CEQr32>;
3419 def : SELECTBinOpImm<setge, R32C, i32ImmSExt10, i32,
3420 SELBr32, ORr32, CGTIr32, CEQIr32>;
3422 def : SELECTBinOpReg<setuge, R8C, i8, SELBr8, ORr8, CLGTBr8, CEQBr8>;
3423 def : SELECTBinOpImm<setuge, R8C, immSExt8, i8,
3424 SELBr8, ORr8, CLGTBIr8, CEQBIr8>;
3426 def : SELECTBinOpReg<setuge, R16C, i16, SELBr16, ORr16, CLGTHr16, CEQHr16>;
3427 def : SELECTBinOpImm<setuge, R16C, i16ImmUns10, i16,
3428 SELBr16, ORr16, CLGTHIr16, CEQHIr16>;
3430 def : SELECTBinOpReg<setuge, R32C, i32, SELBr32, ORr32, CLGTr32, CEQr32>;
3431 def : SELECTBinOpImm<setuge, R32C, i32ImmUns10, i32,
3432 SELBr32, ORr32, CLGTIr32, CEQIr32>;
3434 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3437 // All calls clobber the non-callee-saved registers:
3438 Defs = [R0, R1, R2, R3, R4, R5, R6, R7, R8, R9,
3439 R10,R11,R12,R13,R14,R15,R16,R17,R18,R19,
3440 R20,R21,R22,R23,R24,R25,R26,R27,R28,R29,
3441 R30,R31,R32,R33,R34,R35,R36,R37,R38,R39,
3442 R40,R41,R42,R43,R44,R45,R46,R47,R48,R49,
3443 R50,R51,R52,R53,R54,R55,R56,R57,R58,R59,
3444 R60,R61,R62,R63,R64,R65,R66,R67,R68,R69,
3445 R70,R71,R72,R73,R74,R75,R76,R77,R78,R79],
3446 // All of these instructions use $lr (aka $0)
3448 // Branch relative and set link: Used if we actually know that the target
3449 // is within [-32768, 32767] bytes of the target
3451 BranchSetLink<0b011001100, (outs), (ins relcalltarget:$func, variable_ops),
3452 "brsl\t$$lr, $func",
3453 [(SPUcall (SPUpcrel tglobaladdr:$func, 0))]>;
3455 // Branch absolute and set link: Used if we actually know that the target
3456 // is an absolute address
3458 BranchSetLink<0b011001100, (outs), (ins calltarget:$func, variable_ops),
3459 "brasl\t$$lr, $func",
3460 [(SPUcall (SPUaform tglobaladdr:$func, 0))]>;
3462 // Branch indirect and set link if external data. These instructions are not
3463 // actually generated, matched by an intrinsic:
3464 def BISLED_00: BISLEDForm<0b11, "bisled\t$$lr, $func", [/* empty pattern */]>;
3465 def BISLED_E0: BISLEDForm<0b10, "bisled\t$$lr, $func", [/* empty pattern */]>;
3466 def BISLED_0D: BISLEDForm<0b01, "bisled\t$$lr, $func", [/* empty pattern */]>;
3467 def BISLED_ED: BISLEDForm<0b00, "bisled\t$$lr, $func", [/* empty pattern */]>;
3469 // Branch indirect and set link. This is the "X-form" address version of a
3472 BIForm<0b10010101100, "bisl\t$$lr, $func", [(SPUcall R32C:$func)]>;
3475 // Support calls to external symbols:
3476 def : Pat<(SPUcall (SPUpcrel texternalsym:$func, 0)),
3477 (BRSL texternalsym:$func)>;
3479 def : Pat<(SPUcall (SPUaform texternalsym:$func, 0)),
3480 (BRASL texternalsym:$func)>;
3482 // Unconditional branches:
3483 let isBranch = 1, isTerminator = 1, hasCtrlDep = 1 in {
3484 let isBarrier = 1 in {
3486 UncondBranch<0b001001100, (outs), (ins brtarget:$dest),
3490 // Unconditional, absolute address branch
3492 UncondBranch<0b001100000, (outs), (ins brtarget:$dest),
3494 [/* no pattern */]>;
3498 BIForm<0b00010101100, "bi\t$func", [(brind R32C:$func)]>;
3501 // Conditional branches:
3502 class BRNZInst<dag IOL, list<dag> pattern>:
3503 RI16Form<0b010000100, (outs), IOL, "brnz\t$rCond,$dest",
3504 BranchResolv, pattern>;
3506 class BRNZRegInst<RegisterClass rclass>:
3507 BRNZInst<(ins rclass:$rCond, brtarget:$dest),
3508 [(brcond rclass:$rCond, bb:$dest)]>;
3510 class BRNZVecInst<ValueType vectype>:
3511 BRNZInst<(ins VECREG:$rCond, brtarget:$dest),
3512 [(brcond (vectype VECREG:$rCond), bb:$dest)]>;
3514 multiclass BranchNotZero {
3515 def v4i32 : BRNZVecInst<v4i32>;
3516 def r32 : BRNZRegInst<R32C>;
3519 defm BRNZ : BranchNotZero;
3521 class BRZInst<dag IOL, list<dag> pattern>:
3522 RI16Form<0b000000100, (outs), IOL, "brz\t$rT,$dest",
3523 BranchResolv, pattern>;
3525 class BRZRegInst<RegisterClass rclass>:
3526 BRZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3528 class BRZVecInst<ValueType vectype>:
3529 BRZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3531 multiclass BranchZero {
3532 def v4i32: BRZVecInst<v4i32>;
3533 def r32: BRZRegInst<R32C>;
3536 defm BRZ: BranchZero;
3538 // Note: LLVM doesn't do branch conditional, indirect. Otherwise these would
3541 class BINZInst<dag IOL, list<dag> pattern>:
3542 BICondForm<0b10010100100, (outs), IOL, "binz\t$rA, $dest", pattern>;
3544 class BINZRegInst<RegisterClass rclass>:
3545 BINZInst<(ins rclass:$rA, brtarget:$dest),
3546 [(brcond rclass:$rA, R32C:$dest)]>;
3548 class BINZVecInst<ValueType vectype>:
3549 BINZInst<(ins VECREG:$rA, R32C:$dest),
3550 [(brcond (vectype VECREG:$rA), R32C:$dest)]>;
3552 multiclass BranchNotZeroIndirect {
3553 def v4i32: BINZVecInst<v4i32>;
3554 def r32: BINZRegInst<R32C>;
3557 defm BINZ: BranchNotZeroIndirect;
3559 class BIZInst<dag IOL, list<dag> pattern>:
3560 BICondForm<0b00010100100, (outs), IOL, "biz\t$rA, $func", pattern>;
3562 class BIZRegInst<RegisterClass rclass>:
3563 BIZInst<(ins rclass:$rA, R32C:$func), [/* no pattern */]>;
3565 class BIZVecInst<ValueType vectype>:
3566 BIZInst<(ins VECREG:$rA, R32C:$func), [/* no pattern */]>;
3568 multiclass BranchZeroIndirect {
3569 def v4i32: BIZVecInst<v4i32>;
3570 def r32: BIZRegInst<R32C>;
3573 defm BIZ: BranchZeroIndirect;
3576 class BRHNZInst<dag IOL, list<dag> pattern>:
3577 RI16Form<0b011000100, (outs), IOL, "brhnz\t$rCond,$dest", BranchResolv,
3580 class BRHNZRegInst<RegisterClass rclass>:
3581 BRHNZInst<(ins rclass:$rCond, brtarget:$dest),
3582 [(brcond rclass:$rCond, bb:$dest)]>;
3584 class BRHNZVecInst<ValueType vectype>:
3585 BRHNZInst<(ins VECREG:$rCond, brtarget:$dest), [/* no pattern */]>;
3587 multiclass BranchNotZeroHalfword {
3588 def v8i16: BRHNZVecInst<v8i16>;
3589 def r16: BRHNZRegInst<R16C>;
3592 defm BRHNZ: BranchNotZeroHalfword;
3594 class BRHZInst<dag IOL, list<dag> pattern>:
3595 RI16Form<0b001000100, (outs), IOL, "brhz\t$rT,$dest", BranchResolv,
3598 class BRHZRegInst<RegisterClass rclass>:
3599 BRHZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3601 class BRHZVecInst<ValueType vectype>:
3602 BRHZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3604 multiclass BranchZeroHalfword {
3605 def v8i16: BRHZVecInst<v8i16>;
3606 def r16: BRHZRegInst<R16C>;
3609 defm BRHZ: BranchZeroHalfword;
3612 //===----------------------------------------------------------------------===//
3613 // setcc and brcond patterns:
3614 //===----------------------------------------------------------------------===//
3616 def : Pat<(brcond (i16 (seteq R16C:$rA, 0)), bb:$dest),
3617 (BRHZr16 R16C:$rA, bb:$dest)>;
3618 def : Pat<(brcond (i16 (setne R16C:$rA, 0)), bb:$dest),
3619 (BRHNZr16 R16C:$rA, bb:$dest)>;
3621 def : Pat<(brcond (i32 (seteq R32C:$rA, 0)), bb:$dest),
3622 (BRZr32 R32C:$rA, bb:$dest)>;
3623 def : Pat<(brcond (i32 (setne R32C:$rA, 0)), bb:$dest),
3624 (BRNZr32 R32C:$rA, bb:$dest)>;
3626 multiclass BranchCondEQ<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3628 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3629 (brinst16 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3631 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3632 (brinst16 (CEQHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3634 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3635 (brinst32 (CEQIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3637 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3638 (brinst32 (CEQr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3641 defm BRCONDeq : BranchCondEQ<seteq, BRHNZr16, BRNZr32>;
3642 defm BRCONDne : BranchCondEQ<setne, BRHZr16, BRZr32>;
3644 multiclass BranchCondLGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3646 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3647 (brinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3649 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3650 (brinst16 (CLGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3652 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3653 (brinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3655 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3656 (brinst32 (CLGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3659 defm BRCONDugt : BranchCondLGT<setugt, BRHNZr16, BRNZr32>;
3660 defm BRCONDule : BranchCondLGT<setule, BRHZr16, BRZr32>;
3662 multiclass BranchCondLGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3663 SPUInstr orinst32, SPUInstr brinst32>
3665 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3666 (brinst16 (orinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3667 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3670 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3671 (brinst16 (orinst16 (CLGTHr16 R16C:$rA, R16:$rB),
3672 (CEQHr16 R16C:$rA, R16:$rB)),
3675 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3676 (brinst32 (orinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val),
3677 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3680 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3681 (brinst32 (orinst32 (CLGTr32 R32C:$rA, R32C:$rB),
3682 (CEQr32 R32C:$rA, R32C:$rB)),
3686 defm BRCONDuge : BranchCondLGTEQ<setuge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3687 defm BRCONDult : BranchCondLGTEQ<setult, ORr16, BRHZr16, ORr32, BRZr32>;
3689 multiclass BranchCondGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3691 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3692 (brinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3694 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3695 (brinst16 (CGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3697 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3698 (brinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3700 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3701 (brinst32 (CGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3704 defm BRCONDgt : BranchCondGT<setgt, BRHNZr16, BRNZr32>;
3705 defm BRCONDle : BranchCondGT<setle, BRHZr16, BRZr32>;
3707 multiclass BranchCondGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3708 SPUInstr orinst32, SPUInstr brinst32>
3710 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3711 (brinst16 (orinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3712 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3715 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3716 (brinst16 (orinst16 (CGTHr16 R16C:$rA, R16:$rB),
3717 (CEQHr16 R16C:$rA, R16:$rB)),
3720 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3721 (brinst32 (orinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val),
3722 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3725 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3726 (brinst32 (orinst32 (CGTr32 R32C:$rA, R32C:$rB),
3727 (CEQr32 R32C:$rA, R32C:$rB)),
3731 defm BRCONDge : BranchCondGTEQ<setge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3732 defm BRCONDlt : BranchCondGTEQ<setlt, ORr16, BRHZr16, ORr32, BRZr32>;
3734 let isTerminator = 1, isBarrier = 1 in {
3735 let isReturn = 1 in {
3737 RETForm<"bi\t$$lr", [(retflag)]>;
3741 //===----------------------------------------------------------------------===//
3742 // Single precision floating point instructions
3743 //===----------------------------------------------------------------------===//
3745 class FAInst<dag OOL, dag IOL, list<dag> pattern>:
3746 RRForm<0b01011000100, OOL, IOL, "fa\t$rT, $rA, $rB",
3749 class FAVecInst<ValueType vectype>:
3750 FAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3751 [(set (vectype VECREG:$rT),
3752 (fadd (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3756 def v4f32: FAVecInst<v4f32>;
3757 def f32: FAInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3758 [(set R32FP:$rT, (fadd R32FP:$rA, R32FP:$rB))]>;
3763 class FSInst<dag OOL, dag IOL, list<dag> pattern>:
3764 RRForm<0b01011000100, OOL, IOL, "fs\t$rT, $rA, $rB",
3767 class FSVecInst<ValueType vectype>:
3768 FSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3769 [(set (vectype VECREG:$rT),
3770 (fsub (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3774 def v4f32: FSVecInst<v4f32>;
3775 def f32: FSInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3776 [(set R32FP:$rT, (fsub R32FP:$rA, R32FP:$rB))]>;
3781 class FMInst<dag OOL, dag IOL, list<dag> pattern>:
3782 RRForm<0b01100011010, OOL, IOL,
3783 "fm\t$rT, $rA, $rB", SPrecFP,
3786 class FMVecInst<ValueType type>:
3787 FMInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3788 [(set (type VECREG:$rT),
3789 (fmul (type VECREG:$rA), (type VECREG:$rB)))]>;
3793 def v4f32: FMVecInst<v4f32>;
3794 def f32: FMInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3795 [(set R32FP:$rT, (fmul R32FP:$rA, R32FP:$rB))]>;
3800 // Floating point multiply and add
3801 // e.g. d = c + (a * b)
3803 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
3804 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
3805 [(set (v4f32 VECREG:$rT),
3806 (fadd (v4f32 VECREG:$rC),
3807 (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB))))]>;
3810 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
3811 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
3812 [(set R32FP:$rT, (fadd R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
3814 // FP multiply and subtract
3815 // Subtracts value in rC from product
3818 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
3819 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
3820 [(set (v4f32 VECREG:$rT),
3821 (fsub (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB)),
3822 (v4f32 VECREG:$rC)))]>;
3825 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
3826 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
3828 (fsub (fmul R32FP:$rA, R32FP:$rB), R32FP:$rC))]>;
3830 // Floating Negative Mulitply and Subtract
3831 // Subtracts product from value in rC
3832 // res = fneg(fms a b c)
3835 // NOTE: subtraction order
3839 RRRForm<0b1101, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
3840 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
3841 [(set R32FP:$rT, (fsub R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
3844 RRRForm<0b1101, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
3845 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
3846 [(set (v4f32 VECREG:$rT),
3847 (fsub (v4f32 VECREG:$rC),
3848 (fmul (v4f32 VECREG:$rA),
3849 (v4f32 VECREG:$rB))))]>;
3854 // Floating point reciprocal estimate
3856 class FRESTInst<dag OOL, dag IOL>:
3857 RRForm_1<0b00110111000, OOL, IOL,
3858 "frest\t$rT, $rA", SPrecFP,
3859 [/* no pattern */]>;
3862 FRESTInst<(outs VECREG:$rT), (ins VECREG:$rA)>;
3865 FRESTInst<(outs R32FP:$rT), (ins R32FP:$rA)>;
3867 // Floating point interpolate (used in conjunction with reciprocal estimate)
3869 RRForm<0b00101011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3870 "fi\t$rT, $rA, $rB", SPrecFP,
3871 [/* no pattern */]>;
3874 RRForm<0b00101011110, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3875 "fi\t$rT, $rA, $rB", SPrecFP,
3876 [/* no pattern */]>;
3878 //--------------------------------------------------------------------------
3879 // Basic single precision floating point comparisons:
3881 // Note: There is no support on SPU for single precision NaN. Consequently,
3882 // ordered and unordered comparisons are the same.
3883 //--------------------------------------------------------------------------
3886 RRForm<0b01000011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3887 "fceq\t$rT, $rA, $rB", SPrecFP,
3888 [(set R32C:$rT, (setueq R32FP:$rA, R32FP:$rB))]>;
3890 def : Pat<(setoeq R32FP:$rA, R32FP:$rB),
3891 (FCEQf32 R32FP:$rA, R32FP:$rB)>;
3894 RRForm<0b01010011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3895 "fcmeq\t$rT, $rA, $rB", SPrecFP,
3896 [(set R32C:$rT, (setueq (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
3898 def : Pat<(setoeq (fabs R32FP:$rA), (fabs R32FP:$rB)),
3899 (FCMEQf32 R32FP:$rA, R32FP:$rB)>;
3902 RRForm<0b01000011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3903 "fcgt\t$rT, $rA, $rB", SPrecFP,
3904 [(set R32C:$rT, (setugt R32FP:$rA, R32FP:$rB))]>;
3906 def : Pat<(setugt R32FP:$rA, R32FP:$rB),
3907 (FCGTf32 R32FP:$rA, R32FP:$rB)>;
3910 RRForm<0b01010011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3911 "fcmgt\t$rT, $rA, $rB", SPrecFP,
3912 [(set R32C:$rT, (setugt (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
3914 def : Pat<(setugt (fabs R32FP:$rA), (fabs R32FP:$rB)),
3915 (FCMGTf32 R32FP:$rA, R32FP:$rB)>;
3917 //--------------------------------------------------------------------------
3918 // Single precision floating point comparisons and SETCC equivalents:
3919 //--------------------------------------------------------------------------
3921 def : SETCCNegCondReg<setune, R32FP, i32, XORIr32, FCEQf32>;
3922 def : SETCCNegCondReg<setone, R32FP, i32, XORIr32, FCEQf32>;
3924 def : SETCCBinOpReg<setuge, R32FP, ORr32, FCGTf32, FCEQf32>;
3925 def : SETCCBinOpReg<setoge, R32FP, ORr32, FCGTf32, FCEQf32>;
3927 def : SETCCBinOpReg<setult, R32FP, NORr32, FCGTf32, FCEQf32>;
3928 def : SETCCBinOpReg<setolt, R32FP, NORr32, FCGTf32, FCEQf32>;
3930 def : Pat<(setule R32FP:$rA, R32FP:$rB),
3931 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
3932 def : Pat<(setole R32FP:$rA, R32FP:$rB),
3933 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
3935 // FP Status and Control Register Write
3936 // Why isn't rT a don't care in the ISA?
3937 // Should we create a special RRForm_3 for this guy and zero out the rT?
3939 RRForm_1<0b01011101110, (outs R32FP:$rT), (ins R32FP:$rA),
3940 "fscrwr\t$rA", SPrecFP,
3941 [/* This instruction requires an intrinsic. Note: rT is unused. */]>;
3943 // FP Status and Control Register Read
3945 RRForm_2<0b01011101110, (outs R32FP:$rT), (ins),
3946 "fscrrd\t$rT", SPrecFP,
3947 [/* This instruction requires an intrinsic */]>;
3949 // llvm instruction space
3950 // How do these map onto cell instructions?
3952 // frest rC rB # c = 1/b (both lines)
3954 // fm rD rA rC # d = a * 1/b
3955 // fnms rB rD rB rA # b = - (d * b - a) --should == 0 in a perfect world
3956 // fma rB rB rC rD # b = b * c + d
3957 // = -(d *b -a) * c + d
3958 // = a * c - c ( a *b *c - a)
3963 // These llvm instructions will actually map to library calls.
3964 // All that's needed, then, is to check that the appropriate library is
3965 // imported and do a brsl to the proper function name.
3966 // frem # fmod(x, y): x - (x/y) * y
3967 // (Note: fmod(double, double), fmodf(float,float)
3971 // Unimplemented SPU instruction space
3972 // floating reciprocal absolute square root estimate (frsqest)
3974 // The following are probably just intrinsics
3975 // status and control register write
3976 // status and control register read
3978 //--------------------------------------
3979 // Floating Point Conversions
3980 // Signed conversions:
3982 CVTIntFPForm<0b0101101110, (outs VECREG:$rT), (ins VECREG:$rA),
3983 "csflt\t$rT, $rA, 0", SPrecFP,
3984 [(set (v4f32 VECREG:$rT), (sint_to_fp (v4i32 VECREG:$rA)))]>;
3986 // Convert signed integer to floating point
3988 CVTIntFPForm<0b0101101110, (outs R32FP:$rT), (ins R32C:$rA),
3989 "csflt\t$rT, $rA, 0", SPrecFP,
3990 [(set R32FP:$rT, (sint_to_fp R32C:$rA))]>;
3992 // Convert unsigned into to float
3994 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
3995 "cuflt\t$rT, $rA, 0", SPrecFP,
3996 [(set (v4f32 VECREG:$rT), (uint_to_fp (v4i32 VECREG:$rA)))]>;
3999 CVTIntFPForm<0b1101101110, (outs R32FP:$rT), (ins R32C:$rA),
4000 "cuflt\t$rT, $rA, 0", SPrecFP,
4001 [(set R32FP:$rT, (uint_to_fp R32C:$rA))]>;
4003 // Convert float to unsigned int
4004 // Assume that scale = 0
4007 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4008 "cfltu\t$rT, $rA, 0", SPrecFP,
4009 [(set (v4i32 VECREG:$rT), (fp_to_uint (v4f32 VECREG:$rA)))]>;
4012 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
4013 "cfltu\t$rT, $rA, 0", SPrecFP,
4014 [(set R32C:$rT, (fp_to_uint R32FP:$rA))]>;
4016 // Convert float to signed int
4017 // Assume that scale = 0
4020 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4021 "cflts\t$rT, $rA, 0", SPrecFP,
4022 [(set (v4i32 VECREG:$rT), (fp_to_sint (v4f32 VECREG:$rA)))]>;
4025 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
4026 "cflts\t$rT, $rA, 0", SPrecFP,
4027 [(set R32C:$rT, (fp_to_sint R32FP:$rA))]>;
4029 //===----------------------------------------------------------------------==//
4030 // Single<->Double precision conversions
4031 //===----------------------------------------------------------------------==//
4033 // NOTE: We use "vec" name suffix here to avoid confusion (e.g. input is a
4034 // v4f32, output is v2f64--which goes in the name?)
4036 // Floating point extend single to double
4037 // NOTE: Not sure if passing in v4f32 to FESDvec is correct since it
4038 // operates on two double-word slots (i.e. 1st and 3rd fp numbers
4041 RRForm_1<0b00011101110, (outs VECREG:$rT), (ins VECREG:$rA),
4042 "fesd\t$rT, $rA", SPrecFP,
4043 [/*(set (v2f64 VECREG:$rT), (fextend (v4f32 VECREG:$rA)))*/]>;
4046 RRForm_1<0b00011101110, (outs R64FP:$rT), (ins R32FP:$rA),
4047 "fesd\t$rT, $rA", SPrecFP,
4048 [(set R64FP:$rT, (fextend R32FP:$rA))]>;
4050 // Floating point round double to single
4052 // RRForm_1<0b10011101110, (outs VECREG:$rT), (ins VECREG:$rA),
4053 // "frds\t$rT, $rA,", SPrecFP,
4054 // [(set (v4f32 R32FP:$rT), (fround (v2f64 R64FP:$rA)))]>;
4057 RRForm_1<0b10011101110, (outs R32FP:$rT), (ins R64FP:$rA),
4058 "frds\t$rT, $rA", SPrecFP,
4059 [(set R32FP:$rT, (fround R64FP:$rA))]>;
4061 //ToDo include anyextend?
4063 //===----------------------------------------------------------------------==//
4064 // Double precision floating point instructions
4065 //===----------------------------------------------------------------------==//
4067 RRForm<0b00110011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4068 "dfa\t$rT, $rA, $rB", DPrecFP,
4069 [(set R64FP:$rT, (fadd R64FP:$rA, R64FP:$rB))]>;
4072 RRForm<0b00110011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4073 "dfa\t$rT, $rA, $rB", DPrecFP,
4074 [(set (v2f64 VECREG:$rT), (fadd (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4077 RRForm<0b10100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4078 "dfs\t$rT, $rA, $rB", DPrecFP,
4079 [(set R64FP:$rT, (fsub R64FP:$rA, R64FP:$rB))]>;
4082 RRForm<0b10100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4083 "dfs\t$rT, $rA, $rB", DPrecFP,
4084 [(set (v2f64 VECREG:$rT),
4085 (fsub (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4088 RRForm<0b01100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4089 "dfm\t$rT, $rA, $rB", DPrecFP,
4090 [(set R64FP:$rT, (fmul R64FP:$rA, R64FP:$rB))]>;
4093 RRForm<0b00100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4094 "dfm\t$rT, $rA, $rB", DPrecFP,
4095 [(set (v2f64 VECREG:$rT),
4096 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4099 RRForm<0b00111010110, (outs R64FP:$rT),
4100 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4101 "dfma\t$rT, $rA, $rB", DPrecFP,
4102 [(set R64FP:$rT, (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB)))]>,
4103 RegConstraint<"$rC = $rT">,
4107 RRForm<0b00111010110, (outs VECREG:$rT),
4108 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4109 "dfma\t$rT, $rA, $rB", DPrecFP,
4110 [(set (v2f64 VECREG:$rT),
4111 (fadd (v2f64 VECREG:$rC),
4112 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB))))]>,
4113 RegConstraint<"$rC = $rT">,
4117 RRForm<0b10111010110, (outs R64FP:$rT),
4118 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4119 "dfms\t$rT, $rA, $rB", DPrecFP,
4120 [(set R64FP:$rT, (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC))]>,
4121 RegConstraint<"$rC = $rT">,
4125 RRForm<0b10111010110, (outs VECREG:$rT),
4126 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4127 "dfms\t$rT, $rA, $rB", DPrecFP,
4128 [(set (v2f64 VECREG:$rT),
4129 (fsub (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)),
4130 (v2f64 VECREG:$rC)))]>;
4132 // DFNMS: - (a * b - c)
4133 // - (a * b) + c => c - (a * b)
4135 class DFNMSInst<dag OOL, dag IOL, list<dag> pattern>:
4136 RRForm<0b01111010110, OOL, IOL, "dfnms\t$rT, $rA, $rB",
4138 RegConstraint<"$rC = $rT">,
4141 class DFNMSVecInst<list<dag> pattern>:
4142 DFNMSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4145 class DFNMSRegInst<list<dag> pattern>:
4146 DFNMSInst<(outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4149 multiclass DFMultiplySubtract
4151 def v2f64 : DFNMSVecInst<[(set (v2f64 VECREG:$rT),
4152 (fsub (v2f64 VECREG:$rC),
4153 (fmul (v2f64 VECREG:$rA),
4154 (v2f64 VECREG:$rB))))]>;
4156 def f64 : DFNMSRegInst<[(set R64FP:$rT,
4158 (fmul R64FP:$rA, R64FP:$rB)))]>;
4161 defm DFNMS : DFMultiplySubtract;
4166 RRForm<0b11111010110, (outs R64FP:$rT),
4167 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4168 "dfnma\t$rT, $rA, $rB", DPrecFP,
4169 [(set R64FP:$rT, (fneg (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB))))]>,
4170 RegConstraint<"$rC = $rT">,
4174 RRForm<0b11111010110, (outs VECREG:$rT),
4175 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4176 "dfnma\t$rT, $rA, $rB", DPrecFP,
4177 [(set (v2f64 VECREG:$rT),
4178 (fneg (fadd (v2f64 VECREG:$rC),
4179 (fmul (v2f64 VECREG:$rA),
4180 (v2f64 VECREG:$rB)))))]>,
4181 RegConstraint<"$rC = $rT">,
4184 //===----------------------------------------------------------------------==//
4185 // Floating point negation and absolute value
4186 //===----------------------------------------------------------------------==//
4188 def : Pat<(fneg (v4f32 VECREG:$rA)),
4189 (XORfnegvec (v4f32 VECREG:$rA),
4190 (v4f32 (ILHUv4i32 0x8000)))>;
4192 def : Pat<(fneg R32FP:$rA),
4193 (XORfneg32 R32FP:$rA, (ILHUr32 0x8000))>;
4195 // Floating point absolute value
4196 // Note: f64 fabs is custom-selected.
4198 def : Pat<(fabs R32FP:$rA),
4199 (ANDfabs32 R32FP:$rA, (IOHLr32 (ILHUr32 0x7fff), 0xffff))>;
4201 def : Pat<(fabs (v4f32 VECREG:$rA)),
4202 (ANDfabsvec (v4f32 VECREG:$rA),
4203 (IOHLv4i32 (ILHUv4i32 0x7fff), 0xffff))>;
4205 //===----------------------------------------------------------------------===//
4206 // Hint for branch instructions:
4207 //===----------------------------------------------------------------------===//
4209 /* def HBR : SPUInstr<(outs), (ins), "hbr\t" */
4211 //===----------------------------------------------------------------------===//
4212 // Execution, Load NOP (execute NOPs belong in even pipeline, load NOPs belong
4213 // in the odd pipeline)
4214 //===----------------------------------------------------------------------===//
4216 def ENOP : SPUInstr<(outs), (ins), "enop", ExecNOP> {
4219 let Inst{0-10} = 0b10000000010;
4220 let Inst{11-17} = 0;
4221 let Inst{18-24} = 0;
4222 let Inst{25-31} = 0;
4225 def LNOP : SPUInstr<(outs), (ins), "lnop", LoadNOP> {
4228 let Inst{0-10} = 0b10000000000;
4229 let Inst{11-17} = 0;
4230 let Inst{18-24} = 0;
4231 let Inst{25-31} = 0;
4234 //===----------------------------------------------------------------------===//
4235 // Bit conversions (type conversions between vector/packed types)
4236 // NOTE: Promotions are handled using the XS* instructions.
4237 //===----------------------------------------------------------------------===//
4238 def : Pat<(v16i8 (bitconvert (v8i16 VECREG:$src))), (v16i8 VECREG:$src)>;
4239 def : Pat<(v16i8 (bitconvert (v4i32 VECREG:$src))), (v16i8 VECREG:$src)>;
4240 def : Pat<(v16i8 (bitconvert (v2i64 VECREG:$src))), (v16i8 VECREG:$src)>;
4241 def : Pat<(v16i8 (bitconvert (v4f32 VECREG:$src))), (v16i8 VECREG:$src)>;
4242 def : Pat<(v16i8 (bitconvert (v2f64 VECREG:$src))), (v16i8 VECREG:$src)>;
4244 def : Pat<(v8i16 (bitconvert (v16i8 VECREG:$src))), (v8i16 VECREG:$src)>;
4245 def : Pat<(v8i16 (bitconvert (v4i32 VECREG:$src))), (v8i16 VECREG:$src)>;
4246 def : Pat<(v8i16 (bitconvert (v2i64 VECREG:$src))), (v8i16 VECREG:$src)>;
4247 def : Pat<(v8i16 (bitconvert (v4f32 VECREG:$src))), (v8i16 VECREG:$src)>;
4248 def : Pat<(v8i16 (bitconvert (v2f64 VECREG:$src))), (v8i16 VECREG:$src)>;
4250 def : Pat<(v4i32 (bitconvert (v16i8 VECREG:$src))), (v4i32 VECREG:$src)>;
4251 def : Pat<(v4i32 (bitconvert (v8i16 VECREG:$src))), (v4i32 VECREG:$src)>;
4252 def : Pat<(v4i32 (bitconvert (v2i64 VECREG:$src))), (v4i32 VECREG:$src)>;
4253 def : Pat<(v4i32 (bitconvert (v4f32 VECREG:$src))), (v4i32 VECREG:$src)>;
4254 def : Pat<(v4i32 (bitconvert (v2f64 VECREG:$src))), (v4i32 VECREG:$src)>;
4256 def : Pat<(v2i64 (bitconvert (v16i8 VECREG:$src))), (v2i64 VECREG:$src)>;
4257 def : Pat<(v2i64 (bitconvert (v8i16 VECREG:$src))), (v2i64 VECREG:$src)>;
4258 def : Pat<(v2i64 (bitconvert (v4i32 VECREG:$src))), (v2i64 VECREG:$src)>;
4259 def : Pat<(v2i64 (bitconvert (v4f32 VECREG:$src))), (v2i64 VECREG:$src)>;
4260 def : Pat<(v2i64 (bitconvert (v2f64 VECREG:$src))), (v2i64 VECREG:$src)>;
4262 def : Pat<(v4f32 (bitconvert (v16i8 VECREG:$src))), (v4f32 VECREG:$src)>;
4263 def : Pat<(v4f32 (bitconvert (v8i16 VECREG:$src))), (v4f32 VECREG:$src)>;
4264 def : Pat<(v4f32 (bitconvert (v2i64 VECREG:$src))), (v4f32 VECREG:$src)>;
4265 def : Pat<(v4f32 (bitconvert (v4i32 VECREG:$src))), (v4f32 VECREG:$src)>;
4266 def : Pat<(v4f32 (bitconvert (v2f64 VECREG:$src))), (v4f32 VECREG:$src)>;
4268 def : Pat<(v2f64 (bitconvert (v16i8 VECREG:$src))), (v2f64 VECREG:$src)>;
4269 def : Pat<(v2f64 (bitconvert (v8i16 VECREG:$src))), (v2f64 VECREG:$src)>;
4270 def : Pat<(v2f64 (bitconvert (v4i32 VECREG:$src))), (v2f64 VECREG:$src)>;
4271 def : Pat<(v2f64 (bitconvert (v2i64 VECREG:$src))), (v2f64 VECREG:$src)>;
4272 def : Pat<(v2f64 (bitconvert (v4f32 VECREG:$src))), (v2f64 VECREG:$src)>;
4274 def : Pat<(i128 (bitconvert (v16i8 VECREG:$src))),
4275 (COPY_TO_REGCLASS VECREG:$src, GPRC)>;
4276 def : Pat<(i128 (bitconvert (v8i16 VECREG:$src))),
4277 (COPY_TO_REGCLASS VECREG:$src, GPRC)>;
4278 def : Pat<(i128 (bitconvert (v4i32 VECREG:$src))),
4279 (COPY_TO_REGCLASS VECREG:$src, GPRC)>;
4280 def : Pat<(i128 (bitconvert (v2i64 VECREG:$src))),
4281 (COPY_TO_REGCLASS VECREG:$src, GPRC)>;
4282 def : Pat<(i128 (bitconvert (v4f32 VECREG:$src))),
4283 (COPY_TO_REGCLASS VECREG:$src, GPRC)>;
4284 def : Pat<(i128 (bitconvert (v2f64 VECREG:$src))),
4285 (COPY_TO_REGCLASS VECREG:$src, GPRC)>;
4287 def : Pat<(v16i8 (bitconvert (i128 GPRC:$src))),
4288 (v16i8 (COPY_TO_REGCLASS GPRC:$src, VECREG))>;
4289 def : Pat<(v8i16 (bitconvert (i128 GPRC:$src))),
4290 (v8i16 (COPY_TO_REGCLASS GPRC:$src, VECREG))>;
4291 def : Pat<(v4i32 (bitconvert (i128 GPRC:$src))),
4292 (v4i32 (COPY_TO_REGCLASS GPRC:$src, VECREG))>;
4293 def : Pat<(v2i64 (bitconvert (i128 GPRC:$src))),
4294 (v2i64 (COPY_TO_REGCLASS GPRC:$src, VECREG))>;
4295 def : Pat<(v4f32 (bitconvert (i128 GPRC:$src))),
4296 (v4f32 (COPY_TO_REGCLASS GPRC:$src, VECREG))>;
4297 def : Pat<(v2f64 (bitconvert (i128 GPRC:$src))),
4298 (v2f64 (COPY_TO_REGCLASS GPRC:$src, VECREG))>;
4300 def : Pat<(i32 (bitconvert R32FP:$rA)),
4301 (COPY_TO_REGCLASS R32FP:$rA, R32C)>;
4303 def : Pat<(f32 (bitconvert R32C:$rA)),
4304 (COPY_TO_REGCLASS R32C:$rA, R32FP)>;
4306 def : Pat<(i64 (bitconvert R64FP:$rA)),
4307 (COPY_TO_REGCLASS R64FP:$rA, R64C)>;
4309 def : Pat<(f64 (bitconvert R64C:$rA)),
4310 (COPY_TO_REGCLASS R64C:$rA, R64FP)>;
4313 //===----------------------------------------------------------------------===//
4314 // Instruction patterns:
4315 //===----------------------------------------------------------------------===//
4317 // General 32-bit constants:
4318 def : Pat<(i32 imm:$imm),
4319 (IOHLr32 (ILHUr32 (HI16 imm:$imm)), (LO16 imm:$imm))>;
4321 // Single precision float constants:
4322 def : Pat<(f32 fpimm:$imm),
4323 (IOHLf32 (ILHUf32 (HI16_f32 fpimm:$imm)), (LO16_f32 fpimm:$imm))>;
4325 // General constant 32-bit vectors
4326 def : Pat<(v4i32 v4i32Imm:$imm),
4327 (IOHLv4i32 (v4i32 (ILHUv4i32 (HI16_vec v4i32Imm:$imm))),
4328 (LO16_vec v4i32Imm:$imm))>;
4331 def : Pat<(i8 imm:$imm),
4334 //===----------------------------------------------------------------------===//
4335 // Zero/Any/Sign extensions
4336 //===----------------------------------------------------------------------===//
4338 // sext 8->32: Sign extend bytes to words
4339 def : Pat<(sext_inreg R32C:$rSrc, i8),
4340 (XSHWr32 (XSBHr32 R32C:$rSrc))>;
4342 def : Pat<(i32 (sext R8C:$rSrc)),
4343 (XSHWr16 (XSBHr8 R8C:$rSrc))>;
4345 // sext 8->64: Sign extend bytes to double word
4346 def : Pat<(sext_inreg R64C:$rSrc, i8),
4347 (XSWDr64_inreg (XSHWr64 (XSBHr64 R64C:$rSrc)))>;
4349 def : Pat<(i64 (sext R8C:$rSrc)),
4350 (XSWDr64 (XSHWr16 (XSBHr8 R8C:$rSrc)))>;
4352 // zext 8->16: Zero extend bytes to halfwords
4353 def : Pat<(i16 (zext R8C:$rSrc)),
4354 (ANDHIi8i16 R8C:$rSrc, 0xff)>;
4356 // zext 8->32: Zero extend bytes to words
4357 def : Pat<(i32 (zext R8C:$rSrc)),
4358 (ANDIi8i32 R8C:$rSrc, 0xff)>;
4360 // zext 8->64: Zero extend bytes to double words
4361 def : Pat<(i64 (zext R8C:$rSrc)),
4362 (COPY_TO_REGCLASS (SELBv4i32 (ROTQMBYv4i32
4364 (ANDIi8i32 R8C:$rSrc,0xff), VECREG),
4367 (FSMBIv4i32 0x0f0f)), R64C)>;
4369 // anyext 8->16: Extend 8->16 bits, irrespective of sign, preserves high bits
4370 def : Pat<(i16 (anyext R8C:$rSrc)),
4371 (ORHIi8i16 R8C:$rSrc, 0)>;
4373 // anyext 8->32: Extend 8->32 bits, irrespective of sign, preserves high bits
4374 def : Pat<(i32 (anyext R8C:$rSrc)),
4375 (COPY_TO_REGCLASS R8C:$rSrc, R32C)>;
4377 // sext 16->64: Sign extend halfword to double word
4378 def : Pat<(sext_inreg R64C:$rSrc, i16),
4379 (XSWDr64_inreg (XSHWr64 R64C:$rSrc))>;
4381 def : Pat<(sext R16C:$rSrc),
4382 (XSWDr64 (XSHWr16 R16C:$rSrc))>;
4384 // zext 16->32: Zero extend halfwords to words
4385 def : Pat<(i32 (zext R16C:$rSrc)),
4386 (ANDi16i32 R16C:$rSrc, (ILAr32 0xffff))>;
4388 def : Pat<(i32 (zext (and R16C:$rSrc, 0xf))),
4389 (ANDIi16i32 R16C:$rSrc, 0xf)>;
4391 def : Pat<(i32 (zext (and R16C:$rSrc, 0xff))),
4392 (ANDIi16i32 R16C:$rSrc, 0xff)>;
4394 def : Pat<(i32 (zext (and R16C:$rSrc, 0xfff))),
4395 (ANDIi16i32 R16C:$rSrc, 0xfff)>;
4397 // anyext 16->32: Extend 16->32 bits, irrespective of sign
4398 def : Pat<(i32 (anyext R16C:$rSrc)),
4399 (COPY_TO_REGCLASS R16C:$rSrc, R32C)>;
4401 //===----------------------------------------------------------------------===//
4403 // These truncates are for the SPU's supported types (i8, i16, i32). i64 and
4404 // above are custom lowered.
4405 //===----------------------------------------------------------------------===//
4407 def : Pat<(i8 (trunc GPRC:$src)),
4409 (SHUFBgprc GPRC:$src, GPRC:$src,
4410 (IOHLv4i32 (ILHUv4i32 0x0f0f), 0x0f0f)), R8C)>;
4412 def : Pat<(i8 (trunc R64C:$src)),
4415 (COPY_TO_REGCLASS R64C:$src, VECREG),
4416 (COPY_TO_REGCLASS R64C:$src, VECREG),
4417 (IOHLv4i32 (ILHUv4i32 0x0707), 0x0707)), R8C)>;
4419 def : Pat<(i8 (trunc R32C:$src)),
4422 (COPY_TO_REGCLASS R32C:$src, VECREG),
4423 (COPY_TO_REGCLASS R32C:$src, VECREG),
4424 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)), R8C)>;
4426 def : Pat<(i8 (trunc R16C:$src)),
4429 (COPY_TO_REGCLASS R16C:$src, VECREG),
4430 (COPY_TO_REGCLASS R16C:$src, VECREG),
4431 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)), R8C)>;
4433 def : Pat<(i16 (trunc GPRC:$src)),
4435 (SHUFBgprc GPRC:$src, GPRC:$src,
4436 (IOHLv4i32 (ILHUv4i32 0x0e0f), 0x0e0f)), R16C)>;
4438 def : Pat<(i16 (trunc R64C:$src)),
4441 (COPY_TO_REGCLASS R64C:$src, VECREG),
4442 (COPY_TO_REGCLASS R64C:$src, VECREG),
4443 (IOHLv4i32 (ILHUv4i32 0x0607), 0x0607)), R16C)>;
4445 def : Pat<(i16 (trunc R32C:$src)),
4448 (COPY_TO_REGCLASS R32C:$src, VECREG),
4449 (COPY_TO_REGCLASS R32C:$src, VECREG),
4450 (IOHLv4i32 (ILHUv4i32 0x0203), 0x0203)), R16C)>;
4452 def : Pat<(i32 (trunc GPRC:$src)),
4454 (SHUFBgprc GPRC:$src, GPRC:$src,
4455 (IOHLv4i32 (ILHUv4i32 0x0c0d), 0x0e0f)), R32C)>;
4457 def : Pat<(i32 (trunc R64C:$src)),
4460 (COPY_TO_REGCLASS R64C:$src, VECREG),
4461 (COPY_TO_REGCLASS R64C:$src, VECREG),
4462 (IOHLv4i32 (ILHUv4i32 0x0405), 0x0607)), R32C)>;
4464 //===----------------------------------------------------------------------===//
4465 // Address generation: SPU, like PPC, has to split addresses into high and
4466 // low parts in order to load them into a register.
4467 //===----------------------------------------------------------------------===//
4469 def : Pat<(SPUaform tglobaladdr:$in, 0), (ILAlsa tglobaladdr:$in)>;
4470 def : Pat<(SPUaform texternalsym:$in, 0), (ILAlsa texternalsym:$in)>;
4471 def : Pat<(SPUaform tjumptable:$in, 0), (ILAlsa tjumptable:$in)>;
4472 def : Pat<(SPUaform tconstpool:$in, 0), (ILAlsa tconstpool:$in)>;
4474 def : Pat<(SPUindirect (SPUhi tglobaladdr:$in, 0),
4475 (SPUlo tglobaladdr:$in, 0)),
4476 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4478 def : Pat<(SPUindirect (SPUhi texternalsym:$in, 0),
4479 (SPUlo texternalsym:$in, 0)),
4480 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4482 def : Pat<(SPUindirect (SPUhi tjumptable:$in, 0),
4483 (SPUlo tjumptable:$in, 0)),
4484 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4486 def : Pat<(SPUindirect (SPUhi tconstpool:$in, 0),
4487 (SPUlo tconstpool:$in, 0)),
4488 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4490 def : Pat<(add (SPUhi tglobaladdr:$in, 0), (SPUlo tglobaladdr:$in, 0)),
4491 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4493 def : Pat<(add (SPUhi texternalsym:$in, 0), (SPUlo texternalsym:$in, 0)),
4494 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4496 def : Pat<(add (SPUhi tjumptable:$in, 0), (SPUlo tjumptable:$in, 0)),
4497 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4499 def : Pat<(add (SPUhi tconstpool:$in, 0), (SPUlo tconstpool:$in, 0)),
4500 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4503 include "CellSDKIntrinsics.td"
4504 // Various math operator instruction sequences
4505 include "SPUMathInstr.td"
4506 // 64-bit "instructions"/support
4507 include "SPU64InstrInfo.td"
4508 // 128-bit "instructions"/support
4509 include "SPU128InstrInfo.td"