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 v2i32: LoadDFormVec<v2i32>;
67 def r128: LoadDForm<GPRC>;
68 def r64: LoadDForm<R64C>;
69 def r32: LoadDForm<R32C>;
70 def f32: LoadDForm<R32FP>;
71 def f64: LoadDForm<R64FP>;
72 def r16: LoadDForm<R16C>;
73 def r8: LoadDForm<R8C>;
76 class LoadAFormVec<ValueType vectype>
77 : RI16Form<0b100001100, (outs VECREG:$rT), (ins addr256k:$src),
80 [(set (vectype VECREG:$rT), (load aform_addr:$src))]>
83 class LoadAForm<RegisterClass rclass>
84 : RI16Form<0b100001100, (outs rclass:$rT), (ins addr256k:$src),
87 [(set rclass:$rT, (load aform_addr:$src))]>
92 def v16i8: LoadAFormVec<v16i8>;
93 def v8i16: LoadAFormVec<v8i16>;
94 def v4i32: LoadAFormVec<v4i32>;
95 def v2i64: LoadAFormVec<v2i64>;
96 def v4f32: LoadAFormVec<v4f32>;
97 def v2f64: LoadAFormVec<v2f64>;
99 def v2i32: LoadAFormVec<v2i32>;
101 def r128: LoadAForm<GPRC>;
102 def r64: LoadAForm<R64C>;
103 def r32: LoadAForm<R32C>;
104 def f32: LoadAForm<R32FP>;
105 def f64: LoadAForm<R64FP>;
106 def r16: LoadAForm<R16C>;
107 def r8: LoadAForm<R8C>;
110 class LoadXFormVec<ValueType vectype>
111 : RRForm<0b00100011100, (outs VECREG:$rT), (ins memrr:$src),
114 [(set (vectype VECREG:$rT), (load xform_addr:$src))]>
117 class LoadXForm<RegisterClass rclass>
118 : RRForm<0b00100011100, (outs rclass:$rT), (ins memrr:$src),
121 [(set rclass:$rT, (load xform_addr:$src))]>
124 multiclass LoadXForms
126 def v16i8: LoadXFormVec<v16i8>;
127 def v8i16: LoadXFormVec<v8i16>;
128 def v4i32: LoadXFormVec<v4i32>;
129 def v2i64: LoadXFormVec<v2i64>;
130 def v4f32: LoadXFormVec<v4f32>;
131 def v2f64: LoadXFormVec<v2f64>;
133 def v2i32: LoadXFormVec<v2i32>;
135 def r128: LoadXForm<GPRC>;
136 def r64: LoadXForm<R64C>;
137 def r32: LoadXForm<R32C>;
138 def f32: LoadXForm<R32FP>;
139 def f64: LoadXForm<R64FP>;
140 def r16: LoadXForm<R16C>;
141 def r8: LoadXForm<R8C>;
144 defm LQA : LoadAForms;
145 defm LQD : LoadDForms;
146 defm LQX : LoadXForms;
148 /* Load quadword, PC relative: Not much use at this point in time.
149 Might be of use later for relocatable code. It's effectively the
150 same as LQA, but uses PC-relative addressing.
151 def LQR : RI16Form<0b111001100, (outs VECREG:$rT), (ins s16imm:$disp),
152 "lqr\t$rT, $disp", LoadStore,
153 [(set VECREG:$rT, (load iaddr:$disp))]>;
157 //===----------------------------------------------------------------------===//
159 //===----------------------------------------------------------------------===//
160 class StoreDFormVec<ValueType vectype>
161 : RI10Form<0b00100100, (outs), (ins VECREG:$rT, dformaddr:$src),
164 [(store (vectype VECREG:$rT), dform_addr:$src)]>
167 class StoreDForm<RegisterClass rclass>
168 : RI10Form<0b00100100, (outs), (ins rclass:$rT, dformaddr:$src),
171 [(store rclass:$rT, dform_addr:$src)]>
174 multiclass StoreDForms
176 def v16i8: StoreDFormVec<v16i8>;
177 def v8i16: StoreDFormVec<v8i16>;
178 def v4i32: StoreDFormVec<v4i32>;
179 def v2i64: StoreDFormVec<v2i64>;
180 def v4f32: StoreDFormVec<v4f32>;
181 def v2f64: StoreDFormVec<v2f64>;
183 def v2i32: StoreDFormVec<v2i32>;
185 def r128: StoreDForm<GPRC>;
186 def r64: StoreDForm<R64C>;
187 def r32: StoreDForm<R32C>;
188 def f32: StoreDForm<R32FP>;
189 def f64: StoreDForm<R64FP>;
190 def r16: StoreDForm<R16C>;
191 def r8: StoreDForm<R8C>;
194 class StoreAFormVec<ValueType vectype>
195 : RI16Form<0b0010010, (outs), (ins VECREG:$rT, addr256k:$src),
198 [(store (vectype VECREG:$rT), aform_addr:$src)]>;
200 class StoreAForm<RegisterClass rclass>
201 : RI16Form<0b001001, (outs), (ins rclass:$rT, addr256k:$src),
204 [(store rclass:$rT, aform_addr:$src)]>;
206 multiclass StoreAForms
208 def v16i8: StoreAFormVec<v16i8>;
209 def v8i16: StoreAFormVec<v8i16>;
210 def v4i32: StoreAFormVec<v4i32>;
211 def v2i64: StoreAFormVec<v2i64>;
212 def v4f32: StoreAFormVec<v4f32>;
213 def v2f64: StoreAFormVec<v2f64>;
215 def v2i32: StoreAFormVec<v2i32>;
217 def r128: StoreAForm<GPRC>;
218 def r64: StoreAForm<R64C>;
219 def r32: StoreAForm<R32C>;
220 def f32: StoreAForm<R32FP>;
221 def f64: StoreAForm<R64FP>;
222 def r16: StoreAForm<R16C>;
223 def r8: StoreAForm<R8C>;
226 class StoreXFormVec<ValueType vectype>
227 : RRForm<0b00100100, (outs), (ins VECREG:$rT, memrr:$src),
230 [(store (vectype VECREG:$rT), xform_addr:$src)]>
233 class StoreXForm<RegisterClass rclass>
234 : RRForm<0b00100100, (outs), (ins rclass:$rT, memrr:$src),
237 [(store rclass:$rT, xform_addr:$src)]>
240 multiclass StoreXForms
242 def v16i8: StoreXFormVec<v16i8>;
243 def v8i16: StoreXFormVec<v8i16>;
244 def v4i32: StoreXFormVec<v4i32>;
245 def v2i64: StoreXFormVec<v2i64>;
246 def v4f32: StoreXFormVec<v4f32>;
247 def v2f64: StoreXFormVec<v2f64>;
249 def v2i32: StoreXFormVec<v2i32>;
251 def r128: StoreXForm<GPRC>;
252 def r64: StoreXForm<R64C>;
253 def r32: StoreXForm<R32C>;
254 def f32: StoreXForm<R32FP>;
255 def f64: StoreXForm<R64FP>;
256 def r16: StoreXForm<R16C>;
257 def r8: StoreXForm<R8C>;
260 defm STQD : StoreDForms;
261 defm STQA : StoreAForms;
262 defm STQX : StoreXForms;
264 /* Store quadword, PC relative: Not much use at this point in time. Might
265 be useful for relocatable code.
266 def STQR : RI16Form<0b111000100, (outs), (ins VECREG:$rT, s16imm:$disp),
267 "stqr\t$rT, $disp", LoadStore,
268 [(store VECREG:$rT, iaddr:$disp)]>;
271 //===----------------------------------------------------------------------===//
272 // Generate Controls for Insertion:
273 //===----------------------------------------------------------------------===//
275 def CBD: RI7Form<0b10101111100, (outs VECREG:$rT), (ins shufaddr:$src),
276 "cbd\t$rT, $src", ShuffleOp,
277 [(set (v16i8 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
279 def CBX: RRForm<0b00101011100, (outs VECREG:$rT), (ins memrr:$src),
280 "cbx\t$rT, $src", ShuffleOp,
281 [(set (v16i8 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
283 def CHD: RI7Form<0b10101111100, (outs VECREG:$rT), (ins shufaddr:$src),
284 "chd\t$rT, $src", ShuffleOp,
285 [(set (v8i16 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
287 def CHX: RRForm<0b10101011100, (outs VECREG:$rT), (ins memrr:$src),
288 "chx\t$rT, $src", ShuffleOp,
289 [(set (v8i16 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
291 def CWD: RI7Form<0b01101111100, (outs VECREG:$rT), (ins shufaddr:$src),
292 "cwd\t$rT, $src", ShuffleOp,
293 [(set (v4i32 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
295 def CWX: RRForm<0b01101011100, (outs VECREG:$rT), (ins memrr:$src),
296 "cwx\t$rT, $src", ShuffleOp,
297 [(set (v4i32 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
299 def CWDf32: RI7Form<0b01101111100, (outs VECREG:$rT), (ins shufaddr:$src),
300 "cwd\t$rT, $src", ShuffleOp,
301 [(set (v4f32 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
303 def CWXf32: RRForm<0b01101011100, (outs VECREG:$rT), (ins memrr:$src),
304 "cwx\t$rT, $src", ShuffleOp,
305 [(set (v4f32 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
307 def CDD: RI7Form<0b11101111100, (outs VECREG:$rT), (ins shufaddr:$src),
308 "cdd\t$rT, $src", ShuffleOp,
309 [(set (v2i64 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
311 def CDX: RRForm<0b11101011100, (outs VECREG:$rT), (ins memrr:$src),
312 "cdx\t$rT, $src", ShuffleOp,
313 [(set (v2i64 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
315 def CDDf64: RI7Form<0b11101111100, (outs VECREG:$rT), (ins shufaddr:$src),
316 "cdd\t$rT, $src", ShuffleOp,
317 [(set (v2f64 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
319 def CDXf64: RRForm<0b11101011100, (outs VECREG:$rT), (ins memrr:$src),
320 "cdx\t$rT, $src", ShuffleOp,
321 [(set (v2f64 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
323 //===----------------------------------------------------------------------===//
324 // Constant formation:
325 //===----------------------------------------------------------------------===//
328 RI16Form<0b110000010, (outs VECREG:$rT), (ins s16imm:$val),
329 "ilh\t$rT, $val", ImmLoad,
330 [(set (v8i16 VECREG:$rT), (v8i16 v8i16SExt16Imm:$val))]>;
333 RI16Form<0b110000010, (outs R16C:$rT), (ins s16imm:$val),
334 "ilh\t$rT, $val", ImmLoad,
335 [(set R16C:$rT, immSExt16:$val)]>;
337 // Cell SPU doesn't have a native 8-bit immediate load, but ILH works ("with
338 // the right constant")
340 RI16Form<0b110000010, (outs R8C:$rT), (ins s16imm_i8:$val),
341 "ilh\t$rT, $val", ImmLoad,
342 [(set R8C:$rT, immSExt8:$val)]>;
344 // IL does sign extension!
346 class ILInst<dag OOL, dag IOL, list<dag> pattern>:
347 RI16Form<0b100000010, OOL, IOL, "il\t$rT, $val",
350 class ILVecInst<ValueType vectype, Operand immtype, PatLeaf xform>:
351 ILInst<(outs VECREG:$rT), (ins immtype:$val),
352 [(set (vectype VECREG:$rT), (vectype xform:$val))]>;
354 class ILRegInst<RegisterClass rclass, Operand immtype, PatLeaf xform>:
355 ILInst<(outs rclass:$rT), (ins immtype:$val),
356 [(set rclass:$rT, xform:$val)]>;
358 multiclass ImmediateLoad
360 def v2i64: ILVecInst<v2i64, s16imm_i64, v2i64SExt16Imm>;
361 def v4i32: ILVecInst<v4i32, s16imm_i32, v4i32SExt16Imm>;
363 // TODO: Need v2f64, v4f32
365 def r64: ILRegInst<R64C, s16imm_i64, immSExt16>;
366 def r32: ILRegInst<R32C, s16imm_i32, immSExt16>;
367 def f32: ILRegInst<R32FP, s16imm_f32, fpimmSExt16>;
368 def f64: ILRegInst<R64FP, s16imm_f64, fpimmSExt16>;
371 defm IL : ImmediateLoad;
373 class ILHUInst<dag OOL, dag IOL, list<dag> pattern>:
374 RI16Form<0b010000010, OOL, IOL, "ilhu\t$rT, $val",
377 class ILHUVecInst<ValueType vectype, Operand immtype, PatLeaf xform>:
378 ILHUInst<(outs VECREG:$rT), (ins immtype:$val),
379 [(set (vectype VECREG:$rT), (vectype xform:$val))]>;
381 class ILHURegInst<RegisterClass rclass, Operand immtype, PatLeaf xform>:
382 ILHUInst<(outs rclass:$rT), (ins immtype:$val),
383 [(set rclass:$rT, xform:$val)]>;
385 multiclass ImmLoadHalfwordUpper
387 def v2i64: ILHUVecInst<v2i64, u16imm_i64, immILHUvec_i64>;
388 def v4i32: ILHUVecInst<v4i32, u16imm_i32, immILHUvec>;
390 def r64: ILHURegInst<R64C, u16imm_i64, hi16>;
391 def r32: ILHURegInst<R32C, u16imm_i32, hi16>;
393 // Loads the high portion of an address
394 def hi: ILHURegInst<R32C, symbolHi, hi16>;
396 // Used in custom lowering constant SFP loads:
397 def f32: ILHURegInst<R32FP, f16imm, hi16_f32>;
400 defm ILHU : ImmLoadHalfwordUpper;
402 // Immediate load address (can also be used to load 18-bit unsigned constants,
403 // see the zext 16->32 pattern)
405 class ILAInst<dag OOL, dag IOL, list<dag> pattern>:
406 RI18Form<0b1000010, OOL, IOL, "ila\t$rT, $val",
409 class ILAVecInst<ValueType vectype, Operand immtype, PatLeaf xform>:
410 ILAInst<(outs VECREG:$rT), (ins immtype:$val),
411 [(set (vectype VECREG:$rT), (vectype xform:$val))]>;
413 class ILARegInst<RegisterClass rclass, Operand immtype, PatLeaf xform>:
414 ILAInst<(outs rclass:$rT), (ins immtype:$val),
415 [(set rclass:$rT, xform:$val)]>;
417 multiclass ImmLoadAddress
419 def v2i64: ILAVecInst<v2i64, u18imm, v2i64Uns18Imm>;
420 def v4i32: ILAVecInst<v4i32, u18imm, v4i32Uns18Imm>;
422 def r64: ILARegInst<R64C, u18imm_i64, imm18>;
423 def r32: ILARegInst<R32C, u18imm, imm18>;
424 def f32: ILARegInst<R32FP, f18imm, fpimm18>;
425 def f64: ILARegInst<R64FP, f18imm_f64, fpimm18>;
427 def hi: ILARegInst<R32C, symbolHi, imm18>;
428 def lo: ILARegInst<R32C, symbolLo, imm18>;
430 def lsa: ILAInst<(outs R32C:$rT), (ins symbolLSA:$val),
434 defm ILA : ImmLoadAddress;
436 // Immediate OR, Halfword Lower: The "other" part of loading large constants
437 // into 32-bit registers. See the anonymous pattern Pat<(i32 imm:$imm), ...>
438 // Note that these are really two operand instructions, but they're encoded
439 // as three operands with the first two arguments tied-to each other.
441 class IOHLInst<dag OOL, dag IOL, list<dag> pattern>:
442 RI16Form<0b100000110, OOL, IOL, "iohl\t$rT, $val",
444 RegConstraint<"$rS = $rT">,
447 class IOHLVecInst<ValueType vectype, Operand immtype /* , PatLeaf xform */>:
448 IOHLInst<(outs VECREG:$rT), (ins VECREG:$rS, immtype:$val),
451 class IOHLRegInst<RegisterClass rclass, Operand immtype /* , PatLeaf xform */>:
452 IOHLInst<(outs rclass:$rT), (ins rclass:$rS, immtype:$val),
455 multiclass ImmOrHalfwordLower
457 def v2i64: IOHLVecInst<v2i64, u16imm_i64>;
458 def v4i32: IOHLVecInst<v4i32, u16imm_i32>;
460 def r32: IOHLRegInst<R32C, i32imm>;
461 def f32: IOHLRegInst<R32FP, f32imm>;
463 def lo: IOHLRegInst<R32C, symbolLo>;
466 defm IOHL: ImmOrHalfwordLower;
468 // Form select mask for bytes using immediate, used in conjunction with the
471 class FSMBIVec<ValueType vectype>:
472 RI16Form<0b101001100, (outs VECREG:$rT), (ins u16imm:$val),
475 [(set (vectype VECREG:$rT), (SPUselmask (i16 immU16:$val)))]>;
477 multiclass FormSelectMaskBytesImm
479 def v16i8: FSMBIVec<v16i8>;
480 def v8i16: FSMBIVec<v8i16>;
481 def v4i32: FSMBIVec<v4i32>;
482 def v2i64: FSMBIVec<v2i64>;
485 defm FSMBI : FormSelectMaskBytesImm;
487 // fsmb: Form select mask for bytes. N.B. Input operand, $rA, is 16-bits
488 class FSMBInst<dag OOL, dag IOL, list<dag> pattern>:
489 RRForm_1<0b01101101100, OOL, IOL, "fsmb\t$rT, $rA", SelectOp,
492 class FSMBRegInst<RegisterClass rclass, ValueType vectype>:
493 FSMBInst<(outs VECREG:$rT), (ins rclass:$rA),
494 [(set (vectype VECREG:$rT), (SPUselmask rclass:$rA))]>;
496 class FSMBVecInst<ValueType vectype>:
497 FSMBInst<(outs VECREG:$rT), (ins VECREG:$rA),
498 [(set (vectype VECREG:$rT),
499 (SPUselmask (vectype VECREG:$rA)))]>;
501 multiclass FormSelectMaskBits {
502 def v16i8_r16: FSMBRegInst<R16C, v16i8>;
503 def v16i8: FSMBVecInst<v16i8>;
506 defm FSMB: FormSelectMaskBits;
508 // fsmh: Form select mask for halfwords. N.B., Input operand, $rA, is
509 // only 8-bits wide (even though it's input as 16-bits here)
511 class FSMHInst<dag OOL, dag IOL, list<dag> pattern>:
512 RRForm_1<0b10101101100, OOL, IOL, "fsmh\t$rT, $rA", SelectOp,
515 class FSMHRegInst<RegisterClass rclass, ValueType vectype>:
516 FSMHInst<(outs VECREG:$rT), (ins rclass:$rA),
517 [(set (vectype VECREG:$rT), (SPUselmask rclass:$rA))]>;
519 class FSMHVecInst<ValueType vectype>:
520 FSMHInst<(outs VECREG:$rT), (ins VECREG:$rA),
521 [(set (vectype VECREG:$rT),
522 (SPUselmask (vectype VECREG:$rA)))]>;
524 multiclass FormSelectMaskHalfword {
525 def v8i16_r16: FSMHRegInst<R16C, v8i16>;
526 def v8i16: FSMHVecInst<v8i16>;
529 defm FSMH: FormSelectMaskHalfword;
531 // fsm: Form select mask for words. Like the other fsm* instructions,
532 // only the lower 4 bits of $rA are significant.
534 class FSMInst<dag OOL, dag IOL, list<dag> pattern>:
535 RRForm_1<0b00101101100, OOL, IOL, "fsm\t$rT, $rA", SelectOp,
538 class FSMRegInst<ValueType vectype, RegisterClass rclass>:
539 FSMInst<(outs VECREG:$rT), (ins rclass:$rA),
540 [(set (vectype VECREG:$rT), (SPUselmask rclass:$rA))]>;
542 class FSMVecInst<ValueType vectype>:
543 FSMInst<(outs VECREG:$rT), (ins VECREG:$rA),
544 [(set (vectype VECREG:$rT), (SPUselmask (vectype VECREG:$rA)))]>;
546 multiclass FormSelectMaskWord {
547 def v4i32: FSMVecInst<v4i32>;
549 def r32 : FSMRegInst<v4i32, R32C>;
550 def r16 : FSMRegInst<v4i32, R16C>;
553 defm FSM : FormSelectMaskWord;
555 // Special case when used for i64 math operations
556 multiclass FormSelectMaskWord64 {
557 def r32 : FSMRegInst<v2i64, R32C>;
558 def r16 : FSMRegInst<v2i64, R16C>;
561 defm FSM64 : FormSelectMaskWord64;
563 //===----------------------------------------------------------------------===//
564 // Integer and Logical Operations:
565 //===----------------------------------------------------------------------===//
568 RRForm<0b00010011000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
569 "ah\t$rT, $rA, $rB", IntegerOp,
570 [(set (v8i16 VECREG:$rT), (int_spu_si_ah VECREG:$rA, VECREG:$rB))]>;
572 def : Pat<(add (v8i16 VECREG:$rA), (v8i16 VECREG:$rB)),
573 (AHv8i16 VECREG:$rA, VECREG:$rB)>;
576 RRForm<0b00010011000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
577 "ah\t$rT, $rA, $rB", IntegerOp,
578 [(set R16C:$rT, (add R16C:$rA, R16C:$rB))]>;
581 RI10Form<0b10111000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
582 "ahi\t$rT, $rA, $val", IntegerOp,
583 [(set (v8i16 VECREG:$rT), (add (v8i16 VECREG:$rA),
584 v8i16SExt10Imm:$val))]>;
587 RI10Form<0b10111000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
588 "ahi\t$rT, $rA, $val", IntegerOp,
589 [(set R16C:$rT, (add R16C:$rA, i16ImmSExt10:$val))]>;
591 // v4i32, i32 add instruction:
593 class AInst<dag OOL, dag IOL, list<dag> pattern>:
594 RRForm<0b00000011000, OOL, IOL,
595 "a\t$rT, $rA, $rB", IntegerOp,
598 class AVecInst<ValueType vectype>:
599 AInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
600 [(set (vectype VECREG:$rT), (add (vectype VECREG:$rA),
601 (vectype VECREG:$rB)))]>;
603 class ARegInst<RegisterClass rclass>:
604 AInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
605 [(set rclass:$rT, (add rclass:$rA, rclass:$rB))]>;
607 multiclass AddInstruction {
608 def v4i32: AVecInst<v4i32>;
609 def v16i8: AVecInst<v16i8>;
611 def r32: ARegInst<R32C>;
614 defm A : AddInstruction;
616 class AIInst<dag OOL, dag IOL, list<dag> pattern>:
617 RI10Form<0b00111000, OOL, IOL,
618 "ai\t$rT, $rA, $val", IntegerOp,
621 class AIVecInst<ValueType vectype, PatLeaf immpred>:
622 AIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
623 [(set (vectype VECREG:$rT), (add (vectype VECREG:$rA), immpred:$val))]>;
625 class AIFPVecInst<ValueType vectype, PatLeaf immpred>:
626 AIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
629 class AIRegInst<RegisterClass rclass, PatLeaf immpred>:
630 AIInst<(outs rclass:$rT), (ins rclass:$rA, s10imm_i32:$val),
631 [(set rclass:$rT, (add rclass:$rA, immpred:$val))]>;
633 // This is used to add epsilons to floating point numbers in the f32 fdiv code:
634 class AIFPInst<RegisterClass rclass, PatLeaf immpred>:
635 AIInst<(outs rclass:$rT), (ins rclass:$rA, s10imm_i32:$val),
638 multiclass AddImmediate {
639 def v4i32: AIVecInst<v4i32, v4i32SExt10Imm>;
641 def r32: AIRegInst<R32C, i32ImmSExt10>;
643 def v4f32: AIFPVecInst<v4f32, v4i32SExt10Imm>;
644 def f32: AIFPInst<R32FP, i32ImmSExt10>;
647 defm AI : AddImmediate;
650 RRForm<0b00010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
651 "sfh\t$rT, $rA, $rB", IntegerOp,
652 [(set (v8i16 VECREG:$rT), (sub (v8i16 VECREG:$rA),
653 (v8i16 VECREG:$rB)))]>;
656 RRForm<0b00010010000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
657 "sfh\t$rT, $rA, $rB", IntegerOp,
658 [(set R16C:$rT, (sub R16C:$rA, R16C:$rB))]>;
661 RI10Form<0b10110000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
662 "sfhi\t$rT, $rA, $val", IntegerOp,
663 [(set (v8i16 VECREG:$rT), (sub v8i16SExt10Imm:$val,
664 (v8i16 VECREG:$rA)))]>;
666 def SFHIr16 : RI10Form<0b10110000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
667 "sfhi\t$rT, $rA, $val", IntegerOp,
668 [(set R16C:$rT, (sub i16ImmSExt10:$val, R16C:$rA))]>;
670 def SFvec : RRForm<0b00000010000, (outs VECREG:$rT),
671 (ins VECREG:$rA, VECREG:$rB),
672 "sf\t$rT, $rA, $rB", IntegerOp,
673 [(set (v4i32 VECREG:$rT), (sub (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
675 def SFr32 : RRForm<0b00000010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
676 "sf\t$rT, $rA, $rB", IntegerOp,
677 [(set R32C:$rT, (sub R32C:$rA, R32C:$rB))]>;
680 RI10Form<0b00110000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
681 "sfi\t$rT, $rA, $val", IntegerOp,
682 [(set (v4i32 VECREG:$rT), (sub v4i32SExt10Imm:$val,
683 (v4i32 VECREG:$rA)))]>;
685 def SFIr32 : RI10Form<0b00110000, (outs R32C:$rT),
686 (ins R32C:$rA, s10imm_i32:$val),
687 "sfi\t$rT, $rA, $val", IntegerOp,
688 [(set R32C:$rT, (sub i32ImmSExt10:$val, R32C:$rA))]>;
690 // ADDX: only available in vector form, doesn't match a pattern.
691 class ADDXInst<dag OOL, dag IOL, list<dag> pattern>:
692 RRForm<0b00000010110, OOL, IOL,
693 "addx\t$rT, $rA, $rB",
696 class ADDXVecInst<ValueType vectype>:
697 ADDXInst<(outs VECREG:$rT),
698 (ins VECREG:$rA, VECREG:$rB, VECREG:$rCarry),
700 RegConstraint<"$rCarry = $rT">,
703 class ADDXRegInst<RegisterClass rclass>:
704 ADDXInst<(outs rclass:$rT),
705 (ins rclass:$rA, rclass:$rB, rclass:$rCarry),
707 RegConstraint<"$rCarry = $rT">,
710 multiclass AddExtended {
711 def v2i64 : ADDXVecInst<v2i64>;
712 def v4i32 : ADDXVecInst<v4i32>;
713 def r64 : ADDXRegInst<R64C>;
714 def r32 : ADDXRegInst<R32C>;
717 defm ADDX : AddExtended;
719 // CG: Generate carry for add
720 class CGInst<dag OOL, dag IOL, list<dag> pattern>:
721 RRForm<0b01000011000, OOL, IOL,
725 class CGVecInst<ValueType vectype>:
726 CGInst<(outs VECREG:$rT),
727 (ins VECREG:$rA, VECREG:$rB),
730 class CGRegInst<RegisterClass rclass>:
731 CGInst<(outs rclass:$rT),
732 (ins rclass:$rA, rclass:$rB),
735 multiclass CarryGenerate {
736 def v2i64 : CGVecInst<v2i64>;
737 def v4i32 : CGVecInst<v4i32>;
738 def r64 : CGRegInst<R64C>;
739 def r32 : CGRegInst<R32C>;
742 defm CG : CarryGenerate;
744 // SFX: Subract from, extended. This is used in conjunction with BG to subtract
745 // with carry (borrow, in this case)
746 class SFXInst<dag OOL, dag IOL, list<dag> pattern>:
747 RRForm<0b10000010110, OOL, IOL,
748 "sfx\t$rT, $rA, $rB",
751 class SFXVecInst<ValueType vectype>:
752 SFXInst<(outs VECREG:$rT),
753 (ins VECREG:$rA, VECREG:$rB, VECREG:$rCarry),
755 RegConstraint<"$rCarry = $rT">,
758 class SFXRegInst<RegisterClass rclass>:
759 SFXInst<(outs rclass:$rT),
760 (ins rclass:$rA, rclass:$rB, rclass:$rCarry),
762 RegConstraint<"$rCarry = $rT">,
765 multiclass SubtractExtended {
766 def v2i64 : SFXVecInst<v2i64>;
767 def v4i32 : SFXVecInst<v4i32>;
768 def r64 : SFXRegInst<R64C>;
769 def r32 : SFXRegInst<R32C>;
772 defm SFX : SubtractExtended;
774 // BG: only available in vector form, doesn't match a pattern.
775 class BGInst<dag OOL, dag IOL, list<dag> pattern>:
776 RRForm<0b01000010000, OOL, IOL,
780 class BGVecInst<ValueType vectype>:
781 BGInst<(outs VECREG:$rT),
782 (ins VECREG:$rA, VECREG:$rB),
785 class BGRegInst<RegisterClass rclass>:
786 BGInst<(outs rclass:$rT),
787 (ins rclass:$rA, rclass:$rB),
790 multiclass BorrowGenerate {
791 def v4i32 : BGVecInst<v4i32>;
792 def v2i64 : BGVecInst<v2i64>;
793 def r64 : BGRegInst<R64C>;
794 def r32 : BGRegInst<R32C>;
797 defm BG : BorrowGenerate;
799 // BGX: Borrow generate, extended.
801 RRForm<0b11000010110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB,
803 "bgx\t$rT, $rA, $rB", IntegerOp,
805 RegConstraint<"$rCarry = $rT">,
808 // Halfword multiply variants:
809 // N.B: These can be used to build up larger quantities (16x16 -> 32)
812 RRForm<0b00100011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
813 "mpy\t$rT, $rA, $rB", IntegerMulDiv,
817 RRForm<0b00100011110, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
818 "mpy\t$rT, $rA, $rB", IntegerMulDiv,
819 [(set R16C:$rT, (mul R16C:$rA, R16C:$rB))]>;
821 // Unsigned 16-bit multiply:
823 class MPYUInst<dag OOL, dag IOL, list<dag> pattern>:
824 RRForm<0b00110011110, OOL, IOL,
825 "mpyu\t$rT, $rA, $rB", IntegerMulDiv,
829 MPYUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
833 MPYUInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB),
834 [(set R32C:$rT, (mul (zext R16C:$rA), (zext R16C:$rB)))]>;
837 MPYUInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
840 // mpyi: multiply 16 x s10imm -> 32 result.
842 class MPYIInst<dag OOL, dag IOL, list<dag> pattern>:
843 RI10Form<0b00101110, OOL, IOL,
844 "mpyi\t$rT, $rA, $val", IntegerMulDiv,
848 MPYIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
849 [(set (v8i16 VECREG:$rT),
850 (mul (v8i16 VECREG:$rA), v8i16SExt10Imm:$val))]>;
853 MPYIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
854 [(set R16C:$rT, (mul R16C:$rA, i16ImmSExt10:$val))]>;
856 // mpyui: same issues as other multiplies, plus, this doesn't match a
857 // pattern... but may be used during target DAG selection or lowering
859 class MPYUIInst<dag OOL, dag IOL, list<dag> pattern>:
860 RI10Form<0b10101110, OOL, IOL,
861 "mpyui\t$rT, $rA, $val", IntegerMulDiv,
865 MPYUIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
869 MPYUIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
872 // mpya: 16 x 16 + 16 -> 32 bit result
873 class MPYAInst<dag OOL, dag IOL, list<dag> pattern>:
874 RRRForm<0b0011, OOL, IOL,
875 "mpya\t$rT, $rA, $rB, $rC", IntegerMulDiv,
879 MPYAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
880 [(set (v4i32 VECREG:$rT),
881 (add (v4i32 (bitconvert (mul (v8i16 VECREG:$rA),
882 (v8i16 VECREG:$rB)))),
883 (v4i32 VECREG:$rC)))]>;
886 MPYAInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB, R32C:$rC),
887 [(set R32C:$rT, (add (sext (mul R16C:$rA, R16C:$rB)),
891 MPYAInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB, R32C:$rC),
892 [(set R32C:$rT, (add (mul (sext R16C:$rA), (sext R16C:$rB)),
895 def MPYAr32_sextinreg:
896 MPYAInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB, R32C:$rC),
897 [(set R32C:$rT, (add (mul (sext_inreg R32C:$rA, i16),
898 (sext_inreg R32C:$rB, i16)),
901 // mpyh: multiply high, used to synthesize 32-bit multiplies
902 class MPYHInst<dag OOL, dag IOL, list<dag> pattern>:
903 RRForm<0b10100011110, OOL, IOL,
904 "mpyh\t$rT, $rA, $rB", IntegerMulDiv,
908 MPYHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
912 MPYHInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
915 // mpys: multiply high and shift right (returns the top half of
916 // a 16-bit multiply, sign extended to 32 bits.)
918 class MPYSInst<dag OOL, dag IOL>:
919 RRForm<0b11100011110, OOL, IOL,
920 "mpys\t$rT, $rA, $rB", IntegerMulDiv,
924 MPYSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
927 MPYSInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB)>;
929 // mpyhh: multiply high-high (returns the 32-bit result from multiplying
930 // the top 16 bits of the $rA, $rB)
932 class MPYHHInst<dag OOL, dag IOL>:
933 RRForm<0b01100011110, OOL, IOL,
934 "mpyhh\t$rT, $rA, $rB", IntegerMulDiv,
938 MPYHHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
941 MPYHHInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
943 // mpyhha: Multiply high-high, add to $rT:
945 class MPYHHAInst<dag OOL, dag IOL>:
946 RRForm<0b01100010110, OOL, IOL,
947 "mpyhha\t$rT, $rA, $rB", IntegerMulDiv,
951 MPYHHAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
954 MPYHHAInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
956 // mpyhhu: Multiply high-high, unsigned, e.g.:
958 // +-------+-------+ +-------+-------+ +---------+
959 // | a0 . a1 | x | b0 . b1 | = | a0 x b0 |
960 // +-------+-------+ +-------+-------+ +---------+
962 // where a0, b0 are the upper 16 bits of the 32-bit word
964 class MPYHHUInst<dag OOL, dag IOL>:
965 RRForm<0b01110011110, OOL, IOL,
966 "mpyhhu\t$rT, $rA, $rB", IntegerMulDiv,
970 MPYHHUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
973 MPYHHUInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
975 // mpyhhau: Multiply high-high, unsigned
977 class MPYHHAUInst<dag OOL, dag IOL>:
978 RRForm<0b01110010110, OOL, IOL,
979 "mpyhhau\t$rT, $rA, $rB", IntegerMulDiv,
983 MPYHHAUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
986 MPYHHAUInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
988 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
989 // clz: Count leading zeroes
990 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
991 class CLZInst<dag OOL, dag IOL, list<dag> pattern>:
992 RRForm_1<0b10100101010, OOL, IOL, "clz\t$rT, $rA",
995 class CLZRegInst<RegisterClass rclass>:
996 CLZInst<(outs rclass:$rT), (ins rclass:$rA),
997 [(set rclass:$rT, (ctlz rclass:$rA))]>;
999 class CLZVecInst<ValueType vectype>:
1000 CLZInst<(outs VECREG:$rT), (ins VECREG:$rA),
1001 [(set (vectype VECREG:$rT), (ctlz (vectype VECREG:$rA)))]>;
1003 multiclass CountLeadingZeroes {
1004 def v4i32 : CLZVecInst<v4i32>;
1005 def r32 : CLZRegInst<R32C>;
1008 defm CLZ : CountLeadingZeroes;
1010 // cntb: Count ones in bytes (aka "population count")
1012 // NOTE: This instruction is really a vector instruction, but the custom
1013 // lowering code uses it in unorthodox ways to support CTPOP for other
1017 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
1018 "cntb\t$rT, $rA", IntegerOp,
1019 [(set (v16i8 VECREG:$rT), (SPUcntb (v16i8 VECREG:$rA)))]>;
1022 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
1023 "cntb\t$rT, $rA", IntegerOp,
1024 [(set (v8i16 VECREG:$rT), (SPUcntb (v8i16 VECREG:$rA)))]>;
1027 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
1028 "cntb\t$rT, $rA", IntegerOp,
1029 [(set (v4i32 VECREG:$rT), (SPUcntb (v4i32 VECREG:$rA)))]>;
1031 // gbb: Gather the low order bits from each byte in $rA into a single 16-bit
1032 // quantity stored into $rT's slot 0, upper 16 bits are zeroed, as are
1035 // Note: This instruction "pairs" with the fsmb instruction for all of the
1036 // various types defined here.
1038 // Note 2: The "VecInst" and "RegInst" forms refer to the result being either
1039 // a vector or register.
1041 class GBBInst<dag OOL, dag IOL, list<dag> pattern>:
1042 RRForm_1<0b01001101100, OOL, IOL, "gbb\t$rT, $rA", GatherOp, pattern>;
1044 class GBBRegInst<RegisterClass rclass, ValueType vectype>:
1045 GBBInst<(outs rclass:$rT), (ins VECREG:$rA),
1046 [/* no pattern */]>;
1048 class GBBVecInst<ValueType vectype>:
1049 GBBInst<(outs VECREG:$rT), (ins VECREG:$rA),
1050 [/* no pattern */]>;
1052 multiclass GatherBitsFromBytes {
1053 def v16i8_r32: GBBRegInst<R32C, v16i8>;
1054 def v16i8_r16: GBBRegInst<R16C, v16i8>;
1055 def v16i8: GBBVecInst<v16i8>;
1058 defm GBB: GatherBitsFromBytes;
1060 // gbh: Gather all low order bits from each halfword in $rA into a single
1061 // 8-bit quantity stored in $rT's slot 0, with the upper bits of $rT set to 0
1062 // and slots 1-3 also set to 0.
1064 // See notes for GBBInst, above.
1066 class GBHInst<dag OOL, dag IOL, list<dag> pattern>:
1067 RRForm_1<0b10001101100, OOL, IOL, "gbh\t$rT, $rA", GatherOp,
1070 class GBHRegInst<RegisterClass rclass, ValueType vectype>:
1071 GBHInst<(outs rclass:$rT), (ins VECREG:$rA),
1072 [/* no pattern */]>;
1074 class GBHVecInst<ValueType vectype>:
1075 GBHInst<(outs VECREG:$rT), (ins VECREG:$rA),
1076 [/* no pattern */]>;
1078 multiclass GatherBitsHalfword {
1079 def v8i16_r32: GBHRegInst<R32C, v8i16>;
1080 def v8i16_r16: GBHRegInst<R16C, v8i16>;
1081 def v8i16: GBHVecInst<v8i16>;
1084 defm GBH: GatherBitsHalfword;
1086 // gb: Gather all low order bits from each word in $rA into a single
1087 // 4-bit quantity stored in $rT's slot 0, upper bits in $rT set to 0,
1088 // as well as slots 1-3.
1090 // See notes for gbb, above.
1092 class GBInst<dag OOL, dag IOL, list<dag> pattern>:
1093 RRForm_1<0b00001101100, OOL, IOL, "gb\t$rT, $rA", GatherOp,
1096 class GBRegInst<RegisterClass rclass, ValueType vectype>:
1097 GBInst<(outs rclass:$rT), (ins VECREG:$rA),
1098 [/* no pattern */]>;
1100 class GBVecInst<ValueType vectype>:
1101 GBInst<(outs VECREG:$rT), (ins VECREG:$rA),
1102 [/* no pattern */]>;
1104 multiclass GatherBitsWord {
1105 def v4i32_r32: GBRegInst<R32C, v4i32>;
1106 def v4i32_r16: GBRegInst<R16C, v4i32>;
1107 def v4i32: GBVecInst<v4i32>;
1110 defm GB: GatherBitsWord;
1112 // avgb: average bytes
1114 RRForm<0b11001011000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1115 "avgb\t$rT, $rA, $rB", ByteOp,
1118 // absdb: absolute difference of bytes
1120 RRForm<0b11001010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1121 "absdb\t$rT, $rA, $rB", ByteOp,
1124 // sumb: sum bytes into halfwords
1126 RRForm<0b11001010010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1127 "sumb\t$rT, $rA, $rB", ByteOp,
1130 // Sign extension operations:
1131 class XSBHInst<dag OOL, dag IOL, list<dag> pattern>:
1132 RRForm_1<0b01101101010, OOL, IOL,
1133 "xsbh\t$rDst, $rSrc",
1134 IntegerOp, pattern>;
1136 class XSBHInRegInst<RegisterClass rclass, list<dag> pattern>:
1137 XSBHInst<(outs rclass:$rDst), (ins rclass:$rSrc),
1140 multiclass ExtendByteHalfword {
1141 def v16i8: XSBHInst<(outs VECREG:$rDst), (ins VECREG:$rSrc),
1143 /*(set (v8i16 VECREG:$rDst), (sext (v8i16 VECREG:$rSrc)))*/]>;
1144 def r8: XSBHInst<(outs R16C:$rDst), (ins R8C:$rSrc),
1145 [(set R16C:$rDst, (sext R8C:$rSrc))]>;
1146 def r16: XSBHInRegInst<R16C,
1147 [(set R16C:$rDst, (sext_inreg R16C:$rSrc, i8))]>;
1149 // 32-bit form for XSBH: used to sign extend 8-bit quantities to 16-bit
1150 // quantities to 32-bit quantities via a 32-bit register (see the sext 8->32
1151 // pattern below). Intentionally doesn't match a pattern because we want the
1152 // sext 8->32 pattern to do the work for us, namely because we need the extra
1154 def r32: XSBHInRegInst<R32C, [/* no pattern */]>;
1156 // Same as the 32-bit version, but for i64
1157 def r64: XSBHInRegInst<R64C, [/* no pattern */]>;
1160 defm XSBH : ExtendByteHalfword;
1162 // Sign extend halfwords to words:
1164 class XSHWInst<dag OOL, dag IOL, list<dag> pattern>:
1165 RRForm_1<0b01101101010, OOL, IOL, "xshw\t$rDest, $rSrc",
1166 IntegerOp, pattern>;
1168 class XSHWVecInst<ValueType in_vectype, ValueType out_vectype>:
1169 XSHWInst<(outs VECREG:$rDest), (ins VECREG:$rSrc),
1170 [(set (out_vectype VECREG:$rDest),
1171 (sext (in_vectype VECREG:$rSrc)))]>;
1173 class XSHWInRegInst<RegisterClass rclass, list<dag> pattern>:
1174 XSHWInst<(outs rclass:$rDest), (ins rclass:$rSrc),
1177 class XSHWRegInst<RegisterClass rclass>:
1178 XSHWInst<(outs rclass:$rDest), (ins R16C:$rSrc),
1179 [(set rclass:$rDest, (sext R16C:$rSrc))]>;
1181 multiclass ExtendHalfwordWord {
1182 def v4i32: XSHWVecInst<v4i32, v8i16>;
1184 def r16: XSHWRegInst<R32C>;
1186 def r32: XSHWInRegInst<R32C,
1187 [(set R32C:$rDest, (sext_inreg R32C:$rSrc, i16))]>;
1188 def r64: XSHWInRegInst<R64C, [/* no pattern */]>;
1191 defm XSHW : ExtendHalfwordWord;
1193 // Sign-extend words to doublewords (32->64 bits)
1195 class XSWDInst<dag OOL, dag IOL, list<dag> pattern>:
1196 RRForm_1<0b01100101010, OOL, IOL, "xswd\t$rDst, $rSrc",
1197 IntegerOp, pattern>;
1199 class XSWDVecInst<ValueType in_vectype, ValueType out_vectype>:
1200 XSWDInst<(outs VECREG:$rDst), (ins VECREG:$rSrc),
1201 [/*(set (out_vectype VECREG:$rDst),
1202 (sext (out_vectype VECREG:$rSrc)))*/]>;
1204 class XSWDRegInst<RegisterClass in_rclass, RegisterClass out_rclass>:
1205 XSWDInst<(outs out_rclass:$rDst), (ins in_rclass:$rSrc),
1206 [(set out_rclass:$rDst, (sext in_rclass:$rSrc))]>;
1208 multiclass ExtendWordToDoubleWord {
1209 def v2i64: XSWDVecInst<v4i32, v2i64>;
1210 def r64: XSWDRegInst<R32C, R64C>;
1212 def r64_inreg: XSWDInst<(outs R64C:$rDst), (ins R64C:$rSrc),
1213 [(set R64C:$rDst, (sext_inreg R64C:$rSrc, i32))]>;
1216 defm XSWD : ExtendWordToDoubleWord;
1220 class ANDInst<dag OOL, dag IOL, list<dag> pattern> :
1221 RRForm<0b10000011000, OOL, IOL, "and\t$rT, $rA, $rB",
1222 IntegerOp, pattern>;
1224 class ANDVecInst<ValueType vectype>:
1225 ANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1226 [(set (vectype VECREG:$rT), (and (vectype VECREG:$rA),
1227 (vectype VECREG:$rB)))]>;
1229 class ANDRegInst<RegisterClass rclass>:
1230 ANDInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1231 [(set rclass:$rT, (and rclass:$rA, rclass:$rB))]>;
1233 multiclass BitwiseAnd
1235 def v16i8: ANDVecInst<v16i8>;
1236 def v8i16: ANDVecInst<v8i16>;
1237 def v4i32: ANDVecInst<v4i32>;
1238 def v2i64: ANDVecInst<v2i64>;
1240 def r128: ANDRegInst<GPRC>;
1241 def r64: ANDRegInst<R64C>;
1242 def r32: ANDRegInst<R32C>;
1243 def r16: ANDRegInst<R16C>;
1244 def r8: ANDRegInst<R8C>;
1246 //===---------------------------------------------
1247 // Special instructions to perform the fabs instruction
1248 def fabs32: ANDInst<(outs R32FP:$rT), (ins R32FP:$rA, R32C:$rB),
1249 [/* Intentionally does not match a pattern */]>;
1251 def fabs64: ANDInst<(outs R64FP:$rT), (ins R64FP:$rA, R64C:$rB),
1252 [/* Intentionally does not match a pattern */]>;
1254 def fabsvec: ANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1255 [/* Intentionally does not match a pattern */]>;
1257 //===---------------------------------------------
1259 // Hacked form of AND to zero-extend 16-bit quantities to 32-bit
1260 // quantities -- see 16->32 zext pattern.
1262 // This pattern is somewhat artificial, since it might match some
1263 // compiler generated pattern but it is unlikely to do so.
1265 def i16i32: ANDInst<(outs R32C:$rT), (ins R16C:$rA, R32C:$rB),
1266 [(set R32C:$rT, (and (zext R16C:$rA), R32C:$rB))]>;
1269 defm AND : BitwiseAnd;
1271 // N.B.: vnot_conv is one of those special target selection pattern fragments,
1272 // in which we expect there to be a bit_convert on the constant. Bear in mind
1273 // that llvm translates "not <reg>" to "xor <reg>, -1" (or in this case, a
1274 // constant -1 vector.)
1276 class ANDCInst<dag OOL, dag IOL, list<dag> pattern>:
1277 RRForm<0b10000011010, OOL, IOL, "andc\t$rT, $rA, $rB",
1278 IntegerOp, pattern>;
1280 class ANDCVecInst<ValueType vectype, PatFrag vnot_frag = vnot>:
1281 ANDCInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1282 [(set (vectype VECREG:$rT),
1283 (and (vectype VECREG:$rA),
1284 (vnot_frag (vectype VECREG:$rB))))]>;
1286 class ANDCRegInst<RegisterClass rclass>:
1287 ANDCInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1288 [(set rclass:$rT, (and rclass:$rA, (not rclass:$rB)))]>;
1290 multiclass AndComplement
1292 def v16i8: ANDCVecInst<v16i8>;
1293 def v8i16: ANDCVecInst<v8i16>;
1294 def v4i32: ANDCVecInst<v4i32>;
1295 def v2i64: ANDCVecInst<v2i64>;
1297 def r128: ANDCRegInst<GPRC>;
1298 def r64: ANDCRegInst<R64C>;
1299 def r32: ANDCRegInst<R32C>;
1300 def r16: ANDCRegInst<R16C>;
1301 def r8: ANDCRegInst<R8C>;
1303 // Sometimes, the xor pattern has a bitcast constant:
1304 def v16i8_conv: ANDCVecInst<v16i8, vnot_conv>;
1307 defm ANDC : AndComplement;
1309 class ANDBIInst<dag OOL, dag IOL, list<dag> pattern>:
1310 RI10Form<0b01101000, OOL, IOL, "andbi\t$rT, $rA, $val",
1313 multiclass AndByteImm
1315 def v16i8: ANDBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1316 [(set (v16i8 VECREG:$rT),
1317 (and (v16i8 VECREG:$rA),
1318 (v16i8 v16i8U8Imm:$val)))]>;
1320 def r8: ANDBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1321 [(set R8C:$rT, (and R8C:$rA, immU8:$val))]>;
1324 defm ANDBI : AndByteImm;
1326 class ANDHIInst<dag OOL, dag IOL, list<dag> pattern> :
1327 RI10Form<0b10101000, OOL, IOL, "andhi\t$rT, $rA, $val",
1330 multiclass AndHalfwordImm
1332 def v8i16: ANDHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
1333 [(set (v8i16 VECREG:$rT),
1334 (and (v8i16 VECREG:$rA), v8i16SExt10Imm:$val))]>;
1336 def r16: ANDHIInst<(outs R16C:$rT), (ins R16C:$rA, u10imm:$val),
1337 [(set R16C:$rT, (and R16C:$rA, i16ImmUns10:$val))]>;
1339 // Zero-extend i8 to i16:
1340 def i8i16: ANDHIInst<(outs R16C:$rT), (ins R8C:$rA, u10imm:$val),
1341 [(set R16C:$rT, (and (zext R8C:$rA), i16ImmUns10:$val))]>;
1344 defm ANDHI : AndHalfwordImm;
1346 class ANDIInst<dag OOL, dag IOL, list<dag> pattern> :
1347 RI10Form<0b00101000, OOL, IOL, "andi\t$rT, $rA, $val",
1348 IntegerOp, pattern>;
1350 multiclass AndWordImm
1352 def v4i32: ANDIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
1353 [(set (v4i32 VECREG:$rT),
1354 (and (v4i32 VECREG:$rA), v4i32SExt10Imm:$val))]>;
1356 def r32: ANDIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
1357 [(set R32C:$rT, (and R32C:$rA, i32ImmSExt10:$val))]>;
1359 // Hacked form of ANDI to zero-extend i8 quantities to i32. See the zext 8->32
1361 def i8i32: ANDIInst<(outs R32C:$rT), (ins R8C:$rA, s10imm_i32:$val),
1363 (and (zext R8C:$rA), i32ImmSExt10:$val))]>;
1365 // Hacked form of ANDI to zero-extend i16 quantities to i32. See the
1366 // zext 16->32 pattern below.
1368 // Note that this pattern is somewhat artificial, since it might match
1369 // something the compiler generates but is unlikely to occur in practice.
1370 def i16i32: ANDIInst<(outs R32C:$rT), (ins R16C:$rA, s10imm_i32:$val),
1372 (and (zext R16C:$rA), i32ImmSExt10:$val))]>;
1375 defm ANDI : AndWordImm;
1377 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1378 // Bitwise OR group:
1379 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1381 // Bitwise "or" (N.B.: These are also register-register copy instructions...)
1382 class ORInst<dag OOL, dag IOL, list<dag> pattern>:
1383 RRForm<0b10000010000, OOL, IOL, "or\t$rT, $rA, $rB",
1384 IntegerOp, pattern>;
1386 class ORVecInst<ValueType vectype>:
1387 ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1388 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1389 (vectype VECREG:$rB)))]>;
1391 class ORRegInst<RegisterClass rclass>:
1392 ORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1393 [(set rclass:$rT, (or rclass:$rA, rclass:$rB))]>;
1395 // ORCvtForm: OR conversion form
1397 // This is used to "convert" the preferred slot to its vector equivalent, as
1398 // well as convert a vector back to its preferred slot.
1400 // These are effectively no-ops, but need to exist for proper type conversion
1401 // and type coercion.
1403 class ORCvtForm<dag OOL, dag IOL, list<dag> pattern = [/* no pattern */]>
1404 : SPUInstr<OOL, IOL, "or\t$rT, $rA, $rA", IntegerOp> {
1408 let Pattern = pattern;
1410 let Inst{0-10} = 0b10000010000;
1411 let Inst{11-17} = RA;
1412 let Inst{18-24} = RA;
1413 let Inst{25-31} = RT;
1416 class ORPromoteScalar<RegisterClass rclass>:
1417 ORCvtForm<(outs VECREG:$rT), (ins rclass:$rA)>;
1419 class ORExtractElt<RegisterClass rclass>:
1420 ORCvtForm<(outs rclass:$rT), (ins VECREG:$rA)>;
1422 /* class ORCvtRegGPRC<RegisterClass rclass>:
1423 ORCvtForm<(outs GPRC:$rT), (ins rclass:$rA)>; */
1425 /* class ORCvtGPRCReg<RegisterClass rclass>:
1426 ORCvtForm<(outs rclass:$rT), (ins GPRC:$rA)>; */
1428 class ORCvtFormR32Reg<RegisterClass rclass, list<dag> pattern = [ ]>:
1429 ORCvtForm<(outs rclass:$rT), (ins R32C:$rA), pattern>;
1431 class ORCvtFormRegR32<RegisterClass rclass, list<dag> pattern = [ ]>:
1432 ORCvtForm<(outs R32C:$rT), (ins rclass:$rA), pattern>;
1434 class ORCvtFormR64Reg<RegisterClass rclass, list<dag> pattern = [ ]>:
1435 ORCvtForm<(outs rclass:$rT), (ins R64C:$rA), pattern>;
1437 class ORCvtFormRegR64<RegisterClass rclass, list<dag> pattern = [ ]>:
1438 ORCvtForm<(outs R64C:$rT), (ins rclass:$rA), pattern>;
1441 ORCvtForm<(outs VECREG:$rT), (ins GPRC:$rA)>;
1444 ORCvtForm<(outs GPRC:$rT), (ins VECREG:$rA)>;
1446 multiclass BitwiseOr
1448 def v16i8: ORVecInst<v16i8>;
1449 def v8i16: ORVecInst<v8i16>;
1450 def v4i32: ORVecInst<v4i32>;
1451 def v2i64: ORVecInst<v2i64>;
1453 def v4f32: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1454 [(set (v4f32 VECREG:$rT),
1455 (v4f32 (bitconvert (or (v4i32 VECREG:$rA),
1456 (v4i32 VECREG:$rB)))))]>;
1458 def v2f64: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1459 [(set (v2f64 VECREG:$rT),
1460 (v2f64 (bitconvert (or (v2i64 VECREG:$rA),
1461 (v2i64 VECREG:$rB)))))]>;
1463 def r128: ORRegInst<GPRC>;
1464 def r64: ORRegInst<R64C>;
1465 def r32: ORRegInst<R32C>;
1466 def r16: ORRegInst<R16C>;
1467 def r8: ORRegInst<R8C>;
1469 // OR instructions used to copy f32 and f64 registers.
1470 def f32: ORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
1471 [/* no pattern */]>;
1473 def f64: ORInst<(outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
1474 [/* no pattern */]>;
1476 // scalar->vector promotion, prefslot2vec:
1477 def v16i8_i8: ORPromoteScalar<R8C>;
1478 def v8i16_i16: ORPromoteScalar<R16C>;
1479 def v4i32_i32: ORPromoteScalar<R32C>;
1480 def v2i64_i64: ORPromoteScalar<R64C>;
1481 def v4f32_f32: ORPromoteScalar<R32FP>;
1482 def v2f64_f64: ORPromoteScalar<R64FP>;
1484 // vector->scalar demotion, vec2prefslot:
1485 def i8_v16i8: ORExtractElt<R8C>;
1486 def i16_v8i16: ORExtractElt<R16C>;
1487 def i32_v4i32: ORExtractElt<R32C>;
1488 def i64_v2i64: ORExtractElt<R64C>;
1489 def f32_v4f32: ORExtractElt<R32FP>;
1490 def f64_v2f64: ORExtractElt<R64FP>;
1492 // Conversion from vector to GPRC
1493 def i128_vec: ORCvtVecGPRC;
1495 // Conversion from GPRC to vector
1496 def vec_i128: ORCvtGPRCVec;
1499 // Conversion from register to GPRC
1500 def i128_r64: ORCvtRegGPRC<R64C>;
1501 def i128_f64: ORCvtRegGPRC<R64FP>;
1502 def i128_r32: ORCvtRegGPRC<R32C>;
1503 def i128_f32: ORCvtRegGPRC<R32FP>;
1504 def i128_r16: ORCvtRegGPRC<R16C>;
1505 def i128_r8: ORCvtRegGPRC<R8C>;
1507 // Conversion from GPRC to register
1508 def r64_i128: ORCvtGPRCReg<R64C>;
1509 def f64_i128: ORCvtGPRCReg<R64FP>;
1510 def r32_i128: ORCvtGPRCReg<R32C>;
1511 def f32_i128: ORCvtGPRCReg<R32FP>;
1512 def r16_i128: ORCvtGPRCReg<R16C>;
1513 def r8_i128: ORCvtGPRCReg<R8C>;
1516 // Conversion from register to R32C:
1517 def r32_r16: ORCvtFormRegR32<R16C>;
1518 def r32_r8: ORCvtFormRegR32<R8C>;
1520 // Conversion from R32C to register
1521 def r32_r16: ORCvtFormR32Reg<R16C>;
1522 def r32_r8: ORCvtFormR32Reg<R8C>;
1525 // Conversion from R64C to register:
1526 def r32_r64: ORCvtFormR64Reg<R32C>;
1527 // def r16_r64: ORCvtFormR64Reg<R16C>;
1528 // def r8_r64: ORCvtFormR64Reg<R8C>;
1530 // Conversion to R64C from register:
1531 def r64_r32: ORCvtFormRegR64<R32C>;
1532 // def r64_r16: ORCvtFormRegR64<R16C>;
1533 // def r64_r8: ORCvtFormRegR64<R8C>;
1535 // bitconvert patterns:
1536 def r32_f32: ORCvtFormR32Reg<R32FP,
1537 [(set R32FP:$rT, (bitconvert R32C:$rA))]>;
1538 def f32_r32: ORCvtFormRegR32<R32FP,
1539 [(set R32C:$rT, (bitconvert R32FP:$rA))]>;
1541 def r64_f64: ORCvtFormR64Reg<R64FP,
1542 [(set R64FP:$rT, (bitconvert R64C:$rA))]>;
1543 def f64_r64: ORCvtFormRegR64<R64FP,
1544 [(set R64C:$rT, (bitconvert R64FP:$rA))]>;
1547 defm OR : BitwiseOr;
1549 // scalar->vector promotion patterns (preferred slot to vector):
1550 def : Pat<(v16i8 (SPUprefslot2vec R8C:$rA)),
1551 (ORv16i8_i8 R8C:$rA)>;
1553 def : Pat<(v8i16 (SPUprefslot2vec R16C:$rA)),
1554 (ORv8i16_i16 R16C:$rA)>;
1556 def : Pat<(v4i32 (SPUprefslot2vec R32C:$rA)),
1557 (ORv4i32_i32 R32C:$rA)>;
1559 def : Pat<(v2i64 (SPUprefslot2vec R64C:$rA)),
1560 (ORv2i64_i64 R64C:$rA)>;
1562 def : Pat<(v4f32 (SPUprefslot2vec R32FP:$rA)),
1563 (ORv4f32_f32 R32FP:$rA)>;
1565 def : Pat<(v2f64 (SPUprefslot2vec R64FP:$rA)),
1566 (ORv2f64_f64 R64FP:$rA)>;
1568 // ORi*_v*: Used to extract vector element 0 (the preferred slot), otherwise
1569 // known as converting the vector back to its preferred slot
1571 def : Pat<(SPUvec2prefslot (v16i8 VECREG:$rA)),
1572 (ORi8_v16i8 VECREG:$rA)>;
1574 def : Pat<(SPUvec2prefslot (v8i16 VECREG:$rA)),
1575 (ORi16_v8i16 VECREG:$rA)>;
1577 def : Pat<(SPUvec2prefslot (v4i32 VECREG:$rA)),
1578 (ORi32_v4i32 VECREG:$rA)>;
1580 def : Pat<(SPUvec2prefslot (v2i64 VECREG:$rA)),
1581 (ORi64_v2i64 VECREG:$rA)>;
1583 def : Pat<(SPUvec2prefslot (v4f32 VECREG:$rA)),
1584 (ORf32_v4f32 VECREG:$rA)>;
1586 def : Pat<(SPUvec2prefslot (v2f64 VECREG:$rA)),
1587 (ORf64_v2f64 VECREG:$rA)>;
1589 // Load Register: This is an assembler alias for a bitwise OR of a register
1590 // against itself. It's here because it brings some clarity to assembly
1593 let hasCtrlDep = 1 in {
1594 class LRInst<dag OOL, dag IOL>
1595 : SPUInstr<OOL, IOL, "lr\t$rT, $rA", IntegerOp> {
1599 let Pattern = [/*no pattern*/];
1601 let Inst{0-10} = 0b10000010000; /* It's an OR operation */
1602 let Inst{11-17} = RA;
1603 let Inst{18-24} = RA;
1604 let Inst{25-31} = RT;
1607 class LRVecInst<ValueType vectype>:
1608 LRInst<(outs VECREG:$rT), (ins VECREG:$rA)>;
1610 class LRRegInst<RegisterClass rclass>:
1611 LRInst<(outs rclass:$rT), (ins rclass:$rA)>;
1613 multiclass LoadRegister {
1614 def v2i64: LRVecInst<v2i64>;
1615 def v2f64: LRVecInst<v2f64>;
1616 def v4i32: LRVecInst<v4i32>;
1617 def v4f32: LRVecInst<v4f32>;
1618 def v8i16: LRVecInst<v8i16>;
1619 def v16i8: LRVecInst<v16i8>;
1621 def r128: LRRegInst<GPRC>;
1622 def r64: LRRegInst<R64C>;
1623 def f64: LRRegInst<R64FP>;
1624 def r32: LRRegInst<R32C>;
1625 def f32: LRRegInst<R32FP>;
1626 def r16: LRRegInst<R16C>;
1627 def r8: LRRegInst<R8C>;
1630 defm LR: LoadRegister;
1633 // ORC: Bitwise "or" with complement (c = a | ~b)
1635 class ORCInst<dag OOL, dag IOL, list<dag> pattern>:
1636 RRForm<0b10010010000, OOL, IOL, "orc\t$rT, $rA, $rB",
1637 IntegerOp, pattern>;
1639 class ORCVecInst<ValueType vectype>:
1640 ORCInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1641 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1642 (vnot (vectype VECREG:$rB))))]>;
1644 class ORCRegInst<RegisterClass rclass>:
1645 ORCInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1646 [(set rclass:$rT, (or rclass:$rA, (not rclass:$rB)))]>;
1648 multiclass BitwiseOrComplement
1650 def v16i8: ORCVecInst<v16i8>;
1651 def v8i16: ORCVecInst<v8i16>;
1652 def v4i32: ORCVecInst<v4i32>;
1653 def v2i64: ORCVecInst<v2i64>;
1655 def r128: ORCRegInst<GPRC>;
1656 def r64: ORCRegInst<R64C>;
1657 def r32: ORCRegInst<R32C>;
1658 def r16: ORCRegInst<R16C>;
1659 def r8: ORCRegInst<R8C>;
1662 defm ORC : BitwiseOrComplement;
1664 // OR byte immediate
1665 class ORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1666 RI10Form<0b01100000, OOL, IOL, "orbi\t$rT, $rA, $val",
1667 IntegerOp, pattern>;
1669 class ORBIVecInst<ValueType vectype, PatLeaf immpred>:
1670 ORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1671 [(set (v16i8 VECREG:$rT), (or (vectype VECREG:$rA),
1672 (vectype immpred:$val)))]>;
1674 multiclass BitwiseOrByteImm
1676 def v16i8: ORBIVecInst<v16i8, v16i8U8Imm>;
1678 def r8: ORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1679 [(set R8C:$rT, (or R8C:$rA, immU8:$val))]>;
1682 defm ORBI : BitwiseOrByteImm;
1684 // OR halfword immediate
1685 class ORHIInst<dag OOL, dag IOL, list<dag> pattern>:
1686 RI10Form<0b10100000, OOL, IOL, "orhi\t$rT, $rA, $val",
1687 IntegerOp, pattern>;
1689 class ORHIVecInst<ValueType vectype, PatLeaf immpred>:
1690 ORHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1691 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1694 multiclass BitwiseOrHalfwordImm
1696 def v8i16: ORHIVecInst<v8i16, v8i16Uns10Imm>;
1698 def r16: ORHIInst<(outs R16C:$rT), (ins R16C:$rA, u10imm:$val),
1699 [(set R16C:$rT, (or R16C:$rA, i16ImmUns10:$val))]>;
1701 // Specialized ORHI form used to promote 8-bit registers to 16-bit
1702 def i8i16: ORHIInst<(outs R16C:$rT), (ins R8C:$rA, s10imm:$val),
1703 [(set R16C:$rT, (or (anyext R8C:$rA),
1704 i16ImmSExt10:$val))]>;
1707 defm ORHI : BitwiseOrHalfwordImm;
1709 class ORIInst<dag OOL, dag IOL, list<dag> pattern>:
1710 RI10Form<0b00100000, OOL, IOL, "ori\t$rT, $rA, $val",
1711 IntegerOp, pattern>;
1713 class ORIVecInst<ValueType vectype, PatLeaf immpred>:
1714 ORIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1715 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1718 // Bitwise "or" with immediate
1719 multiclass BitwiseOrImm
1721 def v4i32: ORIVecInst<v4i32, v4i32Uns10Imm>;
1723 def r32: ORIInst<(outs R32C:$rT), (ins R32C:$rA, u10imm_i32:$val),
1724 [(set R32C:$rT, (or R32C:$rA, i32ImmUns10:$val))]>;
1726 // i16i32: hacked version of the ori instruction to extend 16-bit quantities
1727 // to 32-bit quantities. used exclusively to match "anyext" conversions (vide
1728 // infra "anyext 16->32" pattern.)
1729 def i16i32: ORIInst<(outs R32C:$rT), (ins R16C:$rA, s10imm_i32:$val),
1730 [(set R32C:$rT, (or (anyext R16C:$rA),
1731 i32ImmSExt10:$val))]>;
1733 // i8i32: Hacked version of the ORI instruction to extend 16-bit quantities
1734 // to 32-bit quantities. Used exclusively to match "anyext" conversions (vide
1735 // infra "anyext 16->32" pattern.)
1736 def i8i32: ORIInst<(outs R32C:$rT), (ins R8C:$rA, s10imm_i32:$val),
1737 [(set R32C:$rT, (or (anyext R8C:$rA),
1738 i32ImmSExt10:$val))]>;
1741 defm ORI : BitwiseOrImm;
1743 // ORX: "or" across the vector: or's $rA's word slots leaving the result in
1744 // $rT[0], slots 1-3 are zeroed.
1746 // FIXME: Needs to match an intrinsic pattern.
1748 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1749 "orx\t$rT, $rA, $rB", IntegerOp,
1754 class XORInst<dag OOL, dag IOL, list<dag> pattern> :
1755 RRForm<0b10010010000, OOL, IOL, "xor\t$rT, $rA, $rB",
1756 IntegerOp, pattern>;
1758 class XORVecInst<ValueType vectype>:
1759 XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1760 [(set (vectype VECREG:$rT), (xor (vectype VECREG:$rA),
1761 (vectype VECREG:$rB)))]>;
1763 class XORRegInst<RegisterClass rclass>:
1764 XORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1765 [(set rclass:$rT, (xor rclass:$rA, rclass:$rB))]>;
1767 multiclass BitwiseExclusiveOr
1769 def v16i8: XORVecInst<v16i8>;
1770 def v8i16: XORVecInst<v8i16>;
1771 def v4i32: XORVecInst<v4i32>;
1772 def v2i64: XORVecInst<v2i64>;
1774 def r128: XORRegInst<GPRC>;
1775 def r64: XORRegInst<R64C>;
1776 def r32: XORRegInst<R32C>;
1777 def r16: XORRegInst<R16C>;
1778 def r8: XORRegInst<R8C>;
1780 // XOR instructions used to negate f32 and f64 quantities.
1782 def fneg32: XORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32C:$rB),
1783 [/* no pattern */]>;
1785 def fneg64: XORInst<(outs R64FP:$rT), (ins R64FP:$rA, R64C:$rB),
1786 [/* no pattern */]>;
1788 def fnegvec: XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1789 [/* no pattern, see fneg{32,64} */]>;
1792 defm XOR : BitwiseExclusiveOr;
1794 //==----------------------------------------------------------
1796 class XORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1797 RI10Form<0b01100000, OOL, IOL, "xorbi\t$rT, $rA, $val",
1798 IntegerOp, pattern>;
1800 multiclass XorByteImm
1803 XORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1804 [(set (v16i8 VECREG:$rT), (xor (v16i8 VECREG:$rA), v16i8U8Imm:$val))]>;
1807 XORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1808 [(set R8C:$rT, (xor R8C:$rA, immU8:$val))]>;
1811 defm XORBI : XorByteImm;
1814 RI10Form<0b10100000, (outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1815 "xorhi\t$rT, $rA, $val", IntegerOp,
1816 [(set (v8i16 VECREG:$rT), (xor (v8i16 VECREG:$rA),
1817 v8i16SExt10Imm:$val))]>;
1820 RI10Form<0b10100000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
1821 "xorhi\t$rT, $rA, $val", IntegerOp,
1822 [(set R16C:$rT, (xor R16C:$rA, i16ImmSExt10:$val))]>;
1825 RI10Form<0b00100000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm_i32:$val),
1826 "xori\t$rT, $rA, $val", IntegerOp,
1827 [(set (v4i32 VECREG:$rT), (xor (v4i32 VECREG:$rA),
1828 v4i32SExt10Imm:$val))]>;
1831 RI10Form<0b00100000, (outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
1832 "xori\t$rT, $rA, $val", IntegerOp,
1833 [(set R32C:$rT, (xor R32C:$rA, i32ImmSExt10:$val))]>;
1837 class NANDInst<dag OOL, dag IOL, list<dag> pattern>:
1838 RRForm<0b10010011000, OOL, IOL, "nand\t$rT, $rA, $rB",
1839 IntegerOp, pattern>;
1841 class NANDVecInst<ValueType vectype>:
1842 NANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1843 [(set (vectype VECREG:$rT), (vnot (and (vectype VECREG:$rA),
1844 (vectype VECREG:$rB))))]>;
1845 class NANDRegInst<RegisterClass rclass>:
1846 NANDInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1847 [(set rclass:$rT, (not (and rclass:$rA, rclass:$rB)))]>;
1849 multiclass BitwiseNand
1851 def v16i8: NANDVecInst<v16i8>;
1852 def v8i16: NANDVecInst<v8i16>;
1853 def v4i32: NANDVecInst<v4i32>;
1854 def v2i64: NANDVecInst<v2i64>;
1856 def r128: NANDRegInst<GPRC>;
1857 def r64: NANDRegInst<R64C>;
1858 def r32: NANDRegInst<R32C>;
1859 def r16: NANDRegInst<R16C>;
1860 def r8: NANDRegInst<R8C>;
1863 defm NAND : BitwiseNand;
1867 class NORInst<dag OOL, dag IOL, list<dag> pattern>:
1868 RRForm<0b10010010000, OOL, IOL, "nor\t$rT, $rA, $rB",
1869 IntegerOp, pattern>;
1871 class NORVecInst<ValueType vectype>:
1872 NORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1873 [(set (vectype VECREG:$rT), (vnot (or (vectype VECREG:$rA),
1874 (vectype VECREG:$rB))))]>;
1875 class NORRegInst<RegisterClass rclass>:
1876 NORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1877 [(set rclass:$rT, (not (or rclass:$rA, rclass:$rB)))]>;
1879 multiclass BitwiseNor
1881 def v16i8: NORVecInst<v16i8>;
1882 def v8i16: NORVecInst<v8i16>;
1883 def v4i32: NORVecInst<v4i32>;
1884 def v2i64: NORVecInst<v2i64>;
1886 def r128: NORRegInst<GPRC>;
1887 def r64: NORRegInst<R64C>;
1888 def r32: NORRegInst<R32C>;
1889 def r16: NORRegInst<R16C>;
1890 def r8: NORRegInst<R8C>;
1893 defm NOR : BitwiseNor;
1896 class SELBInst<dag OOL, dag IOL, list<dag> pattern>:
1897 RRRForm<0b1000, OOL, IOL, "selb\t$rT, $rA, $rB, $rC",
1898 IntegerOp, pattern>;
1900 class SELBVecInst<ValueType vectype, PatFrag vnot_frag = vnot>:
1901 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1902 [(set (vectype VECREG:$rT),
1903 (or (and (vectype VECREG:$rC), (vectype VECREG:$rB)),
1904 (and (vnot_frag (vectype VECREG:$rC)),
1905 (vectype VECREG:$rA))))]>;
1907 class SELBVecVCondInst<ValueType vectype>:
1908 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1909 [(set (vectype VECREG:$rT),
1910 (select (vectype VECREG:$rC),
1911 (vectype VECREG:$rB),
1912 (vectype VECREG:$rA)))]>;
1914 class SELBVecCondInst<ValueType vectype>:
1915 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, R32C:$rC),
1916 [(set (vectype VECREG:$rT),
1918 (vectype VECREG:$rB),
1919 (vectype VECREG:$rA)))]>;
1921 class SELBRegInst<RegisterClass rclass>:
1922 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rclass:$rC),
1924 (or (and rclass:$rB, rclass:$rC),
1925 (and rclass:$rA, (not rclass:$rC))))]>;
1927 class SELBRegCondInst<RegisterClass rcond, RegisterClass rclass>:
1928 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rcond:$rC),
1930 (select rcond:$rC, rclass:$rB, rclass:$rA))]>;
1932 multiclass SelectBits
1934 def v16i8: SELBVecInst<v16i8>;
1935 def v8i16: SELBVecInst<v8i16>;
1936 def v4i32: SELBVecInst<v4i32>;
1937 def v2i64: SELBVecInst<v2i64, vnot_conv>;
1939 def r128: SELBRegInst<GPRC>;
1940 def r64: SELBRegInst<R64C>;
1941 def r32: SELBRegInst<R32C>;
1942 def r16: SELBRegInst<R16C>;
1943 def r8: SELBRegInst<R8C>;
1945 def v16i8_cond: SELBVecCondInst<v16i8>;
1946 def v8i16_cond: SELBVecCondInst<v8i16>;
1947 def v4i32_cond: SELBVecCondInst<v4i32>;
1948 def v2i64_cond: SELBVecCondInst<v2i64>;
1950 def v16i8_vcond: SELBVecCondInst<v16i8>;
1951 def v8i16_vcond: SELBVecCondInst<v8i16>;
1952 def v4i32_vcond: SELBVecCondInst<v4i32>;
1953 def v2i64_vcond: SELBVecCondInst<v2i64>;
1956 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1957 [(set (v4f32 VECREG:$rT),
1958 (select (v4i32 VECREG:$rC),
1960 (v4f32 VECREG:$rA)))]>;
1962 // SELBr64_cond is defined in SPU64InstrInfo.td
1963 def r32_cond: SELBRegCondInst<R32C, R32C>;
1964 def f32_cond: SELBRegCondInst<R32C, R32FP>;
1965 def r16_cond: SELBRegCondInst<R16C, R16C>;
1966 def r8_cond: SELBRegCondInst<R8C, R8C>;
1969 defm SELB : SelectBits;
1971 class SPUselbPatVec<ValueType vectype, SPUInstr inst>:
1972 Pat<(SPUselb (vectype VECREG:$rA), (vectype VECREG:$rB), (vectype VECREG:$rC)),
1973 (inst VECREG:$rA, VECREG:$rB, VECREG:$rC)>;
1975 def : SPUselbPatVec<v16i8, SELBv16i8>;
1976 def : SPUselbPatVec<v8i16, SELBv8i16>;
1977 def : SPUselbPatVec<v4i32, SELBv4i32>;
1978 def : SPUselbPatVec<v2i64, SELBv2i64>;
1980 class SPUselbPatReg<RegisterClass rclass, SPUInstr inst>:
1981 Pat<(SPUselb rclass:$rA, rclass:$rB, rclass:$rC),
1982 (inst rclass:$rA, rclass:$rB, rclass:$rC)>;
1984 def : SPUselbPatReg<R8C, SELBr8>;
1985 def : SPUselbPatReg<R16C, SELBr16>;
1986 def : SPUselbPatReg<R32C, SELBr32>;
1987 def : SPUselbPatReg<R64C, SELBr64>;
1989 // EQV: Equivalence (1 for each same bit, otherwise 0)
1991 // Note: There are a lot of ways to match this bit operator and these patterns
1992 // attempt to be as exhaustive as possible.
1994 class EQVInst<dag OOL, dag IOL, list<dag> pattern>:
1995 RRForm<0b10010010000, OOL, IOL, "eqv\t$rT, $rA, $rB",
1996 IntegerOp, pattern>;
1998 class EQVVecInst<ValueType vectype>:
1999 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2000 [(set (vectype VECREG:$rT),
2001 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
2002 (and (vnot (vectype VECREG:$rA)),
2003 (vnot (vectype VECREG:$rB)))))]>;
2005 class EQVRegInst<RegisterClass rclass>:
2006 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2007 [(set rclass:$rT, (or (and rclass:$rA, rclass:$rB),
2008 (and (not rclass:$rA), (not rclass:$rB))))]>;
2010 class EQVVecPattern1<ValueType vectype>:
2011 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2012 [(set (vectype VECREG:$rT),
2013 (xor (vectype VECREG:$rA), (vnot (vectype VECREG:$rB))))]>;
2015 class EQVRegPattern1<RegisterClass rclass>:
2016 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2017 [(set rclass:$rT, (xor rclass:$rA, (not rclass:$rB)))]>;
2019 class EQVVecPattern2<ValueType vectype>:
2020 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2021 [(set (vectype VECREG:$rT),
2022 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
2023 (vnot (or (vectype VECREG:$rA), (vectype VECREG:$rB)))))]>;
2025 class EQVRegPattern2<RegisterClass rclass>:
2026 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2028 (or (and rclass:$rA, rclass:$rB),
2029 (not (or rclass:$rA, rclass:$rB))))]>;
2031 class EQVVecPattern3<ValueType vectype>:
2032 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2033 [(set (vectype VECREG:$rT),
2034 (not (xor (vectype VECREG:$rA), (vectype VECREG:$rB))))]>;
2036 class EQVRegPattern3<RegisterClass rclass>:
2037 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2038 [(set rclass:$rT, (not (xor rclass:$rA, rclass:$rB)))]>;
2040 multiclass BitEquivalence
2042 def v16i8: EQVVecInst<v16i8>;
2043 def v8i16: EQVVecInst<v8i16>;
2044 def v4i32: EQVVecInst<v4i32>;
2045 def v2i64: EQVVecInst<v2i64>;
2047 def v16i8_1: EQVVecPattern1<v16i8>;
2048 def v8i16_1: EQVVecPattern1<v8i16>;
2049 def v4i32_1: EQVVecPattern1<v4i32>;
2050 def v2i64_1: EQVVecPattern1<v2i64>;
2052 def v16i8_2: EQVVecPattern2<v16i8>;
2053 def v8i16_2: EQVVecPattern2<v8i16>;
2054 def v4i32_2: EQVVecPattern2<v4i32>;
2055 def v2i64_2: EQVVecPattern2<v2i64>;
2057 def v16i8_3: EQVVecPattern3<v16i8>;
2058 def v8i16_3: EQVVecPattern3<v8i16>;
2059 def v4i32_3: EQVVecPattern3<v4i32>;
2060 def v2i64_3: EQVVecPattern3<v2i64>;
2062 def r128: EQVRegInst<GPRC>;
2063 def r64: EQVRegInst<R64C>;
2064 def r32: EQVRegInst<R32C>;
2065 def r16: EQVRegInst<R16C>;
2066 def r8: EQVRegInst<R8C>;
2068 def r128_1: EQVRegPattern1<GPRC>;
2069 def r64_1: EQVRegPattern1<R64C>;
2070 def r32_1: EQVRegPattern1<R32C>;
2071 def r16_1: EQVRegPattern1<R16C>;
2072 def r8_1: EQVRegPattern1<R8C>;
2074 def r128_2: EQVRegPattern2<GPRC>;
2075 def r64_2: EQVRegPattern2<R64C>;
2076 def r32_2: EQVRegPattern2<R32C>;
2077 def r16_2: EQVRegPattern2<R16C>;
2078 def r8_2: EQVRegPattern2<R8C>;
2080 def r128_3: EQVRegPattern3<GPRC>;
2081 def r64_3: EQVRegPattern3<R64C>;
2082 def r32_3: EQVRegPattern3<R32C>;
2083 def r16_3: EQVRegPattern3<R16C>;
2084 def r8_3: EQVRegPattern3<R8C>;
2087 defm EQV: BitEquivalence;
2089 //===----------------------------------------------------------------------===//
2090 // Vector shuffle...
2091 //===----------------------------------------------------------------------===//
2092 // SPUshuffle is generated in LowerVECTOR_SHUFFLE and gets replaced with SHUFB.
2093 // See the SPUshuffle SDNode operand above, which sets up the DAG pattern
2094 // matcher to emit something when the LowerVECTOR_SHUFFLE generates a node with
2095 // the SPUISD::SHUFB opcode.
2096 //===----------------------------------------------------------------------===//
2098 class SHUFBInst<dag OOL, dag IOL, list<dag> pattern>:
2099 RRRForm<0b1000, OOL, IOL, "shufb\t$rT, $rA, $rB, $rC",
2100 IntegerOp, pattern>;
2102 class SHUFBVecInst<ValueType resultvec, ValueType maskvec>:
2103 SHUFBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
2104 [(set (resultvec VECREG:$rT),
2105 (SPUshuffle (resultvec VECREG:$rA),
2106 (resultvec VECREG:$rB),
2107 (maskvec VECREG:$rC)))]>;
2109 class SHUFBGPRCInst:
2110 SHUFBInst<(outs VECREG:$rT), (ins GPRC:$rA, GPRC:$rB, VECREG:$rC),
2111 [/* no pattern */]>;
2113 multiclass ShuffleBytes
2115 def v16i8 : SHUFBVecInst<v16i8, v16i8>;
2116 def v16i8_m32 : SHUFBVecInst<v16i8, v4i32>;
2117 def v8i16 : SHUFBVecInst<v8i16, v16i8>;
2118 def v8i16_m32 : SHUFBVecInst<v8i16, v4i32>;
2119 def v4i32 : SHUFBVecInst<v4i32, v16i8>;
2120 def v4i32_m32 : SHUFBVecInst<v4i32, v4i32>;
2121 def v2i64 : SHUFBVecInst<v2i64, v16i8>;
2122 def v2i64_m32 : SHUFBVecInst<v2i64, v4i32>;
2124 def v4f32 : SHUFBVecInst<v4f32, v16i8>;
2125 def v4f32_m32 : SHUFBVecInst<v4f32, v4i32>;
2127 def v2f64 : SHUFBVecInst<v2f64, v16i8>;
2128 def v2f64_m32 : SHUFBVecInst<v2f64, v4i32>;
2130 def gprc : SHUFBGPRCInst;
2133 defm SHUFB : ShuffleBytes;
2135 //===----------------------------------------------------------------------===//
2136 // Shift and rotate group:
2137 //===----------------------------------------------------------------------===//
2139 class SHLHInst<dag OOL, dag IOL, list<dag> pattern>:
2140 RRForm<0b11111010000, OOL, IOL, "shlh\t$rT, $rA, $rB",
2141 RotateShift, pattern>;
2143 class SHLHVecInst<ValueType vectype>:
2144 SHLHInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2145 [(set (vectype VECREG:$rT),
2146 (SPUvec_shl (vectype VECREG:$rA), R16C:$rB))]>;
2148 multiclass ShiftLeftHalfword
2150 def v8i16: SHLHVecInst<v8i16>;
2151 def r16: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
2152 [(set R16C:$rT, (shl R16C:$rA, R16C:$rB))]>;
2153 def r16_r32: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2154 [(set R16C:$rT, (shl R16C:$rA, R32C:$rB))]>;
2157 defm SHLH : ShiftLeftHalfword;
2159 //===----------------------------------------------------------------------===//
2161 class SHLHIInst<dag OOL, dag IOL, list<dag> pattern>:
2162 RI7Form<0b11111010000, OOL, IOL, "shlhi\t$rT, $rA, $val",
2163 RotateShift, pattern>;
2165 class SHLHIVecInst<ValueType vectype>:
2166 SHLHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2167 [(set (vectype VECREG:$rT),
2168 (SPUvec_shl (vectype VECREG:$rA), (i16 uimm7:$val)))]>;
2170 multiclass ShiftLeftHalfwordImm
2172 def v8i16: SHLHIVecInst<v8i16>;
2173 def r16: SHLHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2174 [(set R16C:$rT, (shl R16C:$rA, (i16 uimm7:$val)))]>;
2177 defm SHLHI : ShiftLeftHalfwordImm;
2179 def : Pat<(SPUvec_shl (v8i16 VECREG:$rA), (i32 uimm7:$val)),
2180 (SHLHIv8i16 VECREG:$rA, (TO_IMM16 uimm7:$val))>;
2182 def : Pat<(shl R16C:$rA, (i32 uimm7:$val)),
2183 (SHLHIr16 R16C:$rA, (TO_IMM16 uimm7:$val))>;
2185 //===----------------------------------------------------------------------===//
2187 class SHLInst<dag OOL, dag IOL, list<dag> pattern>:
2188 RRForm<0b11111010000, OOL, IOL, "shl\t$rT, $rA, $rB",
2189 RotateShift, pattern>;
2191 multiclass ShiftLeftWord
2194 SHLInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2195 [(set (v4i32 VECREG:$rT),
2196 (SPUvec_shl (v4i32 VECREG:$rA), R16C:$rB))]>;
2198 SHLInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2199 [(set R32C:$rT, (shl R32C:$rA, R32C:$rB))]>;
2202 defm SHL: ShiftLeftWord;
2204 //===----------------------------------------------------------------------===//
2206 class SHLIInst<dag OOL, dag IOL, list<dag> pattern>:
2207 RI7Form<0b11111010000, OOL, IOL, "shli\t$rT, $rA, $val",
2208 RotateShift, pattern>;
2210 multiclass ShiftLeftWordImm
2213 SHLIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2214 [(set (v4i32 VECREG:$rT),
2215 (SPUvec_shl (v4i32 VECREG:$rA), (i32 uimm7:$val)))]>;
2218 SHLIInst<(outs R32C:$rT), (ins R32C:$rA, u7imm_i32:$val),
2219 [(set R32C:$rT, (shl R32C:$rA, (i32 uimm7:$val)))]>;
2222 defm SHLI : ShiftLeftWordImm;
2224 //===----------------------------------------------------------------------===//
2225 // SHLQBI vec form: Note that this will shift the entire vector (the 128-bit
2226 // register) to the left. Vector form is here to ensure type correctness.
2228 // The shift count is in the lowest 3 bits (29-31) of $rB, so only a bit shift
2229 // of 7 bits is actually possible.
2231 // Note also that SHLQBI/SHLQBII are used in conjunction with SHLQBY/SHLQBYI
2232 // to shift i64 and i128. SHLQBI is the residual left over after shifting by
2233 // bytes with SHLQBY.
2235 class SHLQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2236 RRForm<0b11011011100, OOL, IOL, "shlqbi\t$rT, $rA, $rB",
2237 RotateShift, pattern>;
2239 class SHLQBIVecInst<ValueType vectype>:
2240 SHLQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2241 [(set (vectype VECREG:$rT),
2242 (SPUshlquad_l_bits (vectype VECREG:$rA), R32C:$rB))]>;
2244 class SHLQBIRegInst<RegisterClass rclass>:
2245 SHLQBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2246 [/* no pattern */]>;
2248 multiclass ShiftLeftQuadByBits
2250 def v16i8: SHLQBIVecInst<v16i8>;
2251 def v8i16: SHLQBIVecInst<v8i16>;
2252 def v4i32: SHLQBIVecInst<v4i32>;
2253 def v4f32: SHLQBIVecInst<v4f32>;
2254 def v2i64: SHLQBIVecInst<v2i64>;
2255 def v2f64: SHLQBIVecInst<v2f64>;
2257 def r128: SHLQBIRegInst<GPRC>;
2260 defm SHLQBI : ShiftLeftQuadByBits;
2262 // See note above on SHLQBI. In this case, the predicate actually does then
2263 // enforcement, whereas with SHLQBI, we have to "take it on faith."
2264 class SHLQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2265 RI7Form<0b11011111100, OOL, IOL, "shlqbii\t$rT, $rA, $val",
2266 RotateShift, pattern>;
2268 class SHLQBIIVecInst<ValueType vectype>:
2269 SHLQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2270 [(set (vectype VECREG:$rT),
2271 (SPUshlquad_l_bits (vectype VECREG:$rA), (i32 bitshift:$val)))]>;
2273 multiclass ShiftLeftQuadByBitsImm
2275 def v16i8 : SHLQBIIVecInst<v16i8>;
2276 def v8i16 : SHLQBIIVecInst<v8i16>;
2277 def v4i32 : SHLQBIIVecInst<v4i32>;
2278 def v4f32 : SHLQBIIVecInst<v4f32>;
2279 def v2i64 : SHLQBIIVecInst<v2i64>;
2280 def v2f64 : SHLQBIIVecInst<v2f64>;
2283 defm SHLQBII : ShiftLeftQuadByBitsImm;
2285 // SHLQBY, SHLQBYI vector forms: Shift the entire vector to the left by bytes,
2286 // not by bits. See notes above on SHLQBI.
2288 class SHLQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2289 RI7Form<0b11111011100, OOL, IOL, "shlqby\t$rT, $rA, $rB",
2290 RotateShift, pattern>;
2292 class SHLQBYVecInst<ValueType vectype>:
2293 SHLQBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2294 [(set (vectype VECREG:$rT),
2295 (SPUshlquad_l_bytes (vectype VECREG:$rA), R32C:$rB))]>;
2297 multiclass ShiftLeftQuadBytes
2299 def v16i8: SHLQBYVecInst<v16i8>;
2300 def v8i16: SHLQBYVecInst<v8i16>;
2301 def v4i32: SHLQBYVecInst<v4i32>;
2302 def v4f32: SHLQBYVecInst<v4f32>;
2303 def v2i64: SHLQBYVecInst<v2i64>;
2304 def v2f64: SHLQBYVecInst<v2f64>;
2305 def r128: SHLQBYInst<(outs GPRC:$rT), (ins GPRC:$rA, R32C:$rB),
2306 [(set GPRC:$rT, (SPUshlquad_l_bytes GPRC:$rA, R32C:$rB))]>;
2309 defm SHLQBY: ShiftLeftQuadBytes;
2311 class SHLQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2312 RI7Form<0b11111111100, OOL, IOL, "shlqbyi\t$rT, $rA, $val",
2313 RotateShift, pattern>;
2315 class SHLQBYIVecInst<ValueType vectype>:
2316 SHLQBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2317 [(set (vectype VECREG:$rT),
2318 (SPUshlquad_l_bytes (vectype VECREG:$rA), (i32 uimm7:$val)))]>;
2320 multiclass ShiftLeftQuadBytesImm
2322 def v16i8: SHLQBYIVecInst<v16i8>;
2323 def v8i16: SHLQBYIVecInst<v8i16>;
2324 def v4i32: SHLQBYIVecInst<v4i32>;
2325 def v4f32: SHLQBYIVecInst<v4f32>;
2326 def v2i64: SHLQBYIVecInst<v2i64>;
2327 def v2f64: SHLQBYIVecInst<v2f64>;
2328 def r128: SHLQBYIInst<(outs GPRC:$rT), (ins GPRC:$rA, u7imm_i32:$val),
2330 (SPUshlquad_l_bytes GPRC:$rA, (i32 uimm7:$val)))]>;
2333 defm SHLQBYI : ShiftLeftQuadBytesImm;
2335 class SHLQBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2336 RRForm<0b00111001111, OOL, IOL, "shlqbybi\t$rT, $rA, $rB",
2337 RotateShift, pattern>;
2339 class SHLQBYBIVecInst<ValueType vectype>:
2340 SHLQBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2341 [/* no pattern */]>;
2343 class SHLQBYBIRegInst<RegisterClass rclass>:
2344 SHLQBYBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2345 [/* no pattern */]>;
2347 multiclass ShiftLeftQuadBytesBitCount
2349 def v16i8: SHLQBYBIVecInst<v16i8>;
2350 def v8i16: SHLQBYBIVecInst<v8i16>;
2351 def v4i32: SHLQBYBIVecInst<v4i32>;
2352 def v4f32: SHLQBYBIVecInst<v4f32>;
2353 def v2i64: SHLQBYBIVecInst<v2i64>;
2354 def v2f64: SHLQBYBIVecInst<v2f64>;
2356 def r128: SHLQBYBIRegInst<GPRC>;
2359 defm SHLQBYBI : ShiftLeftQuadBytesBitCount;
2361 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2363 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2364 class ROTHInst<dag OOL, dag IOL, list<dag> pattern>:
2365 RRForm<0b00111010000, OOL, IOL, "roth\t$rT, $rA, $rB",
2366 RotateShift, pattern>;
2368 class ROTHVecInst<ValueType vectype>:
2369 ROTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2370 [(set (vectype VECREG:$rT),
2371 (SPUvec_rotl VECREG:$rA, (v8i16 VECREG:$rB)))]>;
2373 class ROTHRegInst<RegisterClass rclass>:
2374 ROTHInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2375 [(set rclass:$rT, (rotl rclass:$rA, rclass:$rB))]>;
2377 multiclass RotateLeftHalfword
2379 def v8i16: ROTHVecInst<v8i16>;
2380 def r16: ROTHRegInst<R16C>;
2383 defm ROTH: RotateLeftHalfword;
2385 def ROTHr16_r32: ROTHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2386 [(set R16C:$rT, (rotl R16C:$rA, R32C:$rB))]>;
2388 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2389 // Rotate halfword, immediate:
2390 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2391 class ROTHIInst<dag OOL, dag IOL, list<dag> pattern>:
2392 RI7Form<0b00111110000, OOL, IOL, "rothi\t$rT, $rA, $val",
2393 RotateShift, pattern>;
2395 class ROTHIVecInst<ValueType vectype>:
2396 ROTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2397 [(set (vectype VECREG:$rT),
2398 (SPUvec_rotl VECREG:$rA, (i16 uimm7:$val)))]>;
2400 multiclass RotateLeftHalfwordImm
2402 def v8i16: ROTHIVecInst<v8i16>;
2403 def r16: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2404 [(set R16C:$rT, (rotl R16C:$rA, (i16 uimm7:$val)))]>;
2405 def r16_r32: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm_i32:$val),
2406 [(set R16C:$rT, (rotl R16C:$rA, (i32 uimm7:$val)))]>;
2409 defm ROTHI: RotateLeftHalfwordImm;
2411 def : Pat<(SPUvec_rotl (v8i16 VECREG:$rA), (i32 uimm7:$val)),
2412 (ROTHIv8i16 VECREG:$rA, (TO_IMM16 imm:$val))>;
2414 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2416 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2418 class ROTInst<dag OOL, dag IOL, list<dag> pattern>:
2419 RRForm<0b00011010000, OOL, IOL, "rot\t$rT, $rA, $rB",
2420 RotateShift, pattern>;
2422 class ROTVecInst<ValueType vectype>:
2423 ROTInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2424 [(set (vectype VECREG:$rT),
2425 (SPUvec_rotl (vectype VECREG:$rA), R32C:$rB))]>;
2427 class ROTRegInst<RegisterClass rclass>:
2428 ROTInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2430 (rotl rclass:$rA, R32C:$rB))]>;
2432 multiclass RotateLeftWord
2434 def v4i32: ROTVecInst<v4i32>;
2435 def r32: ROTRegInst<R32C>;
2438 defm ROT: RotateLeftWord;
2440 // The rotate amount is in the same bits whether we've got an 8-bit, 16-bit or
2442 def ROTr32_r16_anyext:
2443 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R16C:$rB),
2444 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R16C:$rB))))]>;
2446 def : Pat<(rotl R32C:$rA, (i32 (zext R16C:$rB))),
2447 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2449 def : Pat<(rotl R32C:$rA, (i32 (sext R16C:$rB))),
2450 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2452 def ROTr32_r8_anyext:
2453 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R8C:$rB),
2454 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R8C:$rB))))]>;
2456 def : Pat<(rotl R32C:$rA, (i32 (zext R8C:$rB))),
2457 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2459 def : Pat<(rotl R32C:$rA, (i32 (sext R8C:$rB))),
2460 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2462 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2463 // Rotate word, immediate
2464 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2466 class ROTIInst<dag OOL, dag IOL, list<dag> pattern>:
2467 RI7Form<0b00011110000, OOL, IOL, "roti\t$rT, $rA, $val",
2468 RotateShift, pattern>;
2470 class ROTIVecInst<ValueType vectype, Operand optype, ValueType inttype, PatLeaf pred>:
2471 ROTIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2472 [(set (vectype VECREG:$rT),
2473 (SPUvec_rotl (vectype VECREG:$rA), (inttype pred:$val)))]>;
2475 class ROTIRegInst<RegisterClass rclass, Operand optype, ValueType inttype, PatLeaf pred>:
2476 ROTIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2477 [(set rclass:$rT, (rotl rclass:$rA, (inttype pred:$val)))]>;
2479 multiclass RotateLeftWordImm
2481 def v4i32: ROTIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2482 def v4i32_i16: ROTIVecInst<v4i32, u7imm, i16, uimm7>;
2483 def v4i32_i8: ROTIVecInst<v4i32, u7imm_i8, i8, uimm7>;
2485 def r32: ROTIRegInst<R32C, u7imm_i32, i32, uimm7>;
2486 def r32_i16: ROTIRegInst<R32C, u7imm, i16, uimm7>;
2487 def r32_i8: ROTIRegInst<R32C, u7imm_i8, i8, uimm7>;
2490 defm ROTI : RotateLeftWordImm;
2492 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2493 // Rotate quad by byte (count)
2494 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2496 class ROTQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2497 RRForm<0b00111011100, OOL, IOL, "rotqby\t$rT, $rA, $rB",
2498 RotateShift, pattern>;
2500 class ROTQBYVecInst<ValueType vectype>:
2501 ROTQBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2502 [(set (vectype VECREG:$rT),
2503 (SPUrotbytes_left (vectype VECREG:$rA), R32C:$rB))]>;
2505 multiclass RotateQuadLeftByBytes
2507 def v16i8: ROTQBYVecInst<v16i8>;
2508 def v8i16: ROTQBYVecInst<v8i16>;
2509 def v4i32: ROTQBYVecInst<v4i32>;
2510 def v4f32: ROTQBYVecInst<v4f32>;
2511 def v2i64: ROTQBYVecInst<v2i64>;
2512 def v2f64: ROTQBYVecInst<v2f64>;
2515 defm ROTQBY: RotateQuadLeftByBytes;
2517 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2518 // Rotate quad by byte (count), immediate
2519 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2521 class ROTQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2522 RI7Form<0b00111111100, OOL, IOL, "rotqbyi\t$rT, $rA, $val",
2523 RotateShift, pattern>;
2525 class ROTQBYIVecInst<ValueType vectype>:
2526 ROTQBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2527 [(set (vectype VECREG:$rT),
2528 (SPUrotbytes_left (vectype VECREG:$rA), (i16 uimm7:$val)))]>;
2530 multiclass RotateQuadByBytesImm
2532 def v16i8: ROTQBYIVecInst<v16i8>;
2533 def v8i16: ROTQBYIVecInst<v8i16>;
2534 def v4i32: ROTQBYIVecInst<v4i32>;
2535 def v4f32: ROTQBYIVecInst<v4f32>;
2536 def v2i64: ROTQBYIVecInst<v2i64>;
2537 def vfi64: ROTQBYIVecInst<v2f64>;
2540 defm ROTQBYI: RotateQuadByBytesImm;
2542 // See ROTQBY note above.
2543 class ROTQBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2544 RI7Form<0b00110011100, OOL, IOL,
2545 "rotqbybi\t$rT, $rA, $shift",
2546 RotateShift, pattern>;
2548 class ROTQBYBIVecInst<ValueType vectype, RegisterClass rclass>:
2549 ROTQBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, rclass:$shift),
2550 [(set (vectype VECREG:$rT),
2551 (SPUrotbytes_left_bits (vectype VECREG:$rA), rclass:$shift))]>;
2553 multiclass RotateQuadByBytesByBitshift {
2554 def v16i8_r32: ROTQBYBIVecInst<v16i8, R32C>;
2555 def v8i16_r32: ROTQBYBIVecInst<v8i16, R32C>;
2556 def v4i32_r32: ROTQBYBIVecInst<v4i32, R32C>;
2557 def v2i64_r32: ROTQBYBIVecInst<v2i64, R32C>;
2560 defm ROTQBYBI : RotateQuadByBytesByBitshift;
2562 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2563 // See ROTQBY note above.
2565 // Assume that the user of this instruction knows to shift the rotate count
2567 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2569 class ROTQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2570 RRForm<0b00011011100, OOL, IOL, "rotqbi\t$rT, $rA, $rB",
2571 RotateShift, pattern>;
2573 class ROTQBIVecInst<ValueType vectype>:
2574 ROTQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2575 [/* no pattern yet */]>;
2577 class ROTQBIRegInst<RegisterClass rclass>:
2578 ROTQBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2579 [/* no pattern yet */]>;
2581 multiclass RotateQuadByBitCount
2583 def v16i8: ROTQBIVecInst<v16i8>;
2584 def v8i16: ROTQBIVecInst<v8i16>;
2585 def v4i32: ROTQBIVecInst<v4i32>;
2586 def v2i64: ROTQBIVecInst<v2i64>;
2588 def r128: ROTQBIRegInst<GPRC>;
2589 def r64: ROTQBIRegInst<R64C>;
2592 defm ROTQBI: RotateQuadByBitCount;
2594 class ROTQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2595 RI7Form<0b00011111100, OOL, IOL, "rotqbii\t$rT, $rA, $val",
2596 RotateShift, pattern>;
2598 class ROTQBIIVecInst<ValueType vectype, Operand optype, ValueType inttype,
2600 ROTQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2601 [/* no pattern yet */]>;
2603 class ROTQBIIRegInst<RegisterClass rclass, Operand optype, ValueType inttype,
2605 ROTQBIIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2606 [/* no pattern yet */]>;
2608 multiclass RotateQuadByBitCountImm
2610 def v16i8: ROTQBIIVecInst<v16i8, u7imm_i32, i32, uimm7>;
2611 def v8i16: ROTQBIIVecInst<v8i16, u7imm_i32, i32, uimm7>;
2612 def v4i32: ROTQBIIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2613 def v2i64: ROTQBIIVecInst<v2i64, u7imm_i32, i32, uimm7>;
2615 def r128: ROTQBIIRegInst<GPRC, u7imm_i32, i32, uimm7>;
2616 def r64: ROTQBIIRegInst<R64C, u7imm_i32, i32, uimm7>;
2619 defm ROTQBII : RotateQuadByBitCountImm;
2621 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2622 // ROTHM v8i16 form:
2623 // NOTE(1): No vector rotate is generated by the C/C++ frontend (today),
2624 // so this only matches a synthetically generated/lowered code
2626 // NOTE(2): $rB must be negated before the right rotate!
2627 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2629 class ROTHMInst<dag OOL, dag IOL, list<dag> pattern>:
2630 RRForm<0b10111010000, OOL, IOL, "rothm\t$rT, $rA, $rB",
2631 RotateShift, pattern>;
2634 ROTHMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2635 [/* see patterns below - $rB must be negated */]>;
2637 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R32C:$rB),
2638 (ROTHMv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2640 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R16C:$rB),
2641 (ROTHMv8i16 VECREG:$rA,
2642 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2644 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R8C:$rB),
2645 (ROTHMv8i16 VECREG:$rA,
2646 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2648 // ROTHM r16 form: Rotate 16-bit quantity to right, zero fill at the left
2649 // Note: This instruction doesn't match a pattern because rB must be negated
2650 // for the instruction to work. Thus, the pattern below the instruction!
2653 ROTHMInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2654 [/* see patterns below - $rB must be negated! */]>;
2656 def : Pat<(srl R16C:$rA, R32C:$rB),
2657 (ROTHMr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2659 def : Pat<(srl R16C:$rA, R16C:$rB),
2661 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2663 def : Pat<(srl R16C:$rA, R8C:$rB),
2665 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2667 // ROTHMI v8i16 form: See the comment for ROTHM v8i16. The difference here is
2668 // that the immediate can be complemented, so that the user doesn't have to
2671 class ROTHMIInst<dag OOL, dag IOL, list<dag> pattern>:
2672 RI7Form<0b10111110000, OOL, IOL, "rothmi\t$rT, $rA, $val",
2673 RotateShift, pattern>;
2676 ROTHMIInst<(outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2677 [/* no pattern */]>;
2679 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i32 imm:$val)),
2680 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2682 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i16 imm:$val)),
2683 (ROTHMIv8i16 VECREG:$rA, (TO_IMM32 imm:$val))>;
2685 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i8 imm:$val)),
2686 (ROTHMIv8i16 VECREG:$rA, (TO_IMM32 imm:$val))>;
2689 ROTHMIInst<(outs R16C:$rT), (ins R16C:$rA, rothNeg7imm:$val),
2690 [/* no pattern */]>;
2692 def: Pat<(srl R16C:$rA, (i32 uimm7:$val)),
2693 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2695 def: Pat<(srl R16C:$rA, (i16 uimm7:$val)),
2696 (ROTHMIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
2698 def: Pat<(srl R16C:$rA, (i8 uimm7:$val)),
2699 (ROTHMIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
2701 // ROTM v4i32 form: See the ROTHM v8i16 comments.
2702 class ROTMInst<dag OOL, dag IOL, list<dag> pattern>:
2703 RRForm<0b10011010000, OOL, IOL, "rotm\t$rT, $rA, $rB",
2704 RotateShift, pattern>;
2707 ROTMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2708 [/* see patterns below - $rB must be negated */]>;
2710 def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), R32C:$rB),
2711 (ROTMv4i32 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2713 def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), R16C:$rB),
2714 (ROTMv4i32 VECREG:$rA,
2715 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2717 def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), R8C:$rB),
2718 (ROTMv4i32 VECREG:$rA,
2719 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2722 ROTMInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2723 [/* see patterns below - $rB must be negated */]>;
2725 def : Pat<(srl R32C:$rA, R32C:$rB),
2726 (ROTMr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
2728 def : Pat<(srl R32C:$rA, R16C:$rB),
2730 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2732 def : Pat<(srl R32C:$rA, R8C:$rB),
2734 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2736 // ROTMI v4i32 form: See the comment for ROTHM v8i16.
2738 RI7Form<0b10011110000, (outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2739 "rotmi\t$rT, $rA, $val", RotateShift,
2740 [(set (v4i32 VECREG:$rT),
2741 (SPUvec_srl VECREG:$rA, (i32 uimm7:$val)))]>;
2743 def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), (i16 uimm7:$val)),
2744 (ROTMIv4i32 VECREG:$rA, (TO_IMM32 uimm7:$val))>;
2746 def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), (i8 uimm7:$val)),
2747 (ROTMIv4i32 VECREG:$rA, (TO_IMM32 uimm7:$val))>;
2749 // ROTMI r32 form: know how to complement the immediate value.
2751 RI7Form<0b10011110000, (outs R32C:$rT), (ins R32C:$rA, rotNeg7imm:$val),
2752 "rotmi\t$rT, $rA, $val", RotateShift,
2753 [(set R32C:$rT, (srl R32C:$rA, (i32 uimm7:$val)))]>;
2755 def : Pat<(srl R32C:$rA, (i16 imm:$val)),
2756 (ROTMIr32 R32C:$rA, (TO_IMM32 uimm7:$val))>;
2758 def : Pat<(srl R32C:$rA, (i8 imm:$val)),
2759 (ROTMIr32 R32C:$rA, (TO_IMM32 uimm7:$val))>;
2761 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2762 // ROTQMBY: This is a vector form merely so that when used in an
2763 // instruction pattern, type checking will succeed. This instruction assumes
2764 // that the user knew to negate $rB.
2765 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2767 class ROTQMBYInst<dag OOL, dag IOL, list<dag> pattern>:
2768 RRForm<0b10111011100, OOL, IOL, "rotqmby\t$rT, $rA, $rB",
2769 RotateShift, pattern>;
2771 class ROTQMBYVecInst<ValueType vectype>:
2772 ROTQMBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2773 [/* no pattern, $rB must be negated */]>;
2775 class ROTQMBYRegInst<RegisterClass rclass>:
2776 ROTQMBYInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2777 [/* no pattern */]>;
2779 multiclass RotateQuadBytes
2781 def v16i8: ROTQMBYVecInst<v16i8>;
2782 def v8i16: ROTQMBYVecInst<v8i16>;
2783 def v4i32: ROTQMBYVecInst<v4i32>;
2784 def v2i64: ROTQMBYVecInst<v2i64>;
2786 def r128: ROTQMBYRegInst<GPRC>;
2787 def r64: ROTQMBYRegInst<R64C>;
2790 defm ROTQMBY : RotateQuadBytes;
2792 class ROTQMBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2793 RI7Form<0b10111111100, OOL, IOL, "rotqmbyi\t$rT, $rA, $val",
2794 RotateShift, pattern>;
2796 class ROTQMBYIVecInst<ValueType vectype>:
2797 ROTQMBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2798 [/* no pattern */]>;
2800 class ROTQMBYIRegInst<RegisterClass rclass, Operand optype, ValueType inttype,
2802 ROTQMBYIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2803 [/* no pattern */]>;
2805 // 128-bit zero extension form:
2806 class ROTQMBYIZExtInst<RegisterClass rclass, Operand optype, PatLeaf pred>:
2807 ROTQMBYIInst<(outs GPRC:$rT), (ins rclass:$rA, optype:$val),
2808 [/* no pattern */]>;
2810 multiclass RotateQuadBytesImm
2812 def v16i8: ROTQMBYIVecInst<v16i8>;
2813 def v8i16: ROTQMBYIVecInst<v8i16>;
2814 def v4i32: ROTQMBYIVecInst<v4i32>;
2815 def v2i64: ROTQMBYIVecInst<v2i64>;
2817 def r128: ROTQMBYIRegInst<GPRC, rotNeg7imm, i32, uimm7>;
2818 def r64: ROTQMBYIRegInst<R64C, rotNeg7imm, i32, uimm7>;
2820 def r128_zext_r8: ROTQMBYIZExtInst<R8C, rotNeg7imm, uimm7>;
2821 def r128_zext_r16: ROTQMBYIZExtInst<R16C, rotNeg7imm, uimm7>;
2822 def r128_zext_r32: ROTQMBYIZExtInst<R32C, rotNeg7imm, uimm7>;
2823 def r128_zext_r64: ROTQMBYIZExtInst<R64C, rotNeg7imm, uimm7>;
2826 defm ROTQMBYI : RotateQuadBytesImm;
2828 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2829 // Rotate right and mask by bit count
2830 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2832 class ROTQMBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2833 RRForm<0b10110011100, OOL, IOL, "rotqmbybi\t$rT, $rA, $rB",
2834 RotateShift, pattern>;
2836 class ROTQMBYBIVecInst<ValueType vectype>:
2837 ROTQMBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2838 [/* no pattern, */]>;
2840 multiclass RotateMaskQuadByBitCount
2842 def v16i8: ROTQMBYBIVecInst<v16i8>;
2843 def v8i16: ROTQMBYBIVecInst<v8i16>;
2844 def v4i32: ROTQMBYBIVecInst<v4i32>;
2845 def v2i64: ROTQMBYBIVecInst<v2i64>;
2848 defm ROTQMBYBI: RotateMaskQuadByBitCount;
2850 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2851 // Rotate quad and mask by bits
2852 // Note that the rotate amount has to be negated
2853 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2855 class ROTQMBIInst<dag OOL, dag IOL, list<dag> pattern>:
2856 RRForm<0b10011011100, OOL, IOL, "rotqmbi\t$rT, $rA, $rB",
2857 RotateShift, pattern>;
2859 class ROTQMBIVecInst<ValueType vectype>:
2860 ROTQMBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2861 [/* no pattern */]>;
2863 class ROTQMBIRegInst<RegisterClass rclass>:
2864 ROTQMBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2865 [/* no pattern */]>;
2867 multiclass RotateMaskQuadByBits
2869 def v16i8: ROTQMBIVecInst<v16i8>;
2870 def v8i16: ROTQMBIVecInst<v8i16>;
2871 def v4i32: ROTQMBIVecInst<v4i32>;
2872 def v2i64: ROTQMBIVecInst<v2i64>;
2874 def r128: ROTQMBIRegInst<GPRC>;
2875 def r64: ROTQMBIRegInst<R64C>;
2878 defm ROTQMBI: RotateMaskQuadByBits;
2880 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2881 // Rotate quad and mask by bits, immediate
2882 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2884 class ROTQMBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2885 RI7Form<0b10011111100, OOL, IOL, "rotqmbii\t$rT, $rA, $val",
2886 RotateShift, pattern>;
2888 class ROTQMBIIVecInst<ValueType vectype>:
2889 ROTQMBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2890 [/* no pattern */]>;
2892 class ROTQMBIIRegInst<RegisterClass rclass>:
2893 ROTQMBIIInst<(outs rclass:$rT), (ins rclass:$rA, rotNeg7imm:$val),
2894 [/* no pattern */]>;
2896 multiclass RotateMaskQuadByBitsImm
2898 def v16i8: ROTQMBIIVecInst<v16i8>;
2899 def v8i16: ROTQMBIIVecInst<v8i16>;
2900 def v4i32: ROTQMBIIVecInst<v4i32>;
2901 def v2i64: ROTQMBIIVecInst<v2i64>;
2903 def r128: ROTQMBIIRegInst<GPRC>;
2904 def r64: ROTQMBIIRegInst<R64C>;
2907 defm ROTQMBII: RotateMaskQuadByBitsImm;
2909 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2910 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2913 RRForm<0b01111010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2914 "rotmah\t$rT, $rA, $rB", RotateShift,
2915 [/* see patterns below - $rB must be negated */]>;
2917 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), R32C:$rB),
2918 (ROTMAHv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2920 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), R16C:$rB),
2921 (ROTMAHv8i16 VECREG:$rA,
2922 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2924 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), R8C:$rB),
2925 (ROTMAHv8i16 VECREG:$rA,
2926 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2929 RRForm<0b01111010000, (outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2930 "rotmah\t$rT, $rA, $rB", RotateShift,
2931 [/* see patterns below - $rB must be negated */]>;
2933 def : Pat<(sra R16C:$rA, R32C:$rB),
2934 (ROTMAHr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2936 def : Pat<(sra R16C:$rA, R16C:$rB),
2937 (ROTMAHr16 R16C:$rA,
2938 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2940 def : Pat<(sra R16C:$rA, R8C:$rB),
2941 (ROTMAHr16 R16C:$rA,
2942 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2945 RRForm<0b01111110000, (outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2946 "rotmahi\t$rT, $rA, $val", RotateShift,
2947 [(set (v8i16 VECREG:$rT),
2948 (SPUvec_sra (v8i16 VECREG:$rA), (i32 uimm7:$val)))]>;
2950 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i16 uimm7:$val)),
2951 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (TO_IMM32 uimm7:$val))>;
2953 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i8 uimm7:$val)),
2954 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (TO_IMM32 uimm7:$val))>;
2957 RRForm<0b01111110000, (outs R16C:$rT), (ins R16C:$rA, rothNeg7imm_i16:$val),
2958 "rotmahi\t$rT, $rA, $val", RotateShift,
2959 [(set R16C:$rT, (sra R16C:$rA, (i16 uimm7:$val)))]>;
2961 def : Pat<(sra R16C:$rA, (i32 imm:$val)),
2962 (ROTMAHIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
2964 def : Pat<(sra R16C:$rA, (i8 imm:$val)),
2965 (ROTMAHIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
2968 RRForm<0b01011010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2969 "rotma\t$rT, $rA, $rB", RotateShift,
2970 [/* see patterns below - $rB must be negated */]>;
2972 def : Pat<(SPUvec_sra (v4i32 VECREG:$rA), R32C:$rB),
2973 (ROTMAv4i32 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2975 def : Pat<(SPUvec_sra (v4i32 VECREG:$rA), R16C:$rB),
2976 (ROTMAv4i32 VECREG:$rA,
2977 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2979 def : Pat<(SPUvec_sra (v4i32 VECREG:$rA), R8C:$rB),
2980 (ROTMAv4i32 VECREG:$rA,
2981 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2984 RRForm<0b01011010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2985 "rotma\t$rT, $rA, $rB", RotateShift,
2986 [/* see patterns below - $rB must be negated */]>;
2988 def : Pat<(sra R32C:$rA, R32C:$rB),
2989 (ROTMAr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
2991 def : Pat<(sra R32C:$rA, R16C:$rB),
2993 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2995 def : Pat<(sra R32C:$rA, R8C:$rB),
2997 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2999 class ROTMAIInst<dag OOL, dag IOL, list<dag> pattern>:
3000 RRForm<0b01011110000, OOL, IOL,
3001 "rotmai\t$rT, $rA, $val",
3002 RotateShift, pattern>;
3004 class ROTMAIVecInst<ValueType vectype, Operand intop, ValueType inttype>:
3005 ROTMAIInst<(outs VECREG:$rT), (ins VECREG:$rA, intop:$val),
3006 [(set (vectype VECREG:$rT),
3007 (SPUvec_sra VECREG:$rA, (inttype uimm7:$val)))]>;
3009 class ROTMAIRegInst<RegisterClass rclass, Operand intop, ValueType inttype>:
3010 ROTMAIInst<(outs rclass:$rT), (ins rclass:$rA, intop:$val),
3011 [(set rclass:$rT, (sra rclass:$rA, (inttype uimm7:$val)))]>;
3013 multiclass RotateMaskAlgebraicImm {
3014 def v2i64_i32 : ROTMAIVecInst<v2i64, rotNeg7imm, i32>;
3015 def v4i32_i32 : ROTMAIVecInst<v4i32, rotNeg7imm, i32>;
3016 def r64_i32 : ROTMAIRegInst<R64C, rotNeg7imm, i32>;
3017 def r32_i32 : ROTMAIRegInst<R32C, rotNeg7imm, i32>;
3020 defm ROTMAI : RotateMaskAlgebraicImm;
3022 //===----------------------------------------------------------------------===//
3023 // Branch and conditionals:
3024 //===----------------------------------------------------------------------===//
3026 let isTerminator = 1, isBarrier = 1 in {
3027 // Halt If Equal (r32 preferred slot only, no vector form)
3029 RRForm_3<0b00011011110, (outs), (ins R32C:$rA, R32C:$rB),
3030 "heq\t$rA, $rB", BranchResolv,
3031 [/* no pattern to match */]>;
3034 RI10Form_2<0b11111110, (outs), (ins R32C:$rA, s10imm:$val),
3035 "heqi\t$rA, $val", BranchResolv,
3036 [/* no pattern to match */]>;
3038 // HGT/HGTI: These instructions use signed arithmetic for the comparison,
3039 // contrasting with HLGT/HLGTI, which use unsigned comparison:
3041 RRForm_3<0b00011010010, (outs), (ins R32C:$rA, R32C:$rB),
3042 "hgt\t$rA, $rB", BranchResolv,
3043 [/* no pattern to match */]>;
3046 RI10Form_2<0b11110010, (outs), (ins R32C:$rA, s10imm:$val),
3047 "hgti\t$rA, $val", BranchResolv,
3048 [/* no pattern to match */]>;
3051 RRForm_3<0b00011011010, (outs), (ins R32C:$rA, R32C:$rB),
3052 "hlgt\t$rA, $rB", BranchResolv,
3053 [/* no pattern to match */]>;
3056 RI10Form_2<0b11111010, (outs), (ins R32C:$rA, s10imm:$val),
3057 "hlgti\t$rA, $val", BranchResolv,
3058 [/* no pattern to match */]>;
3061 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3062 // Comparison operators for i8, i16 and i32:
3063 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3065 class CEQBInst<dag OOL, dag IOL, list<dag> pattern> :
3066 RRForm<0b00001011110, OOL, IOL, "ceqb\t$rT, $rA, $rB",
3069 multiclass CmpEqualByte
3072 CEQBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3073 [(set (v16i8 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
3074 (v8i16 VECREG:$rB)))]>;
3077 CEQBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3078 [(set R8C:$rT, (seteq R8C:$rA, R8C:$rB))]>;
3081 class CEQBIInst<dag OOL, dag IOL, list<dag> pattern> :
3082 RI10Form<0b01111110, OOL, IOL, "ceqbi\t$rT, $rA, $val",
3085 multiclass CmpEqualByteImm
3088 CEQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3089 [(set (v16i8 VECREG:$rT), (seteq (v16i8 VECREG:$rA),
3090 v16i8SExt8Imm:$val))]>;
3092 CEQBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3093 [(set R8C:$rT, (seteq R8C:$rA, immSExt8:$val))]>;
3096 class CEQHInst<dag OOL, dag IOL, list<dag> pattern> :
3097 RRForm<0b00010011110, OOL, IOL, "ceqh\t$rT, $rA, $rB",
3100 multiclass CmpEqualHalfword
3102 def v8i16 : CEQHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3103 [(set (v8i16 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
3104 (v8i16 VECREG:$rB)))]>;
3106 def r16 : CEQHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3107 [(set R16C:$rT, (seteq R16C:$rA, R16C:$rB))]>;
3110 class CEQHIInst<dag OOL, dag IOL, list<dag> pattern> :
3111 RI10Form<0b10111110, OOL, IOL, "ceqhi\t$rT, $rA, $val",
3114 multiclass CmpEqualHalfwordImm
3116 def v8i16 : CEQHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3117 [(set (v8i16 VECREG:$rT),
3118 (seteq (v8i16 VECREG:$rA),
3119 (v8i16 v8i16SExt10Imm:$val)))]>;
3120 def r16 : CEQHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3121 [(set R16C:$rT, (seteq R16C:$rA, i16ImmSExt10:$val))]>;
3124 class CEQInst<dag OOL, dag IOL, list<dag> pattern> :
3125 RRForm<0b00000011110, OOL, IOL, "ceq\t$rT, $rA, $rB",
3128 multiclass CmpEqualWord
3130 def v4i32 : CEQInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3131 [(set (v4i32 VECREG:$rT),
3132 (seteq (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3134 def r32 : CEQInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3135 [(set R32C:$rT, (seteq R32C:$rA, R32C:$rB))]>;
3138 class CEQIInst<dag OOL, dag IOL, list<dag> pattern> :
3139 RI10Form<0b00111110, OOL, IOL, "ceqi\t$rT, $rA, $val",
3142 multiclass CmpEqualWordImm
3144 def v4i32 : CEQIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3145 [(set (v4i32 VECREG:$rT),
3146 (seteq (v4i32 VECREG:$rA),
3147 (v4i32 v4i32SExt16Imm:$val)))]>;
3149 def r32: CEQIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3150 [(set R32C:$rT, (seteq R32C:$rA, i32ImmSExt10:$val))]>;
3153 class CGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3154 RRForm<0b00001010010, OOL, IOL, "cgtb\t$rT, $rA, $rB",
3157 multiclass CmpGtrByte
3160 CGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3161 [(set (v16i8 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3162 (v8i16 VECREG:$rB)))]>;
3165 CGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3166 [(set R8C:$rT, (setgt R8C:$rA, R8C:$rB))]>;
3169 class CGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3170 RI10Form<0b01110010, OOL, IOL, "cgtbi\t$rT, $rA, $val",
3173 multiclass CmpGtrByteImm
3176 CGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3177 [(set (v16i8 VECREG:$rT), (setgt (v16i8 VECREG:$rA),
3178 v16i8SExt8Imm:$val))]>;
3180 CGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3181 [(set R8C:$rT, (setgt R8C:$rA, immSExt8:$val))]>;
3184 class CGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3185 RRForm<0b00010010010, OOL, IOL, "cgth\t$rT, $rA, $rB",
3188 multiclass CmpGtrHalfword
3190 def v8i16 : CGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3191 [(set (v8i16 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3192 (v8i16 VECREG:$rB)))]>;
3194 def r16 : CGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3195 [(set R16C:$rT, (setgt R16C:$rA, R16C:$rB))]>;
3198 class CGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3199 RI10Form<0b10110010, OOL, IOL, "cgthi\t$rT, $rA, $val",
3202 multiclass CmpGtrHalfwordImm
3204 def v8i16 : CGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3205 [(set (v8i16 VECREG:$rT),
3206 (setgt (v8i16 VECREG:$rA),
3207 (v8i16 v8i16SExt10Imm:$val)))]>;
3208 def r16 : CGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3209 [(set R16C:$rT, (setgt R16C:$rA, i16ImmSExt10:$val))]>;
3212 class CGTInst<dag OOL, dag IOL, list<dag> pattern> :
3213 RRForm<0b00000010010, OOL, IOL, "cgt\t$rT, $rA, $rB",
3216 multiclass CmpGtrWord
3218 def v4i32 : CGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3219 [(set (v4i32 VECREG:$rT),
3220 (setgt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3222 def r32 : CGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3223 [(set R32C:$rT, (setgt R32C:$rA, R32C:$rB))]>;
3226 class CGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3227 RI10Form<0b00110010, OOL, IOL, "cgti\t$rT, $rA, $val",
3230 multiclass CmpGtrWordImm
3232 def v4i32 : CGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3233 [(set (v4i32 VECREG:$rT),
3234 (setgt (v4i32 VECREG:$rA),
3235 (v4i32 v4i32SExt16Imm:$val)))]>;
3237 def r32: CGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3238 [(set R32C:$rT, (setgt R32C:$rA, i32ImmSExt10:$val))]>;
3240 // CGTIv4f32, CGTIf32: These are used in the f32 fdiv instruction sequence:
3241 def v4f32: CGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3242 [(set (v4i32 VECREG:$rT),
3243 (setgt (v4i32 (bitconvert (v4f32 VECREG:$rA))),
3244 (v4i32 v4i32SExt16Imm:$val)))]>;
3246 def f32: CGTIInst<(outs R32C:$rT), (ins R32FP:$rA, s10imm_i32:$val),
3247 [/* no pattern */]>;
3250 class CLGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3251 RRForm<0b00001011010, OOL, IOL, "clgtb\t$rT, $rA, $rB",
3254 multiclass CmpLGtrByte
3257 CLGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3258 [(set (v16i8 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3259 (v8i16 VECREG:$rB)))]>;
3262 CLGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3263 [(set R8C:$rT, (setugt R8C:$rA, R8C:$rB))]>;
3266 class CLGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3267 RI10Form<0b01111010, OOL, IOL, "clgtbi\t$rT, $rA, $val",
3270 multiclass CmpLGtrByteImm
3273 CLGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3274 [(set (v16i8 VECREG:$rT), (setugt (v16i8 VECREG:$rA),
3275 v16i8SExt8Imm:$val))]>;
3277 CLGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3278 [(set R8C:$rT, (setugt R8C:$rA, immSExt8:$val))]>;
3281 class CLGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3282 RRForm<0b00010011010, OOL, IOL, "clgth\t$rT, $rA, $rB",
3285 multiclass CmpLGtrHalfword
3287 def v8i16 : CLGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3288 [(set (v8i16 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3289 (v8i16 VECREG:$rB)))]>;
3291 def r16 : CLGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3292 [(set R16C:$rT, (setugt R16C:$rA, R16C:$rB))]>;
3295 class CLGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3296 RI10Form<0b10111010, OOL, IOL, "clgthi\t$rT, $rA, $val",
3299 multiclass CmpLGtrHalfwordImm
3301 def v8i16 : CLGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3302 [(set (v8i16 VECREG:$rT),
3303 (setugt (v8i16 VECREG:$rA),
3304 (v8i16 v8i16SExt10Imm:$val)))]>;
3305 def r16 : CLGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3306 [(set R16C:$rT, (setugt R16C:$rA, i16ImmSExt10:$val))]>;
3309 class CLGTInst<dag OOL, dag IOL, list<dag> pattern> :
3310 RRForm<0b00000011010, OOL, IOL, "clgt\t$rT, $rA, $rB",
3313 multiclass CmpLGtrWord
3315 def v4i32 : CLGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3316 [(set (v4i32 VECREG:$rT),
3317 (setugt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3319 def r32 : CLGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3320 [(set R32C:$rT, (setugt R32C:$rA, R32C:$rB))]>;
3323 class CLGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3324 RI10Form<0b00111010, OOL, IOL, "clgti\t$rT, $rA, $val",
3327 multiclass CmpLGtrWordImm
3329 def v4i32 : CLGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3330 [(set (v4i32 VECREG:$rT),
3331 (setugt (v4i32 VECREG:$rA),
3332 (v4i32 v4i32SExt16Imm:$val)))]>;
3334 def r32: CLGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3335 [(set R32C:$rT, (setugt R32C:$rA, i32ImmSExt10:$val))]>;
3338 defm CEQB : CmpEqualByte;
3339 defm CEQBI : CmpEqualByteImm;
3340 defm CEQH : CmpEqualHalfword;
3341 defm CEQHI : CmpEqualHalfwordImm;
3342 defm CEQ : CmpEqualWord;
3343 defm CEQI : CmpEqualWordImm;
3344 defm CGTB : CmpGtrByte;
3345 defm CGTBI : CmpGtrByteImm;
3346 defm CGTH : CmpGtrHalfword;
3347 defm CGTHI : CmpGtrHalfwordImm;
3348 defm CGT : CmpGtrWord;
3349 defm CGTI : CmpGtrWordImm;
3350 defm CLGTB : CmpLGtrByte;
3351 defm CLGTBI : CmpLGtrByteImm;
3352 defm CLGTH : CmpLGtrHalfword;
3353 defm CLGTHI : CmpLGtrHalfwordImm;
3354 defm CLGT : CmpLGtrWord;
3355 defm CLGTI : CmpLGtrWordImm;
3357 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3358 // For SETCC primitives not supported above (setlt, setle, setge, etc.)
3359 // define a pattern to generate the right code, as a binary operator
3360 // (in a manner of speaking.)
3363 // 1. This only matches the setcc set of conditionals. Special pattern
3364 // matching is used for select conditionals.
3366 // 2. The "DAG" versions of these classes is almost exclusively used for
3367 // i64 comparisons. See the tblgen fundamentals documentation for what
3368 // ".ResultInstrs[0]" means; see TargetSelectionDAG.td and the Pattern
3369 // class for where ResultInstrs originates.
3370 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3372 class SETCCNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3373 SPUInstr xorinst, SPUInstr cmpare>:
3374 Pat<(cond rclass:$rA, rclass:$rB),
3375 (xorinst (cmpare rclass:$rA, rclass:$rB), (inttype -1))>;
3377 class SETCCNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3378 PatLeaf immpred, SPUInstr xorinst, SPUInstr cmpare>:
3379 Pat<(cond rclass:$rA, (inttype immpred:$imm)),
3380 (xorinst (cmpare rclass:$rA, (inttype immpred:$imm)), (inttype -1))>;
3382 def : SETCCNegCondReg<setne, R8C, i8, XORBIr8, CEQBr8>;
3383 def : SETCCNegCondImm<setne, R8C, i8, immSExt8, XORBIr8, CEQBIr8>;
3385 def : SETCCNegCondReg<setne, R16C, i16, XORHIr16, CEQHr16>;
3386 def : SETCCNegCondImm<setne, R16C, i16, i16ImmSExt10, XORHIr16, CEQHIr16>;
3388 def : SETCCNegCondReg<setne, R32C, i32, XORIr32, CEQr32>;
3389 def : SETCCNegCondImm<setne, R32C, i32, i32ImmSExt10, XORIr32, CEQIr32>;
3391 class SETCCBinOpReg<PatFrag cond, RegisterClass rclass,
3392 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3393 Pat<(cond rclass:$rA, rclass:$rB),
3394 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3395 (cmpOp2 rclass:$rA, rclass:$rB))>;
3397 class SETCCBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3399 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3400 Pat<(cond rclass:$rA, (immtype immpred:$imm)),
3401 (binop (cmpOp1 rclass:$rA, (immtype immpred:$imm)),
3402 (cmpOp2 rclass:$rA, (immtype immpred:$imm)))>;
3404 def : SETCCBinOpReg<setge, R8C, ORr8, CGTBr8, CEQBr8>;
3405 def : SETCCBinOpImm<setge, R8C, immSExt8, i8, ORr8, CGTBIr8, CEQBIr8>;
3406 def : SETCCBinOpReg<setlt, R8C, NORr8, CGTBr8, CEQBr8>;
3407 def : SETCCBinOpImm<setlt, R8C, immSExt8, i8, NORr8, CGTBIr8, CEQBIr8>;
3408 def : Pat<(setle R8C:$rA, R8C:$rB),
3409 (XORBIr8 (CGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3410 def : Pat<(setle R8C:$rA, immU8:$imm),
3411 (XORBIr8 (CGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3413 def : SETCCBinOpReg<setge, R16C, ORr16, CGTHr16, CEQHr16>;
3414 def : SETCCBinOpImm<setge, R16C, i16ImmSExt10, i16,
3415 ORr16, CGTHIr16, CEQHIr16>;
3416 def : SETCCBinOpReg<setlt, R16C, NORr16, CGTHr16, CEQHr16>;
3417 def : SETCCBinOpImm<setlt, R16C, i16ImmSExt10, i16, NORr16, CGTHIr16, CEQHIr16>;
3418 def : Pat<(setle R16C:$rA, R16C:$rB),
3419 (XORHIr16 (CGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3420 def : Pat<(setle R16C:$rA, i16ImmSExt10:$imm),
3421 (XORHIr16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3423 def : SETCCBinOpReg<setge, R32C, ORr32, CGTr32, CEQr32>;
3424 def : SETCCBinOpImm<setge, R32C, i32ImmSExt10, i32,
3425 ORr32, CGTIr32, CEQIr32>;
3426 def : SETCCBinOpReg<setlt, R32C, NORr32, CGTr32, CEQr32>;
3427 def : SETCCBinOpImm<setlt, R32C, i32ImmSExt10, i32, NORr32, CGTIr32, CEQIr32>;
3428 def : Pat<(setle R32C:$rA, R32C:$rB),
3429 (XORIr32 (CGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3430 def : Pat<(setle R32C:$rA, i32ImmSExt10:$imm),
3431 (XORIr32 (CGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3433 def : SETCCBinOpReg<setuge, R8C, ORr8, CLGTBr8, CEQBr8>;
3434 def : SETCCBinOpImm<setuge, R8C, immSExt8, i8, ORr8, CLGTBIr8, CEQBIr8>;
3435 def : SETCCBinOpReg<setult, R8C, NORr8, CLGTBr8, CEQBr8>;
3436 def : SETCCBinOpImm<setult, R8C, immSExt8, i8, NORr8, CLGTBIr8, CEQBIr8>;
3437 def : Pat<(setule R8C:$rA, R8C:$rB),
3438 (XORBIr8 (CLGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3439 def : Pat<(setule R8C:$rA, immU8:$imm),
3440 (XORBIr8 (CLGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3442 def : SETCCBinOpReg<setuge, R16C, ORr16, CLGTHr16, CEQHr16>;
3443 def : SETCCBinOpImm<setuge, R16C, i16ImmSExt10, i16,
3444 ORr16, CLGTHIr16, CEQHIr16>;
3445 def : SETCCBinOpReg<setult, R16C, NORr16, CLGTHr16, CEQHr16>;
3446 def : SETCCBinOpImm<setult, R16C, i16ImmSExt10, i16, NORr16,
3447 CLGTHIr16, CEQHIr16>;
3448 def : Pat<(setule R16C:$rA, R16C:$rB),
3449 (XORHIr16 (CLGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3450 def : Pat<(setule R16C:$rA, i16ImmSExt10:$imm),
3451 (XORHIr16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3453 def : SETCCBinOpReg<setuge, R32C, ORr32, CLGTr32, CEQr32>;
3454 def : SETCCBinOpImm<setuge, R32C, i32ImmSExt10, i32,
3455 ORr32, CLGTIr32, CEQIr32>;
3456 def : SETCCBinOpReg<setult, R32C, NORr32, CLGTr32, CEQr32>;
3457 def : SETCCBinOpImm<setult, R32C, i32ImmSExt10, i32, NORr32, CLGTIr32, CEQIr32>;
3458 def : Pat<(setule R32C:$rA, R32C:$rB),
3459 (XORIr32 (CLGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3460 def : Pat<(setule R32C:$rA, i32ImmSExt10:$imm),
3461 (XORIr32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3463 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3464 // select conditional patterns:
3465 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3467 class SELECTNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3468 SPUInstr selinstr, SPUInstr cmpare>:
3469 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3470 rclass:$rTrue, rclass:$rFalse),
3471 (selinstr rclass:$rTrue, rclass:$rFalse,
3472 (cmpare rclass:$rA, rclass:$rB))>;
3474 class SELECTNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3475 PatLeaf immpred, SPUInstr selinstr, SPUInstr cmpare>:
3476 Pat<(select (inttype (cond rclass:$rA, immpred:$imm)),
3477 rclass:$rTrue, rclass:$rFalse),
3478 (selinstr rclass:$rTrue, rclass:$rFalse,
3479 (cmpare rclass:$rA, immpred:$imm))>;
3481 def : SELECTNegCondReg<setne, R8C, i8, SELBr8, CEQBr8>;
3482 def : SELECTNegCondImm<setne, R8C, i8, immSExt8, SELBr8, CEQBIr8>;
3483 def : SELECTNegCondReg<setle, R8C, i8, SELBr8, CGTBr8>;
3484 def : SELECTNegCondImm<setle, R8C, i8, immSExt8, SELBr8, CGTBr8>;
3485 def : SELECTNegCondReg<setule, R8C, i8, SELBr8, CLGTBr8>;
3486 def : SELECTNegCondImm<setule, R8C, i8, immU8, SELBr8, CLGTBIr8>;
3488 def : SELECTNegCondReg<setne, R16C, i16, SELBr16, CEQHr16>;
3489 def : SELECTNegCondImm<setne, R16C, i16, i16ImmSExt10, SELBr16, CEQHIr16>;
3490 def : SELECTNegCondReg<setle, R16C, i16, SELBr16, CGTHr16>;
3491 def : SELECTNegCondImm<setle, R16C, i16, i16ImmSExt10, SELBr16, CGTHIr16>;
3492 def : SELECTNegCondReg<setule, R16C, i16, SELBr16, CLGTHr16>;
3493 def : SELECTNegCondImm<setule, R16C, i16, i16ImmSExt10, SELBr16, CLGTHIr16>;
3495 def : SELECTNegCondReg<setne, R32C, i32, SELBr32, CEQr32>;
3496 def : SELECTNegCondImm<setne, R32C, i32, i32ImmSExt10, SELBr32, CEQIr32>;
3497 def : SELECTNegCondReg<setle, R32C, i32, SELBr32, CGTr32>;
3498 def : SELECTNegCondImm<setle, R32C, i32, i32ImmSExt10, SELBr32, CGTIr32>;
3499 def : SELECTNegCondReg<setule, R32C, i32, SELBr32, CLGTr32>;
3500 def : SELECTNegCondImm<setule, R32C, i32, i32ImmSExt10, SELBr32, CLGTIr32>;
3502 class SELECTBinOpReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3503 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3505 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3506 rclass:$rTrue, rclass:$rFalse),
3507 (selinstr rclass:$rFalse, rclass:$rTrue,
3508 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3509 (cmpOp2 rclass:$rA, rclass:$rB)))>;
3511 class SELECTBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3513 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3515 Pat<(select (inttype (cond rclass:$rA, (inttype immpred:$imm))),
3516 rclass:$rTrue, rclass:$rFalse),
3517 (selinstr rclass:$rFalse, rclass:$rTrue,
3518 (binop (cmpOp1 rclass:$rA, (inttype immpred:$imm)),
3519 (cmpOp2 rclass:$rA, (inttype immpred:$imm))))>;
3521 def : SELECTBinOpReg<setge, R8C, i8, SELBr8, ORr8, CGTBr8, CEQBr8>;
3522 def : SELECTBinOpImm<setge, R8C, immSExt8, i8,
3523 SELBr8, ORr8, CGTBIr8, CEQBIr8>;
3525 def : SELECTBinOpReg<setge, R16C, i16, SELBr16, ORr16, CGTHr16, CEQHr16>;
3526 def : SELECTBinOpImm<setge, R16C, i16ImmSExt10, i16,
3527 SELBr16, ORr16, CGTHIr16, CEQHIr16>;
3529 def : SELECTBinOpReg<setge, R32C, i32, SELBr32, ORr32, CGTr32, CEQr32>;
3530 def : SELECTBinOpImm<setge, R32C, i32ImmSExt10, i32,
3531 SELBr32, ORr32, CGTIr32, CEQIr32>;
3533 def : SELECTBinOpReg<setuge, R8C, i8, SELBr8, ORr8, CLGTBr8, CEQBr8>;
3534 def : SELECTBinOpImm<setuge, R8C, immSExt8, i8,
3535 SELBr8, ORr8, CLGTBIr8, CEQBIr8>;
3537 def : SELECTBinOpReg<setuge, R16C, i16, SELBr16, ORr16, CLGTHr16, CEQHr16>;
3538 def : SELECTBinOpImm<setuge, R16C, i16ImmUns10, i16,
3539 SELBr16, ORr16, CLGTHIr16, CEQHIr16>;
3541 def : SELECTBinOpReg<setuge, R32C, i32, SELBr32, ORr32, CLGTr32, CEQr32>;
3542 def : SELECTBinOpImm<setuge, R32C, i32ImmUns10, i32,
3543 SELBr32, ORr32, CLGTIr32, CEQIr32>;
3545 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3548 // All calls clobber the non-callee-saved registers:
3549 Defs = [R0, R1, R2, R3, R4, R5, R6, R7, R8, R9,
3550 R10,R11,R12,R13,R14,R15,R16,R17,R18,R19,
3551 R20,R21,R22,R23,R24,R25,R26,R27,R28,R29,
3552 R30,R31,R32,R33,R34,R35,R36,R37,R38,R39,
3553 R40,R41,R42,R43,R44,R45,R46,R47,R48,R49,
3554 R50,R51,R52,R53,R54,R55,R56,R57,R58,R59,
3555 R60,R61,R62,R63,R64,R65,R66,R67,R68,R69,
3556 R70,R71,R72,R73,R74,R75,R76,R77,R78,R79],
3557 // All of these instructions use $lr (aka $0)
3559 // Branch relative and set link: Used if we actually know that the target
3560 // is within [-32768, 32767] bytes of the target
3562 BranchSetLink<0b011001100, (outs), (ins relcalltarget:$func, variable_ops),
3563 "brsl\t$$lr, $func",
3564 [(SPUcall (SPUpcrel tglobaladdr:$func, 0))]>;
3566 // Branch absolute and set link: Used if we actually know that the target
3567 // is an absolute address
3569 BranchSetLink<0b011001100, (outs), (ins calltarget:$func, variable_ops),
3570 "brasl\t$$lr, $func",
3571 [(SPUcall (SPUaform tglobaladdr:$func, 0))]>;
3573 // Branch indirect and set link if external data. These instructions are not
3574 // actually generated, matched by an intrinsic:
3575 def BISLED_00: BISLEDForm<0b11, "bisled\t$$lr, $func", [/* empty pattern */]>;
3576 def BISLED_E0: BISLEDForm<0b10, "bisled\t$$lr, $func", [/* empty pattern */]>;
3577 def BISLED_0D: BISLEDForm<0b01, "bisled\t$$lr, $func", [/* empty pattern */]>;
3578 def BISLED_ED: BISLEDForm<0b00, "bisled\t$$lr, $func", [/* empty pattern */]>;
3580 // Branch indirect and set link. This is the "X-form" address version of a
3583 BIForm<0b10010101100, "bisl\t$$lr, $func", [(SPUcall R32C:$func)]>;
3586 // Support calls to external symbols:
3587 def : Pat<(SPUcall (SPUpcrel texternalsym:$func, 0)),
3588 (BRSL texternalsym:$func)>;
3590 def : Pat<(SPUcall (SPUaform texternalsym:$func, 0)),
3591 (BRASL texternalsym:$func)>;
3593 // Unconditional branches:
3594 let isBranch = 1, isTerminator = 1, hasCtrlDep = 1 in {
3595 let isBarrier = 1 in {
3597 UncondBranch<0b001001100, (outs), (ins brtarget:$dest),
3601 // Unconditional, absolute address branch
3603 UncondBranch<0b001100000, (outs), (ins brtarget:$dest),
3605 [/* no pattern */]>;
3609 BIForm<0b00010101100, "bi\t$func", [(brind R32C:$func)]>;
3612 // Conditional branches:
3613 class BRNZInst<dag IOL, list<dag> pattern>:
3614 RI16Form<0b010000100, (outs), IOL, "brnz\t$rCond,$dest",
3615 BranchResolv, pattern>;
3617 class BRNZRegInst<RegisterClass rclass>:
3618 BRNZInst<(ins rclass:$rCond, brtarget:$dest),
3619 [(brcond rclass:$rCond, bb:$dest)]>;
3621 class BRNZVecInst<ValueType vectype>:
3622 BRNZInst<(ins VECREG:$rCond, brtarget:$dest),
3623 [(brcond (vectype VECREG:$rCond), bb:$dest)]>;
3625 multiclass BranchNotZero {
3626 def v4i32 : BRNZVecInst<v4i32>;
3627 def r32 : BRNZRegInst<R32C>;
3630 defm BRNZ : BranchNotZero;
3632 class BRZInst<dag IOL, list<dag> pattern>:
3633 RI16Form<0b000000100, (outs), IOL, "brz\t$rT,$dest",
3634 BranchResolv, pattern>;
3636 class BRZRegInst<RegisterClass rclass>:
3637 BRZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3639 class BRZVecInst<ValueType vectype>:
3640 BRZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3642 multiclass BranchZero {
3643 def v4i32: BRZVecInst<v4i32>;
3644 def r32: BRZRegInst<R32C>;
3647 defm BRZ: BranchZero;
3649 // Note: LLVM doesn't do branch conditional, indirect. Otherwise these would
3652 class BINZInst<dag IOL, list<dag> pattern>:
3653 BICondForm<0b10010100100, (outs), IOL, "binz\t$rA, $dest", pattern>;
3655 class BINZRegInst<RegisterClass rclass>:
3656 BINZInst<(ins rclass:$rA, brtarget:$dest),
3657 [(brcond rclass:$rA, R32C:$dest)]>;
3659 class BINZVecInst<ValueType vectype>:
3660 BINZInst<(ins VECREG:$rA, R32C:$dest),
3661 [(brcond (vectype VECREG:$rA), R32C:$dest)]>;
3663 multiclass BranchNotZeroIndirect {
3664 def v4i32: BINZVecInst<v4i32>;
3665 def r32: BINZRegInst<R32C>;
3668 defm BINZ: BranchNotZeroIndirect;
3670 class BIZInst<dag IOL, list<dag> pattern>:
3671 BICondForm<0b00010100100, (outs), IOL, "biz\t$rA, $func", pattern>;
3673 class BIZRegInst<RegisterClass rclass>:
3674 BIZInst<(ins rclass:$rA, R32C:$func), [/* no pattern */]>;
3676 class BIZVecInst<ValueType vectype>:
3677 BIZInst<(ins VECREG:$rA, R32C:$func), [/* no pattern */]>;
3679 multiclass BranchZeroIndirect {
3680 def v4i32: BIZVecInst<v4i32>;
3681 def r32: BIZRegInst<R32C>;
3684 defm BIZ: BranchZeroIndirect;
3687 class BRHNZInst<dag IOL, list<dag> pattern>:
3688 RI16Form<0b011000100, (outs), IOL, "brhnz\t$rCond,$dest", BranchResolv,
3691 class BRHNZRegInst<RegisterClass rclass>:
3692 BRHNZInst<(ins rclass:$rCond, brtarget:$dest),
3693 [(brcond rclass:$rCond, bb:$dest)]>;
3695 class BRHNZVecInst<ValueType vectype>:
3696 BRHNZInst<(ins VECREG:$rCond, brtarget:$dest), [/* no pattern */]>;
3698 multiclass BranchNotZeroHalfword {
3699 def v8i16: BRHNZVecInst<v8i16>;
3700 def r16: BRHNZRegInst<R16C>;
3703 defm BRHNZ: BranchNotZeroHalfword;
3705 class BRHZInst<dag IOL, list<dag> pattern>:
3706 RI16Form<0b001000100, (outs), IOL, "brhz\t$rT,$dest", BranchResolv,
3709 class BRHZRegInst<RegisterClass rclass>:
3710 BRHZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3712 class BRHZVecInst<ValueType vectype>:
3713 BRHZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3715 multiclass BranchZeroHalfword {
3716 def v8i16: BRHZVecInst<v8i16>;
3717 def r16: BRHZRegInst<R16C>;
3720 defm BRHZ: BranchZeroHalfword;
3723 //===----------------------------------------------------------------------===//
3724 // setcc and brcond patterns:
3725 //===----------------------------------------------------------------------===//
3727 def : Pat<(brcond (i16 (seteq R16C:$rA, 0)), bb:$dest),
3728 (BRHZr16 R16C:$rA, bb:$dest)>;
3729 def : Pat<(brcond (i16 (setne R16C:$rA, 0)), bb:$dest),
3730 (BRHNZr16 R16C:$rA, bb:$dest)>;
3732 def : Pat<(brcond (i32 (seteq R32C:$rA, 0)), bb:$dest),
3733 (BRZr32 R32C:$rA, bb:$dest)>;
3734 def : Pat<(brcond (i32 (setne R32C:$rA, 0)), bb:$dest),
3735 (BRNZr32 R32C:$rA, bb:$dest)>;
3737 multiclass BranchCondEQ<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3739 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3740 (brinst16 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3742 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3743 (brinst16 (CEQHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3745 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3746 (brinst32 (CEQIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3748 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3749 (brinst32 (CEQr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3752 defm BRCONDeq : BranchCondEQ<seteq, BRHNZr16, BRNZr32>;
3753 defm BRCONDne : BranchCondEQ<setne, BRHZr16, BRZr32>;
3755 multiclass BranchCondLGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3757 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3758 (brinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3760 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3761 (brinst16 (CLGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3763 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3764 (brinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3766 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3767 (brinst32 (CLGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3770 defm BRCONDugt : BranchCondLGT<setugt, BRHNZr16, BRNZr32>;
3771 defm BRCONDule : BranchCondLGT<setule, BRHZr16, BRZr32>;
3773 multiclass BranchCondLGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3774 SPUInstr orinst32, SPUInstr brinst32>
3776 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3777 (brinst16 (orinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3778 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3781 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3782 (brinst16 (orinst16 (CLGTHr16 R16C:$rA, R16:$rB),
3783 (CEQHr16 R16C:$rA, R16:$rB)),
3786 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3787 (brinst32 (orinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val),
3788 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3791 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3792 (brinst32 (orinst32 (CLGTr32 R32C:$rA, R32C:$rB),
3793 (CEQr32 R32C:$rA, R32C:$rB)),
3797 defm BRCONDuge : BranchCondLGTEQ<setuge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3798 defm BRCONDult : BranchCondLGTEQ<setult, ORr16, BRHZr16, ORr32, BRZr32>;
3800 multiclass BranchCondGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3802 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3803 (brinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3805 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3806 (brinst16 (CGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3808 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3809 (brinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3811 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3812 (brinst32 (CGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3815 defm BRCONDgt : BranchCondGT<setgt, BRHNZr16, BRNZr32>;
3816 defm BRCONDle : BranchCondGT<setle, BRHZr16, BRZr32>;
3818 multiclass BranchCondGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3819 SPUInstr orinst32, SPUInstr brinst32>
3821 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3822 (brinst16 (orinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3823 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3826 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3827 (brinst16 (orinst16 (CGTHr16 R16C:$rA, R16:$rB),
3828 (CEQHr16 R16C:$rA, R16:$rB)),
3831 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3832 (brinst32 (orinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val),
3833 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3836 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3837 (brinst32 (orinst32 (CGTr32 R32C:$rA, R32C:$rB),
3838 (CEQr32 R32C:$rA, R32C:$rB)),
3842 defm BRCONDge : BranchCondGTEQ<setge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3843 defm BRCONDlt : BranchCondGTEQ<setlt, ORr16, BRHZr16, ORr32, BRZr32>;
3845 let isTerminator = 1, isBarrier = 1 in {
3846 let isReturn = 1 in {
3848 RETForm<"bi\t$$lr", [(retflag)]>;
3852 //===----------------------------------------------------------------------===//
3853 // Single precision floating point instructions
3854 //===----------------------------------------------------------------------===//
3856 class FAInst<dag OOL, dag IOL, list<dag> pattern>:
3857 RRForm<0b01011000100, OOL, IOL, "fa\t$rT, $rA, $rB",
3860 class FAVecInst<ValueType vectype>:
3861 FAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3862 [(set (vectype VECREG:$rT),
3863 (fadd (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3867 def v4f32: FAVecInst<v4f32>;
3868 def f32: FAInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3869 [(set R32FP:$rT, (fadd R32FP:$rA, R32FP:$rB))]>;
3874 class FSInst<dag OOL, dag IOL, list<dag> pattern>:
3875 RRForm<0b01011000100, OOL, IOL, "fs\t$rT, $rA, $rB",
3878 class FSVecInst<ValueType vectype>:
3879 FSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3880 [(set (vectype VECREG:$rT),
3881 (fsub (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3885 def v4f32: FSVecInst<v4f32>;
3886 def f32: FSInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3887 [(set R32FP:$rT, (fsub R32FP:$rA, R32FP:$rB))]>;
3892 // Floating point reciprocal estimate
3894 class FRESTInst<dag OOL, dag IOL>:
3895 RRForm_1<0b00110111000, OOL, IOL,
3896 "frest\t$rT, $rA", SPrecFP,
3897 [/* no pattern */]>;
3900 FRESTInst<(outs VECREG:$rT), (ins VECREG:$rA)>;
3903 FRESTInst<(outs R32FP:$rT), (ins R32FP:$rA)>;
3905 // Floating point interpolate (used in conjunction with reciprocal estimate)
3907 RRForm<0b00101011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3908 "fi\t$rT, $rA, $rB", SPrecFP,
3909 [/* no pattern */]>;
3912 RRForm<0b00101011110, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3913 "fi\t$rT, $rA, $rB", SPrecFP,
3914 [/* no pattern */]>;
3916 //--------------------------------------------------------------------------
3917 // Basic single precision floating point comparisons:
3919 // Note: There is no support on SPU for single precision NaN. Consequently,
3920 // ordered and unordered comparisons are the same.
3921 //--------------------------------------------------------------------------
3924 RRForm<0b01000011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3925 "fceq\t$rT, $rA, $rB", SPrecFP,
3926 [(set R32C:$rT, (setueq R32FP:$rA, R32FP:$rB))]>;
3928 def : Pat<(setoeq R32FP:$rA, R32FP:$rB),
3929 (FCEQf32 R32FP:$rA, R32FP:$rB)>;
3932 RRForm<0b01010011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3933 "fcmeq\t$rT, $rA, $rB", SPrecFP,
3934 [(set R32C:$rT, (setueq (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
3936 def : Pat<(setoeq (fabs R32FP:$rA), (fabs R32FP:$rB)),
3937 (FCMEQf32 R32FP:$rA, R32FP:$rB)>;
3940 RRForm<0b01000011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3941 "fcgt\t$rT, $rA, $rB", SPrecFP,
3942 [(set R32C:$rT, (setugt R32FP:$rA, R32FP:$rB))]>;
3944 def : Pat<(setugt R32FP:$rA, R32FP:$rB),
3945 (FCGTf32 R32FP:$rA, R32FP:$rB)>;
3948 RRForm<0b01010011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3949 "fcmgt\t$rT, $rA, $rB", SPrecFP,
3950 [(set R32C:$rT, (setugt (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
3952 def : Pat<(setugt (fabs R32FP:$rA), (fabs R32FP:$rB)),
3953 (FCMGTf32 R32FP:$rA, R32FP:$rB)>;
3955 //--------------------------------------------------------------------------
3956 // Single precision floating point comparisons and SETCC equivalents:
3957 //--------------------------------------------------------------------------
3959 def : SETCCNegCondReg<setune, R32FP, i32, XORIr32, FCEQf32>;
3960 def : SETCCNegCondReg<setone, R32FP, i32, XORIr32, FCEQf32>;
3962 def : SETCCBinOpReg<setuge, R32FP, ORr32, FCGTf32, FCEQf32>;
3963 def : SETCCBinOpReg<setoge, R32FP, ORr32, FCGTf32, FCEQf32>;
3965 def : SETCCBinOpReg<setult, R32FP, NORr32, FCGTf32, FCEQf32>;
3966 def : SETCCBinOpReg<setolt, R32FP, NORr32, FCGTf32, FCEQf32>;
3968 def : Pat<(setule R32FP:$rA, R32FP:$rB),
3969 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
3970 def : Pat<(setole R32FP:$rA, R32FP:$rB),
3971 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
3973 // FP Status and Control Register Write
3974 // Why isn't rT a don't care in the ISA?
3975 // Should we create a special RRForm_3 for this guy and zero out the rT?
3977 RRForm_1<0b01011101110, (outs R32FP:$rT), (ins R32FP:$rA),
3978 "fscrwr\t$rA", SPrecFP,
3979 [/* This instruction requires an intrinsic. Note: rT is unused. */]>;
3981 // FP Status and Control Register Read
3983 RRForm_2<0b01011101110, (outs R32FP:$rT), (ins),
3984 "fscrrd\t$rT", SPrecFP,
3985 [/* This instruction requires an intrinsic */]>;
3987 // llvm instruction space
3988 // How do these map onto cell instructions?
3990 // frest rC rB # c = 1/b (both lines)
3992 // fm rD rA rC # d = a * 1/b
3993 // fnms rB rD rB rA # b = - (d * b - a) --should == 0 in a perfect world
3994 // fma rB rB rC rD # b = b * c + d
3995 // = -(d *b -a) * c + d
3996 // = a * c - c ( a *b *c - a)
4001 // These llvm instructions will actually map to library calls.
4002 // All that's needed, then, is to check that the appropriate library is
4003 // imported and do a brsl to the proper function name.
4004 // frem # fmod(x, y): x - (x/y) * y
4005 // (Note: fmod(double, double), fmodf(float,float)
4009 // Unimplemented SPU instruction space
4010 // floating reciprocal absolute square root estimate (frsqest)
4012 // The following are probably just intrinsics
4013 // status and control register write
4014 // status and control register read
4016 //--------------------------------------
4017 // Floating point multiply instructions
4018 //--------------------------------------
4021 RRForm<0b00100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4022 "fm\t$rT, $rA, $rB", SPrecFP,
4023 [(set (v4f32 VECREG:$rT), (fmul (v4f32 VECREG:$rA),
4024 (v4f32 VECREG:$rB)))]>;
4027 RRForm<0b01100011010, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
4028 "fm\t$rT, $rA, $rB", SPrecFP,
4029 [(set R32FP:$rT, (fmul R32FP:$rA, R32FP:$rB))]>;
4031 // Floating point multiply and add
4032 // e.g. d = c + (a * b)
4034 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4035 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
4036 [(set (v4f32 VECREG:$rT),
4037 (fadd (v4f32 VECREG:$rC),
4038 (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB))))]>;
4041 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
4042 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
4043 [(set R32FP:$rT, (fadd R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
4045 // FP multiply and subtract
4046 // Subtracts value in rC from product
4049 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4050 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
4051 [(set (v4f32 VECREG:$rT),
4052 (fsub (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB)),
4053 (v4f32 VECREG:$rC)))]>;
4056 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
4057 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
4059 (fsub (fmul R32FP:$rA, R32FP:$rB), R32FP:$rC))]>;
4061 // Floating Negative Mulitply and Subtract
4062 // Subtracts product from value in rC
4063 // res = fneg(fms a b c)
4066 // NOTE: subtraction order
4070 RRRForm<0b1101, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
4071 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
4072 [(set R32FP:$rT, (fsub R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
4075 RRRForm<0b1101, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4076 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
4077 [(set (v4f32 VECREG:$rT),
4078 (fsub (v4f32 VECREG:$rC),
4079 (fmul (v4f32 VECREG:$rA),
4080 (v4f32 VECREG:$rB))))]>;
4082 //--------------------------------------
4083 // Floating Point Conversions
4084 // Signed conversions:
4086 CVTIntFPForm<0b0101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4087 "csflt\t$rT, $rA, 0", SPrecFP,
4088 [(set (v4f32 VECREG:$rT), (sint_to_fp (v4i32 VECREG:$rA)))]>;
4090 // Convert signed integer to floating point
4092 CVTIntFPForm<0b0101101110, (outs R32FP:$rT), (ins R32C:$rA),
4093 "csflt\t$rT, $rA, 0", SPrecFP,
4094 [(set R32FP:$rT, (sint_to_fp R32C:$rA))]>;
4096 // Convert unsigned into to float
4098 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4099 "cuflt\t$rT, $rA, 0", SPrecFP,
4100 [(set (v4f32 VECREG:$rT), (uint_to_fp (v4i32 VECREG:$rA)))]>;
4103 CVTIntFPForm<0b1101101110, (outs R32FP:$rT), (ins R32C:$rA),
4104 "cuflt\t$rT, $rA, 0", SPrecFP,
4105 [(set R32FP:$rT, (uint_to_fp R32C:$rA))]>;
4107 // Convert float to unsigned int
4108 // Assume that scale = 0
4111 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4112 "cfltu\t$rT, $rA, 0", SPrecFP,
4113 [(set (v4i32 VECREG:$rT), (fp_to_uint (v4f32 VECREG:$rA)))]>;
4116 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
4117 "cfltu\t$rT, $rA, 0", SPrecFP,
4118 [(set R32C:$rT, (fp_to_uint R32FP:$rA))]>;
4120 // Convert float to signed int
4121 // Assume that scale = 0
4124 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4125 "cflts\t$rT, $rA, 0", SPrecFP,
4126 [(set (v4i32 VECREG:$rT), (fp_to_sint (v4f32 VECREG:$rA)))]>;
4129 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
4130 "cflts\t$rT, $rA, 0", SPrecFP,
4131 [(set R32C:$rT, (fp_to_sint R32FP:$rA))]>;
4133 //===----------------------------------------------------------------------==//
4134 // Single<->Double precision conversions
4135 //===----------------------------------------------------------------------==//
4137 // NOTE: We use "vec" name suffix here to avoid confusion (e.g. input is a
4138 // v4f32, output is v2f64--which goes in the name?)
4140 // Floating point extend single to double
4141 // NOTE: Not sure if passing in v4f32 to FESDvec is correct since it
4142 // operates on two double-word slots (i.e. 1st and 3rd fp numbers
4145 RRForm_1<0b00011101110, (outs VECREG:$rT), (ins VECREG:$rA),
4146 "fesd\t$rT, $rA", SPrecFP,
4147 [/*(set (v2f64 VECREG:$rT), (fextend (v4f32 VECREG:$rA)))*/]>;
4150 RRForm_1<0b00011101110, (outs R64FP:$rT), (ins R32FP:$rA),
4151 "fesd\t$rT, $rA", SPrecFP,
4152 [(set R64FP:$rT, (fextend R32FP:$rA))]>;
4154 // Floating point round double to single
4156 // RRForm_1<0b10011101110, (outs VECREG:$rT), (ins VECREG:$rA),
4157 // "frds\t$rT, $rA,", SPrecFP,
4158 // [(set (v4f32 R32FP:$rT), (fround (v2f64 R64FP:$rA)))]>;
4161 RRForm_1<0b10011101110, (outs R32FP:$rT), (ins R64FP:$rA),
4162 "frds\t$rT, $rA", SPrecFP,
4163 [(set R32FP:$rT, (fround R64FP:$rA))]>;
4165 //ToDo include anyextend?
4167 //===----------------------------------------------------------------------==//
4168 // Double precision floating point instructions
4169 //===----------------------------------------------------------------------==//
4171 RRForm<0b00110011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4172 "dfa\t$rT, $rA, $rB", DPrecFP,
4173 [(set R64FP:$rT, (fadd R64FP:$rA, R64FP:$rB))]>;
4176 RRForm<0b00110011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4177 "dfa\t$rT, $rA, $rB", DPrecFP,
4178 [(set (v2f64 VECREG:$rT), (fadd (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4181 RRForm<0b10100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4182 "dfs\t$rT, $rA, $rB", DPrecFP,
4183 [(set R64FP:$rT, (fsub R64FP:$rA, R64FP:$rB))]>;
4186 RRForm<0b10100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4187 "dfs\t$rT, $rA, $rB", DPrecFP,
4188 [(set (v2f64 VECREG:$rT),
4189 (fsub (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4192 RRForm<0b01100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4193 "dfm\t$rT, $rA, $rB", DPrecFP,
4194 [(set R64FP:$rT, (fmul R64FP:$rA, R64FP:$rB))]>;
4197 RRForm<0b00100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4198 "dfm\t$rT, $rA, $rB", DPrecFP,
4199 [(set (v2f64 VECREG:$rT),
4200 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4203 RRForm<0b00111010110, (outs R64FP:$rT),
4204 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4205 "dfma\t$rT, $rA, $rB", DPrecFP,
4206 [(set R64FP:$rT, (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB)))]>,
4207 RegConstraint<"$rC = $rT">,
4211 RRForm<0b00111010110, (outs VECREG:$rT),
4212 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4213 "dfma\t$rT, $rA, $rB", DPrecFP,
4214 [(set (v2f64 VECREG:$rT),
4215 (fadd (v2f64 VECREG:$rC),
4216 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB))))]>,
4217 RegConstraint<"$rC = $rT">,
4221 RRForm<0b10111010110, (outs R64FP:$rT),
4222 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4223 "dfms\t$rT, $rA, $rB", DPrecFP,
4224 [(set R64FP:$rT, (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC))]>,
4225 RegConstraint<"$rC = $rT">,
4229 RRForm<0b10111010110, (outs VECREG:$rT),
4230 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4231 "dfms\t$rT, $rA, $rB", DPrecFP,
4232 [(set (v2f64 VECREG:$rT),
4233 (fsub (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)),
4234 (v2f64 VECREG:$rC)))]>;
4236 // DFNMS: - (a * b - c)
4237 // - (a * b) + c => c - (a * b)
4239 class DFNMSInst<dag OOL, dag IOL, list<dag> pattern>:
4240 RRForm<0b01111010110, OOL, IOL, "dfnms\t$rT, $rA, $rB",
4242 RegConstraint<"$rC = $rT">,
4245 class DFNMSVecInst<list<dag> pattern>:
4246 DFNMSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4249 class DFNMSRegInst<list<dag> pattern>:
4250 DFNMSInst<(outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4253 multiclass DFMultiplySubtract
4255 def v2f64 : DFNMSVecInst<[(set (v2f64 VECREG:$rT),
4256 (fsub (v2f64 VECREG:$rC),
4257 (fmul (v2f64 VECREG:$rA),
4258 (v2f64 VECREG:$rB))))]>;
4260 def f64 : DFNMSRegInst<[(set R64FP:$rT,
4262 (fmul R64FP:$rA, R64FP:$rB)))]>;
4265 defm DFNMS : DFMultiplySubtract;
4270 RRForm<0b11111010110, (outs R64FP:$rT),
4271 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4272 "dfnma\t$rT, $rA, $rB", DPrecFP,
4273 [(set R64FP:$rT, (fneg (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB))))]>,
4274 RegConstraint<"$rC = $rT">,
4278 RRForm<0b11111010110, (outs VECREG:$rT),
4279 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4280 "dfnma\t$rT, $rA, $rB", DPrecFP,
4281 [(set (v2f64 VECREG:$rT),
4282 (fneg (fadd (v2f64 VECREG:$rC),
4283 (fmul (v2f64 VECREG:$rA),
4284 (v2f64 VECREG:$rB)))))]>,
4285 RegConstraint<"$rC = $rT">,
4288 //===----------------------------------------------------------------------==//
4289 // Floating point negation and absolute value
4290 //===----------------------------------------------------------------------==//
4292 def : Pat<(fneg (v4f32 VECREG:$rA)),
4293 (XORfnegvec (v4f32 VECREG:$rA),
4294 (v4f32 (ILHUv4i32 0x8000)))>;
4296 def : Pat<(fneg R32FP:$rA),
4297 (XORfneg32 R32FP:$rA, (ILHUr32 0x8000))>;
4299 // Floating point absolute value
4300 // Note: f64 fabs is custom-selected.
4302 def : Pat<(fabs R32FP:$rA),
4303 (ANDfabs32 R32FP:$rA, (IOHLr32 (ILHUr32 0x7fff), 0xffff))>;
4305 def : Pat<(fabs (v4f32 VECREG:$rA)),
4306 (ANDfabsvec (v4f32 VECREG:$rA),
4307 (IOHLv4i32 (ILHUv4i32 0x7fff), 0xffff))>;
4309 //===----------------------------------------------------------------------===//
4310 // Hint for branch instructions:
4311 //===----------------------------------------------------------------------===//
4313 /* def HBR : SPUInstr<(outs), (ins), "hbr\t" */
4315 //===----------------------------------------------------------------------===//
4316 // Execution, Load NOP (execute NOPs belong in even pipeline, load NOPs belong
4317 // in the odd pipeline)
4318 //===----------------------------------------------------------------------===//
4320 def ENOP : SPUInstr<(outs), (ins), "enop", ExecNOP> {
4323 let Inst{0-10} = 0b10000000010;
4324 let Inst{11-17} = 0;
4325 let Inst{18-24} = 0;
4326 let Inst{25-31} = 0;
4329 def LNOP : SPUInstr<(outs), (ins), "lnop", LoadNOP> {
4332 let Inst{0-10} = 0b10000000000;
4333 let Inst{11-17} = 0;
4334 let Inst{18-24} = 0;
4335 let Inst{25-31} = 0;
4338 //===----------------------------------------------------------------------===//
4339 // Bit conversions (type conversions between vector/packed types)
4340 // NOTE: Promotions are handled using the XS* instructions.
4341 //===----------------------------------------------------------------------===//
4342 def : Pat<(v16i8 (bitconvert (v8i16 VECREG:$src))), (v16i8 VECREG:$src)>;
4343 def : Pat<(v16i8 (bitconvert (v4i32 VECREG:$src))), (v16i8 VECREG:$src)>;
4344 def : Pat<(v16i8 (bitconvert (v2i64 VECREG:$src))), (v16i8 VECREG:$src)>;
4345 def : Pat<(v16i8 (bitconvert (v4f32 VECREG:$src))), (v16i8 VECREG:$src)>;
4346 def : Pat<(v16i8 (bitconvert (v2f64 VECREG:$src))), (v16i8 VECREG:$src)>;
4348 def : Pat<(v8i16 (bitconvert (v16i8 VECREG:$src))), (v8i16 VECREG:$src)>;
4349 def : Pat<(v8i16 (bitconvert (v4i32 VECREG:$src))), (v8i16 VECREG:$src)>;
4350 def : Pat<(v8i16 (bitconvert (v2i64 VECREG:$src))), (v8i16 VECREG:$src)>;
4351 def : Pat<(v8i16 (bitconvert (v4f32 VECREG:$src))), (v8i16 VECREG:$src)>;
4352 def : Pat<(v8i16 (bitconvert (v2f64 VECREG:$src))), (v8i16 VECREG:$src)>;
4354 def : Pat<(v4i32 (bitconvert (v16i8 VECREG:$src))), (v4i32 VECREG:$src)>;
4355 def : Pat<(v4i32 (bitconvert (v8i16 VECREG:$src))), (v4i32 VECREG:$src)>;
4356 def : Pat<(v4i32 (bitconvert (v2i64 VECREG:$src))), (v4i32 VECREG:$src)>;
4357 def : Pat<(v4i32 (bitconvert (v4f32 VECREG:$src))), (v4i32 VECREG:$src)>;
4358 def : Pat<(v4i32 (bitconvert (v2f64 VECREG:$src))), (v4i32 VECREG:$src)>;
4360 def : Pat<(v2i64 (bitconvert (v16i8 VECREG:$src))), (v2i64 VECREG:$src)>;
4361 def : Pat<(v2i64 (bitconvert (v8i16 VECREG:$src))), (v2i64 VECREG:$src)>;
4362 def : Pat<(v2i64 (bitconvert (v4i32 VECREG:$src))), (v2i64 VECREG:$src)>;
4363 def : Pat<(v2i64 (bitconvert (v4f32 VECREG:$src))), (v2i64 VECREG:$src)>;
4364 def : Pat<(v2i64 (bitconvert (v2f64 VECREG:$src))), (v2i64 VECREG:$src)>;
4366 def : Pat<(v4f32 (bitconvert (v16i8 VECREG:$src))), (v4f32 VECREG:$src)>;
4367 def : Pat<(v4f32 (bitconvert (v8i16 VECREG:$src))), (v4f32 VECREG:$src)>;
4368 def : Pat<(v4f32 (bitconvert (v2i64 VECREG:$src))), (v4f32 VECREG:$src)>;
4369 def : Pat<(v4f32 (bitconvert (v4i32 VECREG:$src))), (v4f32 VECREG:$src)>;
4370 def : Pat<(v4f32 (bitconvert (v2f64 VECREG:$src))), (v4f32 VECREG:$src)>;
4372 def : Pat<(v2f64 (bitconvert (v16i8 VECREG:$src))), (v2f64 VECREG:$src)>;
4373 def : Pat<(v2f64 (bitconvert (v8i16 VECREG:$src))), (v2f64 VECREG:$src)>;
4374 def : Pat<(v2f64 (bitconvert (v4i32 VECREG:$src))), (v2f64 VECREG:$src)>;
4375 def : Pat<(v2f64 (bitconvert (v2i64 VECREG:$src))), (v2f64 VECREG:$src)>;
4376 def : Pat<(v2f64 (bitconvert (v2f64 VECREG:$src))), (v2f64 VECREG:$src)>;
4378 def : Pat<(i128 (bitconvert (v16i8 VECREG:$src))),
4379 (ORi128_vec VECREG:$src)>;
4380 def : Pat<(i128 (bitconvert (v8i16 VECREG:$src))),
4381 (ORi128_vec VECREG:$src)>;
4382 def : Pat<(i128 (bitconvert (v4i32 VECREG:$src))),
4383 (ORi128_vec VECREG:$src)>;
4384 def : Pat<(i128 (bitconvert (v2i64 VECREG:$src))),
4385 (ORi128_vec VECREG:$src)>;
4386 def : Pat<(i128 (bitconvert (v4f32 VECREG:$src))),
4387 (ORi128_vec VECREG:$src)>;
4388 def : Pat<(i128 (bitconvert (v2f64 VECREG:$src))),
4389 (ORi128_vec VECREG:$src)>;
4391 def : Pat<(v16i8 (bitconvert (i128 GPRC:$src))),
4392 (v16i8 (ORvec_i128 GPRC:$src))>;
4393 def : Pat<(v8i16 (bitconvert (i128 GPRC:$src))),
4394 (v8i16 (ORvec_i128 GPRC:$src))>;
4395 def : Pat<(v4i32 (bitconvert (i128 GPRC:$src))),
4396 (v4i32 (ORvec_i128 GPRC:$src))>;
4397 def : Pat<(v2i64 (bitconvert (i128 GPRC:$src))),
4398 (v2i64 (ORvec_i128 GPRC:$src))>;
4399 def : Pat<(v4f32 (bitconvert (i128 GPRC:$src))),
4400 (v4f32 (ORvec_i128 GPRC:$src))>;
4401 def : Pat<(v2f64 (bitconvert (i128 GPRC:$src))),
4402 (v2f64 (ORvec_i128 GPRC:$src))>;
4404 //===----------------------------------------------------------------------===//
4405 // Instruction patterns:
4406 //===----------------------------------------------------------------------===//
4408 // General 32-bit constants:
4409 def : Pat<(i32 imm:$imm),
4410 (IOHLr32 (ILHUr32 (HI16 imm:$imm)), (LO16 imm:$imm))>;
4412 // Single precision float constants:
4413 def : Pat<(f32 fpimm:$imm),
4414 (IOHLf32 (ILHUf32 (HI16_f32 fpimm:$imm)), (LO16_f32 fpimm:$imm))>;
4416 // General constant 32-bit vectors
4417 def : Pat<(v4i32 v4i32Imm:$imm),
4418 (IOHLv4i32 (v4i32 (ILHUv4i32 (HI16_vec v4i32Imm:$imm))),
4419 (LO16_vec v4i32Imm:$imm))>;
4422 def : Pat<(i8 imm:$imm),
4425 //===----------------------------------------------------------------------===//
4426 // Zero/Any/Sign extensions
4427 //===----------------------------------------------------------------------===//
4429 // sext 8->32: Sign extend bytes to words
4430 def : Pat<(sext_inreg R32C:$rSrc, i8),
4431 (XSHWr32 (XSBHr32 R32C:$rSrc))>;
4433 def : Pat<(i32 (sext R8C:$rSrc)),
4434 (XSHWr16 (XSBHr8 R8C:$rSrc))>;
4436 // sext 8->64: Sign extend bytes to double word
4437 def : Pat<(sext_inreg R64C:$rSrc, i8),
4438 (XSWDr64_inreg (XSHWr64 (XSBHr64 R64C:$rSrc)))>;
4440 def : Pat<(i64 (sext R8C:$rSrc)),
4441 (XSWDr64 (XSHWr16 (XSBHr8 R8C:$rSrc)))>;
4443 // zext 8->16: Zero extend bytes to halfwords
4444 def : Pat<(i16 (zext R8C:$rSrc)),
4445 (ANDHIi8i16 R8C:$rSrc, 0xff)>;
4447 // zext 8->32: Zero extend bytes to words
4448 def : Pat<(i32 (zext R8C:$rSrc)),
4449 (ANDIi8i32 R8C:$rSrc, 0xff)>;
4451 // zext 8->64: Zero extend bytes to double words
4452 def : Pat<(i64 (zext R8C:$rSrc)),
4453 (ORi64_v2i64 (SELBv4i32 (ROTQMBYv4i32
4454 (ORv4i32_i32 (ANDIi8i32 R8C:$rSrc, 0xff)),
4457 (FSMBIv4i32 0x0f0f)))>;
4459 // anyext 8->16: Extend 8->16 bits, irrespective of sign, preserves high bits
4460 def : Pat<(i16 (anyext R8C:$rSrc)),
4461 (ORHIi8i16 R8C:$rSrc, 0)>;
4463 // anyext 8->32: Extend 8->32 bits, irrespective of sign, preserves high bits
4464 def : Pat<(i32 (anyext R8C:$rSrc)),
4465 (ORIi8i32 R8C:$rSrc, 0)>;
4467 // sext 16->64: Sign extend halfword to double word
4468 def : Pat<(sext_inreg R64C:$rSrc, i16),
4469 (XSWDr64_inreg (XSHWr64 R64C:$rSrc))>;
4471 def : Pat<(sext R16C:$rSrc),
4472 (XSWDr64 (XSHWr16 R16C:$rSrc))>;
4474 // zext 16->32: Zero extend halfwords to words
4475 def : Pat<(i32 (zext R16C:$rSrc)),
4476 (ANDi16i32 R16C:$rSrc, (ILAr32 0xffff))>;
4478 def : Pat<(i32 (zext (and R16C:$rSrc, 0xf))),
4479 (ANDIi16i32 R16C:$rSrc, 0xf)>;
4481 def : Pat<(i32 (zext (and R16C:$rSrc, 0xff))),
4482 (ANDIi16i32 R16C:$rSrc, 0xff)>;
4484 def : Pat<(i32 (zext (and R16C:$rSrc, 0xfff))),
4485 (ANDIi16i32 R16C:$rSrc, 0xfff)>;
4487 // anyext 16->32: Extend 16->32 bits, irrespective of sign
4488 def : Pat<(i32 (anyext R16C:$rSrc)),
4489 (ORIi16i32 R16C:$rSrc, 0)>;
4491 //===----------------------------------------------------------------------===//
4493 // These truncates are for the SPU's supported types (i8, i16, i32). i64 and
4494 // above are custom lowered.
4495 //===----------------------------------------------------------------------===//
4497 def : Pat<(i8 (trunc GPRC:$src)),
4499 (SHUFBgprc GPRC:$src, GPRC:$src,
4500 (IOHLv4i32 (ILHUv4i32 0x0f0f), 0x0f0f)))>;
4502 def : Pat<(i8 (trunc R64C:$src)),
4505 (ORv2i64_i64 R64C:$src),
4506 (ORv2i64_i64 R64C:$src),
4507 (IOHLv4i32 (ILHUv4i32 0x0707), 0x0707)))>;
4509 def : Pat<(i8 (trunc R32C:$src)),
4512 (ORv4i32_i32 R32C:$src),
4513 (ORv4i32_i32 R32C:$src),
4514 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)))>;
4516 def : Pat<(i8 (trunc R16C:$src)),
4519 (ORv8i16_i16 R16C:$src),
4520 (ORv8i16_i16 R16C:$src),
4521 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)))>;
4523 def : Pat<(i16 (trunc GPRC:$src)),
4525 (SHUFBgprc GPRC:$src, GPRC:$src,
4526 (IOHLv4i32 (ILHUv4i32 0x0e0f), 0x0e0f)))>;
4528 def : Pat<(i16 (trunc R64C:$src)),
4531 (ORv2i64_i64 R64C:$src),
4532 (ORv2i64_i64 R64C:$src),
4533 (IOHLv4i32 (ILHUv4i32 0x0607), 0x0607)))>;
4535 def : Pat<(i16 (trunc R32C:$src)),
4538 (ORv4i32_i32 R32C:$src),
4539 (ORv4i32_i32 R32C:$src),
4540 (IOHLv4i32 (ILHUv4i32 0x0203), 0x0203)))>;
4542 def : Pat<(i32 (trunc GPRC:$src)),
4544 (SHUFBgprc GPRC:$src, GPRC:$src,
4545 (IOHLv4i32 (ILHUv4i32 0x0c0d), 0x0e0f)))>;
4547 def : Pat<(i32 (trunc R64C:$src)),
4550 (ORv2i64_i64 R64C:$src),
4551 (ORv2i64_i64 R64C:$src),
4552 (IOHLv4i32 (ILHUv4i32 0x0405), 0x0607)))>;
4554 //===----------------------------------------------------------------------===//
4555 // Address generation: SPU, like PPC, has to split addresses into high and
4556 // low parts in order to load them into a register.
4557 //===----------------------------------------------------------------------===//
4559 def : Pat<(SPUaform tglobaladdr:$in, 0), (ILAlsa tglobaladdr:$in)>;
4560 def : Pat<(SPUaform texternalsym:$in, 0), (ILAlsa texternalsym:$in)>;
4561 def : Pat<(SPUaform tjumptable:$in, 0), (ILAlsa tjumptable:$in)>;
4562 def : Pat<(SPUaform tconstpool:$in, 0), (ILAlsa tconstpool:$in)>;
4564 def : Pat<(SPUindirect (SPUhi tglobaladdr:$in, 0),
4565 (SPUlo tglobaladdr:$in, 0)),
4566 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4568 def : Pat<(SPUindirect (SPUhi texternalsym:$in, 0),
4569 (SPUlo texternalsym:$in, 0)),
4570 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4572 def : Pat<(SPUindirect (SPUhi tjumptable:$in, 0),
4573 (SPUlo tjumptable:$in, 0)),
4574 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4576 def : Pat<(SPUindirect (SPUhi tconstpool:$in, 0),
4577 (SPUlo tconstpool:$in, 0)),
4578 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4580 def : Pat<(add (SPUhi tglobaladdr:$in, 0), (SPUlo tglobaladdr:$in, 0)),
4581 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4583 def : Pat<(add (SPUhi texternalsym:$in, 0), (SPUlo texternalsym:$in, 0)),
4584 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4586 def : Pat<(add (SPUhi tjumptable:$in, 0), (SPUlo tjumptable:$in, 0)),
4587 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4589 def : Pat<(add (SPUhi tconstpool:$in, 0), (SPUlo tconstpool:$in, 0)),
4590 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4593 include "CellSDKIntrinsics.td"
4594 // Various math operator instruction sequences
4595 include "SPUMathInstr.td"
4596 // 64-bit "instructions"/support
4597 include "SPU64InstrInfo.td"
4598 // 128-bit "instructions"/support
4599 include "SPU128InstrInfo.td"