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 //===----------------------------------------------------------------------===//
34 // DWARF debugging Pseudo Instructions
35 //===----------------------------------------------------------------------===//
37 def DWARF_LOC : Pseudo<(outs), (ins i32imm:$line, i32imm:$col, i32imm:$file),
38 "${:comment} .loc $file, $line, $col",
39 [(dwarf_loc (i32 imm:$line), (i32 imm:$col),
42 //===----------------------------------------------------------------------===//
44 // NB: The ordering is actually important, since the instruction selection
45 // will try each of the instructions in sequence, i.e., the D-form first with
46 // the 10-bit displacement, then the A-form with the 16 bit displacement, and
47 // finally the X-form with the register-register.
48 //===----------------------------------------------------------------------===//
50 let canFoldAsLoad = 1 in {
51 class LoadDFormVec<ValueType vectype>
52 : RI10Form<0b00101100, (outs VECREG:$rT), (ins dformaddr:$src),
55 [(set (vectype VECREG:$rT), (load dform_addr:$src))]>
58 class LoadDForm<RegisterClass rclass>
59 : RI10Form<0b00101100, (outs rclass:$rT), (ins dformaddr:$src),
62 [(set rclass:$rT, (load dform_addr:$src))]>
67 def v16i8: LoadDFormVec<v16i8>;
68 def v8i16: LoadDFormVec<v8i16>;
69 def v4i32: LoadDFormVec<v4i32>;
70 def v2i64: LoadDFormVec<v2i64>;
71 def v4f32: LoadDFormVec<v4f32>;
72 def v2f64: LoadDFormVec<v2f64>;
74 def r128: LoadDForm<GPRC>;
75 def r64: LoadDForm<R64C>;
76 def r32: LoadDForm<R32C>;
77 def f32: LoadDForm<R32FP>;
78 def f64: LoadDForm<R64FP>;
79 def r16: LoadDForm<R16C>;
80 def r8: LoadDForm<R8C>;
83 class LoadAFormVec<ValueType vectype>
84 : RI16Form<0b100001100, (outs VECREG:$rT), (ins addr256k:$src),
87 [(set (vectype VECREG:$rT), (load aform_addr:$src))]>
90 class LoadAForm<RegisterClass rclass>
91 : RI16Form<0b100001100, (outs rclass:$rT), (ins addr256k:$src),
94 [(set rclass:$rT, (load aform_addr:$src))]>
99 def v16i8: LoadAFormVec<v16i8>;
100 def v8i16: LoadAFormVec<v8i16>;
101 def v4i32: LoadAFormVec<v4i32>;
102 def v2i64: LoadAFormVec<v2i64>;
103 def v4f32: LoadAFormVec<v4f32>;
104 def v2f64: LoadAFormVec<v2f64>;
106 def r128: LoadAForm<GPRC>;
107 def r64: LoadAForm<R64C>;
108 def r32: LoadAForm<R32C>;
109 def f32: LoadAForm<R32FP>;
110 def f64: LoadAForm<R64FP>;
111 def r16: LoadAForm<R16C>;
112 def r8: LoadAForm<R8C>;
115 class LoadXFormVec<ValueType vectype>
116 : RRForm<0b00100011100, (outs VECREG:$rT), (ins memrr:$src),
119 [(set (vectype VECREG:$rT), (load xform_addr:$src))]>
122 class LoadXForm<RegisterClass rclass>
123 : RRForm<0b00100011100, (outs rclass:$rT), (ins memrr:$src),
126 [(set rclass:$rT, (load xform_addr:$src))]>
129 multiclass LoadXForms
131 def v16i8: LoadXFormVec<v16i8>;
132 def v8i16: LoadXFormVec<v8i16>;
133 def v4i32: LoadXFormVec<v4i32>;
134 def v2i64: LoadXFormVec<v2i64>;
135 def v4f32: LoadXFormVec<v4f32>;
136 def v2f64: LoadXFormVec<v2f64>;
138 def r128: LoadXForm<GPRC>;
139 def r64: LoadXForm<R64C>;
140 def r32: LoadXForm<R32C>;
141 def f32: LoadXForm<R32FP>;
142 def f64: LoadXForm<R64FP>;
143 def r16: LoadXForm<R16C>;
144 def r8: LoadXForm<R8C>;
147 defm LQA : LoadAForms;
148 defm LQD : LoadDForms;
149 defm LQX : LoadXForms;
151 /* Load quadword, PC relative: Not much use at this point in time.
152 Might be of use later for relocatable code. It's effectively the
153 same as LQA, but uses PC-relative addressing.
154 def LQR : RI16Form<0b111001100, (outs VECREG:$rT), (ins s16imm:$disp),
155 "lqr\t$rT, $disp", LoadStore,
156 [(set VECREG:$rT, (load iaddr:$disp))]>;
160 //===----------------------------------------------------------------------===//
162 //===----------------------------------------------------------------------===//
163 class StoreDFormVec<ValueType vectype>
164 : RI10Form<0b00100100, (outs), (ins VECREG:$rT, dformaddr:$src),
167 [(store (vectype VECREG:$rT), dform_addr:$src)]>
170 class StoreDForm<RegisterClass rclass>
171 : RI10Form<0b00100100, (outs), (ins rclass:$rT, dformaddr:$src),
174 [(store rclass:$rT, dform_addr:$src)]>
177 multiclass StoreDForms
179 def v16i8: StoreDFormVec<v16i8>;
180 def v8i16: StoreDFormVec<v8i16>;
181 def v4i32: StoreDFormVec<v4i32>;
182 def v2i64: StoreDFormVec<v2i64>;
183 def v4f32: StoreDFormVec<v4f32>;
184 def v2f64: StoreDFormVec<v2f64>;
186 def r128: StoreDForm<GPRC>;
187 def r64: StoreDForm<R64C>;
188 def r32: StoreDForm<R32C>;
189 def f32: StoreDForm<R32FP>;
190 def f64: StoreDForm<R64FP>;
191 def r16: StoreDForm<R16C>;
192 def r8: StoreDForm<R8C>;
195 class StoreAFormVec<ValueType vectype>
196 : RI16Form<0b0010010, (outs), (ins VECREG:$rT, addr256k:$src),
199 [(store (vectype VECREG:$rT), aform_addr:$src)]>;
201 class StoreAForm<RegisterClass rclass>
202 : RI16Form<0b001001, (outs), (ins rclass:$rT, addr256k:$src),
205 [(store rclass:$rT, aform_addr:$src)]>;
207 multiclass StoreAForms
209 def v16i8: StoreAFormVec<v16i8>;
210 def v8i16: StoreAFormVec<v8i16>;
211 def v4i32: StoreAFormVec<v4i32>;
212 def v2i64: StoreAFormVec<v2i64>;
213 def v4f32: StoreAFormVec<v4f32>;
214 def v2f64: StoreAFormVec<v2f64>;
216 def r128: StoreAForm<GPRC>;
217 def r64: StoreAForm<R64C>;
218 def r32: StoreAForm<R32C>;
219 def f32: StoreAForm<R32FP>;
220 def f64: StoreAForm<R64FP>;
221 def r16: StoreAForm<R16C>;
222 def r8: StoreAForm<R8C>;
225 class StoreXFormVec<ValueType vectype>
226 : RRForm<0b00100100, (outs), (ins VECREG:$rT, memrr:$src),
229 [(store (vectype VECREG:$rT), xform_addr:$src)]>
232 class StoreXForm<RegisterClass rclass>
233 : RRForm<0b00100100, (outs), (ins rclass:$rT, memrr:$src),
236 [(store rclass:$rT, xform_addr:$src)]>
239 multiclass StoreXForms
241 def v16i8: StoreXFormVec<v16i8>;
242 def v8i16: StoreXFormVec<v8i16>;
243 def v4i32: StoreXFormVec<v4i32>;
244 def v2i64: StoreXFormVec<v2i64>;
245 def v4f32: StoreXFormVec<v4f32>;
246 def v2f64: StoreXFormVec<v2f64>;
248 def r128: StoreXForm<GPRC>;
249 def r64: StoreXForm<R64C>;
250 def r32: StoreXForm<R32C>;
251 def f32: StoreXForm<R32FP>;
252 def f64: StoreXForm<R64FP>;
253 def r16: StoreXForm<R16C>;
254 def r8: StoreXForm<R8C>;
257 defm STQD : StoreDForms;
258 defm STQA : StoreAForms;
259 defm STQX : StoreXForms;
261 /* Store quadword, PC relative: Not much use at this point in time. Might
262 be useful for relocatable code.
263 def STQR : RI16Form<0b111000100, (outs), (ins VECREG:$rT, s16imm:$disp),
264 "stqr\t$rT, $disp", LoadStore,
265 [(store VECREG:$rT, iaddr:$disp)]>;
268 //===----------------------------------------------------------------------===//
269 // Generate Controls for Insertion:
270 //===----------------------------------------------------------------------===//
272 def CBD: RI7Form<0b10101111100, (outs VECREG:$rT), (ins shufaddr:$src),
273 "cbd\t$rT, $src", ShuffleOp,
274 [(set (v16i8 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
276 def CBX: RRForm<0b00101011100, (outs VECREG:$rT), (ins memrr:$src),
277 "cbx\t$rT, $src", ShuffleOp,
278 [(set (v16i8 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
280 def CHD: RI7Form<0b10101111100, (outs VECREG:$rT), (ins shufaddr:$src),
281 "chd\t$rT, $src", ShuffleOp,
282 [(set (v8i16 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
284 def CHX: RRForm<0b10101011100, (outs VECREG:$rT), (ins memrr:$src),
285 "chx\t$rT, $src", ShuffleOp,
286 [(set (v8i16 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
288 def CWD: RI7Form<0b01101111100, (outs VECREG:$rT), (ins shufaddr:$src),
289 "cwd\t$rT, $src", ShuffleOp,
290 [(set (v4i32 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
292 def CWX: RRForm<0b01101011100, (outs VECREG:$rT), (ins memrr:$src),
293 "cwx\t$rT, $src", ShuffleOp,
294 [(set (v4i32 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
296 def CWDf32: RI7Form<0b01101111100, (outs VECREG:$rT), (ins shufaddr:$src),
297 "cwd\t$rT, $src", ShuffleOp,
298 [(set (v4f32 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
300 def CWXf32: RRForm<0b01101011100, (outs VECREG:$rT), (ins memrr:$src),
301 "cwx\t$rT, $src", ShuffleOp,
302 [(set (v4f32 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
304 def CDD: RI7Form<0b11101111100, (outs VECREG:$rT), (ins shufaddr:$src),
305 "cdd\t$rT, $src", ShuffleOp,
306 [(set (v2i64 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
308 def CDX: RRForm<0b11101011100, (outs VECREG:$rT), (ins memrr:$src),
309 "cdx\t$rT, $src", ShuffleOp,
310 [(set (v2i64 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
312 def CDDf64: RI7Form<0b11101111100, (outs VECREG:$rT), (ins shufaddr:$src),
313 "cdd\t$rT, $src", ShuffleOp,
314 [(set (v2f64 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
316 def CDXf64: RRForm<0b11101011100, (outs VECREG:$rT), (ins memrr:$src),
317 "cdx\t$rT, $src", ShuffleOp,
318 [(set (v2f64 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
320 //===----------------------------------------------------------------------===//
321 // Constant formation:
322 //===----------------------------------------------------------------------===//
325 RI16Form<0b110000010, (outs VECREG:$rT), (ins s16imm:$val),
326 "ilh\t$rT, $val", ImmLoad,
327 [(set (v8i16 VECREG:$rT), (v8i16 v8i16SExt16Imm:$val))]>;
330 RI16Form<0b110000010, (outs R16C:$rT), (ins s16imm:$val),
331 "ilh\t$rT, $val", ImmLoad,
332 [(set R16C:$rT, immSExt16:$val)]>;
334 // Cell SPU doesn't have a native 8-bit immediate load, but ILH works ("with
335 // the right constant")
337 RI16Form<0b110000010, (outs R8C:$rT), (ins s16imm_i8:$val),
338 "ilh\t$rT, $val", ImmLoad,
339 [(set R8C:$rT, immSExt8:$val)]>;
341 // IL does sign extension!
343 class ILInst<dag OOL, dag IOL, list<dag> pattern>:
344 RI16Form<0b100000010, OOL, IOL, "il\t$rT, $val",
347 class ILVecInst<ValueType vectype, Operand immtype, PatLeaf xform>:
348 ILInst<(outs VECREG:$rT), (ins immtype:$val),
349 [(set (vectype VECREG:$rT), (vectype xform:$val))]>;
351 class ILRegInst<RegisterClass rclass, Operand immtype, PatLeaf xform>:
352 ILInst<(outs rclass:$rT), (ins immtype:$val),
353 [(set rclass:$rT, xform:$val)]>;
355 multiclass ImmediateLoad
357 def v2i64: ILVecInst<v2i64, s16imm_i64, v2i64SExt16Imm>;
358 def v4i32: ILVecInst<v4i32, s16imm_i32, v4i32SExt16Imm>;
360 // TODO: Need v2f64, v4f32
362 def r64: ILRegInst<R64C, s16imm_i64, immSExt16>;
363 def r32: ILRegInst<R32C, s16imm_i32, immSExt16>;
364 def f32: ILRegInst<R32FP, s16imm_f32, fpimmSExt16>;
365 def f64: ILRegInst<R64FP, s16imm_f64, fpimmSExt16>;
368 defm IL : ImmediateLoad;
370 class ILHUInst<dag OOL, dag IOL, list<dag> pattern>:
371 RI16Form<0b010000010, OOL, IOL, "ilhu\t$rT, $val",
374 class ILHUVecInst<ValueType vectype, Operand immtype, PatLeaf xform>:
375 ILHUInst<(outs VECREG:$rT), (ins immtype:$val),
376 [(set (vectype VECREG:$rT), (vectype xform:$val))]>;
378 class ILHURegInst<RegisterClass rclass, Operand immtype, PatLeaf xform>:
379 ILHUInst<(outs rclass:$rT), (ins immtype:$val),
380 [(set rclass:$rT, xform:$val)]>;
382 multiclass ImmLoadHalfwordUpper
384 def v2i64: ILHUVecInst<v2i64, u16imm_i64, immILHUvec_i64>;
385 def v4i32: ILHUVecInst<v4i32, u16imm_i32, immILHUvec>;
387 def r64: ILHURegInst<R64C, u16imm_i64, hi16>;
388 def r32: ILHURegInst<R32C, u16imm_i32, hi16>;
390 // Loads the high portion of an address
391 def hi: ILHURegInst<R32C, symbolHi, hi16>;
393 // Used in custom lowering constant SFP loads:
394 def f32: ILHURegInst<R32FP, f16imm, hi16_f32>;
397 defm ILHU : ImmLoadHalfwordUpper;
399 // Immediate load address (can also be used to load 18-bit unsigned constants,
400 // see the zext 16->32 pattern)
402 class ILAInst<dag OOL, dag IOL, list<dag> pattern>:
403 RI18Form<0b1000010, OOL, IOL, "ila\t$rT, $val",
406 class ILAVecInst<ValueType vectype, Operand immtype, PatLeaf xform>:
407 ILAInst<(outs VECREG:$rT), (ins immtype:$val),
408 [(set (vectype VECREG:$rT), (vectype xform:$val))]>;
410 class ILARegInst<RegisterClass rclass, Operand immtype, PatLeaf xform>:
411 ILAInst<(outs rclass:$rT), (ins immtype:$val),
412 [(set rclass:$rT, xform:$val)]>;
414 multiclass ImmLoadAddress
416 def v2i64: ILAVecInst<v2i64, u18imm, v2i64Uns18Imm>;
417 def v4i32: ILAVecInst<v4i32, u18imm, v4i32Uns18Imm>;
419 def r64: ILARegInst<R64C, u18imm_i64, imm18>;
420 def r32: ILARegInst<R32C, u18imm, imm18>;
421 def f32: ILARegInst<R32FP, f18imm, fpimm18>;
422 def f64: ILARegInst<R64FP, f18imm_f64, fpimm18>;
424 def hi: ILARegInst<R32C, symbolHi, imm18>;
425 def lo: ILARegInst<R32C, symbolLo, imm18>;
427 def lsa: ILAInst<(outs R32C:$rT), (ins symbolLSA:$val),
431 defm ILA : ImmLoadAddress;
433 // Immediate OR, Halfword Lower: The "other" part of loading large constants
434 // into 32-bit registers. See the anonymous pattern Pat<(i32 imm:$imm), ...>
435 // Note that these are really two operand instructions, but they're encoded
436 // as three operands with the first two arguments tied-to each other.
438 class IOHLInst<dag OOL, dag IOL, list<dag> pattern>:
439 RI16Form<0b100000110, OOL, IOL, "iohl\t$rT, $val",
441 RegConstraint<"$rS = $rT">,
444 class IOHLVecInst<ValueType vectype, Operand immtype /* , PatLeaf xform */>:
445 IOHLInst<(outs VECREG:$rT), (ins VECREG:$rS, immtype:$val),
448 class IOHLRegInst<RegisterClass rclass, Operand immtype /* , PatLeaf xform */>:
449 IOHLInst<(outs rclass:$rT), (ins rclass:$rS, immtype:$val),
452 multiclass ImmOrHalfwordLower
454 def v2i64: IOHLVecInst<v2i64, u16imm_i64>;
455 def v4i32: IOHLVecInst<v4i32, u16imm_i32>;
457 def r32: IOHLRegInst<R32C, i32imm>;
458 def f32: IOHLRegInst<R32FP, f32imm>;
460 def lo: IOHLRegInst<R32C, symbolLo>;
463 defm IOHL: ImmOrHalfwordLower;
465 // Form select mask for bytes using immediate, used in conjunction with the
468 class FSMBIVec<ValueType vectype>:
469 RI16Form<0b101001100, (outs VECREG:$rT), (ins u16imm:$val),
472 [(set (vectype VECREG:$rT), (SPUselmask (i16 immU16:$val)))]>;
474 multiclass FormSelectMaskBytesImm
476 def v16i8: FSMBIVec<v16i8>;
477 def v8i16: FSMBIVec<v8i16>;
478 def v4i32: FSMBIVec<v4i32>;
479 def v2i64: FSMBIVec<v2i64>;
482 defm FSMBI : FormSelectMaskBytesImm;
484 // fsmb: Form select mask for bytes. N.B. Input operand, $rA, is 16-bits
485 class FSMBInst<dag OOL, dag IOL, list<dag> pattern>:
486 RRForm_1<0b01101101100, OOL, IOL, "fsmb\t$rT, $rA", SelectOp,
489 class FSMBRegInst<RegisterClass rclass, ValueType vectype>:
490 FSMBInst<(outs VECREG:$rT), (ins rclass:$rA),
491 [(set (vectype VECREG:$rT), (SPUselmask rclass:$rA))]>;
493 class FSMBVecInst<ValueType vectype>:
494 FSMBInst<(outs VECREG:$rT), (ins VECREG:$rA),
495 [(set (vectype VECREG:$rT),
496 (SPUselmask (vectype VECREG:$rA)))]>;
498 multiclass FormSelectMaskBits {
499 def v16i8_r16: FSMBRegInst<R16C, v16i8>;
500 def v16i8: FSMBVecInst<v16i8>;
503 defm FSMB: FormSelectMaskBits;
505 // fsmh: Form select mask for halfwords. N.B., Input operand, $rA, is
506 // only 8-bits wide (even though it's input as 16-bits here)
508 class FSMHInst<dag OOL, dag IOL, list<dag> pattern>:
509 RRForm_1<0b10101101100, OOL, IOL, "fsmh\t$rT, $rA", SelectOp,
512 class FSMHRegInst<RegisterClass rclass, ValueType vectype>:
513 FSMHInst<(outs VECREG:$rT), (ins rclass:$rA),
514 [(set (vectype VECREG:$rT), (SPUselmask rclass:$rA))]>;
516 class FSMHVecInst<ValueType vectype>:
517 FSMHInst<(outs VECREG:$rT), (ins VECREG:$rA),
518 [(set (vectype VECREG:$rT),
519 (SPUselmask (vectype VECREG:$rA)))]>;
521 multiclass FormSelectMaskHalfword {
522 def v8i16_r16: FSMHRegInst<R16C, v8i16>;
523 def v8i16: FSMHVecInst<v8i16>;
526 defm FSMH: FormSelectMaskHalfword;
528 // fsm: Form select mask for words. Like the other fsm* instructions,
529 // only the lower 4 bits of $rA are significant.
531 class FSMInst<dag OOL, dag IOL, list<dag> pattern>:
532 RRForm_1<0b00101101100, OOL, IOL, "fsm\t$rT, $rA", SelectOp,
535 class FSMRegInst<ValueType vectype, RegisterClass rclass>:
536 FSMInst<(outs VECREG:$rT), (ins rclass:$rA),
537 [(set (vectype VECREG:$rT), (SPUselmask rclass:$rA))]>;
539 class FSMVecInst<ValueType vectype>:
540 FSMInst<(outs VECREG:$rT), (ins VECREG:$rA),
541 [(set (vectype VECREG:$rT), (SPUselmask (vectype VECREG:$rA)))]>;
543 multiclass FormSelectMaskWord {
544 def v4i32: FSMVecInst<v4i32>;
546 def r32 : FSMRegInst<v4i32, R32C>;
547 def r16 : FSMRegInst<v4i32, R16C>;
550 defm FSM : FormSelectMaskWord;
552 // Special case when used for i64 math operations
553 multiclass FormSelectMaskWord64 {
554 def r32 : FSMRegInst<v2i64, R32C>;
555 def r16 : FSMRegInst<v2i64, R16C>;
558 defm FSM64 : FormSelectMaskWord64;
560 //===----------------------------------------------------------------------===//
561 // Integer and Logical Operations:
562 //===----------------------------------------------------------------------===//
565 RRForm<0b00010011000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
566 "ah\t$rT, $rA, $rB", IntegerOp,
567 [(set (v8i16 VECREG:$rT), (int_spu_si_ah VECREG:$rA, VECREG:$rB))]>;
569 def : Pat<(add (v8i16 VECREG:$rA), (v8i16 VECREG:$rB)),
570 (AHv8i16 VECREG:$rA, VECREG:$rB)>;
573 RRForm<0b00010011000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
574 "ah\t$rT, $rA, $rB", IntegerOp,
575 [(set R16C:$rT, (add R16C:$rA, R16C:$rB))]>;
578 RI10Form<0b10111000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
579 "ahi\t$rT, $rA, $val", IntegerOp,
580 [(set (v8i16 VECREG:$rT), (add (v8i16 VECREG:$rA),
581 v8i16SExt10Imm:$val))]>;
584 RI10Form<0b10111000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
585 "ahi\t$rT, $rA, $val", IntegerOp,
586 [(set R16C:$rT, (add R16C:$rA, v8i16SExt10Imm:$val))]>;
589 RRForm<0b00000011000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
590 "a\t$rT, $rA, $rB", IntegerOp,
591 [(set (v4i32 VECREG:$rT), (add (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
593 def : Pat<(add (v16i8 VECREG:$rA), (v16i8 VECREG:$rB)),
594 (Avec VECREG:$rA, VECREG:$rB)>;
597 RRForm<0b00000011000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
598 "a\t$rT, $rA, $rB", IntegerOp,
599 [(set R32C:$rT, (add R32C:$rA, R32C:$rB))]>;
602 RRForm<0b00000011000, (outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
603 "a\t$rT, $rA, $rB", IntegerOp,
607 RI10Form<0b00111000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
608 "ai\t$rT, $rA, $val", IntegerOp,
609 [(set (v4i32 VECREG:$rT), (add (v4i32 VECREG:$rA),
610 v4i32SExt10Imm:$val))]>;
613 RI10Form<0b00111000, (outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
614 "ai\t$rT, $rA, $val", IntegerOp,
615 [(set R32C:$rT, (add R32C:$rA, i32ImmSExt10:$val))]>;
618 RRForm<0b00010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
619 "sfh\t$rT, $rA, $rB", IntegerOp,
620 [(set (v8i16 VECREG:$rT), (sub (v8i16 VECREG:$rA),
621 (v8i16 VECREG:$rB)))]>;
624 RRForm<0b00010010000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
625 "sfh\t$rT, $rA, $rB", IntegerOp,
626 [(set R16C:$rT, (sub R16C:$rA, R16C:$rB))]>;
629 RI10Form<0b10110000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
630 "sfhi\t$rT, $rA, $val", IntegerOp,
631 [(set (v8i16 VECREG:$rT), (sub v8i16SExt10Imm:$val,
632 (v8i16 VECREG:$rA)))]>;
634 def SFHIr16 : RI10Form<0b10110000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
635 "sfhi\t$rT, $rA, $val", IntegerOp,
636 [(set R16C:$rT, (sub i16ImmSExt10:$val, R16C:$rA))]>;
638 def SFvec : RRForm<0b00000010000, (outs VECREG:$rT),
639 (ins VECREG:$rA, VECREG:$rB),
640 "sf\t$rT, $rA, $rB", IntegerOp,
641 [(set (v4i32 VECREG:$rT), (sub (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
643 def SFr32 : RRForm<0b00000010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
644 "sf\t$rT, $rA, $rB", IntegerOp,
645 [(set R32C:$rT, (sub R32C:$rA, R32C:$rB))]>;
648 RI10Form<0b00110000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
649 "sfi\t$rT, $rA, $val", IntegerOp,
650 [(set (v4i32 VECREG:$rT), (sub v4i32SExt10Imm:$val,
651 (v4i32 VECREG:$rA)))]>;
653 def SFIr32 : RI10Form<0b00110000, (outs R32C:$rT),
654 (ins R32C:$rA, s10imm_i32:$val),
655 "sfi\t$rT, $rA, $val", IntegerOp,
656 [(set R32C:$rT, (sub i32ImmSExt10:$val, R32C:$rA))]>;
658 // ADDX: only available in vector form, doesn't match a pattern.
659 class ADDXInst<dag OOL, dag IOL, list<dag> pattern>:
660 RRForm<0b00000010110, OOL, IOL,
661 "addx\t$rT, $rA, $rB",
664 class ADDXVecInst<ValueType vectype>:
665 ADDXInst<(outs VECREG:$rT),
666 (ins VECREG:$rA, VECREG:$rB, VECREG:$rCarry),
667 [(set (vectype VECREG:$rT),
668 (SPUaddx (vectype VECREG:$rA), (vectype VECREG:$rB),
669 (vectype VECREG:$rCarry)))]>,
670 RegConstraint<"$rCarry = $rT">,
673 class ADDXRegInst<RegisterClass rclass>:
674 ADDXInst<(outs rclass:$rT),
675 (ins rclass:$rA, rclass:$rB, rclass:$rCarry),
677 (SPUaddx rclass:$rA, rclass:$rB, rclass:$rCarry))]>,
678 RegConstraint<"$rCarry = $rT">,
681 multiclass AddExtended {
682 def v2i64 : ADDXVecInst<v2i64>;
683 def v4i32 : ADDXVecInst<v4i32>;
684 def r64 : ADDXRegInst<R64C>;
685 def r32 : ADDXRegInst<R32C>;
688 defm ADDX : AddExtended;
690 // CG: Generate carry for add
691 class CGInst<dag OOL, dag IOL, list<dag> pattern>:
692 RRForm<0b01000011000, OOL, IOL,
696 class CGVecInst<ValueType vectype>:
697 CGInst<(outs VECREG:$rT),
698 (ins VECREG:$rA, VECREG:$rB),
699 [(set (vectype VECREG:$rT),
700 (SPUcarry_gen (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
702 class CGRegInst<RegisterClass rclass>:
703 CGInst<(outs rclass:$rT),
704 (ins rclass:$rA, rclass:$rB),
706 (SPUcarry_gen rclass:$rA, rclass:$rB))]>;
708 multiclass CarryGenerate {
709 def v2i64 : CGVecInst<v2i64>;
710 def v4i32 : CGVecInst<v4i32>;
711 def r64 : CGRegInst<R64C>;
712 def r32 : CGRegInst<R32C>;
715 defm CG : CarryGenerate;
717 // SFX: Subract from, extended. This is used in conjunction with BG to subtract
718 // with carry (borrow, in this case)
719 class SFXInst<dag OOL, dag IOL, list<dag> pattern>:
720 RRForm<0b10000010110, OOL, IOL,
721 "sfx\t$rT, $rA, $rB",
724 class SFXVecInst<ValueType vectype>:
725 SFXInst<(outs VECREG:$rT),
726 (ins VECREG:$rA, VECREG:$rB, VECREG:$rCarry),
727 [(set (vectype VECREG:$rT),
728 (SPUsubx (vectype VECREG:$rA), (vectype VECREG:$rB),
729 (vectype VECREG:$rCarry)))]>,
730 RegConstraint<"$rCarry = $rT">,
733 class SFXRegInst<RegisterClass rclass>:
734 SFXInst<(outs rclass:$rT),
735 (ins rclass:$rA, rclass:$rB, rclass:$rCarry),
737 (SPUsubx rclass:$rA, rclass:$rB, rclass:$rCarry))]>,
738 RegConstraint<"$rCarry = $rT">,
741 multiclass SubtractExtended {
742 def v2i64 : SFXVecInst<v2i64>;
743 def v4i32 : SFXVecInst<v4i32>;
744 def r64 : SFXRegInst<R64C>;
745 def r32 : SFXRegInst<R32C>;
748 defm SFX : SubtractExtended;
750 // BG: only available in vector form, doesn't match a pattern.
751 class BGInst<dag OOL, dag IOL, list<dag> pattern>:
752 RRForm<0b01000010000, OOL, IOL,
756 class BGVecInst<ValueType vectype>:
757 BGInst<(outs VECREG:$rT),
758 (ins VECREG:$rA, VECREG:$rB),
759 [(set (vectype VECREG:$rT),
760 (SPUborrow_gen (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
762 class BGRegInst<RegisterClass rclass>:
763 BGInst<(outs rclass:$rT),
764 (ins rclass:$rA, rclass:$rB),
766 (SPUborrow_gen rclass:$rA, rclass:$rB))]>;
768 multiclass BorrowGenerate {
769 def v4i32 : BGVecInst<v4i32>;
770 def v2i64 : BGVecInst<v2i64>;
771 def r64 : BGRegInst<R64C>;
772 def r32 : BGRegInst<R32C>;
775 defm BG : BorrowGenerate;
777 // BGX: Borrow generate, extended.
779 RRForm<0b11000010110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB,
781 "bgx\t$rT, $rA, $rB", IntegerOp,
783 RegConstraint<"$rCarry = $rT">,
786 // Halfword multiply variants:
787 // N.B: These can be used to build up larger quantities (16x16 -> 32)
790 RRForm<0b00100011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
791 "mpy\t$rT, $rA, $rB", IntegerMulDiv,
792 [(set (v8i16 VECREG:$rT), (SPUmpy_v8i16 (v8i16 VECREG:$rA),
793 (v8i16 VECREG:$rB)))]>;
796 RRForm<0b00100011110, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
797 "mpy\t$rT, $rA, $rB", IntegerMulDiv,
798 [(set R16C:$rT, (mul R16C:$rA, R16C:$rB))]>;
801 RRForm<0b00110011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
802 "mpyu\t$rT, $rA, $rB", IntegerMulDiv,
803 [(set (v4i32 VECREG:$rT),
804 (SPUmpyu_v4i32 (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
807 RRForm<0b00110011110, (outs R32C:$rT), (ins R16C:$rA, R16C:$rB),
808 "mpyu\t$rT, $rA, $rB", IntegerMulDiv,
809 [(set R32C:$rT, (mul (zext R16C:$rA),
813 RRForm<0b00110011110, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
814 "mpyu\t$rT, $rA, $rB", IntegerMulDiv,
815 [(set R32C:$rT, (SPUmpyu_i32 R32C:$rA, R32C:$rB))]>;
817 // mpyi: multiply 16 x s10imm -> 32 result (custom lowering for 32 bit result,
818 // this only produces the lower 16 bits)
820 RI10Form<0b00101110, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
821 "mpyi\t$rT, $rA, $val", IntegerMulDiv,
822 [(set (v8i16 VECREG:$rT), (mul (v8i16 VECREG:$rA), v8i16SExt10Imm:$val))]>;
825 RI10Form<0b00101110, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
826 "mpyi\t$rT, $rA, $val", IntegerMulDiv,
827 [(set R16C:$rT, (mul R16C:$rA, i16ImmSExt10:$val))]>;
829 // mpyui: same issues as other multiplies, plus, this doesn't match a
830 // pattern... but may be used during target DAG selection or lowering
832 RI10Form<0b10101110, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
833 "mpyui\t$rT, $rA, $val", IntegerMulDiv,
837 RI10Form<0b10101110, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
838 "mpyui\t$rT, $rA, $val", IntegerMulDiv,
841 // mpya: 16 x 16 + 16 -> 32 bit result
843 RRRForm<0b0011, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
844 "mpya\t$rT, $rA, $rB, $rC", IntegerMulDiv,
845 [(set (v4i32 VECREG:$rT), (add (v4i32 (bitconvert (mul (v8i16 VECREG:$rA),
846 (v8i16 VECREG:$rB)))),
847 (v4i32 VECREG:$rC)))]>;
850 RRRForm<0b0011, (outs R32C:$rT), (ins R16C:$rA, R16C:$rB, R32C:$rC),
851 "mpya\t$rT, $rA, $rB, $rC", IntegerMulDiv,
852 [(set R32C:$rT, (add (sext (mul R16C:$rA, R16C:$rB)),
855 def : Pat<(add (mul (sext R16C:$rA), (sext R16C:$rB)), R32C:$rC),
856 (MPYAr32 R16C:$rA, R16C:$rB, R32C:$rC)>;
858 def MPYAr32_sextinreg:
859 RRRForm<0b0011, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB, R32C:$rC),
860 "mpya\t$rT, $rA, $rB, $rC", IntegerMulDiv,
861 [(set R32C:$rT, (add (mul (sext_inreg R32C:$rA, i16),
862 (sext_inreg R32C:$rB, i16)),
866 // RRRForm<0b0011, (outs R32C:$rT), (ins R16C:$rA, R16C:$rB, R32C:$rC),
867 // "mpya\t$rT, $rA, $rB, $rC", IntegerMulDiv,
868 // [(set R32C:$rT, (add (sext (mul R16C:$rA, R16C:$rB)),
871 // mpyh: multiply high, used to synthesize 32-bit multiplies
873 RRForm<0b10100011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
874 "mpyh\t$rT, $rA, $rB", IntegerMulDiv,
875 [(set (v4i32 VECREG:$rT),
876 (SPUmpyh_v4i32 (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
879 RRForm<0b10100011110, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
880 "mpyh\t$rT, $rA, $rB", IntegerMulDiv,
881 [(set R32C:$rT, (SPUmpyh_i32 R32C:$rA, R32C:$rB))]>;
883 // mpys: multiply high and shift right (returns the top half of
884 // a 16-bit multiply, sign extended to 32 bits.)
886 RRForm<0b11100011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
887 "mpys\t$rT, $rA, $rB", IntegerMulDiv,
891 RRForm<0b11100011110, (outs R32C:$rT), (ins R16C:$rA, R16C:$rB),
892 "mpys\t$rT, $rA, $rB", IntegerMulDiv,
895 // mpyhh: multiply high-high (returns the 32-bit result from multiplying
896 // the top 16 bits of the $rA, $rB)
898 RRForm<0b01100011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
899 "mpyhh\t$rT, $rA, $rB", IntegerMulDiv,
900 [(set (v8i16 VECREG:$rT),
901 (SPUmpyhh_v8i16 (v8i16 VECREG:$rA), (v8i16 VECREG:$rB)))]>;
904 RRForm<0b01100011110, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
905 "mpyhh\t$rT, $rA, $rB", IntegerMulDiv,
908 // mpyhha: Multiply high-high, add to $rT:
910 RRForm<0b01100010110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
911 "mpyhha\t$rT, $rA, $rB", IntegerMulDiv,
915 RRForm<0b01100010110, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
916 "mpyhha\t$rT, $rA, $rB", IntegerMulDiv,
919 // mpyhhu: Multiply high-high, unsigned
921 RRForm<0b01110011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
922 "mpyhhu\t$rT, $rA, $rB", IntegerMulDiv,
926 RRForm<0b01110011110, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
927 "mpyhhu\t$rT, $rA, $rB", IntegerMulDiv,
930 // mpyhhau: Multiply high-high, unsigned
932 RRForm<0b01110010110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
933 "mpyhhau\t$rT, $rA, $rB", IntegerMulDiv,
937 RRForm<0b01110010110, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
938 "mpyhhau\t$rT, $rA, $rB", IntegerMulDiv,
941 // clz: Count leading zeroes
942 class CLZInst<dag OOL, dag IOL, list<dag> pattern>:
943 RRForm_1<0b10100101010, OOL, IOL, "clz\t$rT, $rA",
946 class CLZRegInst<RegisterClass rclass>:
947 CLZInst<(outs rclass:$rT), (ins rclass:$rA),
948 [(set rclass:$rT, (ctlz rclass:$rA))]>;
950 class CLZVecInst<ValueType vectype>:
951 CLZInst<(outs VECREG:$rT), (ins VECREG:$rA),
952 [(set (vectype VECREG:$rT), (ctlz (vectype VECREG:$rA)))]>;
954 multiclass CountLeadingZeroes {
955 def v4i32 : CLZVecInst<v4i32>;
956 def r32 : CLZRegInst<R32C>;
959 defm CLZ : CountLeadingZeroes;
961 // cntb: Count ones in bytes (aka "population count")
963 // NOTE: This instruction is really a vector instruction, but the custom
964 // lowering code uses it in unorthodox ways to support CTPOP for other
968 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
969 "cntb\t$rT, $rA", IntegerOp,
970 [(set (v16i8 VECREG:$rT), (SPUcntb (v16i8 VECREG:$rA)))]>;
973 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
974 "cntb\t$rT, $rA", IntegerOp,
975 [(set (v8i16 VECREG:$rT), (SPUcntb (v8i16 VECREG:$rA)))]>;
978 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
979 "cntb\t$rT, $rA", IntegerOp,
980 [(set (v4i32 VECREG:$rT), (SPUcntb (v4i32 VECREG:$rA)))]>;
982 // gbb: Gather the low order bits from each byte in $rA into a single 16-bit
983 // quantity stored into $rT's slot 0, upper 16 bits are zeroed, as are
986 // Note: This instruction "pairs" with the fsmb instruction for all of the
987 // various types defined here.
989 // Note 2: The "VecInst" and "RegInst" forms refer to the result being either
990 // a vector or register.
992 class GBBInst<dag OOL, dag IOL, list<dag> pattern>:
993 RRForm_1<0b01001101100, OOL, IOL, "gbb\t$rT, $rA", GatherOp, pattern>;
995 class GBBRegInst<RegisterClass rclass, ValueType vectype>:
996 GBBInst<(outs rclass:$rT), (ins VECREG:$rA),
997 [(set rclass:$rT, (SPUgatherbits (vectype VECREG:$rA)))]>;
999 class GBBVecInst<ValueType vectype>:
1000 GBBInst<(outs VECREG:$rT), (ins VECREG:$rA),
1001 [(set (vectype VECREG:$rT), (SPUgatherbits (vectype VECREG:$rA)))]>;
1003 multiclass GatherBitsFromBytes {
1004 def v16i8_r32: GBBRegInst<R32C, v16i8>;
1005 def v16i8_r16: GBBRegInst<R16C, v16i8>;
1006 def v16i8: GBBVecInst<v16i8>;
1009 defm GBB: GatherBitsFromBytes;
1011 // gbh: Gather all low order bits from each halfword in $rA into a single
1012 // 8-bit quantity stored in $rT's slot 0, with the upper bits of $rT set to 0
1013 // and slots 1-3 also set to 0.
1015 // See notes for GBBInst, above.
1017 class GBHInst<dag OOL, dag IOL, list<dag> pattern>:
1018 RRForm_1<0b10001101100, OOL, IOL, "gbh\t$rT, $rA", GatherOp,
1021 class GBHRegInst<RegisterClass rclass, ValueType vectype>:
1022 GBHInst<(outs rclass:$rT), (ins VECREG:$rA),
1023 [(set rclass:$rT, (SPUgatherbits (vectype VECREG:$rA)))]>;
1025 class GBHVecInst<ValueType vectype>:
1026 GBHInst<(outs VECREG:$rT), (ins VECREG:$rA),
1027 [(set (vectype VECREG:$rT),
1028 (SPUgatherbits (vectype VECREG:$rA)))]>;
1030 multiclass GatherBitsHalfword {
1031 def v8i16_r32: GBHRegInst<R32C, v8i16>;
1032 def v8i16_r16: GBHRegInst<R16C, v8i16>;
1033 def v8i16: GBHVecInst<v8i16>;
1036 defm GBH: GatherBitsHalfword;
1038 // gb: Gather all low order bits from each word in $rA into a single
1039 // 4-bit quantity stored in $rT's slot 0, upper bits in $rT set to 0,
1040 // as well as slots 1-3.
1042 // See notes for gbb, above.
1044 class GBInst<dag OOL, dag IOL, list<dag> pattern>:
1045 RRForm_1<0b00001101100, OOL, IOL, "gb\t$rT, $rA", GatherOp,
1048 class GBRegInst<RegisterClass rclass, ValueType vectype>:
1049 GBInst<(outs rclass:$rT), (ins VECREG:$rA),
1050 [(set rclass:$rT, (SPUgatherbits (vectype VECREG:$rA)))]>;
1052 class GBVecInst<ValueType vectype>:
1053 GBInst<(outs VECREG:$rT), (ins VECREG:$rA),
1054 [(set (vectype VECREG:$rT),
1055 (SPUgatherbits (vectype VECREG:$rA)))]>;
1057 multiclass GatherBitsWord {
1058 def v4i32_r32: GBRegInst<R32C, v4i32>;
1059 def v4i32_r16: GBRegInst<R16C, v4i32>;
1060 def v4i32: GBVecInst<v4i32>;
1063 defm GB: GatherBitsWord;
1065 // avgb: average bytes
1067 RRForm<0b11001011000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1068 "avgb\t$rT, $rA, $rB", ByteOp,
1071 // absdb: absolute difference of bytes
1073 RRForm<0b11001010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1074 "absdb\t$rT, $rA, $rB", ByteOp,
1077 // sumb: sum bytes into halfwords
1079 RRForm<0b11001010010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1080 "sumb\t$rT, $rA, $rB", ByteOp,
1083 // Sign extension operations:
1084 class XSBHInst<dag OOL, dag IOL, list<dag> pattern>:
1085 RRForm_1<0b01101101010, OOL, IOL,
1086 "xsbh\t$rDst, $rSrc",
1087 IntegerOp, pattern>;
1089 class XSBHVecInst<ValueType vectype>:
1090 XSBHInst<(outs VECREG:$rDst), (ins VECREG:$rSrc),
1091 [(set (v8i16 VECREG:$rDst), (sext (vectype VECREG:$rSrc)))]>;
1093 class XSBHInRegInst<RegisterClass rclass>:
1094 XSBHInst<(outs rclass:$rDst), (ins rclass:$rSrc),
1095 [(set rclass:$rDst, (sext_inreg rclass:$rSrc, i8))]>;
1097 multiclass ExtendByteHalfword {
1098 def v16i8: XSBHVecInst<v8i16>;
1099 def r16: XSBHInRegInst<R16C>;
1100 def r8: XSBHInst<(outs R16C:$rDst), (ins R8C:$rSrc),
1101 [(set R16C:$rDst, (sext R8C:$rSrc))]>;
1103 // 32-bit form for XSBH: used to sign extend 8-bit quantities to 16-bit
1104 // quantities to 32-bit quantities via a 32-bit register (see the sext 8->32
1105 // pattern below). Intentionally doesn't match a pattern because we want the
1106 // sext 8->32 pattern to do the work for us, namely because we need the extra
1108 def r32: XSBHInRegInst<R32C>;
1111 defm XSBH : ExtendByteHalfword;
1113 // Sign extend halfwords to words:
1115 RRForm_1<0b01101101010, (outs VECREG:$rDest), (ins VECREG:$rSrc),
1116 "xshw\t$rDest, $rSrc", IntegerOp,
1117 [(set (v4i32 VECREG:$rDest), (sext (v8i16 VECREG:$rSrc)))]>;
1120 RRForm_1<0b01101101010, (outs R32C:$rDst), (ins R32C:$rSrc),
1121 "xshw\t$rDst, $rSrc", IntegerOp,
1122 [(set R32C:$rDst, (sext_inreg R32C:$rSrc, i16))]>;
1125 RRForm_1<0b01101101010, (outs R32C:$rDst), (ins R16C:$rSrc),
1126 "xshw\t$rDst, $rSrc", IntegerOp,
1127 [(set R32C:$rDst, (sext R16C:$rSrc))]>;
1130 RRForm_1<0b01100101010, (outs VECREG:$rDst), (ins VECREG:$rSrc),
1131 "xswd\t$rDst, $rSrc", IntegerOp,
1132 [(set (v2i64 VECREG:$rDst), (sext (v4i32 VECREG:$rSrc)))]>;
1135 RRForm_1<0b01100101010, (outs R64C:$rDst), (ins R64C:$rSrc),
1136 "xswd\t$rDst, $rSrc", IntegerOp,
1137 [(set R64C:$rDst, (sext_inreg R64C:$rSrc, i32))]>;
1140 RRForm_1<0b01100101010, (outs R64C:$rDst), (ins R32C:$rSrc),
1141 "xswd\t$rDst, $rSrc", IntegerOp,
1142 [(set R64C:$rDst, (SPUsext32_to_64 R32C:$rSrc))]>;
1144 def : Pat<(sext R32C:$inp),
1145 (XSWDr32 R32C:$inp)>;
1149 class ANDInst<dag OOL, dag IOL, list<dag> pattern> :
1150 RRForm<0b10000011000, OOL, IOL, "and\t$rT, $rA, $rB",
1151 IntegerOp, pattern>;
1153 class ANDVecInst<ValueType vectype>:
1154 ANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1155 [(set (vectype VECREG:$rT), (and (vectype VECREG:$rA),
1156 (vectype VECREG:$rB)))]>;
1158 class ANDRegInst<RegisterClass rclass>:
1159 ANDInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1160 [(set rclass:$rT, (and rclass:$rA, rclass:$rB))]>;
1162 multiclass BitwiseAnd
1164 def v16i8: ANDVecInst<v16i8>;
1165 def v8i16: ANDVecInst<v8i16>;
1166 def v4i32: ANDVecInst<v4i32>;
1167 def v2i64: ANDVecInst<v2i64>;
1169 def r128: ANDRegInst<GPRC>;
1170 def r64: ANDRegInst<R64C>;
1171 def r32: ANDRegInst<R32C>;
1172 def r16: ANDRegInst<R16C>;
1173 def r8: ANDRegInst<R8C>;
1175 //===---------------------------------------------
1176 // Special instructions to perform the fabs instruction
1177 def fabs32: ANDInst<(outs R32FP:$rT), (ins R32FP:$rA, R32C:$rB),
1178 [/* Intentionally does not match a pattern */]>;
1180 def fabs64: ANDInst<(outs R64FP:$rT), (ins R64FP:$rA, VECREG:$rB),
1181 [/* Intentionally does not match a pattern */]>;
1183 // Could use v4i32, but won't for clarity
1184 def fabsvec: ANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1185 [/* Intentionally does not match a pattern */]>;
1187 //===---------------------------------------------
1189 // Hacked form of AND to zero-extend 16-bit quantities to 32-bit
1190 // quantities -- see 16->32 zext pattern.
1192 // This pattern is somewhat artificial, since it might match some
1193 // compiler generated pattern but it is unlikely to do so.
1195 def i16i32: ANDInst<(outs R32C:$rT), (ins R16C:$rA, R32C:$rB),
1196 [(set R32C:$rT, (and (zext R16C:$rA), R32C:$rB))]>;
1199 defm AND : BitwiseAnd;
1201 // N.B.: vnot_conv is one of those special target selection pattern fragments,
1202 // in which we expect there to be a bit_convert on the constant. Bear in mind
1203 // that llvm translates "not <reg>" to "xor <reg>, -1" (or in this case, a
1204 // constant -1 vector.)
1206 class ANDCInst<dag OOL, dag IOL, list<dag> pattern>:
1207 RRForm<0b10000011010, OOL, IOL, "andc\t$rT, $rA, $rB",
1208 IntegerOp, pattern>;
1210 class ANDCVecInst<ValueType vectype>:
1211 ANDCInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1212 [(set (vectype VECREG:$rT), (and (vectype VECREG:$rA),
1213 (vnot (vectype VECREG:$rB))))]>;
1215 class ANDCRegInst<RegisterClass rclass>:
1216 ANDCInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1217 [(set rclass:$rT, (and rclass:$rA, (not rclass:$rB)))]>;
1219 multiclass AndComplement
1221 def v16i8: ANDCVecInst<v16i8>;
1222 def v8i16: ANDCVecInst<v8i16>;
1223 def v4i32: ANDCVecInst<v4i32>;
1224 def v2i64: ANDCVecInst<v2i64>;
1226 def r128: ANDCRegInst<GPRC>;
1227 def r64: ANDCRegInst<R64C>;
1228 def r32: ANDCRegInst<R32C>;
1229 def r16: ANDCRegInst<R16C>;
1230 def r8: ANDCRegInst<R8C>;
1233 defm ANDC : AndComplement;
1235 class ANDBIInst<dag OOL, dag IOL, list<dag> pattern>:
1236 RI10Form<0b01101000, OOL, IOL, "andbi\t$rT, $rA, $val",
1239 multiclass AndByteImm
1241 def v16i8: ANDBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1242 [(set (v16i8 VECREG:$rT),
1243 (and (v16i8 VECREG:$rA),
1244 (v16i8 v16i8U8Imm:$val)))]>;
1246 def r8: ANDBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1247 [(set R8C:$rT, (and R8C:$rA, immU8:$val))]>;
1250 defm ANDBI : AndByteImm;
1252 class ANDHIInst<dag OOL, dag IOL, list<dag> pattern> :
1253 RI10Form<0b10101000, OOL, IOL, "andhi\t$rT, $rA, $val",
1256 multiclass AndHalfwordImm
1258 def v8i16: ANDHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
1259 [(set (v8i16 VECREG:$rT),
1260 (and (v8i16 VECREG:$rA), v8i16SExt10Imm:$val))]>;
1262 def r16: ANDHIInst<(outs R16C:$rT), (ins R16C:$rA, u10imm:$val),
1263 [(set R16C:$rT, (and R16C:$rA, i16ImmUns10:$val))]>;
1265 // Zero-extend i8 to i16:
1266 def i8i16: ANDHIInst<(outs R16C:$rT), (ins R8C:$rA, u10imm:$val),
1267 [(set R16C:$rT, (and (zext R8C:$rA), i16ImmUns10:$val))]>;
1270 defm ANDHI : AndHalfwordImm;
1272 class ANDIInst<dag OOL, dag IOL, list<dag> pattern> :
1273 RI10Form<0b00101000, OOL, IOL, "andi\t$rT, $rA, $val",
1274 IntegerOp, pattern>;
1276 multiclass AndWordImm
1278 def v4i32: ANDIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
1279 [(set (v4i32 VECREG:$rT),
1280 (and (v4i32 VECREG:$rA), v4i32SExt10Imm:$val))]>;
1282 def r32: ANDIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
1283 [(set R32C:$rT, (and R32C:$rA, i32ImmSExt10:$val))]>;
1285 // Hacked form of ANDI to zero-extend i8 quantities to i32. See the zext 8->32
1287 def i8i32: ANDIInst<(outs R32C:$rT), (ins R8C:$rA, s10imm_i32:$val),
1289 (and (zext R8C:$rA), i32ImmSExt10:$val))]>;
1291 // Hacked form of ANDI to zero-extend i16 quantities to i32. See the
1292 // zext 16->32 pattern below.
1294 // Note that this pattern is somewhat artificial, since it might match
1295 // something the compiler generates but is unlikely to occur in practice.
1296 def i16i32: ANDIInst<(outs R32C:$rT), (ins R16C:$rA, s10imm_i32:$val),
1298 (and (zext R16C:$rA), i32ImmSExt10:$val))]>;
1301 defm ANDI : AndWordImm;
1303 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1304 // Bitwise OR group:
1305 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1307 // Bitwise "or" (N.B.: These are also register-register copy instructions...)
1308 class ORInst<dag OOL, dag IOL, list<dag> pattern>:
1309 RRForm<0b10000010000, OOL, IOL, "or\t$rT, $rA, $rB",
1310 IntegerOp, pattern>;
1312 class ORVecInst<ValueType vectype>:
1313 ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1314 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1315 (vectype VECREG:$rB)))]>;
1317 class ORRegInst<RegisterClass rclass>:
1318 ORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1319 [(set rclass:$rT, (or rclass:$rA, rclass:$rB))]>;
1321 // ORCvtForm: OR conversion form
1323 // This is used to "convert" the preferred slot to its vector equivalent, as
1324 // well as convert a vector back to its preferred slot.
1326 // These are effectively no-ops, but need to exist for proper type conversion
1327 // and type coercion.
1329 class ORCvtForm<dag OOL, dag IOL>
1330 : SPUInstr<OOL, IOL, "or\t$rT, $rA, $rA", IntegerOp> {
1334 let Pattern = [/* no pattern */];
1336 let Inst{0-10} = 0b10000010000;
1337 let Inst{11-17} = RA;
1338 let Inst{18-24} = RA;
1339 let Inst{25-31} = RT;
1342 class ORPromoteScalar<RegisterClass rclass>:
1343 ORCvtForm<(outs VECREG:$rT), (ins rclass:$rA)>;
1345 class ORExtractElt<RegisterClass rclass>:
1346 ORCvtForm<(outs rclass:$rT), (ins VECREG:$rA)>;
1348 class ORCvtRegGPRC<RegisterClass rclass>:
1349 ORCvtForm<(outs GPRC:$rT), (ins rclass:$rA)>;
1352 ORCvtForm<(outs GPRC:$rT), (ins VECREG:$rA)>;
1354 class ORCvtGPRCReg<RegisterClass rclass>:
1355 ORCvtForm<(outs rclass:$rT), (ins GPRC:$rA)>;
1358 ORCvtForm<(outs VECREG:$rT), (ins GPRC:$rA)>;
1360 multiclass BitwiseOr
1362 def v16i8: ORVecInst<v16i8>;
1363 def v8i16: ORVecInst<v8i16>;
1364 def v4i32: ORVecInst<v4i32>;
1365 def v2i64: ORVecInst<v2i64>;
1367 def v4f32: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1368 [(set (v4f32 VECREG:$rT),
1369 (v4f32 (bitconvert (or (v4i32 VECREG:$rA),
1370 (v4i32 VECREG:$rB)))))]>;
1372 def v2f64: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1373 [(set (v2f64 VECREG:$rT),
1374 (v2f64 (bitconvert (or (v2i64 VECREG:$rA),
1375 (v2i64 VECREG:$rB)))))]>;
1377 def r64: ORRegInst<R64C>;
1378 def r32: ORRegInst<R32C>;
1379 def r16: ORRegInst<R16C>;
1380 def r8: ORRegInst<R8C>;
1382 // OR instructions used to copy f32 and f64 registers.
1383 def f32: ORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
1384 [/* no pattern */]>;
1386 def f64: ORInst<(outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
1387 [/* no pattern */]>;
1389 // scalar->vector promotion:
1390 def v16i8_i8: ORPromoteScalar<R8C>;
1391 def v8i16_i16: ORPromoteScalar<R16C>;
1392 def v4i32_i32: ORPromoteScalar<R32C>;
1393 def v2i64_i64: ORPromoteScalar<R64C>;
1394 def v4f32_f32: ORPromoteScalar<R32FP>;
1395 def v2f64_f64: ORPromoteScalar<R64FP>;
1397 // extract element 0:
1398 def i8_v16i8: ORExtractElt<R8C>;
1399 def i16_v8i16: ORExtractElt<R16C>;
1400 def i32_v4i32: ORExtractElt<R32C>;
1401 def i64_v2i64: ORExtractElt<R64C>;
1402 def f32_v4f32: ORExtractElt<R32FP>;
1403 def f64_v2f64: ORExtractElt<R64FP>;
1405 // Conversion from GPRC to register
1406 def i128_r64: ORCvtRegGPRC<R64C>;
1407 def i128_f64: ORCvtRegGPRC<R64FP>;
1408 def i128_r32: ORCvtRegGPRC<R32C>;
1409 def i128_f32: ORCvtRegGPRC<R32FP>;
1410 def i128_r16: ORCvtRegGPRC<R16C>;
1411 def i128_r8: ORCvtRegGPRC<R8C>;
1413 // Conversion from GPRC to vector
1414 def i128_vec: ORCvtVecGPRC;
1416 // Conversion from register to GPRC
1417 def r64_i128: ORCvtGPRCReg<R64C>;
1418 def f64_i128: ORCvtGPRCReg<R64FP>;
1419 def r32_i128: ORCvtGPRCReg<R32C>;
1420 def f32_i128: ORCvtGPRCReg<R32FP>;
1421 def r16_i128: ORCvtGPRCReg<R16C>;
1422 def r8_i128: ORCvtGPRCReg<R8C>;
1424 // Conversion from vector to GPRC
1425 def vec_i128: ORCvtGPRCVec;
1428 defm OR : BitwiseOr;
1430 // scalar->vector promotion patterns (preferred slot to vector):
1431 def : Pat<(v16i8 (SPUprefslot2vec R8C:$rA)),
1432 (ORv16i8_i8 R8C:$rA)>;
1434 def : Pat<(v8i16 (SPUprefslot2vec R16C:$rA)),
1435 (ORv8i16_i16 R16C:$rA)>;
1437 def : Pat<(v4i32 (SPUprefslot2vec R32C:$rA)),
1438 (ORv4i32_i32 R32C:$rA)>;
1440 def : Pat<(v2i64 (SPUprefslot2vec R64C:$rA)),
1441 (ORv2i64_i64 R64C:$rA)>;
1443 def : Pat<(v4f32 (SPUprefslot2vec R32FP:$rA)),
1444 (ORv4f32_f32 R32FP:$rA)>;
1446 def : Pat<(v2f64 (SPUprefslot2vec R64FP:$rA)),
1447 (ORv2f64_f64 R64FP:$rA)>;
1449 // ORi*_v*: Used to extract vector element 0 (the preferred slot), otherwise
1450 // known as converting the vector back to its preferred slot
1452 def : Pat<(SPUvec2prefslot (v16i8 VECREG:$rA)),
1453 (ORi8_v16i8 VECREG:$rA)>;
1455 def : Pat<(SPUvec2prefslot (v8i16 VECREG:$rA)),
1456 (ORi16_v8i16 VECREG:$rA)>;
1458 def : Pat<(SPUvec2prefslot (v4i32 VECREG:$rA)),
1459 (ORi32_v4i32 VECREG:$rA)>;
1461 def : Pat<(SPUvec2prefslot (v2i64 VECREG:$rA)),
1462 (ORi64_v2i64 VECREG:$rA)>;
1464 def : Pat<(SPUvec2prefslot (v4f32 VECREG:$rA)),
1465 (ORf32_v4f32 VECREG:$rA)>;
1467 def : Pat<(SPUvec2prefslot (v2f64 VECREG:$rA)),
1468 (ORf64_v2f64 VECREG:$rA)>;
1470 // Load Register: This is an assembler alias for a bitwise OR of a register
1471 // against itself. It's here because it brings some clarity to assembly
1474 let hasCtrlDep = 1 in {
1475 class LRInst<dag OOL, dag IOL>
1476 : SPUInstr<OOL, IOL, "lr\t$rT, $rA", IntegerOp> {
1480 let Pattern = [/*no pattern*/];
1482 let Inst{0-10} = 0b10000010000; /* It's an OR operation */
1483 let Inst{11-17} = RA;
1484 let Inst{18-24} = RA;
1485 let Inst{25-31} = RT;
1488 class LRVecInst<ValueType vectype>:
1489 LRInst<(outs VECREG:$rT), (ins VECREG:$rA)>;
1491 class LRRegInst<RegisterClass rclass>:
1492 LRInst<(outs rclass:$rT), (ins rclass:$rA)>;
1494 multiclass LoadRegister {
1495 def v2i64: LRVecInst<v2i64>;
1496 def v2f64: LRVecInst<v2f64>;
1497 def v4i32: LRVecInst<v4i32>;
1498 def v4f32: LRVecInst<v4f32>;
1499 def v8i16: LRVecInst<v8i16>;
1500 def v16i8: LRVecInst<v16i8>;
1502 def r128: LRRegInst<GPRC>;
1503 def r64: LRRegInst<R64C>;
1504 def f64: LRRegInst<R64FP>;
1505 def r32: LRRegInst<R32C>;
1506 def f32: LRRegInst<R32FP>;
1507 def r16: LRRegInst<R16C>;
1508 def r8: LRRegInst<R8C>;
1511 defm LR: LoadRegister;
1514 // ORC: Bitwise "or" with complement (c = a | ~b)
1516 class ORCInst<dag OOL, dag IOL, list<dag> pattern>:
1517 RRForm<0b10010010000, OOL, IOL, "orc\t$rT, $rA, $rB",
1518 IntegerOp, pattern>;
1520 class ORCVecInst<ValueType vectype>:
1521 ORCInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1522 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1523 (vnot (vectype VECREG:$rB))))]>;
1525 class ORCRegInst<RegisterClass rclass>:
1526 ORCInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1527 [(set rclass:$rT, (or rclass:$rA, (not rclass:$rB)))]>;
1529 multiclass BitwiseOrComplement
1531 def v16i8: ORCVecInst<v16i8>;
1532 def v8i16: ORCVecInst<v8i16>;
1533 def v4i32: ORCVecInst<v4i32>;
1534 def v2i64: ORCVecInst<v2i64>;
1536 def r64: ORCRegInst<R64C>;
1537 def r32: ORCRegInst<R32C>;
1538 def r16: ORCRegInst<R16C>;
1539 def r8: ORCRegInst<R8C>;
1542 defm ORC : BitwiseOrComplement;
1544 // OR byte immediate
1545 class ORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1546 RI10Form<0b01100000, OOL, IOL, "orbi\t$rT, $rA, $val",
1547 IntegerOp, pattern>;
1549 class ORBIVecInst<ValueType vectype, PatLeaf immpred>:
1550 ORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1551 [(set (v16i8 VECREG:$rT), (or (vectype VECREG:$rA),
1552 (vectype immpred:$val)))]>;
1554 multiclass BitwiseOrByteImm
1556 def v16i8: ORBIVecInst<v16i8, v16i8U8Imm>;
1558 def r8: ORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1559 [(set R8C:$rT, (or R8C:$rA, immU8:$val))]>;
1562 defm ORBI : BitwiseOrByteImm;
1564 // OR halfword immediate
1565 class ORHIInst<dag OOL, dag IOL, list<dag> pattern>:
1566 RI10Form<0b10100000, OOL, IOL, "orhi\t$rT, $rA, $val",
1567 IntegerOp, pattern>;
1569 class ORHIVecInst<ValueType vectype, PatLeaf immpred>:
1570 ORHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1571 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1574 multiclass BitwiseOrHalfwordImm
1576 def v8i16: ORHIVecInst<v8i16, v8i16Uns10Imm>;
1578 def r16: ORHIInst<(outs R16C:$rT), (ins R16C:$rA, u10imm:$val),
1579 [(set R16C:$rT, (or R16C:$rA, i16ImmUns10:$val))]>;
1581 // Specialized ORHI form used to promote 8-bit registers to 16-bit
1582 def i8i16: ORHIInst<(outs R16C:$rT), (ins R8C:$rA, s10imm:$val),
1583 [(set R16C:$rT, (or (anyext R8C:$rA),
1584 i16ImmSExt10:$val))]>;
1587 defm ORHI : BitwiseOrHalfwordImm;
1589 class ORIInst<dag OOL, dag IOL, list<dag> pattern>:
1590 RI10Form<0b00100000, OOL, IOL, "ori\t$rT, $rA, $val",
1591 IntegerOp, pattern>;
1593 class ORIVecInst<ValueType vectype, PatLeaf immpred>:
1594 ORIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1595 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1598 // Bitwise "or" with immediate
1599 multiclass BitwiseOrImm
1601 def v4i32: ORIVecInst<v4i32, v4i32Uns10Imm>;
1603 def r32: ORIInst<(outs R32C:$rT), (ins R32C:$rA, u10imm_i32:$val),
1604 [(set R32C:$rT, (or R32C:$rA, i32ImmUns10:$val))]>;
1606 // i16i32: hacked version of the ori instruction to extend 16-bit quantities
1607 // to 32-bit quantities. used exclusively to match "anyext" conversions (vide
1608 // infra "anyext 16->32" pattern.)
1609 def i16i32: ORIInst<(outs R32C:$rT), (ins R16C:$rA, s10imm_i32:$val),
1610 [(set R32C:$rT, (or (anyext R16C:$rA),
1611 i32ImmSExt10:$val))]>;
1613 // i8i32: Hacked version of the ORI instruction to extend 16-bit quantities
1614 // to 32-bit quantities. Used exclusively to match "anyext" conversions (vide
1615 // infra "anyext 16->32" pattern.)
1616 def i8i32: ORIInst<(outs R32C:$rT), (ins R8C:$rA, s10imm_i32:$val),
1617 [(set R32C:$rT, (or (anyext R8C:$rA),
1618 i32ImmSExt10:$val))]>;
1621 defm ORI : BitwiseOrImm;
1623 // ORX: "or" across the vector: or's $rA's word slots leaving the result in
1624 // $rT[0], slots 1-3 are zeroed.
1626 // FIXME: Needs to match an intrinsic pattern.
1628 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1629 "orx\t$rT, $rA, $rB", IntegerOp,
1634 class XORInst<dag OOL, dag IOL, list<dag> pattern> :
1635 RRForm<0b10010010000, OOL, IOL, "xor\t$rT, $rA, $rB",
1636 IntegerOp, pattern>;
1638 class XORVecInst<ValueType vectype>:
1639 XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1640 [(set (vectype VECREG:$rT), (xor (vectype VECREG:$rA),
1641 (vectype VECREG:$rB)))]>;
1643 class XORRegInst<RegisterClass rclass>:
1644 XORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1645 [(set rclass:$rT, (xor rclass:$rA, rclass:$rB))]>;
1647 multiclass BitwiseExclusiveOr
1649 def v16i8: XORVecInst<v16i8>;
1650 def v8i16: XORVecInst<v8i16>;
1651 def v4i32: XORVecInst<v4i32>;
1652 def v2i64: XORVecInst<v2i64>;
1654 def r128: XORRegInst<GPRC>;
1655 def r64: XORRegInst<R64C>;
1656 def r32: XORRegInst<R32C>;
1657 def r16: XORRegInst<R16C>;
1658 def r8: XORRegInst<R8C>;
1660 // Special forms for floating point instructions.
1661 // fneg and fabs require bitwise logical ops to manipulate the sign bit.
1663 def fneg32: XORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32C:$rB),
1664 [/* no pattern */]>;
1666 def fneg64: XORInst<(outs R64FP:$rT), (ins R64FP:$rA, VECREG:$rB),
1667 [/* no pattern */]>;
1669 def fnegvec: XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1670 [/* no pattern, see fneg{32,64} */]>;
1673 defm XOR : BitwiseExclusiveOr;
1675 //==----------------------------------------------------------
1677 class XORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1678 RI10Form<0b01100000, OOL, IOL, "xorbi\t$rT, $rA, $val",
1679 IntegerOp, pattern>;
1681 multiclass XorByteImm
1684 XORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1685 [(set (v16i8 VECREG:$rT), (xor (v16i8 VECREG:$rA), v16i8U8Imm:$val))]>;
1688 XORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1689 [(set R8C:$rT, (xor R8C:$rA, immU8:$val))]>;
1692 defm XORBI : XorByteImm;
1695 RI10Form<0b10100000, (outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1696 "xorhi\t$rT, $rA, $val", IntegerOp,
1697 [(set (v8i16 VECREG:$rT), (xor (v8i16 VECREG:$rA),
1698 v8i16SExt10Imm:$val))]>;
1701 RI10Form<0b10100000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
1702 "xorhi\t$rT, $rA, $val", IntegerOp,
1703 [(set R16C:$rT, (xor R16C:$rA, i16ImmSExt10:$val))]>;
1706 RI10Form<0b00100000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm_i32:$val),
1707 "xori\t$rT, $rA, $val", IntegerOp,
1708 [(set (v4i32 VECREG:$rT), (xor (v4i32 VECREG:$rA),
1709 v4i32SExt10Imm:$val))]>;
1712 RI10Form<0b00100000, (outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
1713 "xori\t$rT, $rA, $val", IntegerOp,
1714 [(set R32C:$rT, (xor R32C:$rA, i32ImmSExt10:$val))]>;
1718 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1719 "nand\t$rT, $rA, $rB", IntegerOp,
1720 [(set (v16i8 VECREG:$rT), (vnot (and (v16i8 VECREG:$rA),
1721 (v16i8 VECREG:$rB))))]>;
1724 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1725 "nand\t$rT, $rA, $rB", IntegerOp,
1726 [(set (v8i16 VECREG:$rT), (vnot (and (v8i16 VECREG:$rA),
1727 (v8i16 VECREG:$rB))))]>;
1730 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1731 "nand\t$rT, $rA, $rB", IntegerOp,
1732 [(set (v4i32 VECREG:$rT), (vnot (and (v4i32 VECREG:$rA),
1733 (v4i32 VECREG:$rB))))]>;
1736 RRForm<0b10010010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
1737 "nand\t$rT, $rA, $rB", IntegerOp,
1738 [(set R32C:$rT, (not (and R32C:$rA, R32C:$rB)))]>;
1741 RRForm<0b10010010000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
1742 "nand\t$rT, $rA, $rB", IntegerOp,
1743 [(set R16C:$rT, (not (and R16C:$rA, R16C:$rB)))]>;
1746 RRForm<0b10010010000, (outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
1747 "nand\t$rT, $rA, $rB", IntegerOp,
1748 [(set R8C:$rT, (not (and R8C:$rA, R8C:$rB)))]>;
1752 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1753 "nor\t$rT, $rA, $rB", IntegerOp,
1754 [(set (v16i8 VECREG:$rT), (vnot (or (v16i8 VECREG:$rA),
1755 (v16i8 VECREG:$rB))))]>;
1758 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1759 "nor\t$rT, $rA, $rB", IntegerOp,
1760 [(set (v8i16 VECREG:$rT), (vnot (or (v8i16 VECREG:$rA),
1761 (v8i16 VECREG:$rB))))]>;
1764 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1765 "nor\t$rT, $rA, $rB", IntegerOp,
1766 [(set (v4i32 VECREG:$rT), (vnot (or (v4i32 VECREG:$rA),
1767 (v4i32 VECREG:$rB))))]>;
1770 RRForm<0b10010010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
1771 "nor\t$rT, $rA, $rB", IntegerOp,
1772 [(set R32C:$rT, (not (or R32C:$rA, R32C:$rB)))]>;
1775 RRForm<0b10010010000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
1776 "nor\t$rT, $rA, $rB", IntegerOp,
1777 [(set R16C:$rT, (not (or R16C:$rA, R16C:$rB)))]>;
1780 RRForm<0b10010010000, (outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
1781 "nor\t$rT, $rA, $rB", IntegerOp,
1782 [(set R8C:$rT, (not (or R8C:$rA, R8C:$rB)))]>;
1785 class SELBInst<dag OOL, dag IOL, list<dag> pattern>:
1786 RRRForm<0b1000, OOL, IOL, "selb\t$rT, $rA, $rB, $rC",
1787 IntegerOp, pattern>;
1789 class SELBVecInst<ValueType vectype>:
1790 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1791 [(set (vectype VECREG:$rT),
1792 (or (and (vectype VECREG:$rC), (vectype VECREG:$rB)),
1793 (and (vnot (vectype VECREG:$rC)),
1794 (vectype VECREG:$rA))))]>;
1796 class SELBVecCondInst<ValueType vectype>:
1797 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, R32C:$rC),
1798 [(set (vectype VECREG:$rT),
1800 (vectype VECREG:$rB),
1801 (vectype VECREG:$rA)))]>;
1803 class SELBRegInst<RegisterClass rclass>:
1804 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rclass:$rC),
1806 (or (and rclass:$rA, rclass:$rC),
1807 (and rclass:$rB, (not rclass:$rC))))]>;
1809 class SELBRegCondInst<RegisterClass rcond, RegisterClass rclass>:
1810 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rcond:$rC),
1812 (select rcond:$rC, rclass:$rB, rclass:$rA))]>;
1814 multiclass SelectBits
1816 def v16i8: SELBVecInst<v16i8>;
1817 def v8i16: SELBVecInst<v8i16>;
1818 def v4i32: SELBVecInst<v4i32>;
1819 def v2i64: SELBVecInst<v2i64>;
1821 def r128: SELBRegInst<GPRC>;
1822 def r64: SELBRegInst<R64C>;
1823 def r32: SELBRegInst<R32C>;
1824 def r16: SELBRegInst<R16C>;
1825 def r8: SELBRegInst<R8C>;
1827 def v16i8_cond: SELBVecCondInst<v16i8>;
1828 def v8i16_cond: SELBVecCondInst<v8i16>;
1829 def v4i32_cond: SELBVecCondInst<v4i32>;
1830 def v2i64_cond: SELBVecCondInst<v2i64>;
1832 // SELBr64_cond is defined further down, look for i64 comparisons
1833 def r32_cond: SELBRegCondInst<R32C, R32C>;
1834 def r16_cond: SELBRegCondInst<R16C, R16C>;
1835 def r8_cond: SELBRegCondInst<R8C, R8C>;
1838 defm SELB : SelectBits;
1840 class SPUselbPatVec<ValueType vectype, SPUInstr inst>:
1841 Pat<(SPUselb (vectype VECREG:$rA), (vectype VECREG:$rB), (vectype VECREG:$rC)),
1842 (inst VECREG:$rA, VECREG:$rB, VECREG:$rC)>;
1844 def : SPUselbPatVec<v16i8, SELBv16i8>;
1845 def : SPUselbPatVec<v8i16, SELBv8i16>;
1846 def : SPUselbPatVec<v4i32, SELBv4i32>;
1847 def : SPUselbPatVec<v2i64, SELBv2i64>;
1849 class SPUselbPatReg<RegisterClass rclass, SPUInstr inst>:
1850 Pat<(SPUselb rclass:$rA, rclass:$rB, rclass:$rC),
1851 (inst rclass:$rA, rclass:$rB, rclass:$rC)>;
1853 def : SPUselbPatReg<R8C, SELBr8>;
1854 def : SPUselbPatReg<R16C, SELBr16>;
1855 def : SPUselbPatReg<R32C, SELBr32>;
1856 def : SPUselbPatReg<R64C, SELBr64>;
1858 // EQV: Equivalence (1 for each same bit, otherwise 0)
1860 // Note: There are a lot of ways to match this bit operator and these patterns
1861 // attempt to be as exhaustive as possible.
1863 class EQVInst<dag OOL, dag IOL, list<dag> pattern>:
1864 RRForm<0b10010010000, OOL, IOL, "eqv\t$rT, $rA, $rB",
1865 IntegerOp, pattern>;
1867 class EQVVecInst<ValueType vectype>:
1868 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1869 [(set (vectype VECREG:$rT),
1870 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
1871 (and (vnot (vectype VECREG:$rA)),
1872 (vnot (vectype VECREG:$rB)))))]>;
1874 class EQVRegInst<RegisterClass rclass>:
1875 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1876 [(set rclass:$rT, (or (and rclass:$rA, rclass:$rB),
1877 (and (not rclass:$rA), (not rclass:$rB))))]>;
1879 class EQVVecPattern1<ValueType vectype>:
1880 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1881 [(set (vectype VECREG:$rT),
1882 (xor (vectype VECREG:$rA), (vnot (vectype VECREG:$rB))))]>;
1884 class EQVRegPattern1<RegisterClass rclass>:
1885 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1886 [(set rclass:$rT, (xor rclass:$rA, (not rclass:$rB)))]>;
1888 class EQVVecPattern2<ValueType vectype>:
1889 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1890 [(set (vectype VECREG:$rT),
1891 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
1892 (vnot (or (vectype VECREG:$rA), (vectype VECREG:$rB)))))]>;
1894 class EQVRegPattern2<RegisterClass rclass>:
1895 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1897 (or (and rclass:$rA, rclass:$rB),
1898 (not (or rclass:$rA, rclass:$rB))))]>;
1900 class EQVVecPattern3<ValueType vectype>:
1901 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1902 [(set (vectype VECREG:$rT),
1903 (not (xor (vectype VECREG:$rA), (vectype VECREG:$rB))))]>;
1905 class EQVRegPattern3<RegisterClass rclass>:
1906 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1907 [(set rclass:$rT, (not (xor rclass:$rA, rclass:$rB)))]>;
1909 multiclass BitEquivalence
1911 def v16i8: EQVVecInst<v16i8>;
1912 def v8i16: EQVVecInst<v8i16>;
1913 def v4i32: EQVVecInst<v4i32>;
1914 def v2i64: EQVVecInst<v2i64>;
1916 def v16i8_1: EQVVecPattern1<v16i8>;
1917 def v8i16_1: EQVVecPattern1<v8i16>;
1918 def v4i32_1: EQVVecPattern1<v4i32>;
1919 def v2i64_1: EQVVecPattern1<v2i64>;
1921 def v16i8_2: EQVVecPattern2<v16i8>;
1922 def v8i16_2: EQVVecPattern2<v8i16>;
1923 def v4i32_2: EQVVecPattern2<v4i32>;
1924 def v2i64_2: EQVVecPattern2<v2i64>;
1926 def v16i8_3: EQVVecPattern3<v16i8>;
1927 def v8i16_3: EQVVecPattern3<v8i16>;
1928 def v4i32_3: EQVVecPattern3<v4i32>;
1929 def v2i64_3: EQVVecPattern3<v2i64>;
1931 def r128: EQVRegInst<GPRC>;
1932 def r64: EQVRegInst<R64C>;
1933 def r32: EQVRegInst<R32C>;
1934 def r16: EQVRegInst<R16C>;
1935 def r8: EQVRegInst<R8C>;
1937 def r128_1: EQVRegPattern1<GPRC>;
1938 def r64_1: EQVRegPattern1<R64C>;
1939 def r32_1: EQVRegPattern1<R32C>;
1940 def r16_1: EQVRegPattern1<R16C>;
1941 def r8_1: EQVRegPattern1<R8C>;
1943 def r128_2: EQVRegPattern2<GPRC>;
1944 def r64_2: EQVRegPattern2<R64C>;
1945 def r32_2: EQVRegPattern2<R32C>;
1946 def r16_2: EQVRegPattern2<R16C>;
1947 def r8_2: EQVRegPattern2<R8C>;
1949 def r128_3: EQVRegPattern3<GPRC>;
1950 def r64_3: EQVRegPattern3<R64C>;
1951 def r32_3: EQVRegPattern3<R32C>;
1952 def r16_3: EQVRegPattern3<R16C>;
1953 def r8_3: EQVRegPattern3<R8C>;
1956 defm EQV: BitEquivalence;
1958 //===----------------------------------------------------------------------===//
1959 // Vector shuffle...
1960 //===----------------------------------------------------------------------===//
1961 // SPUshuffle is generated in LowerVECTOR_SHUFFLE and gets replaced with SHUFB.
1962 // See the SPUshuffle SDNode operand above, which sets up the DAG pattern
1963 // matcher to emit something when the LowerVECTOR_SHUFFLE generates a node with
1964 // the SPUISD::SHUFB opcode.
1965 //===----------------------------------------------------------------------===//
1967 class SHUFBInst<dag OOL, dag IOL, list<dag> pattern>:
1968 RRRForm<0b1000, OOL, IOL, "shufb\t$rT, $rA, $rB, $rC",
1969 IntegerOp, pattern>;
1971 class SHUFBVecInst<ValueType resultvec, ValueType maskvec>:
1972 SHUFBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1973 [(set (resultvec VECREG:$rT),
1974 (SPUshuffle (resultvec VECREG:$rA),
1975 (resultvec VECREG:$rB),
1976 (maskvec VECREG:$rC)))]>;
1978 class SHUFBGPRCInst:
1979 SHUFBInst<(outs VECREG:$rT), (ins GPRC:$rA, GPRC:$rB, VECREG:$rC),
1980 [/* no pattern */]>;
1982 multiclass ShuffleBytes
1984 def v16i8 : SHUFBVecInst<v16i8, v16i8>;
1985 def v16i8_m32 : SHUFBVecInst<v16i8, v4i32>;
1986 def v8i16 : SHUFBVecInst<v8i16, v16i8>;
1987 def v8i16_m32 : SHUFBVecInst<v8i16, v4i32>;
1988 def v4i32 : SHUFBVecInst<v4i32, v16i8>;
1989 def v4i32_m32 : SHUFBVecInst<v4i32, v4i32>;
1990 def v2i64 : SHUFBVecInst<v2i64, v16i8>;
1991 def v2i64_m32 : SHUFBVecInst<v2i64, v4i32>;
1993 def v4f32 : SHUFBVecInst<v4f32, v16i8>;
1994 def v4f32_m32 : SHUFBVecInst<v4f32, v4i32>;
1996 def v2f64 : SHUFBVecInst<v2f64, v16i8>;
1997 def v2f64_m32 : SHUFBVecInst<v2f64, v4i32>;
1999 def gprc : SHUFBGPRCInst;
2002 defm SHUFB : ShuffleBytes;
2004 //===----------------------------------------------------------------------===//
2005 // Shift and rotate group:
2006 //===----------------------------------------------------------------------===//
2008 class SHLHInst<dag OOL, dag IOL, list<dag> pattern>:
2009 RRForm<0b11111010000, OOL, IOL, "shlh\t$rT, $rA, $rB",
2010 RotateShift, pattern>;
2012 class SHLHVecInst<ValueType vectype>:
2013 SHLHInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2014 [(set (vectype VECREG:$rT),
2015 (SPUvec_shl (vectype VECREG:$rA), R16C:$rB))]>;
2017 // $rB gets promoted to 32-bit register type when confronted with
2018 // this llvm assembly code:
2020 // define i16 @shlh_i16_1(i16 %arg1, i16 %arg2) {
2021 // %A = shl i16 %arg1, %arg2
2025 multiclass ShiftLeftHalfword
2027 def v8i16: SHLHVecInst<v8i16>;
2028 def r16: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
2029 [(set R16C:$rT, (shl R16C:$rA, R16C:$rB))]>;
2030 def r16_r32: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2031 [(set R16C:$rT, (shl R16C:$rA, R32C:$rB))]>;
2034 defm SHLH : ShiftLeftHalfword;
2036 //===----------------------------------------------------------------------===//
2038 class SHLHIInst<dag OOL, dag IOL, list<dag> pattern>:
2039 RI7Form<0b11111010000, OOL, IOL, "shlhi\t$rT, $rA, $val",
2040 RotateShift, pattern>;
2042 class SHLHIVecInst<ValueType vectype>:
2043 SHLHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2044 [(set (vectype VECREG:$rT),
2045 (SPUvec_shl (vectype VECREG:$rA), (i16 uimm7:$val)))]>;
2047 multiclass ShiftLeftHalfwordImm
2049 def v8i16: SHLHIVecInst<v8i16>;
2050 def r16: SHLHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2051 [(set R16C:$rT, (shl R16C:$rA, (i16 uimm7:$val)))]>;
2054 defm SHLHI : ShiftLeftHalfwordImm;
2056 def : Pat<(SPUvec_shl (v8i16 VECREG:$rA), (i32 uimm7:$val)),
2057 (SHLHIv8i16 VECREG:$rA, uimm7:$val)>;
2059 def : Pat<(shl R16C:$rA, (i32 uimm7:$val)),
2060 (SHLHIr16 R16C:$rA, uimm7:$val)>;
2062 //===----------------------------------------------------------------------===//
2064 class SHLInst<dag OOL, dag IOL, list<dag> pattern>:
2065 RRForm<0b11111010000, OOL, IOL, "shl\t$rT, $rA, $rB",
2066 RotateShift, pattern>;
2068 multiclass ShiftLeftWord
2071 SHLInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2072 [(set (v4i32 VECREG:$rT),
2073 (SPUvec_shl (v4i32 VECREG:$rA), R16C:$rB))]>;
2075 SHLInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2076 [(set R32C:$rT, (shl R32C:$rA, R32C:$rB))]>;
2079 defm SHL: ShiftLeftWord;
2081 //===----------------------------------------------------------------------===//
2083 class SHLIInst<dag OOL, dag IOL, list<dag> pattern>:
2084 RI7Form<0b11111010000, OOL, IOL, "shli\t$rT, $rA, $val",
2085 RotateShift, pattern>;
2087 multiclass ShiftLeftWordImm
2090 SHLIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2091 [(set (v4i32 VECREG:$rT),
2092 (SPUvec_shl (v4i32 VECREG:$rA), (i32 uimm7:$val)))]>;
2095 SHLIInst<(outs R32C:$rT), (ins R32C:$rA, u7imm_i32:$val),
2096 [(set R32C:$rT, (shl R32C:$rA, (i32 uimm7:$val)))]>;
2099 defm SHLI : ShiftLeftWordImm;
2101 //===----------------------------------------------------------------------===//
2102 // SHLQBI vec form: Note that this will shift the entire vector (the 128-bit
2103 // register) to the left. Vector form is here to ensure type correctness.
2105 // The shift count is in the lowest 3 bits (29-31) of $rB, so only a bit shift
2106 // of 7 bits is actually possible.
2108 // Note also that SHLQBI/SHLQBII are used in conjunction with SHLQBY/SHLQBYI
2109 // to shift i64 and i128. SHLQBI is the residual left over after shifting by
2110 // bytes with SHLQBY.
2112 class SHLQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2113 RRForm<0b11011011100, OOL, IOL, "shlqbi\t$rT, $rA, $rB",
2114 RotateShift, pattern>;
2116 class SHLQBIVecInst<ValueType vectype>:
2117 SHLQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2118 [(set (vectype VECREG:$rT),
2119 (SPUshlquad_l_bits (vectype VECREG:$rA), R32C:$rB))]>;
2121 multiclass ShiftLeftQuadByBits
2123 def v16i8: SHLQBIVecInst<v16i8>;
2124 def v8i16: SHLQBIVecInst<v8i16>;
2125 def v4i32: SHLQBIVecInst<v4i32>;
2126 def v4f32: SHLQBIVecInst<v4f32>;
2127 def v2i64: SHLQBIVecInst<v2i64>;
2128 def v2f64: SHLQBIVecInst<v2f64>;
2131 defm SHLQBI : ShiftLeftQuadByBits;
2133 // See note above on SHLQBI. In this case, the predicate actually does then
2134 // enforcement, whereas with SHLQBI, we have to "take it on faith."
2135 class SHLQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2136 RI7Form<0b11011111100, OOL, IOL, "shlqbii\t$rT, $rA, $val",
2137 RotateShift, pattern>;
2139 class SHLQBIIVecInst<ValueType vectype>:
2140 SHLQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2141 [(set (vectype VECREG:$rT),
2142 (SPUshlquad_l_bits (vectype VECREG:$rA), (i32 bitshift:$val)))]>;
2144 multiclass ShiftLeftQuadByBitsImm
2146 def v16i8 : SHLQBIIVecInst<v16i8>;
2147 def v8i16 : SHLQBIIVecInst<v8i16>;
2148 def v4i32 : SHLQBIIVecInst<v4i32>;
2149 def v4f32 : SHLQBIIVecInst<v4f32>;
2150 def v2i64 : SHLQBIIVecInst<v2i64>;
2151 def v2f64 : SHLQBIIVecInst<v2f64>;
2154 defm SHLQBII : ShiftLeftQuadByBitsImm;
2156 // SHLQBY, SHLQBYI vector forms: Shift the entire vector to the left by bytes,
2157 // not by bits. See notes above on SHLQBI.
2159 class SHLQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2160 RI7Form<0b11111011100, OOL, IOL, "shlqby\t$rT, $rA, $rB",
2161 RotateShift, pattern>;
2163 class SHLQBYVecInst<ValueType vectype>:
2164 SHLQBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2165 [(set (vectype VECREG:$rT),
2166 (SPUshlquad_l_bytes (vectype VECREG:$rA), R32C:$rB))]>;
2168 multiclass ShiftLeftQuadBytes
2170 def v16i8: SHLQBYVecInst<v16i8>;
2171 def v8i16: SHLQBYVecInst<v8i16>;
2172 def v4i32: SHLQBYVecInst<v4i32>;
2173 def v4f32: SHLQBYVecInst<v4f32>;
2174 def v2i64: SHLQBYVecInst<v2i64>;
2175 def v2f64: SHLQBYVecInst<v2f64>;
2176 def r128: SHLQBYInst<(outs GPRC:$rT), (ins GPRC:$rA, R32C:$rB),
2177 [(set GPRC:$rT, (SPUshlquad_l_bytes GPRC:$rA, R32C:$rB))]>;
2180 defm SHLQBY: ShiftLeftQuadBytes;
2182 class SHLQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2183 RI7Form<0b11111111100, OOL, IOL, "shlqbyi\t$rT, $rA, $val",
2184 RotateShift, pattern>;
2186 class SHLQBYIVecInst<ValueType vectype>:
2187 SHLQBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2188 [(set (vectype VECREG:$rT),
2189 (SPUshlquad_l_bytes (vectype VECREG:$rA), (i32 uimm7:$val)))]>;
2191 multiclass ShiftLeftQuadBytesImm
2193 def v16i8: SHLQBYIVecInst<v16i8>;
2194 def v8i16: SHLQBYIVecInst<v8i16>;
2195 def v4i32: SHLQBYIVecInst<v4i32>;
2196 def v4f32: SHLQBYIVecInst<v4f32>;
2197 def v2i64: SHLQBYIVecInst<v2i64>;
2198 def v2f64: SHLQBYIVecInst<v2f64>;
2199 def r128: SHLQBYIInst<(outs GPRC:$rT), (ins GPRC:$rA, u7imm_i32:$val),
2201 (SPUshlquad_l_bytes GPRC:$rA, (i32 uimm7:$val)))]>;
2204 defm SHLQBYI : ShiftLeftQuadBytesImm;
2206 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2208 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2209 class ROTHInst<dag OOL, dag IOL, list<dag> pattern>:
2210 RRForm<0b00111010000, OOL, IOL, "roth\t$rT, $rA, $rB",
2211 RotateShift, pattern>;
2213 class ROTHVecInst<ValueType vectype>:
2214 ROTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2215 [(set (vectype VECREG:$rT),
2216 (SPUvec_rotl VECREG:$rA, VECREG:$rB))]>;
2218 class ROTHRegInst<RegisterClass rclass>:
2219 ROTHInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2220 [(set rclass:$rT, (rotl rclass:$rA, rclass:$rB))]>;
2222 multiclass RotateLeftHalfword
2224 def v8i16: ROTHVecInst<v8i16>;
2225 def r16: ROTHRegInst<R16C>;
2228 defm ROTH: RotateLeftHalfword;
2230 def ROTHr16_r32: ROTHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2231 [(set R16C:$rT, (rotl R16C:$rA, R32C:$rB))]>;
2233 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2234 // Rotate halfword, immediate:
2235 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2236 class ROTHIInst<dag OOL, dag IOL, list<dag> pattern>:
2237 RI7Form<0b00111110000, OOL, IOL, "rothi\t$rT, $rA, $val",
2238 RotateShift, pattern>;
2240 class ROTHIVecInst<ValueType vectype>:
2241 ROTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2242 [(set (vectype VECREG:$rT),
2243 (SPUvec_rotl VECREG:$rA, (i16 uimm7:$val)))]>;
2245 multiclass RotateLeftHalfwordImm
2247 def v8i16: ROTHIVecInst<v8i16>;
2248 def r16: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2249 [(set R16C:$rT, (rotl R16C:$rA, (i16 uimm7:$val)))]>;
2250 def r16_r32: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm_i32:$val),
2251 [(set R16C:$rT, (rotl R16C:$rA, (i32 uimm7:$val)))]>;
2254 defm ROTHI: RotateLeftHalfwordImm;
2256 def : Pat<(SPUvec_rotl VECREG:$rA, (i32 uimm7:$val)),
2257 (ROTHIv8i16 VECREG:$rA, imm:$val)>;
2259 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2261 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2263 class ROTInst<dag OOL, dag IOL, list<dag> pattern>:
2264 RRForm<0b00011010000, OOL, IOL, "rot\t$rT, $rA, $rB",
2265 RotateShift, pattern>;
2267 class ROTVecInst<ValueType vectype>:
2268 ROTInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2269 [(set (vectype VECREG:$rT),
2270 (SPUvec_rotl (vectype VECREG:$rA), R32C:$rB))]>;
2272 class ROTRegInst<RegisterClass rclass>:
2273 ROTInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2275 (rotl rclass:$rA, R32C:$rB))]>;
2277 multiclass RotateLeftWord
2279 def v4i32: ROTVecInst<v4i32>;
2280 def r32: ROTRegInst<R32C>;
2283 defm ROT: RotateLeftWord;
2285 // The rotate amount is in the same bits whether we've got an 8-bit, 16-bit or
2287 def ROTr32_r16_anyext:
2288 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R16C:$rB),
2289 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R16C:$rB))))]>;
2291 def : Pat<(rotl R32C:$rA, (i32 (zext R16C:$rB))),
2292 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2294 def : Pat<(rotl R32C:$rA, (i32 (sext R16C:$rB))),
2295 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2297 def ROTr32_r8_anyext:
2298 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R8C:$rB),
2299 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R8C:$rB))))]>;
2301 def : Pat<(rotl R32C:$rA, (i32 (zext R8C:$rB))),
2302 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2304 def : Pat<(rotl R32C:$rA, (i32 (sext R8C:$rB))),
2305 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2307 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2308 // Rotate word, immediate
2309 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2311 class ROTIInst<dag OOL, dag IOL, list<dag> pattern>:
2312 RI7Form<0b00011110000, OOL, IOL, "roti\t$rT, $rA, $val",
2313 RotateShift, pattern>;
2315 class ROTIVecInst<ValueType vectype, Operand optype, ValueType inttype, PatLeaf pred>:
2316 ROTIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2317 [(set (vectype VECREG:$rT),
2318 (SPUvec_rotl (vectype VECREG:$rA), (inttype pred:$val)))]>;
2320 class ROTIRegInst<RegisterClass rclass, Operand optype, ValueType inttype, PatLeaf pred>:
2321 ROTIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2322 [(set rclass:$rT, (rotl rclass:$rA, (inttype pred:$val)))]>;
2324 multiclass RotateLeftWordImm
2326 def v4i32: ROTIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2327 def v4i32_i16: ROTIVecInst<v4i32, u7imm, i16, uimm7>;
2328 def v4i32_i8: ROTIVecInst<v4i32, u7imm_i8, i8, uimm7>;
2330 def r32: ROTIRegInst<R32C, u7imm_i32, i32, uimm7>;
2331 def r32_i16: ROTIRegInst<R32C, u7imm, i16, uimm7>;
2332 def r32_i8: ROTIRegInst<R32C, u7imm_i8, i8, uimm7>;
2335 defm ROTI : RotateLeftWordImm;
2337 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2338 // Rotate quad by byte (count)
2339 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2341 class ROTQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2342 RRForm<0b00111011100, OOL, IOL, "rotqby\t$rT, $rA, $rB",
2343 RotateShift, pattern>;
2345 class ROTQBYVecInst<ValueType vectype>:
2346 ROTQBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2347 [(set (vectype VECREG:$rT),
2348 (SPUrotbytes_left (vectype VECREG:$rA), R32C:$rB))]>;
2350 multiclass RotateQuadLeftByBytes
2352 def v16i8: ROTQBYVecInst<v16i8>;
2353 def v8i16: ROTQBYVecInst<v8i16>;
2354 def v4i32: ROTQBYVecInst<v4i32>;
2355 def v4f32: ROTQBYVecInst<v4f32>;
2356 def v2i64: ROTQBYVecInst<v2i64>;
2357 def v2f64: ROTQBYVecInst<v2f64>;
2360 defm ROTQBY: RotateQuadLeftByBytes;
2362 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2363 // Rotate quad by byte (count), immediate
2364 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2366 class ROTQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2367 RI7Form<0b00111111100, OOL, IOL, "rotqbyi\t$rT, $rA, $val",
2368 RotateShift, pattern>;
2370 class ROTQBYIVecInst<ValueType vectype>:
2371 ROTQBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2372 [(set (vectype VECREG:$rT),
2373 (SPUrotbytes_left (vectype VECREG:$rA), (i16 uimm7:$val)))]>;
2375 multiclass RotateQuadByBytesImm
2377 def v16i8: ROTQBYIVecInst<v16i8>;
2378 def v8i16: ROTQBYIVecInst<v8i16>;
2379 def v4i32: ROTQBYIVecInst<v4i32>;
2380 def v4f32: ROTQBYIVecInst<v4f32>;
2381 def v2i64: ROTQBYIVecInst<v2i64>;
2382 def vfi64: ROTQBYIVecInst<v2f64>;
2385 defm ROTQBYI: RotateQuadByBytesImm;
2387 // See ROTQBY note above.
2388 class ROTQBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2389 RI7Form<0b00110011100, OOL, IOL,
2390 "rotqbybi\t$rT, $rA, $shift",
2391 RotateShift, pattern>;
2393 class ROTQBYBIVecInst<ValueType vectype, RegisterClass rclass>:
2394 ROTQBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, rclass:$shift),
2395 [(set (vectype VECREG:$rT),
2396 (SPUrotbytes_left_bits (vectype VECREG:$rA), rclass:$shift))]>;
2398 multiclass RotateQuadByBytesByBitshift {
2399 def v16i8_r32: ROTQBYBIVecInst<v16i8, R32C>;
2400 def v8i16_r32: ROTQBYBIVecInst<v8i16, R32C>;
2401 def v4i32_r32: ROTQBYBIVecInst<v4i32, R32C>;
2402 def v2i64_r32: ROTQBYBIVecInst<v2i64, R32C>;
2405 defm ROTQBYBI : RotateQuadByBytesByBitshift;
2407 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2408 // See ROTQBY note above.
2410 // Assume that the user of this instruction knows to shift the rotate count
2412 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2414 class ROTQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2415 RRForm<0b00011011100, OOL, IOL, "rotqbi\t$rT, $rA, $rB",
2416 RotateShift, pattern>;
2418 class ROTQBIVecInst<ValueType vectype>:
2419 ROTQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2420 [/* no pattern yet */]>;
2422 class ROTQBIRegInst<RegisterClass rclass>:
2423 ROTQBIInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2424 [/* no pattern yet */]>;
2426 multiclass RotateQuadByBitCount
2428 def v16i8: ROTQBIVecInst<v16i8>;
2429 def v8i16: ROTQBIVecInst<v8i16>;
2430 def v4i32: ROTQBIVecInst<v4i32>;
2431 def v2i64: ROTQBIVecInst<v2i64>;
2433 def r128: ROTQBIRegInst<GPRC>;
2434 def r64: ROTQBIRegInst<R64C>;
2437 defm ROTQBI: RotateQuadByBitCount;
2439 class ROTQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2440 RI7Form<0b00011111100, OOL, IOL, "rotqbii\t$rT, $rA, $val",
2441 RotateShift, pattern>;
2443 class ROTQBIIVecInst<ValueType vectype, Operand optype, ValueType inttype,
2445 ROTQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2446 [/* no pattern yet */]>;
2448 class ROTQBIIRegInst<RegisterClass rclass, Operand optype, ValueType inttype,
2450 ROTQBIIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2451 [/* no pattern yet */]>;
2453 multiclass RotateQuadByBitCountImm
2455 def v16i8: ROTQBIIVecInst<v16i8, u7imm_i32, i32, uimm7>;
2456 def v8i16: ROTQBIIVecInst<v8i16, u7imm_i32, i32, uimm7>;
2457 def v4i32: ROTQBIIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2458 def v2i64: ROTQBIIVecInst<v2i64, u7imm_i32, i32, uimm7>;
2460 def r128: ROTQBIIRegInst<GPRC, u7imm_i32, i32, uimm7>;
2461 def r64: ROTQBIIRegInst<R64C, u7imm_i32, i32, uimm7>;
2464 defm ROTQBII : RotateQuadByBitCountImm;
2466 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2467 // ROTHM v8i16 form:
2468 // NOTE(1): No vector rotate is generated by the C/C++ frontend (today),
2469 // so this only matches a synthetically generated/lowered code
2471 // NOTE(2): $rB must be negated before the right rotate!
2472 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2474 class ROTHMInst<dag OOL, dag IOL, list<dag> pattern>:
2475 RRForm<0b10111010000, OOL, IOL, "rothm\t$rT, $rA, $rB",
2476 RotateShift, pattern>;
2479 ROTHMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2480 [/* see patterns below - $rB must be negated */]>;
2482 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R32C:$rB),
2483 (ROTHMv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2485 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R16C:$rB),
2486 (ROTHMv8i16 VECREG:$rA,
2487 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2489 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R8C:$rB),
2490 (ROTHMv8i16 VECREG:$rA,
2491 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2493 // ROTHM r16 form: Rotate 16-bit quantity to right, zero fill at the left
2494 // Note: This instruction doesn't match a pattern because rB must be negated
2495 // for the instruction to work. Thus, the pattern below the instruction!
2498 ROTHMInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2499 [/* see patterns below - $rB must be negated! */]>;
2501 def : Pat<(srl R16C:$rA, R32C:$rB),
2502 (ROTHMr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2504 def : Pat<(srl R16C:$rA, R16C:$rB),
2506 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2508 def : Pat<(srl R16C:$rA, R8C:$rB),
2510 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2512 // ROTHMI v8i16 form: See the comment for ROTHM v8i16. The difference here is
2513 // that the immediate can be complemented, so that the user doesn't have to
2516 class ROTHMIInst<dag OOL, dag IOL, list<dag> pattern>:
2517 RI7Form<0b10111110000, OOL, IOL, "rothmi\t$rT, $rA, $val",
2518 RotateShift, pattern>;
2521 ROTHMIInst<(outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2522 [/* no pattern */]>;
2524 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i32 imm:$val)),
2525 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2527 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i16 imm:$val)),
2528 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2530 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i8 imm:$val)),
2531 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2534 ROTHMIInst<(outs R16C:$rT), (ins R16C:$rA, rothNeg7imm:$val),
2535 [/* no pattern */]>;
2537 def: Pat<(srl R16C:$rA, (i32 uimm7:$val)),
2538 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2540 def: Pat<(srl R16C:$rA, (i16 uimm7:$val)),
2541 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2543 def: Pat<(srl R16C:$rA, (i8 uimm7:$val)),
2544 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2546 // ROTM v4i32 form: See the ROTHM v8i16 comments.
2547 class ROTMInst<dag OOL, dag IOL, list<dag> pattern>:
2548 RRForm<0b10011010000, OOL, IOL, "rotm\t$rT, $rA, $rB",
2549 RotateShift, pattern>;
2552 ROTMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2553 [/* see patterns below - $rB must be negated */]>;
2555 def : Pat<(SPUvec_srl VECREG:$rA, R32C:$rB),
2556 (ROTMv4i32 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2558 def : Pat<(SPUvec_srl VECREG:$rA, R16C:$rB),
2559 (ROTMv4i32 VECREG:$rA,
2560 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2562 def : Pat<(SPUvec_srl VECREG:$rA, R8C:$rB),
2563 (ROTMv4i32 VECREG:$rA,
2564 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2567 ROTMInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2568 [/* see patterns below - $rB must be negated */]>;
2570 def : Pat<(srl R32C:$rA, R32C:$rB),
2571 (ROTMr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
2573 def : Pat<(srl R32C:$rA, R16C:$rB),
2575 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2577 def : Pat<(srl R32C:$rA, R8C:$rB),
2579 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2581 // ROTMI v4i32 form: See the comment for ROTHM v8i16.
2583 RI7Form<0b10011110000, (outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2584 "rotmi\t$rT, $rA, $val", RotateShift,
2585 [(set (v4i32 VECREG:$rT),
2586 (SPUvec_srl VECREG:$rA, (i32 uimm7:$val)))]>;
2588 def : Pat<(SPUvec_srl VECREG:$rA, (i16 uimm7:$val)),
2589 (ROTMIv4i32 VECREG:$rA, uimm7:$val)>;
2591 def : Pat<(SPUvec_srl VECREG:$rA, (i8 uimm7:$val)),
2592 (ROTMIv4i32 VECREG:$rA, uimm7:$val)>;
2594 // ROTMI r32 form: know how to complement the immediate value.
2596 RI7Form<0b10011110000, (outs R32C:$rT), (ins R32C:$rA, rotNeg7imm:$val),
2597 "rotmi\t$rT, $rA, $val", RotateShift,
2598 [(set R32C:$rT, (srl R32C:$rA, (i32 uimm7:$val)))]>;
2600 def : Pat<(srl R32C:$rA, (i16 imm:$val)),
2601 (ROTMIr32 R32C:$rA, uimm7:$val)>;
2603 def : Pat<(srl R32C:$rA, (i8 imm:$val)),
2604 (ROTMIr32 R32C:$rA, uimm7:$val)>;
2606 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2607 // ROTQMBYvec: This is a vector form merely so that when used in an
2608 // instruction pattern, type checking will succeed. This instruction assumes
2609 // that the user knew to negate $rB.
2611 // Using the SPUrotquad_rz_bytes target-specific DAG node, the patterns
2612 // ensure that $rB is negated.
2613 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2615 class ROTQMBYInst<dag OOL, dag IOL, list<dag> pattern>:
2616 RRForm<0b10111011100, OOL, IOL, "rotqmby\t$rT, $rA, $rB",
2617 RotateShift, pattern>;
2619 class ROTQMBYVecInst<ValueType vectype>:
2620 ROTQMBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2621 [/* no pattern, $rB must be negated */]>;
2623 class ROTQMBYRegInst<RegisterClass rclass>:
2624 ROTQMBYInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2626 (SPUrotquad_rz_bytes rclass:$rA, R32C:$rB))]>;
2628 multiclass RotateQuadBytes
2630 def v16i8: ROTQMBYVecInst<v16i8>;
2631 def v8i16: ROTQMBYVecInst<v8i16>;
2632 def v4i32: ROTQMBYVecInst<v4i32>;
2633 def v2i64: ROTQMBYVecInst<v2i64>;
2635 def r128: ROTQMBYRegInst<GPRC>;
2636 def r64: ROTQMBYRegInst<R64C>;
2639 defm ROTQMBY : RotateQuadBytes;
2641 def : Pat<(SPUrotquad_rz_bytes (v16i8 VECREG:$rA), R32C:$rB),
2642 (ROTQMBYv16i8 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2643 def : Pat<(SPUrotquad_rz_bytes (v8i16 VECREG:$rA), R32C:$rB),
2644 (ROTQMBYv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2645 def : Pat<(SPUrotquad_rz_bytes (v4i32 VECREG:$rA), R32C:$rB),
2646 (ROTQMBYv4i32 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2647 def : Pat<(SPUrotquad_rz_bytes (v2i64 VECREG:$rA), R32C:$rB),
2648 (ROTQMBYv2i64 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2649 def : Pat<(SPUrotquad_rz_bytes GPRC:$rA, R32C:$rB),
2650 (ROTQMBYr128 GPRC:$rA, (SFIr32 R32C:$rB, 0))>;
2651 def : Pat<(SPUrotquad_rz_bytes R64C:$rA, R32C:$rB),
2652 (ROTQMBYr64 R64C:$rA, (SFIr32 R32C:$rB, 0))>;
2654 class ROTQMBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2655 RI7Form<0b10111111100, OOL, IOL, "rotqmbyi\t$rT, $rA, $val",
2656 RotateShift, pattern>;
2658 class ROTQMBYIVecInst<ValueType vectype>:
2659 ROTQMBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2660 [(set (vectype VECREG:$rT),
2661 (SPUrotquad_rz_bytes (vectype VECREG:$rA), (i32 uimm7:$val)))]>;
2663 class ROTQMBYIRegInst<RegisterClass rclass, Operand optype, ValueType inttype, PatLeaf pred>:
2664 ROTQMBYIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2666 (SPUrotquad_rz_bytes rclass:$rA, (inttype pred:$val)))]>;
2668 multiclass RotateQuadBytesImm
2670 def v16i8: ROTQMBYIVecInst<v16i8>;
2671 def v8i16: ROTQMBYIVecInst<v8i16>;
2672 def v4i32: ROTQMBYIVecInst<v4i32>;
2673 def v2i64: ROTQMBYIVecInst<v2i64>;
2675 def r128: ROTQMBYIRegInst<GPRC, rotNeg7imm, i32, uimm7>;
2676 def r64: ROTQMBYIRegInst<R64C, rotNeg7imm, i32, uimm7>;
2679 defm ROTQMBYI : RotateQuadBytesImm;
2681 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2682 // Rotate right and mask by bit count
2683 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2685 class ROTQMBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2686 RRForm<0b10110011100, OOL, IOL, "rotqmbybi\t$rT, $rA, $rB",
2687 RotateShift, pattern>;
2689 class ROTQMBYBIVecInst<ValueType vectype>:
2690 ROTQMBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2691 [/* no pattern, intrinsic? */]>;
2693 multiclass RotateMaskQuadByBitCount
2695 def v16i8: ROTQMBYBIVecInst<v16i8>;
2696 def v8i16: ROTQMBYBIVecInst<v8i16>;
2697 def v4i32: ROTQMBYBIVecInst<v4i32>;
2698 def v2i64: ROTQMBYBIVecInst<v2i64>;
2701 defm ROTQMBYBI: RotateMaskQuadByBitCount;
2703 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2704 // Rotate quad and mask by bits
2705 // Note that the rotate amount has to be negated
2706 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2708 class ROTQMBIInst<dag OOL, dag IOL, list<dag> pattern>:
2709 RRForm<0b10011011100, OOL, IOL, "rotqmbi\t$rT, $rA, $rB",
2710 RotateShift, pattern>;
2712 class ROTQMBIVecInst<ValueType vectype>:
2713 ROTQMBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2714 [/* no pattern */]>;
2716 class ROTQMBIRegInst<RegisterClass rclass>:
2717 ROTQMBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2718 [/* no pattern */]>;
2720 multiclass RotateMaskQuadByBits
2722 def v16i8: ROTQMBIVecInst<v16i8>;
2723 def v8i16: ROTQMBIVecInst<v8i16>;
2724 def v4i32: ROTQMBIVecInst<v4i32>;
2725 def v2i64: ROTQMBIVecInst<v2i64>;
2727 def r128: ROTQMBIRegInst<GPRC>;
2728 def r64: ROTQMBIRegInst<R64C>;
2731 defm ROTQMBI: RotateMaskQuadByBits;
2733 def : Pat<(SPUrotquad_rz_bits (v16i8 VECREG:$rA), R32C:$rB),
2734 (ROTQMBIv16i8 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2735 def : Pat<(SPUrotquad_rz_bits (v8i16 VECREG:$rA), R32C:$rB),
2736 (ROTQMBIv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2737 def : Pat<(SPUrotquad_rz_bits (v4i32 VECREG:$rA), R32C:$rB),
2738 (ROTQMBIv4i32 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2739 def : Pat<(SPUrotquad_rz_bits (v2i64 VECREG:$rA), R32C:$rB),
2740 (ROTQMBIv2i64 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2741 def : Pat<(SPUrotquad_rz_bits GPRC:$rA, R32C:$rB),
2742 (ROTQMBIr128 GPRC:$rA, (SFIr32 R32C:$rB, 0))>;
2743 def : Pat<(SPUrotquad_rz_bits R64C:$rA, R32C:$rB),
2744 (ROTQMBIr64 R64C:$rA, (SFIr32 R32C:$rB, 0))>;
2746 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2747 // Rotate quad and mask by bits, immediate
2748 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2750 class ROTQMBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2751 RI7Form<0b10011111100, OOL, IOL, "rotqmbii\t$rT, $rA, $val",
2752 RotateShift, pattern>;
2754 class ROTQMBIIVecInst<ValueType vectype>:
2755 ROTQMBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2756 [(set (vectype VECREG:$rT),
2757 (SPUrotquad_rz_bits (vectype VECREG:$rA), (i32 uimm7:$val)))]>;
2759 class ROTQMBIIRegInst<RegisterClass rclass>:
2760 ROTQMBIIInst<(outs rclass:$rT), (ins rclass:$rA, rotNeg7imm:$val),
2762 (SPUrotquad_rz_bits rclass:$rA, (i32 uimm7:$val)))]>;
2764 multiclass RotateMaskQuadByBitsImm
2766 def v16i8: ROTQMBIIVecInst<v16i8>;
2767 def v8i16: ROTQMBIIVecInst<v8i16>;
2768 def v4i32: ROTQMBIIVecInst<v4i32>;
2769 def v2i64: ROTQMBIIVecInst<v2i64>;
2771 def r128: ROTQMBIIRegInst<GPRC>;
2772 def r64: ROTQMBIIRegInst<R64C>;
2775 defm ROTQMBII: RotateMaskQuadByBitsImm;
2777 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2778 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2781 RRForm<0b01111010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2782 "rotmah\t$rT, $rA, $rB", RotateShift,
2783 [/* see patterns below - $rB must be negated */]>;
2785 def : Pat<(SPUvec_sra VECREG:$rA, R32C:$rB),
2786 (ROTMAHv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2788 def : Pat<(SPUvec_sra VECREG:$rA, R16C:$rB),
2789 (ROTMAHv8i16 VECREG:$rA,
2790 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2792 def : Pat<(SPUvec_sra VECREG:$rA, R8C:$rB),
2793 (ROTMAHv8i16 VECREG:$rA,
2794 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2797 RRForm<0b01111010000, (outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2798 "rotmah\t$rT, $rA, $rB", RotateShift,
2799 [/* see patterns below - $rB must be negated */]>;
2801 def : Pat<(sra R16C:$rA, R32C:$rB),
2802 (ROTMAHr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2804 def : Pat<(sra R16C:$rA, R16C:$rB),
2805 (ROTMAHr16 R16C:$rA,
2806 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2808 def : Pat<(sra R16C:$rA, R8C:$rB),
2809 (ROTMAHr16 R16C:$rA,
2810 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2813 RRForm<0b01111110000, (outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2814 "rotmahi\t$rT, $rA, $val", RotateShift,
2815 [(set (v8i16 VECREG:$rT),
2816 (SPUvec_sra (v8i16 VECREG:$rA), (i32 uimm7:$val)))]>;
2818 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i16 uimm7:$val)),
2819 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (i32 uimm7:$val))>;
2821 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i8 uimm7:$val)),
2822 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (i32 uimm7:$val))>;
2825 RRForm<0b01111110000, (outs R16C:$rT), (ins R16C:$rA, rothNeg7imm_i16:$val),
2826 "rotmahi\t$rT, $rA, $val", RotateShift,
2827 [(set R16C:$rT, (sra R16C:$rA, (i16 uimm7:$val)))]>;
2829 def : Pat<(sra R16C:$rA, (i32 imm:$val)),
2830 (ROTMAHIr16 R16C:$rA, uimm7:$val)>;
2832 def : Pat<(sra R16C:$rA, (i8 imm:$val)),
2833 (ROTMAHIr16 R16C:$rA, uimm7:$val)>;
2836 RRForm<0b01011010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2837 "rotma\t$rT, $rA, $rB", RotateShift,
2838 [/* see patterns below - $rB must be negated */]>;
2840 def : Pat<(SPUvec_sra VECREG:$rA, R32C:$rB),
2841 (ROTMAv4i32 (v4i32 VECREG:$rA), (SFIr32 R32C:$rB, 0))>;
2843 def : Pat<(SPUvec_sra VECREG:$rA, R16C:$rB),
2844 (ROTMAv4i32 (v4i32 VECREG:$rA),
2845 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2847 def : Pat<(SPUvec_sra VECREG:$rA, R8C:$rB),
2848 (ROTMAv4i32 (v4i32 VECREG:$rA),
2849 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2852 RRForm<0b01011010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2853 "rotma\t$rT, $rA, $rB", RotateShift,
2854 [/* see patterns below - $rB must be negated */]>;
2856 def : Pat<(sra R32C:$rA, R32C:$rB),
2857 (ROTMAr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
2859 def : Pat<(sra R32C:$rA, R16C:$rB),
2861 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2863 def : Pat<(sra R32C:$rA, R8C:$rB),
2865 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2867 class ROTMAIInst<dag OOL, dag IOL, list<dag> pattern>:
2868 RRForm<0b01011110000, OOL, IOL,
2869 "rotmai\t$rT, $rA, $val",
2870 RotateShift, pattern>;
2872 class ROTMAIVecInst<ValueType vectype, Operand intop, ValueType inttype>:
2873 ROTMAIInst<(outs VECREG:$rT), (ins VECREG:$rA, intop:$val),
2874 [(set (vectype VECREG:$rT),
2875 (SPUvec_sra VECREG:$rA, (inttype uimm7:$val)))]>;
2877 class ROTMAIRegInst<RegisterClass rclass, Operand intop, ValueType inttype>:
2878 ROTMAIInst<(outs rclass:$rT), (ins rclass:$rA, intop:$val),
2879 [(set rclass:$rT, (sra rclass:$rA, (inttype uimm7:$val)))]>;
2881 multiclass RotateMaskAlgebraicImm {
2882 def v2i64_i32 : ROTMAIVecInst<v2i64, rotNeg7imm, i32>;
2883 def v4i32_i32 : ROTMAIVecInst<v4i32, rotNeg7imm, i32>;
2884 def r64_i32 : ROTMAIRegInst<R64C, rotNeg7imm, i32>;
2885 def r32_i32 : ROTMAIRegInst<R32C, rotNeg7imm, i32>;
2888 defm ROTMAI : RotateMaskAlgebraicImm;
2890 //===----------------------------------------------------------------------===//
2891 // Branch and conditionals:
2892 //===----------------------------------------------------------------------===//
2894 let isTerminator = 1, isBarrier = 1 in {
2895 // Halt If Equal (r32 preferred slot only, no vector form)
2897 RRForm_3<0b00011011110, (outs), (ins R32C:$rA, R32C:$rB),
2898 "heq\t$rA, $rB", BranchResolv,
2899 [/* no pattern to match */]>;
2902 RI10Form_2<0b11111110, (outs), (ins R32C:$rA, s10imm:$val),
2903 "heqi\t$rA, $val", BranchResolv,
2904 [/* no pattern to match */]>;
2906 // HGT/HGTI: These instructions use signed arithmetic for the comparison,
2907 // contrasting with HLGT/HLGTI, which use unsigned comparison:
2909 RRForm_3<0b00011010010, (outs), (ins R32C:$rA, R32C:$rB),
2910 "hgt\t$rA, $rB", BranchResolv,
2911 [/* no pattern to match */]>;
2914 RI10Form_2<0b11110010, (outs), (ins R32C:$rA, s10imm:$val),
2915 "hgti\t$rA, $val", BranchResolv,
2916 [/* no pattern to match */]>;
2919 RRForm_3<0b00011011010, (outs), (ins R32C:$rA, R32C:$rB),
2920 "hlgt\t$rA, $rB", BranchResolv,
2921 [/* no pattern to match */]>;
2924 RI10Form_2<0b11111010, (outs), (ins R32C:$rA, s10imm:$val),
2925 "hlgti\t$rA, $val", BranchResolv,
2926 [/* no pattern to match */]>;
2929 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2930 // Comparison operators for i8, i16 and i32:
2931 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2933 class CEQBInst<dag OOL, dag IOL, list<dag> pattern> :
2934 RRForm<0b00001011110, OOL, IOL, "ceqb\t$rT, $rA, $rB",
2937 multiclass CmpEqualByte
2940 CEQBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2941 [(set (v16i8 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
2942 (v8i16 VECREG:$rB)))]>;
2945 CEQBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
2946 [(set R8C:$rT, (seteq R8C:$rA, R8C:$rB))]>;
2949 class CEQBIInst<dag OOL, dag IOL, list<dag> pattern> :
2950 RI10Form<0b01111110, OOL, IOL, "ceqbi\t$rT, $rA, $val",
2953 multiclass CmpEqualByteImm
2956 CEQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
2957 [(set (v16i8 VECREG:$rT), (seteq (v16i8 VECREG:$rA),
2958 v16i8SExt8Imm:$val))]>;
2960 CEQBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
2961 [(set R8C:$rT, (seteq R8C:$rA, immSExt8:$val))]>;
2964 class CEQHInst<dag OOL, dag IOL, list<dag> pattern> :
2965 RRForm<0b00010011110, OOL, IOL, "ceqh\t$rT, $rA, $rB",
2968 multiclass CmpEqualHalfword
2970 def v8i16 : CEQHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2971 [(set (v8i16 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
2972 (v8i16 VECREG:$rB)))]>;
2974 def r16 : CEQHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
2975 [(set R16C:$rT, (seteq R16C:$rA, R16C:$rB))]>;
2978 class CEQHIInst<dag OOL, dag IOL, list<dag> pattern> :
2979 RI10Form<0b10111110, OOL, IOL, "ceqhi\t$rT, $rA, $val",
2982 multiclass CmpEqualHalfwordImm
2984 def v8i16 : CEQHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
2985 [(set (v8i16 VECREG:$rT),
2986 (seteq (v8i16 VECREG:$rA),
2987 (v8i16 v8i16SExt10Imm:$val)))]>;
2988 def r16 : CEQHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
2989 [(set R16C:$rT, (seteq R16C:$rA, i16ImmSExt10:$val))]>;
2992 class CEQInst<dag OOL, dag IOL, list<dag> pattern> :
2993 RRForm<0b00000011110, OOL, IOL, "ceq\t$rT, $rA, $rB",
2996 multiclass CmpEqualWord
2998 def v4i32 : CEQInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2999 [(set (v4i32 VECREG:$rT),
3000 (seteq (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3002 def r32 : CEQInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3003 [(set R32C:$rT, (seteq R32C:$rA, R32C:$rB))]>;
3006 class CEQIInst<dag OOL, dag IOL, list<dag> pattern> :
3007 RI10Form<0b00111110, OOL, IOL, "ceqi\t$rT, $rA, $val",
3010 multiclass CmpEqualWordImm
3012 def v4i32 : CEQIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3013 [(set (v4i32 VECREG:$rT),
3014 (seteq (v4i32 VECREG:$rA),
3015 (v4i32 v4i32SExt16Imm:$val)))]>;
3017 def r32: CEQIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3018 [(set R32C:$rT, (seteq R32C:$rA, i32ImmSExt10:$val))]>;
3021 class CGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3022 RRForm<0b00001010010, OOL, IOL, "cgtb\t$rT, $rA, $rB",
3025 multiclass CmpGtrByte
3028 CGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3029 [(set (v16i8 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3030 (v8i16 VECREG:$rB)))]>;
3033 CGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3034 [(set R8C:$rT, (setgt R8C:$rA, R8C:$rB))]>;
3037 class CGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3038 RI10Form<0b01110010, OOL, IOL, "cgtbi\t$rT, $rA, $val",
3041 multiclass CmpGtrByteImm
3044 CGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3045 [(set (v16i8 VECREG:$rT), (setgt (v16i8 VECREG:$rA),
3046 v16i8SExt8Imm:$val))]>;
3048 CGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3049 [(set R8C:$rT, (setgt R8C:$rA, immSExt8:$val))]>;
3052 class CGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3053 RRForm<0b00010010010, OOL, IOL, "cgth\t$rT, $rA, $rB",
3056 multiclass CmpGtrHalfword
3058 def v8i16 : CGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3059 [(set (v8i16 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3060 (v8i16 VECREG:$rB)))]>;
3062 def r16 : CGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3063 [(set R16C:$rT, (setgt R16C:$rA, R16C:$rB))]>;
3066 class CGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3067 RI10Form<0b10110010, OOL, IOL, "cgthi\t$rT, $rA, $val",
3070 multiclass CmpGtrHalfwordImm
3072 def v8i16 : CGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3073 [(set (v8i16 VECREG:$rT),
3074 (setgt (v8i16 VECREG:$rA),
3075 (v8i16 v8i16SExt10Imm:$val)))]>;
3076 def r16 : CGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3077 [(set R16C:$rT, (setgt R16C:$rA, i16ImmSExt10:$val))]>;
3080 class CGTInst<dag OOL, dag IOL, list<dag> pattern> :
3081 RRForm<0b00000010010, OOL, IOL, "cgt\t$rT, $rA, $rB",
3084 multiclass CmpGtrWord
3086 def v4i32 : CGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3087 [(set (v4i32 VECREG:$rT),
3088 (setgt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3090 def r32 : CGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3091 [(set R32C:$rT, (setgt R32C:$rA, R32C:$rB))]>;
3094 class CGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3095 RI10Form<0b00110010, OOL, IOL, "cgti\t$rT, $rA, $val",
3098 multiclass CmpGtrWordImm
3100 def v4i32 : CGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3101 [(set (v4i32 VECREG:$rT),
3102 (setgt (v4i32 VECREG:$rA),
3103 (v4i32 v4i32SExt16Imm:$val)))]>;
3105 def r32: CGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3106 [(set R32C:$rT, (setgt R32C:$rA, i32ImmSExt10:$val))]>;
3109 class CLGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3110 RRForm<0b00001011010, OOL, IOL, "clgtb\t$rT, $rA, $rB",
3113 multiclass CmpLGtrByte
3116 CLGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3117 [(set (v16i8 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3118 (v8i16 VECREG:$rB)))]>;
3121 CLGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3122 [(set R8C:$rT, (setugt R8C:$rA, R8C:$rB))]>;
3125 class CLGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3126 RI10Form<0b01111010, OOL, IOL, "clgtbi\t$rT, $rA, $val",
3129 multiclass CmpLGtrByteImm
3132 CLGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3133 [(set (v16i8 VECREG:$rT), (setugt (v16i8 VECREG:$rA),
3134 v16i8SExt8Imm:$val))]>;
3136 CLGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3137 [(set R8C:$rT, (setugt R8C:$rA, immSExt8:$val))]>;
3140 class CLGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3141 RRForm<0b00010011010, OOL, IOL, "clgth\t$rT, $rA, $rB",
3144 multiclass CmpLGtrHalfword
3146 def v8i16 : CLGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3147 [(set (v8i16 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3148 (v8i16 VECREG:$rB)))]>;
3150 def r16 : CLGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3151 [(set R16C:$rT, (setugt R16C:$rA, R16C:$rB))]>;
3154 class CLGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3155 RI10Form<0b10111010, OOL, IOL, "clgthi\t$rT, $rA, $val",
3158 multiclass CmpLGtrHalfwordImm
3160 def v8i16 : CLGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3161 [(set (v8i16 VECREG:$rT),
3162 (setugt (v8i16 VECREG:$rA),
3163 (v8i16 v8i16SExt10Imm:$val)))]>;
3164 def r16 : CLGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3165 [(set R16C:$rT, (setugt R16C:$rA, i16ImmSExt10:$val))]>;
3168 class CLGTInst<dag OOL, dag IOL, list<dag> pattern> :
3169 RRForm<0b00000011010, OOL, IOL, "clgt\t$rT, $rA, $rB",
3172 multiclass CmpLGtrWord
3174 def v4i32 : CLGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3175 [(set (v4i32 VECREG:$rT),
3176 (setugt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3178 def r32 : CLGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3179 [(set R32C:$rT, (setugt R32C:$rA, R32C:$rB))]>;
3182 class CLGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3183 RI10Form<0b00111010, OOL, IOL, "clgti\t$rT, $rA, $val",
3186 multiclass CmpLGtrWordImm
3188 def v4i32 : CLGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3189 [(set (v4i32 VECREG:$rT),
3190 (setugt (v4i32 VECREG:$rA),
3191 (v4i32 v4i32SExt16Imm:$val)))]>;
3193 def r32: CLGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3194 [(set R32C:$rT, (setugt R32C:$rA, i32ImmSExt10:$val))]>;
3197 defm CEQB : CmpEqualByte;
3198 defm CEQBI : CmpEqualByteImm;
3199 defm CEQH : CmpEqualHalfword;
3200 defm CEQHI : CmpEqualHalfwordImm;
3201 defm CEQ : CmpEqualWord;
3202 defm CEQI : CmpEqualWordImm;
3203 defm CGTB : CmpGtrByte;
3204 defm CGTBI : CmpGtrByteImm;
3205 defm CGTH : CmpGtrHalfword;
3206 defm CGTHI : CmpGtrHalfwordImm;
3207 defm CGT : CmpGtrWord;
3208 defm CGTI : CmpGtrWordImm;
3209 defm CLGTB : CmpLGtrByte;
3210 defm CLGTBI : CmpLGtrByteImm;
3211 defm CLGTH : CmpLGtrHalfword;
3212 defm CLGTHI : CmpLGtrHalfwordImm;
3213 defm CLGT : CmpLGtrWord;
3214 defm CLGTI : CmpLGtrWordImm;
3216 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3217 // For SETCC primitives not supported above (setlt, setle, setge, etc.)
3218 // define a pattern to generate the right code, as a binary operator
3219 // (in a manner of speaking.)
3222 // 1. This only matches the setcc set of conditionals. Special pattern
3223 // matching is used for select conditionals.
3225 // 2. The "DAG" versions of these classes is almost exclusively used for
3226 // i64 comparisons. See the tblgen fundamentals documentation for what
3227 // ".ResultInstrs[0]" means; see TargetSelectionDAG.td and the Pattern
3228 // class for where ResultInstrs originates.
3229 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3231 class SETCCNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3232 SPUInstr xorinst, SPUInstr cmpare>:
3233 Pat<(cond rclass:$rA, rclass:$rB),
3234 (xorinst (cmpare rclass:$rA, rclass:$rB), (inttype -1))>;
3236 class SETCCNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3237 PatLeaf immpred, SPUInstr xorinst, SPUInstr cmpare>:
3238 Pat<(cond rclass:$rA, (inttype immpred:$imm)),
3239 (xorinst (cmpare rclass:$rA, (inttype immpred:$imm)), (inttype -1))>;
3241 def : SETCCNegCondReg<setne, R8C, i8, XORBIr8, CEQBr8>;
3242 def : SETCCNegCondImm<setne, R8C, i8, immSExt8, XORBIr8, CEQBIr8>;
3244 def : SETCCNegCondReg<setne, R16C, i16, XORHIr16, CEQHr16>;
3245 def : SETCCNegCondImm<setne, R16C, i16, i16ImmSExt10, XORHIr16, CEQHIr16>;
3247 def : SETCCNegCondReg<setne, R32C, i32, XORIr32, CEQr32>;
3248 def : SETCCNegCondImm<setne, R32C, i32, i32ImmSExt10, XORIr32, CEQIr32>;
3250 class SETCCBinOpReg<PatFrag cond, RegisterClass rclass,
3251 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3252 Pat<(cond rclass:$rA, rclass:$rB),
3253 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3254 (cmpOp2 rclass:$rA, rclass:$rB))>;
3256 class SETCCBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3258 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3259 Pat<(cond rclass:$rA, (immtype immpred:$imm)),
3260 (binop (cmpOp1 rclass:$rA, (immtype immpred:$imm)),
3261 (cmpOp2 rclass:$rA, (immtype immpred:$imm)))>;
3263 def : SETCCBinOpReg<setge, R8C, ORr8, CGTBr8, CEQBr8>;
3264 def : SETCCBinOpImm<setge, R8C, immSExt8, i8, ORr8, CGTBIr8, CEQBIr8>;
3265 def : SETCCBinOpReg<setlt, R8C, NORr8, CGTBr8, CEQBr8>;
3266 def : SETCCBinOpImm<setlt, R8C, immSExt8, i8, NORr8, CGTBIr8, CEQBIr8>;
3267 def : Pat<(setle R8C:$rA, R8C:$rB),
3268 (XORBIr8 (CGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3269 def : Pat<(setle R8C:$rA, immU8:$imm),
3270 (XORBIr8 (CGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3272 def : SETCCBinOpReg<setge, R16C, ORr16, CGTHr16, CEQHr16>;
3273 def : SETCCBinOpImm<setge, R16C, i16ImmSExt10, i16,
3274 ORr16, CGTHIr16, CEQHIr16>;
3275 def : SETCCBinOpReg<setlt, R16C, NORr16, CGTHr16, CEQHr16>;
3276 def : SETCCBinOpImm<setlt, R16C, i16ImmSExt10, i16, NORr16, CGTHIr16, CEQHIr16>;
3277 def : Pat<(setle R16C:$rA, R16C:$rB),
3278 (XORHIr16 (CGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3279 def : Pat<(setle R16C:$rA, i16ImmSExt10:$imm),
3280 (XORHIr16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3282 def : SETCCBinOpReg<setge, R32C, ORr32, CGTr32, CEQr32>;
3283 def : SETCCBinOpImm<setge, R32C, i32ImmSExt10, i32,
3284 ORr32, CGTIr32, CEQIr32>;
3285 def : SETCCBinOpReg<setlt, R32C, NORr32, CGTr32, CEQr32>;
3286 def : SETCCBinOpImm<setlt, R32C, i32ImmSExt10, i32, NORr32, CGTIr32, CEQIr32>;
3287 def : Pat<(setle R32C:$rA, R32C:$rB),
3288 (XORIr32 (CGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3289 def : Pat<(setle R32C:$rA, i32ImmSExt10:$imm),
3290 (XORIr32 (CGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3292 def : SETCCBinOpReg<setuge, R8C, ORr8, CLGTBr8, CEQBr8>;
3293 def : SETCCBinOpImm<setuge, R8C, immSExt8, i8, ORr8, CLGTBIr8, CEQBIr8>;
3294 def : SETCCBinOpReg<setult, R8C, NORr8, CLGTBr8, CEQBr8>;
3295 def : SETCCBinOpImm<setult, R8C, immSExt8, i8, NORr8, CLGTBIr8, CEQBIr8>;
3296 def : Pat<(setule R8C:$rA, R8C:$rB),
3297 (XORBIr8 (CLGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3298 def : Pat<(setule R8C:$rA, immU8:$imm),
3299 (XORBIr8 (CLGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3301 def : SETCCBinOpReg<setuge, R16C, ORr16, CLGTHr16, CEQHr16>;
3302 def : SETCCBinOpImm<setuge, R16C, i16ImmSExt10, i16,
3303 ORr16, CLGTHIr16, CEQHIr16>;
3304 def : SETCCBinOpReg<setult, R16C, NORr16, CLGTHr16, CEQHr16>;
3305 def : SETCCBinOpImm<setult, R16C, i16ImmSExt10, i16, NORr16,
3306 CLGTHIr16, CEQHIr16>;
3307 def : Pat<(setule R16C:$rA, R16C:$rB),
3308 (XORHIr16 (CLGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3309 def : Pat<(setule R16C:$rA, i16ImmSExt10:$imm),
3310 (XORHIr16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3312 def : SETCCBinOpReg<setuge, R32C, ORr32, CLGTr32, CEQr32>;
3313 def : SETCCBinOpImm<setuge, R32C, i32ImmSExt10, i32,
3314 ORr32, CLGTIr32, CEQIr32>;
3315 def : SETCCBinOpReg<setult, R32C, NORr32, CLGTr32, CEQr32>;
3316 def : SETCCBinOpImm<setult, R32C, i32ImmSExt10, i32, NORr32, CLGTIr32, CEQIr32>;
3317 def : Pat<(setule R32C:$rA, R32C:$rB),
3318 (XORIr32 (CLGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3319 def : Pat<(setule R32C:$rA, i32ImmSExt10:$imm),
3320 (XORIr32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3322 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3323 // select conditional patterns:
3324 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3326 class SELECTNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3327 SPUInstr selinstr, SPUInstr cmpare>:
3328 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3329 rclass:$rTrue, rclass:$rFalse),
3330 (selinstr rclass:$rTrue, rclass:$rFalse,
3331 (cmpare rclass:$rA, rclass:$rB))>;
3333 class SELECTNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3334 PatLeaf immpred, SPUInstr selinstr, SPUInstr cmpare>:
3335 Pat<(select (inttype (cond rclass:$rA, immpred:$imm)),
3336 rclass:$rTrue, rclass:$rFalse),
3337 (selinstr rclass:$rTrue, rclass:$rFalse,
3338 (cmpare rclass:$rA, immpred:$imm))>;
3340 def : SELECTNegCondReg<setne, R8C, i8, SELBr8, CEQBr8>;
3341 def : SELECTNegCondImm<setne, R8C, i8, immSExt8, SELBr8, CEQBIr8>;
3342 def : SELECTNegCondReg<setle, R8C, i8, SELBr8, CGTBr8>;
3343 def : SELECTNegCondImm<setle, R8C, i8, immSExt8, SELBr8, CGTBr8>;
3344 def : SELECTNegCondReg<setule, R8C, i8, SELBr8, CLGTBr8>;
3345 def : SELECTNegCondImm<setule, R8C, i8, immU8, SELBr8, CLGTBIr8>;
3347 def : SELECTNegCondReg<setne, R16C, i16, SELBr16, CEQHr16>;
3348 def : SELECTNegCondImm<setne, R16C, i16, i16ImmSExt10, SELBr16, CEQHIr16>;
3349 def : SELECTNegCondReg<setle, R16C, i16, SELBr16, CGTHr16>;
3350 def : SELECTNegCondImm<setle, R16C, i16, i16ImmSExt10, SELBr16, CGTHIr16>;
3351 def : SELECTNegCondReg<setule, R16C, i16, SELBr16, CLGTHr16>;
3352 def : SELECTNegCondImm<setule, R16C, i16, i16ImmSExt10, SELBr16, CLGTHIr16>;
3354 def : SELECTNegCondReg<setne, R32C, i32, SELBr32, CEQr32>;
3355 def : SELECTNegCondImm<setne, R32C, i32, i32ImmSExt10, SELBr32, CEQIr32>;
3356 def : SELECTNegCondReg<setle, R32C, i32, SELBr32, CGTr32>;
3357 def : SELECTNegCondImm<setle, R32C, i32, i32ImmSExt10, SELBr32, CGTIr32>;
3358 def : SELECTNegCondReg<setule, R32C, i32, SELBr32, CLGTr32>;
3359 def : SELECTNegCondImm<setule, R32C, i32, i32ImmSExt10, SELBr32, CLGTIr32>;
3361 class SELECTBinOpReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3362 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3364 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3365 rclass:$rTrue, rclass:$rFalse),
3366 (selinstr rclass:$rFalse, rclass:$rTrue,
3367 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3368 (cmpOp2 rclass:$rA, rclass:$rB)))>;
3370 class SELECTBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3372 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3374 Pat<(select (inttype (cond rclass:$rA, (inttype immpred:$imm))),
3375 rclass:$rTrue, rclass:$rFalse),
3376 (selinstr rclass:$rFalse, rclass:$rTrue,
3377 (binop (cmpOp1 rclass:$rA, (inttype immpred:$imm)),
3378 (cmpOp2 rclass:$rA, (inttype immpred:$imm))))>;
3380 def : SELECTBinOpReg<setge, R8C, i8, SELBr8, ORr8, CGTBr8, CEQBr8>;
3381 def : SELECTBinOpImm<setge, R8C, immSExt8, i8,
3382 SELBr8, ORr8, CGTBIr8, CEQBIr8>;
3384 def : SELECTBinOpReg<setge, R16C, i16, SELBr16, ORr16, CGTHr16, CEQHr16>;
3385 def : SELECTBinOpImm<setge, R16C, i16ImmSExt10, i16,
3386 SELBr16, ORr16, CGTHIr16, CEQHIr16>;
3388 def : SELECTBinOpReg<setge, R32C, i32, SELBr32, ORr32, CGTr32, CEQr32>;
3389 def : SELECTBinOpImm<setge, R32C, i32ImmSExt10, i32,
3390 SELBr32, ORr32, CGTIr32, CEQIr32>;
3392 def : SELECTBinOpReg<setuge, R8C, i8, SELBr8, ORr8, CLGTBr8, CEQBr8>;
3393 def : SELECTBinOpImm<setuge, R8C, immSExt8, i8,
3394 SELBr8, ORr8, CLGTBIr8, CEQBIr8>;
3396 def : SELECTBinOpReg<setuge, R16C, i16, SELBr16, ORr16, CLGTHr16, CEQHr16>;
3397 def : SELECTBinOpImm<setuge, R16C, i16ImmUns10, i16,
3398 SELBr16, ORr16, CLGTHIr16, CEQHIr16>;
3400 def : SELECTBinOpReg<setuge, R32C, i32, SELBr32, ORr32, CLGTr32, CEQr32>;
3401 def : SELECTBinOpImm<setuge, R32C, i32ImmUns10, i32,
3402 SELBr32, ORr32, CLGTIr32, CEQIr32>;
3404 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3407 // All calls clobber the non-callee-saved registers:
3408 Defs = [R0, R1, R2, R3, R4, R5, R6, R7, R8, R9,
3409 R10,R11,R12,R13,R14,R15,R16,R17,R18,R19,
3410 R20,R21,R22,R23,R24,R25,R26,R27,R28,R29,
3411 R30,R31,R32,R33,R34,R35,R36,R37,R38,R39,
3412 R40,R41,R42,R43,R44,R45,R46,R47,R48,R49,
3413 R50,R51,R52,R53,R54,R55,R56,R57,R58,R59,
3414 R60,R61,R62,R63,R64,R65,R66,R67,R68,R69,
3415 R70,R71,R72,R73,R74,R75,R76,R77,R78,R79],
3416 // All of these instructions use $lr (aka $0)
3418 // Branch relative and set link: Used if we actually know that the target
3419 // is within [-32768, 32767] bytes of the target
3421 BranchSetLink<0b011001100, (outs), (ins relcalltarget:$func, variable_ops),
3422 "brsl\t$$lr, $func",
3423 [(SPUcall (SPUpcrel tglobaladdr:$func, 0))]>;
3425 // Branch absolute and set link: Used if we actually know that the target
3426 // is an absolute address
3428 BranchSetLink<0b011001100, (outs), (ins calltarget:$func, variable_ops),
3429 "brasl\t$$lr, $func",
3430 [(SPUcall (SPUaform tglobaladdr:$func, 0))]>;
3432 // Branch indirect and set link if external data. These instructions are not
3433 // actually generated, matched by an intrinsic:
3434 def BISLED_00: BISLEDForm<0b11, "bisled\t$$lr, $func", [/* empty pattern */]>;
3435 def BISLED_E0: BISLEDForm<0b10, "bisled\t$$lr, $func", [/* empty pattern */]>;
3436 def BISLED_0D: BISLEDForm<0b01, "bisled\t$$lr, $func", [/* empty pattern */]>;
3437 def BISLED_ED: BISLEDForm<0b00, "bisled\t$$lr, $func", [/* empty pattern */]>;
3439 // Branch indirect and set link. This is the "X-form" address version of a
3442 BIForm<0b10010101100, "bisl\t$$lr, $func", [(SPUcall R32C:$func)]>;
3445 // Unconditional branches:
3446 let isBranch = 1, isTerminator = 1, hasCtrlDep = 1, isBarrier = 1 in {
3448 UncondBranch<0b001001100, (outs), (ins brtarget:$dest),
3452 // Unconditional, absolute address branch
3454 UncondBranch<0b001100000, (outs), (ins brtarget:$dest),
3456 [/* no pattern */]>;
3460 BIForm<0b00010101100, "bi\t$func", [(brind R32C:$func)]>;
3462 // Various branches:
3463 class BRNZInst<dag IOL, list<dag> pattern>:
3464 RI16Form<0b010000100, (outs), IOL, "brnz\t$rCond,$dest",
3465 BranchResolv, pattern>;
3467 class BRNZRegInst<RegisterClass rclass>:
3468 BRNZInst<(ins rclass:$rCond, brtarget:$dest),
3469 [(brcond rclass:$rCond, bb:$dest)]>;
3471 class BRNZVecInst<ValueType vectype>:
3472 BRNZInst<(ins VECREG:$rCond, brtarget:$dest),
3473 [(brcond (vectype VECREG:$rCond), bb:$dest)]>;
3475 multiclass BranchNotZero {
3476 def v4i32 : BRNZVecInst<v4i32>;
3477 def r32 : BRNZRegInst<R32C>;
3480 defm BRNZ : BranchNotZero;
3482 class BRZInst<dag IOL, list<dag> pattern>:
3483 RI16Form<0b000000100, (outs), IOL, "brz\t$rT,$dest",
3484 BranchResolv, pattern>;
3486 class BRZRegInst<RegisterClass rclass>:
3487 BRZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3489 class BRZVecInst<ValueType vectype>:
3490 BRZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3492 multiclass BranchZero {
3493 def v4i32: BRZVecInst<v4i32>;
3494 def r32: BRZRegInst<R32C>;
3497 defm BRZ: BranchZero;
3499 // Note: LLVM doesn't do branch conditional, indirect. Otherwise these would
3502 class BINZInst<dag IOL, list<dag> pattern>:
3503 BICondForm<0b10010100100, (outs), IOL, "binz\t$rA, $dest", pattern>;
3505 class BINZRegInst<RegisterClass rclass>:
3506 BINZInst<(ins rclass:$rA, brtarget:$dest),
3507 [(brcond rclass:$rA, R32C:$dest)]>;
3509 class BINZVecInst<ValueType vectype>:
3510 BINZInst<(ins VECREG:$rA, R32C:$dest),
3511 [(brcond (vectype VECREG:$rA), R32C:$dest)]>;
3513 multiclass BranchNotZeroIndirect {
3514 def v4i32: BINZVecInst<v4i32>;
3515 def r32: BINZRegInst<R32C>;
3518 defm BINZ: BranchNotZeroIndirect;
3520 class BIZInst<dag IOL, list<dag> pattern>:
3521 BICondForm<0b00010100100, (outs), IOL, "biz\t$rA, $func", pattern>;
3523 class BIZRegInst<RegisterClass rclass>:
3524 BIZInst<(ins rclass:$rA, R32C:$func), [/* no pattern */]>;
3526 class BIZVecInst<ValueType vectype>:
3527 BIZInst<(ins VECREG:$rA, R32C:$func), [/* no pattern */]>;
3529 multiclass BranchZeroIndirect {
3530 def v4i32: BIZVecInst<v4i32>;
3531 def r32: BIZRegInst<R32C>;
3534 defm BIZ: BranchZeroIndirect;
3537 class BRHNZInst<dag IOL, list<dag> pattern>:
3538 RI16Form<0b011000100, (outs), IOL, "brhnz\t$rCond,$dest", BranchResolv,
3541 class BRHNZRegInst<RegisterClass rclass>:
3542 BRHNZInst<(ins rclass:$rCond, brtarget:$dest),
3543 [(brcond rclass:$rCond, bb:$dest)]>;
3545 class BRHNZVecInst<ValueType vectype>:
3546 BRHNZInst<(ins VECREG:$rCond, brtarget:$dest), [/* no pattern */]>;
3548 multiclass BranchNotZeroHalfword {
3549 def v8i16: BRHNZVecInst<v8i16>;
3550 def r16: BRHNZRegInst<R16C>;
3553 defm BRHNZ: BranchNotZeroHalfword;
3555 class BRHZInst<dag IOL, list<dag> pattern>:
3556 RI16Form<0b001000100, (outs), IOL, "brhz\t$rT,$dest", BranchResolv,
3559 class BRHZRegInst<RegisterClass rclass>:
3560 BRHZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3562 class BRHZVecInst<ValueType vectype>:
3563 BRHZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3565 multiclass BranchZeroHalfword {
3566 def v8i16: BRHZVecInst<v8i16>;
3567 def r16: BRHZRegInst<R16C>;
3570 defm BRHZ: BranchZeroHalfword;
3573 //===----------------------------------------------------------------------===//
3574 // setcc and brcond patterns:
3575 //===----------------------------------------------------------------------===//
3577 def : Pat<(brcond (i16 (seteq R16C:$rA, 0)), bb:$dest),
3578 (BRHZr16 R16C:$rA, bb:$dest)>;
3579 def : Pat<(brcond (i16 (setne R16C:$rA, 0)), bb:$dest),
3580 (BRHNZr16 R16C:$rA, bb:$dest)>;
3582 def : Pat<(brcond (i32 (seteq R32C:$rA, 0)), bb:$dest),
3583 (BRZr32 R32C:$rA, bb:$dest)>;
3584 def : Pat<(brcond (i32 (setne R32C:$rA, 0)), bb:$dest),
3585 (BRNZr32 R32C:$rA, bb:$dest)>;
3587 multiclass BranchCondEQ<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3589 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3590 (brinst16 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3592 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3593 (brinst16 (CEQHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3595 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3596 (brinst32 (CEQIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3598 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3599 (brinst32 (CEQr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3602 defm BRCONDeq : BranchCondEQ<seteq, BRHZr16, BRZr32>;
3603 defm BRCONDne : BranchCondEQ<setne, BRHNZr16, BRNZr32>;
3605 multiclass BranchCondLGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3607 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3608 (brinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3610 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3611 (brinst16 (CLGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3613 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3614 (brinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3616 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3617 (brinst32 (CLGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3620 defm BRCONDugt : BranchCondLGT<setugt, BRHNZr16, BRNZr32>;
3621 defm BRCONDule : BranchCondLGT<setule, BRHZr16, BRZr32>;
3623 multiclass BranchCondLGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3624 SPUInstr orinst32, SPUInstr brinst32>
3626 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3627 (brinst16 (orinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3628 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3631 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3632 (brinst16 (orinst16 (CLGTHr16 R16C:$rA, R16:$rB),
3633 (CEQHr16 R16C:$rA, R16:$rB)),
3636 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3637 (brinst32 (orinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val),
3638 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3641 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3642 (brinst32 (orinst32 (CLGTr32 R32C:$rA, R32C:$rB),
3643 (CEQr32 R32C:$rA, R32C:$rB)),
3647 defm BRCONDuge : BranchCondLGTEQ<setuge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3648 defm BRCONDult : BranchCondLGTEQ<setult, ORr16, BRHZr16, ORr32, BRZr32>;
3650 multiclass BranchCondGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3652 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3653 (brinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3655 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3656 (brinst16 (CGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3658 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3659 (brinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3661 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3662 (brinst32 (CGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3665 defm BRCONDgt : BranchCondGT<setgt, BRHNZr16, BRNZr32>;
3666 defm BRCONDle : BranchCondGT<setle, BRHZr16, BRZr32>;
3668 multiclass BranchCondGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3669 SPUInstr orinst32, SPUInstr brinst32>
3671 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3672 (brinst16 (orinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3673 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3676 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3677 (brinst16 (orinst16 (CGTHr16 R16C:$rA, R16:$rB),
3678 (CEQHr16 R16C:$rA, R16:$rB)),
3681 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3682 (brinst32 (orinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val),
3683 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3686 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3687 (brinst32 (orinst32 (CGTr32 R32C:$rA, R32C:$rB),
3688 (CEQr32 R32C:$rA, R32C:$rB)),
3692 defm BRCONDge : BranchCondGTEQ<setge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3693 defm BRCONDlt : BranchCondGTEQ<setlt, ORr16, BRHZr16, ORr32, BRZr32>;
3695 let isTerminator = 1, isBarrier = 1 in {
3696 let isReturn = 1 in {
3698 RETForm<"bi\t$$lr", [(retflag)]>;
3702 //===----------------------------------------------------------------------===//
3703 // Single precision floating point instructions
3704 //===----------------------------------------------------------------------===//
3706 class FAInst<dag OOL, dag IOL, list<dag> pattern>:
3707 RRForm<0b01011000100, OOL, IOL, "fa\t$rT, $rA, $rB",
3710 class FAVecInst<ValueType vectype>:
3711 FAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3712 [(set (vectype VECREG:$rT),
3713 (fadd (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3717 def v4f32: FAVecInst<v4f32>;
3718 def r32: FAInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3719 [(set R32FP:$rT, (fadd R32FP:$rA, R32FP:$rB))]>;
3724 class FSInst<dag OOL, dag IOL, list<dag> pattern>:
3725 RRForm<0b01011000100, OOL, IOL, "fs\t$rT, $rA, $rB",
3728 class FSVecInst<ValueType vectype>:
3729 FSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3730 [(set (vectype VECREG:$rT),
3731 (fsub (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3735 def v4f32: FSVecInst<v4f32>;
3736 def r32: FSInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3737 [(set R32FP:$rT, (fsub R32FP:$rA, R32FP:$rB))]>;
3742 // Floating point reciprocal estimate
3744 RRForm_1<0b00011101100, (outs VECREG:$rT), (ins VECREG:$rA),
3745 "frest\t$rT, $rA", SPrecFP,
3746 [(set (v4f32 VECREG:$rT), (SPUreciprocalEst (v4f32 VECREG:$rA)))]>;
3749 RRForm_1<0b00011101100, (outs R32FP:$rT), (ins R32FP:$rA),
3750 "frest\t$rT, $rA", SPrecFP,
3751 [(set R32FP:$rT, (SPUreciprocalEst R32FP:$rA))]>;
3753 // Floating point interpolate (used in conjunction with reciprocal estimate)
3755 RRForm<0b00101011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3756 "fi\t$rT, $rA, $rB", SPrecFP,
3757 [(set (v4f32 VECREG:$rT), (SPUinterpolate (v4f32 VECREG:$rA),
3758 (v4f32 VECREG:$rB)))]>;
3761 RRForm<0b00101011110, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3762 "fi\t$rT, $rA, $rB", SPrecFP,
3763 [(set R32FP:$rT, (SPUinterpolate R32FP:$rA, R32FP:$rB))]>;
3765 //--------------------------------------------------------------------------
3766 // Basic single precision floating point comparisons:
3768 // Note: There is no support on SPU for single precision NaN. Consequently,
3769 // ordered and unordered comparisons are the same.
3770 //--------------------------------------------------------------------------
3773 RRForm<0b01000011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3774 "fceq\t$rT, $rA, $rB", SPrecFP,
3775 [(set R32C:$rT, (setueq R32FP:$rA, R32FP:$rB))]>;
3777 def : Pat<(setoeq R32FP:$rA, R32FP:$rB),
3778 (FCEQf32 R32FP:$rA, R32FP:$rB)>;
3781 RRForm<0b01010011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3782 "fcmeq\t$rT, $rA, $rB", SPrecFP,
3783 [(set R32C:$rT, (setueq (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
3785 def : Pat<(setoeq (fabs R32FP:$rA), (fabs R32FP:$rB)),
3786 (FCMEQf32 R32FP:$rA, R32FP:$rB)>;
3789 RRForm<0b01000011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3790 "fcgt\t$rT, $rA, $rB", SPrecFP,
3791 [(set R32C:$rT, (setugt R32FP:$rA, R32FP:$rB))]>;
3793 def : Pat<(setugt R32FP:$rA, R32FP:$rB),
3794 (FCGTf32 R32FP:$rA, R32FP:$rB)>;
3797 RRForm<0b01010011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3798 "fcmgt\t$rT, $rA, $rB", SPrecFP,
3799 [(set R32C:$rT, (setugt (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
3801 def : Pat<(setugt (fabs R32FP:$rA), (fabs R32FP:$rB)),
3802 (FCMGTf32 R32FP:$rA, R32FP:$rB)>;
3804 //--------------------------------------------------------------------------
3805 // Single precision floating point comparisons and SETCC equivalents:
3806 //--------------------------------------------------------------------------
3808 def : SETCCNegCondReg<setune, R32FP, i32, XORIr32, FCEQf32>;
3809 def : SETCCNegCondReg<setone, R32FP, i32, XORIr32, FCEQf32>;
3811 def : SETCCBinOpReg<setuge, R32FP, ORr32, FCGTf32, FCEQf32>;
3812 def : SETCCBinOpReg<setoge, R32FP, ORr32, FCGTf32, FCEQf32>;
3814 def : SETCCBinOpReg<setult, R32FP, NORr32, FCGTf32, FCEQf32>;
3815 def : SETCCBinOpReg<setolt, R32FP, NORr32, FCGTf32, FCEQf32>;
3817 def : Pat<(setule R32FP:$rA, R32FP:$rB),
3818 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
3819 def : Pat<(setole R32FP:$rA, R32FP:$rB),
3820 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
3822 // FP Status and Control Register Write
3823 // Why isn't rT a don't care in the ISA?
3824 // Should we create a special RRForm_3 for this guy and zero out the rT?
3826 RRForm_1<0b01011101110, (outs R32FP:$rT), (ins R32FP:$rA),
3827 "fscrwr\t$rA", SPrecFP,
3828 [/* This instruction requires an intrinsic. Note: rT is unused. */]>;
3830 // FP Status and Control Register Read
3832 RRForm_2<0b01011101110, (outs R32FP:$rT), (ins),
3833 "fscrrd\t$rT", SPrecFP,
3834 [/* This instruction requires an intrinsic */]>;
3836 // llvm instruction space
3837 // How do these map onto cell instructions?
3839 // frest rC rB # c = 1/b (both lines)
3841 // fm rD rA rC # d = a * 1/b
3842 // fnms rB rD rB rA # b = - (d * b - a) --should == 0 in a perfect world
3843 // fma rB rB rC rD # b = b * c + d
3844 // = -(d *b -a) * c + d
3845 // = a * c - c ( a *b *c - a)
3850 // These llvm instructions will actually map to library calls.
3851 // All that's needed, then, is to check that the appropriate library is
3852 // imported and do a brsl to the proper function name.
3853 // frem # fmod(x, y): x - (x/y) * y
3854 // (Note: fmod(double, double), fmodf(float,float)
3858 // Unimplemented SPU instruction space
3859 // floating reciprocal absolute square root estimate (frsqest)
3861 // The following are probably just intrinsics
3862 // status and control register write
3863 // status and control register read
3865 //--------------------------------------
3866 // Floating point multiply instructions
3867 //--------------------------------------
3870 RRForm<0b00100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3871 "fm\t$rT, $rA, $rB", SPrecFP,
3872 [(set (v4f32 VECREG:$rT), (fmul (v4f32 VECREG:$rA),
3873 (v4f32 VECREG:$rB)))]>;
3876 RRForm<0b01100011010, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3877 "fm\t$rT, $rA, $rB", SPrecFP,
3878 [(set R32FP:$rT, (fmul R32FP:$rA, R32FP:$rB))]>;
3880 // Floating point multiply and add
3881 // e.g. d = c + (a * b)
3883 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
3884 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
3885 [(set (v4f32 VECREG:$rT),
3886 (fadd (v4f32 VECREG:$rC),
3887 (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB))))]>;
3890 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
3891 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
3892 [(set R32FP:$rT, (fadd R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
3894 // FP multiply and subtract
3895 // Subtracts value in rC from product
3898 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
3899 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
3900 [(set (v4f32 VECREG:$rT),
3901 (fsub (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB)),
3902 (v4f32 VECREG:$rC)))]>;
3905 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
3906 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
3908 (fsub (fmul R32FP:$rA, R32FP:$rB), R32FP:$rC))]>;
3910 // Floating Negative Mulitply and Subtract
3911 // Subtracts product from value in rC
3912 // res = fneg(fms a b c)
3915 // NOTE: subtraction order
3919 RRRForm<0b1101, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
3920 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
3921 [(set R32FP:$rT, (fsub R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
3924 RRRForm<0b1101, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
3925 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
3926 [(set (v4f32 VECREG:$rT),
3927 (fsub (v4f32 VECREG:$rC),
3928 (fmul (v4f32 VECREG:$rA),
3929 (v4f32 VECREG:$rB))))]>;
3931 //--------------------------------------
3932 // Floating Point Conversions
3933 // Signed conversions:
3935 CVTIntFPForm<0b0101101110, (outs VECREG:$rT), (ins VECREG:$rA),
3936 "csflt\t$rT, $rA, 0", SPrecFP,
3937 [(set (v4f32 VECREG:$rT), (sint_to_fp (v4i32 VECREG:$rA)))]>;
3939 // Convert signed integer to floating point
3941 CVTIntFPForm<0b0101101110, (outs R32FP:$rT), (ins R32C:$rA),
3942 "csflt\t$rT, $rA, 0", SPrecFP,
3943 [(set R32FP:$rT, (sint_to_fp R32C:$rA))]>;
3945 // Convert unsigned into to float
3947 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
3948 "cuflt\t$rT, $rA, 0", SPrecFP,
3949 [(set (v4f32 VECREG:$rT), (uint_to_fp (v4i32 VECREG:$rA)))]>;
3952 CVTIntFPForm<0b1101101110, (outs R32FP:$rT), (ins R32C:$rA),
3953 "cuflt\t$rT, $rA, 0", SPrecFP,
3954 [(set R32FP:$rT, (uint_to_fp R32C:$rA))]>;
3956 // Convert float to unsigned int
3957 // Assume that scale = 0
3960 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
3961 "cfltu\t$rT, $rA, 0", SPrecFP,
3962 [(set (v4i32 VECREG:$rT), (fp_to_uint (v4f32 VECREG:$rA)))]>;
3965 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
3966 "cfltu\t$rT, $rA, 0", SPrecFP,
3967 [(set R32C:$rT, (fp_to_uint R32FP:$rA))]>;
3969 // Convert float to signed int
3970 // Assume that scale = 0
3973 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
3974 "cflts\t$rT, $rA, 0", SPrecFP,
3975 [(set (v4i32 VECREG:$rT), (fp_to_sint (v4f32 VECREG:$rA)))]>;
3978 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
3979 "cflts\t$rT, $rA, 0", SPrecFP,
3980 [(set R32C:$rT, (fp_to_sint R32FP:$rA))]>;
3982 //===----------------------------------------------------------------------==//
3983 // Single<->Double precision conversions
3984 //===----------------------------------------------------------------------==//
3986 // NOTE: We use "vec" name suffix here to avoid confusion (e.g. input is a
3987 // v4f32, output is v2f64--which goes in the name?)
3989 // Floating point extend single to double
3990 // NOTE: Not sure if passing in v4f32 to FESDvec is correct since it
3991 // operates on two double-word slots (i.e. 1st and 3rd fp numbers
3994 RRForm_1<0b00011101110, (outs VECREG:$rT), (ins VECREG:$rA),
3995 "fesd\t$rT, $rA", SPrecFP,
3996 [(set (v2f64 VECREG:$rT), (fextend (v4f32 VECREG:$rA)))]>;
3999 RRForm_1<0b00011101110, (outs R64FP:$rT), (ins R32FP:$rA),
4000 "fesd\t$rT, $rA", SPrecFP,
4001 [(set R64FP:$rT, (fextend R32FP:$rA))]>;
4003 // Floating point round double to single
4005 // RRForm_1<0b10011101110, (outs VECREG:$rT), (ins VECREG:$rA),
4006 // "frds\t$rT, $rA,", SPrecFP,
4007 // [(set (v4f32 R32FP:$rT), (fround (v2f64 R64FP:$rA)))]>;
4010 RRForm_1<0b10011101110, (outs R32FP:$rT), (ins R64FP:$rA),
4011 "frds\t$rT, $rA", SPrecFP,
4012 [(set R32FP:$rT, (fround R64FP:$rA))]>;
4014 //ToDo include anyextend?
4016 //===----------------------------------------------------------------------==//
4017 // Double precision floating point instructions
4018 //===----------------------------------------------------------------------==//
4020 RRForm<0b00110011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4021 "dfa\t$rT, $rA, $rB", DPrecFP,
4022 [(set R64FP:$rT, (fadd R64FP:$rA, R64FP:$rB))]>;
4025 RRForm<0b00110011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4026 "dfa\t$rT, $rA, $rB", DPrecFP,
4027 [(set (v2f64 VECREG:$rT), (fadd (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4030 RRForm<0b10100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4031 "dfs\t$rT, $rA, $rB", DPrecFP,
4032 [(set R64FP:$rT, (fsub R64FP:$rA, R64FP:$rB))]>;
4035 RRForm<0b10100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4036 "dfs\t$rT, $rA, $rB", DPrecFP,
4037 [(set (v2f64 VECREG:$rT),
4038 (fsub (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4041 RRForm<0b01100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4042 "dfm\t$rT, $rA, $rB", DPrecFP,
4043 [(set R64FP:$rT, (fmul R64FP:$rA, R64FP:$rB))]>;
4046 RRForm<0b00100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4047 "dfm\t$rT, $rA, $rB", DPrecFP,
4048 [(set (v2f64 VECREG:$rT),
4049 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4052 RRForm<0b00111010110, (outs R64FP:$rT),
4053 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4054 "dfma\t$rT, $rA, $rB", DPrecFP,
4055 [(set R64FP:$rT, (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB)))]>,
4056 RegConstraint<"$rC = $rT">,
4060 RRForm<0b00111010110, (outs VECREG:$rT),
4061 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4062 "dfma\t$rT, $rA, $rB", DPrecFP,
4063 [(set (v2f64 VECREG:$rT),
4064 (fadd (v2f64 VECREG:$rC),
4065 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB))))]>,
4066 RegConstraint<"$rC = $rT">,
4070 RRForm<0b10111010110, (outs R64FP:$rT),
4071 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4072 "dfms\t$rT, $rA, $rB", DPrecFP,
4073 [(set R64FP:$rT, (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC))]>,
4074 RegConstraint<"$rC = $rT">,
4078 RRForm<0b10111010110, (outs VECREG:$rT),
4079 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4080 "dfms\t$rT, $rA, $rB", DPrecFP,
4081 [(set (v2f64 VECREG:$rT),
4082 (fsub (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)),
4083 (v2f64 VECREG:$rC)))]>;
4085 // FNMS: - (a * b - c)
4086 // - (a * b) + c => c - (a * b)
4088 RRForm<0b01111010110, (outs R64FP:$rT),
4089 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4090 "dfnms\t$rT, $rA, $rB", DPrecFP,
4091 [(set R64FP:$rT, (fsub R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB)))]>,
4092 RegConstraint<"$rC = $rT">,
4095 def : Pat<(fneg (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC)),
4096 (FNMSf64 R64FP:$rA, R64FP:$rB, R64FP:$rC)>;
4099 RRForm<0b01111010110, (outs VECREG:$rT),
4100 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4101 "dfnms\t$rT, $rA, $rB", DPrecFP,
4102 [(set (v2f64 VECREG:$rT),
4103 (fsub (v2f64 VECREG:$rC),
4104 (fmul (v2f64 VECREG:$rA),
4105 (v2f64 VECREG:$rB))))]>,
4106 RegConstraint<"$rC = $rT">,
4109 def : Pat<(fneg (fsub (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)),
4110 (v2f64 VECREG:$rC))),
4111 (FNMSv2f64 VECREG:$rA, VECREG:$rB, VECREG:$rC)>;
4116 RRForm<0b11111010110, (outs R64FP:$rT),
4117 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4118 "dfnma\t$rT, $rA, $rB", DPrecFP,
4119 [(set R64FP:$rT, (fneg (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB))))]>,
4120 RegConstraint<"$rC = $rT">,
4124 RRForm<0b11111010110, (outs VECREG:$rT),
4125 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4126 "dfnma\t$rT, $rA, $rB", DPrecFP,
4127 [(set (v2f64 VECREG:$rT),
4128 (fneg (fadd (v2f64 VECREG:$rC),
4129 (fmul (v2f64 VECREG:$rA),
4130 (v2f64 VECREG:$rB)))))]>,
4131 RegConstraint<"$rC = $rT">,
4134 //===----------------------------------------------------------------------==//
4135 // Floating point negation and absolute value
4136 //===----------------------------------------------------------------------==//
4138 def : Pat<(fneg (v4f32 VECREG:$rA)),
4139 (XORfnegvec (v4f32 VECREG:$rA),
4140 (v4f32 (ILHUv4i32 0x8000)))>;
4142 def : Pat<(fneg R32FP:$rA),
4143 (XORfneg32 R32FP:$rA, (ILHUr32 0x8000))>;
4145 def : Pat<(fneg (v2f64 VECREG:$rA)),
4146 (XORfnegvec (v2f64 VECREG:$rA),
4147 (v2f64 (ANDBIv16i8 (FSMBIv16i8 0x8080), 0x80)))>;
4149 def : Pat<(fneg R64FP:$rA),
4150 (XORfneg64 R64FP:$rA,
4151 (ANDBIv16i8 (FSMBIv16i8 0x8080), 0x80))>;
4153 // Floating point absolute value
4155 def : Pat<(fabs R32FP:$rA),
4156 (ANDfabs32 R32FP:$rA, (IOHLr32 (ILHUr32 0x7fff), 0xffff))>;
4158 def : Pat<(fabs (v4f32 VECREG:$rA)),
4159 (ANDfabsvec (v4f32 VECREG:$rA),
4160 (v4f32 (ANDBIv16i8 (FSMBIv16i8 0xffff), 0x7f)))>;
4162 def : Pat<(fabs R64FP:$rA),
4163 (ANDfabs64 R64FP:$rA, (ANDBIv16i8 (FSMBIv16i8 0xffff), 0x7f))>;
4165 def : Pat<(fabs (v2f64 VECREG:$rA)),
4166 (ANDfabsvec (v2f64 VECREG:$rA),
4167 (v2f64 (ANDBIv16i8 (FSMBIv16i8 0xffff), 0x7f)))>;
4169 //===----------------------------------------------------------------------===//
4170 // Hint for branch instructions:
4171 //===----------------------------------------------------------------------===//
4173 /* def HBR : SPUInstr<(outs), (ins), "hbr\t" */
4175 //===----------------------------------------------------------------------===//
4176 // Execution, Load NOP (execute NOPs belong in even pipeline, load NOPs belong
4177 // in the odd pipeline)
4178 //===----------------------------------------------------------------------===//
4180 def ENOP : SPUInstr<(outs), (ins), "enop", ExecNOP> {
4183 let Inst{0-10} = 0b10000000010;
4184 let Inst{11-17} = 0;
4185 let Inst{18-24} = 0;
4186 let Inst{25-31} = 0;
4189 def LNOP : SPUInstr<(outs), (ins), "lnop", LoadNOP> {
4192 let Inst{0-10} = 0b10000000000;
4193 let Inst{11-17} = 0;
4194 let Inst{18-24} = 0;
4195 let Inst{25-31} = 0;
4198 //===----------------------------------------------------------------------===//
4199 // Bit conversions (type conversions between vector/packed types)
4200 // NOTE: Promotions are handled using the XS* instructions. Truncation
4202 //===----------------------------------------------------------------------===//
4203 def : Pat<(v16i8 (bitconvert (v8i16 VECREG:$src))), (v16i8 VECREG:$src)>;
4204 def : Pat<(v16i8 (bitconvert (v4i32 VECREG:$src))), (v16i8 VECREG:$src)>;
4205 def : Pat<(v16i8 (bitconvert (v2i64 VECREG:$src))), (v16i8 VECREG:$src)>;
4206 def : Pat<(v16i8 (bitconvert (v4f32 VECREG:$src))), (v16i8 VECREG:$src)>;
4207 def : Pat<(v16i8 (bitconvert (v2f64 VECREG:$src))), (v16i8 VECREG:$src)>;
4209 def : Pat<(v8i16 (bitconvert (v16i8 VECREG:$src))), (v8i16 VECREG:$src)>;
4210 def : Pat<(v8i16 (bitconvert (v4i32 VECREG:$src))), (v8i16 VECREG:$src)>;
4211 def : Pat<(v8i16 (bitconvert (v2i64 VECREG:$src))), (v8i16 VECREG:$src)>;
4212 def : Pat<(v8i16 (bitconvert (v4f32 VECREG:$src))), (v8i16 VECREG:$src)>;
4213 def : Pat<(v8i16 (bitconvert (v2f64 VECREG:$src))), (v8i16 VECREG:$src)>;
4215 def : Pat<(v4i32 (bitconvert (v16i8 VECREG:$src))), (v4i32 VECREG:$src)>;
4216 def : Pat<(v4i32 (bitconvert (v8i16 VECREG:$src))), (v4i32 VECREG:$src)>;
4217 def : Pat<(v4i32 (bitconvert (v2i64 VECREG:$src))), (v4i32 VECREG:$src)>;
4218 def : Pat<(v4i32 (bitconvert (v4f32 VECREG:$src))), (v4i32 VECREG:$src)>;
4219 def : Pat<(v4i32 (bitconvert (v2f64 VECREG:$src))), (v4i32 VECREG:$src)>;
4221 def : Pat<(v2i64 (bitconvert (v16i8 VECREG:$src))), (v2i64 VECREG:$src)>;
4222 def : Pat<(v2i64 (bitconvert (v8i16 VECREG:$src))), (v2i64 VECREG:$src)>;
4223 def : Pat<(v2i64 (bitconvert (v4i32 VECREG:$src))), (v2i64 VECREG:$src)>;
4224 def : Pat<(v2i64 (bitconvert (v4f32 VECREG:$src))), (v2i64 VECREG:$src)>;
4225 def : Pat<(v2i64 (bitconvert (v2f64 VECREG:$src))), (v2i64 VECREG:$src)>;
4227 def : Pat<(v4f32 (bitconvert (v16i8 VECREG:$src))), (v4f32 VECREG:$src)>;
4228 def : Pat<(v4f32 (bitconvert (v8i16 VECREG:$src))), (v4f32 VECREG:$src)>;
4229 def : Pat<(v4f32 (bitconvert (v2i64 VECREG:$src))), (v4f32 VECREG:$src)>;
4230 def : Pat<(v4f32 (bitconvert (v4i32 VECREG:$src))), (v4f32 VECREG:$src)>;
4231 def : Pat<(v4f32 (bitconvert (v2f64 VECREG:$src))), (v4f32 VECREG:$src)>;
4233 def : Pat<(v2f64 (bitconvert (v16i8 VECREG:$src))), (v2f64 VECREG:$src)>;
4234 def : Pat<(v2f64 (bitconvert (v8i16 VECREG:$src))), (v2f64 VECREG:$src)>;
4235 def : Pat<(v2f64 (bitconvert (v4i32 VECREG:$src))), (v2f64 VECREG:$src)>;
4236 def : Pat<(v2f64 (bitconvert (v2i64 VECREG:$src))), (v2f64 VECREG:$src)>;
4237 def : Pat<(v2f64 (bitconvert (v2f64 VECREG:$src))), (v2f64 VECREG:$src)>;
4239 def : Pat<(f32 (bitconvert (i32 R32C:$src))), (f32 R32FP:$src)>;
4240 def : Pat<(f64 (bitconvert (i64 R64C:$src))), (f64 R64FP:$src)>;
4242 //===----------------------------------------------------------------------===//
4243 // Instruction patterns:
4244 //===----------------------------------------------------------------------===//
4246 // General 32-bit constants:
4247 def : Pat<(i32 imm:$imm),
4248 (IOHLr32 (ILHUr32 (HI16 imm:$imm)), (LO16 imm:$imm))>;
4250 // Single precision float constants:
4251 def : Pat<(f32 fpimm:$imm),
4252 (IOHLf32 (ILHUf32 (HI16_f32 fpimm:$imm)), (LO16_f32 fpimm:$imm))>;
4254 // General constant 32-bit vectors
4255 def : Pat<(v4i32 v4i32Imm:$imm),
4256 (IOHLv4i32 (v4i32 (ILHUv4i32 (HI16_vec v4i32Imm:$imm))),
4257 (LO16_vec v4i32Imm:$imm))>;
4260 def : Pat<(i8 imm:$imm),
4263 //===----------------------------------------------------------------------===//
4264 // Call instruction patterns:
4265 //===----------------------------------------------------------------------===//
4270 //===----------------------------------------------------------------------===//
4271 // Zero/Any/Sign extensions
4272 //===----------------------------------------------------------------------===//
4274 // sext 8->32: Sign extend bytes to words
4275 def : Pat<(sext_inreg R32C:$rSrc, i8),
4276 (XSHWr32 (XSBHr32 R32C:$rSrc))>;
4278 def : Pat<(i32 (sext R8C:$rSrc)),
4279 (XSHWr16 (XSBHr8 R8C:$rSrc))>;
4281 // zext 8->16: Zero extend bytes to halfwords
4282 def : Pat<(i16 (zext R8C:$rSrc)),
4283 (ANDHIi8i16 R8C:$rSrc, 0xff)>;
4285 // zext 8->32: Zero extend bytes to words
4286 def : Pat<(i32 (zext R8C:$rSrc)),
4287 (ANDIi8i32 R8C:$rSrc, 0xff)>;
4289 // anyext 8->16: Extend 8->16 bits, irrespective of sign
4290 def : Pat<(i16 (anyext R8C:$rSrc)),
4291 (ORHIi8i16 R8C:$rSrc, 0)>;
4293 // anyext 8->32: Extend 8->32 bits, irrespective of sign
4294 def : Pat<(i32 (anyext R8C:$rSrc)),
4295 (ORIi8i32 R8C:$rSrc, 0)>;
4297 // zext 16->32: Zero extend halfwords to words
4298 def : Pat<(i32 (zext R16C:$rSrc)),
4299 (ANDi16i32 R16C:$rSrc, (ILAr32 0xffff))>;
4301 def : Pat<(i32 (zext (and R16C:$rSrc, 0xf))),
4302 (ANDIi16i32 R16C:$rSrc, 0xf)>;
4304 def : Pat<(i32 (zext (and R16C:$rSrc, 0xff))),
4305 (ANDIi16i32 R16C:$rSrc, 0xff)>;
4307 def : Pat<(i32 (zext (and R16C:$rSrc, 0xfff))),
4308 (ANDIi16i32 R16C:$rSrc, 0xfff)>;
4310 // anyext 16->32: Extend 16->32 bits, irrespective of sign
4311 def : Pat<(i32 (anyext R16C:$rSrc)),
4312 (ORIi16i32 R16C:$rSrc, 0)>;
4314 //===----------------------------------------------------------------------===//
4316 // These truncates are for the SPU's supported types (i8, i16, i32). i64 and
4317 // above are custom lowered.
4318 //===----------------------------------------------------------------------===//
4320 def : Pat<(i8 (trunc GPRC:$src)),
4322 (SHUFBgprc GPRC:$src, GPRC:$src,
4323 (IOHLv4i32 (ILHUv4i32 0x0f0f), 0x0f0f)))>;
4325 def : Pat<(i8 (trunc R64C:$src)),
4328 (ORv2i64_i64 R64C:$src),
4329 (ORv2i64_i64 R64C:$src),
4330 (IOHLv4i32 (ILHUv4i32 0x0707), 0x0707)))>;
4332 def : Pat<(i8 (trunc R32C:$src)),
4335 (ORv4i32_i32 R32C:$src),
4336 (ORv4i32_i32 R32C:$src),
4337 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)))>;
4339 def : Pat<(i8 (trunc R16C:$src)),
4342 (ORv8i16_i16 R16C:$src),
4343 (ORv8i16_i16 R16C:$src),
4344 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)))>;
4346 def : Pat<(i16 (trunc GPRC:$src)),
4348 (SHUFBgprc GPRC:$src, GPRC:$src,
4349 (IOHLv4i32 (ILHUv4i32 0x0e0f), 0x0e0f)))>;
4351 def : Pat<(i16 (trunc R64C:$src)),
4354 (ORv2i64_i64 R64C:$src),
4355 (ORv2i64_i64 R64C:$src),
4356 (IOHLv4i32 (ILHUv4i32 0x0607), 0x0607)))>;
4358 def : Pat<(i16 (trunc R32C:$src)),
4361 (ORv4i32_i32 R32C:$src),
4362 (ORv4i32_i32 R32C:$src),
4363 (IOHLv4i32 (ILHUv4i32 0x0203), 0x0203)))>;
4365 def : Pat<(i32 (trunc GPRC:$src)),
4367 (SHUFBgprc GPRC:$src, GPRC:$src,
4368 (IOHLv4i32 (ILHUv4i32 0x0c0d), 0x0e0f)))>;
4370 def : Pat<(i32 (trunc R64C:$src)),
4373 (ORv2i64_i64 R64C:$src),
4374 (ORv2i64_i64 R64C:$src),
4375 (IOHLv4i32 (ILHUv4i32 0x0405), 0x0607)))>;
4377 //===----------------------------------------------------------------------===//
4378 // Address generation: SPU, like PPC, has to split addresses into high and
4379 // low parts in order to load them into a register.
4380 //===----------------------------------------------------------------------===//
4382 def : Pat<(SPUaform tglobaladdr:$in, 0), (ILAlsa tglobaladdr:$in)>;
4383 def : Pat<(SPUaform texternalsym:$in, 0), (ILAlsa texternalsym:$in)>;
4384 def : Pat<(SPUaform tjumptable:$in, 0), (ILAlsa tjumptable:$in)>;
4385 def : Pat<(SPUaform tconstpool:$in, 0), (ILAlsa tconstpool:$in)>;
4387 def : Pat<(SPUindirect (SPUhi tglobaladdr:$in, 0),
4388 (SPUlo tglobaladdr:$in, 0)),
4389 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4391 def : Pat<(SPUindirect (SPUhi texternalsym:$in, 0),
4392 (SPUlo texternalsym:$in, 0)),
4393 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4395 def : Pat<(SPUindirect (SPUhi tjumptable:$in, 0),
4396 (SPUlo tjumptable:$in, 0)),
4397 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4399 def : Pat<(SPUindirect (SPUhi tconstpool:$in, 0),
4400 (SPUlo tconstpool:$in, 0)),
4401 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4403 def : Pat<(SPUindirect R32C:$sp, i32ImmSExt10:$imm),
4404 (AIr32 R32C:$sp, i32ImmSExt10:$imm)>;
4406 def : Pat<(SPUindirect R32C:$sp, imm:$imm),
4408 (IOHLr32 (ILHUr32 (HI16 imm:$imm)), (LO16 imm:$imm)))>;
4410 def : Pat<(add (SPUhi tglobaladdr:$in, 0), (SPUlo tglobaladdr:$in, 0)),
4411 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4413 def : Pat<(add (SPUhi texternalsym:$in, 0), (SPUlo texternalsym:$in, 0)),
4414 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4416 def : Pat<(add (SPUhi tjumptable:$in, 0), (SPUlo tjumptable:$in, 0)),
4417 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4419 def : Pat<(add (SPUhi tconstpool:$in, 0), (SPUlo tconstpool:$in, 0)),
4420 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4423 include "CellSDKIntrinsics.td"
4424 // 64-bit "instructions"/support
4425 include "SPU64InstrInfo.td"